Predicting PM2.5 Concentrations Based on Vehicle Ownership Using Big Data Models and Machine Learning

Abstract

This study integrates multi-source satellite-remote-sensed PM2.5 data and provincial vehicle ownership data in China (1998-2023) to construct a high-precision prediction model. We developed a spatiotemporally coupled hybrid model that fuses time series analysis (Prophet) with graph convolutional networks, and incorporates a reinforcement learning framework for dynamic feature optimization. The model achieves an R² of 0.95, with a mean absolute error of 1.97 g/m³ and a root mean square error of 2.67 g/m³ on the 2022-2023 test set, representing a 12% improvement over single models. Analysis reveals significant regional heterogeneity, indicating a stronger PM2.5 suppression effect from new energy vehicles in eastern coastal areas compared to inland regions. This study provides a robust quantitative tool for vehicle pollution control and offers a methodological framework for multi-scale environmental modeling.