Short-term load forecasting using artificial neural networks

The deregulation of the power system industry has made short term load forecasting increasingly important. This paper presents an artificial neural network based hour ahead load forecasting model that improves upon previous models by using the entire load profile of the previous day, rather than making potentially unjustified assumptions about the functional relationship between past hours load and current load. Historical load data for the ISO-New England control area was used to test the proposed model. The mean absolute percentage error for the hour ahead load forecasting was found to be 0.439%, which compares favorably to previous models. In addition, seasonal changes and weekends appear to have relatively small effects on the network performance. This suggests that the use of the 24 past hours load as input variables can potentially create better hour-ahead forecasting models.