Advanced Air Quality Prediction in Metropolitan Delhi via Graph Multi-Attention Network and Bayesian Hyperparameter Optimization

Air pollution is a major public health risk, especially for urban environments such as Delhi, where pollutant levels often surpass safe thresholds. It is essential to accurately predict the concentration of carbon dioxide (CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>) to facilitate real-time air quality monitoring, policy, and environmental management. This paper introduces a prediction model based on deep learning with the Graph Multi-Attention Network (GMAN) that can efficiently extract intricate spatiotemporal relationships from multivariate inputs and make interpretable predictions. The model is trained and tested on hourly CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> observations at the IGI Delhi monitoring station for one year (April 2023 to April 2024). The solution includes large-scale data preprocessing, lag-feature engineering from lag, and SHAP-based feature selection to enhance the quality of input. GMAN employs quantile regression loss to produce uncertainty-aware and robust CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> concentration predictions. Bayesian hyperparameter optimization is further used for model parameter effective tuning to enhance model accuracy and generalization. Comparative performances against benchmark models such as LSTM, CNN-LSTM, and GNN show that the presented GMAN performs better than all of them in RMSE, MAPE, and R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> performance metrics. In addition, the model's interpretability aspects are capable of identifying the contributing factors of CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> variations, which are valuable inputs for environmental regulation and urban planning. This work highlights the potential of cutting-edge neural network architectures for data-intensive air quality forecasting in complex urban environments.