CVformer: Hybrid CEEMDAN-VMD Decomposition and Transformer for Coal Price Prediction

Aiming at the problem of insufficient prediction accuracy caused by the non-stationarity and multi-scale characteristics of coal price series, a hybrid prediction model combining CEEMDAN-VMD with Transformer, named CVformer, is proposed. Firstly, the original price series are decomposed by CEEMDAN to extract the multi-scale feature components. Secondly, the high-frequency components are decomposed by VMD to suppress modal aliasing and reconstruct the global and local feature matrices. Finally, the sliding-window attention mechanism is designed based on Transformer to dynamically integrate the multi-scale time-series dependencies. Experiments show that the model achieves superior performance on coal price datasets with multiple time scales, and its prediction accuracy and robustness significantly outperform the existing single- and mixed-models. Furthermore, the ablation experiment demonstrates that the multi-scale decomposition and dynamic attention fusion strategy effectively improve the ability to capture the long-term trend and short-term fluctuations in a synergistic manner, which verifies the effectiveness of the framework for the complex time-series prediction task.