Explainable Water Quality Index Forecasting Framework Using Hybrid Graph Attention and Transformer Networks for Intelligent Early Warning Systems

Abstract

The importance of water quality monitoring for aquatic ecosystem protection cannot be overemphasized, but the traditional machine learning algorithms used for monitoring data analysis have shortcomings in failing to account for the spatiotemporal interactions of the data. This study seeks to develop a spatiotemporal deep learning framework that employs Graph Attention Networks (GAT), Temporal Fusion Transformers (TFT), and SHAP-based interpretability for WQI forecasting and early warning purposes. Firstly, preprocessing of multisite physicochemical data involves filling in missing values using interpolation, removal of anomalies, and Min-Max normalization. Secondly, a spatial graph of monitoring stations is created, whereby the GAT layer captures the spatial information by allocating attention scores to the neighboring nodes, and TFT is employed to capture long-term temporal interactions and nonlinear trends. Lastly, a persistence-based early warning system identifies pollution degradation risks, while SHAP gives insight into parameter contribution. Empirical tests have revealed that the GAT-TFT model significantly outperforms conventional techniques (ANN, LSTM, Hybrid LSTM, and basic GNNs), exhibiting superior prediction efficiency with MAE = 0.048, RMSE = 0.079, and  = 0.96. Moreover, the early warning system exhibits an accuracy of 0.92, precision of 0.89, recall of 0.94, an F1 score of 0.91, and a False Alarm Rate of 0.08, where dissolved oxygen, turbidity, and pH were found to be the most significant factors. In conclusion, the presented framework offers environmental agencies a powerful and highly efficient method for proactive management of water resources.