Carbon Price Point–Interval Forecasting Based on Two-Layer Decomposition and Deep Learning Combined Model Using Weight Assignment

Abstract

With the intensification of global warming, the demand for carbon emissions reduction has gradually increased in various countries. Carbon price is crucial for promoting the activation of the carbon trading market and facilitating emissions reduction. However, the current carbon price has non-linear characteristics, large fluctuations, and high complexity, making accurate predictions challenging. To effectively predict the trends and change of carbon price, this study proposed a hybrid deep learning point–interval prediction model. First, an improved variational mode decomposition–symplectic geometry mode decomposition (IVMD–SGMD) two-layer decomposition model was constructed to decompose the carbon prices into regular subsequences. Then, attention–temporal convolutional network–bidirectional gated recursive unit (Attention-TCN-BiGRU) and Encoder–Decoder long short-term memory (LSTM) combined prediction models were constructed for the prediction of subsequences. The entropy method (EM) was used to assign weights to the predictions of two models to achieve model complementarity and a linear reconstruction of the models’ results. Then the error correction was performed to obtain the final prediction results. This study conducted an experiment on carbon prices in the Guangdong and Shenzhen markets. The mean absolute error (MAE) of the two datasets were reduced by 89.69% and 87.43% lower than that for LSTM. To demonstrate the model's adaptability, prediction experiments conducted on natural gas and crude oil prices were employed, confirming its strong predictive accuracy in energy price forecasting. Based on the point prediction error, the interval prediction using the improved kernel density estimation (IKDE) provides more carbon market information for decision makers. The proposed model aids government energy policy formulation and fosters ongoing efforts to reduce carbon emissions.