A robust black carbon prediction model derived from observational datasets in the Yangtze River Delta region, China

Black carbon (BC) is a short-lived pollutant with significant environment and human health impacts. Monitoring BC is important, but its spatial coverage is limited. Therefore, predicting BC concentration is crucial in densely populated regions like the Yangtze River Delta (YRD). This study explores the machine learning (ML) models, including IAP, LASSO, RF, and SNN, to develop a robust BC prediction model for the YRD, based on BC behavior at the Dianshan Lake (DSL) site. The annual BC concentration at DSL was 1.37 microg/m³, mainly from liquid fuel combustion (68 %), biomass burning (16 %), and coal combustion contributing (16 %). The ML model was first trained with DSL data (June 2021-June 2022) and then applied to other YRD sites in four scenarios based on air mass trajectories. The SNN model had R² values over 0.80 for all scenarios at the DSL site. The model performed better during peak traffic periods, highlighting traffic's impact on BC levels. These models were applied to the YRD sites with R² from 0.45 to 0.84. Hangzhou (HZ, 0.75-0.81), Nanjing (NJ, 0.78-0.80), and Pudong (PD, 0.76-0.84) showed the highest results. This suggests the model effectively captures BC concentrations from sources with similar emissions and meteorological conditions.