Rainfall-Runoff Prediction at Multiple Timescales with a Single Long Short-Term Memory Network
Martin Gauch, Frederik Kratzert, Daniel Klotz, Grey Nearing, Jimmy Lin, and Sepp Hochreiter
The paper introduces a deep learning architecture for rainfall-runoff predictions at multiple time scales. The architecture is based on the Long Short-Term Memory (LSTM) network, a recurrent neural network designed to model long-term dependencies. LSTMs process the input sequentially; longer series result in longer training times. The multi-timescale LSTM architecture processes past inputs at a coarser temporal resolution than more recent inputs, thus significantly reducing the training time. The typical time scales for hydrology modeling are hourly and daily data; short-term forecasts are relevant for events with high water flow fluctuations, such as floods. The presented multi-timescale LSTM model generalizes to varying (in number and resolution) timescales.
The authors evaluate two approaches: 1) the same LSTM is used for daily and hourly predictions; 2) individual LSTMs are used per time scale. The second approach is more general and accepts different input variables for different time scales. The models are compared to the previously developed LSTM model adapted to hourly predictions and a traditional hydrological model, the US National Water Model by the National Oceanic and Atmospheric Administration (NOAA). Hourly data are obtained from the United States Geological Survey (USGS) Water Information System, which are then averaged to obtain daily values. Only the basins included into the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset are selected. The results show that multi-timescale LSTM is computationally more efficient with no loss in accuracy compared to standard LSTMs and has a higher predictive skill compared to the National Water Model.
Hydrology and Earth System Sciences, 25, 4, 2045-2062, 2021-04-19.