Bias correcting and extending the PM forecast by CMAQ up to 7 days using deep convolutional neural networks

Abstract

With rising levels of air-pollution, air-quality forecasting has become integral to the dissemination of human health advisories and the preparation of mitigation strategies. To achieve more accurate forecasts, researchers around the globe are developing mathematical modeling techniques to obtain more accurate forecasts. In this study, we explore the capabilities of a deep neural network (DNN) model in conjunction with conventional, more reliable chemical transport models i) to improve the performance of the chemical transport model (e.g., CMAQ) and; ii) to extend the forecast period to seven days. Using a generalized deep convolutional neural network (CNN) model, we forecast air pollutants such as PM2.5, PM10, and NO2 up to seven days in advance. The CNN model bias-corrects hourly concentrations of air pollutants from the CMAQ model on the first-day and forecasts the remaining six days. Our results show improved performance of the average yearly index of agreement (IOA) from the CMAQ to the CNN model by 13% for PM2.5, 22% for PM10, and 43% for NO2 for the first-day bias correction; and the seventh-day forecast of NO2 by the CNN model was more accurate than the first-day forecast of the CMAQ model. The forecasts for PM2.5 and PM10, however, are reliable only up to two days in advance. The trained model is also capable of forecasting pollutants at stations not included in the training. The increase in the average yearly IOA at such stations is 13% for PM2.5, 22% for PM10, and 40% for NO2. Although the CNN model enhances the performance of the CMAQ model, it can be further improved by adding remote sensing data.