Short-Term Load Forecasting Method Based on PE-SFTCN-BILSTM Network

As the global energy internet infrastructure advances and electricity market reforms progress, short-term electricity load forecasting has become increasingly vital for the scheduling of the power market. Given the evident daily and weekly periodicity in short-term electricity load, this paper introduces a novel time encoding approach based on Periodic Encoding (PE). In order to comprehensively explore the temporal characteristics of electricity load and the influence of external meteorological factors on it, a short-term load forecasting model is proposed, which combines Time Convolutional Network (TCN), Feature Selection Network (SFNET), and Bidirectional Long Short-Term Memory Network (BILSTM). Firstly, the electricity load data is encoded using the PE-based time encoding method. Subsequently, input features for the forecasting model are selected using Pearson correlation coefficients and maximum mutual information coefficients, and a feature matrix is constructed using a sliding window approach. Finally, the SFNET is employed to extract information from different input features in the load data, while the BILSTM network extracts temporal features and performs the forecasting. Using the real data of a place in Australia as an example, the prediction accuracy reached 96.82%, and compared with the common load forecasting methods, the accuracy was the highest, which verified the higher prediction accuracy of the proposed model.