A Machine Learning Approach for High-Resolution Modeling and Apportionment of Black Carbon Concentrations in an Impacted Community

Abstract

High spatial resolution is needed in air quality models to resolve near-source gradients in primary air pollutant concentrations. Models that are used to support analyses of chronic exposure to air pollution also need to be run over long time scales (i.e., seasonal to annual). The resulting applications of Eulerian (grid-based) air quality models at high spatial resolution and over long time periods are computationally expensive due to the large number of computational cells and the small model timestep that must be used to maintain numerical stability in the representation of the meteorology. Computational costs are further increased by the need to quantify source contributions to overall air pollutant concentrations in order to design effective control strategies. A neural network-based model for source-resolved non-reactive primary air pollutants is developed and evaluated in this study. High-resolution (50 m) maps of source-resolved black carbon (BC) concentrations from diesel and gasoline exhaust in West Oakland, California, were developed using the Weather Research and Forecasting (WRF) model. Daily maps derived from two weeks of WRF modeling were used to train a neural network to predict BC concentrations for 100 summer days, using meteorological variables and high-resolution BC emission maps as inputs. The neural network predicts absolute BC concentrations with similar accuracy and relative source contributions to results obtained using full WRF model simulations, with an 80% reduction in the computational cost.