Enhanced urban PM2.5 prediction: Applying quadtree division and time-series transformer with WRF-chem

Abstract

Accurately predicting air quality levels in urban landscapes is a key environmental and public health challenge due to the complex interactions between pollutant emissions, atmospheric chemistry, and meteorological conditions. Chemical transport models such as Community Multiscale Air Quality (CMAQ) and Weather Research and Forecasting Modeling with Chemistry (WRF-Chem) provide fundamental insights into the behavior of atmospheric pollutants but face limitations in spatial resolution and prediction accuracy. Machine learning combined with low-cost sensors can improve prediction accuracy compared to numerical models, but uneven sensor distribution may result in biased prediction results. To address these limitations, this study introduces an innovative framework that leverages a quadtree space division to refine the numerical simulations of air pollutants into higher spatial resolutions by dynamically adjusting grid sizes based on the distribution of increased sensor deployments. Using a variable-resolution grid, the study first aggregates variables into a grid and then calculates spatial dependence based on grid cells. The deep-learning-based Time Series Transformer model is then used to perform detailed temporal predictions of PM2.5 concentration across the entire grid. The spatial and temporal modeling framework is designed to use the past 48 h' aggregated data, including meteorological variables from WRF-Chem numerical simulation and PM2.5 concentrations from ground monitoring stations and sensors, to predict future 48-h’ PM2.5 concentrations and tested across entire grids within Los Angeles County from January to June 2020. The results show that the performance of variable resolution grids is better than that of fixed grids regarding four metrics: minimum resolution, overall spatiotemporal RMSE, average bias, and prediction coverage. Among them, the variable resolution grids with up to 4 sensors showed the best performance, with an overall spatiotemporal RMSE of 1.12 μg/m3, which was about 40% lower than 2-sensor and 8-sensor variable resolution grids, a significant reduction of 55% compared to a 3 km fixed grid configuration. Future work will involve additional relevant factors such as emission sources and sinks to improve prediction accuracy.