Health risks connected with fine particulate matter (PM2.5) pollutants are well documented; increased risks of asthma, heart attack and heart failure are a few of the effects associated with PM2.5. Accurately forecasting PM2.5 is crucial for state agencies directed to devise State Implementation Plans (SIPS) to deal with National Ambient Air Quality Standards (NAAQS) exceedances. In previous work, we explored the application of multi-temporal data-driven neural networks (NNs) to forecasting PM2.5. Our work showed that under different input conditions, the NN approach achieves higher forecasting scores for local (12 km) resolution when compared to the other Chemical Transport Model forecast models, such as the Community Multi-Scale Air Quality system (CMAQ). Critical to our approach was the inclusion of prior PM2.5 concentrations, retrieved from ground monitoring stations, as part of the input dataset for the NN. The NN approach can provide high-level forecasting accuracy; however, because of the dependency on ground monitoring stations, the forecast coverage is sparse. Here, we extend our previous station-specific efforts by forecasting hourly PM2.5 values that are spatially continuous through the use of a deep neural network (DNN). The DNN approach combines spatial Kriging with additional local source variables to interpolate the measured PM2.5 concentrations across non-station locations. These interpolated PM2.5 values are used as inputs in the original forecasting NN. Cross-validation testing, using all New York State AirNow PM2.5 stations, showed that this forecast approach achieves accurate results, with a regression coefficient (R2) of 0.59, and a root mean square error (RMSE) of 2.22 . Additionally, herein we demonstrate the usefulness of this approach on specific temporal events where significant dynamics of PM2.5 were observed; particularly, we show that even bias-corrected CMAQ forecasts do not track these transients and our NN method.
Deep Air Quality Forecasts: Suspended Particulate Matter Modeling With Convolutional Neural and Long Short-Term Memory Networks