An interpretable short-term electrical load forecasting model based on SHapley Additive exPlanations——A case study in Haidian, Beijing

In recent years, with the rapid development of modern society, the demand for electric power is booming. Accurate forecasting of electrical peak load is crucial for maintaining a stable electricity supply. This study developed daily electrical peak load forecasting models for Haidian by statistical (Stepwise Multiple Linear Regression, SMLR) and machine learning (Extreme Gradient Boosting, XGBoost, and Least Absolute Shrinkage and Selection Operator-Extreme Gradient Boosting, LASSO-XGBoost) algorithms under two scenarios, and conducted an interpretability analysis on the optimal models to quantify the contributions of the critical features for daily electrical peak load through SHapley Additive exPlanations (SHAP) and Partial Dependence Relationship (PDR) analysis. The results showed that XGBoost and LASSO-XGBoost achieved superior predictive accuracy (R²>0.93), compared with SMLR model. Meteorological factors, past electricity peak load, and holiday parameter were critical features for peak load forecast: Peak load in Haidian was significantly influenced by the "weekend effect", which was notably lower on holidays compared to workdays. Sensible temperature was the most critical meteorological factor for peak load, which increased rapidly when it exceeded 29.5°C. Electricity load exhibited short-term temporal dependence, and previous peak loads were also important contributors to the prediction result. Our results provided valuable insights for the electricity sector in formulating strategies for electricity production and dispatching.