AI-Driven Dual-Scale Building Energy Benchmarking for Decarbonization

Abstract

Energy benchmarking is a critical approach for policymakers, building owners, and architects to monitor energy consumption and implement energy conservation strategies. However, challenges exist, including limited data availability for annual and monthly energy usage, the complexity of machine learning models that are difficult for users to interpret, and difficulties in classifying buildings based on their energy use patterns. This study addresses these issues by proposing a data-driven building energy benchmarking framework, utilizing white-box and grey-box AI models at both annual and monthly scales. The study focuses on building energy data from Washington, DC, in the United States. The first stage of the study involves clustering the energy benchmarking data using an unsupervised K-Means algorithm. This clustering group buildings into four distinct categories based on their monthly energy usage patterns, allowing for better insight into the energy use characteristics of each cluster. Subsequently, two robust models, a white-box model (multi-linear regression) and a grey-box model (LGBM), were employed to predict building energy consumption on both an annual and monthly basis. The final stage of the study investigates the sensitivity of various building attributes, such as building characteristics, to understand their impact on energy usage and carbon emissions. The results reveal that the LGBM model provides more accurate predictions compared to the multi-linear regression model, though both models demonstrate relatively similar sensitivity to key building attributes. Notably, Energy Star ® ratings, building type, weather conditions, and building area emerged as the most significant input factors influencing energy consumption. This research not only develops a reliable building classification framework but also highlights the effectiveness of combining annual and monthly energy benchmarking for the dual-scale use of advanced AI techniques. It also offers a replicable approach that can be generalized to other cities, particularly those without existing energy benchmarking programs. By advancing AI-driven methodologies for building energy benchmarking, this study contributes to the broader goal of achieving net-zero carbon emissions by 2050 and supports efforts toward decarbonization and climate resilience across the United States.

Keywords: AI-Driven, Building Energy Benchmarking, Machine Learning, Model sensitivity, Decarbonization