A deep learning-based probabilistic framework for streamflow prediction and water quality assessment

This study presents an integrated modelling framework designed to support water quality assessment and regulatory compliance analysis in river basins. The proposed approach combines deep learning-based streamflow prediction, stochastic uncertainty propagation, hydrological spatial disaggregation, and process-based water quality modelling at the basin scale. The framework is applied to the Júcar River Basin (eastern Spain), a highly regulated Mediterranean system characterized by strong hydrological variability and significant water quality pressures. Streamflow is predicted using a Long Short-Term Memory (LSTM) network, and uncertainty is propagated through Monte Carlo simulations to generate ensemble inflow predictions. These inflows are spatially disaggregated using variation factors derived from an HBV hydrological model, providing distributed inputs to a water quality model that simulates concentrations of biochemical oxygen demand (BOD), phosphorus, and ammonium. The framework successfully reproduces the main spatial and temporal patterns of pollutant concentrations while explicitly quantifying the uncertainty associated with hydrological predictions. Water chemical status classifications show high agreements with observations, achieving accuracies of 90% for BOD, 84.4% for phosphorus, and 85.3% for ammonium. By providing probabilistic estimates of pollutant concentrations, the framework also enables identification of periods with elevated risk of regulatory threshold exceedance. Overall, the proposed approach offers a robust and transferable methodology for integrating probabilistic hydrological forecasting with water quality assessment, supporting risk-informed river basin management under uncertainty.