Reservoir operation is a complex human-controlled activity that substantially alters surface hydrology. A reservoir operator needs to account for reservoir gains and losses, and other upstream and downstream conditions, in order to make decisions about daily releases that meet operating targets. The National Water Model (NWM) version 2.0 is anticipated to include representation of approximately 5,300 reservoirs. By utilizing a levelpool routing, the current NWM does not represent complex reservoir dynamics, which negatively affects the accuracy of the model’s simulations and forecasts.
Lynker scientists at the NWC are utilizing machine learning techniques based on historical records of release decisions to improve the NWM’s operational forecasting skill. Deep neural networks (i.e., feed-forward backpropagation with three hidden layers and a considerable number of neurons) have been utilized to train the NWM to think like a reservoir operator. K-fold cross-validation and dropout techniques were considered to train generalized models and avoid the overfitting problem, which is commonly ignored in the application of ML techniques in hydrology and water resources fields. There is no need for rule curves to train the model and the proposed technique has the potential to scale well to the national level. Application of the developed models for two study basins, Apalachicola-Chattahoochee-Flint (ACF) and Colorado Headwaters, have shown promising results by mainly utilizing observed previous day release, reservoir inflow, previous day storage and day of the year as inputs to the model.