Evaluating regression models and machine learning for long-term energy demand forecasting – Germany

Reliable long-term energy demand forecasting underpins energy-system planning and decarbonization policy design, yet most established methods are calibrated for short horizons and may not capture structural change. This study compares four forecasting models: multiple linear regression (MLR), a vector error correction model (VECM), support vector regression (SVR), and an artificial neural network (ANN). Germany’s final energy consumption (FEC) is projected to 2050 from annual data spanning 1992–2021. The SVR and ANN were tuned using particle swarm optimization and Bayesian optimization. Accuracy is assessed using mean absolute error (MAE), mean squared error (MSE), and the coefficient of determination (R²). The SVR achieves the highest test accuracy (MAE = 9.19, R² = 0.974), but its long-term projections substantially exceed the historical range, suggesting extrapolation instability. The ANN attains comparable test accuracy (MAE = 10.05, R² = 0.965) with stronger generalization and more plausible projections, and is therefore adopted for the base-case and scenario analyses. MLR is transparent but less accurate, whereas VECM forecasts are sensitive to the assumed cointegration rank in the small-sample setting. The ANN base-case indicates an approximately 12% decline in FEC by 2050 relative to 2008, driven primarily by fuel substitution and efficiency improvements. Across five decarbonization scenarios, projected 2050 FEC ranges from 1476 to 2195 TWh, demonstrating that transition assumptions dominate long-term forecast uncertainty. Even the most ambitious pathway, approximately 39% below 2008, falls short of Germany’s 45% FEC reduction target for 2045. The results indicate that generalization and extrapolation behavior are as important as predictive accuracy in model selection for long-term energy transition analysis.