Integrating Deep Neural Network with Numerical Models to Have Better Weather and Air Quality Forecast Both Spatially and Temporally

Numerical models are excellent tools for forecasting future weather events or air quality. The scientific and computational advancements in numerical models provided us with more accurate forecasts while promoting an understanding of various physical and chemical processes in the atmosphere. However, due to the simplified implementation of such processes, the understanding of modeling uncertainties was limited. In addition to this, these numerical models require significant computational resources and time to have a quality forecast. This thesis tries to mitigate the uncertainties which lead to large biases, using advanced deep neural network (DNN) models and integrate them into the existing numerical model to have better and faster weather and air quality forecasts. In this study, a long-term forecasting system based on a deep Convolutional Neural Network (CNN) was developed for the air pollutants such as ozone, NO2, PM2.5, and PM10 for up to two weeks. An optimized deep learning algorithm was used to develop species-specific and location-neutral models. These models used the simulation outputs of the Weather Research and Forecasting (WRF) model and Community Multiscale Air Quality (CMAQ) model at the first step, and then train a deep neural network (DNN) model for each air pollutant in the second step. Once trained, these models forecasted the next 24-hour in advance for all species across the geographical domain for up to two weeks. In the final task, a deep CNN system was developed to bias-correct and reduce systematic uncertainties in the simulation of meteorology, such as wind speed and direction, surface pressure, temperature, humidity, etc., in the WRF model.