Comparing CMAQ Forecasts with a Neural Network Forecast Model for PM2.5 in New York

Human health is strongly affected by the concentration of fine particulate matter (PM2.5). The need to forecast unhealthy conditions has driven the development of Chemical Transport Models such as Community Multi-Scale Air Quality (CMAQ). These models attempt to simulate the complex dynamics of chemical transport by combined meteorology, emission inventories (EI's), and gas/ particle chemistry and dynamics. Ultimately, the goal is to establish useful forecasts that could provide vulnerable members of the population with warnings. In the simplest utilization, any forecast should focus on next day pollution levels, and should be provided by the end of the business day (5 p.m. local). This paper explores the potential of different approaches in providing these forecasts. First, we assess the potential of CMAQ forecasts at the single grid cell level (12 km), and show that significant variability not encountered in the field measurements occurs. This observation motivates the exploration of other data driven approaches, in particular, a neural network (NN) approach. This approach makes use of meteorology and PM2.5 observations as model predictors. We find that this approach generally results in a more accurate prediction of future pollution levels at the 12 km spatial resolution scale of CMAQ. Furthermore, we find that the NN is able to adjust to the sharp transitions encountered in pollution transported events, such as smoke plumes from forest fires, more accurately than CMAQ.