Scholarlink

<p>Accurate prediction of carbon emissions is essential for enabling government agencies to set scientific emission reduction targets and formulate targeted green development strategies. To enhance the accuracy and reliability of carbon emission forecasting, this study proposes a deep learning framework based on multi-dimensional feature fusion. By leveraging feature enhancement techniques, a multi-dimensional feature space is constructed, integrating the local feature extraction capability of Convolutional Neural Network(CNN) with the temporal dependency modeling strength of Bidirectional Long Short-Term Memory (BiLSTM). This enables collaborative spatio-temporal feature extraction, while an attention mechanism dynamically allocates weights to key time-series nodes. Using China’s daily carbon emission data from 2019 to 2025 as the research subject, a multi-dimensional feature analysis framework is established. Through feature engineering, three types of variables, temporal, statistical, and time-series features, are systematically extracted to preliminarily explore emission fluctuation patterns and periodic trends. The integrated feature set is then fed into the hybrid model, and feature perturbation analysis is applied to quantify the contribution of each feature. Experimental results highlight the central role of time-series differences, rolling statistics, and lag features in carbon emission prediction. To validate the model’s effectiveness, a comparative experiment is designed using CNN, BiLSTM, and BiLSTM-Attention as benchmark models under identical test conditions. The proposed model achieves a Root Mean Square Error of 0.865 on the test set, representing an average reduction of 30% compared to baseline models, confirming its superior performance. Based on the findings, strategic recommendations are offered to guide optimized carbon emission control.</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 hybrid CNN-BiLSTM-attention model with feature engineering for accurate carbon emission forecasting

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Shuixiang Yu,
  • Yaozong Tang,
  • Xinzhou Xu,
  • Tongxuan Liu

摘要

Accurate prediction of carbon emissions is essential for enabling government agencies to set scientific emission reduction targets and formulate targeted green development strategies. To enhance the accuracy and reliability of carbon emission forecasting, this study proposes a deep learning framework based on multi-dimensional feature fusion. By leveraging feature enhancement techniques, a multi-dimensional feature space is constructed, integrating the local feature extraction capability of Convolutional Neural Network(CNN) with the temporal dependency modeling strength of Bidirectional Long Short-Term Memory (BiLSTM). This enables collaborative spatio-temporal feature extraction, while an attention mechanism dynamically allocates weights to key time-series nodes. Using China’s daily carbon emission data from 2019 to 2025 as the research subject, a multi-dimensional feature analysis framework is established. Through feature engineering, three types of variables, temporal, statistical, and time-series features, are systematically extracted to preliminarily explore emission fluctuation patterns and periodic trends. The integrated feature set is then fed into the hybrid model, and feature perturbation analysis is applied to quantify the contribution of each feature. Experimental results highlight the central role of time-series differences, rolling statistics, and lag features in carbon emission prediction. To validate the model’s effectiveness, a comparative experiment is designed using CNN, BiLSTM, and BiLSTM-Attention as benchmark models under identical test conditions. The proposed model achieves a Root Mean Square Error of 0.865 on the test set, representing an average reduction of 30% compared to baseline models, confirming its superior performance. Based on the findings, strategic recommendations are offered to guide optimized carbon emission control.