Short‐term load forecasting based on <scp>LSTNet</scp> in power system
Rong Liu, Luan Chen, Weihao Hu, Qi Huang
Abstract
Accurate short-term power load forecasting is very important in power grid decision-making operations and users power management. However, due to the nonlinear and random behavior of users, the electrical load curve is a complex signal. Although a lot of research has been done in this field, a predictive model with good accuracy and stability is still needed. To further improve this situation, this article proposes a novel model: long-term and short-term time series network (LSTNet) to predict load. This article firstly analyzes the correlation between other variables and load and uses Spearman correlation coefficient to measure the impact of other variables on the load and does autocorrelation analysis of load itself. Then, this article designs a load forecasting model based on LSTNet. The model uses a convolutional layer composed of a convolutional neural network (CNN) to capture short-term characteristics of load and short-term dependencies of variables, while a recurrent layer and recurrent-skip layer composed of long-term short-term memory network (LSTM) to capture long-term characteristics and variables of load long-term dependence, the adaptive regression part composed of autoregressive model (AR) to improve the robustness of the model. The experiment results show that the LSTNet model has better performance.