Ensemble Deep Learning Method for Short-Term Load Forecasting

Abstract

Short-term load forecasting (STLF) is the basis for the economic operation of the power system, and accurate STLF can optimize the power company's generation scheduling and improve the economics and safety of power grid operation. Classical regression-based models are mainly developed for stationary time series, while power load is typical nonstationary one. Shallow neural network model usually cannot capture complicated non-linear pattern efficiently, while power load features complicated varying patterns due to the numerous factors such as region, climate, economics, industry. Deep neural network, especially recurrent neural network (RNN) methods, like long short-term memory (LSTM), can model complicated pattern efficiently with the state-of-the-art erformance, but the training of the deep network becomes much harder with the increase of input sequence length. Since the power load holds large span of periodicity from daily through yearly, LSTM cannot fully exploit the inner correlation of power load. In this paper, ensemble deep learning method is proposed to exploit both non-linear pattern by LSTM and large-span period by similar day method. The proposed method integrates several LSTM networks, and each network is fed with different input time sequences which are selected regarding the similarity of load pattern. Experiment results show the effectiveness of the proposed method when comparing with exiting methods.