A novel interval-valued carbon price prediction system based on deep learning ensemble and multi-source data integration

This paper presents an interval-valued carbon price prediction system based on multi-source data fusion, dual decomposition, and an ensemble of state-of-the-art deep learning techniques. It takes the historical carbon price data and a broad set of external driving factors as input for capturing the complex dynamics and interdependencies of carbon markets. This study focuses on two major Chinese carbon markets: using 1,357 daily interval samples from the Guangzhou market (December 1, 2014, to March 3, 2021) and 1,200 daily interval samples from the Hubei market (January 11, 2016, to January 21, 2021). The proposed dual decomposition method combines CEEMDAN and VMD optimized by IWOA to extract multi-frequency signal components, thereby preserving fine-grained information. Embracing three deep learning state-of-the-art models, each designed to capture different aspects of carbon price patterns ranging from local to long-term temporal dependencies, the prediction module contains PatchTST, DLinear, and iTransformer. This framework introduces a multi-layer perceptron to combine single-model outputs optimally for strong and precise interval-valued predictions. Extensive experiments on Guangzhou and Hubei carbon market datasets demonstrate the system's superiority over traditional benchmarks, achieving both higher predictive accuracy and improved computational efficiency. The system demonstrates strong generalizability to different market conditions and time periods, reflected in the consistent performance improvement on a set of evaluation metrics. The proposed system will be of practical value for the market participants, policymakers, and other stakeholders to support their informed decision-making processes concerning carbon trading strategies, risk management, and regulatory interventions. The modular architecture allows flexibility for incorporation of nascent carbon markets at a global scale.