Neural system for power load prediction in a week time horizon

Abstract

In this paper a neural system for predicting electric power load in Poland in a week time horizon is presented. The system consists of seven multi-layer neural networks that have common input. Each network is dedicated to predict the total load in one of the seven successive days. Various form of input vectors as well as various ways of encoding them were tested. Verification which type of input data are crucial as well as which periodic aspects should be taken into account in data representation in week prediction was studied. Various numbers of neurons in a hidden layer were tested as well. The mean absolute percentage error (MAPE) is equal to 2.6% for the most effective system.