Operational use of the NASA SPoRT Machine Learning Hydrologic Forecasts at the National Weather Service River Forecast Centers

The National Aeronautics and Space Administration (NASA) Short-Term Prediction Research and Transition Center (SPoRT) has been collaborating with the National Weather Service (NWS) River Forecast Centers (RFCs) on the use of machine learning based hydrologic forecast information in operational use. NASA SPoRT forecasts are trained using a Long Short-Term Memory (LSTM) network using their Land Information System (SPoRT-LIS) relative soil moisture with 6-hr temporal resolution Multi-Radar Multi-Sensor quantitative precipitation estimates (MRMS QPE) and gauge data. LSTM Forecasts are provided for three different quantitative precipitation forecast (QPF) scenarios derived from the NWS Weather Prediction Center (WPC), the NCEP Global Forecast System (GFS) and National Blend of Models (NBM).

The NWS RFCs provide stage/elevation based forecast guidance to the NWS Weather Forecast Offices (WFOs) for dissemination to the public in routine forecasts served on the Advanced Hydrologic Prediction Services (AHPS) web pages and are used in official river flood watches and warnings. In some cases the WFOs also provide similar hydrologic forecasts to the public for headwater locations that are quicker responding than are readily provided by RFC models. The LSTM forecasts are made available to the NWS via publicly available web pages and internally within NWS hydrologic applications. As a result LSTM forecasts are evaluated alongside internal NWS hydrologic model simulations developed by RFC and WFO hydrologists and used to provide hydrologic forecast guidance used in river forecasts, watches and warnings. Evaluation and use of LSTM forecasts in RFC and WFO operations have demonstrated this technique to be a valuable addition to the available information for hydrologic forecast services and look to expand this further in the future to other sites and offices. In locations where LSTM forecasts are effective, development of new locations can be a savings in personnel time and resources that are normally spent in the manual calibration of hydrologic models. Manual calibration of hydrologic models is time intensive and requires training and expertise that is a consideration within field offices. At both the RFCs and WFOs, LSTM forecasts are a successful example of a research based product that have been incorporated into NWS operations to assist in the NWS Mission of providing “services for the protection of life and property and enhancement of the national economy”. This presentation shares the experience and perspectives of the field office use of LSTM as a hydrologic forecast tool.