<p>Water quality degradation poses an increasing challenge for the sustainable management of surface water resources in Iran’s semi-arid regions. This study employed an integrated approach, combining the Shannon’s Entropy Water Quality Index (EWQI), Monte Carlo simulation (MCS) and time series forecasting models, to evaluate and forecast the water quality dynamics in the Dez Dam reservoir in Iran. Monthly water quality data (2016–2025) from three monitoring stations were analyzed for key physicochemical parameters. The EWQI ranged from 82 to 157, corresponding to poor to excellent water quality classes, with nutrient enrichment emerging as the dominant factor influencing water quality variability. Monte Carlo simulations quantified uncertainty in parameter weights and yielded narrow 95% prediction uncertainty bands (d-factor &lt; 0.5), confirming the model’s robustness. To forecast EWQI one year ahead (April 2025–March 2026), four deep learning (DL) architectures including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Simple Recurrent Neural Network (RNN), and Convolutional Neural Network (CNN) were benchmarked against a Seasonal ARIMA model. Among these, LSTM achieved the best accuracy (RMSE = 2.05, NSE = 0.93, KGE = 0.88), effectively capturing temporal dependencies and non-linear trends. Forecast results indicate that the reservoir will maintain a “medium” water quality status, emphasizing the need for continued nutrient control and adaptive management strategies. The proposed EWQI–MCS–DL framework offers a reliable and transferable approach for water quality assessment and forecasting in dam reservoirs under environmental change.</p> Graphical Abstract <p></p> <p>This graphical abstract visually illustrates the step-by-step framework adopted in this study for uncertainty-aware water quality assessment and forecasting of the Dez Dam Reservoir. The process begins with the collection of raw water quality parameters, including Total Dissolved Solids (TDS), Electrical Conductivity (EC), Dissolved Oxygen (DO), potential of hydrogen (pH), chlorophyll-a (Chl-a), Total Nitrogen (TN), and Total Phosphorus (TP), which are statistically summarized to represent the fundamental characteristics of the monitoring data. These parameters form the primary input layer of the proposed methodology. In the next stage, Shannon’s entropy method combined with Monte Carlo simulation (MCS) is applied to derive objective entropy-based weights for each parameter. This step explicitly accounts for uncertainty in parameter importance and reduces subjectivity in index construction. The weighted parameters are then integrated to calculate the Entropy-based Water Quality Index (EWQI), providing a comprehensive and quantitative representation of the reservoir’s water quality status. The graphical abstract highlights the temporal variability of EWQI along with uncertainty bounds generated through repeated simulations. Subsequently, the EWQI time series is used as input for the forecasting module, where deep learning models and a classical Seasonal ARIMA model are implemented and compared. The forecasting block illustrates how different modeling approaches capture temporal dynamics, with deep learning models particularly LSTM demonstrating superior performance in representing non-linear behavior. To further enhance reliability, a bootstrap-based uncertainty analysis is conducted, producing confidence intervals around the forecasts. Finally, the combined forecast integrates model predictions and uncertainty bands to deliver an interpretable and decision-oriented outlook of future water quality conditions. Overall, the graphical abstract emphasizes the logical flow from raw data to uncertainty-aware forecasting, demonstrating the effectiveness of the integrated entropy–deep learning framework for supporting sustainable water quality management in reservoir systems.</p>