<p>Given the significance of cities as key execution units in carbon reduction initiatives, their role is crucial for the building sector and for China to achieve its “carbon peak and carbon neutrality” goals. Based on historical data on carbon emissions in the municipal-level building sector, this study employs spatiotemporal analysis models—such as Moran’s Index, kernel density estimation, and center of gravity migration analysis—to capture the spatial heterogeneity and dynamic evolution of carbon emissions from 2015 to 2020. By integrating an extended Kaya identity with the <i>k</i>-means clustering algorithm and embedding a machine learning algorithm with optimal test performance, the study identifies the heterogeneity in development types and influencing factors among different cities. The results indicate that carbon emissions in the municipal-level building sector across Chinese cities increasingly concentrate on higher emission ranges. North China serves as the high-value center for carbon emissions, whereas Southwest China is the low-value center. Analysis of center of gravity migration trajectories reveals that factors such as energy consumption per unit of GDP significantly drive the growth of carbon emissions in the building sector, while the general public budget expenditure intensity exerts an inhibitory effect. Additionally, Chinese cities are classified into five distinct development types. Compared to traditional nonlinear models and other machine learning models, XGBoost achieves the best predictive performance for carbon emissions across of buildings different city types, with <i>R</i>² values of 0.998, 0.992, 0.983, 0.995, and 0.986. These findings underscore the need to develop differentiated carbon reduction strategies that align with the unique development characteristics and influencing factors of each city type.</p>