Frederik Kratzert, Daniel Klotz, Günter Klambauer, Sepp Hochreiter, and Grey Stephen Nearing
An outstanding problem in the hydrological modeling community is how to effectively use large-scale data sets for enhancing rainfall-runoff models. Currently, hydrological models perform best when calibrated for a specific basin, ignoring additional information that could be extracted from other (similar behaving) basins.
In this talk, we present recent results [Kratzert et al. 2019a] which show that models based on the Long Short-Term Memory network (LSTM) contradict this trend. A single LSTM-based model trained simultaneously on meteorological time series data and static catchment attribute data from hundreds of basins significantly outperforms a set of benchmark models that are either calibrated per basin (SAC-SMA, VIC, FUSE, mHM and HBV ensemble) or regionally using MPR (VIC and mHM). We also present the results of a follow-on paper [Kratzert et al. 2019b], which showed that this approach also generalizes to prediction in ungauged basins. That is, a single LSTM-based model trained on a subset of all basins and evaluated on a hold-out test set (i.e., simulating the ungauged setting in the sense that it was used for prediction in basins that supplied no training data) on average outperformed the SAC-SMA hydrological model calibrated separately for each basin using long data records (i.e., in a gauged setting). Our ungauged LSTM also out-performs that US National Water Model.
Besides showing the superior performance of this approach, we also demonstrate its potential for advancing hydrological catchment understanding. Concretely, we adapted the standard LSTM architecture to include an input gate that responds to observable catchment characteristics and groups catchments by learned similarities. This affords a level of interpretability that helps us understand how the trained model learns to transfer information between basins along complex, interactions between catchment observables.
- F. Kratzert, D. Klotz, G. Shalev, G. Klambauer, S. Hochreiter, and G. Nearing (2019) Benchmarking a Catchment-Aware Long Short-Term Memory Network (LSTM) for Large-Scale Hydrological Modeling. arXiv, 1907.08456v1, Machine Learning (cs.LG), 2019-07-19. (more)
- F. Kratzert, D. Klotz, A. K Sampson, S. Hochreiter, and G. Nearing (2019) Prediction in Ungauged Basins with Long Short-term Memory Networks. EarthArXiv. doi:10.31223/osf.io/4rysp, 2019-08-26. (more)
American Geophysical Union, AGU Fall Meeting 2019, San Francisco, 9-13 Dec.