A data driven framework for forecasting post-retrofit HVAC and lighting energy consumption

Buildings account for a substantial share of global energy consumption, yet significant uncertainty persists in predicting energy savings following retrofit interventions, contributing to the widely reported energy performance gap. Existing prediction approaches primarily rely on physics-based simulation or short-term operational forecasting, with limited research addressing cross-building prediction of post-retrofit energy performance. This study develops a data-driven cross-building machine learning framework to forecast post-retrofit heating, ventilation, and air conditioning (HVAC) and lighting energy consumption using pre-retrofit building characteristics and operational variables. A Random Forest regression model was trained on a large-scale dataset comprising energy and operational records from 9450 commercial buildings, incorporating building characteristics, occupancy patterns, system specifications, and climatic variables. The model was validated using two independent commercial office buildings excluded from training to evaluate predictive generalization under realistic deployment conditions. Model performance was assessed using mean absolute percentage error (MAPE), root mean squared error (RMSE), coefficient of variation of RMSE (CVRMSE), and coefficient of determination. The proposed framework achieved strong predictive accuracy, with MAPE values below 7% and CVRMSE below 10% across validation cases, outperforming commonly reported benchmarks in building energy forecasting. Results demonstrate that cross-building learning captures transferable relationships between pre-retrofit conditions and post-retrofit energy performance without requiring building-specific calibration. Feature importance analysis identified baseline HVAC consumption, occupancy-related variables, building size, and climatic factors as dominant drivers of post-retrofit energy demand. This study contributes a scalable and reliable predictive approach for retrofit performance forecasting, reducing uncertainty in energy savings estimation and supporting evidence-based investment and measurement and verification processes. The findings highlight the potential of data-driven cross-building models to enhance the credibility and effectiveness of large-scale building energy retrofit programs.