Time Series Prediction With Neural Networks. Application To Electric Energy Demand

Abstract. Electric energy demand forecasting represents a fundamental information to plan the activities of the companies that generate and distribute it. So a good prediction of its demand will provide an invaluable tool to plan their production and growth policies. This demand may be seen as a temporal series when its data are conveniently arranged. In this way the prediction of a future value may be performed studying the past ones. Neural networks have proved to be a very powerful tool to do this. They are mathematical structures that mimic that of the nervous system of living beings and are used extensively for system identification and prediction of their future evolution. In this work a neural network is presented to forecast the evolution of the monthly demand of electric consumption. Two strategies are proposed: the first uses a network that is trained once an then used to predict future values of the time series, while in the second the network is trained with all the past data every time a prediction is to be performed. The Spanish monthly consumption from 1975 to 2002 has been used to validate the models proposed. Errors smaller than 5% have been obtained in most of the predictions.