Research on Load Forecasting Method Based on SHAP Value Analysis and Improved Transformer Model

In the context of increasingly complex global energy systems and continuously growing electricity demand, electric load forecasting has become a critical component for ensuring the stable operation and optimal scheduling of power systems. This paper proposes an electric load forecasting method based on a combined XLSTM-Transformer model and adaptive bandwidth kernel density estimation (ABKDE), which employs SHAP values for feature selection to enhance the accuracy and efficiency of the model. Initially, the SHAP values of each feature are calculated using an LSTM model, and features with minimal impact on forecast outcomes are eliminated. Subsequently, a combined XLSTM- Transformer model is utilized to capture long-term dependencies and intricate nonlinear relationships within time series data. The ABKDE approach is then applied to perform probabilistic forecasting, thereby quantifying the uncertainty in forecast results and generating probabilistic predictions across different confidence intervals. Finally, simulation verification is conducted on the dataset from the Tempe campus of Arizona State University. Results indicate that the proposed method not only improves point forecasting accuracy but also better quantifies the probability density forecasting of load.