Causal Heterogeneity of Freight Co₂ Emissions: Toward Advancing Prediction and Targeted Mitigation

Abstract

Identifying the causal determinants of freight vehicle emissions is essential for enhancing predictive performance and optimizing low-emission operational strategies. In this paper, instantaneous emission data were collected using PEMS in representative routes in Xi’an. A hybrid deep generative causal framework is proposed to quantify causal effects between CO₂ emissions and influencing variables. Additionally, a GRU-based causal-optimized prediction model (COPM) is developed to contextually fuse causal dependencies, while a learnable nonlinear functional decomposition is employed to redefine the hidden state. Experimental findings identify speed and engine speed as the dominant causal determinants among ten key variables. Moreover, all examined factors exhibit causal heterogeneity, each possessing an operating interval with a relatively minor impact. COPM achieves superior predictive accuracy, reducing RMSE by 18.76% and MAE by 25.95% over baseline models. This framework supports optimizing transportation system management and dynamic vehicle control by maintaining operations within low-impact regions, enhancing efficiency, and reducing emissions.

Keywords: Deep learning, Causal heterogeneity, Carbon emission prediction, PEMS, Freight vehicle