Symbolic and connectionist machine learning techniques for short-term electric load forecasting

Abstract

This work applies connectionist neural network learning techniques and symbolic machine learning techniques to the problem of short-term electric load forecasting. The short-term electric load forecasting problem considered here is the prediction of bus loads one day ahead. The forecast quantities of interest are average integrated daily load and daily peak load.

The primary objectives of this work are two-fold: to determine the forces driving the load demand and produce a human intelligible model, and use of this model to forecast load for new, unseen scenarios.

In the first part of this work, connectionist techniques for modeling bus load is presented. Critical design issues for neural network modeling and implementation such as neural network architecture, training database creation, training dataset selection, training data normalization are presented in context of nonlinear modeling in general and electric load forecasting in particular. Local function approximation and nearest neighbor norms techniques are applied to this task. Simulations are performed for forecast of average bus loads of the town of Blacksburg, Virginia, U.S.A; the connectionist model is able to forecast integrated average daily load with an accuracy of about 2.5%. Connectionist neural network knowledge acquisition algorithms are however, not mature enough, presently, to handle complex real world problems such as knowledge extraction from large databases. Presence of symbolic along with numeric data in input and output poses problems for data pre-processing for neural network training. Only at the time of completion of this thesis are researchers discussing the possibility of using special techniques to present symbolic data for neural networks. Also, multilayer feedforward networks trained by the backpropagation algorithm perform poorly in forecasting chaotic patterns such as those encountered in peak load demand.

Symbolic machine learning techniques are powerful concept acquisition techniques that extract underlying knowledge from large databases. They are sufficiently powerful to accept symbolic and numeric data. Inductive learning algorithms employing a statistical 72 test as the splitting criterion are applied to extract load dependency information. The extracted patterns are expressed as graphic decision trees and equivalent human intelligible high level language if-then rules. Implementation details of the statistical decision algorithm are discussed and simulations are performed to construct decision trees. Using this model, new cases are forecast. This algorithm is capable of forecasting holiday and weekend loads too. The proposed algorithm is robust enough to handle raw, unprocessed databases which contain missing data. The peak load forecasting problem is solved using a simple methodology that combines the robustness of decision trees and the numerical accuracy of connectionist models.

The two paradigms, connectionist and symbolic learning techniques are compared from a knowledge acquisition and forecasting perspective and directions for further work suggested.