A Prototype Day-Ahead Forecast System for Rapid Wind Ramp Events

Document Sample
A Prototype Day-Ahead Forecast System for Rapid Wind Ramp Events Powered By Docstoc
					A Prototype Day-Ahead Forecast
System for Rapid Wind Ramp Events



 Presented by Eric Grimit, 3TIER
 October 2, 2007
 for CanWEA 2007

 Thanks to:
 Bonneville Power Administration
OVERVIEW


• Rapid Ramp Events (RREs)
  – Impact
  – Characteristics (Frequency, Causes, Definitions)
  – Prediction (Limitations of state-of-the-art)

• A New Tool for Day-Ahead RRE Prediction
  – Probability-based (good for managing risk)

• Challenges and Solutions

• Summary
THE IMPACT OF RAPID RAMP EVENTS

• To operate a power system, it is important to
  match the generation and the load.
• To do this, you must forecast the load, but also
  any non-dispatchable generation (e.g. wind).
• Regulation reserves are designed to
  accommodate deviations from the schedule.
• However, rapid wind ramps result in significant
  changes from the schedule over a short period.
• An RRE forecast would help minimize the need
  for excessive regulation reserves.
Does Spatial Diversity
 Eliminate Ramps?
BPA RAMP STUDY




        3770 MW: 18 projects
10-MINUTE RAMPS

    Percent of 10-minute ramps larger than 10% in magnitude




                                         All 3770 MW = 0.03%
HOURLY RAMPS

    Percent of hourly ramps larger than 10% in magnitude




                                         All 3770 MW = 6%
                                   Ten Ramps per week > 375 MW
RAPID RAMP EVENT DEFINITION


• A ramp in wind energy production has three attributes:
  time, magnitude, and direction
   – We define the duration of a ramp to be 1 hour, since that is the
     natural temporal resolution of day-ahead wind energy forecasts

   – Therefore, the magnitude of a ramp is the absolute change in
     hourly-average wind energy

   – The direction of the change can be positive or negative

     We defined a “rapid ramp event” as a ramp rate exceeding 10%
     of nameplate per hour, which is consistent with the report1

   1. “Wind Power Modeling and Analysis of Output for Existing and Planned Wind
      Farms in the Klondike/Dalles Region [BPA, 2006]”
Forecasting Rapid
  Ramp Events
STATE-OF-THE-ART FORECAST
WHY A NEW TOOL IS NEEDED


• Normal goal for forecasting is to minimize large
  deviations from the observations
  – “Best” forecast is often chosen so that the root mean
    square error (RMSE) is minimized
  – Tends to be conservative – hedges toward the mean
  – Yields smooth forecasts and therefore results in
    serious under-prediction of ramps

• Need a forecast system that is tuned to optimize
  prediction of ramp events
EVENT VERIFICATION TECHNIQUES


• With categorical (yes/no) forecasts of a
  binary event, there are four possible
  outcomes.
                           Ramp Observed
                         Yes           No
   Ramp Forecast




                                   Correct
                   No




                         Miss
                                   Negative
                   Yes




                         Hit         False
RAMP VERIFICATION EXAMPLE

         State-of-the-Art DA Forecast for Nine Canyon
                 Hits        Misses   Falses   Correct Negatives




    10     15           20       25    30      35     40       45   50

                Size of Ramp [% of Maximum Capacity]
CHALLENGES FOR PREDICTION

• Timing
  – Example: Day-ahead forecasts of winter-time cold
    fronts on the west coast can be off by 6 hours or more
• Intensity/Strength
  – Example: Small errors in offshore-onshore pressure
    gradients can result in poor land/sea breeze intensity
• Location
  – Example: Thunderstorm locations are difficult to
    forecast due to poor simulation of their initiation,
    development and decay processes
  – The locations of larger-scale weather features (e.g.,
    fronts) are generally not as problematic
BENEFITS OF HIGHER RESOLUTION


                      High resolution
                      provides better
                      modeling of the
                      intensity and
                      location of wind
                      ramp events

                         A = WRF, G = MM5

                     2 = 15km, 3 = 5km, 4 = 1.67km
TIMING ERRORS
 ALLOWING FOR TIMING ERROR
       State-of-the-Art DA Forecast for Nine Canyon
               Hits        Misses   Falses   Correct Negatives




                                                                               Timing “Window” = 3 hrs
                                                                            State-of-the-Art DA Forecast for Nine Canyon
                                                                                    Hits        Misses   Falses   Correct Negatives




  10     15           20       25    30      35     40       45   50

              Size of Ramp [% of Maximum Capacity]



Timing “Window” = 0 hrs


                                                                       10     15           20       25    30      35     40       45   50

                                                                                   Size of Ramp [% of Maximum Capacity]
RRE PREDICTION METHODOLOGY
PROTOTYPE 8-MEMBER ENSEMBLE


Description   Boundary    Land      Radiation       Cloud
               Layer     Surface                 Microphysics
WRF – A         YSU       5LTD     Dudhia/RRTM    FERRIER
WRF – B         MYJ       5LTD     Dudhia/RRT     FERRIER
WRF – C         YSU      NLSM      Dudhia/RRTM      WSM5
WRF – D         MYJ      NLSM      Dudhia/RRTM      WSM5
WRF – E         YSU       5LTD        CAM3          WSM5
WRF – F         MYJ       5LTD        CAM3          WSM5
WRF – G         YSU      NLSM         CAM3        FERRIER
WRF – H         MYJ      NLSM         CAM3        FERRIER
STATISTICAL CALIBRATION

• Forecast Ensembles Are Not Perfect
  – They are a limited sample of the weather uncertainty
  – Members are not always equally likely, especially if
    multiple models are used


• Bayesian Model Averaging (BMA)
  – BMA accounts for imperfect ensembles by weighting
    each forecast member by its relative skill over a
    relevant training set
  – BMA calibrates the probabilistic forecast and provides
    prediction intervals that are both dependable and
    sharp
TESTING/TRAINING DATA TIMELINE


                    2006                                2007
May   Jun   Jul   Aug   Sep   Oct   Nov   Dec   Jan   Feb   Mar   Apr


                        Nov   Dec   Jan   Feb

                        Training    Testing

 • The prototype was developed on a limited
   training set of four months:
      – two months for training and
      – two months for testing
RAMP FORECAST TIME SERIES
PRELIMINARY RESULTS




                      Generally over forecasting,
                      especially for the larger ramps
PRELIMINARY RESULTS


                         The largest contributor to
                         forecast errors is timing




 Allowing for a timing
 error of 3 hours
CHALLENGES AND SOLUTIONS


• Tendency for over-forecasting of RRE
  frequencies
  – Take better account of timing errors by:
     • using a larger forecast ensemble
     • building it into the BMA post-processing
  – Use a larger training sample


• Low bias in probabilistic power forecasts
  – Remove with larger training sample
CONCLUSIONS


• 3TIER has developed a prototype day-ahead,
  rapid ramp event (RRE) forecast system
  – The system is fully probabilistic; implicitly taking account of both
    wind forecast and power conversion uncertainties
  – 3TIER has identified challenges from the prototype results and
    has proposed solutions
  – With these improvements, 3TIER anticipates delivering ramp
    forecasts that improve decision-making capability for day-ahead
    scheduling

• Stay Tuned
  – Further testing and verification results at CanWEA 2008
 www.3tiergroup.com


egrimit@3tiergroup.com
 (206) 325-1573 x 125
cpotter@3tiergroup.com
 (206) 325-1573 x 123