Daily streamflow prediction using an LSTM neural network in Alpine catchments

<p>Prediction of rainfall-runoff response in Alpine catchments is complex because hydrological processes vary strongly in space and time, they are elevation and temperature dependent, subsurface water stores are heterogeneous, snow plays an important role, and runoff response is fast. As a result, the transformation of rainfall into runoff is highly nonlinear. Machine Learning (ML) methods are suitable for reproducing such nonlinearities between input and output data and have been used for streamflow prediction. Recurrent Neural Networks (RNNs) with memory states, such as Long and Short-Term Memory (LSTM) models, are particularly suitable for hydrological variables that are dependent in time. An example of a recent application of LSTM to the rainfall-runoff transformation in many catchments in the USA showed that the LSTM model can learn physically meaningful catchment embeddings from precipitation-temperature-streamflow data, and performs comparably to widely used conceptual hydrological models (Kratzert et al., 2019).</p><p>In this study, we tested the LSTM approach on high-quality daily data from 23 Alpine catchments in Switzerland with three goals in mind. First, the LSTM model was trained and validated using daily climate variables (precipitation, air temperature, sunshine duration) and streamflow data on all catchments individually and the performance was compared to a distributed hydrological model (PREVAH). The performance of the LSTM model was in many (but not in all) cases better than the hydrological model. Second, a single LSTM model was trained in all catchments simultaneously, embedding terrain attributes extracted from the Digital Elevation Model (DEM). In this way differences between catchments related to the elevation and temperature dependent hydrological processes, such as snow accumulation and melt, evapotranspiration, runoff generation, etc., can be captured. We show the performance of this model and evaluate the regionalization potential provided by it. Third, the LSTM model was applied in an ensemble forecasting context, and we discuss the benefits and limitations this application brings compared to forecasting with a process-based hydrological model.</p>