River Flood Prediction Using a Long Short-Term Memory Recurrent Neural Network

River flooding and the impacts to the surrounding areas are a concern for decision makers throughout the United States. Accurate medium range (~3-5 days) forecasts of river flooding events are particularly challenging for forecasters. This is because streamflow forecasts provided by the River Forecast Centers may not incorporate quantitative precipitation forecasts beyond day 1 or day 2. Therefore, stream height forecasts will essentially show no change in values if a heavy rainfall event further than one or two days is expected to impact the area. Additionally, forecasts of gauge height are not available for all streams and rivers, some of which exhibit problematic flooding. Therefore, forecasters have to rely on local knowledge to estimate the likelihood of a river entering flood stage in these situations. Due to the non-linear relationships between river flooding, precipitation and rainfall infiltration rates, human forecast estimates can be extremely crude. While humans are unable to consider all possible combinations between variables, Long Short-Term Memory (LSTM) recurrent neural networks can be trained on historical data in order to learn the relationships between the data to accurately predict future flooding events. Furthermore, LSTM models are much less computationally expensive than hydrologic models, making longer term predictions of streams and rivers possible. To predict river gauge height, the LSTM time-lagged input features include: gauge height to initialize the model, the NASA Short-term Prediction Research and Transition Center’s instance of the Land Information System (SPoRT-LIS) relative soil moisture to describe the rainfall infiltration rates and 6-hr quantitative precipitation forecasts from the NWS’s Weather Prediction Center or any NCEP numerical prediction model (NWP) forecast. The developed LSTM produces a 5-day forecast in 6-hr segments. The network has been tested on a variety of river basins across the Tennessee Valley and Mid-Atlantic regions resulting in a river gauge height mean absolute error of less than 7 inches on day 5 of the forecast when evaluated over the Jan – June 2019 time period.