Deep LSTM and GAN based Short-term Load Forecasting Method at the Zone Level

Abstract

Accurate short-term load forecasting (STLF) with time horizon ranging from seconds to hour is essential to revealing the deviation of day-ahead load forecasting. Especially, under the competitive market environment, accurate forecasting result can reduce the expensive cost on the activation of spinning reserve. Method based on deep long short-term memory RNN (LSTM) is promising to implement STLF. However, over-fitting is a common problem capping such model's generalization ability. To cope with it, generative adversarial networks (GAN) and deep LSTM based model is proposed in this paper. The deep LSTM is concatenated with the generator whose parameters are initialized by the pre-trained GAN. Essentially, the model aims to address the overfitting issue by the synthetic load produced by the generator. It has the effect of the regularization and thus the deep LSTM can make better forecasting on the unseen data. The results demonstrate that the proposed algorithm has better generalization ability.