Short Term Load Forecasting Using Neural Nets

Load forecasting is decisive in the operation of power systems, for economic and security reasons. Many techniques have been proposed in the last two decades [1]. This work presents a short-term load forecasting system (whose main objective is to maintain the generation-load balance) using Neural Networks. Neural Networks have demonstrated to be a very efficient technique to time series forecasting, particularly in load series [2]. In the application shown in this paper, a Neural Network is used to learn the daily load behaviour of a real electrical system (CEMIG, Brazil, 1993). The network inputs are: past load data, the forecasting hour and the type of day (weekday or weekend). The windowing technique [3] is used to identify the series characteristics. Many neural nets with different architectures were tested and the results evaluated in terms of forecasting errors. We achieved an average forecasting error close to 1.5%. The forecasting system was developed in C programming language and includes the pre-processing of the input data, the network training and the forecasting. This system offers to the user options such as: tuning of some network parameters (learning rate, momentum term, number of processors in any layer), usage or not of the forecasted values as network inputs, adjustment of the size of the training window etc.