Nonlinearity in the relationships between urban form and residential energy use intensity

The influence of urban form on building energy use is a crucial issue for energy-efficient urban planning and management. Nevertheless, traditional simulation and statistical methods often struggle to accurately represent and analyze this influence, which typically involves complex and nonlinear relationships. Conversely, while emerging machine learning studies excel in modeling such complexities, they often prioritize prediction over interpretation. This study seeks to deal with these limitations by employing gradient boosting decision trees (GBDT), along with interpretability analysis to reveal the nonlinear relationships between residential energy use intensity and various influential factors in Chicago, with a particular focus on urban form measures. Through a rigorous training procedure that accounts for the effects of randomness in hyperparameter tuning, the study develops and interprets final GBDT models using the feature importance and partial dependence plot (PDP) analyses. The results indicate that urban form factors have important predictive power for annual electricity use intensity, summer electricity use intensity, and winter gas use intensity. The PDPs reveal three distinct patterns in these factors: nonvisible, smooth, and complex nonlinear relationships. Notably, tree canopy coverage exhibits a smooth nonlinear relationship with annual electricity use intensity, whereas building density presents a complex nonlinear relationship. Additionally, the study develops traditional regression models using the same dataset and compares these with the GBDT models to highlight differences in how each approach assesses the importance of variables and complex relationships. Although machine learning models generally achieve higher accuracy, this study suggests they should primarily serve as exploratory data analysis rather than definitive tests of the relationships, given the approximative nature of the interpretability analysis. The findings from this study are useful in developing planning strategies and management policies to improve residential energy efficiency in urban areas.