CMAQ-CNN: A new-generation of post-processing techniques for chemical transport models using deep neural networks

Abstract

Chemical transport models simulate ambient air pollution concentrations by considering emission, transport and deposition mechanism, and other physical processes. Despite their advancements over the past decades, the models exhibit significant model-measurement errors, especially during high concentration episodes. Furthermore, biases within meteorological predictions, emission estimates, initial and boundary conditions, and other model inputs can propagate into the final predictions. This research introduces a new generation of post-processing application for real-time air quality forecasts that employ the physical process of a numerical model in conjunction with an advanced deep learning algorithm. We use a deep convolutional neural network (CNN) to map ozone precursors from Community Multiscale Air Quality (CMAQ) and meteorological parameters from the weather research and forecasting (WRF) model (as inputs variables) to observed hourly ozone concentrations at a monitoring station (as a target). Our results show that the CMAQ-CNN model significantly improves the performance of the CMAQ model in both accuracy and bias. The absolute correlation coefficient is improved by 0.16 on average. The CMAQ-CNN model improves simulated ozone peaks for almost all cases and reduces the bias of CMAQ predictions by an average of more than 20 ppb (or 40%). Systematic improvements in CMAQ-CNN simulations suggest that the deep learning model is effective at reproducing accurate estimates of ground-level air quality concentrations. While this study focuses on ozone in the United States and outputs of CMAQ, the proposed approach can be applied to any measured air pollution parameters or numerical model in a mesoscale resolution.