Study on the characteristics and prediction of concrete carbon emissions based on a machine learning approach with spatiotemporal heterogeneity analysis

Although concrete remains indispensable in construction for its compressive strength and durability, its production constitutes a predominant source of building-sector carbon emissions. Current emission accounting methods frequently neglect fine-grained spatiotemporal variations, resulting in prediction inaccuracies that hinder effective decarbonization policy formulation. This study addresses this gap by elucidating how regional and seasonal factors modulate concrete emissions and establishing a high-precision dynamic prediction model. Focusing on Shandong Province, China’s northern hub for construction material production, we quantified disparities in mix proportions, raw material logistics (0–130 km), and production energy consumption intensity across five subregions (East, West, South, Central, and North Shandong). We conducted a systematic comparison of four mainstream machine learning algorithms: Random Forest (RF), Gradient Boosting Decision Tree (GBDT), XGBoost, and Support Vector Machine (SVM). The GBDT model achieved superior performance (R² = 0.958, RMSE = 20.05 kgCO₂), outperforming alternatives by 12–23 % in accuracy. SHAP analysis revealed cement content, aggregate type, fly ash, and transport distance as dominant predictors. Our findings demonstrate that: (1) Emission heterogeneity across subregions reaches 28.6 % due to localized supply chains and seasonal energy mixes. (2) GBDT’s ensemble learning effectively captures nonlinear interactions in material–process–emission relationships. This work advances low-carbon concrete strategies by providing a spatiotemporally adaptive prediction tool, with implications for lifecycle carbon management in construction. The methodology is transferable to other emission-intensive regions, supporting low-carbon goals through data-driven industrial optimization.