Wang Jingyi, Chang Chunyi
Under the "dual carbon" goals, the low-carbon transformation of the transportation sector is urgently required. To accurately assess the future trends of carbon emissions from Beijing's transportation industry and support the formulation of emission reduction policies, this paper aims to construct a highly accurate and robust carbon emission prediction model.Firstly, a "top-down" emission factor method was employed to calculate the CO2 emissions from Beijing's transportation industry from 1997 to 2023. Secondly, 25 influencing factors were screened as input features through a literature review. Thirdly, the prediction performance of four individual models—BP neural network, GRU, LSTM, and Transformer—was compared. Finally, a combined model integrating temporal feature extraction (LSTM+Attention) and static feature extraction (BP) was constructed and its effectiveness was validated.The results indicate that among the individual models, BP and LSTM demonstrated superior performance. The constructed combined model outperformed all individual models across the three metrics of MAPE, MAE, and RMSE. Specifically, its MAPE was 12.18% lower than that of the best individual model, significantly enhancing prediction accuracy and stability. This combined model effectively integrates temporal and non-temporal features, making it suitable for scenario-based forecasting of future carbon emissions from Beijing's transportation industry and capable of providing a scientific basis for relevant authorities to formulate differentiated emission reduction pathways.
Transportation industry; Carbon emissions; Deep learning