New Graph-Based and Transformer Deep Learning Models for River Dissolved Oxygen Forecasting

Abstract

Dissolved oxygen (DO) is a key indicator of water quality and the health of an aquatic ecosystem. Aspiring to reach a more accurate forecasting approach for DO levels of natural streams, the present work proposes new graph-based and transformer-based deep learning models. The models were trained and validated using a network of real-time hydrometric and water quality monitoring stations for the Credit River Watershed, Ontario, Canada, and the results were compared with both benchmarking and state-of-the-art approaches. The proposed new Graph Neural Network Sample and Aggregate (GNN-SAGE) model was the best-performing approach, reaching coefficient of determination (R2) and root mean squared error (RMSE) values of 97% and 0.34 mg/L, respectively, when compared with benchmarking models. The findings from the Shapley additive explanations (SHAP) indicated that the GNN-SAGE benefited from spatiotemporal information from the surrounding stations, improving the model’s results. Furthermore, temperature has been found to be a major input attribute for determining future DO levels. The results established that the proposed GNN-SAGE model outperforms the accuracy of existing models for DO forecasting, with great potential for real-time water quality management in urban watersheds.