Predicting the operational carbon emissions of urban community service units by using a two-stage explainable machine learning method: a case study in Nanjing, China

Low-carbon community development is fundamental to urban sustainability. With accelerating urbanization, the growing demand for community services which occur in non-residential units (e.g. catering, shopping, education, and healthcare) increasingly contributes to community-level carbon emissions, yet remains largely unexplored. This study proposes a two-stage explainable machine learning method to predict time-series carbon emissions from service units, integrating ensemble learning algorithms, Bayesian optimization, Shapley Additive exPlanations (SHAP), and a recursive strategy. A case study in China utilizes emissions data compiling from 2013 to 2024 with community-level features. Results show that the models achieve satisfactory performance, with R2 of 0.98 in both stages. Floor area and service type are the dominant features contributing to units’ carbon emissions. Furthermore, under varying feature values (e.g. service type, population, and housing price), we predict future emission trajectories in three scenarios: baseline, development, and decline. Future emissions will inevitably rise with community evolution, suggesting the need for addressing shifts of service structure and tailored energy management in carbon-intensive units. By advancing carbon emission prediction methods, this study offers data-driven insights on emission trajectories with key features of evolving communities, informing low-carbon management for sustainable community development.