Deep Learning for Operational Streamflow Forecasts, Or More Specifically: Long Short-Term Memory Networks As a Rainfall-Runoff Modulefor the U.S. National Water Model

This dissertation investigates deep learning (DL) and combining hydrologic process-based (PB) models with DL for a hybrid (HB) modeling approach (often referred to as ''physics-informed machine learning" or ''theory-guided learning") for improving the predictive performance of streamflow in the U.S. National Water Model. An in-depth analysis is made of the benefits of DL and the potential drawbacks of the HB models. No evidence is found supporting the use HB models over the "pure" DL models in the use cases analyzed. The performance of the HB models is found to degrade in ungauged basins, whereas the DL models do not. The DL models are the best performing models for predicting extremely high runoff events, even when such events are not included in the training set. Adding physics inspired constraints to data-driven models causes a loss of system information relative to the DL models. As such, a "pure" DL model, specifically the Long Short-Term Memory (LSTM), is chosen as one of the core modules for the Next Generation (Nextgen) U.S. National Water Model. The LSTM (via Nextgen) is applied to simulate streamflow for a three-year period across the 191,020 km^2 New England region.