A Hybrid SA-GA-BP Model with Feature Selection for Forecasting CO2 Emissions: Scenario Analysis for China’s Carbon Peak and Sustainable Development

Accurate forecasting of CO2 emissions is essential for evaluating emission-reduction pathways and supporting progress toward carbon-peaking and sustainable-development goals. Such forecasting is often challenging because macroeconomic datasets are small and exhibit strong nonlinear relationships. This study develops SA-GA-BP, a hybrid forecasting framework that integrates sensitivity analysis, Genetic Algorithm, and a Back Propagation neural network. A leave-one-variable-out sensitivity analysis uses changes in model performance to identify seven variables with high predictive importance and reduce input redundancy. GA then optimizes the BP network’s initial parameters to reduce the risk of local optima during small-sample training. Under the train–test split used in this study, the SA-GA-BP model yielded an R2 of 0.98766 and an RMSE of 2.6598, showing favorable numerical prediction performance among the compared models. In the low-carbon scenario, emissions peak in 2028 at 12.18 Gt CO2; under the high-carbon scenario, no peak occurs by 2030. The results suggest that accelerated energy-efficiency improvement, reduced fossil-fuel dependence, and continued industrial restructuring are important for achieving an earlier and lower carbon peak. The framework provides a practical approach to small-sample emission forecasting and supports scenario-based assessment of China’s carbon-peak pathway and sustainable development.