Scholarlink

<p>Under the increasingly severe global climate challenges, accurate carbon dioxide emission price prediction has become a critical tool for optimizing resource allocation and accelerating low-carbon transitions. Prevailing forecasting approaches, which predominantly rely on point estimates or intervals constructed from single-valued data, often fail to adequately model price dynamics influenced by a complex interplay of multidimensional factors, thereby constraining their real-world utility. This study introduces a novel forecasting framework designed specifically for interval-valued carbon prices, which synthesizes advanced multi-factor feature engineering, clustering-based ensemble learning, and a refined Transformer model. The methodology begins by transforming daily high-low price ranges into complementary center and radius sequences, thereby encapsulating essential intraday volatility information. Next, a multidimensional factor screening strategy based on a two-stage hierarchical architecture is applied to identify optimal feature factors, followed by hierarchical clustering to partition heterogeneous feature subsets. For each cluster, an enhanced Transformer sub-model is trained to synergize local feature extraction and global dependency modeling. Finally, two-stage nonlinear integration algorithm aggregates predictions across subsets to generate final results. An illustrative empirical analysis of two carbon trading markets in China is conducted to show that the presented model can effectively capture the intrinsic features of the series and improve the accuracy of the forecasting results.</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 multi-factor clustering integration interval-valued carbon dioxide emission price prediction based on two-stage feature engineering and improved transformer

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Qian Li,
  • Jujie Wang,
  • Yuxuan Lu,
  • Qiuzi Lei

摘要

Under the increasingly severe global climate challenges, accurate carbon dioxide emission price prediction has become a critical tool for optimizing resource allocation and accelerating low-carbon transitions. Prevailing forecasting approaches, which predominantly rely on point estimates or intervals constructed from single-valued data, often fail to adequately model price dynamics influenced by a complex interplay of multidimensional factors, thereby constraining their real-world utility. This study introduces a novel forecasting framework designed specifically for interval-valued carbon prices, which synthesizes advanced multi-factor feature engineering, clustering-based ensemble learning, and a refined Transformer model. The methodology begins by transforming daily high-low price ranges into complementary center and radius sequences, thereby encapsulating essential intraday volatility information. Next, a multidimensional factor screening strategy based on a two-stage hierarchical architecture is applied to identify optimal feature factors, followed by hierarchical clustering to partition heterogeneous feature subsets. For each cluster, an enhanced Transformer sub-model is trained to synergize local feature extraction and global dependency modeling. Finally, two-stage nonlinear integration algorithm aggregates predictions across subsets to generate final results. An illustrative empirical analysis of two carbon trading markets in China is conducted to show that the presented model can effectively capture the intrinsic features of the series and improve the accuracy of the forecasting results.