AGATNet: An Adaptive Graph Attention Network for Bias Correction of CMAQ‐Forecasted PM 2.5 Concentrations Over South Korea

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 3
• Main Authors: Dimri, Rijul, Choi, Yunsoo, Salman, Ahmed Khan, Park, Jincheol, Singh, Deveshwar
• Format: Journal Article
• Language: English
• Published: 01.09.2024
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000244

Accurate forecasting of surface PM 2.5 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 PM 2.5 ‐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 PM 2.5 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 PM 2.5 ‐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. Air pollution monitoring stations in South Korea are unevenly distributed across various geographic locations, challenging traditional deep learning models like convolution networks, which assume uniformly spaced data points. Graph neural networks (GNN) are better equipped to handle data that is not uniformly distributed, making them particularly suitable for tasks like air pollution forecasting. AGATNet, a graph neural network, was developed to enhance the accuracy of air pollution forecasts by utilizing data from the previous day and making predictions for the next 3 days to refine PM 2.5 concentration forecasts. Our results show that AGATNet significantly outperforms both traditional forecasting methods and other advanced models, particularly in predicting low‐pollution events that occur frequently and high‐pollution events that occur less frequently. Although AGATNet faces challenges in areas with minimal variation in pollution levels, it still surpasses other methods in performance. This progress highlights AGATNet's potential as an effective tool for air quality monitoring and demonstrates the advantages of using GNN for station data. We integrated graph attention and temporal convolution networks to correct biases in CMAQ‐forecasted surface PM 2.5 concentrations in Korea AGATNet's adaptive learning autonomously understands node connectivity & temporal sequence, eliminating the need for a pre‐existing graph AGATNet outperformed multi‐layer perceptron, long short‐term memory network, and PM 2.5 ‐GNN in correcting biases in 72‐hr advance forecasts of surface PM 2.5 concentrations