# STGATN: A novel spatiotemporal graph attention network for predicting pollutant concentrations at multiple stations

## TL;DR

This paper introduces STGATN, a new AI model that improves air pollution predictions by capturing spatial and temporal patterns across monitoring stations.

## Contribution

STGATN introduces a novel spatiotemporal graph attention network with a fusion gate and transformer attention to avoid error propagation in pollution prediction.

## Key findings

- STGATN outperforms existing methods in predicting pollutant concentrations in Beijing.
- The model effectively captures dynamic spatiotemporal correlations and temporal causal relationships.
- The fusion gate and transformer attention reduce error accumulation in long-term predictions.

## Abstract

Accurately predicting air pollutant concentrations can reduce health risks and provide crucial references for environmental governance. In pollution prediction tasks, three key factors are essential: (1) dynamic dependencies among global monitoring stations should be considered in spatial feature extraction due to the diffusion properties of air pollutants; (2) precise temporal correlation modeling is critical because pollutant concentrations change dynamically and periodically; (3) it is vital to avoid propagation of long-term prediction errors across spatiotemporal dimensions. To address these challenges, we propose STGATN, a novel spatiotemporal graph attention network with an encoder-decoder architecture. Both the encoder and decoder incorporate a spatiotemporal embedding mechanism, a spatiotemporal graph attention block, a gated temporal convolutional network, and a fusion gate. Specifically, the spatiotemporal graph attention module is designed to use temporal and graph attention networks to extract dynamic spatiotemporal correlations. The gated temporal convolutional network is constructed to capture the long-term temporal causal relationships. The fusion gate adaptively fuses the spatiotemporal correlations and temporal causal relationships. In addition, a spatiotemporal embedding mechanism, including positional and temporal information, is added to account for pollutants’ periodicity and station-specific properties. Moreover, this paper proposes a transformer attention that establishes direct dependencies between future and historical time steps to avoid prediction error accumulation in the dynamic decoding process. The experimental results show that the proposed prediction model significantly outperforms the latest baseline methods on the air pollution dataset from actual monitoring stations in Beijing City.