A meta-learning framework with temporal feature integration for electricity load forecasting

Accurate electricity load forecasting is essential for the stability, efficiency, and sustainability of modern power systems. However, individual forecasting models often lack generalization across temporal and regional variations and offer limited interpretability. This study proposes a comprehensive meta-learning-based forecast combination framework to enhance both prediction accuracy and model transparency. Using hourly load data from 20 European countries spanning 2018 to 2024, the framework incorporates time-aware features such as hour of the day, day of the week, month, and public holidays. Ten diverse base models—including XGBoost, LightGBM, Random Forest, and LSTM—are trained globally, from which the top five performers are selected (based on R², MAE, and MAPE) and fed into five meta-learners: Ridge Regression, Lasso, Random Forest, Gradient Boosting, and MLP. These meta-models are trained using both model predictions and engineered time features. Experimental results demonstrate superior performance, with the best-performing meta-learner (Random Forest Regressor) achieving a coefficient of determination (R²) of 0.9998 and a Mean Absolute Percentage Error (MAPE) of 0.79%, significantly outperforming traditional ensemble methods. Furthermore, the inclusion of lag features and 5-fold cross-validation led to substantial improvements across all models, including dramatic reductions in MAE (up to 87%), MAPE (up to 88%), and MSE (up to 97%), along with near-perfect R² scores (~ 1.000). Additionally, SHAP-based explainability reveals the contribution of individual time-based features and the influence of each base model within the ensemble, thereby enhancing transparency and supporting practical decision-making.