An Interpretable Comparative Framework for Monthly Electricity Generation Forecasting Using Statistical and Machine Learning Models

Abstract

Precise forecasting of electricity generation for planning in the power systems, scheduling its operations, coordinating in markets and analysing energy policies are crucial to success. There are also growing needs for excellent and interpretable forecasting frameworks that have become essential as electricity systems become more complex due to seasonal demand variability, infrastructure transitions, and changes in production patterns. We introduce a comparative forecasting framework to predict monthly electricity generation using the open-access Electricity Generation Time Series data set. The forecasting target is the United States total electricity generation series, and analysis bridges classical statistical forecasting methods to modern machine learning approaches within a unified evaluation pipeline. In the study, we assess six forecasting models: ARIMA, Holt-Winters ETS, SVR, RF, XGBoost and LightGBM. This supervised representation from past observations, rolling statistics and seasonal encodings are extracted to improve prediction performance. A rolling-origin validation strategy is used for realistic temporal evaluation and to avoid the issue of information leakage through random train–test splitting. The performance of the forecasts, which is evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Symmetric Mean Absolute Percentage Error (sMAPE) and coefficient of determination (R²). The comparison is further bolstered by bootstrapped confidence intervals, Diebold–Mariano significance testing; and Ljung-Box residual diagnostics. Moreover, model interpretability and robustness are also assessed via SHAP-like feature importance analysis, lag sensitivity experiments, noise perturbation tests and ablation analyses. Random Forest had the best overall predictive performance among models as per these results, with LightGBM and XGBoost following closely behind in predictive accuracy and SVR being comparatively weak in generalizing on monthly electricity generation series. Analysis of statistical significance established that even though none of the differences were large, still some improvements over baseline models were significant and not just incidental. Interpretability analysis also showed that seasonal lag variables, especially these annual lag terms and rolling mean statistics, were most predictive in forecasting performance.