Scholarlink

<p>Accurate carbon price forecasting is important for informed policy-making and investment decisions in the expanding carbon trading market. This study proposes an improved deep learning model, TCN-LSTM-Self-Attention, that integrates decomposition and error correction to enhance predictive precision. We first decompose the close price and error series into subsequences, which are then separately predicted using the TCN-LSTM-Self-Attention framework. Next, the initial predictions are refined through error predictions, resulting in final corrected forecasts. Empirical results demonstrate notable accuracy improvements compared to traditional methods, achieving a coefficient of determination of 0.982, a root mean square error of 0.646, and a mean absolute percentage error of 0.777%. These findings demonstrate the model’s ability to capture complex short-term and long-term dynamics of carbon price fluctuations. Validation across Shenzhen, Guangdong, and Hubei carbon trading pilots further indicates its robustness and applicability. We recommend this integrated model as a useful tool for policymakers and market participants seeking accurate and stable carbon price forecasts.</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 carbon price prediction model based on decomposition and dual-channel attention network

  • 期刊论文

  • 发表日期：

    2025年6月14日

查看全文
  • Zhonglin Ma,
  • Chao Wang,
  • Hong Qi,
  • Jacob Wood

摘要

Accurate carbon price forecasting is important for informed policy-making and investment decisions in the expanding carbon trading market. This study proposes an improved deep learning model, TCN-LSTM-Self-Attention, that integrates decomposition and error correction to enhance predictive precision. We first decompose the close price and error series into subsequences, which are then separately predicted using the TCN-LSTM-Self-Attention framework. Next, the initial predictions are refined through error predictions, resulting in final corrected forecasts. Empirical results demonstrate notable accuracy improvements compared to traditional methods, achieving a coefficient of determination of 0.982, a root mean square error of 0.646, and a mean absolute percentage error of 0.777%. These findings demonstrate the model’s ability to capture complex short-term and long-term dynamics of carbon price fluctuations. Validation across Shenzhen, Guangdong, and Hubei carbon trading pilots further indicates its robustness and applicability. We recommend this integrated model as a useful tool for policymakers and market participants seeking accurate and stable carbon price forecasts.