Accurate representation of short-term reservoir water-level dynamics is essential for operational analysis and scenario-based assessment under prescribed inflow–outflow conditions. In many practical applications, physically based modelling is limited by incomplete process knowledge, unavailable boundary conditions, or insufficient temporal resolution of input data. This study presents a data-driven framework for hourly conditional simulation of reservoir water level based on a hybrid Conv1D–LSTM architecture. The model learns nonlinear relationships among hydraulic forcing, operational control, and system state from historical observations, and is evaluated in a recursive multi-step simulation (rollout) mode to reflect its intended use and capture error accumulation over time. A systematic analysis of input sequence length and activation function is performed to identify a robust model configuration. On the test set, the selected configuration (L=24, GELU) achieved RMSE = 0.1057 m, MAE = 0.0881 m, and R² = 0.972 in rollout evaluation. The proposed framework is designed for scenario-based simulation rather than one-step deterministic forecasting, enabling rapid operational screening of alternative inflow–outflow regimes. Unlike many previous studies that emphasize one-step predictive accuracy, this work explicitly assesses model stability in recursive multi-step simulation, which is more relevant for reservoir scenario analysis.

Accurate multi-horizon forecasting of reservoir water levels is essential for effective water resource management and flood risk mitigation. While deep learning models have demonstrated strong predictive capabilities, they often struggle to adequately represent uncertainty and extreme hydrological events, particularly at longer forecast horizons. This study proposes four variants of a Conv1D–LSTM–Temporal Attention (CLTA) architecture for probabilistic multi-horizon forecasting, differing exclusively in loss function design. The models incorporate non-crossing constraints, horizon-aware weighting, and event-aware weighting to address key limitations of standard quantile regression approaches. All models are trained on hourly water level data from May 2021 to October 2022 and evaluated on a fully unseen dataset spanning December 2022 to May 2023. The results demonstrate that horizon-aware weighting achieves the lowest average RMSE (0.0149) and the most stable performance across forecast horizons on unseen data, while event-aware weighting improves representation of extreme hydrological events and achieves the highest coefficient of determination ($R^2=0.9961$). However, a controlled experiment further reveals that model performance is strongly influenced by the data partitioning strategy, even when architecture and loss formulation are held constant. Overall, the findings indicate that loss function design, in interaction with data partitioning strategy, is a key contributing factor to model performance in deep learning-based hydrological forecasting. A Multi-Criteria Decision Analysis (MCDA) framework identifies the horizon-weighted model as the most robust general-purpose solution, while the event-aware model is preferable for applications focused on extreme event representation. These results highlight the importance of integrating domain knowledge into both model design and evaluation strategy, offering a scalable and computationally efficient alternative to increasing architectural complexity.

Accurate representation of short-term reservoir water-level dynamics is essential for operational analysis and scenario-based assessment under prescribed inflow–outflow conditions. In many practical applications, physically based modelling is limited by incomplete process knowledge, unavailable boundary conditions, or insufficient temporal resolution of input data. This study presents a data-driven framework for hourly conditional simulation of reservoir water level based on a hybrid Conv1D–LSTM architecture. The model learns nonlinear relationships among hydraulic forcing, operational control, and system state from historical observations, and is evaluated in a recursive multi-step simulation (rollout) mode to reflect its intended use and capture error accumulation over time. A systematic analysis of input sequence length and activation function is performed to identify a robust model configuration. On the test set, the selected configuration ($L = 24$, GELU) achieved RMSE = 0.1057 m, MAE = 0.0881 m, and R² = 0.972 in rollout evaluation. The proposed framework is designed for scenario-based simulation rather than one-step deterministic forecasting, enabling rapid operational screening of alternative inflow–outflow regimes. Unlike many previous studies that emphasize one-step predictive accuracy, this work explicitly assesses model stability in recursive multi-step simulation, which is more relevant for reservoir scenario analysis.