Carbon Emission Forecasting Using Multi-Scale Temporal Patches

Accurate carbon emission forecasting is crucial for achieving the “dual-carbon” goals and supporting effective emission-reduction strategies. However, carbon emissions are affected by multiple factors (e.g., industrial structure, energy consumption, and meteorological conditions), resulting in strong nonlinearity and time-varying dynamics. Traditional forecasting models often struggle to jointly capture long-term dependencies and local variations in time series. To address this challenge, this paper proposes a novel deep learning framework based on Multi-Scale Temporal Patches (MSTP). The framework segments long sequences into interrelated temporal patches to enable fine-grained representation and multi-scale feature extraction while maintaining cross-patch dependencies. The model combines a Mamba backbone with an enhanced Local Window Transformer (LWT) mechanism, supporting carbon emission forecasting for 48–720 future time steps. Experiments against baseline models, including the standard Transformer and its variants, show that the proposed method achieves an average MSE of 0.1288 across multiple horizons, yielding an approximately 50.4% relative reduction in MSE compared with the strongest baseline, Informer. Ablation studies further confirm the substantial contributions of MSTP, Mamba and LWT to prediction accuracy. Overall, this hybrid architecture improves forecasting performance and provides technical support for sustainable energy planning and real-time emission scheduling.