Improving Carbon Emission Prediction with Multiscale Decomposition and Transformer Networks

With the intensification of global climate change, carbon emissions have become a critical challenge in achieving sustainable development. Accurate carbon emission forecasting plays a key role in formulating carbon neutrality policies and optimizing energy structures. However, current mainstream forecasting methods often face limitations in handling complex nonlinear time series data, including insufficient modeling capabilities, susceptibility to mode mixing, and low prediction accuracy. To address these issues, this paper proposes a carbon emission forecasting model that combines Variational Mode Decomposition (VMD) with an optimized Transformer-ResLSTM structure. Specifically, VMD is first employed to decompose the original time series into a set of Intrinsic Mode Functions (IMFs) with distinct frequency characteristics, effectively mitigating mode mixing issues. Then, a specially designed Transformer-ResLSTM network is used to model each IMF individually, extracting critical information from both global and local features. Finally, the prediction values are generated through a reconstruction and error correction module, improving both prediction accuracy and robustness. Experiments conducted on the CARMA and GEFCom datasets demonstrate that the proposed model outperforms multiple baseline models across all evaluation metrics, such as RMSE, MAE, MAPE, and R2, highlighting its superiority and practicality in carbon emission forecasting tasks. This model provides an effective solution to the challenges of nonlinear modeling in carbon emission forecasting and offers robust technical support for achieving carbon neutrality goals.