Abstract:

Accurate forecasting of short-term spot electricity and carbon prices is crucial for decision-making in electricity market trading, carbon market operation, and their coordinated management. However, electricity and carbon price series are influenced by multiple complex factors, including energy structure, policy regulation, and renewable energy fluctuations, exhibiting high volatility and nonlinearity, which poses significant challenges to forecasting accuracy. Therefore, this paper proposes a dual-attention temporal convolutional network, bidirectional gated recurrent network (DA-TCN-BiGRU) short-term electricity-carbon price coupling forecasting method considering time-series fluctuation information mining. First, the central collision optimization-based variational mode decomposition algorithm is used to decompose the electricity and carbon price series into multi-frequency subsequences, so as to fully extract their fluctuation modes at different time scales. Second, the correlation strength between each feature in the high-dimensional feature set and electricity-carbon prices is evaluated based on the maximal information coefficient, and key features are selected. On this basis, a dual-attention TCN-BiGRU deep learning model is constructed to forecast carbon prices, and the predicted carbon price values are further input into the same framework as key exogenous variables to predict electricity prices, achieving progressive coupling forecasting of electricity-carbon prices. Finally, the case study based on actual data from the Hubei Province electricity-carbon market shows that the proposed method has higher accuracy and stronger stability in electricity-carbon price prediction, verifying the effectiveness and superiority of the model.

Key words: electricity-carbon price, optimized variational mode decomposition, maximal information coefficient, dual attention mechanism, temporal convolutional network, bidirectional gated recurrent network