Synergizing Machine Learning with Conceptual Model for Daily Reservoir Storage Predictions in the Data-scarce catchments

Reservoirs play a crucial role in water resources management. However, quantifying reservoir storage prediction is challenging, especially when reservoir inflow data is unavailable. In recent years, Machine Learning (ML) models have shown their successful application in hydrological predictions, although these models are criticized for their inability to follow physical constraints. On the other hand, conceptual models are applied widely in various hydrological studies due to their simplistic structure yet inclusive of various hydrological processes. However, these conceptual models show limited predictive skills. Thus, synergizing domain knowledge from a conceptual model with the predictive ability of the ML model can help for better physical consistent outputs. We developed the Physics Informed Machine Learning (PIML) model for reservoir storage predictions. This model combines the predictability of Long Short Term Memory (LSTM) with domain understanding of the conceptual (SIMHYD) model. The applicability of the PIML model is demonstrated on two United States reservoirs where reservoir inflow data is unavailable. Our results show that the PIML model outperforms the SIMHYD model in the reservoir storage predictions while being mindful that reservoir storage will not be more than the maximum storage capacity. This study may be helpful in better-informed reservoir operation in the data-scarce catchments.