Hybrid Vector Auto Regression and Tree-Based Ensembles for Energy Load Forecasting: A Case Study

The increasing depletion of non-renewable energy resources due to population growth, urbanization, and technological progress has heightened the need for accurate energy consumption forecasting to facilitate sustainable infrastructure and efficient resource management. Numerous conventional unimodal models have difficulty capturing non-linear trends in energy consumption patterns. We propose a two-stage hybrid forecasting framework that synergistically integrates Vector Autoregression (VAR) for capturing linear data and tree-based ensembles (Random Forest and XGBoost) trained on the residuals to capture nonlinearities. Results suggest that the VAR model is able to captures most of the energy patterns, while the hybrid extensions capture non-linearity. Across all evaluation metrics, the base VAR model and its hybrid extensions achieved strong predictive performance, with normalized root mean square errors (NRMSE) as low as 0.6977 and R² values exceeding 0.5 in the most predictable settings. Although the base VAR model handled most temporal dynamics well, the hybrid model incorporating tree-based ensemble offered extra versatility and marginal gain, especially with more intricate circumstances. Most importantly, our hybrid framework performed better than or closely resembled top-performing deep learning and ensemble-based strategies published in the recent literature by achieving competitive accuracy levels with reduced error rates.