Long-Term Water Quality Prediction With Transformer-Based Spatial-Temporal Graph Fusion

Over the past decades of rapid development, the global water pollution problem became prominent. Accurate water quality prediction can detect the trend and anomaly of water quality changes in advance, thereby taking timely measures to avoid water quality problems. Traditional statistical methods for water quality prediction tend to fail to capture the complex relationship among multiple water quality variables. Deep learning models face a challenge to capture both temporal dependence and spatial correlation of the water quality series data. To solve the above problems, this work proposes an adaptive and dynamic graph fusion water quality prediction model based on a spatiotemporal attention mechanism named Spatial-Temporal Graph Fusion Transformer (STGFT). It integrates a spatial attention encoder, a temporal attention encoder, an adaptive dynamic adjacency matrix generator, and a multi-graph fusion layer. Among them, the first two are adopted to capture the spatial correlations and temporal characteristics among different water quality monitoring stations, respectively. The generator can produce adaptive and dynamic adjacency matrices to reflect potential spatial relationships in a river network. Experimental results with real-life water quality datasets reveal that the prediction accuracy of STGFT outperforms the existing state-of-the-art models. Note to Practitioners—This paper is motivated by the problem of long-term water quality prediction. The highly volatile water quality data and the nonlinear characteristics of the time series greatly affect the accuracy of the forecasting task. Existing approaches fail to simultaneously capture spatial correlations and temporal characteristics among different water quality monitoring stations, affecting the accuracy of water quality predictions. This work proposes a water quality prediction method that captures the spatial correlations and temporal characteristics among different water quality monitoring stations. Moreover, it produces adaptive and dynamic adjacency matrices to reflect potential spatial relationships in a river network. Experimental results from three real-world datasets show that this approach is feasible and obtains more accurate prediction results. Furthermore, this method can also be applied to other areas of time series prediction, including finance, traffic, and smart manufacturing.