TCN-SGConv-LSTM with Attention Mechanism for Power Load Prediction

The goal of this study is to improve power load forecasting accuracy through the integration of the Time Convolutional Network (TCN), Structured Global Convolution (SGConv), and Long Short-Term Memory (LSTM) modules. The proposed model, TCN-SGConv-LSTM, leverages TCN for capturing long-term temporal dependencies and combines the memory capabilities of LSTM with SGConv to enhance global feature extraction. The model further integrates the Squeeze-and-Excitation (SE) attention mechanism to better recognize key features. Experiments using power load data from central air conditioners of a mall, electric vehicle charging station, and energy storage station demonstrate superior performance over traditional LSTM and TCN-LSTM models, significantly reducing the Mean Absolute Percentage Error (MAPE). This approach offers a new and practical solution for power load forecasting with broad potential applications.