Introduction Accurately forecasting carbon emissions is essential for China’s carbon neutrality goals, yet the country’s vast disparities in economic development and energy structures create complex spatiotemporal heterogeneity that traditional centralized models often fail to capture. Methods To address this challenge, we developed a personalized federated learning (pFL) framework based on Long Short-Term Memory networks with an adaptive sparse attention mechanism (LSTM-ASA). We implemented a geography-based partitioning strategy that divides the nation into five macro-regions (e.g., Eastern, Northwestern) and applied a performance-weighted aggregation strategy to optimize provincial-level interval forecasting and uncertainty quantification. Results Experimental results using provincial carbon emission data from 2021 to 2025 demonstrate that the pFL framework consistently outperforms centralized baselines. Specifically, the proposed method achieved a reduction in MAE ranging from 2.68% (Eastern) to 17.91% (Northwestern) and an improvement in of R 2 up to 8.90% (Southwestern). Furthermore, the framework maintained high interval reliability with a PICP consistently exceeding 96%, effectively addressing regional diversity and spatiotemporal heterogeneity. Discussion These findings validate the robustness and adaptability of integrating regional partitioning with federated learning for environmental modeling. The study offers a novel technical foundation for policymakers to formulate differentiated, region-specific carbon reduction strategies.
Xiao et al. (2026) studied this question.