Applications of Deep Learning in Atmospheric Sciences: Air Quality Forecasting, Post-Processing, and Hurricane Tracking

This study employs deep learning-based models for developing: fast, real-time air quality forecasting systems; a post-processing tool for bias-correcting the chemical transport model; and a reliable hybrid hurricane tracking model. A deep convolutional neural network (CNN) algorithm, which is an advanced deep learning algorithm, was employed to predict the hourly ozone concentrations each day (24 hours in advance) for the entire year using several meteorological variables and air pollution concentrations from the previous day. The CNN model showed a reasonable performance with an average index of agreement (IOA) of 0.84-0.89 and a Pearson correlation coefficient of 0.74-0.81. Although the CNN model successfully captured daily trends of the ozone concentrations, it notably underpredictd high ozone peaks during the summer. To address this issue, six generalized machine leaning ensemble models were developed to regularize low- and high-ozone episodes. By resampling the training dataset based on the daily peaks, the ‘best’ ensemble model reduced the ozone peak prediction error by 5 to 30 ppb during summer. Another deep CNN model was developed to post-process the results of the Community Multiscale Air Quality (CMAQ) model. The CNN model significantly improved the performance of the CMAQ model by improving its absolute correlation coefficient by 0.16 and reducing its prediction bias by more than 20 ppb on average. To improve the prediction of hurricane models, a novel, hybrid approach was proposed using a CNN and an ensemble Kalman filter (EnKF). First, a three-step CNN ensemble model was developed to predict direction, distance traveled, and intensity of a hurricane. Then, an EnKF was applied as a post-processing step. The results of the hybrid model for 17 tropical storms in 2017 showed statistical advantages over official National Hurricane Center (NHC) 24-hour-ahead forecasts (i.e., ~13% and ~34% improvement in track and intensity forecast biases, respectively).