Temporal Impact Analysis of Carbon Emissions in China’s Power Sector Using N-BEATS Forecasting and TimeSHAP Interpretation

China’s power sector plays a pivotal role in national carbon emissions, making accurate forecasting essential for effective energy management and policy-making. In this study, we present a deep learning approach to forecast daily carbon emissions from January 1, 2019, to December 31, 2024, using the N-BEATS model. To address the challenge of interpretability inherent in deep learning models, we integrate TimeSHAP—a time series adaptation of SHAP values—to quantify the temporal influence on the forecast. This method not only reveals the overall behavior of the forecasting model but also highlights the relative importance of specific time periods (e.g., months) in driving future predictions. Our findings demonstrate that even in the absence of additional explanatory variables, the combination of N-BEATS forecasting and TimeSHAP interpretability can uncover intrinsic temporal patterns, thereby providing valuable insights for targeted energy management strategies.