A VMD-TCN-GRU Hybrid Model for Multi-Step Forecasting of Industrial Carbon Emissions in China

Accurate forecasting of carbon emissions is critical for the design and implementation of effective climate policies. In this study, we propose a deep learning model—VMD-TCN-GRU—to predict daily industrial carbon emission rates in China. The model integrates Variational Mode Decomposition for noise reduction and signal smoothing, Temporal Convolutional Networks for multi-scale temporal feature extraction, and Gated Recurrent Units for capturing long-term dependencies. Using daily carbon emission data from the Carbon Monitor platform between January 1, 2023 and December 31, 2024, the model was evaluated under 1-day, 3-day, and 5-day forecasting scenarios. Experimental results demonstrate that the proposed VMD-TCNGRU model consistently outperforms baseline models including BILSTM, BILSTM with self-attention, CNN-BILSTM, LightGBM, and Transformer in terms of MSE, MAE, and MAPE. These findings validate the effectiveness and robustness of the proposed model for short- and medium-term carbon emission forecasting.