Urban Building Energy Modeling: A Comparative Study of Process-Driven and Data-Driven Models

Abstract

This study investigates the predictive capabilities of process-driven (PD) energy modeling and Machine Learning techniques, specifically Light Gradient Boosting Machine (LGBM) and Random Forest (RF) algorithms, in analyzing building energy consumption patterns. Leveraging a comprehensive dataset encompassing diverse building characteristics, energy-related variables, and operational configurations, the comparative performances of these methodologies is explored. Results reveal that while all approaches demonstrate promising predictive accuracies, LGBM exhibits a slight advantage over RF and the process-driven model. Moreover, the process-driven model showcases efficacy in colder seasons and for buildings of extreme ages, while encountering limitations in accurately modeling energy consumption for structures constructed during 1970s to 1990s. Conversely, Machine Learning models demonstrate consistent performance (with relative errors of 5–10%) across varied building ages, underscoring their adaptability and potential for capturing nuanced energy dynamics. However, a notable constraint lies in the availability of sufficient data for training Machine Learning models, posing challenges for model testing. These findings contribute to advancing our understanding of energy modeling methodologies at urban scale and offer insights for optimizing building energy efficiency strategies for a sustainable development of urban environments.