Short-term Load Forecasting for Distribution Substations Based on Residual Neutral Networks and Long Short-Term Memory Neutral Networks with Attention Mechanism
Hailong Li, Pan Zhang, Chunhui Li
Abstract
Electric load at distribution substation level has strong volatility due to various customer characteristics, making it challenging for short-term load forecasting. This paper proposes a novel short-term load forecasting method for distribution substations based on combined residual neutral networks (ResNet) and Long Short-Term Memory (LSTM) neutral networks with attention mechanism (ResNet-LSTM-Attention) where the advantages of ResNet and LSTM are combined together. A 34-layer ResNet is built to extract the latent features of data. Then, a two-layer LSTM is added to learn the time series characteristics of the extracted features, followed by an attention mechanism to selectively pay attention to the state of the hidden layer. The K-fold cross-validation is further adopted to make full use of data and improve the generalization ability of the model. Finally, the data set from North Carolina and the smart meter energy consumption data from the Low Carbon London project are employed to verify the validity of the method. Compared with traditional LSTM model, the proposed ResNet-LSTM-Attention method shows smaller mean absolute percentage error and better performance on load forecasting.