Modeling Black Carbon Concentrations at High Spatial Resolution

Abstract:

Concentrations of directly-emitted air pollutants vary on fine temporal and spatial scales, depending on meteorological conditions and proximity to pollution sources. Accurate knowledge of pollution concentrations is needed to understand air pollution exposures and to protect vulnerable communities from related adverse health effects. At sub-kilometer spatial resolutions needed to resolve near-source variations in pollutant concentrations, Eulerian models are computationally expensive, and furthermore alternative approaches to parameterization of turbulent mixing are required, as “K theory” is not suitable at sub-kilometer length scales. In this research, a neural network was trained to predict high-resolution (150 m) concentration maps from a combination of lower-resolution (4 km) meteorological model outputs and high-resolution (150 m) emission maps. This approach provides predictions of concentrations with an order of magnitude reduction in computational cost.

The Weather Research and Forecasting model (WRF) was used to model concentrations of black carbon in West Oakland, California, at 150 m spatial resolution for eight different sources of black carbon separately for a 100-day period during the summer of 2017. WRF was run in a nested configuration with grid resolutions ranging from 4 km to 150 m, with large eddy simulation turbulence closures on the 150 m domain. WRF outputs were used to train a neural network to predict daily average concentrations at 150 m resolution for eight individual mostly diesel-related black carbon (BC) sources including trucks, ships, and railroad locomotives. The model was used to map total BC concentrations and to apportion individual source contributions to the total. The neural network was trained using one month of the high-resolution WRF output and evaluated against the other 70 days of WRF output, with an RMSE of 0.2 micrograms per cubic meter for total black carbon concentrations.