Daily Load Forecasting Using Explainable AI-Driven Method With Integrated Error Correlation

Abstract

This paper presents a novel, explainable, and weather-independent day-ahead load forecasting (DALF) method. The proposed methodology integrates calendar-based segmentation using Classification and Regression Trees (CART), hybrid modeling with Multiple Linear Regression (MLR) and Gaussian Process Regression (GPR), and post-hoc error correction via Generalized Least Squares with Autoregressive Residuals (GLSAR). To enhance interpretability, Shapley Additive exPlanations (SHAP) are employed to quantify the influence of lagged load features and calendar variables across segmented forecasts. The framework is validated using half-hourly data from Thailand’s Electricity Generating Authority (EGAT) and cross-tested against France’s national load data, encompassing both regular and holiday-specific load profiles. Experimental results demonstrate that the MLR–GPR–GLSAR configuration achieves superior performance, with a Mean Absolute Percentage Error (MAPE) of 0.72% and an R2 of 0.99, outperforming Generalized Additive Models (GAM), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LGB), Convolutional Long Short-Term Memory (CNN-LSTM), and Neural Hierarchical Interpolation for Time Series (NHITS), as baselines. The proposed method offers a scalable, interpretable solution suited to data-sparse operational environments. It highlights the value of correcting residual autocorrelation and using calendar-aware segmentation in improving load-forecasting accuracy.