Reliable Spatial Downscaling of PM2.5 Forecasts of WRF-Chem in Flexible Grids using Deep Learning

Abstract

Accurate PM2.5 concentration forecasting is significant for air pollution prevention and control, such as traffic restriction, factory emissions limitation control, and citizen protection measures (Shi et al., 2019; Yuan et al., 2019). While there have been extensive studies dedicated to PM2.5 prediction, achieving precision, particularly in urban contexts, continues to be an elusive goal. Numerical models like CMAQ and WRF-Chem, crucial for simulating atmospheric pollutants, rely on complex differential equations and assumptions that can lead to inherent uncertainties in their predictive accuracy. Machine learning methods, including deep learning, Random Forest, and deep ensemble models, have proven effective in estimating PM2.5 levels, but their focus on broader scales often leads to inadequate representation of finer urban pollution details and uneven sensor distribution. This research aims to overcome these limitations by integrating deep learning and downscaling techniques. The approach begins with the downscaling of spatial resolution using an unstructured grid, enhancing data resolution from ground-based sensors. The study then synergizes outputs from WRF-Chem models and transformer neural networks to capture complex non-linear relationships and temporal patterns in PM2.5 data. This study enhances PM2.5 prediction accuracy and efficiency in urban areas by integrating transformer networks and quadtrees for pattern recognition and data management, addressing sensor distribution issues and aligning model outputs with ground observations.