Abstract
Abstract
Accurate forecasting of surface PM2.5 concentrations is essential for enhancing air quality insights and enabling informed decision‐making in a timely manner. Traditional numerical models often exhibit biases originating from uncertainties in input parameters and oversimplified parameterization. This study introduces AGATNet, a graph‐based neural network aimed at correcting such biases by adaptively learning the spatial connections between air quality monitoring stations and associated temporal dependency of input features, leveraging masked self‐attentional layers and causal dilated 1D convolution. Trained with PM2.5‐contributing input features provided for the past 24 hr and future 72 hr during the years from 2016 to 2019, AGATNet effectively corrected CMAQ's 72‐hr advance forecasts of surface PM2.5 concentrations in South Korea for 2021. Across 183 monitoring stations, the application of AGATNet resulted in a substantial improvement in forecast accuracy, with index of agreement increased from 0.67 to 0.96 on +1 hr and root mean square error decreased by 51.56% on average throughout 2021, outperforming other machine learning models such as PM2.5‐GNN, multi‐layer perceptron, and long short‐term memory network. Notably, AGATNet demonstrated the most reliable hit rates for both the highly‐polluted episodes as well as relatively pristine conditions across South Korea, the distributions and occurrences of which were spatially and temporally more closely aligned to the observed values. AGATNet's success across diverse terrains and pollution scenarios in South Korea underscores its robust adaptability as well as the utility of graph neural networks in capturing spatial and temporal variabilities in input features more effectively.