Frederik Kratzert, Daniel Klotz, Günter Klambauer, Sepp Hochreiter, and Grey Stephen Nearing
Effective use of large-scale datasets for enhancing rainfall-runoff models is an outstanding problem in hydrological modeling. Currently, hydrological models perform best when calibrated for a specific basin, without additional information that could be extracted from other similar basins.
Here, we present recent results showing that models based on the Long Short-Term Memory network (LSTM) overcome this restriction. A single LSTM-based model calibrated on data from hundreds of basins significantly outperforms models calibrated per basin or regionally. The model is trained simultaneously on meteorological time series data and static catchment attributes.
We then present the results of a follow-up paper demonstrating that this approach also generalizes to predictions in ungauged basins. In this setup, a single LSTM-based model is trained on a subset of all basins and evaluated on basins that supplied no training data. We show that the ungauged LSTM model on average outperforms the hydrological model calibrated separately for each basin in a gauged setting as well as the distributed US National Water Model.
We also demonstrate the potential of this approach for advancing knowledge of hydrological catchments. Concretely, we adapt the standard LSTM architecture to evaluate observable catchment characteristics and group catchments by learned similarities. Our results show that LSTM-type models are able to extract information from observable catchment characteristics to differentiate between different rainfall-runoff regimes.
American Geophysical Union, AGU Fall Meeting 2019, San Francisco, 9-13 Dec.