A multi-step hourly load forecasting system using neural nets

This work investigates the performance of a neural network-based hourly load forecasting system. Tests are made varying the forecasting leading time from 1 to 744 hours ahead. Forecasting electric load for long periods ahead (i.e., over 24 hours) requires the neural network to feed itself with predicted load values (multi-step prediction) in order to forecast the next period. The results obtained in these tests are very good when compared with single-step prediction, which uses only the actual load values available for the next prediction. This feature is a key result in power systems operation since it allows accurate prediction with large leading times. In the experiments we use real load data from the Electric State Company of Minas Gerais (CEMIG) and predict load for a whole year (from March/1993 to February/l994). The results are evaluated using three error figures: MAPE (Mean Absolute Percentage Error), RMSE (Root Mean Squared Error) and Theil's U (rate between the RMSE of the actual forecasting system and the RMSE of a naive forecasting system). In many cases, results exhibit a MAPE below 2%. Temperature and other weather data are not considered in these predictions.