An adaptive recurrent neural network system for multi-step-ahead hourly prediction of power system loads

In this paper a new recurrent neural network (RNN) based system for hourly prediction of power system loads for up to two days ahead is developed. The system is a modular one consisting of 24 non-fully connected RNNs. Each RNN predicts the one and two-day-ahead load values of a particular hour of the day. The RNNs are trained with a backpropagation through time algorithm using a teacher forcing strategy. To handle non-stationarities, an adaptive scheme is used to adjust the RNN weights during the forecasting phase. The performance of the forecaster is tested on one year of real data from two utilities and the results are excellent. This recurrent system outperforms another modular feedforward NN-based forecaster which is in beta testing at several electric utilities.<>