Abstract
Accurate carbon price prediction can help the government establish an effective and stable carbon trading market mechanism, which researchers are increasingly focusing on. However, much research on carbon price prediction has ignored the impacts of multiple factors on the carbon price. A novel ensemble deep learning prediction model, termed CEEMDAN-Attention-RNN, which considers multiple influencing factors, has been proposed to improve the accuracy of carbon price prediction. Firstly, raw data such as carbon price and external variables are decomposed by complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) into multiple sub-components with different frequencies. Then, the recurrent neural network (RNN) enhanced by LSTM and GRU is combined with the attention mechanism to form a prediction model. Finally, several evaluation indicators are used to obtain the final prediction accuracy, and the model is applied to 3 pilot areas of carbon trading in the Yangtze River basin. The results indicate that the mean absolute percentage errors of the proposed model are 1.8872%, 1.5686%, and 5.2548% in Shanghai Municipality, Hubei Province, and Chongqing Municipality, respectively, and its forecasting ability is better than that of other carbon price forecasting models. Therefore, the proposed model is an excellent method for carbon trading price prediction due to its high accuracy. In addition, high-precision carbon trading price forecasting technology is of great significance for the government to formulate emission reduction policies.