Carbon Emission Prediction for Sichuan Province Using a SHAP-Explained Machine Learning Framework

This study develops a provincial carbon accounting and forecasting framework for Sichuan Province covering 2005–2022 under a consumption-based boundary. Direct fossil fuel emissions and indirect electricity-related emissions are clearly distinguished to ensure accounting consistency and additive closure. On this basis, a Transformer-based machine learning model integrated with SHAP is constructed to predict carbon emissions and identify peak characteristics. The results show high historical fitting accuracy and project that Sichuan’s carbon emissions will reach a plateau-type peak around 2031, followed by a gradual decline. SHAP-based interpretation indicates that total energy consumption and the carbon emission factor are the dominant drivers, while economic scale and energy intensity exert secondary but significant effects. Robustness tests confirm the stability of the explanatory structure. The proposed framework integrates accounting consistency, predictive modeling, and interpretability analysis, providing empirical support for regional peak management and low-carbon transition planning.