Learning spatiotemporal dependencies using adaptive hierarchical graph convolutional neural network for air quality prediction

• Published in: Journal of cleaner production Vol. 459; p. 142541
• Main Authors: Hu, Wei, Zhang, Zhen, Zhang, Shiqing, Chen, Caimei, Yuan, Jiwei, Yao, Jun, Zhao, Shuchang, Guo, Lin
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 25.06.2024
• Subjects: Air quality; data collection; Deep learning; Graph convolutional networks; Multi-level aggregation; neural networks; PM2.5; prediction; Spatiotemporal; topology; Deep learning; Multi-level aggregation; Air quality; Graph convolutional networks; Spatiotemporal; PM2.5
• ISSN: 0959-6526; 1879-1786
• DOI: 10.1016/j.jclepro.2024.142541

Air quality prediction has garnered significant attention from both governmental bodies and the general public due to its close connection with people’s daily lives. Learning spatio-temporal dependencies within air quality data are crucial for accurate air quality forecasting. Graph Convolutional Network (GCN) has been widely utilized in spatiotemporal data prediction due to its remarkable ability of capturing non-Euclidean spatial dependencies among nodes. However, previous research on Spatiotemporal Graph Neural Networks (ST-GNNs) has either utilized graph adjacency matrices predefined by heuristic rules or learned through trainable parameters. These approaches fail to accurately represent the true spatial relationships between stations and adaptively capture the required multi-level spatial dependencies for prediction. To address these limitations, this paper proposes an Adaptive Hierarchical Graph Convolutional Neural Network (AHGCNN) for PM2.5 prediction. More specially, an adaptive hierarchical graph convolutional structure is initially developed for dynamically extracting multi-level spatial dependencies. Then, a hierarchical aggregation scheme is designed to consolidate the unique topological structures of graph neural networks across different levels. Finally, a unified Adaptive Hierarchical Graph Convolution Gated Recurrent Unit (AHGC-GRU) model is proposed by integrating Gated Recurrent Unit (GRU) with hierarchical graph convolution in an effort to capture spatiotemporal dependencies of air quality data. Extensive experiments are conducted on real-world datasets, and the results demonstrate that the proposed model outperforms existing models.