Energies, Vol. 14, Pages 5873: A Two-Stage Short-Term Load Forecasting Method Using Long Short-Term Memory and Multilayer Perceptron (Energies)

Load forecasting is an essential task in the operation management of a power system. Electric power companies utilize short-term load forecasting (STLF) technology to make reasonable power generation plans. A forecasting model with low prediction errors helps reduce operating costs and risks for the operators. In recent years, machine learning has become one of the most popular technologies for load forecasting. In this paper, a two-stage STLF model based on long short-term memory (LSTM) and multilayer perceptron (MLP), which improves the forecasting accuracy over the entire time horizon, is proposed. In the first stage, a sequence-to-sequence (seq2seq) architecture, which can handle a multi-sequence of input to extract more features of historical data than that of single sequence, is used to make multistep predictions. In the second stage, the MLP is used for residual modification by perceiving other information that the LSTM cannot. To construct the model, we collected the electrical load, calendar, and meteorological records of Kanto region in Japan for four years. Unlike other LSTM-based hybrid architectures, the proposed model uses two independent neural networks instead of making the neural network deeper by concatenating a series of LSTM cells and convolutional neural networks (CNNs). Therefore, the proposed model is easy to be trained and more interpretable. The seq2seq module performs well in the first few hours of the predictions. The MLP inherits the advantage of the seq2seq module and improves the results by feeding artificially selected features both from historical data and information of the target day. Compared to the LSTM-AM model and single MLP model, the mean absolute percentage error (MAPE) of the proposed model decreases from 2.82% and 2.65% to 2%, respectively. The results demonstrate that the MLP helps improve the prediction accuracy of seq2seq module and the proposed model achieves better performance than other popular models. In addition, this paper also reveals the reason why the MLP achieves the improvement.