Urban road BC emissions of LDGVs: Machine learning models using OBD/PEMS data

Abstract

Urban Black Carbon (BC) emissions from light-duty gasoline vehicles (LDGVs) are challenging to quantify in real-world settings. This study employed a Portable Emission Measurement System (PEMS) to assess BC emissions from five LDGVs on urban roads. We also developed five machine learning (ML) models based on On-Board Diagnostics (OBD) data to predict BC emissions. Among these, the Random Forest (RF) model consistently demonstrates the best ability to predict BC emissions across all tested LDGVs, with R2 values exceeding 0.6. Integrating OBD-based ML models within vehicles could enable real-time BC monitoring and aid emission reduction strategies. We observed a strong correlation between BC emissions and engine parameters, such as engine speed and load (R2 values between 0.5 and 0.9). Furthermore, China VI standard-compliant LDGVs showed minor differences in BC emissions across urban road types. Vehicles equipped with gasoline direct injection (GDI) engines registered BC emission factors (EFs) of 0.141 ± 0.038 mg/km, an increase of 23.7% compared to their port fuel injection (PFI) counterparts, which averaged 0.114 ± 0.049 mg/km.