Frederik Kratzert, Daniel Klotz, Guy Shalev, Günter Klambauer, Sepp Hochreiter, and Grey Nearing
Regional rainfall-runoff modeling remains a largely unsolved problem in hydrological sciences. This is because traditional hydrological models degrade significantly in performance when calibrated for a set of multiple basins compared to a single basin. Here, we use Long Short-Term Memory networks (LSTMs) and demonstrate that this problem can be overcome with data-driven models.
We are able to significantly improve performance compared to several different hydrological benchmark models. We train a single LSTM model on 531 basins from the Catchment Attributes and Meteorology for Large-Sample Studies (CAMELS) dataset using meteorological time series data and catchment features. Our model significantly outperforms the models calibrated regionally, and achieves better performance compared to models calibrated for each basin individually.
Furthermore, we propose an adaption to the standard LSTM architecture, which we call an Entity-Aware-LSTM (EA-LSTM). The EA-LSTM model allows evaluating the catchment characteristics and embedding it as a feature layer in a deep learning model. The EA-LSTM model is also able to learn complex interactions between the catchments attributes that allows grouping of different basins.
arXiv, 1907.08456v1, Machine Learning (cs.LG), 2019-07-19.