An adaptive and modular recurrent neural network based power system load forecaster

Abstract

This paper describes a recurrent neural network (RNN) based hourly load forecaster for hourly prediction of power system loads. The system is modular, consisting of 24 RNNs, one for each hour of the day. The RNNs considered are sigmoid type neural networks with a single hidden layer. Two types of recurrency are considered: one has connections between the hidden layer nodes, and the other has feedback from output to hidden layer nodes. The hours of the day are divided into four categories and a different set of load and temperature input variables is defined for the RNNs of each category. The RNNs are trained with Pineda's recurrent backpropagation algorithm. To handle non-stationarities, an adaptive scheme is used to adjust the RNN weights during the online forecasting phase. The performance of the forecaster was evaluated on real data from two electric utilities with excellent results.