A neural network architecture for load forecasting

Abstract

Neural networks offer superior performance for predicting the future behaviour of pseudo-random time series. The authors present a neural network architecture for load forecasting which is capable of capturing the relevant relationships and weather trends. The proposed architecture is tested by training three neural networks, which in turn are tested with weather data form the same four-day period. The network is made up of a series of subnetworks each connected to its immediate neighbors in a way that takes into consideration not only current weather conditions but also the weather trend around the hour for which the forecast is being made. The neural network forecasts were very close to the actual values despite the facts that only a small sample was used and there were errors in the data. A more comprehensive study is being contemplated for the next phase. One of the issues to be addressed is the expansion of the scope of the research to include data from a complete season (three consecutive months) over several years.<>