AI-Driven Building Energy and Carbon Emissions Benchmarking at Multiple Scales

Leveraging data analytics for building energy benchmarking stands as a highly effective strategy to enhance energy conservation, minimize carbon footprint, and optimize energy management practices. Yet, data limitations, especially for the monthly and end-use loads, are a major challenge for most areas in the United States. Meanwhile, current benchmarking classifications are typically conducted by primary types and in particular climate zones. However, multiple building attributes impact energy consumption, leading to identical building types that may have distinct energy use patterns. Additionally, sophisticated artificial intelligence (AI) models, including classic machine learning, ensemble learning, and deep learning models, have been discussed in existing studies, revealing varied performance levels. Because of “black-box” issues of most existing AI approaches, studies have not provided a comprehensive post-prediction analysis to open the black-box models, limiting insights into building types, climate zones, and energy with model performance.To fill in the three gaps addressed above, this study proposes an AI-driven generalizable building energy and carbon emissions benchmarking approach that is applicable to any contiguous U.S. city in three layers, including annual total, annual end-use, and monthly electricity and natural gas energy usage. Building energy benchmarking classifications are conducted by intelligent clustering algorithms that challenge classic building classification models to improve the model performance. Furthermore, this study creates a monthly benchmarking approach that can accurately predict monthly electricity and natural gas loads from annual total energy use, improving energy efficiency and management. Meanwhile, the research develops a post-prediction model evaluation method to gain insights into the performance of multiple AI models. During post-prediction, a statistical acceptance interval evaluates and classifies the actual and predicted values into well-estimated, underestimated, and overestimated clusters. This approach also assesses prediction performance associated with energy patterns, building types, and climate zones, providing a deeper reference for future studies regarding model enhancement and energy conservation development.