An AI Hybrid Building Energy Benchmarking Framework Across Two Time Scales

Buildings account for approximately one-third of global energy usage and associated carbon emissions, making energy benchmarking a crucial tool for advancing decarbonization. Current benchmarking studies have often been limited to mainly the annual scale, relied heavily on simulation-based approaches, or employed regression methods that fail to capture the complexity of diverse building stock. These limitations hinder the interpretability, generalizability, and actionable value of existing models. This study introduces a hybrid AI framework for building energy benchmarking across two time scales—annual and monthly. The framework integrates supervised learning models, including white- and gray-box models, to predict annual and monthly energy consumption, combined with unsupervised learning through neural network-based Self-Organizing Maps (SOM), to classify heterogeneous building stocks. The supervised models provide interpretable and accurate predictions at both aggregated annual and fine-grained monthly levels. The model is trained using a six-year dataset from Washington, D.C., incorporating multiple building attributes and high-resolution weather data. Additionally, the generalizability and robustness have been validated via the real-world dataset from a different climate zone in Pittsburgh, PA. Followed by unsupervised learning models, the SOM clustering preserves topological relationships in high-dimensional data, enabling more nuanced classification compared to centroid-based methods. Results demonstrate that the hybrid approach significantly improves predictive accuracy compared to conventional regression methods, with the proposed model achieving over 80% R2 at the annual scale and robust performance across seasonal monthly predictions. White-box sensitivity highlights that building type and energy use patterns are the most influential variables, while the gray-box analysis using SHAP values further reveals that Energy Star® rating, Natural Gas (%), and Electricity Use (%) are the three most influential predictors, contributing mean SHAP values of 8.69, 8.46, and 6.47, respectively. SOM results reveal that categorized buildings within the same cluster often share similar energy-use patterns—underscoring the value of data-driven classification. The proposed hybrid framework provides policymakers, building managers, and designers with a scalable, transparent, and transferable tool for identifying energy-saving opportunities, prioritizing retrofit strategies, and accelerating progress toward net-zero carbon buildings.