Scholarlink

<p>Carbon emission reduction in China’s transportation sector is crucial for achieving the dual-carbon targets. However, existing research on carbon emission prediction often neglects the role of unstructured data and fails to capture the complex dynamic characteristics of carbon emissions. To address these limitations, this study proposes a hybrid forecasting model based on multimodal data fusion and multi-scale decomposition. First, a multimodal dataset is constructed by integrating structured data with unstructured data quantified through policy text analysis. Second, the multivariate variational mode decomposition method is employed to decompose the multimodal data, with fuzzy entropy used to assess complexity. Subsequently, different scenarios are established based on data types, decomposition methods, and prediction models to forecast carbon emissions, with evaluation metrics applied to assess the results. Finally, baseline, pessimistic, and optimistic scenarios are simulated to project transportation carbon emissions from 2023 to 2050. The results demonstrate that: (1) Incorporating policy number, effectiveness, and sentiment enhances prediction accuracy with cumulative effects. Under the S5, S10, and S15 scenarios, the CNN-GRU-Attention model reduces RMSE (MAPE) by 2.756% (2.749%), 2.180% (6.131%), and 2.650% (47.309%), respectively. (2) The MVMD decomposition method significantly improves prediction accuracy. Under the S6-S10 scenarios, the CNN-GRU-Attention model reduces RMSE (MAPE) by 2.423% (0.451%), 2.121% (0.730%), 2.218% (2.652%), 1.638% (-1.393%), and 1.846% (3.912%), respectively. (3) The hybrid model demonstrates superior forecasting performance. Under the S11-S15 scenarios, the CNN-GRU-Attention model reduces RMSE (MAPE) by 0.886% (12.475%), 1.367% (9.942%), 0.981% (1.673%), 1.210% (3.501%), and 1.362% (50.871%), respectively. This study not only provides a high-precision carbon emission forecasting framework but also provides a scientific basis for formulating differentiated emission reduction strategies.</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

Hybrid forecasting and multiscenario analysis of carbon emissions from transportation sector based on multimodal data fusion and multiscale decomposition

  • 期刊论文

  • 发表日期：

    2025年10月20日

查看全文
  • Xizhen Xu,
  • Yuming Liu,
  • Guoliang Ou

摘要

Carbon emission reduction in China’s transportation sector is crucial for achieving the dual-carbon targets. However, existing research on carbon emission prediction often neglects the role of unstructured data and fails to capture the complex dynamic characteristics of carbon emissions. To address these limitations, this study proposes a hybrid forecasting model based on multimodal data fusion and multi-scale decomposition. First, a multimodal dataset is constructed by integrating structured data with unstructured data quantified through policy text analysis. Second, the multivariate variational mode decomposition method is employed to decompose the multimodal data, with fuzzy entropy used to assess complexity. Subsequently, different scenarios are established based on data types, decomposition methods, and prediction models to forecast carbon emissions, with evaluation metrics applied to assess the results. Finally, baseline, pessimistic, and optimistic scenarios are simulated to project transportation carbon emissions from 2023 to 2050. The results demonstrate that: (1) Incorporating policy number, effectiveness, and sentiment enhances prediction accuracy with cumulative effects. Under the S5, S10, and S15 scenarios, the CNN-GRU-Attention model reduces RMSE (MAPE) by 2.756% (2.749%), 2.180% (6.131%), and 2.650% (47.309%), respectively. (2) The MVMD decomposition method significantly improves prediction accuracy. Under the S6-S10 scenarios, the CNN-GRU-Attention model reduces RMSE (MAPE) by 2.423% (0.451%), 2.121% (0.730%), 2.218% (2.652%), 1.638% (-1.393%), and 1.846% (3.912%), respectively. (3) The hybrid model demonstrates superior forecasting performance. Under the S11-S15 scenarios, the CNN-GRU-Attention model reduces RMSE (MAPE) by 0.886% (12.475%), 1.367% (9.942%), 0.981% (1.673%), 1.210% (3.501%), and 1.362% (50.871%), respectively. This study not only provides a high-precision carbon emission forecasting framework but also provides a scientific basis for formulating differentiated emission reduction strategies.