The National Weather Service (NWS) forecasts floods at approximately 3,600 locations across the United States (U.S.). However, the river network, as defined by the 1:100,000 scale National Hydrography Dataset-Plus (NHDPlus) dataset, consists of 2.7 million river segments. Through the National Flood Interoperability Experiment, a continental scale streamflow simulation and forecast system was implemented and continuously operated through the summer of 2015. This system leveraged the WRF-Hydro framework, initialized on a 3-km grid, the Routing Application for the Parallel Computation of Discharge river routing model, operating on the NHDPlus, and real-time atmospheric forcing to continuously forecast streamflow. Although this system produced forecasts, this paper presents a study of the three-month nowcast to demonstrate the capacity to seamlessly predict reach scale streamflow at the continental scale. In addition, this paper evaluates the impact of reservoirs, through a case study in Texas. Validation of the uncalibrated model using observed hourly streamflow at 5,701 U.S. Geological Survey gages shows 26% demonstrate PBias ≤ |25%|, 11% demonstrate Nash-Sutcliffe Efficiency (NSE) ≥ 0.25, and 6% demonstrate both PBias ≤ |25%| and NSE ≥ 0.25. When evaluating the impact of reservoirs, the analysis shows when reservoirs are included, NSE ≥ 0.25 for 56% of the gages downstream while NSE ≥ 0.25 for 11% when they are not. The results presented here provide a benchmark for the evolving hydrology program within the NWS and supports their efforts to develop a reach scale flood forecasting system for the country.(KEY TERMS: continental scale river dynamics; streamflow prediction; surface water hydrology; flood forecasting.)

We build three long short‐term memory (LSTM) daily streamflow prediction models (deep learning networks) for 531 basins across the contiguous United States (CONUS), and compare their performance: (1) a LSTM post‐processor trained on the United States National Water Model (NWM) outputs (LSTM_PP), (2) a LSTM post‐processor trained on the NWM outputs and atmospheric forcings (LSTM_PPA), and (3) a LSTM model trained only on atmospheric forcing (LSTM_A). We trained the LSTMs for the period 2004–2014 and evaluated on 1994–2002, and compared several performance metrics to the NWM reanalysis. Overall performance of the three LSTMs is similar, with median NSE scores of 0.73 (LSTM_PP), 0.75 (LSTM_PPA), and 0.74 (LSTM_A), and all three LSTMs outperform the NWM validation scores of 0.62. Additionally, LSTM_A outperforms LSTM_PP and LSTM_PPA in ungauged basins. While LSTM as a post‐processor improves NWM predictions substantially, we achieved comparable performance with the LSTM trained without the NWM outputs (LSTM_A). Finally, we performed a sensitivity analysis to diagnose the land surface component of the NWM as the source of mass bias error and the channel router as a source of simulation timing error. This indicates that the NWM channel routing scheme should be considered a priority for NWM improvement.

The NOAA National Water Model (NWM) became operational in August 2016, producing the first ever real-time, distributed, continuous set of hydrologic forecasts over the continental United States (CONUS). This project uses integrated hydrometeorological assessment methods to investigate the utility of the NWM to predict catastrophic flooding associated with an extreme rainfall event that occurred in Ellicott City, Maryland, on 27–28 May 2018. Short-range forecasts (0–18-h lead time) from the NWM version 1.2 are explored, focusing on the quantitative precipitation forecast (QPF) forcing from the High-Resolution Rapid Refresh (HRRR) model and the corresponding NWM streamflow forecast. A comprehensive assessment of multiscale hydrometeorological processes are considered using a combination of object-based, grid-based, and hydrologic point-based verification. Results highlight the benefits and risks of using a distributed hydrologic modeling tool such as the NWM to connect operational CONUS-scale atmospheric forcings to local impact predictions. For the Ellicott City event, reasonably skillful QPF in several HRRR model forecast cycles produced NWM streamflow forecasts in the small Ellicott City basin that were suggestive of flash flood potential. In larger surrounding basins, the NWM streamflow response was more complex, and errors were found to be governed by both hydrologic process representation, as well as forcing errors. The integrated, hydrometeorological multiscale analysis method demonstrated here guides both research and ongoing model development efforts, along with providing user education and engagement to ultimately engender improved flash flood prediction.

The U.S. National Water Model (NWM) is a large scale hydrology simulator. Although NWM achieves coupling of multi-scale hydrological processes, its predictability at individual catchments can be improved. Hydrologic post-processing is an approach to reduce systematic simulation errors with statistical models, and has been shown to improve forecast accuracy of both calibrated and uncalibrated models. In this experiment we trained a Long Short-Term Memory (LSTM) network to post-process the NWM output, and tested performance at 531 basins across the continental United States. The LSTM post-processor provided a significant benefit to nearly all aspects of NWM streamflow predictions. The LSTM also benefited from NWM input - in particular, representation of hydrologic signatures improved, which indicates better representation of physical flow patterns.

No abstract