Data-Driven Identification and Neural Projection of Carbon Emission Drivers in Sichuan Province for Sustainable Development: A Coupled PLSR–RNN Framework

In the context of global climate mitigation and China’s “Dual Carbon” targets, accurately identifying emission drivers and forecasting long-term trajectories at the provincial level has become a critical research priority for designing sustainable transition pathways. This study investigates carbon emissions in Sichuan Province over the period 1997–2022 using an integrated analytical workflow that sequentially couples Pareto-based feature selection and robust partial least squares regression (PLSR) for explanatory driver identification, with recurrent neural network (RNN) modeling for independent trend projection. Pareto filtering is employed to isolate 34 high-impact variables that collectively explain 85% of total emissions. To address numerical instability in the explanatory phase, a log-transformed robust PLSR is implemented. Separately, to extrapolate near-term trajectories, a dynamically switched long short-term memory (LSTM) and gated recurrent unit (GRU) architecture is utilized under stability assumptions, operating within a data-driven framework with empirically calibrated physical constraints. This integrated modeling approach demonstrates strong explanatory performance (training R2 = 0.9237, cross-validation Q2 = 0.7241) and yields a mean absolute percentage error (MAPE) of 3.11–3.53% on a chronologically separated five-year validation period (2018–2022). This low error—achieved over a holdout interval encompassing the Double Carbon policy announcement and COVID-19 disruptions—provides preliminary evidence of the model’s capacity to capture the dominant inter-annual emission dynamics for short-horizon projection, despite the inherent constraint of limited annual observations (N = 26). Among the identified drivers, the utilization of by-product gases in heavy industry and gasoline consumption in urban transportation exhibit the strongest positive influence, with variable importance in projection (VIP) scores exceeding 1.5. Conversely, coal use in urban residential sectors exerts a modest negative effect on emission growth (elasticity = −0.023). Under empirically bounded stability assumptions, the constrained baseline projection suggests that total emissions may edge down to 259.84 Mt by 2027. However, this apparent trajectory, is strictly conditioned by the imposed year-to-year change limits. It must be interpreted as a restricted scenario outcome rather than an independent predictive discovery of an approaching carbon-peaking plateau. While the explanatory fit of the PLSR component does not mathematically validate the subsequent neural-network forecasts, this research offers a practical methodological reference for the application-specific integration of established statistical and deep-learning tools in regional emission diagnosis.