Quantifying Weather’s Share in Dynamic Grid Emission Factors via SHAP: A Multi-Timescale Attribution Framework

Accurately quantifying the impact of weather on dynamic grid carbon intensity is crucial for power system decarbonization. This study proposes a novel, interpretable machine learning framework integrating tree-based models with SHapley Additive exPlanations (SHAP) to quantify this impact across multiple timescales via a standardized “Weather Share” metric. Applied to city-level hourly data from China, the analysis reveals that meteorological variables collectively explain 21.64% of the hourly variation in carbon intensity, with air temperature and solar irradiance being the dominant drivers. Significant temporal variations are observed: the weather share is higher in summer (29.8%) and winter (23.5%) than in transition seasons and increases markedly to 32.7% during extreme high-temperature events. The proposed framework provides a robust, quantitative tool for grid operators, offering actionable insights for weather-aware carbon reduction strategies and highlighting critical time windows for targeted interventions.