Docstoc

Something

Document Sample
Something Powered By Docstoc
					                               Effective December 6, 2006, this report has been made publicly available
                               in accordance with Section 734.3(b)(3) and published in accordance with
                               Section 734.7 of the U.S. Export Administration Regulations. As a result of
                               this publication, this report is subject to only copyright protection and does
                               not require any license agreement from EPRI. This notice supersedes the
                               export control restrictions and any proprietary licensed material notices
                               embedded in the document prior to publication.




Wind Energy Forecasting Technology Update: 2005

                     1011884
Wind Energy Forecasting Technology Update: 2005
                                                  1011884

                               Technical Update, March 2006




                                      EPRI Project Manager

                                               C. McGowin




                                    ELECTRIC POWER RESEARCH INSTITUTE
    3420 Hillview Avenue, Palo Alto, California 94304-1395 ▪ PO Box 10412, Palo Alto, California 94303-0813 ▪ USA
                           800.313.3774 ▪ 650.855.2121 ▪ askepri@epri.com ▪ www.epri.com
DISCLAIMER OF WARRANTIES AND LIMITATION OF LIABILITIES
THIS DOCUMENT WAS PREPARED BY THE ORGANIZATION(S) NAMED BELOW AS AN ACCOUNT OF
WORK SPONSORED OR COSPONSORED BY THE ELECTRIC POWER RESEARCH INSTITUTE, INC. (EPRI).
NEITHER EPRI, ANY MEMBER OF EPRI, ANY COSPONSOR, THE ORGANIZATION(S) BELOW, NOR ANY
PERSON ACTING ON BEHALF OF ANY OF THEM:

(A) MAKES ANY WARRANTY OR REPRESENTATION WHATSOEVER, EXPRESS OR IMPLIED, (I) WITH
RESPECT TO THE USE OF ANY INFORMATION, APPARATUS, METHOD, PROCESS, OR SIMILAR ITEM
DISCLOSED IN THIS DOCUMENT, INCLUDING MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE, OR (II) THAT SUCH USE DOES NOT INFRINGE ON OR INTERFERE WITH PRIVATELY OWNED
RIGHTS, INCLUDING ANY PARTY'S INTELLECTUAL PROPERTY, OR (III) THAT THIS DOCUMENT IS
SUITABLE TO ANY PARTICULAR USER'S CIRCUMSTANCE; OR

(B) ASSUMES RESPONSIBILITY FOR ANY DAMAGES OR OTHER LIABILITY WHATSOEVER (INCLUDING
ANY CONSEQUENTIAL DAMAGES, EVEN IF EPRI OR ANY EPRI REPRESENTATIVE HAS BEEN ADVISED
OF THE POSSIBILITY OF SUCH DAMAGES) RESULTING FROM YOUR SELECTION OR USE OF THIS
DOCUMENT OR ANY INFORMATION, APPARATUS, METHOD, PROCESS, OR SIMILAR ITEM DISCLOSED IN
THIS DOCUMENT.

ORGANIZATION(S) THAT PREPARED THIS DOCUMENT

Electric Power Research Institute, Inc.




This is an EPRI Technical Update report. A Technical Update report is intended as an informal report of
continuing research, a meeting, or a topical study. It is not a final EPRI technical report.




ORDERING INFORMATION
Requests for copies of this report should be directed to EPRI Orders and Conferences, 1355 Willow Way, Suite 278,
Concord, CA 94520. Toll-free number: 800.313.3774, press 2, or internally x5379; voice: 925.609.9169; fax:
925.609.1310.

Electric Power Research Institute and EPRI are registered service marks of the Electric Power Research Institute, Inc.

Copyright © 2006 Electric Power Research Institute, Inc. All rights reserved.
CITATIONS

This report was prepared by

Electric Power Research Institute, Inc.
3420 Hillview Avenue
Palo Alto, CA 94304

Principal Investigator
C. McGowin


This report describes research sponsored by EPRI.

The report is a corporate document that should be cited in the literature in the following manner:
Wind Energy Forecasting Technology Update: 2005, EPRI, Palo Alto, CA: 2006. 1011884.




                                                                                                iii
PRODUCT DESCRIPTION


The worldwide installed wind generation capacity increased by 25% and reached almost
60,000 MW worldwide during 2005. As wind capacity continues to grow and large regional
concentrations of wind generation emerge, utilities and regional transmission organizations will
increasingly need accurate same-day and next-day forecasts of wind energy generation to
dispatch system generation and transmission resource and anticipate rapid changes of wind
generation.

Results
The report summarizes selected results of the California Regional Wind Energy Forecasting
Project including (1) development and testing of a next-hour regional forecasting algorithm
based on a new two-stage artificial neural network (ANN) algorithm; (2) enhancement and
testing of the existing 48-hour forecasting algorithm; (3) wind tunnel and high-resolution
numerical modeling of wind flow over complex terrain to determine the plant-scale power curve;
and (4) development of the California Wind Energy Research Dataset (CARD). Significant
advances are reported in all four areas.

Challenges and Objectives
The major challenges of developing accurate wind energy forecasting systems include the
uncertainty of forecasting local weather conditions beyond a few hours, the complexity of
translating weather forecasts into the fine structure of wind flow over a wind plant site,
especially in complex terrain, and the lack of availability of real-time wind speed, direction,
energy generation, and wind turbine availability data from a network of meteorological stations
surrounding the wind plants. The objective is to summarize the results of the California Energy
Commission-EPRI Regional Wind Energy Forecasting System Development project completed
in December 2005. The project developed and tested wind energy forecasting algorithms for both
same-day and next-day hourly forecasts of wind speed and energy generation for the principal
wind resource areas of California and for five wind plants.

EPRI Perspective
Accurate same-day and next-day wind energy forecast will become especially important in the
future if development of new wind generation facilities continues at the current rapid pace and
results in large regional concentrations of wind generation like those that already exist in
California, Texas, the upper Midwest, and Europe. The utilities and regional transmission
organizations that operate the electricity system need accurate forecasts of hourly ramp rates
with sufficient lead time to be ready to adjust other generation up and down and compensate for
rapid changes of wind generation. EPRI is planning a new project to implement same-day and
next-day forecast algorithms such as those developed in the California Regional Wind Energy

                                                                                                   v
Forecasting Project completed in 2005 in a real-time forecast workstation. The objective is to
work with the system operators to refine the forecast technology, the method of delivery, and the
visual display of real-time forecasts to meet the needs of the operators.

Approach
Researchers developed a two-stage artificial neural network (ANN) algorithm to generate same-
day, next-hour forecasts of hourly or shorter-term wind energy generation based on real-time
wind resource and power generation data; they then tested a portion of the algorithm using an
autoregressive method. Researchers also conducted a screening evaluation of six potential
enhancements of the existing next-day, 48-hour forecast algorithm tested in the previous
EPRI-CEC wind energy forecasting project and tested the most promising enhancements over a
12-month period. Researchers also investigated use of wind tunnel and high-resolution modeling
of wind flow over complex terrain to develop a more accurate plant-scale power curve used in
the wind forecasting algorithm. Finally, researchers developed the California Wind Generation
Research Dataset (CARD). CARD contains one year of simulated hourly wind speed, direction,
specific energy, and other data at multiple elevations over two horizontal grids with 5-km grid
point spacing, one over northern California and one over southern California.

Key Words
Wind power
Wind energy forecasting
Weather modeling
Artificial neural networks
Wind integration impacts




vi
ABSTRACT

The worldwide installed wind generation capacity increased by 25% and reached almost
60,000 MW worldwide during 2005. As wind capacity continues to grow and large regional
concentrations of wind generation emerge, utilities and regional transmission organizations will
increasingly need accurate same-day and next-day forecasts of wind energy generation to
dispatch system generation and transmission resource and anticipate rapid changes of wind
generation.
The project objective is to summarize the results of the California Energy Commission-EPRI
Regional Wind Energy Forecasting System Development project completed in December 2005.
The project developed and tested wind energy forecasting algorithms for both same-day and
next-day hourly forecasts of wind speed and energy generation for the principal wind resource
areas of California and for five wind plants. The report summarizes selected results including
(1) development and testing of a next-hour regional forecasting algorithm based on a new
two-stage artificial neural network (ANN) algorithm; (2) enhancement and testing of the existing
48-hour forecast algorithm used in the previous CEC-EPI project completed in 2002; (3) wind
tunnel and high-resolution numerical modeling of wind flow over complex terrain to determine
the plant-scale power curve; and (4) development of the California Wind Generation Research
Dataset (CARD). Significant advances are reported in all four areas.




                                                                                               vii
ACKNOWLEDGEMENTS



We gratefully acknowledge several individuals and organizations who contributed to the
CEC-EPRI California Regional Wind Energy Forecast System Development project. They
include Dr. Dora Yen Nakafuji and Michael Kane at the California Energy Commission;
David Hawkins at the California Independent System Operator; Dr. Marc Schwartz at the
National Renewable Energy Laboratory; Dr. Robert Farber at Southern California Edison; John
Zack at AWS Truewind, LLC; Professor Bruce White and Dr. David Lubitz at the University of
California at Davis; and Dr. Steve Chin at Lawrence Livermore National Laboratory.




                                                                                          ix
CONTENTS


1 INTRODUCTION ....................................................................................................................1-1
    Objectives and Scope ...........................................................................................................1-1

2 WIND ENERGY FORECASTING SYSTEMS AND DEVELOPERS.......................................2-1
    Wind Energy Forecasting Providers......................................................................................2-1
    Risoe National Laboratory – Prediktor ..................................................................................2-2
        Numerical Weather Prediction Data .................................................................................2-2
        WAsP................................................................................................................................2-3
        PARK................................................................................................................................2-4
        Model Output Statistics (MOS) .........................................................................................2-4
        Meteorological Forecast Data Retrieval System ..............................................................2-4
        Power Prediction System .................................................................................................2-5
        Meteorological Forecast Data Archive..............................................................................2-5
        Prediction Output and Web-Display .................................................................................2-5
        Example Forecast Output.................................................................................................2-6
    AWS Truewind - eWind .........................................................................................................2-6
        Physics-Based Atmospheric Numerical Models ...............................................................2-7
        Adaptive Statistical Models...............................................................................................2-9
        Plant Output Models .........................................................................................................2-9
        Forecast Delivery System...............................................................................................2-10
    Applied Modeling - WEFS ...................................................................................................2-10
        Mesoscale Model MM5...................................................................................................2-11
        Diagnostic Wind Model...................................................................................................2-11
        Adaptive Statistical Model ..............................................................................................2-12
        Forecast Access by Users..............................................................................................2-12
    3TIER Environmental Forecast Group ................................................................................2-12
        Hour-Ahead Forecasting ................................................................................................2-13
        Day-Ahead Forecasting..................................................................................................2-13

                                                                                                                                            xi
3 WIND ENERGY FORECASTING INTEGRATION INTO ELECTRICITY GRID
OPERATIONS ...........................................................................................................................3-1
      Characteristics of Wind Generation.......................................................................................3-1
      How Wind Generation Affects Electricity Grid Operations ....................................................3-2
          System Operators and Balancing Authorities...................................................................3-2
          Area Control Error (ACE)..................................................................................................3-2
          Control Area Objectives....................................................................................................3-4
      System Impacts and Challenges...........................................................................................3-6
      Importance of Wind Energy Forecasting ...............................................................................3-6
          Forecasting Wind Generation Ramp Rates......................................................................3-7
      Customizing Wind Energy Forecasts for System Operators .................................................3-8
      Conclusions.........................................................................................................................3-10

4 CALIFORNIA REGIONAL WIND ENERGY FORECASTING SYSTEM
DEVELOPMENT........................................................................................................................4-1
      Introduction ...........................................................................................................................4-1
          Objectives and Scope.......................................................................................................4-2
          Project Participants...........................................................................................................4-4
          Major Project Tasks..........................................................................................................4-4
      Next-Hour Regional Wind Energy Forecasting System Development and Testing ..............4-4
          Objectives and Scope.......................................................................................................4-4
          Approach ..........................................................................................................................4-5
          Forecast System Design ..................................................................................................4-6
          Results..............................................................................................................................4-8
          Conclusions ....................................................................................................................4-10
      Recommendations ..............................................................................................................4-11
      Next-Day Wind Plant Energy Forecasting System Development and Testing....................4-11
          Objectives.......................................................................................................................4-11
          Scope .............................................................................................................................4-11
              Approach ...................................................................................................................4-12
          Phase 1: Screening of Improved Data and Forecast Methodologies .............................4-13
              Results .......................................................................................................................4-15
          Phase 2: Evaluation of Improved Forecast Algorithm at Five Wind Projects .................4-17
              Approach ...................................................................................................................4-17
              Results .......................................................................................................................4-18


xii
           Variability of Power Generation Forecast Error .........................................................4-19
                Ensemble Forecasts .............................................................................................4-20
       Recommendations..........................................................................................................4-23
   Numerical and Wind Tunnel Modeling of Wind Flow and Plant-Scale Power Curve
   over Complex Terrain..........................................................................................................4-24
       Objectives.......................................................................................................................4-24
       Scope .............................................................................................................................4-25
       Wind Tunnel Modeling of Wind Flow over Complex Terrain...........................................4-25
           Approach ...................................................................................................................4-25
           Results .......................................................................................................................4-25
           Conclusions ...............................................................................................................4-28
       Numerical Modeling of Wind Flow over Complex Terrain ..............................................4-28
           Approach ...................................................................................................................4-28
           Results .......................................................................................................................4-29
                Turbine Wind Speed Ratios ..................................................................................4-29
                Atmospheric Stability.............................................................................................4-30
           Conclusions ...............................................................................................................4-32
       Overall Conclusions and Recommendations..................................................................4-32
   High-Resolution Weather and Wind Flow Forecasting........................................................4-33
       Introduction.....................................................................................................................4-33
       COAMPS Model and Experiment Design.......................................................................4-33
       Results............................................................................................................................4-34
       Conclusions ....................................................................................................................4-37
       Objective.........................................................................................................................4-38
       Approach ........................................................................................................................4-38
       Results............................................................................................................................4-39

5 SUMMARY AND CONCLUSIONS .........................................................................................5-1
       Short-Term, Same-Day Forecasting.................................................................................5-1
       Long-Term, Next-Day Forecasting ...................................................................................5-2
       Wind Tunnel and Numerical Modeling..............................................................................5-3
       California Wind Generation Research Dataset (CARD) ...................................................5-3
   Conclusions...........................................................................................................................5-4




                                                                                                                                         xiii
    Recommendations ................................................................................................................5-5
        Field Implementation of Wind Forecast Workstation ........................................................5-5
        Next-Hour and Same-Day Forecasting Research ............................................................5-6
        Next-Day and Longer Forecasting Research ...................................................................5-6

6 REFERENCES .......................................................................................................................6-1




xiv
LIST OF FIGURES

Figure 2-1 Risoe Prediktor Model Schematic ............................................................................2-3
Figure 2-2 A schematic representation of the major components of the eWind forecast
     system (Source: AWS Truewind, LLC) ..............................................................................2-7
Figure 2-3 Schematic of Applied Modeling, Inc. Wind Energy Forecasting System
     (WEFS) ............................................................................................................................2-10
Figure 3-1 Typical Variation of Total and Regional One-Minute Wind Generation in
     California on a Summer Day [California ISO, March 2005]................................................3-3
Figure 3-2 Buffalo Ridge 115-kV Bus Voltage in kV (Top Trace and Left Scale) and
     Transformer Output in MW (Bottom Trace and Right Scale) (NREL) ................................3-3
Figure 3-3 Relationship between an Individual Control Area and Rest of Interconnection
     and Definition of Area Control Error [California ISO, 2005]................................................3-4
Figure 3-4 Control Area Objectives Focus on Balancing Load & Interchange vs.
     Generation [California ISO, 2005] ......................................................................................3-5
Figure 3-5 Balancing Authority Balances Generation Resources vs. Real-Time Load
     [California ISO, 2005].........................................................................................................3-5
Figure 3-6 Hourly Wind Generation for Selected Days at Tehachapi, April 2005 Data
     Adjusted to 4500 MW Rated Capacity [California ISO, 2005] ............................................3-8
Figure 3-7 Hourly Wind Generation (MW) and Ramp Rates (MW/hr) at Tehachapi for
     April 8, 2005, Adjusted to 4500 MW Rated Capacity [California ISO, 2005]......................3-9
Figure 3-8 Example Display of Real-Time Forecasts for Oak Creek Energy Systems
     Wind Project in Tehachapi, California [AWS Truewind, 2005] .........................................3-11
Figure 4-1 California Mean Wind Power Map at 50-m Elevation and 2005 Rated
     Capacity of Wind Generation at Principle Wind Resource Areas [California Energy
     Commission] ......................................................................................................................4-2
Figure 4-2 Schematic of the proposed two-stage short-term (0- to 3 hour) forecast
     system................................................................................................................................4-7
Figure 4-3 Mean absolute errors of regional power generation forecasts (% of rated
     capacity) vs. look-ahead period for the warm (May-Oct) and cold (Jan-Apr, Nov-
     Dec) season months of 2004, ANN-based autoregressive forecasts (stage 1 –
     method 1), and the four largest California wind resource areas.........................................4-9
Figure 4-4 Skill scores of regional power generation forecasts relative to persistence
     forecasts vs. look-ahead period for the warm (May-Oct) and cold (Jan-Apr, Nov-
     Dec) season months of 2004, ANN-based autoregressive forecasts, (stage 1 –
     method 1) and the four largest California wind resource areas..........................................4-9
Figure 4-5 Schematic of the eWind forecast system ..............................................................4-13



                                                                                                                                          xv
Figure 4-6 Terrain model of Altamont Pass used in Atmospheric Boundary Layer Wind
     Tunnel testing at the University of California at Davis. Wind power densities were
     measured at the met-tower and wind turbine locations vs. wind direction in the wind
     tunnel. View is from the prevailing wind direction, west-southwest (240°), toward the
     east northeast (60°). The highest elevation is in the foreground......................................4-27
Figure 4-7 Evolution of individual turbine wind speed ratios over the six-hour, 100-m
     simulation beginning at 0000 UTC 16 July 2002, Altamont Pass Turbine Cluster
     M127 ................................................................................................................................4-29
Figure 4-8 Comparison of three methods of predicting the plant power output using
     forecast M127 wind speeds from the December 2001 to September 2002 dataset ........4-30
Figure 4-9 Comparison of four methods of predicting the plant power output using
     observed M127 wind speeds from the ten 6-hr periods simulated by the high-
     resolution numerical model. "Mean Ratios" is the simulation method using wind
     speed ratios between the turbine locations and M127 that are averages over the ten
     cases with a range of stabilities. “Stability-modified” refers to predictions in which
     the turbine wind speed ratios were adjusted for the stability of each case. .....................4-31
Figure 4-10 Resolution of terrain elevation (meters) vs. grid size (km) of the nested
     domain: (a) ∆x = 12 km (nest_2), (b) ∆x = 4 km (nest_3), (c) ∆x = 1.333 km
     (nest_4), and (4) ∆x = 0.444 km (nest_5). The letters mark the locations of the met
     towers used in this study..................................................................................................4-35
Figure 4-11 Weekly mean absolute errors of NARAC forecasts averaged over the
     selected stations for June 1-7, 2005. The colored lines represent the results for
     different horizontal resolutions (36, 12, and 4km, respectively). The left panels
     present MAEs for wind speed forecasts, and the right panels for wind direction. The
     top panels (a and b) are the forecast errors using the measurements from all
     stations, the middle ones (c and d) using five reliable stations, and the bottom plots
     (e and f) using uncertain measurements from the remaining six stations. .......................4-36
Figure 4-12 The geographical domain covered by the 80 X 80 matrix of 5-km grid cells
     used to produce the high-resolution MASS 6 simulations over northern California for
     the project. .......................................................................................................................4-40
Figure 4-14 The geographical domain covered by the 80 X 80 matrix of 5-km grid cells
     used to produce the high-resolution MASS 6 simulations over southern California
     for the project ...................................................................................................................4-41
Figure 4-15 The directory structure of the CARD database. The shaded boxes denote
     directories. Names not enclosed in boxes are files. The numbers below the
     directory or file names indicate the number of the number of files or directories with
     the name above the number. See text for explanation of directory and file names. ........4-42




xvi
LIST OF TABLES

Table 2-1 Selected Wind Energy Forecasting Providers ..........................................................2-1
Table 4-1 Focus Areas and Associated Experiments ..............................................................4-15
Table 4-2 Summary of the Improvements in the Power Production Forecast MAE
    associated with the Forecast System Modifications in each Focus Area.........................4-16
Table 4-3. Mean absolute errors and skill scores of the best-performing power
    production and wind speed forecast methods for each wind plant, 48-hour
    forecasts, and July 1, 2004 to June 30, 2005 forecast period..........................................4-18
Table 4-4 Mean absolute errors of ensemble power production (% of rated capacity)
    and wind speed (m/s) forecasts for the five participating wind plants ..............................4-21
Table 4-5 Mean error (ME) and mean absolute error (MAE) of UC Davis plant power
    curve methods. Values are percentage of wind farm capacity.........................................4-28




                                                                                                              xvii
1
INTRODUCTION



Worldwide installed capacity of wind power reached almost 60,000 MW during 2005 and
continues to grow at 20% to 30% per year. Because wind generation is intermittent and not
dispatchable, the addition of large blocks of wind generation creates a challenge for control area
operators, independent system operators (ISO) and Regional Transmission Operators (RTO).
This is especially true when the local grid is isolated and not interconnected with the regional
grid, such as on remote islands, and when the wind generation facilities are concentrated in areas
far from the population centers and the transmission system is constrained and unable to carry
the load during peak wind generation periods.

Wind energy forecasting uses sophisticated numerical weather forecasts, wind plant power
generation models, and statistical methods to predict wind energy generation at five-minute to
one-hour intervals over periods up to 48 to 72 hours in advance, and for seasonal and annual
periods. Forecasting has great potential to both help manage the interface between large wind
generation facilities and the electricity grid and address the needs of the ISOs and RTOs, as well
as the utilities, wind plant operators, power marketers and buyers, and utility system dispatch
personnel.

This report provides an update of the status, capabilities, and specific needs for wind energy
forecasting and builds on the information presented in the 2004 EPRI report, Wind Energy
Forecasting Technology Update: 2004 (1008389, April 2005). The specific topics addressed
include wind energy forecasting integration into electricity grid operations and the results of the
California Regional Wind Energy Forecasting System Development project, completed in
December 2005.


Objectives and Scope
The overall objective is to update the status of short-term, intermediate-term, and long-term wind
energy forecasting technologies and their applications to power marketing, grid operations, wind
project planning, project developments, and strategic planning. The approach is to gather and
analyze information on recent wind energy forecasting system developments and applications.




                                                                                                 1-1
Introduction


Report Organization
The report consists of five sections, including:

      •   Section 1, Introduction.

      •   Section 2: Wind Energy Forecasting Systems and Developers summarizes the principle
          developers of wind energy forecasting technology in the U.S. and summarizes their
          forecast methodologies.

      •   Section 3: Wind Energy Forecasting Integration into Electricity Grid Operations
          addresses the growing need to provide accurate same-day and next day forecasts of high
          wind ramp rates; and the next steps to develop, test, and implement a real-time forecast
          system and display in collaboration with the electricity system operators.

      •   Section 4: California Regional Wind Energy Forecasting System Development
          summarizes the same-day and next-day forecast algorithm development and testing in the
          EPRI-CEC PIER project completed in 2005 [CEC PIER Program 2006a, 2006b, 2006c,
          and 2006d].

      •   Section 5: Conclusions and Recommendations draws conclusions and makes
          recommendations for future based on the research completed during the 2005.

Sections 3 and 4 are excerpted from the final report of the CEC-EPRI California Regional Wind
Energy Forecasting System Development Project [EPRI-CEC PIER Program 2006a and 2006b].




1-2
2
WIND ENERGY FORECASTING SYSTEMS AND
DEVELOPERS


Wind Energy Forecasting Providers

Table 2-1 lists several of the companies and groups offering wind energy services throughout the
world. They include Applied Modeling, AWS Truewind, 3TIER Environmental Forecast Group,
and Wind Logics in the U.S., Garrad Hassan in the UK, Risoe National Laboratory in Denmark,
ISET in Germany, and CSIRO in Australia. In addition, the European Commission ANEMOS
Project is tracking several wind energy forecasting projects in Europe.


   Table 2-1
   Selected Wind Energy Forecasting Providers


    Organization                      Forecast Model Description                Internet Address

    Applied Modeling, Henderson, NV   48-Hour Forecasts (WEFS)                  www.amiace.com

    AWS Truewind LLC, Albany, NY      1-Hour, 48-Hour, and Annual Forecasts     www.awstruewind.com
                                      (eWind); Wind Mapping



    3TIER Environmental Forecast      1-Hour and 48-Hour Forecasts              www.3tiergroup.com
    Group, Seattle, WA

    WindLogics, St. Paul, MN          1-Hour and 48-Hour Forecasts              www.windlogics.com

    Risoe National Laboratory,        1-Hour and 48-Hour Forecasts              www.riso.dk
    Roskilde, DK                      (Prediktor and Zephyr)

    ISET, Kassel, Germany             1-Hour and 48-Hour Forecasts              www.iset.uni-kassel.de

    Garrad Hassan, UK                 1-Hour and 48-Hour Forecasts              www.garradhassan.com
                                      (GH Forecaster)

    CSIRO, Australia                  Annual Forecasts, Wind Mapping            www.csiro.au
                                      (WindScape)

    European Union Consortium         European Commission ANEMOS                anemos.cma.fr
                                      project tracks wind energy forecasting.


                                                                                                         2-1
Wind Energy Forecasting Systems and Developers


Risoe National Laboratory – Prediktor

Risoe National Laboratory in Denmark has been developing the Prediktor physical power
prediction model for wind plants over the last five years. Staff at Risoe recognized that there
were many areas in Europe, especially in Denmark, where the penetration of wind energy is so
great that the fluctuations of the wind energy delivered to the electricity grid affect the control
and dispatch of the system.

Two types of wind power prediction models have been proposed, physical models and statistical
models. Risoe elected to develop a physical model to predict the output of a wind farm. In order
to forecast wind plant output at specific times in the future, one must first forecast the wind
speed and wind direction and then the wind energy generation can be estimated. The integration
of numerical weather prediction information with a diagnosis of local effects and the specific
characteristics of the wind turbine are the key features of the physical approach. The key model
components include:

      •   Wind speed and wind direction data from a Numerical Weather Prediction (NWP) model.

      •   A description of the site (orography, roughness, obstacles).

      •   A description of the wind turbine (hub height, power curve, thrust curve).

Figure 2-1 presents a schematic of the Risoe Prediktor model. Predictions of wind speed and
wind direction from the NWP Model (HIRLAM Wind) are modified using the geostrophic drag
law and the logarithmic wind profile to produce an estimate of the surface wind speed and
direction. This estimate is then used in the Wind Atlas Analysis and Application Program
(WAsP) to generate a local wind speed estimate. The program PARK is then applied to simulate
the wake and array effects on the each individual wind turbine. The power production of the
wind park is based on the calculated array efficiency for each wind direction sector. In addition,
local corrections are applied to the local wind speed and wind direction and to the estimated
power production. Thus, some historical wind resources and the power generation are needed to
calibrate the model and improve the accuracy of the forecasts.

Numerical Weather Prediction Data
Numerical weather prediction models use current conditions and past trends to predict the future
behavior of the atmosphere. This behavior at discrete time intervals into the future is the first key
ingredient of the Risoe Wind Plant power prediction model. The data are usually derived from
the regional atmospheric forecasting models operated by national weather services. They may be
either obtained directly from the weather service as point information for the wind farm location
or retrieved as grid information from public access ftp-servers. The choice depends on the wind
farm and the practice of the weather service.




2-2
                                                        Wind Energy Forecasting Systems and Developers




   Figure 2-1
   Risoe Prediktor Model Schematic


Surface Wind
The idea behind the physical model is that the predicted wind speed and wind direction from
ETA model, which is the wind value specific to a 20 km grid cell, can be transformed using the
geostrophic drag law and the logarithmic wind profile. The drag law is expressed as:

       G = µ*/κ (√ [ln (µ*/f zo)-A]2 + B2                                            (2-1)

where G is the geostrophic wind, equal to the ETA wind speed, µ* is the friction velocity, κ is
the Von Karman constant (0.4), f is the Coriolis parameter, zo is the roughness length, and A and
B are constants (1.8 and 4.5, respectively). The logarithmic wind profile is expressed as:
       u(z) = (µ*/κ) (ln (z/zo))                                                     (2-2)
where u(z) is the wind in the surface boundary layer at height z.

WAsP

The wind speed and wind direction calculated from ETA Model is valid for a very large area and
must be corrected for local effects. This is done using the Wind Atlas Analysis and Siting

                                                                                                  2-3
Wind Energy Forecasting Systems and Developers


program (WAsP) (Mortensen, et al, 1993). WAsP modifies the local wind field for the effects of
obstacles (structures, wind breaks, etc.), the effects of surface roughness and the changes in
surface roughness, and the effects of orography. The application of WAsP requires a digitized
terrain file, a file of surface roughness and change in surface roughness as a function of the wind
direction sector, and an obstacle file, which contains a description of each obstacle, affecting the
wind speed.

PARK

The PARK Program estimates the impact of wake turbulence created by upwind turbines on the
total energy generation. The required data inputs include the wind turbine coordinates, the wind
turbine power curve, and the wind turbine thrust curve. PARK calculates a wind park efficiency
factor vs. wind direction sector (twelve 30-degree sectors).

Model Output Statistics (MOS)

To correct for effects not explained by the models, Model Output Statistics (MOS) are applied in
two stages. In the first stage, a MOS corrections in the form of a simple linear function is applied
to the wind speed.

        y(final, sector) = y(model, sector)a(sector) + b(sector)                       (2-3)

where y(final, sector) is the final wind speed prediction , y(model, sector) is the wind speed
forecast from the physical models, and a(sector) and b(sector) are the direction dependent
constants of the linear function. The directionally dependency is according to twelve 30 degree
sectors.

In the second MOS stage corrections for any other biases in the model predictions of power
output are applied, c is selected such that:

        PMOS = PMODEL + c                                                              (2-4)

where c is a constant value (power) not sector dependent.

Meteorological Forecast Data Retrieval System

The ETA forecast model operated by NOAA is used to provide meteorological data. ETA is a
numerical weather prediction model, which became operational in the mid 1990s. The model
employs 60 vertical layers with a 20 kilometer grid resolution and is run four times each day out
to 48 hours in the future. The model grid covers North America from the pole to the Caribbean
Ocean. It is operated by National Center for Environmental Prediction (NCEP) in Washington,
DC (http://www.erh.noaa.gov/er/bgm/models.htm#ETA).

Output from this model is downloaded automatically by a PERL script. The script runs every
20 minutes and checks for the existence of the latest forecast file at the primary ftp-site. If the

2-4
                                                          Wind Energy Forecasting Systems and Developers

forecast file is not present there, it checks a secondary ftp-site as a back up. The forecast files are
in the GRIB format and each one is approximately five MB in size. For every forecast cycle run,
17 files are downloaded; the forecast data has a three-hour time increment out to 48 hours. From
these files, wind and temperature at several elevations and locations around the wind farm are
extracted for a 60 x 60 km region around the Mountain View and Altamont wind farms. The data
are written to file for use in the Prediktor Power Prediction System.

The retrieval system includes a checking script to help ensure the downloading of forecasts is
being carried out and to restart it if for some reason it has stopped. It also has the capability of
triggering the Power Prediction System when the appropriate forecast data have been collected.

Power Prediction System

The farm power prediction program is written in JAVA and uses object-oriented programming
structures. For example, wind farms and forecast winds are program classes and the calculation
of power or use of the geostrophic drag law are methods within those classes. Input parameters
include the forecast wind files, farm attributes, turbines, power curve, etc, and the farms’ wind
and power parameters derived from statistical models. For large wind farms with many turbines,
the physical modeling is sometimes minimized, because such detailed modeling of individual
turbine production requires extensive set-up time and calculations.

The system operates by looking up the most recent forecast downloaded for each farm and
determining whether a power prediction calculation has been applied to that forecast. If it has
not, the program proceeds. The way the program is written allows for the modification of the
wind farm attributes and the addition of further wind farms. Output from the program is handled
by the web-site prediction display scripts.

Meteorological Forecast Data Archive

Each day, an automated PERL script collects the previous day’s forecast data and compresses
and stores the data in a standard way. Such an archive is essential for finding relationships
between forecast wind, observed wind and wind farm power output.

Prediction Output and Web-Display

Up-to-date prediction information is posted on password-protected web-sites four times a day
that the algorithm checks for the existence of a new power prediction. If a new power prediction
exists, a text file is created that reports what meteorological forecast cycle has been used for this
prediction. Other summary information could be written to this file, such as time of maximum
power output, periods of no power production, extreme winds etc. The forecast is then sent by
FTP to a directory on our web server so that it can be displayed in a text box on the wind farm
web-site. The script also collates the power prediction data into a power history file. This
contains the power predictions for the last three predictions using older meteorological forecasts.
These data are used to create a graph of predicted power output against time. The degree of
agreement of the power predictions from successive meteorological forecasts can give some

                                                                                                    2-5
Wind Energy Forecasting Systems and Developers


indication of the predictability of the wind conditions. The last eight predictions could be used
for this purpose (48 hour lead time, six hourly forecasts). The older predictions would tend to be
less accurate and could therefore be given less weight.

The predicted wind speed and farm output are also written to a file, called PredFile, in a standard
format defined by EPRI. Each time this file is updated, it is posted to the appropriate EPRI ftp
sites.

Example Forecast Output

Table 2-2 presents an example forecast generated by the Risoe Prediktor Model. The forecasts
are issued four times each day and cover the period from zero hour zero to 48 hours in advance.
Each forecast is specifically identified by the Wind Plant, the Contractor, the Forecast Issue
Time in both Universal or Greenwich Mean Time, the forecast hour interval (every three hours),
and the valid time for the forecast. The forecast parameters include wind speed, wind direction,
temperature, and power output.
      Table 2-2
      Example of Forecast at SW Mesa Generated by the Risoe Forecast Model

                             Forecast Issued             Forecast Hour      Wind Spd    Dir    Temp Wind Gen
 Plant   Contractor     UCT Time        Local Time   Hr #     Local Time     m/sec   degrees   deg F    kWh
 SWM      RISOE       200204010000 200203311900        0     200203311900      3.7     75.5     N/A    8873.4
 SWM      RISOE       200204010000 200203311900        3     200203312200       7      112      N/A    21451
 SWM      RISOE       200204010000 200203311900        6     200204010100      7.1    145.5     N/A   21326.8
 SWM      RISOE       200204010000 200203311900        9     200204010400      8.7    158.8     N/A   37483.5
 SWM      RISOE       200204010000 200203311900       12     200204010700       9     162.8     N/A   42172.8
 SWM      RISOE       200204010000 200203311900       15     200204011000      7.6     188      N/A    25470
 SWM      RISOE       200204010000 200203311900       18     200204011300      5.6    190.2     N/A   14528.3
 SWM      RISOE       200204010000 200203311900       21     200204011600      4.3    165.6     N/A   10559.8
 SWM      RISOE       200204010000 200203311900       24     200204011900      6.9    149.9     N/A    19343
 SWM      RISOE       200204010000 200203311900       27     200204012200     11.7    160.8     N/A   70423.1
 SWM      RISOE       200204010000 200203311900       30     200204020100     10.8    179.7     N/A   56920.4
 SWM      RISOE       200204010000 200203311900       33     200204020400       8     195.4     N/A    28181
 SWM      RISOE       200204010000 200203311900       36     200204020700      8.1    229.7     N/A   29442.6
 SWM      RISOE       200204010000 200203311900       39     200204021000      7.2    262.9     N/A   23325.3
 SWM      RISOE       200204010000 200203311900       42     200204021300      4.7    247.9     N/A   10109.5
 SWM      RISOE       200204010000 200203311900       45     200204021600      3.8    247.2     N/A    8901.8



AWS Truewind - eWind

AWS Truewind LLC (formerly TrueWind Solutions LLC) has developed a state-of-the-art
system, called eWind, to forecast the power output of a wind plant as well as a wide range of
meteorological parameters in the vicinity of the plant. The eWind system consists of four basic
components: (1) a set of high-resolution three-dimensional physics-based atmospheric numerical
models; (2) adaptive statistical models; (3) plant output models; and (4) a forecast delivery
system. Figure 2-2 presents a schematic representation of the components of the eWind forecast
system. The following three subsections provide an overview of the major components of the
eWind system.




2-6
                                                             Wind Energy Forecasting Systems and Developers




 Reg io nal / Glo b al
                         eWind                                 Statistical
 Weat her Models                                                Models               Measured
                                                   Wind
                                                                                      Wind
                                                  Forecast
                                                                                      Dat a
                                                                          Adjusted
                                                                           Wind
                                                                          Forecast
                                                             Adjusted
                                             Physical         Wind
                                             Models          Forecast                 USER
                                                                          Po wer
                                                                          Out put
                                                                         Fore cast
                                                                                     Po wer
                                                  Wind                               Out put
                                                 Forecast                            Dat a
      Raw                    Surf ace
  At mo sp heri c        Ch ar act e rist i cs
                                                             Plant Output
     Dat a                     Dat a                           Models




    Figure 2-2
    A schematic representation of the major components of the eWind forecast system
    (Source: AWS Truewind, LLC)


Physics-Based Atmospheric Numerical Models

Physics-based atmospheric numerical models are a set of mathematical equations that represent
the basic physical principles of conservation of mass, momentum and energy and the equation of
state for moist air. The model’s equations are solved on a three-dimensional computational grid.
These models are conceptually similar to those used by government-operated weather forecast
centers (such as the National Center for Environmental Prediction (NCEP) in the United States)
throughout the world. However, the eWind models are run at a higher resolution (i.e., smaller
grid cells) and the physics and data assimilation schemes used by the eWind models have been
specifically configured for high-resolution simulations for wind power forecasting applications.
In the current standard configuration of the eWind system, a single physics-based model, known
as the Mesoscale Atmospheric Simulation System (MASS), was used to generate the forecasts.
MASS is a non-hydrostatic atmospheric model that has been developed and used for a variety of
applications by MESO, Inc. (one of the principals of TrueWind Solutions) since 1985.

Recent research has demonstrated that a composite of forecasts from an appropriate ensemble of
simulations from physics-based atmospheric models is generally superior to a forecast based on
one simulation (i.e., one member of the ensemble). There are two fundamental strategies that can

                                                                                                       2-7
Wind Energy Forecasting Systems and Developers


be used to generate an ensemble of forecasts. One strategy is to use the same atmospheric model
and vary the input data (initial and boundary conditions) within their range of uncertainty. The
other strategy is to use the same input data and to employ an ensemble of models or different
configurations of the same model. The relative value of either strategy depends upon the sources
of uncertainty in the forecast simulations. If a greater amount of the uncertainty is related to the
input data then the strategy of executing a set of simulations by perturbing the input data will be
more valuable. If the uncertainty is mostly related to the model formulation, then an ensemble of
models will be more useful. In practice, the magnitude of the sources of uncertainty varies with
location, season, spatial scale of the forecast and other factors. Therefore, the choice of ensemble
must be determined from experience and experimentation. The eWind system can employ the
ensemble approach if sufficient computational resources are available for a specific application.
If the ensemble approach cannot be employed because of insufficient computational resources
for a particular application, then the single-simulation approach is utilized.

The ensemble approach is facilitated by the fact that AWS Truewind has extensive experience
and expertise with the development and/or use of a number of atmospheric models. In addition
to the MASS model, AWS Truewind also has implemented several other physics-based
atmospheric models from a variety of sources for execution on its computational platforms.
These models include the FOREWIND, MM5, WRF, COAMPS, workstation-ETA and OMEGA
models. Each of these models has unique attributes that can bring additional information to an
ensemble of numerical forecast simulations.

FOREWIND is a high-resolution boundary layer model that has been developed by AWS
Truewind Solutions. This model is intended to run at very high resolution over a layer extending
from the surface of the earth to approximately 3 km while accepting data about the state of the
atmosphere above 3 km from another model. The ability to limit the domain to the atmospheric
boundary layer permits higher vertical and horizontal resolution to be used in this model while
still achieving the execution time required for a forecast simulation to be useful.

The MM5 model is a public-domain non-hydrostatic three-dimensional mesoscale model
developed by the Pennsylvania State University and the National Center for Atmospheric
Research (NCAR). Due to its availability in the public domain, it has been widely used and has
become a de facto standard for mesoscale models. Its main strength is that it has been widely
used and verified and incorporates a variety of atmospheric physics formulations from different
sources. However, its numerical techniques and data assimilation options are somewhat out of
date.

The WRF model is a next-generation atmospheric model and is currently being jointly developed
by NCAR and NCEP. An early version of this model is now available. It incorporates many of
the physical sub-models (long and short wave radiation, boundary layer turbulence etc.) from
MM5, but utilizes more advanced numerical techniques and will ultimately have an advanced
data assimilation system.

COAMPS is a high-resolution model developed by the U.S. Navy. Its most unique feature is that
it models both the oceans and the atmosphere, whereas all of the other models in this group treat
bodies of water as a specified (from an external source) lower boundary condition.


2-8
                                                         Wind Energy Forecasting Systems and Developers

The workstation-Eta model is a version of the National Weather Service’s ETA model that has
been designed to run on a computer workstation rather than in the supercomputer environment at
NCEP. It uses a different vertical coordinate system (Eta) from most of the other mesoscale
models. This is especially beneficial in regions of steep and complex terrain.

The OMEGA model is a very unique unstructured adaptive grid model that has been
co-developed by SAIC and MESO, Inc. It is the first atmospheric model to use an unstructured
grid (i.e., a grid in which the grid cells have no predefined relationship to each other). The
unique grid structure permits the use of a continuously variable (in space and time) grid
resolution within the model domain and permits the adaptation of the grid structure to geographic
(e.g., land-water boundaries) or atmospheric features (e.g., frontal zones).

All of these models can currently be executed on computational platforms at AWS Truewind.
However, it is not cost-effective to run an ensemble of simulations that utilizes all of these
models and a variety of initial state perturbations. Therefore, a manageable set of initial
perturbation and model configuration strategies are customized for each forecast application.

Adaptive Statistical Models

The adaptive statistical models are used to build a set of empirical relationships between the
output of the physics-based atmospheric models and specific parameters to be forecasted for a
particular location. In the Southwest Mesa application, the specific meteorological parameters
are the wind speed and direction and air density at the location of the wind plant. The role of the
statistical models is to adjust the output of the physics-based models to account for sub-grid scale
and other processes that cannot be resolved or otherwise adequately simulated by the physical
models.

Two types of statistical models are currently used in the eWind system. The first is a traditional
multiple screening linear regression model. The second is a Bayesian neural network model.
As with the linear regression approach, this scheme uses a training sample to determine a set of
parameters that define the equations that relate the inputs (i.e., the predictors) to the target
(i.e., the predictand). The difference is essentially in the form of the functional relationships and
the method used to estimate the parameters. The neural network scheme used in eWind employs
a Markov Chain Monte Carlo training method. This method trains an ensemble of networks by
taking many samples from the distribution of the network parameters. These network parameters
are the weights and biases that determine the input-to-target function. They are the conceptual
equivalents of slopes and intercepts in a linear regression. The distributions used for sampling
depend on the training data, the “noise model” (i.e., the assumed distribution of the noise in the
data) and the “priors” (additional information provided to the scheme about the expected
smoothness of the functions, i.e., the ranges of the network parameters).

Plant Output Models

The third component of the system is the plant output model. This model is a relationship
between the critical atmospheric variables and the plant output. It is possible to use either a

                                                                                                   2-9
Wind Energy Forecasting Systems and Developers


physical or statistical approach in the formulation of the plant output model. The eWind system
utilizes only the statistical approach. The eWind statistical plant output model can be a fixed
relationship derived from a single usually long-term dataset of measured meteorological
parameters and the plant energy output or it can be a dynamic relationship derived from recent
(e.g., the last 60 days) measured data from the plant. However, even a statistical plant output
model can incorporate information about the layout of the plant into the model’s framework.
For example different sets of relationships may be used for cases in which the wind is blowing
parallel to a row of turbines than in cases when the wind is blowing perpendicular to the row.

Forecast Delivery System

The final piece is the forecast delivery system. The user has the option of receiving the forecast
information via email, an ftp transmission, a faxed page or on a password-protected web page
display.


Applied Modeling - WEFS

Applied Modeling Inc. (AMI) has developed and tested a state of the art forecasting system
based on the most advanced numerical weather prediction models and wind modeling
technologies. As shown in Figure 2-3, the AMI Wind Energy Forecasting System (WEFS)
consists of the following modules: (1) a mesoscale model; (2) a diagnostic wind model; (3) an
adaptive statistical model; and (4) the forecast access by users. The following sections describe
the development and operation of the WEFS modules.




                                                                     Real-Time
                                                                       Data




  Regional             Mesoscale                 Diagnostic          Adaptive               Wind &
   Model                Model                    Wind Model          Statistics             Energy
   ETA                   MM5                       DWM                                     Forecasts

Grid size ~ 80 km     Grid sizes                 Grid sizes
                    45,15,5,1.67 km              100, 30 m

                                                                                           USER
                                                                                        By e-mail, FTP



    Figure 2-3
    Schematic of Applied Modeling, Inc. Wind Energy Forecasting System (WEFS)



2-10
                                                       Wind Energy Forecasting Systems and Developers



Mesoscale Model MM5

The first WEFS module is a PC-based version of the advanced, three-dimensional mesoscale
model MM5 (Version 3), developed by Applied Modeling Inc. (AMI). The Fifth Generation
Mesoscale Model (MM5) is a limited-area, nonhydrostatic, terrain-following, sigma-coordinate
model designed to simulate or predict mesoscale and regional-scale atmospheric circulation.
It was originally developed as a community mesoscale model at Penn State University and
National Center of Atmospheric Research (NCAR) during the 1970s. It is continuously being
improved by contributions from several universities and government laboratories as well as
private consulting firms around the world. As a result, the MM5 model offers several advanced
parameterizations for turbulence, cloud and precipitation. It also incorporates detailed
topographical and land use databases, both for the U.S. and elsewhere. The MM5 model is the
most widely-used and verified mesoscale model today. It is used in diverse applications, from air
pollution modeling studies to forecasting storms and tornadoes. AMI is currently using the MM5
model to forecast tropical storms in Southeast Asia and delivers four forecasts daily via the
Internet.

Mesoscale models such as MM5 are traditionally run on large computer systems, such as
supercomputers or Unix workstations. Recently, AMI developed a PC-based MM5 version
running with the Linux operating system. Realistic tests conducted by AMI indicate that a PC
equipped with two or more CPUs (i.e., multiprocessor) is as fast as the best Unix workstations.
Yet the Linux PC only costs a small fraction of these expensive computers.

The MM5 forecasts use the outputs of a regional or global model for initial and boundary
conditions. In its current configuration, the MM5 model can use the outputs from either the
regional-scale ETA model or the global-scale AVN model. All ETA/AVN outputs are
downloaded from the FTP server of the NOAA National Center of Environmental Prediction
(NCEP) in Maryland. To enhance the accuracy of the forecasts, the MM5 model is generally
configured to use a modeling domain with several nested grids with varying spatial resolutions.
For sites located in complex terrain, it is necessary to deploy modeling grids with the lowest
possible resolution (a few kilometers or less). AMI has a global topographical and land use data
base with a one-km resolution for use with MM5.

Diagnostic Wind Model

To further resolve the local topography and microscale flow effects, the MM5 predictions are
coupled with a diagnostic wind model (DWM) developed by AMI. The DWM model can derive
mass-consistent, three-dimensional wind fields that includes treatment for localized flow
phenomena such as terrain channeling, thermal drainage and over land/over water transition.
A refined resolution of 100 m or less is frequently used in the DWM simulations. The same
number of vertical layers is used in both MM5 and DWM simulations, including those at the
wind anemometer and turbine hub heights. Hourly-averaged predictions from the MM5 model
serve as inputs to the DWM model. To enhance the accuracy of short-term (e.g., next-hour)
forecasts, the DWM model can also accept onsite real-time wind measurements as inputs.


                                                                                               2-11
Wind Energy Forecasting Systems and Developers


Adaptive Statistical Model

Even with the best available models such as MM5 and DWM, forecast errors are still present and
can be caused by both systematic and non-systematic factors. Non-systematic or inherent errors
include those due to random atmospheric turbulence. While there is little that can be done about
these inherent errors, systematic biases in the forecasts can be characterized and at least partially
addressed. AMI devised an adaptive and efficient statistical scheme to minimize biases towards
either overpredictions or underpredictions. For each forecast, the statistical model computes
simple linear regression equations using recent actual measurements at the facility. Monitoring
data (wind, power and temperature) from the last 10 days or less are used to derive the regression
equations. Separate equations can be easily generated for different wind speed intervals for wind
and power predictions or time of day for temperature predictions.

The AMI scheme is fully dynamic and adaptive since new regression equations are derived for
each new forecast and take into consideration the most recent model biases. Unlike the
traditional MOS (Model Output Statistics) approach, the AMI statistical scheme does not require
long sampling time and extensive monitoring data. Furthermore, it is much simpler to implement
than MOS, which requires extensive recalculations due to changes in the forecast models,
weather conditions or wind plant configurations.

Forecast Access by Users

Upon completion of the forecast calculations, the forecast wind speed, direction, ambient
temperature, and energy generation data are sent via e-mail or other electronic means to the host
and other organizations. The forecasts can also be uploaded to the host FTP server along with
appropriate statistics designed to evaluate the accuracy and skill of the forecasts.


3TIER Environmental Forecast Group
3TIER Environmental Forecast Group (3TIER) has developed and tested a state of the art
forecasting system based on the most advanced numerical weather prediction models and wind
modeling technologies. 3TIER is currently the exclusive forecast provider for 800 MW of
generation capacity throughout the Western and Central U.S. 3TIER develops and operates a
forecasting strategy that delivers a range of forecast products at various lead times via a secure
web site for each wind project. The system has ample built-in redundancies, including
independent communication methods, this to insure that the end-user gets the best forecast, no
matter what real-time data is available. Inline with 3TIER’s philosophy of using proven state of
the art modeling technologies to solve forecast problems, 3TIER has based its forecast system on
a combination of mesoscale numerical weather prediction models and self learning (neural
network) models. These models use on and off site observations to generate both short and long
term forecasts. In our experience, different sites require different techniques and varying
amounts of off-site data. Our past experience with wind forecasting in the Pacific Northwest has
demonstrated that integration of relevant (i.e., highly correlated to changes in wind at the site
itself) high-temporal resolution off-site observations into the forecast process is essential to
achieve performance that is considerably better than that obtained by techniques using only data
from a single site.

2-12
                                                       Wind Energy Forecasting Systems and Developers


Hour-Ahead Forecasting

The hour-ahead forecasts will be based on a neural network model of wind speed and power
production trained on the leading predictors of wind speed and power at each site. The model
will be flexible so that many model inputs can be tested, and the model inputs varied by time of
the year. It has been shown that properly trained neural network models can perform
considerably better than more simple models (such as ARIMA) for forecasting wind and power
production. The neural network model forecasts are nudged so that mean monthly bias is below
the threshold set by the user.

Day-Ahead Forecasting

Day-ahead forecasts are based on the output from a mesoscale non-hydrostatic numerical
weather prediction model configured to run for the area around the wind plant or wind resource
area. Based on our experience in orographically complex regions in the West, the model is
usually configured with a horizontal resolution of 5 km or less. Thus it is capable of explicitly
representing the complex wind fields that are often encountered at the various sites. The
mesoscale model output is bias-corrected using a statistical correction based on Model Output
Statistics (MOS) or other statistical correction methods. An ensemble based forecast strategy,
employing a combination of initial condition, model physics, and temporal uncertainty methods,
may eventually be employed to control for synoptic scale forecast error. As an ensemble-based
approach adds significantly to the complexity and expense of running numerical weather
prediction models, the decision to adopt an ensemble based approach is usually made in
consultation with the user.




                                                                                               2-13
3
WIND ENERGY FORECASTING INTEGRATION INTO
ELECTRICITY GRID OPERATIONS


This section addresses the characteristics of wind generation; how wind generation affects
electricity system operations; why wind energy forecasting is important for next-hour and next-
day dispatching of system resources; and the content, format, and method of delivery of forecasts
that would be most useful to system operators.

It is excerpted from the final report of California Regional Wind Energy Forecasting project,
completed by the California Energy Commission and EPRI in December 2002, California
Regional Wind Energy Forecasting System Development, Volume 2: Wind Energy Forecasting
System Development and Testing and Numerical Modeling of Wind Flow over Complex Terrain
[EPRI-CEC PIER, 2006b].


Characteristics of Wind Generation

Important characteristics of wind energy that affect system operation include:

   •   Commercial wind turbines employ mechanical and electrical components to convert wind
       energy to electricity and use induction generators and sophisticated electronic controls to
       produce 50 or 60 Hz power, while thermal generation uses fuel combustion to either
       generate steam and drive a single large steam turbine generator or drive a combustion
       turbine.
   •   A large wind-generation facility consists of a large number of relatively small generating
       units, which are connected together and act like a distributed generation system.
   •   The “prime mover” for a wind plant is, obviously, the wind. The characteristics are
       determined by Mother Nature, not plant operators or system dispatchers. A particular
       wind regime can be characterized quite well in a statistical sense but does not lend itself
       well to deterministic analyses. Consequently, wind plants cannot be reliably scheduled in
       advance, and the capacity value is typically zero to 15 percent of the rated capacity.
   •   The wind resource and the resulting output of a large wind plant fluctuate over nearly all
       time scales of potential interest for power system designers and operators: seconds,
       minutes, hours, days, seasons, and so on. Especially when coupled with the uncertainties
       mentioned above, such behavior challenges the validity of applying well-established
       methodologies and analytical approaches for power system engineering.



                                                                                               3-1
Wind Energy Forecasting Integration into Electricity Grid Operations


How Wind Generation Affects Electricity Grid Operations

As shown in Figures 3-1 and 3-2, wind generation exhibits both short- and longer-term
fluctuations over periods from seconds to longer periods due the intermittent nature of wind.
Figure 3-1 illustrates the fluctuations of one-minute wind power delivered to the California grid
during a single 24-hour period on a summer day in California [California ISO, 2005]. The chart
presents both aggregated wind power data for the state and data for each of the five wind
resource areas (Solano, Altamont, Pacheco, Tehachapi, and San Gorgonio).

Similarly, Figure 3-2 illustrates the short-term fluctuations of both wind energy generation and
115-kV line voltage over a three-day period in March 2002 for 230 MW of wind generators in
the Buffalo Ridge area of southwest Minnesota [EPRI, 2005a]. As the wind speed varied over a
wide range, wind energy generation fluctuated between about 0 and 200 MW and the line
voltage varied between about 114 and 123 kV. Experience in Germany, where more than
15,000 MW of wind generation is in place, indicates that very large changes in wind output can
occur in minutes as weather fronts pass through [E.ON Netz, 2004].

The California diurnal wind generation profile in Figure 3-1 is typical of a summer day in the
areas affected by the coastal marine-layer. Wind speed and generation build during the afternoon
as the marine-layer spreads from the high-pressure region over the cool Pacific Ocean to the
low-pressure regions over the hot interior valleys. It then reaches a peak in early evening, and
begins to fall off during the early morning hours reaching a minimum between about 10:00 am
and 2:00 pm.

Unfortunately, the diurnal patterns of wind power and system load do not match well. This is
especially true during the evening, when system load decreases while wind generation is
reaching its peak. The system operator or balancing authority must reduce other generation to
balance generation and load using decremental bids (800 MW in the example). In the morning,
when load is building and wind generation diminishes to its minimum, the reverse situation
exists. The system operator must dispatch additional generation or non-spinning reserve to
balance the system (1000 MW in the example).

System Operators and Balancing Authorities

System operators are often single utilities responsible for operating the generation, transmission,
and distribution system of single and sometimes multiple control areas. Balancing authorities
like the California Independent System Operator and other regional transmission operators
(RTOs) balance the generation and transmission resources for a group of control areas covering
large regions, as illustrated in Figure 3-3.

Area Control Error (ACE)

As defined in Figure 3-3, the Area Control Error (ACE) is an algebraic function of the average
deviations vs. schedule of generation, load, frequency, and net interchanges with other control


3-2
                                                                                                     Wind Energy Forecasting Integration into Electricity Grid Operations



                                                                                              Total California Generation

     1200




     1000




      800                                 Need to Dispatch 1000 MW
                                          Of Additional Generation Or                                                                                                                                                                TOTAL
                                        Activate Non-Spinning Reserves                                                                                                                                                               Pacheco
                                                                                                                                                                                                                                     Solano
MW




      600
                                                                                                                                                                                                                                     Tehachapi
                                                                                                                                                                                                                                     Altamont
                                                                                                                                                                                                                                     San Gorgonio

      400


                                                                                                               Need to Decrease Generation By
      200                                                                                                      800 MW Using Decremental Bids


       0
            0:01

                    1:01

                                      2:01

                                             3:01

                                                    4:01

                                                           5:01

                                                                  6:01

                                                                         7:01

                                                                                8:01

                                                                                       9:01

                                                                                                     10:01

                                                                                                             11:01

                                                                                                                     12:01

                                                                                                                             13:01

                                                                                                                                     14:01

                                                                                                                                             15:01

                                                                                                                                                     16:01

                                                                                                                                                             17:01

                                                                                                                                                                         18:01

                                                                                                                                                                                 19:01

                                                                                                                                                                                         20:01

                                                                                                                                                                                                 21:01

                                                                                                                                                                                                         22:01

                                                                                                                                                                                                                 23:01
     Figure 3-1 Typical Variation of Total and Regional One-Minute Wind Generation in
     California on a Summer Day [California ISO, March 2005]




                                                                                                                                                                                                                                       MW
            kV




                   120                                                                                                                                                                                                           200




                   116                                                                                                                                                                                                           100




                   112                                                                                                                                                                                                           0
                           1-Mar-02




                                                                                                                                                             21-Mar-02




                                                                                                                                                                                                                         31-Mar-02
                                                                                         11-Mar-02




     Figure 3-2 Buffalo Ridge 115-kV Bus Voltage in kV (Top Trace and Left Scale) and
     Transformer Output in MW (Bottom Trace and Right Scale) (NREL)


                                                                                                                                                                                                                                              3-3
Wind Energy Forecasting Integration into Electricity Grid Operations




           Area Control Error - ACE
                                                             ACE = ∆I − 10 B ⋅ ∆F
                                                                  = ∆G − ∆L − 10 B ⋅ ∆F

                                                                       F - frequency, Hz
                F                                                      I - net interchange, MW
                                                                       G - generation, MW
                                                                       L - load, MW
                                   Rest of Interconnection             ACE - area control error, MW
              Area 1
                                                                       B - bias setting, MW/0.1 Hz
              ACE
                                                                       ∆ - deviation from schedule

                         I
                                                                            Tie-line meter
                                                                            Frequency meter




      Figure 3-3
      Relationship between an Individual Control Area and Rest of Interconnection and
      Definition of Area Control Error [California ISO, 2005]


areas. Each control area is responsible for maintaining ACE within two ranges, referred to as
CPS1 and CPS2.

Control Area Objectives

Figure 3-4 illustrates the principal objectives of the area and regional control operators, which
are to operate transmission within thermal limits, maintain voltage within voltage stability limits,
observe transient stability limits, and balance generation against load. In addition, the regional
balancing authority balances the overall generation load, maintains scheduled interchanges, and
supports interconnection frequency.

In order to meet these objectives, the operator is responsible for scheduling generation and
transmission resources for both day-ahead and hour-ahead periods, load following, and system
voltage and frequency regulation to balance generation and load and meet other objectives.
Figure 3-5 illustrates the duration of system load vs. upward and downward regulation;
day-ahead, hour-ahead, incremental and decremental scheduling; and five-minute-dispatch of
generation resources.




3-4
                                 Wind Energy Forecasting Integration into Electricity Grid Operations




Figure 3-4 Control Area Objectives Focus on Balancing Load & Interchange vs. Generation
[California ISO, 2005]




Figure 3-5 Balancing Authority Balances Generation Resources vs. Real-Time Load
[California ISO, 2005]



                                                                                                 3-5
Wind Energy Forecasting Integration into Electricity Grid Operations



System Impacts and Challenges

The system impacts and operator challenges created by wind generation result from the
intermittency and short- and longer-term fluctuations of wind generation, due to gusting winds,
the resulting high ramp rates of wind generation, diurnal and seasonal wind speed and system
load variations, and movement of weather systems.

The system impacts include:
      •   Unit commitments and scheduling
      •   Voltage regulation/reactive power control
      •   Reserve margins for security and reliability
      •   Transmission bottlenecks during windy periods
      •   Frequency control and regulating reserves
      •   Load following and energy balance

The magnitudes of the impacts can vary over wide ranges and depend on several important
factors:
      •   Percentage penetration of rated wind capacity in the generation mix
      •   Geographical dispersion of wind capacity
      •   Diurnal and seasonal correlations between wind generation and system load
      •   Penetrations and types of other generation resources in the mix
      •   Presence of hydro, pumped storage hydro, and peaking capacity in the mix
      •   Adequacy of transmission resources to transmit wind energy during periods of peak
          generation to the population centers

Importance of Wind Energy Forecasting

Wind energy forecasting is one of several mitigation measures available to reduce the impacts of
wind on power system operation and control [EPRI, 2003c, 2004a, 2005a, and 2005b].

Other mitigation measures include:
      •   Power electronics and line compensation to absorb short-term fluctuations and control
          power factor.
      •   Integration with hydroelectric generation to absorb both short-term fluctuations and store
          off-peak wind energy generated at night when demand is low.
      •   Addition of storage and flow batteries, compressed air energy storage, pumped hydro,
          and other energy storage facilities.

3-6
                                       Wind Energy Forecasting Integration into Electricity Grid Operations


   •   Transmission upgrades to relieve bottlenecks during windy periods; and wind energy
       forecasting.

In combination with load forecasting, wind energy forecasting can support optimal dispatching
of intermediate and peaking generation, including hydro, fossil, and other dispatchable
generating units; dispatching of transmission resources; scheduling next-hour and next-day wind
energy deliveries to the grid; markets for green power and green certificate trading; and other
uses. The ongoing California ISO Participating Intermittent Resources Program (PIRP) provides
hourly forecasts of next-hour and next-day wind generation to participating wind plant operators
for use in scheduling next-hour wind energy deliveries to the California grid [California ISO,
2005].

Balancing authorities like California ISO need accurate next-hour and next-day forecasts of
hourly wind generation to anticipate high ramp rates [California ISO, 2005].

For the next-hour market, California ISO uses five-hour forecasts to prepare accurate dispatch
notices to send to quick-start generators and they need to receive the forecasts in time to send the
notices at T-270 min. (270 minutes before the start time). In addition, the next-hour forecasts are
due at T-75 min. (75 minutes before the hour) when the next-hour market closes.

For the next-day market, the forecast are used to assess hourly energy production and anticipated
hourly ramps. The wind energy forecast influences the procurement of ancillary services
(regulation and operating reserves), preparation of accurate day head generation dispatch notices,
and impact of wind generation schedules on transmission congestion.

Forecasting Wind Generation Ramp Rates

Forecasting of wind generation ramp rates will be particularly important if the rated wind
capacity in the Tehachapi Mountains increases as forecast by about 4000 about MW from
600 MW now to 4500 MW in the future. For example, Figure 3-6 forecasts the daily variation of
wind generation at Tehachapi at 4500 MW capacity, based on wind generation data for selected
days during April 2005 [California ISO, 2005]. The daily variations cover the entire range from
zero to 4500 MW during the 30-day period.

Very high positive and negative ramp rates occurred on several of the days. For example,
Figure 3-7 summarizes the extreme range of hourly wind generation and ramp rates that would
have occurred on April 8, 2005 with 4500 MW of wind capacity at Tehachapi. During the
24 hour period, wind generation varies between zero and 4500 MW, and the ramp rate exceeded
minus 1000 MW/hr during two hours, and reached or exceeded plus or minus 600 MW/hr during
11 of the 24 hours. The two minus-1000 MW/hr ramps occur at 7:00 and 8:00 AM, precisely at
the time when load is building rapidly.

As a result of the above and other internal studies, California ISO has concluded that it is
important to accelerate development and validation of wind energy forecast tools now so that
they are ready to use when the wind generation capacity in the state grows to a level that reach
two to four times the current 2100-MW level [California ISO, 2005].

                                                                                                       3-7
Wind Energy Forecasting Integration into Electricity Grid Operations



              Tehachapi Region – 4500 MW Rated Capacity




      Figure 3-6
      Hourly Wind Generation for Selected Days at Tehachapi, April 2005 Data Adjusted to
      4500 MW Rated Capacity [California ISO, 2005]



Customizing Wind Energy Forecasts for System Operators

Even the most accurate next-hour and next-day wind energy forecasts for a region will be almost
useless to the system operator if they are delivered, for example, as simple tables of wind speed
and wind energy generation forecasts vs. time, without additional information that directly
addresses the needs of the operator with regard to content, format, and method of delivery.

For example, other forecast information and data may at times be of greater interest to the system
operator than the actual wind speed and energy forecast, e.g., the forecast ramp rate of regional
wind generation, the impact of forecast errors on control area CPS1 and CPS2 compliance, and
the cumulative monthly imbalance of scheduled vs. delivered wind energy.

Figure 2-8 presents an example of a web-page display developed by AWS Truewind during the
development of the intermediate-term forecasting system described in Section 4, Next-Day
Forecast System Development and Testing. It displays the most recent 48-hour forecast of hourly
wind generation, in this case for a specific wind project, together with the observed wind
generation for the site through the most recent hour, plus forecast performance metrics for
various time periods through the present.




3-8
                                      Wind Energy Forecasting Integration into Electricity Grid Operations




            Tehachapi Region – 4500 MW Rated Capacity




   Figure 3-7
   Hourly Wind Generation (MW) and Ramp Rates (MW/hr) at Tehachapi for April 8, 2005,
   Adjusted to 4500 MW Rated Capacity [California ISO, 2005]


Examples of other information that could be provided in such a graphical display include:

   •   Relative confidence in forecast accuracy based on weather conditions.

   •   Range of uncertainty of forecast vs. time superimposed on forecast chart.

   •   Archived charts comparing forecast and observed wind speeds and energy generation.



                                                                                                      3-9
Wind Energy Forecasting Integration into Electricity Grid Operations


    •   Archived forecast performance metrics vs. observed data, e.g., mean error (bias, ME),
        mean absolute error (MAE), mean square error (MSE), and skill scores (Skill) over
        different time periods to present (day, week, month, season, and year).

    •   Customized forecast and other information, such as next-hour and next-day ramp rates of
        regional wind generation; estimated impacts of forecast errors on CPS1 and CPS2 area
        control error requirements, and the cumulative monthly imbalance between scheduled
        and delivered wind energy.

In order for the system operator to “buy in” and trust the forecasts, the operator should also be
involved in the system customization and initial trial period. That way the forecast information
provided and needs of the operator are more likely to merge.


Conclusions

Wind energy forecasting will become especially important to control area operators and
balancing authorities like the California ISO in the future as local concentration of wind
generation reach thousands of megawatts.

Accurate forecasting systems are needed to generate both next-hour and next-day and longer
forecasts for several reasons:
    •   Provide early warning of high hourly ramp rates of wind generation for planning both
        same and next day dispatching of generation and transmission resources.
    •   Support markets for ancillary services to support intermittent wind generation.
    •   Support issuing accurate dispatch notices to quick-start generators.
    •   Support scheduling of next-hour and next-day deliveries of wind energy to the system.

Development and validation of the forecasting algorithms needed to meet the needs of the
control area operators and balancing authorities should begin now.

The system operators should be actively involved in the development and testing of the
algorithms to ensure that their needs are met.




3-10
                                 Wind Energy Forecasting Integration into Electricity Grid Operations




Figure 3-8
Example Display of Real-Time Forecasts for Oak Creek Energy Systems Wind Project in
Tehachapi, California [AWS Truewind, 2005]




                                                                                               3-11
4
CALIFORNIA REGIONAL WIND ENERGY
FORECASTING SYSTEM DEVELOPMENT


This section summarizes the results of the California Regional Wind Energy Forecasting System
Development project, completed by the California Energy Commission and EPRI in December
2002. It is excerpted from the final report, California Regional Wind Energy Forecasting System
Development, Volume 1: Executive Summary [EPRI-CEC PIER, 2006a].


Introduction

California has good potential for developing new wind generation capacity beyond the current
2030 MW, which is distributed between the five principle wind resource areas of the state
(Solano, Altamont, Pacheco, Tehachapi, and San Gorgonio), as shown in the California wind
power density map in Figure 4-1. Wind generation capacity is expect to grow rapidly in the
future in response to the California Renewable Portfolio Standard, which calls for 20%
renewables in the generation mix by 2017. For example, about 4000 MW of wind generation is
expected to be added to the existing 609 MW at Tehachapi.

Because wind generation is an intermittent resource and large concentrations of wind generation
can affect electricity grid operations and reserve requirements, development of accurate wind
energy forecasting tools will become an increasingly critical need for managing wind and other
intermittent generation resources connected to the California grid. Accurate next-hour and next-
day forecasts will make it possible to optimize the response to rapid changes in wind generation
to balance load and supply reserve and regulation resources to the grid.

In 2002, the California Energy Commission (CEC) and EPRI completed testing of two
forecasting systems at Altamont and at San Gorgonio [EPRI-CEC PIER, 2003a and 2003b;
EPRI 2003a]. Two wind energy forecasting system developers, Risoe National Laboratory and
TrueWind Solutions, applied their meteorology-based, meso-scale modeling algorithms to
generate twice-daily, 48-hour forecasts of hourly wind speed and energy generation, during a
12-month period.

The host wind projects were the 90-MW Wind Power Partners/WindWorks project, operated by
PowerWorks at Altamont Pass, and the 66.6-MW Mountain View 1 and 2 wind project, operated
by Seawest at San Gorgonio Pass. Based on the monthly and annual mean absolute errors (MAE)
of the forecast vs. observed data, the Risoe and TrueWind forecasts performed better than simple
persistence and climatology forecasts. However, the forecast errors were still significant,
indicating that additional research is needed to incorporate improved forecast technology and
forecast performance.

                                                                                             4-1
California Regional Wind Energy Forecasting System Development




  Rated Capacity – 2005 (MW)
  Solano County            180.8
  Altamont Pass            548.3
  Pacheco Pass              16.0
  Tehachapi Pass           608.7
  San Gorgonio Pass        615.9
  Others                     0.7
   TOTAL                 2030.4




      Figure 4-1
      California Mean Wind Power Map at 50-m Elevation and 2005 Rated Capacity of Wind
      Generation at Principle Wind Resource Areas [California Energy Commission]


In 2004, CEC, EPRI, and California Independent System Operator (Ca ISO) initiated a new
18-month project to build on the first project and develop and test improved wind energy
forecast algorithms for both short-term forecasts (regional five-minute forecasts over three
hours) and intermediate-term forecasts (hourly wind plant forecasts over 48 hours) in the
principle wind resource areas of the state. The project was completed during December 2005,
and the results are presented in the four-volume report, California Regional Wind Energy
Forecasting System Development and Testing [EPRI-CEC PIER, 2006a, 2006b, 2006c, and
2006d].

Objectives and Scope

The overall project objectives include both economic and technical goals.



4-2
                                         California Regional Wind Energy Forecasting System Development


The overall economic goals are:

    •   Support the California Independent System Operator’s (Ca ISO) development of a viable
        competitive market for intermittent wind resources.

    •   Pave the way for increasing market penetration of renewable resources.

    The overall technical goals are:

    •   Leverage the experiences gained under the prior forecasting efforts to improve forecast
        accuracy

    •   Provide capability to generate accurate forecasts for both short-term and longer-term
        forecast timeframes.

The specific objectives include:

    •   Develop and test short-term forecasting algorithms with higher accuracy than persistence
        forecasts to provide real-time forecasting capability and support system real-time
        updates to meet dispatching needs.

    •   Determine the sources of forecast error and assess methods to reduce errors for both
        next-hour and next-day forecasts, e.g., improved input data, finer grid sizes in meso-
        scale models, and improved statistical models for short-term forecasting and model
        operating statistics.

    •   Investigate wind flow and wind plant power curve variations over complex terrain via
        wind tunnel and numerical modeling.

The project scope includes:

   •    Generate real-time weather forecasts real time over a 4-km grid in both northern and
        southern California using the COAMPS meso-scale model.

   •    Develop and test wind energy forecast systems to provide forecasts for two “look-ahead”
        time horizons: (1) short-term forecasts of five-minute wind energy generation over a
        three-hour period to be issued every five minutes for the principle wind resource areas of
        the state (Solano, Altamont, Tehachapi, and San Gorgonio); and (2) intermediate-term
        forecasts of hourly wind generation over the a 48-hour period issued twice daily or every
        12 hours for wind plants in each of the principle wind resource areas.

   •    Conduct numerical and wind tunnel modeling of wind flow and power density at each
        wind turbine location vs. wind speed at a reference meteorological tower to investigate
        the variation of wind flow and wind plant power curve with wind speed and direction,
        atmospheric stability, and other conditions.

                                                                                                   4-3
California Regional Wind Energy Forecasting System Development


      •   Generate the California Wind Generation Research Dataset (CARD), a data base of daily
          forecasts of hourly wind generation at multiple elevations over 5-km grids in northern
          and southern California.

The project was conducted over the eighteen-month period, July 2004 through December 2005.

Project Participants

The project participants included the California Energy Commission as program manager, the
Electric Power Research Institute (EPRI) as project manager, EPRI subcontractors AWS
Truewind LLC, the University of California at Davis, and UC Davis subcontractor, Lawrence
Livermore Laboratory; project advisors, California Independent System Operator, National
Renewable Energy Laboratory, Southern California Edison, and five wind plant operators who
together with Ca ISO also provided wind resource and power data for their respective wind
projects, Sacramento Municipal Power District, PPM/High Winds, PowerWorks, Oak Creek
Energy Systems, and BMR/Mountain View 1 & 2.

Major Project Tasks

The project consisted of six major tasks: Task 1: Project Review and Reporting; Task 2: Wind
Resource Data Collection and Analysis; Task 3: Rapid-Update Wind Speed and Direction
Forecast Algorithm; Task 4: Regional Short-Term Wind Energy Forecasting System
Development and Testing; Task 5: Long-Term Wind Energy Forecasting System Development
and Testing; and Task 6: Wind Tunnel Testing Coupled with Advanced Numerical Model Data.

The project was conducted over the eighteen-month period, July 2004 through December 2005.

The following sections summarize the results of Tasks 2 through 6.


Next-Hour Regional Wind Energy Forecasting System Development and
Testing

This section summarizes the conceptual design of a two-stage short-term forecasting algorithm,
based on Artificial Neural Networks, and application of a portion of the algorithm to generate
five-minute forecasts over three hours for the principal California wind resource areas. The
detailed results are presented in [EPRI-CEC PIER, 2006b].

Objectives and Scope

The overall goal of the “Next-Hour Regional Wind Energy Forecasting System Development
and Testing” task was to produce an initial prototype of an integrated wind generation
forecasting system that can provide short-term regional power production forecasts in California.
The specific functional objective for this system is that it be capable of producing forecasts of the
regional power production in five-minute intervals for the next three hours after forecast

4-4
                                          California Regional Wind Energy Forecasting System Development

delivery. Thus each forecast should consist of predictions for the next 36 five-minute intervals.
Furthermore, the system should be capable of producing an update every five minutes.

The short-term forecast system is intended to provide a low-cost regional wind generation
forecasting tool to be used by Control Area Operators and Scheduling Coordinators at the
California ISO for more accurate and economical supplemental energy scheduling, real-time
dispatch, load following and AGC control. The ultimate objective is to augment traditional short-
term statistical forecasting techniques to include a consideration of local and regional
atmospheric predictors of short-term changes in weather conditions. The intention is to develop a
system that will generate regional power production forecasts even in the absence of site-specific
monitored data by utilizing all of the available local and regional forecasted and observed
weather data.

The work on this task was divided into two phases. The first phase designed a short-term forecast
system based on a review of a variety of forecasting methods and AWS Truewind’s experience
with operational forecast systems and ongoing forecast system research and development for
longer-term forecasts. The second phase implemented a portion of the system designed in the
first phase and assessed its performance using five-minute regional wind power production data
from the California ISO system.

Approach

The design process began with a comprehensive review of the key options for the forecast
system with a focus on what data could be used as input into the system and which techniques
could be used to produce the forecasts. It was recognized that zero- to three-hour forecasts in
five-minute intervals will necessarily rely heavily upon statistical time series prediction tools
since it will not be possible to gather sufficient data to initialize a high resolution physics-based
atmospheric model on the scale required to make zero- to three-hour forecasts in five-minute
intervals nor would it be possible to execute such models quickly enough to have their output be
useful in the forecast process. However, physics-based models run at moderately high resolution
every few hours may provide some useful trend information for forecasts for the zero- to three-
hour period. Therefore, the formulation of the forecast system began with a focus on statistical
methods for time series prediction that could be employed in a short-term forecast system. The
methods that were reviewed included classical time series prediction methods such as multiple
linear regression and ARIMA as well as newer techniques based on recent advances in learning
theory such as Artificial Neural Networks (ANN) and Support Vector Regression (SVR).

The information gathered during this review provided the basis for two major design decisions.
First, it was decided that the ANN technique should be used as the primary tool for the statistical
components of the forecast system. Second, it was decided that the system should not be based
on a single forecast procedure that could be argued to be the best overall choice, but instead
should be based upon an ensemble of three forecast methods that have significantly different data
utilization and model formulation characteristics. The concept is that each component of the
system will have its strengths and that the composite of the forecasts produced by each of the
component methods will yield better overall performance than using only one forecast
procedure.

                                                                                                    4-5
California Regional Wind Energy Forecasting System Development


In order to incorporate this ensemble approach, the design of the system included four forecast
subsystems. Three of the subsystems produce quasi-independent preliminary power production
predictions for all of the 36 five-minute intervals in the three-hour forecast window. The fourth
subsystem blends the three separate preliminary forecasts into the final forecast by weighting
each forecast in accordance with its performance characteristics (i.e., placing a greater weight on
the forecast that is likely to do better under a specific set of conditions).

As noted previously, the design objective for the short-term forecast system was to go beyond
the basic use of power production time series information from the individual wind resource
regions, subregions or wind plants. In addition to this power production time series data, it is
envisioned that the following information will also be used in the forecast system: (1) time
series information of meteorological parameters from meteorological towers operated by wind
generators or other members of the wind energy community; (2) meteorological data from
surface weather observing sites operated by the National Weather Service and other
organizations; (3) meteorological data from remote sensing systems such as wind profilers,
Doppler radars and satellite-based sensors; and (4) short-term forecast data from high resolution
atmospheric physics-based models run in a rapid update cycle mode (i.e., assimilation of new
data and execution of short-term forecasts every few hours or possibly every hour).

Forecast System Design

An initial forecast system design was formulated to address the previously noted design
objectives. Figure 4-2 presents a schematic of the system. The circles represent input or output
data, while the rectangles depict algorithms (i.e., numerical models) that operate on the data.
The three columns of boxes that originate below the row of circles represent three forecast
subsystems. The leftmost column is the autoregressive subsystem. This system generates a
forecast based solely on the recent behavior (most recent level of production, trends, rate of
change of trends etc.) of the production in the region (or wind plant). The second subsystem
utilizes recent meteorological data within and in proximity to the region of interest. It finds and
utilizes spatial and temporal meteorological relationships that have predictive value for the
regional power production. The third (rightmost) subsystem constructs a regional power
production forecast from the output of frequently updated very high resolution physics-based
forecast simulations of the wind over the region of interest. Each of these subsystems supply an
independent forecast to the fourth subsystem, which constructs the final forecast by weighting
each of the three input forecasts according to their recent performance characteristics.

Due to limitations in the availability of data and the resources available for this task, it was only
possible to implement and test one of the three subsystems. The implemented and tested
subsystem was based on the approach that employed an ANN with autoregressive-type inputs
from the time series of the regional power production. No meteorological data are used in this
approach. The performance of this approach was evaluated with regional power production data
for the year 2004 supplied by the California ISO for the San Gorgonio, Tehachapi, Pacheco,
Altamont and Solano regions.




4-6
                                      California Regional Wind Energy Forecasting System Development




      Power              Wind Plant
                         Met Tower                Regional                  Remotely
    Production                                    Surface
                           Time                                              Sensed
      Time                                        Met Data
      Series              Series                                            Met Data




                                                               Local-area
                                                             Physics-based
      ANN                         ANN                         Atmospheric
                                                                 Model




                                Regional                        Model
       Power                                                    Output
     Production                  Power
                                                               Statistics
      Forecast                   Model                          (MOS)




                                  Power                        Regional
                                Production                      Power
                                 Forecast                        Model




                                                                 Power
                                                               Production
             ANN                                                Forecast




              Final
             Power
           Production
            Forecast


Figure 4-2
Schematic of the proposed two-stage short-term (0- to 3 hour) forecast system




                                                                                                4-7
California Regional Wind Energy Forecasting System Development



Results

After the forecast sensitivity experiments were completed, zero- to three-hour ANN-based
forecasts in five-minute intervals were produced for four of the five regions for which data were
available for the entire year of 2004. The one-year test indicated that the forecast performance
was substantially different during the cold and warm seasons. This is not very surprising since
the wind regimes in California are substantially different between the cold and warm seasons.

Figures 4-3 and 4-4 present the average mean absolute errors and skill scores (i.e., the percentage
reduction in the mean absolute error) relative to a persistence forecast vs. look-ahead time for all
four regions and for the warm and cold seasons. The two charts indicate that the reduction in
MAE relative to a persistence forecast (i.e., the skill score) is much higher during the warm
season than during the cold season. However, the actual MAE values are typically somewhat
lower in the cold season.

This dichotomy between the MAE values and skill scores is mostly attributable to the fact that
the cold season is characterized by a wind regime that consists of relatively long periods of light
winds (often below the turbine start-up speeds) with little variability and that are interrupted by
short periods of high wind speeds with high variability. The short periods with high wind speeds
are associated with storms that typically move into California from the Pacific Ocean, and the
longer periods with light winds correspond to the quiescent period between storm events.
A persistence forecast typically performs quite well during the light-wind periods since the
power production is often zero for many consecutive hours during those periods. Persistence will
perform poorly during the storm events, but unfortunately, a forecast system based solely upon
time series of power production data will also perform poorly since the historical time series of
data will not provide much information about approaching storm systems since storms are
irregular events. It is necessary to utilize real-time meteorological data from a variety of regional
sites and physics-based atmospheric model output data to obtain some predictive skill for these
types of events. The use of these two additional forecasting resources is included in the short-
term forecast system design. Since the overall level of power production is low during the cold
season, the level of forecast performance is typically not as critical.

The high levels of production typically achieved during the warm season make it more critical
for a forecast system to perform well during that portion of the year. As noted earlier, the MAE
values of the autoregressive-type ANN forecast system component are typically slightly higher
(as a % of capacity) in the warm season but the skill scores are also significantly higher. The
average skill scores shown in Figure 4-4 indicate that a persistence forecast outperforms the
autoregressive ANN-based forecast for the first two forecast intervals (0 to 10 minutes) for both
seasons. However, the autoregressive forecasts outperform persistence by an increasing margin
from the 15-minute mark through the end of the three-hour forecast period during the warm
season. During this period, the average skill score relative to a persistence forecast rises from just
above 0% to approximately 17%.




4-8
                                           California Regional Wind Energy Forecasting System Development


                              Mean Absolute Error: 0-3 hr Forecasts
                                      4-Region Average
                                            Cold Season   Warm Season

                 10
                  9
                  8
                  7
                  6
                  5
                  4
                  3
                  2
                  1
                  0




                      110
                      115
                       10
                       15
                       20
                       25
                       30
                       35
                       40
                       45
                       50
                       55
                       60
                       65
                       70
                       75
                       80
                       85
                       90
                       95
                        5




                      100
                      105


                      120
                      125
                      130
                      135
                      140
                      145
                      150
                      155
                      160
                      165
                      170
                      175
                      180
                                  Forecast "Look-Ahead" Period (minutes)


Figure 4-3 Mean absolute errors of regional power generation forecasts (% of rated
capacity) vs. look-ahead period for the warm (May-Oct) and cold (Jan-Apr, Nov-Dec)
season months of 2004, ANN-based autoregressive forecasts (stage 1 – method 1), and the
four largest California wind resource areas.


                          Skill Score vs. Persistence: 0-3 hr Forecasts
                                        4-Region Average and Aggregate
                          Cold Season      Warm Season    4-Region Warm Season Aggregate

                 35%
                 30%
                 25%
                 20%
                 15%
                 10%
                  5%
                  0%
                  -5%
                 -10%
                 -15%
                        110
                        115
                         10
                         15
                         20
                         25
                         30
                         35
                         40
                         45
                         50
                         55
                         60
                         65
                         70
                         75
                         80
                         85
                         90
                         95
                          5




                        100
                        105

                        120
                        125
                        130
                        135
                        140
                        145
                        150
                        155
                        160
                        165
                        170
                        175
                        180




                                    Forecast "Look-Ahead" Period (minutes)



Figure 4-4 Skill scores of regional power generation forecasts relative to persistence
forecasts vs. look-ahead period for the warm (May-Oct) and cold (Jan-Apr, Nov-Dec)
season months of 2004, ANN-based autoregressive forecasts, (stage 1 – method 1) and the
four largest California wind resource areas




                                                                                                     4-9
California Regional Wind Energy Forecasting System Development


It should be noted that the skill scores in Figure 4-4 are unweighted mean values. The average
skill score would be somewhat higher if a capacity-weighted average was employed, since the
regions with lower rated capacity (e.g., Solano) tend to have lower skill scores.

The forecast performance for the aggregated power production from all four regions is
substantially better than those for the individual regions.

During the warm season, the MAE for forecasts of the four-region aggregate power production is
slightly under 0.5% for a five-minute forecast and rises to only about 4% for a three-hour ahead
forecast of the five-minute power production. The warm season skill scores for the four-region
aggregate forecast are quite impressive (Figure 4-4). They rise from just under 0% for a five-
minute-ahead forecast to approximately 30% for a three-hour-ahead forecast.

Conclusions

The key results of the short-term forecast system development and testing are:

    1. Review of statistical and physics-based forecast methods and analysis of the spatial and
       temporal characteristics of the wind speed and direction variability to provide the
       foundation for the design of a robust short-term forecasting system capable of producing
       forecasts of the 5-minute regional energy production for the next 3-hour period updated
       every 5 minutes;

    2. Design of a two-stage forecast system with the first stage consisting of a mini-ensemble
       of three forecast methods that each produce an independent power production forecast,
       and a second stage that weights each of the three forecasts from the first stage based on
       their recent performance to produce a single composite forecast and an estimate of
       forecast uncertainty; and

    3. Initial testing of one of the three forecast methods (the autoregressive method) in the
       first-stage of the proposed forecast system with one year (2004) of regional wind power
       production data supplied by the California ISO.

The main conclusions are:

    1. The autoregressive component of the three-method first stage of the proposed forecast
       system produced a reduction in the regional power production forecast error of a
       persistence forecast for the warm season (May–October) that ranged from near 5 to 10%
       for the first 30 to 60 minutes of the 3-hr forecast period to the 15% to 20% range during
       the latter stages of the period; and

    2. The autoregressive method showed virtually no improvement over persistence for the
       cold season, but that is to be expected due to the character of the wind during the
       California cold season, and is one of the reasons for incorporating two other methods into
       the forecast system.


4-10
                                         California Regional Wind Energy Forecasting System Development


Recommendations

The recommended next steps are:

   1. Implement and evaluate the two other component forecast subsystems and the ensemble
      compositing subsystem that were not implemented in this project;

   2. Implement and evaluate the performance of the entire forecast system for several months
      in different seasons for each of the wind resource areas in California; and

   3. Test the system in an operational environment and obtain feedback from CA ISO
      personnel.

An alternative approach would be to test and implement the same subsystem tested in this project
and then develop the other components while the single subsystem forecast system is in
production. The results produced in this project suggest that this approach could yield a 5% to
20% improvement over a persistence forecast, especially for the warm season.


Next-Day Wind Plant Energy Forecasting System Development and Testing

This section summarizes testing of various improvements in data, physics-based atmospheric
numerical models, and other features of the next-day forecast system and application of the
resulting improved forecast system to five wind projects in the principle wind resource areas.
The detailed results are presented in [EPRI-CEC PIER, 2006b].


Objectives

There are two primary objectives to the “Next-day Wind Plant Energy Forecasting System
Development and Testing” task. The first was to formulate and evaluate new forecast methods
and datasets that have the potential to improve the forecast performance of next-day wind power
production forecasts relative to the performance in the previous CEC-EPRI wind energy
forecasting project [EPRI, 2003a, 2003b, and 2003c]. The second was to test the performance of
the modified forecasting system over a one-year period for a set of California wind plants.


Scope

The task was conducted in two phases: a screening phase in which a variety of potential forecast
system improvements were tested using data from the previous project; and an evaluation phase
in which forecasts were generated by a modified version of AWS Truewind’s eWind forecast
system and evaluated for a set of five participating California wind plants.




                                                                                                 4-11
California Regional Wind Energy Forecasting System Development


Approach

The screening phase of the project was built upon the data and results from the forecast
evaluation experiment conducted during the 2001-2002 period as part of the previous CEC-EPRI
forecasting project [EPRI, 2003a, 2003b, and 2003c]. In the previous project, 48-hour forecasts
were generated for two California wind plants twice each day for a one year period (October
2001 to September 2002) by two forecast providers: AWS Truewind, LLC (AWST) and Risoe
National Laboratory (Risoe) of Denmark. Each forecast provider generated and delivered two
48-hour forecasts per day of the hourly average power production (kW) and the average hourly
wind speed (m/s) and direction for one meteorological tower for each wind plant. The morning
forecast was delivered at 8:00 AM PST and the forecast period extended until the hour ending at
8:00 AM PST two days after the forecast was delivered. The evening forecast was delivered
12 hours later at 8:00 PM PST and its forecast period ended 12 hours after the end of the
morning forecast.

It is important to note that these forecasts were generated in “next-day” mode. This means that
the forecasts were produced without real-time data from the wind plants. Real-time data from the
wind plants are very important for the performance of forecasts during approximately the first six
to nine hours of the forecast period. However, real-time data from the plant have little impact on
forecasts beyond this period.

The two wind plants that participated in the previous project were the Mountain View 1 and 2
wind plant in the San Gorgonio Pass of southern California and the PowerWorks plant located in
the Altamont Pass. The rated capacities of the Mountain View and PowerWorks wind plants are
66.6 and 90 MW, respectively. The screening phase of the current project was designed to test
potential improvements to the forecast system on the months in the earlier project that had the
largest forecast errors.

The evaluation phase of the current task was designed to be similar in structure to that of the
previous project. Forecasts of the hourly power production and wind speed for each of the next
48 hours for each participating wind plant were simulated for twice daily delivery. The forecasts
were generated by a modified version of AWS Truewind’s eWind system that employed many of
the system enhancements that were found to be useful in the screening phase of the project.
However, the number of participating wind plants was expanded from two in the previous
project to five in this project. Fortunately, the two plants that participated in the previous project
also participated in the current project. This provided an opportunity to directly assess the change
in forecast performance between the previous and current projects for the same wind plants. The
evaluation metrics used in the current project were also kept the same as those used in the earlier
project to facilitate comparison. These included, the mean error (ME), the mean absolute error as
a percentage of installed rated capacity and also as a percentage of actual production, and the
skill scores (i.e., the percentage reduction in MAE) relative to persistence and climatology
forecasts.




4-12
                                          California Regional Wind Energy Forecasting System Development


Phase 1: Screening of Improved Data and Forecast Methodologies

The Phase 1 screening assessment of potential improvements in the input data and forecast
methods was based on the measurement and forecast data from the previous CEC-EPRI project
for the Mountain View 1 and 2 and PowerWorks wind projects.

It was not practical to execute a large number of forecast method experiments for the full
12 months of the previous forecast evaluation period for both wind plants. Therefore, it was
decided to test each forecast system improvement using a three-month data sample for each of
the two wind plants. The test months were selected independently for the PowerWorks and
Mountain View plants. The main selection criteria were that (1) plant data were available for a
large fraction of the hours in the months (i.e., low lost data rate); (2) the forecast performance
was below the average for that plant; and (3) the months should be selected such that they
include a winter, summer and transition season month.

Figure 4-5 is a schematic overview of the components of the eWind forecast system. The system
consists of two major components: data and numerical models. There are three fundamental
types of data used in the forecasting process: (1) regional weather data, which consists of data
from a variety of sources including surface-based meteorological sensor arrays at airports,
rawinsonde balloons and satellite-based sensors; (2) time series of power production and




   Figure 4-5 Schematic of the eWind forecast system




                                                                                                  4-13
California Regional Wind Energy Forecasting System Development


meteorological data from the wind plant; and (3) off-site local meteorological data. There are
also three major types of numerical models used in the forecast system: (1) physics-based
atmospheric models; (2) statistical prediction models and (3) a plant output model.

There are almost an infinite number of different configurations and new techniques that can be
implemented within the general framework of the eWind system. Obviously, it was only possible
to test a very small subset of the possible changes within the resource limitations of this project.
The modifications tested within this task were based upon two criteria: (1) those that appeared to
have the greatest promise of improving the forecasts; and (2) those that could be tested with the
data available for this project and within the resource limitations of the project. The forecast
system experiments were designed to test potential improvements to each part of the forecast
system. Modifications were tested within the six focus areas listed in Table 4-1.

Focus Area 1 addressed the input data used by the physics-based model. There are many
emerging datasets that have the potential to improve the simulation of winds in physics-based
models. This project evaluated the impact of using higher-resolution water surface temperatures
from the current generation of satellite sensors.

Focus Area 2 assessed the impact of the resolution of the physics-based model grid. Smaller grid
cells mean that the physics-based model can simulate smaller scale weather features that
determine the evolution of the wind in the vicinity of the wind plant. The impact of utilizing a
higher resolution grid for the physics-based model was evaluated.

Focus Area 3 assessed the impact of employing a “next generation” physics-based atmospheric
model in place of the MASS model currently used in the eWind system. The next generation
model known as WRF was used to assess the impact of this potential improvement. The WRF
model is an open-source community model being jointly developed by the National Center for
Atmospheric Research (NCAR) and the U.S. National Weather Service.

Focus Area 4 examined the impact of using more sophisticated statistical techniques for the
Model Output Statistics (MOS) component of the forecast system.

Focus Area 5 investigated the sensitivity of the forecast performance to the formulation of the
plant scale power curve.




4-14
                                               California Regional Wind Energy Forecasting System Development

   Table 4-1
   Focus Areas and Associated Experiments

                      Focus Area                                             Experiments
   1. Additional or improved input data for the          4-km MODIS and Pathfinder Water Surface
      physics-based simulations                          Temperature (WST) was used to initialize the
                                                         physics-based model
   2. Higher resolution (i.e., smaller grid cells) for   4-km and 1-km physics-based simulations were
      physics-based model                                executed
   3. “Next-generation” physics-based models             Forecast simulations were executed with the WRF
                                                         model in place of the MASS model in the eWind
                                                         system
   4. Advanced statistical models for MOS                Stratified multiple linear regression and artificial
                                                         neural network methods were used for MOS in
                                                         place of the screening multiple linear regression
   5. More sophisticated formulation for plant           Median-based power curve plus residual
      output model                                       formulation was used
   6. Forecast Ensembles                                 An ensemble of MASS and WRF model forecasts
                                                         was used in place of the single MASS forecast

Finally, Focus Area 6 examined the impact of employing a forecast based on a composite of
individual forecasts from different methods.

Results

Table 4-2 summarizes the impact of each of the forecast system modifications on the wind
energy forecast performance. The MAE percentage reductions in Table 4-2 are the net reductions
in the MAE for all six of the evaluation months (for both the Mountain View and PowerWorks
wind plants) or as many months as were available for a particular experiment. For a variety of
reasons, the baseline method is not the same for each focus area. Therefore, Table 4-2 also notes
the baseline method for each focus area.




                                                                                                           4-15
California Regional Wind Energy Forecasting System Development

    Table 4-2
    Summary of the Improvements in the Power Production Forecast MAE associated with the
    Forecast System Modifications in each Focus Area


        Focus                          Forecast System                           MAE
         Area                            Modification                        Reduction (%)
                    MODIS and Pathfinder WST data (4 km)
           1                                                                      12.3%
                    BASELINE: NCEP OI WST (110 km)
                    1 km physics-based model grid
           2                                                                      -4.7%
                    BASELINE: 10 km physics-based model grid
                    WRF with 40 km grid as the physics-based model
           3                                                                      4.0%
                    BASELINE: MASS with 40 km grid
                    Stratified 2-stage SMLR scheme
           4                                                                      15.8%
                    BASELINE: Screening Multiple Linear Regression
                    Model deviations from power curve
           5                                                                      4.6%
                    BASELINE: Speed-based plant-scale power curve
                    Mean of an ensemble of forecasts
           6                                                                      0.8%
                    BASELINE: "Best" single forecast method

The results indicate that the most significant improvements in forecast performance were
associated with two of the six enhancements in Table 4-2. They are enhancements 1 and 4, which
reduced mean absolute error by 12.3% and 15.8%, respectively. Enhancement 1 uses
higher-resolution water surface temperature data derived from satellite-based sensors to initialize
the surface temperature parameter in the physics-based model; and Enhancement 4 uses the
stratified two-stage screening multiple linear regression (SMLR) approach to generate the MOS
factors in the forecast system. The other forecast system modifications yielded smaller
improvements in forecast performance.

Interestingly, the use of higher-resolution grids in the physics-based model did not improve the
ultimate wind speed or power production forecasts. Although the higher resolution grids did
improve the performance of the raw wind speed forecasts produced by the physics-based model,
after applying the MOS procedure, the additional forecast performance improvement was greatly
diminished. That is, the MOS procedure and higher-resolution grid appear to contribute similar
information and result in similar forecast performance improvements of the raw physics-based
forecasts.

Another noteworthy observation is that the ensemble approach may have greater potential for
improving forecast performance than indicated by this experiment. Others studies have shown
that the ensemble approach tends to produce more significant forecast improvement with a larger
set of ensemble members than used in the screening evaluation reported here. Note that the
forecast evaluation phase reported in the next section did use a larger set of ensemble members.




4-16
                                          California Regional Wind Energy Forecasting System Development


Phase 2: Evaluation of Improved Forecast Algorithm at Five Wind Projects

Approach

Phase 2 assessed the performance of the modified forecast system for a set of California wind
plants for a one-year evaluation period extending from July 1, 2004 to June 30, 2005.

Five wind plants agreed to provide power production and meteorological data for use in the
generation and evaluation of power production and wind speed forecasts for their plants. The
participating plants were: (1) the Mountain View plant, a 66.6 MW facility in the San Gorgonio
Pass of southern California; (2) the Oak Creek plant, a 34.5 MW plant in the Tehachapi Pass,
which is adjacent to the Mojave Desert; (3) the PowerWorks wind plant with a rated capacity of
90 MW in the Altamont Pass, which is located just to the east of the San Francisco Bay Area;
(4) the 15.18 MW SMUD wind plant located in the Montezuma Hills in Solano County; and
(5) the 162 MW High Winds Energy Center which is adjacent to the SMUD plant in Solano
County.

The protocols for the forecast system test were the same as those used for the 2001-2002
evaluation period in the previous CEC-EPRI project (EPRI, 2003a, 2003b, and 2003c): 48-hour
forecasts of the hourly power production and wind speed were generated on a twice-daily cycle
with scheduled delivery times of 8:00 AM and 8:00 PM PST each day. Three different physics-
based models and four different Model Output Statistics (MOS) procedures were used to produce
an ensemble of 12 different forecasts for each of the wind plants. This fairly large set of forecasts
provided an opportunity to evaluate a number of different forecast methods as well as the
performance of a forecast produced by averaging all 12 of the individual forecasts (an ensemble
mean).

The forecasts were generated via a mixture of real-time and historical modes, although all
forecasts only used data that were or would have been available to the forecast system in an
operational real-time mode. In order to construct a more detailed evaluation of the forecast
performance, a comprehensive analysis was only done for the forecasts generated during the
morning forecast cycle (i.e., the forecasts scheduled for delivery at 8:00 AM each day).

The performance statistics indicate that the forecast performance differences between the
morning and afternoon cycles were not significant and that all of the significant conclusions from
the analysis of the performance of the morning forecast cycle would apply to the combined pool
of both the afternoon and morning forecast cycles.

There was a potential of forecast hours was 17,520 forecast hours (365 days x 48 forecast hours
per day) in the forecast evaluation pool for each wind plant and for the morning forecast cycle.
Unfortunately, the actual number of hours in the verification pool varied, and in many cases was
substantially less than the maximum possible number of forecast hours. The most common
contributor to missed forecasts was unavailable (i.e., missing) observed data from the wind
plants. The number of unavailable data varied substantially between the participating wind
plants.


                                                                                                  4-17
California Regional Wind Energy Forecasting System Development


Results

Table 4-3 presents the annual Mean Absolute Errors (MAE) and skill scores for all 48 forecast
hours for both wind speed and energy forecasts for each of the five participating wind plants.
The results correspond to the best-performing forecast method for each wind plant over the entire
one-year period.

The average MAE of the power production forecasts for all five participating wind plants was
14.5% of installed capacity and 52.7% of the average production. The skill scores of the power
production forecasts were 33.2% vs. persistence and 29.5% vs. climatology.

It should be noted that the skill scores relative to climatology are most likely too low. This is
because no climatological data were available for three of the wind plants and the actual monthly
mean wind energy generation for each hour of the day was assumed for the climatology value
instead of independently-collected climatology data. Therefore, the forecast error of the resulting
climatology forecast is lower than would be expected if actual climatology data had been used.
The average mean absolute error of the wind speed forecast was 2.27 m/s, which is 34.1% of the
average wind speed.

Two of the five participating plants (Mountain View and PowerWorks) also participated in the
forecast evaluation experiment conducted in the previous CEC-EPRI forecasting research project
[EPRI-CEC PIER 2003q, 2003b, and 2003c]. This provided an opportunity to compare the
results from this project to those obtained in the previous project. This comparison indicated that
there was considerable improvement in almost all of the forecast performance statistics.

For the Mountain View wind plant, the wind energy forecast MAE decreased from 16.6% of
rated capacity in the previous project to 13.0%, and the wind speed MAE for the Catellus Tower
decreased from 3.05 m/s to 2.65 m/s. In addition, the skill score vs. persistence increased from
37.5% to 40.5% and the skill score vs. climatology increased from 36.4% to 47.7%.

For the PowerWorks wind plant, the wind energy forecast MAE decreased from 14.1% of rated
capacity to 11.9%, while the wind speed forecast MAE for PowerWorks Tower M438 decreased
from 1.93 m/s to 1.78 m/s. In addition, the skill score of the power production forecasts vs.
climatology increased from 30.9% to 38.9%. These statistics indicate that the forecast system
enhancements substantially improved forecast performance relative to the previous project.
    Table 4-3. Mean absolute errors and skill scores of the best-performing power production
    and wind speed forecast methods for each wind plant, 48-hour forecasts, and July 1, 2004
    to June 30, 2005 forecast period

                            Power Production Forecast MAE                           Wind Speed Forecast MAE
       Site          MAE           MAE           Skill          Skill        MAE        MAE           Skill          Skill
                  % of Capcity   % of Prod   vs Persistence vs Climatology   m/s     % of Speed   vs Persistence vs Climatology
  Mountain View     13.0%         36.6%         40.5%          47.7%         2.65      27.5%         40.1%          27.0%
    Oak Creek       15.0%         57.1%         33.2%          27.1%         2.03      40.6%         32.8%          43.8%
   PowerWorks       11.9%         58.5%         26.5%          38.9%         2.52      34.4%         28.2%          12.4%
      SMUD          16.0%         60.8%         37.1%          16.1%         1.98      35.9%         31.3%          10.9%
    HighWinds       16.8%         50.7%         28.6%          17.5%         2.16      31.9%         27.1%           9.9%
     Overall        14.5%         52.7%         33.2%          29.5%         2.27      34.1%         31.9%          20.8%


4-18
                                          California Regional Wind Energy Forecasting System Development



Variability of Power Generation Forecast Error

The forecast performance results raise a number of issues and questions about the nature of the
variability in forecast performance. For example, the annual power-generation forecast MAE
exhibits considerable variability between the wind plants, ranging from 11.9% of rated capacity
at the PowerWorks plant to 16.8% at the High Winds plant.

Furthermore, the performance statistics indicate that the wind speed forecast MAEs are not
highly correlated with the wind energy forecast MAEs. For example, the SMUD wind plant had
the lowest wind speed MAE but the second highest power production MAE, while the Mountain
View wind plant had the highest wind speed MAE but the second lowest wind energy forecast
MAE. Clearly, the magnitude of the wind speed MAE is not the determining factor for the wind
energy MAE.

An obvious question is: what factors are responsible for this variability? It is clearly of critical
importance to understand the factors that contribute to forecast performance variability in order
to determine the direction of future efforts to improve forecast performance.

Therefore, a detailed analysis of the forecast errors for three wind plants was conducted. By
eliminating the differences in wind speed errors between the wind plants, it was possible to gain
considerable insight. This was done by assuming a constant two meter/sec wind speed error,
randomly distributed between positive and negative deviations from the observed values for all
hours of the year. The resulting wind energy forecast MAEs were 8.8% of rated capacity at the
Mountain View 1 and 2 wind plant and 13.1% and 14.7% at the High Winds and SMUD wind
plants, respectively. Thus, even with similar wind speed errors, the wind energy generation
MAEs still vary significantly between the wind plants.

The principle reason for the energy-forecast MAE variation between plants appears to be related
to differences in the maximum slopes of the plant-scale power curves and the wind speed
frequency distributions between the plants. Figure 4-2 shows empirical plant-scale power curves
for each of the three wind plants. The general shapes of the curves imply that the sensitivity of
wind energy forecast error to wind speed forecast error varies with wind speed. When the wind
speed is in the range corresponding to the steeply-sloped middle section of the plant-scale power
curve, the wind energy forecast error should be most-sensitive to wind speed forecast errors.
Thus, the maximum wind energy forecast errors should occur at sites where the wind speed
spends many hours in the middle range of the power curve.

However, the maximum slopes of the power curves also vary between the wind plants, and
plants with higher maximum slopes should also exhibit higher wind energy forecast errors. The
slopes are mostly determined by the correlation between the wind speeds at the individual wind
turbine locations. This was verified by the “constant 2 m/s error experiment”. The SMUD wind
plant with the steepest-slope power curve also exhibited the highest wind energy forecast MAE
when a constant wind speed forecast error of 2 m/s was assumed. The Mountain View plant with
the lowest maximum power curve slope exhibited the lowest energy forecast MAE.



                                                                                                  4-19
California Regional Wind Energy Forecasting System Development



                     Empirical Plant-Scale Power Curves
                             Mountain View       SMUD        High Winds
          120%

          100%

           80%

           60%

           40%

           20%

             0%
                   0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
                                   Anemometer Wind Speeds (m/s)

    Figure 4-2
    Empirical median plant-scale power curves derived from measured wind energy
    generation and wind speed data for three of the participating wind plants


Ensemble Forecasts

Applications and testing of composite forecasts based on an ensemble of individual forecast
methods have demonstrated improved forecast performance in other meteorological forecast
applications. In order to test the value of this concept in wind power production forecasting, a
composite forecast was generated from the ensemble of 12 individual forecasts that were
generated in this project. The 12 individual forecasts consisted of four different MOS procedures
applied to each of three physics-based models. The full suite of methods was only available for
the four-month period extending from March 2005 through June 2005 so the evaluation of the
ensemble approach was limited to that period.

Table 4-4 summarizes the wind speed and energy forecast MAEs for four combinations of power
generation and wind speed forecasts from the 12-member ensemble. The first column of the
power production and wind speed sections of the table lists the MAE of the ensemble-mean
forecast. This forecast was constructed by calculating the average of all available members of the
12-member ensemble for each hour of every forecast cycle.

The next column to the right, labeled “Best Overall Method,” lists the MAE of the method that
had the lowest MAE over the entire four-month period for each wind plant. This is the method
that would typically be used in an operational forecast environment, since it would most likely be
classified as the best method when reviewing performance statistics compiled over a long period.
4-20
                                                        California Regional Wind Energy Forecasting System Development

     Table 4-4
     Mean absolute errors of ensemble power production (% of rated capacity) and wind speed
     (m/s) forecasts for the five participating wind plants

                               Power Production Forecast MAE                              Wind Speed Forecast MAE
      Month          Ensemble-12   Best Overall   Best Monthly   Average    Ensemble-12    Best Overall   Best Monthly   Average
                        Mean         Method         Method       of MAEs       Mean          Method         Method       of MAEs
  Mountain View        16.4%         16.0%          15.7%        17.9%         2.62           2.66           2.63         2.89
    Oak Creek          18.9%         17.3%          17.3%        20.7%         2.07           2.14           2.05         2.25
   PowerWorks          13.9%         14.1%          13.2%        14.5%         2.41           2.48           2.38         2.57
      SMUD             17.1%         17.4%          17.3%        18.6%         1.82           1.87           1.80         1.94
    HighWinds          18.4%         18.3%          17.6%        19.3%         1.95           2.02           1.95         2.11
    4-months           17.0%         16.6%          16.2%        18.2%         2.17           2.24           2.16         2.35


Notes: The ensemble forecasts include: (1) averaging all the forecasts in a 12-member ensemble (“Ensemble-12 Mean”), (2) the
       individual method with the lowest MAE for the 4-month period (“Best Overall Method”), (3) a composite of the individual
       methods with lowest MAE in each month (“Best Monthly Method”); and (4) the average MAE of all 12 members of the
       ensemble (“Average of MAEs”).


The next column to the right, labeled ”Best Monthly Method,” lists the composite MAE of the
forecast created by taking the method with the lowest MAE for each of the four months. It would
be difficult to use this approach in an operational environment, since one would have to identify
which method was going to perform best for a particular month before the month began.
However, it may be possible to develop an “intelligent ensemble composite” that varies the
weight placed on different members of the ensemble in the ensemble-composite forecast as a
function of parameters which indicate which ensemble members are likely to perform better for a
particular forecast cycle. This is a possible area for future research.

The rightmost column, labeled “Average of MAEs,” represents the average MAE of all 12
ensemble members. This is the forecast MAE that is most likely if one had no knowledge about
the relative performance of the forecast methods and randomly selected a forecast method each
day.

The results indicate that the performance of the ensemble mean forecast was substantially better
for the wind speed forecasts than for the power production forecasts. For the wind speed
forecasts, the ensemble-mean method clearly outperformed the best overall method. The five-
plant average MAE for the ensemble-mean wind speed forecasts was 3.1% lower than the
composite MAE of the best overall forecast for each wind plant. In fact, the MAE of the
ensemble-mean was only very slightly higher than the MAE of the “Best Monthly Forecast”.
Furthermore, the ensemble-mean forecast had a lower MAE than the best overall method for all
five of the wind plants. The ensemble-mean forecast also had an MAE that was almost 8% lower
than that of the average MAE of all of the ensemble members.

However, the ensemble-mean power production forecast had a slightly higher four-month
average MAE (17.0%) than the composite of the best overall method for each plant (16.6%).
Much of the difference is attributable to very poor forecast performance for the Oak Creek plant,
because one plant output model significantly outperformed all of the other types due to unusual
systematic variations in the plant-scale power curve. Excluding the Oak Creek plant, the
performance of the ensemble-mean power production forecast was about the same as that of the
best overall method for each plant. This is still not as good as the performance of the ensemble-

                                                                                                                             4-21
California Regional Wind Energy Forecasting System Development


mean for the wind speed forecasts. The reason for this is not clear but may be related to the fact
that there was less diversity in plant output models than in other parts of the system. This will
require further investigation.

These limited tests indicate that the ensemble approach appears to have at least modest potential
to improve wind speed forecast performance.

The simple approach of constructing an average wind-speed forecast from all ensemble members
could reduce wind speed MAE by 3% to 5% relative to the best overall forecast method. The
improvement could increase if an intelligent ensemble composite can be constructed by
weighting certain ensemble members more heavily when key parameters indicate they are more
likely to produce better forecasts.

For power production forecasts, further refinements of the ensemble forecasting approach are
needed. One possible approach is to apply adjustments to the plant-scale power curve to account
for atmospheric stability based on rapid-update, high-resolution simulations of the wind flow
field around the wind turbines, as discussed in the next section.

Conclusions

The results of the forecast experiments conducted in the screening phase indicated that:

    1. The use of improved higher resolution water surface temperature values as input to the
       physics-based model and a more sophisticated statistical procedure in Model Output
       Statistics (MOS) component of the forecast system had the greatest positive impact on
       forecast performance;

    2. The use of a higher resolution grid (i.e., smaller grid cells) for the physics-based model
       generally improved the performance of the raw physics-based model forecasts but did not
       significantly improve the performance of forecasts after the MOS procedure was applied;

    3. The use of a next generation physics-based model (the WRF) and a more sophisticated
       plant output model yielded modest improvements in forecast performance in the
       experimental sample, but the significance of these improvements for a larger sample was
       questionable; and

    4. Forecasts generated by computing the mean from an ensemble of two forecasts from
       different physics-based models produced an insignificant improvement in performance,
       but this disappointing performance may have been attributable to the small size and
       limited diversity of the ensemble.

The results of the one-year forecast evaluation experiment indicated that:

    1. The average mean absolute error (MAE) of the 48-hour power production forecasts was
       14.5% of installed capacity and 52.7% of average production for the entire one-year
       period and the five participating wind plants;


4-22
                                         California Regional Wind Energy Forecasting System Development

   2. The average MAE of the 48-hour wind speed forecasts was 2.27 m/s or 34.1% of the
      mean wind speed for the one-year period at the meteorological tower sites within each
      wind plant;

   3. There was a considerable improvement in forecast performance between the forecast
      evaluation conducted in the previous (2001-02 period) and current (2004-05 period)
      projects for the two wind projects that participated in both projects; the MAE of the
      annual power production forecast MAE decreased from 16.6% to 13.0% for the Mountain
      View 1 and 2 plant at San Gorgonio, and from 14.1% to 11.9% for the PowerWorks plant
      at Altamont Pass;

   4. Ensemble forecasting yielded a 2% to 5% reduction in the MAE of the wind speed
      forecasts, but there was no significant improvement in the MAE of the power production
      forecasts.

   5. The annual power production forecast MAEs varied significantly among the wind plants,
      ranging from 11.9% to 16.8%;

   6. The annual power production MAE was not well-correlated with the wind speed forecast
      MAE, with low power production MAE values often occurring with high wind speed
      MAE and vice versa; and

   7. A substantial portion of the power production forecast MAE variability is due to
      differences of the wind speed frequency distribution and the maximum slope of the wind
      plant-scale power curve. In addition, the high-resolution wind flow simulations described
      in the next section indicate that atmospheric stability in the surface boundary layer affects
      the variation of wind speeds and power generation between individual wind turbines and
      thus the plant-scale power curve.


Recommendations

Although significant progress was made in the improvement of the accuracy of power
production forecasts and the understanding of the characteristics of forecast errors in this
project, there are still many promising paths to pursue to further improve day-ahead
power production forecast performance.

Ideally, the development of power production forecasting technology should be viewed
as an ongoing cyclic process. The first step in each development cycle is a forecast
evaluation experiment for a set of wind plants similar to the one performed in the second
phase of this task. Next the data from this evaluation experiment should be thoroughly
analyzed to understand the error ranges and characteristics of the current state-of-the-art
technology. The combination of an understanding of the error characteristics of the
current state-of-the art forecast systems and awareness of new data or modeling
technology will enable modifications to the forecast system that can be tested in the
evaluation phase of a new development cycle.
                                                                                                 4-23
California Regional Wind Energy Forecasting System Development


At present, the most promising opportunity for further improvement of forecast
performance is the emergence of higher-resolution and more accurate measurements of
atmospheric and ground and water surface variables by satellite-based and ground-based
remote sensing systems. The challenge will be to effectively use the high volumes of data
produced by these systems to obtain the maximum possible improvement in the
performance of power production forecasts on the day-ahead or other look-ahead time
scales. It is recommended that the next phase of wind energy forecasting research
continue in California focus on use recently available remotely-sensed data.

Numerical and Wind Tunnel Modeling of Wind Flow and Plant-Scale Power
Curve over Complex Terrain

This section summarizes the parallel efforts by AWS Truewind and University of California at
Davis to improve the modeling of wind plant power curves by utilizing numerical and wind
tunnel models to evaluate wind flow, atmospheric stability, and other conditions over the
complex terrain at Altamont Pass. The detailed numerical and wind tunnel modeling results are
presented in [EPRI-CEC PIER, 2006b] and [EPRI-CEC PIER, 2006c], respectively.

Objectives

A significant source of uncertainty and error in wind power production forecasts is attributable to
the scatter in the relationship between the average wind speed over a prescribed time interval
(e.g., an hour) measured at one or more meteorological towers within a wind plant and the
average plant power production over that interval. The data scatter indicate that, even with a
perfect wind speed and direction forecast at each location of the meteorological tower(s), there
will still be a substantial error in the power production forecasts.

Typically, given a perfect wind speed and direction forecast, the energy forecast MAE is about
5% of a rated capacity. This represents a substantial portion of the typical 15% to 20% MAE
of state-of-the-art next-day energy forecasts. Some of the energy forecast error is related to
non-meteorological factors such as variations in turbine availability and turbine performance.
However, most of this error is attributable to the variation in wind speed and direction within the
plant’s domain, which frequently causes the wind speed experienced by each turbine to be
different from that measured at the meteorological tower.

The objective of the wind flow modeling tasks is to improve the accuracy of plant-scale wind
power production forecasts by constructing generic (i.e., not case specific) relationships between
the wind speed and direction at a plant’s meteorological tower and the wind speed and direction
at each turbine location within the farm. The results presented here are based on both wind
tunnel and very high-resolution physics-based numerical simulations of the wind flow within and
in the vicinity of the wind plant for a representative sample of cases.




4-24
                                         California Regional Wind Energy Forecasting System Development


Scope

The vision is that once relationships have been constructed, they can be used to estimate the
wind speed at each turbine location from the measured or forecasted speed at the meteorological
tower. The wind speed at each turbine location can then be used to calculate the power
production of the turbine through the use of the manufacturer’s power curve. Ultimately, the
calculated power production for each turbine can be aggregated to yield a power production
estimate for the entire wind plant.

The following sections present the research approach, results, and conclusions for the wind
tunnel and numerical modeling research and the overall conclusions and recommendations.

Wind Tunnel Modeling of Wind Flow over Complex Terrain

Approach

Researchers at the University of California at Davis (UC Davis) used an atmospheric boundary
layer wind tunnel (ABLWT) to model the air flow over the cluster of wind turbines associated
with Tower M127 at the PowerWorks wind plant in the Altamont Pass of California. Figure 4-6
illustrates the terrain model of Altamont Pass used to measure wind power density at the met
tower and wind turbine locations in the wind tunnel. The view is from the prevailing wind
direction, west-southwest (240 degrees), toward east-southeast (60 degrees), and the highest
terrain elevation is in the foreground.

The UC Davis group calculated the ratio of wind speeds at each turbine to the speed at the
meteorological tower by measuring the simulated steady-state wind speed in the wind tunnel at
the meteorological tower and each of 87 turbine locations. They calculated the wind speed ratios
for four different wind directions by rotating the terrain model within the wind tunnel. The
measurements were interpolated to create a database of relative winds for each turbine location
over all wind directions. Given a wind speed and direction at the meteorological tower, this
database can be used to extrapolate the wind speed at each turbine location. The turbine power
curve supplied by the manufacturer is then applied at each turbine location to predict the power
production of each turbine, which when summed, yields the total power prediction for the plant.

Results

In addition to the plant power curve based on a database of wind-tunnel measurements, several
other methods were developed for comparison purposes. A numerical simulation method that
required minimal use of computational resources was developed at UC Davis to meet the goal of
developing an “empirical” power curve that could be applied without requiring wind-tunnel
measurements or computationally intensive numerical models. This method, called
“Potential10,” used ensembles of two-dimensional potential flow simulations at the turbine
locations to develop a relative wind database equivalent to the one based on wind-tunnel
measurements. Both the ABLWT and Potential10 methods can be applied with or without using
the wind direction at the meteorological tower, or a correction for air density. Finally, for the

                                                                                                 4-25
California Regional Wind Energy Forecasting System Development


ABLWT method an “optimum” version was developed in which the method was “tuned” to
minimize overall error. Additionally, two power curve methods were applied that used historical
plant wind and power production data. The first used multiple linear regression (MLR) to fit a
power curve to the historical data, while the second used the medians of historical power
production binned according to wind speed.

Each of the power curve methods was applied to predict the power production of the M127
turbine cluster based on the hourly observed Tower M127 wind speeds and directions between
June 25, 2001 and June 11, 2005. Observations were excluded if the wind direction measurement
appeared inaccurate. Table 4-4 presents the resulting mean errors (ME) and mean absolute errors
(MAE) of the M127 power predictions (% of rated capacity)for the various plant power curve
methods.

Generally, the ABLWT method outperformed the Potential10 method, mainly because the
Potential10 tends to over-predict power production over a broad range of wind speeds. It should
be noted that the ABLWT method has been more extensively developed than the Potential10
method. Therefore, it may be possible to improve the Potential10 method by refining the model,
especially the treatment of hill wake effects.




4-26
                                    California Regional Wind Energy Forecasting System Development




Figure 4-6
Terrain model of Altamont Pass used in Atmospheric Boundary Layer Wind Tunnel testing
at the University of California at Davis. Wind power densities were measured at the met-
tower and wind turbine locations vs. wind direction in the wind tunnel. View is from the
prevailing wind direction, west-southwest (240°), toward the east northeast (60°). The
highest elevation is in the foreground.



                                                                                            4-27
California Regional Wind Energy Forecasting System Development

    Table 4-5
    Mean error (ME) and mean absolute error (MAE) of UC Davis plant power curve methods.
    Values are percentage of wind farm capacity.


       Prediction Method                                          ME            MAE
       ABLWT (No Wind Direction)                                 2.55           6.18
       ABLWT (w/ Wind Direction)                                 2.06           5.96
       ABLWT (w/ Wind Direction and Density)                     1.89           6.01
       ABLWT (Optimum)                                           -0.38          5.80
       Potential10 (No Wind Direction)                           4.17           7.21
       Potential10 (w/ Wind Direction)                           3.10           6.77
       Potential10 (w/ Wind Direction and Density)               2.92           6.74
       Median Historical Data Fit                                0.07           5.49
       MLR Historical Data Fit                                   -0.34          6.49



Conclusions

The UC Davis plant power curve methods were all similar in overall accuracy. Perhaps the most
significant finding is that the ABLWT and Potential10 methods are capable of predicting the
wind farm power production to a similar degree of accuracy as other methods, but do not require
any historical data from the site for power curve development. This suggests that these methods
may be especially useful for wind resource assessment and siting of turbines at proposed wind
plant locations.


Numerical Modeling of Wind Flow over Complex Terrain

Approach

In this task, the same turbine/meteorological tower wind speed ratios, calculated in the boundary
layer wind tunnel experiments, were inferred via a set of high-resolution, physics-based,
atmospheric numerical model simulations. The question is whether a numerical model can
correctly simulate wind speed differences between individual turbines, and whether the
simulated differences calculated from a few detailed simulations can improve the prediction of
the plant power output over a longer period of time.

Version 6.4 of the Mesoscale Atmospheric Simulation System (MASS) was used to perform the
simulations. MASS was originally developed in the 1980’s (Kaplan et al. 1982) as a research
model. It has since been used for a wide range of research and commercial applications, and in
recent years some of the model’s databases and physical parameterizations have been optimized
for a number of wind energy applications. MASS is a three-dimensional physics-based model
which uses a set of mathematical equations to represent the basic physical principles of
conservation of mass, momentum and energy and the equation of state for moist air.




4-28
                                           California Regional Wind Energy Forecasting System Development

Results

The MASS model was used to generate six-hour simulations for each of ten cases distributed
over three months (December 2001, May and July 2002), using a 100-meter grid centered over
the M127 cluster of turbines at Altamont Pass. Thirty-three grid cells covered the turbine cluster,
with one to five turbines falling within each grid cell. Most of the turbines are in locations with
less favorable winds than the met tower, i.e., terrain blocking and other effects will tend to
reduce the wind speed ratios (turbine/met tower) for most of the wind turbines to less than unity,
especially under more stable atmospheric conditions.

Turbine Wind Speed Ratios

Figure 4-7 shows the evolution of wind speed ratios for each of the 100-m cells over the six-hour
simulation for 16 July 2002, one of the more stable cases. For cases with less atmospheric
stability, there is much less variation between turbine locations, and the ratios are closer to one.

The ratio of the wind speed at each turbine location to the speed at M127 was calculated at each
time and then averaged to yield a set of mean turbine wind speed ratios. These ratios were quite
different from those obtained from the UC Davis boundary layer wind tunnel. The numerical
simulation-derived ratios were used to infer the wind speeds at each turbine location from the
wind speed at M127. The wind speeds were then used to calculate the power production of each
turbine using the manufacturer’s power curve, and the turbine power outputs were then added to
yield the total cluster power output.

                                       16 July 2002

          2

        1.8

        1.6

        1.4

        1.2

          1

        0.8

        0.6

        0.4

        0.2

          0
              0        1           2            3            4           5            6
                                            hour UTC



   Figure 4-7
   Evolution of individual turbine wind speed ratios over the six-hour, 100-m simulation
   beginning at 0000 UTC 16 July 2002, Altamont Pass Turbine Cluster M127



                                                                                                   4-29
California Regional Wind Energy Forecasting System Development


Using two datasets of observed wind speeds covering either ten months or three years, this
forecast method exhibited a mean absolute error (MAE) from 0.3% to 0.4% of rated capacity
higher than those obtained using an empirical (statistical) plant-scale power curve, and 0.1% to
0.3% higher than those obtained using the wind tunnel method. The mean error (ME, bias) of the
numerical simulation method however, was about 2% lower than those of the other two methods.

As shown in Figure 4-8, using ten months of forecast wind speeds, the MAEs of the three
methods were almost indistinguishable.

Atmospheric Stability

The ten simulated cases exhibited a wide range of atmospheric stabilities in the lowest few
hundred meters of the atmosphere. The high wind speed cases in July exhibited relatively high
atmospheric stability. Some of the December cases were also very stable, probably a result of the
presence of a wintertime cold stable air mass, and strong radiational cooling at the surface when
skies were clear. One of the December cases and two of the May cases were much less stable,
and under near-neutral stability, it would be expected that the numerical simulations would
behave more like the neutral stability flow simulated in the wind tunnel.

Further investigation showed that forecast performance improved, by modifying the turbine wind
speed ratios in a very simple way using a stability parameter calculated from the 100-m


                                              Plant Power Prediction by Three Methods
                                             Dec 2001-Sep 2002 Forecasted Wind Speeds
                                                                  ME    MAE

                                        20


                                                       15.3             15.4           15.3
       Plant Power Error (% Capacity)




                                        15



                                        10



                                         5


                                                0.3                             0.5
                                         0
                                             Plant Power Curve   Simulations    Wind Tunnel
                                                                 -2.5
                                        -5


    Figure 4-8 Comparison of three methods of predicting the plant power output using
    forecast M127 wind speeds from the December 2001 to September 2002 dataset




4-30
                                                                         California Regional Wind Energy Forecasting System Development



                                                  Stability-Modified Predictions from only
                                                        High Resolution Simulations
                                                                              ME   MAE

                                       20
      Plant Power Error (% Capacity)




                                       15
                                                      13.0                                                     13.1
                                                                       12.1                    11.7
                                                9.9                                                      9.9
                                       10
                                                                 7.2

                                                                                         5.1
                                        5



                                        0
                                            Plant Power Curve   Mean Ratios        Stability-modified   Wind Tunnel


                                       -5



   Figure 4-9
   Comparison of four methods of predicting the plant power output using observed M127
   wind speeds from the ten 6-hr periods simulated by the high-resolution numerical model.
   "Mean Ratios" is the simulation method using wind speed ratios between the turbine
   locations and M127 that are averages over the ten cases with a range of stabilities.
   “Stability-modified” refers to predictions in which the turbine wind speed ratios were
   adjusted for the stability of each case.


simulation data. As shown in Figure 4-9, the MAE of the stability-corrected power forecast was
1.3% of rated capacity lower than the other two methods, and the ME was 4% lower.

This encouraging result is tempered by the fact that the same stability correction didn’t help
when applied outside the time frame of the 100-m simulations. In that test, the stability at the met
tower was extracted from much coarser 8-km simulations. The lack of improvement suggested
that the 8-km runs do not resolve the meteorology of the Altamont Pass well enough to provide a
useful source of stability information.

One strategy would be to run operational simulations to as high a resolution as possible, perhaps
at 2-km or 1-km resolution. The atmospheric stability estimates provided by these higher-
resolution runs may be sufficiently dependable to apply a stability correction to the turbine wind
speed ratios in order to improve forecast performance.

Another approach would be to apply the understanding of the effect of stability on the M127
wind speed-plant power relationship to improve the empirical plant power curve. Perhaps a set of
stability classes could be defined, and an empirical plant power curve could be derived for each
stability class.


                                                                                                                                 4-31
California Regional Wind Energy Forecasting System Development


Conclusions

In summary, the results from the very high-resolution numerical simulations of the wind flow in
the Altamont Pass and their comparison to the results from the wind tunnel simulations indicate
that:

    1. The wind tunnel and numerical simulation wind speed ratios are substantially different,
       and the ratios derived from the numerical simulations are generally lower than those
       derived from the wind tunnel data;

    2. Variations of the atmospheric boundary layer stability between the cases are the most
       likely cause of the differences between the wind speed ratios derived from the numerical
       and wind tunnel modeling. Only the numerical modeling considers atmospheric stability;

    3. The met tower wind speed ratios exhibit substantial temporal variability on the time
       scales of hours as well as between days and seasons;

    4. Forecasts using the numerical simulation and wind tunnel ratios achieved about the same
       level of performance, and the performance was not significantly different from forecasts
       generating using the commonly-used empirical power curve method; and

    4. The use of stability-modified wind speed ratios appears to offer some potential for
       improvement based on theoretical arguments and preliminary tests.

Overall Conclusions and Recommendations

The similarity of the accuracy resulting from the various UC Davis plant power curve methods
suggests that the an overall accuracy limit is being approached, and that it will be difficult to
achieve substantially better accuracy based on the wind speed and direction at a single point
within complex terrain. This is because many factors such as atmospheric stability and local
forcing conditions cannot be factored into any plant power curve based only on the wind speed
and direction at a single location.

Future research should adapt the power curve methods addressed in this project to predict wind
power output based on meteorological data from multiple local sites. Ideally, a wind tunnel test
would be performed along with high resolution modeling to develop the plant-scale power curve
using wind speed and direction measurements from the nacelle anemometer and vane and wind
energy measurements for each individual wind turbine and wind speed and direction data from
multiple meteorological towers.

The results from the high-resolution simulations clearly showed that the stability of the lower
atmosphere has a significant effect on the spatial differences in wind speeds between turbines in
the Altamont Pass M127 cluster, and therefore on the relationship between the met tower wind
speed and the wind plant power output. Although the stability routinely varies from day to day in
a complex way (it’s not a simple seasonal effect), it is not used in any way to alter the plant-scale
power curve and therefore the prediction of wind plant power generation.


4-32
                                         California Regional Wind Energy Forecasting System Development

Recommendations for additional work on the impact of atmospheric stability include:

    1. Investigate and find the best way to quantify atmospheric stability based on observations
       and high-resolution numerical simulations;

    2. Whether a stability parameter can be used to improve the plant-scale power curve; and

    3. Whether numerical simulations with relatively coarse grid resolution (one to four
       kilometers) can generate the critical atmospheric stability information needed to adjust
       the plant-scale power curve.


High-Resolution Weather and Wind Flow Forecasting

This section describes high-resolution forecasting of the weather conditions in northern and
southern California used to perform the numerical modeling described in the previous section.
The detailed results are presented in [EPRI-CEC PIER, 2006b].

Introduction

Accurate wind energy forecasting begins with an accurate forecast of the weather conditions in
the region and accounts for local terrain, bodies of water, and other surface features that affect
the wind speed and direction at the location of each wind turbine. The National Atmospheric
Release Advisory Center (NARAC) at the Lawrence Livermore National Laboratory (LLNL) has
the capability to generate accurate weather and wind speed and direction forecasts at very fine
grid resolution using the latest meso-scale models. As a result, LLNL joined the project team to
generate high resolution wind speed and direction forecasts for use in the project.

The principle objectives of LLNL’s involvement were to provide real-time wind speed and
direction forecasts for use in development of improved wind energy forecasting algorithms and
to support wind tunnel and numerical modeling of wind flow over complex terrain.

COAMPS Model and Experiment Design

A modified version of the Naval Research Laboratory's (NRL’s) three-dimensional Coupled
Ocean/Atmosphere Mesoscale Prediction System (COAMPS), Version 2.0.15 was used in this
study [Chin et al. 2000, 2001 and 2005]. COAMPS consists of a data assimilation system, a
nonhydrostatic atmospheric forecast model, and a hydrostatic ocean model.

In this study, we used only the data assimilation and the atmospheric model to provide real-time
forecasts.
•   The atmospheric forecast model is composed of a compressible form of the dynamics, nest-
    grid capability, and parameterizations of subgrid-scale turbulence, surface momentum and
    heat fluxes, explicit ice microphysics, subgrid-scale cumulus clouds, shortwave and
    longwave radiation, and urban canopy physics.

                                                                                                 4-33
California Regional Wind Energy Forecasting System Development


•   The terrain-following vertical coordinate is also used to simulate airflow over an irregular
    surface. The model terrain is given from one-km resolution terrain database using Silhouette
    average method with filter.
•   The model domain contains 31 grid points in the vertical direction, with the grid size varied
    to maximize resolution at lower levels. The vertical consists of nine grid points below 127-m
    elevation, with grid spacings of 4, 4, 4, 6, 10, 16, 24, 34 and 50 meters, starting at ground
    level. The grid spacing aloft gradually increases to 800 m at 3.152-km altitude. Above this
    level, the grid size is uniformly set at 800 m up to 7.592-km altitude. Then, the grid size
    gradually increases to 5.0 km at 24.352-km altitude.
•   In the horizontal direction, a total of three nested domains are used. Both zonal and
    meridianal coordinates have 61 grid points for all nested grids. A uniform grid size of 36 km
    is used for the outer coarse grid (nest_1), the grid sizes of the inner grids are each one third of
    the size of the previous grid, e.g., the middle grid size is 12 km (nest_2), and the inner grid
    size is 4 km (nest_3) with a constant size ratio of three to define the inner nest grids.
•   The time series of the forecasts use time steps of 90 and 45 seconds for non-sound and sound
    wave calculations, respectively. The time steps for the finer-grid domains are reduced in
    proportion to the nest-grid size ratios.
•   The rigid boundary condition is imposed at the vertical boundary. A sponge-damping layer is
    placed above 10.052 km to minimize the reflection of internal gravity waves off the rigid
    upper boundary. The Davies (1976) boundary condition is applied to the lateral boundaries
    with a nudging zone of seven grid points at each lateral boundary. A time filter with a
    coefficient of 0.2 is applied to control computational instability associated with the leapfrog
    time approximation in the model.

In this study, two watches of 48-h forecasts (00Z and 12 Z, respectively) are performed daily
over California for a grid centered at the Altamont Pass and during the 12 months from July
2004 to June 2005. However, due to the size limit of huge forecast data storage, only the nested-
grid data for the first week of each month are stored and used to assess the forecast errors with
respect to the measurements at 11 available tower observations. Nonetheless, the yearly forecast
data of the finest grids were stored at UC Davis for a separate study to evaluate the wind energy
forecast errors. In this study, the forecast errors are measured by the mean absolute errors of the
wind speed and direction forecast vs. the observed data for each forecast hour, which avoids the
self-canceling effect of under-and over-prediction present in the mean forecast error.

Results

Figure 4-10 shows the locations of the 11 meteorological towers used to estimate wind speed
forecast errors and the resolution of the terrain elevation as a function of the four grid sizes tested
in the evaluation (12, 4, 1.33, and 0.444 km). The mean absolute errors of forecast wind speed
and direction were derived for each month using the forecast and observed data from 11
meteorological towers at Altamont Pass for each forecast hour throughout the 48-hour forecast
period, and averaging over the available met towers and 14 weekly forecasts (two forecasts per
day, seven days per week). Normally, the mean absolute errors were calculated using the


4-34
                                         California Regional Wind Energy Forecasting System Development




   Figure 4-10
   Resolution of terrain elevation (meters) vs. grid size (km) of the nested domain: (a) ∆x = 12
   km (nest_2), (b) ∆x = 4 km (nest_3), (c) ∆x = 1.333 km (nest_4), and (4) ∆x = 0.444 km
   (nest_5). The letters mark the locations of the met towers used in this study.


forecasts for the first week of each month. Occasionally, the forecast period was shifted several
days to accommodate the availability of station measurements.

As shown in Figure 4-11, for a warm month (June 2005), the wind speed forecast errors clearly
exhibit dependence on grid resolution using the data from all stations (Figure 4-11a), but the
wind direction error does not show the same dependence (Figure 4-11b). The observed station
wind speed of the warm month sometimes reaches 20 m/s. The absolute error of forecast wind


                                                                                                 4-35
California Regional Wind Energy Forecasting System Development




    Figure 4-11
    Weekly mean absolute errors of NARAC forecasts averaged over the selected stations for
    June 1-7, 2005. The colored lines represent the results for different horizontal resolutions
    (36, 12, and 4km, respectively). The left panels present MAEs for wind speed forecasts,
    and the right panels for wind direction. The top panels (a and b) are the forecast errors
    using the measurements from all stations, the middle ones (c and d) using five reliable
    stations, and the bottom plots (e and f) using uncertain measurements from the remaining
    six stations.

4-36
                                          California Regional Wind Energy Forecasting System Development

speed is about 6 m/s with the coarser grid resolution (36 km) and can decrease to 4 m/s in the
higher resolution forecast (4 km). With the separation of UNC and REL stations, the impact of
grid resolution on wind speed errors remains unchanged (Figures 4-11c and 4-11e). In contrast,
the wind direction error is significantly reduced with increasing grid resolution using REL
stations, while the error is fairly large at UNC stations and shows an opposite dependence on the
grid resolution (Figures 4-11d and 4-11f).
In contrast, the forecast errors for a cold month (December 2004) show much different
patterns relative to those of the warm-months. Unlike the warm months, there is no clear
dependence of forecast errors on grid resolution in the cold month, even with the separation
of REL and UNC stations. In addition, the magnitude of wind speed errors is noticeably
reduced in the cold month as a result of weaker winds (< 10 m/s, except for the storm
periods). Although the resolution impact is weak during the cold months, the separation of
UNC and REL stations still exhibits qualitative improvement in the forecast error for both
wind speed and direction.


Conclusions

The month-to-month variation of wind forecast errors clearly exhibits a semi-annual fluctuation
with prominent dependence on the grid resolution in the warm months (i.e., strong wind power
period) when the frontal activity is weak; the large forecast errors are systematically reduced
with increasing grid resolution for both wind speed and direction. However, this dependence
diminishes when synoptic-scale frontal activity prevails in the cold months.

The remaining question to be addressed from this research outcome is whether the grid
resolution dependence would continue with decreasing grid size or if this dependence tendency
converges at a certain grid size in the strong wind power period. Although the increasing grid
resolution can resolve a better representation of model terrain geometry (magnitude and shape),
further study of Silhouette terrain representation and the use of finer resolution terrain database
are highly recommended to improve the forecast accuracy for the airflow over the complex
terrain.

California Wind Generation Research Dataset (CARD)

This section summarizes the California Wind Generation Research Dataset (CARD) of wind
energy forecasts at multiple elevations over 5-km grids in northern and southern California. The
forecasts were generated using NOAA-NCEP forecast data and AWS Truewind’s MASS 6
model for the period July 1, 2004 through June 30, 2005. The CARD database is described in
[EPRI-CEC PIER, 2006d].




                                                                                                  4-37
California Regional Wind Energy Forecasting System Development


Objective

The California Energy Commission and other state wind industry participants identified the need
for a detailed four-dimensional database (i.e., space and time) of atmospheric parameters
relevant to wind power development, operations, and forecasting and used in a range of
California wind engineering, economic and policy models. The California Wind Generation
Research Dataset (CARD) was formulated in response to this need.

The original vision was that the CARD database would consist of a wide variety of measurement
data from both proprietary and public sources, including data from both in situ and remote-
sensing measurement devices. However, a number of issues arose that would make it difficult to
implement the concept.

First, the in-situ data relevant to the wind industry are mostly proprietary and access to the data
typically comes with a variety of restrictions on the permitted uses and users. It would be
difficult to include such data in a database while adhering to the restrictions.

Second, the data from different measurement systems often have substantially different accuracy,
scales of representativeness, and spatial and temporal coverage characteristics. Thus, it is often
difficult to construct a composite representation of the behavior of the wind for a specific region
(e.g., the area of a wind plant) from different sensors. The most common approach is for the user
to ignore unfamiliar data types and rely on those that have characteristics that are familiar to the
user.

Third, the spatial and temporal coverage provided by measurement data is typically characterized
by large gaps for which no data is available due to the limitations of the measuring devices.

An approach that addresses many of these issues is to create a database of numerically-simulated
data using a physics-based atmospheric model. Physics-based atmospheric models use data from
a wide variety of atmospheric sensors and create a physically consistent three-dimensional
dataset of all of the basic atmospheric variables.

Therefore, it was decided to use the numerical simulation approach to create the CARD database.
The task objectives were to design, create and document the CARD database.

The structure of the database was designed to take advantage of the physics-based numerical
simulations that were generated during the production of the one-year of day-ahead forecasts for
the five participating wind plants.

Approach

The CARD dataset was generated via numerical atmospheric simulations executed during the
production of the one-year of day-ahead wind power production forecasting for the June 2004 to
July 2005 period for five California wind plants in California. The algorithms used to generate
the forecasts required that high-resolution physics-based simulations of the atmosphere be


4-38
                                         California Regional Wind Energy Forecasting System Development

generated over regional areas surrounding each wind plant for which forecasts were to be made.
The output from these simulations was used to create the CARD database.

The atmospheric model used to generate the simulations was Version 6 of the Mesoscale
Atmospheric Simulation System (MASS) model. The MASS model was developed in the 1980s
as part of NASA’s research activities in the development of new remote sensing systems (Kaplan
et al, 1982). The model has evolved over the ensuing 20 years by incorporating new
representations of various physical processes as they became available. Version 5 of the MASS
model was used to generate the forecast simulations for the previous CEC-EPRI forecast
evaluation project (EPRI-CEC PIER, 2003a and 2003b).

The MASS 6 forecast simulations were generated on a nested grid system consisting of three
grids. An outer grid of 100 by 80 grid cells with a cell size of 20 km was used to simulate the
larger scale flow over the southwestern United States and the adjacent Pacific Ocean and
Mexico. This grid is referred to as the “A” grid. Two higher-resolution grids were nested inside
of the A grid. Both of the high-resolution grids employed an 80 by 80 matrix of grid cells and a
grid cell size of 5 km.

Figures 4-12 and 4-13 show the regions covered by the inner northern and southern California
grids, respectively. The northern grid in Figure 4-12 is centered over the Bay Area of northern
California and was used to generate forecast data for the PowerWorks, SMUD and High Winds
wind plants. The southern gird in Figure 4-13 was used to generate forecast data for the Oak
Creek and Mountain View wind plants. The physics-based simulations were initialized twice per
day.

The initialization times were 0000 UTC (4:00 PM PST) and 1200 UTC (4:00 AM PST). The
data required to specify the initial conditions and the lateral boundary conditions for the MASS
model simulations was extracted from initialization analysis and forecast data from NCEP’s
Global Forecast System (GFS).

Results

The CARD database consists of two-dimensional fields (i.e., grid point values) of selected
variables from the 5-km forecast grids over northern and southern California (Figures 4-12 and
4-13) at one-hour intervals. The data were extracted from hours 9 through 32 of the numerical
simulations initialized at 0000 UTC of each day. Since 0000 UTC corresponds to 4:00 PM PST,
hour 9 of the simulation corresponds to the hour ending at 1:00 AM on the day after the
initialization of the simulation, and hour 32 corresponds to the hour ending at the following
midnight. Thus, the database is a concatenation of the 24 hours of physics-based model output
representing the period from hour 9 (1:00 AM) to hour 32 (midnight) of each day’s numerical
simulation. The database variables are wind direction (degrees), wind speed m/s), air
density(kg/m2) , temperature (C), water vapor mixing ratio (kg/kg) at height levels of 10, 30, 50,
70, 100, 300, 600 and 1000 meters above ground level and the wind power density at 10, 30, 50
and 70 meters above ground level.



                                                                                                 4-39
California Regional Wind Energy Forecasting System Development


Figure 4-14 presents the structure of the CARD dataset. The top level of the directory structure is
named CALR1. There are two subdirectories within the CALR1 directory, one for the northern
California B grid and one for the southern California C grid. Each of the individual grid
directories (B and C) contains 365 dated directories (with a directory naming convention of
YYYYMMDDHH where YYYY is the four-digit year, MM is the two-digit month, DD is the
two-digit day and HH is the two-digit UTC time of the initialization time of the model simulation




    Figure 4-12
    The geographical domain covered by the 80 X 80 matrix of 5-km grid cells used to produce
    the high-resolution MASS 6 simulations over northern California for the project.




4-40
                                    California Regional Wind Energy Forecasting System Development




Figure 4-13
The geographical domain covered by the 80 X 80 matrix of 5-km grid cells used to produce
the high-resolution MASS 6 simulations over southern California for the project




                                                                                            4-41
California Regional Wind Energy Forecasting System Development




    Figure 4-14
    The directory structure of the CARD database. The shaded boxes denote directories.
    Names not enclosed in boxes are files. The numbers below the directory or file names
    indicate the number of the number of files or directories with the name above the number.
    See text for explanation of directory and file names.


that produced the data in the directory), an ASTAT directory and a STATIC_DATA file. Each
dated directory contains a total of 24 individual data files. Each of these files contains grid point
values for all of the database variables for one hour of the PST calendar day denoted by the
directory name.

The ASTAT directory contains files of summary statistics for each month of the year and for the
year as a whole. The summary statistics files provide monthly and annual statistical data, which
consists of the mean, max, min, and standard deviation for each of the database variables as well
as the prevailing (most frequently occurring) wind direction for each database height level.

The STATIC_DATA file contains two-dimensional arrays of the terrain elevation and latitude
and longitude for each grid point.




4-42
5
SUMMARY AND CONCLUSIONS


Summary

The worldwide installed wind generation capacity increased by 25% and reached almost
60,000 MW worldwide during 2005. As wind capacity continues to grow and large regional
concentrations of wind generation emerge, utilities and regional transmission organizations will
increasingly need accurate same-day and next-day forecasts of wind energy generation to
dispatch system generation and transmission resource and anticipate rapid changes of wind
generation.

The project objectives were to summarize the status of wind energy forecasting, address the
integration of wind energy forecasting into the utility and regional electricity system operations;
and summarize the results of the California Energy Commission-EPRI Regional Wind Energy
Forecasting System Development project completed in December 2005.

The CEC-EPRI forecasting project developed and tested wind energy forecasting algorithms for
both same-day and next-day hourly forecasts of wind speed and energy generation for the
principal wind resource areas of California and for five wind plants. The report summarizes
selected results including (1) development and testing of a next-hour regional forecasting
algorithm based on a new two-stage, artificial-neural-network (ANN) algorithm;
(2) enhancement and testing of the existing 48-hour forecast algorithm; (3) wind tunnel and
high-resolution numerical modeling of wind flow over complex terrain and power generation of
individual wind turbines; and (4) development of the California Wind Generation Research
Dataset (CARD). The following sections summarize the project results.


Short-Term, Same-Day Forecasting

The project team first developed a design concept for a two-stage forecast system for short-term,
same-day forecasting based upon artificial neural network (ANN) techniques. The first stage
consists of a mini-ensemble of three different forecast methods that exploit different input
datasets and predictive tools. The second stage is composed an artificial neural network that
weights each of the three forecasts from the first stage according to their recent performance
characteristics and creates an “optimal” composite forecast.

A limited version of the forecast system was tested using five-minute regional wind power
generation data for 2004 provided by the California Independent System Operator (CaISO). The
forecast system used only one of the three forecast methods in the first stage of the system, an


                                                                                                 5-1
Summary and Conclusions


autoregressive ANN technique, to forecast five-minute power generation over a moving
three-hour forecast period.

The limited forecast system exhibited considerable skill (up to 20%) relative to persistence
during the warm season, but virtually no skill during the cold season when the wind speeds are
low most of the time. The lack of skill during the cold season was expected, and it is anticipated
that the other two methods in the first stage of the system will provide most of the forecast skill
during the cold season. The next step is to implement and test the entire two-stage forecast
system.

Long-Term, Next-Day Forecasting

The project team evaluated improvements of the existing 48-hour forecast algorithm used in the
previous CEC-EPRI project in two phases.

Phase 1 screened several potential input data and forecast model enhancements via a series of
forecast experiments using wind resource and power production data for selected months from
the previous CEC-EPRI project [EPRI-CEC PIER 2003a and 2003b].

Several of the enhancements resulted in significant improvement in forecast performance relative
to the previous project, including the use of high-resolution surface water temperature data from
satellite-based sensors and a more sophisticated statistical model to adjust the raw forecasts of
the physics-based model.

Phase 2 incorporated several of the enhancements into a modified eWind forecast system and
generated daily 48-hour forecasts of hourly wind speed and energy generation at each of five
participating California wind plants for a one-year period. Two of the five participating plants
also participated in the previous CEC-EPRI forecasting project, the 66.6 MW Mountain View 1
and 2 project in San Gorgonio Pass and the 90 MW Wind Energy Partners/WindWorks project at
Altamont Pass. The three other wind plants were located in the Solano (2) and Tehachapi wind
resource areas.

The annual mean absolute error (MAE) of the 48-hour power generation forecasts was 14.5% of
rated capacity and 52.7% of the actual wind generation for all five wind plants. The annual wind
speed forecast MAE for the same plants and time periods was 2.27 m/s, which was 34.1% of the
average speed.

In addition, application of an ensemble of forecast methods and a simple arithmetic average of
the forecasts reduced the wind-speed forecast MAE by 2% to 5%. However, ensemble
forecasting did not significantly reduce the power production forecast MAE, because the power
production forecast error is sensitive to the wind speed forecast error due to the shape and slope
of the plant-scale power curve.

Another likely reason is that none of the forecast methods in the ensemble address the impact of
atmospheric stability in the surface boundary layer on wind flow and the variation of power
generation between individual wind turbines, especially in complex terrain. The high-resolution

5-2
                                                                             Summary and Conclusions

wind flow simulations showed that the variability of individual turbine power generation and
therefore the plant-scale power curve are affected by whether the surface boundary layer
unstable, neutral, or stable conditions, as indicated by the vertical temperature profile.

The enhancements reduced the annual mean absolute errors of the wind power generation
forecasts and provided significant insights into the characteristics and sources of the wind speed
and power generation forecast errors. The results will help plan the direction of future research to
improve day-ahead forecast performance.


Wind Tunnel and Numerical Modeling

The project team applied two different approaches to very high-resolution modeling of wind
flow and power generation in complex terrain. First, the atmospheric boundary layer wind tunnel
at the University of California at Davis was used to simulate the variation of wind speeds at
different turbine locations within the Altamont Pass M127 cluster. Second, high-resolution,
physics-based, atmospheric numerical model simulations were used to simulate the same wind
speed variations.

Both approaches were used to infer the wind speeds at individual turbine locations from the wind
speed at the M127 tower, apply the turbine manufacturer’s power curve to estimate the power
production of each wind turbine, and aggregate the power production of the turbines to calculate
the total hourly power output. The wind tunnel and numerical simulation methods each generated
wind plant power forecasts that were very close to, but not quite as accurate as those generated
by the widely-used empirical plant-scale power curve approach. This activity was assisted by
high-resolution numerical weather forecasts provided by the National Atmospheric Release
Advisory Center (NARAC) at the Lawrence Livermore National Laboratory (LLNL).

The numerical simulation results also indicate that atmospheric stability of the lower atmosphere
significantly affects the relationship between the wind speeds at individual turbine locations. By
modeling atmospheric stability during the high-resolution simulations, the accuracy of the wind
plant power forecasts improved.

The importance of atmospheric stability suggests that additional research to incorporate
atmospheric stability into the forecast models could further improve the accuracy of the next-day
wind power production forecasts.


California Wind Generation Research Dataset (CARD)

The project team developed the California Wind Generation Research Database (CARD). The
CARD dataset contains one continuous year of hourly data for a set of meteorological variables.
The data are provided at multiple levels on two horizontal grids, each with a grid cell size of
5 km, one located in northern California and one in southern California. The dataset was
generated for the period July 1, 2004 through June 30, 2005, using the MASS mesoscale model,
and it is not based directly on measured data.


                                                                                                5-3
Summary and Conclusions


The database variables are wind direction (degrees), wind speed (m/s), air density (kg/m2),
temperature (C), water vapor mixing ratio (kg/kg) at height levels of 10, 30, 50, 70, 100, 300,
600 and 1000 meters above ground level and the wind power density at 10, 30, 50 and 70 meters
above ground level.

The CARD database can be used to simulate the hourly operation of a wind project at candidate
sites, test wind energy forecasting methods, and conduct other wind power research and planning
activities.


Conclusions

Three factors shape the need for wind energy forecasting in California and other regions where
wind generation is growing rapidly:
      •   The expected large additions of wind generation in California;
      •   The impacts of large regional concentrations of intermittent wind generation on wind
          generation ramp rates and electricity system operations and costs; and
      •   The resulting need for accurate same-day and next-day forecasts to forecast rapid ramp
          rates of wind generation and to support grid operations and wind energy markets.

The California Regional Wind Energy Forecasting Project achieved all of its objectives and
made significant advances in the following areas:

New Two-Stage Artificial Neural Network Forecast Algorithm: The algorithm developed in the
project has excellent potential for short-term wind power generation forecasting over the zero to
six hour forecast time window. Since only a portion of the first stage was tested and it
demonstrated improved forecast performance vs. persistence only during the warm season,
additional testing of the full algorithm is needed for both the warm and cold seasons.

Enhancement of 48-Hour Forecast Algorithm: The most promising enhancements include use of
higher -resolution water temperature data available from satellite sensors; advanced statistical
techniques to develop MOS (model operating statistics) to adjust raw forecasts, and ensemble
forecasting using different combinations of meso-scale models, plant-scale power curve models,
and statistical methods and averaging the resulting forecasts.

Ensemble Forecasting: Ensemble forecasting improved 48-hour wind speed but not power
generation forecast performance. Therefore, additional development and testing are needed, most
likely to address the high sensitivity of power generation errors to wind speed forecast errors due
to the shape of the plant-scale power curve and to account for atmospheric stability and other
factors that affect the plant-scale power curve used in the forecast algorithm.

Wind Tunnel and High-Resolution Numerical Modeling of Wind Flow over Complex Terrain:
The forecast performance was similar for wind power production forecasts based on data from
wind tunnel and high-resolution numerical modeling of wind flow over a portion of Altamont
Pass. However, the variation of wind turbine/reference meteorological tower wind-speed ratios

5-4
                                                                            Summary and Conclusions

between individual wind turbines indicated by the wind tunnel and numerical model simulations
were different, possibly due to the impact of atmospheric stability as discussed below.

Impact of Atmospheric Stability: The most significant finding of the numerical modeling work is
that atmospheric stability in the surface boundary layer affects the variations of wind
turbine/reference meteorological tower wind-speed ratios between individual wind turbines.
Initial testing of a simple algorithm that adjusts wind speed ratios for atmospheric stability
yielded improved power generation forecast performance. Thus further development and testing
is needed of both a plant-scale power curve algorithm that accounts for atmospheric stability and
a rapid-update high-resolution meso-scale model to support the algorithm.

California Wind Generation Research Dataset (CARD): The CARD database contains
8,760 hours of hourly wind speed, direction, wind power density, and other data at multiple
levels over two 5-km grids in northern and southern California. It was generated via a numerical
simulation of hourly weather conditions during the period July 1, 2004 through June 30 2005 and
is not representative of long-term average data. CARD can be used to simulate the hourly
operation of a wind project at candidate sites, test wind energy forecasting methods, and conduct
other wind power research and planning activities.


Recommendations

In order to provide accurate wind energy forecasting technology when it is needed within 5 to
10 years , it is strongly recommended that: (1) the California wind energy forecasting research
program should continue with active participation by the California utilities and the California
Independent System Operator, California Energy Commission, and Electric Power Research
Institute; and (2) the program should focus on two specific activities:
   •   Implement the same-day and next-day forecast algorithms developed in the CEC-EPRI
       project via development of a real-time wind energy forecast workstation in collaboration
       with participating utilities, ISOs, and RTOs; and
   •   Continue to improve the forecast performance of both same-day and next-day forecast
       algorithms, especially with regard to accurately forecasting high hourly ramp rates of
       wind generation.

The recommended field implementation and research activities can be conducted in parallel.

Field Implementation of Wind Forecast Workstation
   1. Incorporate forecast algorithms from the California Regional Wind Energy Forecast
      System Development Project into a simplified forecast workstation to display real-time
      same-day and next-day forecasts, forecast performance statistics, and forecast confidence
      indicator.
   2. Meet with utility, ISO, and RTO system operators to demonstrate workstation and invite
      input on forecast needs for responding to high wind ramp rates and other impacts.



                                                                                                5-5
Summary and Conclusions


      3. Develop specifications for prototype wind forecast workstation in collaboration with
         system operators.
      4. Develop, build, and test prototype forecast workstation; implement same-day and next-
         day forecast algorithm enhancements in workstation as they become available.
      5. Implement forecast improvements from research as they become available.

Next-Hour and Same-Day Forecasting Research
      1. Complete development and testing of the two-stage ANN regional forecast algorithm.
      2. Focus development on improving forecast accuracy and providing accurate forecasts of
         high ramp rates over the three to six-hour time period.
      3. Develop methodology for optimal placement of meteorological towers within and around
         each wind resource area.
      4. Facilitate development of, install, and maintain supplemental network of real-time
         meteorological towers at optimal locations within each wind resource area.
      5. Develop and test rapid-update meteorological forecast model to support same day
         forecasting.

Next-Day and Longer Forecasting Research
      1. Continue research on use of remote-sensing data, enhancements of the meso-scale
         weather models, plant-scale power curves, statistical MOS, and ensemble forecast
         techniques to improve forecast performance.
      2. Investigate the impacts of atmospheric stability in the surface boundary layer on the
         variability of turbine/met tower wind speed ratios.
      3. Develop and test methodology to apply adjustment for atmospheric stability to
         turbine/met tower wind speed ratios and translate ratios into the plant-scale power curve.
      4. Develop and test methodology to account for moisture evaporation from irrigated fields
         and other effects.




5-6
6
REFERENCES


AWS Truewind, 2005, Zack, John, Monthly Progress Report to EPRI, June 2005.

California ISO, 2005, Hawkins, David Wind Generation Forecasting: A Balancing Authority
View, presentation to Utility Wind Integration Group Technical Workshop, Sacramento, CA,
November 2005.

3TIER, 2003, Westrick, Kenneth, Kristin Larson, Bob Baker, and Tilmann Gneiting,
“Description and Results from a Comprehensive Wind Energy Forecast System in the Pacific
Northwestern U.S.”, Wind Power 2003 Conference, American Wind Energy Association,
Austin, TX, May 2003.

EPRI, 2003a, Wind Energy Forecasting Applications in Texas and California, EPRI Palo Alto,
CA: 2003. 1004038.

EPRI, 2003b, Short-Term Wind Generation Forecasting Using Artificial Neural Networks, EPRI
Palo Alto, CA: October 2003. 1009219.

EPRI 2003c, Characterizing the Impacts of Significant Wind Generation Facilities on Bulk
Power System Operations Planning, Utility Wind Interest Group - Xcel Energy-North Case
Study, EPRI, Palo Alto, CA, Utility Wind Interest Group, Springfield, VA, Xcel Energy,
Minneapolis, MN, NRECA Cooperative Research Network, Washington, D.C., American Public
Power Association, DEED, Washington, DC, Western Area Power Administration, Denver, CO:
2003. 1004807.

EPRI, 2004a, Wind Power Integration Technology Assessment and Case Studies, EPRI, Palo
Alto, CA: 2004. 1004806.

EPRI, 2004b, Texas Wind Energy Forecasting System Development and Testing Phase 2:
12-Month Testing, EPRI Palo Alto, CA: 2004. 1008033.

EPRI, 2004c, Debs, A., C. Hansen Y. Makarov, D. Hawkins and Peter Hirsch, “Wind Power
Forecasting in California Based on the EPRI ANNSTLF,” Balkan Power Conference, University
of Tuzla, Sarajevo, Bosnia & Herzegovina, May 26-28, 2004.

EPRI, 2005a, Wind Energy Forecasting Technology Update: 2004, EPRI, Palo Alto, CA: 2005.
1008389.



                                                                                           6-1
References


EPRI, 2005b, Wind Power Integration: Smoothing Short-Term Power Fluctuations, Electric
Power Research Institute, Palo Alto, CA: 2004. 1004806.

EPRI, 2005c, Wind Power Integration: Energy Storage for Firming and Shaping, Electric Power
Research Institute, Palo Alto, CA: 2004. 1004806.

EPRI-CEC PIER, 2003a, California Wind Energy Forecasting System Development and Testing
Phase 1: Initial Testing, EPRI Palo Alto, CA: 2003. 1003778.

EPRI-CEC PIER, 2003b, California Wind Energy Forecasting System Development and Testing
Phase 2: 12-Month Testing, EPRI Palo Alto, CA: 2003. 1003779.

EPRI-CEC PIER, 2006a, California Regional Wind Energy Forecasting System Development,
Volume 1: Executive Summary, EPRI Palo Alto, CA: 2006. 1013262.

EPRI-CEC PIER, 2006b, California Regional Wind Energy Forecasting System Development,
Volume 2: Executive Summary, EPRI Palo Alto, CA: 2006. 1013263.

EPRI-CEC PIER, 2006c, California Regional Wind Energy Forecasting System Development,
Volume 3: Executive Summary, EPRI Palo Alto, CA: 2006. 1013264.

EPRI-CEC PIER, 2006d, California Regional Wind Energy Forecasting System Development,
Volume 4: Executive Summary, EPRI Palo Alto, CA: 2006. 1013265.

Khotanzad, 1998, A. Khotanzad, R. Afkhami-Rohani, D. Maratukulam. ANNSTLF – Artificial
Neural Network Short-Term Load Forecaster – Generation Three. IEEE Transactions on Power
Systems, Vol. 13, No. 4, 1998, 1413-1422.

Xcel, 2004, Xcel Energy and the Minnesota Department of Commerce Wind Integration Study,
Final Report,




6-2
Export Control Restrictions                                                The Electric Power Research Institute (EPRI)
Access to and use of EPRI Intellectual Property is                         The Electric Power Research Institute (EPRI), with
granted with the specific understanding and                                major locations in Palo Alto, California, and Charlotte,
requirement that responsibility for ensuring full                          North Carolina, was established in 1973 as an
compliance with all applicable U.S. and foreign export                     independent, nonprofit center for public interest energy
laws and regulations is being undertaken by you and                        and environmental research. EPRI brings together
your company. This includes an obligation to ensure                        members, participants, the Institute’s scientists and
that any individual receiving access hereunder who is                      engineers, and other leading experts to work
not a U.S. citizen or permanent U.S. resident is                           collaboratively on solutions to the challenges of electric
permitted access under applicable U.S. and foreign                         power. These solutions span nearly every area of
export laws and regulations. In the event you are                          electricity generation, delivery, and use, including
uncertain whether you or your company may lawfully                         health, safety, and environment. EPRI’s members
obtain access to this EPRI Intellectual Property, you                      represent over 90% of the electricity generated in the
acknowledge that it is your obligation to consult with                     United States. International participation represents
your company’s legal counsel to determine whether                          nearly 15% of EPRI’s total research, development, and
this access is lawful. Although EPRI may make                              demonstration program.
available on a case-by-case basis an informal
assessment of the applicable U.S. export classification                    Together…Shaping the Future of Electricity
for specific EPRI Intellectual Property, you and your
company acknowledge that this assessment is solely
for informational purposes and not for reliance
purposes. You and your company acknowledge that it
is still the obligation of you and your company to make
your own assessment of the applicable U.S. export
classification and ensure compliance accordingly. You
and your company understand and acknowledge your
obligations to make a prompt report to EPRI and the
appropriate authorities regarding any access to or use
of EPRI Intellectual Property hereunder that may be in
violation of applicable U.S. or foreign export laws or
regulations.




© 2006 Electric Power Research Institute (EPRI), Inc. All rights
reserved. Electric Power Research Institute and EPRI are registered
service marks of the Electric Power Research Institute, Inc.

    Printed on recycled paper in the United States of America                                                                     1011884




                                                   ELECTRIC POWER RESEARCH INSTITUTE
                  3420 Hillview Avenue, Palo Alto, California 94304-1395 • PO Box 10412, Palo Alto, California 94303-0813 • USA
                                        800.313.3774 • 650.855.2121 • askepri@epri.com • www.epri.com

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:41
posted:5/18/2011
language:English
pages:100