Short-Term Load Forecasting

1
Short-Term Load Forecasting In Electricity Market
Acknowledge Dr. S. N. Singh (EE) Dr. S. K. Singh
(IIM-L)
N. M. Pindoriya Ph. D. Student (EE)
2
TALK OUTLINE

• Importance of STLF
• Approaches to STLF
• Wavelet Neural Network (WNN)
• Case Study and Forecasting Results

3
Introduction

• Electricity Market (Power Industry Restructuring)
• Objective Competition costumers choice
• Trading Instruments
• 1) The pool
• 2) Bilateral Contract
• 3) Multilateral contract
• Energy Markets
• 1) Day-Ahead (Forward) Market
• 2) Hour-Ahead market
• 3) Real-Time (Spot) Market

REACH Symposium 2008 1
4
Types of Load Forecasting
In electricity markets, the load has to be
predicted with the highest possible precision in
different time horizons.
(one hour to a week)
REACH Symposium 2008 2
5
Importance of STLF
System Operator

• Economic load dispatch
• Hydro-thermal coordination
• System security assessment

Generators

• Unit commitment
• Strategic bidding
• Cost effective-risk management

STLF
LSE

• Load scheduling
• Optimal bidding

REACH Symposium 2008 3
6
Input data sources for STLF
Real time data base
Historical Load weather data
Weather Forecast
Measured load
STLF
Information display
EMS
REACH Symposium 2008 4
7
Approaches to STLF

• Hard computing techniques
• Multiple linear regression,
• Time series (AR, MA, ARIMA, etc.)
• State space and kalman filter.
• Limited abilities to capture non-linear and
  non-stationary characteristics of the hourly load
  series.

REACH Symposium 2008 5
8
Approaches to STLF

• Soft computing techniques
• Artificial Neural Networks (ANNs),
• Fuzzy logic (FL), ANFIS, SVM, etc
• Hybrid approach like Wavelet-based ANN

REACH Symposium 2008 6
9
Wavelet Neural Network
WNN combines the time-frequency localization
characteristic of wavelet and learning ability of
ANN into a single unit.

WNN
Adaptive WNN Fixed grid WNN
Activation function (CWT) Activation function (DWT)
Wavelet parameters and weights are optimized during training Wavelet parameters are predefined and only weights are optimized
REACH Symposium 2008 7
10
Adaptive Wavelet Neural Network (AWNN)
Input Layer
Wavelet Layer
Output Layer
Product Layer
?ij
?j
?
w1
x1
v1
?
w2
?
wm
g
v2
xn
?

• BP training algorithm has been used for training
  of the networks.

REACH Symposium 2008 8
11
Mexican hat wavelet (a) Translated (b) Dilated
REACH Symposium 2008
12
Case study
California Electricity Market, Year 2007
(http//oasis.caiso.com/ )

• Data sets for Training and Testing

Seasons Winter Summer
Historical hourly load data (Training) Jan. 2 Feb. 18 July 3 Aug. 19
Test weeks Feb. 19 Feb. 25 Aug. 20 Aug. 26
REACH Symposium 2008 9
13
Case study

• Selection of input variables

• The hourly load series exhibits multiple seasonal
  patterns corresponding to daily and weekly
  seasonality.

REACH Symposium 2008 10
14
Case study

• Input variables to be used to forecast the load
  Lh at hour h,

Hourly load Trend
Hourly load Daily and weekly Seasonality
Temperature Exogenous variable
REACH Symposium 2008 11
15
Case study
REACH Symposium 2008 12
16
Case study

• Winter test week

REACH Symposium 2008 13
17
Case study

• Summer test week

REACH Symposium 2008 14
18
Case study

• Statistical error measures

WMAPE WMAPE WMAPE Weekly variance (10-4) Weekly variance (10-4) Weekly variance (10-4) R-Squared error R-Squared error R-Squared error
CAISO ANN AWNN CAISO ANN AWNN CAISO ANN AWNN
Winter 1.774 1.849 0.825 2.429 3.220 0.713 0.9697 0.9540 0.9917
Summer 1.358 1.252 0.799 2.115 1.109 0.369 0.9889 0.9923 0.9975
Average 1.566 1.551 0.812 2.272 2.164 0.541 0.9793 0.9732 0.9946
REACH Symposium 2008 15
19

• Thank you