Advancing Urban Building Energy Modeling: The Role of Hybrid Energy Modeling in Enhancing Energy Consumption Predictions

Urban building energy modeling (UBEM) is essential for understanding energy consumption and developing sustainable policies at the city scale. However, current UBEM approaches overlook spatial and temporal interactions and lack generalizability across diverse urban contexts. This study introduces a hybrid framework that integrates physics-based simulations with machine learning based residual learning to enhance prediction accuracy using real energy consumption data. The methodology incorporates GIS-supported data collection and processing. Multiple ML models were applied to predict monthly consumption and validate their performance. Meanwhile, a physics-based model is used to simulate hourly energy consumption. The best performing ML model was later used for daily residual learning to calibrate physics-based simulation outputs. The framework was tested on residential buildings connected to the District Heating Network in Turin, Italy. Results showed LGBM achieved the highest performance with a R2 of 0.883 and a MAPE below 15% in most months. Residual learning reduced daily prediction error in 80% of cases, with up to 75% improvement in extreme cases. After model calibration, 65% of buildings achieved a daily MAPE below 30%, and 55% fell below 20%, demonstrating consistent error reduction across varied building types and consumption levels. This confirms the effectiveness of the hybrid approach in enhancing accuracy and reliability at the urban scale.