Improving sub seasonal streamflow prediction for hydrological drought forecasting using machine learning in the Rhine River basin in the Netherlands

Tefera B Shibeshi , Shankar karuppannan

Accurate sub seasonal streamflow forecasts are essential for hydrological drought preparedness in large river basins with competing water demands. Traditional process based hydrological models often struggle to represent non linear, non stationary catchment behavior under changing climate and land use conditions. In this study, we develop Long Short Term Memory (LSTM) networks to improve sub seasonal streamflow forecasting in the Rhine River Basin, focusing on low flow conditions at the Lobith outlet in the Netherlands. We evaluate several univariate and multivariate LSTM configurations that ingest combinations of streamflow from multiple upstream gauges and basin average precipitation and temperature derived from GRDC and ERA5 data. Models are trained to predict daily discharge, weekly averages, and weekly minimum flows for lead times of 1–20 days and 1–4 weeks, and forecast skill is assessed using the Nash–Sutcliffe efficiency (NSE) and mean absolute percentage error (MAPE) relative to climatological benchmarks. The best multivariate LSTM model achieves NSE values up to 0.99 at short lead times and substantially reduces MAPE compared with climatology, particularly for drought relevant low flow metrics, while predictive skill gradually declines as the forecast horizon increases. These results highlight the potential of deep learning to enhance sub seasonal low flow forecasting and support early warning systems for hydrological drought risk management in large European river basins.