Carbon emissions trading price forecasting based on temporal-spatial multidimensional collaborative attention network and segment imbalance regression

Accurate prediction of carbon emission trading prices is highly significant for developing low-carbon economies and constructing carbon markets. Making accurate predictions is challenging because the carbon price is highly nonlinear and non-stationary due to intricate influencing factors in multi-regions. Although current studies offer various solutions, there is still lack of in-depth consideration of the interaction in different regions. To address the problem, a novel hybrid forecasting framework (RTBSC) is proposed to predict carbon prices in different carbon markets, which combines reconstruction complete ensemble empirical mode decomposition module (RCMD), temporal-spatial multidimensional collaborative attention network (TSMA) and segment imbalance regression (SIR) method with crisscross optimization with chaotic map (CMCSO). First, to improve noise resistance and decomposition efficiency, CEEMDAN decomposes historical price to obtain Intrinsic Mode Function (IMF) in RCMD module. The complexity of IMFs is approximately classified by calculating the sample entropy, and similar IMFs are reconstructed by PCA to solve the problem of redundancy and the overfitting of carbon price subsequences. Then, a temporal-spatial multidimensional attention mechanism is proposed to capture the correlation between multiple regions. Multiple attention mechanisms are introduced to adaptively calculate input attention at different levels, capturing regional temporal-spatial relationships and feature importance. On this basis, aiming to extract more feature information by extracting deep temporal-spatial features with TSMA, Bidirectional Gate Recurrent Unit (BIGRU) is used to predict the carbon price. Thereafter, SIR-CMCSO is used to train the TSMA-BIGRU network and enhance the adaptive training ability. Multiple experiments are conducted, and the results demonstrate the model outperforms other advanced methods by comprehensively considering multi-regional data correlations, improving carbon emission trading price prediction accuracy, and providing a new approach for carbon emission trading price prediction.