Scholarlink

<p>Currently, Carbon emission forecasting faces two issues: capturing local abrupt changes and global trends simultaneously, and aligning model forecasting performance with regional economic development levels. To address these, we propose a hybrid forecasting architecture based on dimensionality transformation and deep cascading. To evaluate the rationality of the proposed architecture, we construct a spatial correlation network using daily carbon emission data spanning 30 Chinese provinces from 2019 to 2024. Extensive comparative experiments demonstrate that our model consistently and significantly outperforms all competing baselines across multiple evaluation criteria, including RMSE, R², MASE, MSE and SMAPE. The proposed architecture generally achieved an average reduction of approximately 10% in SMAPE across the 30 provinces, with a maximum reduction of about 27%, and the Diebold-Mariano test confirmed that the performance improvements in most provinces are statistically significant at the 1% or 5% level. These results provide robust evidence of its effectiveness and superiority. In terms of computational efficiency, the architecture attains an average training time of approximately 3&#xa0;min per province, demonstrating its practical viability. The indispensability of each constituent module and the rationality of the designated cascading sequence are rigorously corroborated through ablation experiments and computational efficiency evaluations. To address the notable regional heterogeneity in model improvement effectiveness, this paper incorporates a four-quadrant analytical approach alongside regional economic development levels to establish a differentiated governance framework. The robustness and applicability of this tiered strategy are further validated through sensitivity analysis based on multi-year GDP data. Overall, this study not only provides a rigorous and efficient technical pathway for high-precision carbon emission forecasting, but also offers scientific decision support for formulating emission reduction policies that balance economic feasibility with implementation efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Framework Based on Transformer for Carbon Emission Forecasting and Differentiated Decision Support

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Suli Cheng,
  • Xuanlong Lu,
  • Jie Li

摘要

Currently, Carbon emission forecasting faces two issues: capturing local abrupt changes and global trends simultaneously, and aligning model forecasting performance with regional economic development levels. To address these, we propose a hybrid forecasting architecture based on dimensionality transformation and deep cascading. To evaluate the rationality of the proposed architecture, we construct a spatial correlation network using daily carbon emission data spanning 30 Chinese provinces from 2019 to 2024. Extensive comparative experiments demonstrate that our model consistently and significantly outperforms all competing baselines across multiple evaluation criteria, including RMSE, R², MASE, MSE and SMAPE. The proposed architecture generally achieved an average reduction of approximately 10% in SMAPE across the 30 provinces, with a maximum reduction of about 27%, and the Diebold-Mariano test confirmed that the performance improvements in most provinces are statistically significant at the 1% or 5% level. These results provide robust evidence of its effectiveness and superiority. In terms of computational efficiency, the architecture attains an average training time of approximately 3 min per province, demonstrating its practical viability. The indispensability of each constituent module and the rationality of the designated cascading sequence are rigorously corroborated through ablation experiments and computational efficiency evaluations. To address the notable regional heterogeneity in model improvement effectiveness, this paper incorporates a four-quadrant analytical approach alongside regional economic development levels to establish a differentiated governance framework. The robustness and applicability of this tiered strategy are further validated through sensitivity analysis based on multi-year GDP data. Overall, this study not only provides a rigorous and efficient technical pathway for high-precision carbon emission forecasting, but also offers scientific decision support for formulating emission reduction policies that balance economic feasibility with implementation efficiency.