Spatiotemporal hierarchical transmit neural network for regional-level air-quality prediction

Abstract

Air pollution has become a growing problem owing to its harmful effects on human health, making accurate air quality prediction an important task to avoid serious environmental impacts. Recently, graph-neural-network-based methods have emerged as promising approaches to modeling the spatial dependencies between adjacent stations. However, they have limitations in capturing the hierarchical temporal features of the time series, which include trends, and periodic components. They also overlook dynamic correlations between stations and multiple types of connections, resulting in generated graphs that lack dynamics. In this paper, we proposed a spatiotemporal hierarchical transmit neural network for the prediction of air quality by extracting long-term periodic features and short-term spatiotemporal dependencies. It incorporates a periodic feature extraction component (PFEC), a scene dynamic graph module (SDGM), a spatiotemporal extraction component (STEC), and a transmit attention (TransATT) component. The PFEC applies discrete Fourier transform and trend decomposition techniques to extract long-term periodic features from the spatiotemporal graph. The SDGM generates dynamic graphs by combining node features of time series with predefined graphs to encode diverse station relationships. The STEC comprises of two convolutional operations and attention mechanisms, enabling the model to capture the short-term spatiotemporal dependencies. TransATT integrates the extracted short-term spatiotemporal dependencies and long-term periodic features, allowing the model to transmit with short- and long-term features. To demonstrate the effectiveness of the proposed model, we conducted experiments on three real-world datasets and found that our approach outperforms state-of-the-art methods.