Short Term Load Forecasting with Artificial Neural Network via case study
1R.M.Holmukhe2Mrs.Sunita Dhurnale, 3Mr.P.S. Chaudhari, 4Mr.P.P.Kulkarni
Bharati Vidyapeeth Deemed University,
College of Engineering,
Abstract-In the decade 2000-2010, Indian • Enforcement of rationalized tariff and
Electricity sector has acknowledged the ABT.
major changes in statutory framework, • Major grants (more than 500 Billions
operational framework, and policies of INR) under 11th plan under R-APDRP.
Governments. Main reasons of such
• Introduction of competition and Open
developments and alterations could be
• Formation of Indian Energy
• Worldwide need of fuel or natural Exchange.
energy resources saving.
• Meeting regular energy requirement I. INTRODUCTION
of the India. The paper emphasizes the requirement,
• Eliminating monopoly of the SEBs. types and development of load forecasting
• Rationalization of Tariff, introduction applications tools to meet routine regulatory
of power market I open access to render and operational requirements I compliances
benefits to consumers. by Transmission Supply Users (TSUs) of
• Increasing level of Standard of the changed regime of Indian Power Sector
Performance, qualityof supply and such as Generation Companies,
services. Transmission Companies, Independent
• Reduction in Transmission and Power Plants, Distribution Companies,
Distribution losses. Trading Companies, Open Access Users
• Contribution to reducing global and consultants.
warming and support to global CDM.
II. OBJECTIVE SCOPE
Major developments of this decade are ~ To highlight on the load forecasting
listed below. needs and importance to operational
• Enactment of Electricity Act 2003, mechanisms under changed regime of
Finalization of National Electricity Policy power sector in India.
and National ~ To reach detailed enlightenment on
Tariff Policy. available models I tools Itechniques
Imethodologies of load forecasting.
• Restructuring of SEBs and
unbundling of DISCOMs. ~ Understanding practical adoption of
load forecasting tool through case
• Formation of SERCs, CERC, ATE and
study presentation of one of the Utility
making them operational.
Paper is organized in main sections. Jon these schedules Generators prepare
••• Need of load forecasting. generation schedules. SLDC controls and
••• Load forecasting types, load monitors these operations.ABT mechanism
forecasting tools and methodologies enforces incentive and disincentives to
in details. Generators and TSUs as per the
••• Case study and concluding compliances and non-compliances to the
annotations. submitted schedules. UI calculations and
balance and settlement is done on the ABT
III. NEED OF LOAD FORECASTING compliant meter data and tariff determined by
AND LOAD FORECASTING TYPES ERCs.
a) Need of Load Forecasting: b) Types of Load Forecasting:
1. Planning: Load forecasting types are worked out as per
Every participant viz GENCO, TRANSCO, the operational requirements.
DISCO and Traders need load forecast 1. Long term load forecasting (LTLF):
inputs to prepare new schemes of extension Applicable for system and long term network
or enhancements or capacity additions or planning.
infrastructure development. The network and 2. Mid term Load Forecasting (MTLF):
system planning is always based on load Applicable for quarterly, half yearly and
requirements. Advanced load forecasting yearly LF
tools or applications gives appropriate future needs.
long term load requirements. 3. Short term Load Forecasting (STLF):
A major component of the ARR is investment Applicable for day ahead and week ahead LF
CAPEX in Infrastructure Projects for network needs.
enhancement and restructuring for loss
reduction. Load forecasting results give IV. LOAD FORECASTING TOOLS &
accurate inputs for optimal CAPEX proposals METHODOLOGY:
and competitive tariff to consumers.
2. Estimating sales forecast for ARR. 1. Artificial Neural network.(ANN)
Under new statutory regime, DISCOMS have 2. Fuzzy logic (FL).
to furnish ARR to ERCs for Tariff 3. Autoregressive model.
Determination. Tariff is determined on the 4. Similar day approach.
basis of sales forecast. Accurate realization 5. Time series.
1"7'"0 or recovery as per recovery is always hinged 6. Expert system.
to accurate sales forecasting sale forecasting 7. Support vector machine.
is estimated on the basis of load forecasting
ANN and FL are the popular and commonly
used mathematical tools for LF applications.
3. Furnishing Day ahead Schedule to
LDCs by DISCOMs. Detailed explanations and calculations are
placed below for ANN LF models.
GENCOs get fixed charges for its asset
implementation and maintenance in the ARR. 4.1 Artificial Neural Network.(ANN):
However, they get recurring and FCAs Introduction:
through monthly energy charges from Short term load forecasting is an essential
DISCOMs. tool in operation and planning of the power
Apart from above DISCOMs and other TSUs system. It helps in coordinating the
are governed by ABT mechanism. Under generation and area interchange to meet the
ABT, DISCOMs and receiving TSUs have to load demand. It also helps in security
submit day ahead schedule to LDCs. Based \ assessment, dynamic state estimation, load
management and other related functions. In '.----------------:~.
the last few decades, various methods for SINGLE PROCESSING UNIT (PE)
short term load forecasting have been
proposed. The methods vary from simple NEURON
regression and extrapolation of fading
memory Kalman filter and knowledge based
Among the various methods available in .
the literature, most can be classified into two
categories. In the first category are the
methods, which rely solely on the past data
and fit the load pattern as a time series. In
the second category are the methods, which 4.1.1 ARTIFICIAL NEURAL NETWORK
give emphasis to the weather variables, i.e., (ANN):
temperature, humidity, light intensity, etc, and Artificial Neural Networks are
find a functional relationship between these increasingly finding use as alternative
variables and the load demand. computational paradigm for solving complex
problems like pattern recognition etc.
Recently, Artificial Neural Networks Neurons in ANN can be viewed as simple
(ANN) has been used for short term load processing elements (PE). A commonly used
forecasting. Both time series models and PE representation of an artificial neuron is
weather dependent models have been used shown in Fig 1. The PEs can be
in ANN based short tem load forecasting. In interconnected in various topologies.
this paper, a short-term load forecasting Depending on the various topologies,
method using the ANN is proposed. A activation functions and weight change
multilayered feed forward (MLFF) neural strategies, a large number of ANN
network with back propagation learning architectures have been developed, e.g.,
algorithm has been used because of its Back propagation, Hotfield net, Kohonen net,
simplicity and good generalization property. etc. Among the various ANN architectures
The input, to the neural network is based only available in the literature, the multilayer feed
on past load data and are heuristically forward (MLFF) network with error back
chosen in such a manner that they inherently propagation learning algorithm has been
reflect all the major components, such as, selected for this problem mainly because (i) it
trend, type of day, load shape as well as is the most simple and comprehensive neural
weather which influence the system load. approval for model based prediction and/or
The main contributions of this paper are: control and (ii) it has the generalisation
(i) Heuristic choice of a small set of input which capability.
inherently represents the major components of
the load pattern (ii) introduction of a stopping 4.1.2 Multilayered Feed Forward Network
criteria during learning phase to avoid over (MLFF):
fitting of the network to learning examples, and In MLFF network the PEs are arranged in
(iii) A detailed analysis of the generalisation layers and only in adjacent layers are
properties like interpolation/extrapolation ability connected. It has a minimum of layers of
of the ANN, working life of a trained network, PEs; (i) the input layer, (ii) the middle or
ie, useful period of a network after which a hidden layer(s), (iii) the output layer. The
retaining is required etc information propagation is only in the forward
direction and there are no feedback loops. A
MLFF network topology is shown in Fig 2. \
In order to obtain bounded output from PEs a 8j = OJ(~- OJ)L8kWkj
sigmoidal activation function is chosen where
output is limited to (0, 1) for the input range (- where k is over all PE's in the layer above the
oc,oc). jth layer of PE and 11,the learning rate, oc, the
The MLFF network uses separate momentum factor. The momentum term
stages for learning and operation. The helps in faster convergence of the algorithm.
learning problem can be stated as: given a Once the network gets trained, the resulting
set of input-output pairs (11 • 01), ..... (In . On), connection weights are frozen. In the
find the interconnection weights Wjj for each operation stage the network is used to
interconnection of ANN such that the network compute an output from a set of inputs.
maps Ij to OJfor i 1, 2, 3, = , n, as closely
4.1.3 PROPOSED METHOD
Characteristics of the Load Data
The error back propagation learning In order to reflect the load behavior in
algorithm, the interconnection weights are the input information, the historical hourly
adjusted such that the error function E = - load data for 1 year of a number of systems
(1/2) L k (tk - Ok) is minimized, where, were analyzed. It was observed that the load
tk desired output for unit in layer k data exhibits a daily and weekly periodicity. It
Ok actual output for unit in layer k was also observed that the daily load pattern
for the working days showed marked
similarity whereas the holiday load patterns
SHEMATIC ILLUSTRATION OF MULTILAYER
FEED l'"ORW ARD (MLFF) NETWORK
were quite different from those of the working
days. Therefore, hourly loads for working
days and holidays were treated separately.
Auto-correlation of hurly load was obtained
L n-k(Yt- y-)(Yt+k y-)
rk = ---------------------------
L n (Yt_ y-)2
t 1 =
rk = auto-correlation factor for time lag k
INPUT PATTERN I
I .,.(1) ··1-
The minimisation process is based on gradient .
descent algorithm. The interconnecting weights
between jth layer (upper layer) neurons and ith
-I.O() --J:-.-.-".......~~~..........,.... -:r:rr""........,..,.....,..!
layer (lower layer) neurons is modified using H(),-,t,. •.•.._
the following relationship.
Figure 3Auto-Correlation Factor (rk) for
Wjj (new) = Wjj (old) + 118jOJ+ o;
two weeks load on best system
Where, if, PEj is an output layer PE, then
8j OJ(tj - OJ)(1 - OJ)
if, PEj is an hidden layer PE, then
n total number of available data ie, input layer, one hidden layer and output
y- = mean value of that available data layer. The number of neurons in the input
Yh hth hour data layer is equal to the number of variables in
for two weeks (336 hours) data for the test the input data. The output layer consists of
systems and is shown in Fig 3. Loads for 24 one neuron. Although, the choice of number
hours and 168 hours are highly correlated of hidden layer neurons is arbitrary and a
and based on these observations. Five hourly optimal number of hidden layer neurons is
loads were heuristically chosen and used as generally obtained through trail and error. On
input information. These inputs are as the basis of a large number of simulations a
follows: (i) previous hour load (L1), (ii) large number of neurons in the hidden layer
previous to previous hour 10ad(L leads to large training time, as well as, it
2),(iii)previous day(same day type) same hour creates a grandmother network. The new
load(L24), network memorizes the learning patterns very
well but does not perform well for new set of
(iv) previous week same day and same hour input. Whereas, with too small number of
load(L168), (v) previous week same day but hidden layer neurons, the network has
previous hour load(L169). difficulty in learning, as it is unable to create
the required complex decision boundaries.
Therefore, a good starting point for optimal
Among these, L24 and L168 reflect the daily choice of hidden layer neuron by trail and
and weekly periodicity of the hourly load. L1- error is to use geometric mean of the input
, L2, L168 and L169 reflect the trend of the and output layer neurons.
hourly load pattern and L1 and L2 also Stopping Criteria:
implicitly reflect the weather effect.
Fig. 4 shows the convergence
Scaling of the Input and Output Data
characterizes of the learning algorithm for
\ The input and output variables for the
IEEE 24 bus system. The testing was done
neural network will have very different ranges
after every iteration during learning. Initially,
if the actual hourly load data is directly used.
the Mean Square error (MSE) for both the
This may cause convergence problem during
training and testing set decreases gradually.
the learning process. To avoid this, the input
But after some iterations, is, around 2000
and output load data were scaled such that
iterations, the MSE for the testing examples
they were within the range (0,1), with majority
increases, through, the MSE for the learning
of the data having values near to 0.5. For this
examples still decreases, is, network starts
purpose the actual load was scaled using the
over fitting for the training set from this point.
Thus, the learning should be stopped at this
L - Lmin point.
Ls = ------------------------ Simulation and Results
Lmax - Lmin
Where,L = the actual load Test Systems:
L,= the scaled load which is used as
input to the net The developed algorithm was tested
Lmax= the maximum load, 1.5 to 2 times with hourly load data for the following
the peak load for the whole year systems: (i) OSEB (Orissa State Electricity
Lmin = the minimum load, 0.5 to 0.75 Board, India); (ii) IEEE 24 bus reliability
times the valley system. The two systems have quite different
ANN Architecture: daily load patterns. The load data of OSEB
The artificial neural network architecture used system for the year 1990 has peak load in
is a feed forward network with three layers, August and a valley load in March. While the
IEEE 24 bus reliability test system is a winter In order to test the generalization
peaking system with peak load in December, property or exploitation and interpolation
it has a second peak in June at 90% of capability of the net in more details, the hourly
annual peak. load data was divided into four groups, ie, A, B,
C and D. Four distinct training and testing sets
As daily load pattern for normal working days
were prepared as detailed in Table 3. The
were quite different from those of weekend
results are presented in Table 3. From Table 4
days and holidays. The load data for each
it can be seen that the network is able to
system was divided into two groups, ie,
perform both interpolation and extrapolation
normal working days and weekend days and
quite well with less than 5% average error.
holidays. These two sets were treated
Extrapolation ability is of particular interest as it
shows that network can predict even for
unknown situations. Only for few stray cases,
the errors have been more than 5%.
V. CASE STUDY
Development and Implementation of LF
application for one of the distribution utilities
.•........ at Mumbai.
.•..... .•...•..•.. '"""' •..•............•
;. -••..........•...•~ ...•...•
.; " Requirement:
2.0 Short term load forecasting reports for
e , fS """~o~oW"'1'"'t''''~b'o<r'--'-'-'' ~otQd •• -;000
submitting day ahead schedule at 15 minutes
NO.: of H~,.Qtfon ~---
interval to SLDC, Kalwa for ABT
l"ll",," .'~\ .".~. 'IJI<:H_. _
t t.' .....!t~!.J
r .• ~!'I~-:'!.U,~,.r~IN;-
' , compliances.
Figure4Convergence characteristics of @, Data required for application development:
MLFF Neural Network 1. Three years past load data at 15
Supervised Learning minutes interval.
2. Three years congruent weather data of
The ANNs used for forecasting hourly humidity, rainfall, temperature and wind
load consists of five input neurons, two 25 direction.
3. City holiday calendar.
25hidden layer neurons and one output layer
@; Data required for running load forecast:
neuron. Twenty-four separate ANNs one for
1. Weather forecast at 15 minutes interval.
each hour forecast, were trained using the 2. Actual weather of past day at 15
input-output data pairs. Separate ANNs were minutes interval.
trained for weekdays and weekends. Thus, a 3. Major network and load changes.
total of 48 ANNs were trained for each
system. On the basis of a large number of
simultaneous optimal values for the learning
coefficientm) and momentum factor(a) used
for training of each ANN was obtained. After
the convergence of the training algorithm,
each ANN was tested using input output pairs
from the test set data.
Testing Generalization Property
DESCRIPTION FOR TRAINING AND TESTING TABLE 4
DATA SET FOR ANN GENERALISATION
TEST RESULTS FROM GENERALISED
Training Set Data Testing Set Data
Case I Taken randomly Taken randomly Training Error (%) Testing Error (%)
from region S&C from region A&D
extrapolating ability in
Max. mm avo max
Case II Taken randomly Taken randomly Case I - 3.62 0.213 1.487 - 5.334
from region C&D from region A&S
extrapolating ability in
Case II - 2.93 0.085 0.986 - 6.269
Case III Taken randomly Taken randomly
Testing Case III - 3.19 - 0.040 1.40 4.915
from region A&S from region C&D
extrapolating ability in 0.12 3.36
downward direction 1.26
Case IV - 3.91 0.048 - 4.594
Case IV Taken randomly Taken randomly -0.01 0.97
from region A,S, from region A,S
@; SCREEN SHOTS of the LF applications.
2. G.Gross and F.D.Galiana. 'Short-term Load Forecasting',
Proceedings of IEEE, vol 75, nol2, December 1987, p 1558-1573.
3. AK Mahalnobis, D.P.Kothari and S.I.Ahson. 'Computer Aided
Power System Analysis and Control', Tata Mcgraw Hill Publication
Co, New Delhi. 1988.
4. Paper on LF by EugeneAR?inberg State University
I. COM: Clean Development Mechanism.
2. SEB: State Electricity Board
3. SERC: State Electricity Regulatory Commission.
4. CERC: Central Electricity Regulatory Commission.
5. CEA: Central Electricity Authority.
6. DISCOM: Distribution Company.
7. TRANSCO: Transmission Company.
8. GENCOs: Generation Companies.
9. STU: State Transmission Utility.
10. ABT: Availability Based Tariff.
• II. R-APDRP: Restructured Accelerated Power Development and
" t- - 12. INR: Indian Rupees ..
....... - ===
....... __ _ _._ _ ..••-_ _ _ __
.......••... ........• .. .. .
Bharati Vidyapeeth University, College of Engineering, Pune
I. CEA, Mop and RAPDRP Website.