Prediction of water quality parameters and pollution exceedance analysis in typical rivers of semi-arid regions based on interpretable deep learning models

Abstract

Deep learning models that integrate environmental characteristics provide a powerful means for high-precision water quality prediction; however, their black-box nature can limit interpretability and reliability. We proposed an interpretable Attention-Gated Recurrent Unit (AT-GRU) model that integrates water quality, meteorological, and hydrological data from the semi-arid Dahei River Basin, to improve prediction accuracy and transparency of results. The model achieved superior daily-scale prediction accuracy (average R² = 0.907) over traditional machine learning and deep learning approaches. To enhance interpretability, SHapley Additive exPlanations (SHAP) analysis was conducted to identify key drivers behind the predictions. Results indicated that ammonia nitrogen (NH₃N), population count, and river flow were the dominant predictors of total nitrogen (TN) and total phosphorus (TP), while meteorological factors had limited influence under high-pollution conditions. Extreme precipitation events were found to temporarily elevate nutrient concentrations. Analysis of exceedances and extremes further highlighted specific periods of most effective regulatory interventions. Overall, our study contributes a data- and mechanism-informed modeling framework that supports targeted pollution control, early warning, and adaptive water quality management strategies in semi-arid regions.