A Hybrid Time-Series Simulation Framework for Provincial Carbon Emissions Using Multi-Factor Decomposition and Deep Learning

Accurate time-series simulation of carbon emissions for both the whole society and the electricity industry is pivotal for realizing China’s “Dual Carbon” goals. This research constructs a hybrid simulation architecture integrating factor decomposition with deep learning to quantify emission trajectories for both the whole society and the electricity industry in Anhui Province. First, the extended Kaya identity and Logarithmic Mean Divisia Index (LMDI) are employed to analyze socioeconomic drivers. The decomposition analysis indicates that per capita income is the primary driver of carbon emissions, whereas energy intensity exerts the strongest inhibitory effect. Subsequently, Variational Mode Decomposition (VMD) is applied to the nonstationary emission series to produce multi-scale sub-signals, which are then fed into a predictive model comprising a Bayesian-optimized (BO) Transformer coupled with Long Short-Term Memory (LSTM) networks. The study establishes three distinct evolution scenarios: Moderate Sustainability (MS), Business as Usual (BAU), and Strong Economic Growth (SEG). Simulation results indicate that under MS, carbon emissions from the whole society and the electricity industry peak in 2029 at 435.2 Mt and 2030 at 281.2 Mt, respectively. Conversely, the SEG scenario delays the peak of the whole society to 2034, while the electricity industry fails to peak before 2035. These findings reveal significant risks of temporal asynchrony between the whole society and the electricity industry peaks, providing robust methodological support for regional decarbonization planning.