Towards learning universal, regional, and local hydrologicalbehaviors via machine learning applied tolarge-sample datasets
Sprache des Titels:
Regional rainfall?runoff modeling is an old butstill mostly outstanding problem in the hydrological sci-ences. The problem currently is that traditional hydrologi-cal models degrade significantly in performance when cal-ibrated for multiple basins together instead of for a singlebasin alone. In this paper, we propose a novel, data-driven ap-proach using Long Short-Term Memory networks (LSTMs)and demonstrate that under a ?big data? paradigm, this is notnecessarily the case. By training a single LSTM model on531 basins from the CAMELS dataset using meteorologicaltime series data and static catchment attributes, we were ableto significantly improve performance compared to a set ofseveral different hydrological benchmark models. Our pro-posed approach not only significantly outperforms hydrolog-ical models that were calibrated regionally, but also achievesbetter performance than hydrological models that were cali-brated for each basin individually. Furthermore, we proposean adaption to the standard LSTM architecture, which we callan Entity-Aware-LSTM (EA-LSTM), that allows for learn-ing catchment similarities as a feature layer in a deep learn-ing model. We show that these learned catchment similaritiescorrespond well to what we would expect from prior hydro-logical understanding.