Tackling carbon price volatility: A synchronous joint-decomposition and multi-scale spatiotemporal feature learning for interval-valued forecasting

Precise carbon prices forecasting is important for the optimal allocation of resources within carbon markets. While research has shifted from point forecasting to interval forecasting, the non-linear, non-stationary, and dynamical interval-valued carbon price volatility continue to pose fundamental challenges. To resolve them, this article proposes an innovative hybrid architecture that integrates interval representation, synchronous joint decomposition, and multi-scale spatiotemporal feature learning, aiming at high-accuracy interval-valued carbon price prediction. The raw data are first transformed into midpoint and radius sequences, which yields a finer and more interpretable representation of uncertainty. The optimized successive variational mode decomposition (SVMD) method is then applied to jointly decompose and refactoring the dual-sequences. This synchronous decomposition mechanism keeps the physical consistency between the two components intact and blocks information leakage. Furthermore, the enhanced spatiotemporal Transformer integrating long short-term memory (LSTM) and convolutional neural network (CNN) is designed to simultaneously extract long-term dependencies, short-term variations, and local features from the decomposed sub-sequences, substantially boosting prediction accuracy. Empirical evaluations conducted on the Guangdong and Hubei carbon markets demonstrate the model's superior performance, achieving mean absolute percentage errors of 0.03661 and 0.02045, respectively. Additional comparative experiments and Z-score analysis confirm the accuracy and robustness of the framework.