Professional Documents
Culture Documents
2006/11/22
Liyao,Hou
Introduction
Neural network input variables
The ANN used: type and structure
ANN training: data and algorithm
Results
Forecasting for a local load
Conclusion
Contribution
Applied artificial neural networks (ANN) for short term load
forecasting in the power system
Demonstrate ANN capabilities in load forecasting without using
load history as an input.
Keywords:
Power systems; Load forecasting; Artificial neural networks
Backgrounds:
Short term load forecasting is
important to power systems:
• Cost savings and secure operational conditions
• Optimize energy prices
Literature review
ANNs have been applied to many power system analysis and control.
• Load forecasting, static and dynamic security assessment,
• dynamic load modeling, alarm processing and fault diagnosis
The authors demonstrated ANN capabilities in short term load
forecasting using real load and weather data from the Hydro-Quebec
databases(1999)
Three types of variables were used as inputs to the neural network:
• (a) hour and day indicators
• (b) weather related inputs
• (c) historical loads
Motivation
Generally, load history is not available, so estimated values of this
load are used instead.
However, a small error in estimated values may grow to a serious
problem since it is fed back as an input to the load forecasting
procedure.
Input variables:
Statistical analysis helps determine input variables
General input variables :
• (a) hour and day indicators
• (b) weather related inputs
• (c) historical loads.
Weather variables
Correlation matrix for all the variables
• L (load), T (temperature),W (wind velocity) and C (cloud cover).
Weather variables
Effect of weather variables on cloud cover load.
Relationship between cloud
cover and the load
Result analysis
Only the temperature is considered in this application.
Other weather variables (wind velocity and cloud cover) are neglected,
ANN structure
ANN has 32 inputs,
65 hidden neurons
1 neuron in the output layer
Inputs are
• a day indicator,
• an hour indicator,
• and the temperature
The ANN output is
• forecasted total load at hour k.
Training data
Error back propagation algorithm, is used to train the layered
perception-type ANN.
1994 data is divided into several windows
• ½ (non-consecutive ones) is used for training
• ½ is used for testing the ANN.
Training was done for all the data widows at the same time;
Objective
• Adjust the ANN weights to minimize the error between desired and ANN outputs.
Training algorithm
Taylor expansion of the
error as weight function
Objective is to
eliminate this error
(to make it zero)
Multiplying both
sides of Eq. (5) by J and
solving for DW gives (6).
W -- vector containing all the weights in the network,
J -- Jacobian vector of derivatives of the error to each weight.
l is a scalar and [I] is the unity matrix
ANN performance
ANN output compared to the actual load and to the Hydro-Quebec
estimation of this load.
Contribution
Using ANNs for short term load forecasting for the Hydro-Quebec
system
Demonstrate ANN capabilities in load forecasting without using load
history as an input.
Limitation
The possibility for better results exists and can be achieved by using:
• (1) advanced types of ANN
• (2) better selection of input variables
• (3) better ANN architecture
• (4) better selection of the training set