An adaptive prediction framework for NOX high-emissions: evidence from Nanjing, China

Abstract

The efficient predictions of NO_X high-emissions from heavy-duty diesel vehicles are crucial to facilitate proactive emission reduction rather than post-intervention analysis. Based on 839,303 real-world driving records, a spatial–temporal guided deep transfer learning framework was developed and evaluated, where shared-determinants of NO_X high-emissions were frozen and vehicle-specific characteristics were adaptively fine-tuned. The results indicate that the 95th percentile can serve as an identification threshold for NO_X high-emissions, and corresponding hotspots and heterogeneous areas were identified. Compared with isolated spikes, prolonged NO_X high-emissions deserve more attention, where frequent fluctuations in driving behaviors and engine conditions are primary contributors. Finally, the optimal combination of model structure and frozen strategy is recommended to facilitate NO_X high-emission predictions across individual vehicles, with stable macro-average F1-socres of 0.92. Such findings provide environmental authorities with a deep understanding of NO_X high-emissions, and offer technical supports for developing early warning systems and achieving proactive emission interventions.