Spatio-temporal graph neural networks for the forecast of natural water systems

ABSTRACT Accurately forecasting natural water systems is a complex task due to their interconnected structure, where both spatial and temporal dependencies play a critical role. In this work, we applied spatio-temporal graph neural networks of varying complexity to forecast the flow of rivers and the total releases of reservoirs in the Upper Colorado River Basin. Since prolonged droughts driven by climate change can reduce water levels in hydrological systems to critical thresholds, it is essential to forecast to mitigate their negative consequences. The models were trained using five years of historical time series data from directly connected sensor points within a river basin. We evaluated six models and compared their forecasting performance using mean squared error, overall, in boxplots. The graph convolutional recurrent network model performed the best compared to the other five models in the case study, which indicates that the graph convolution with the Chebyshev polynomial has the best forecast accuracy in water system forecasting.