Scholarlink

<p>Against the backdrop of the steady advancement of the dual-carbon strategy, accurate assessment of regional carbon emission trends is a crucial prerequisite for formulating scientific emission reduction policies. Current carbon emission prediction studies mostly employ a globally unified modeling approach, generally neglecting the emission heterogeneity existing within provincial regions. Furthermore, they struggle to simultaneously address practical issues such as multicollinearity of driving variables, limited short-term sample sizes, and nonlinear data characteristics, leaving significant room for improvement in the overall prediction framework. To address these shortcomings and practical modeling challenges, this study constructs an integrated hybrid analysis and prediction framework. First, an improved K-means clustering method, incorporating maximum-minimum distance and silhouette coefficient optimization, is used to delineate provincial emission regions, distinguishing the emission development characteristics of different provinces. Second, a prediction model combining grey ridge regression and backpropagation neural networks is constructed for various cluster subsets. Grey ridge regression is used to process small sample data and resolve multicollinearity issues, while neural networks learn prediction residuals to uncover inherent nonlinear relationships within the data. Empirical results show that the improved clustering algorithm achieves a silhouette coefficient of 0.68, and related clustering evaluation indicators are significantly improved compared to traditional algorithms. The combined prediction model achieves a coefficient of determination of 0.92, a mean absolute percentage error of 6.8%, and a root mean square error of 0.29. Even under strong noise interference, the model maintains a high predictive level, and the rolling prediction results from 2019 to 2022 show stable fluctuations, demonstrating good robustness and stability overall. The method proposed in this study can effectively distinguish regional carbon emission differences and accurately predict time-series emission data. It can provide reliable data and technical support for various regions to identify emission reduction pathways and implement differentiated low-carbon management measures.</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

Carbon emission prediction framework based on improved K-means and grey ridge regression BP neural network hybrid model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • Tianhai Lu

摘要

Against the backdrop of the steady advancement of the dual-carbon strategy, accurate assessment of regional carbon emission trends is a crucial prerequisite for formulating scientific emission reduction policies. Current carbon emission prediction studies mostly employ a globally unified modeling approach, generally neglecting the emission heterogeneity existing within provincial regions. Furthermore, they struggle to simultaneously address practical issues such as multicollinearity of driving variables, limited short-term sample sizes, and nonlinear data characteristics, leaving significant room for improvement in the overall prediction framework. To address these shortcomings and practical modeling challenges, this study constructs an integrated hybrid analysis and prediction framework. First, an improved K-means clustering method, incorporating maximum-minimum distance and silhouette coefficient optimization, is used to delineate provincial emission regions, distinguishing the emission development characteristics of different provinces. Second, a prediction model combining grey ridge regression and backpropagation neural networks is constructed for various cluster subsets. Grey ridge regression is used to process small sample data and resolve multicollinearity issues, while neural networks learn prediction residuals to uncover inherent nonlinear relationships within the data. Empirical results show that the improved clustering algorithm achieves a silhouette coefficient of 0.68, and related clustering evaluation indicators are significantly improved compared to traditional algorithms. The combined prediction model achieves a coefficient of determination of 0.92, a mean absolute percentage error of 6.8%, and a root mean square error of 0.29. Even under strong noise interference, the model maintains a high predictive level, and the rolling prediction results from 2019 to 2022 show stable fluctuations, demonstrating good robustness and stability overall. The method proposed in this study can effectively distinguish regional carbon emission differences and accurately predict time-series emission data. It can provide reliable data and technical support for various regions to identify emission reduction pathways and implement differentiated low-carbon management measures.