Assessing the adequacy of traditional hydrological models for climate change impact studies: a case for long short-term memory (LSTM) neural networks

Abstract. Climate change impact studies are essential for understanding the effects of changing climate conditions on water resources. This paper assesses the effectiveness of long short-term memory (LSTM) neural networks compared to traditional hydrological models for these studies. Traditional hydrological models, which rely on simplified process parameterization with a limited number of parameters, are examined for their capability to accurately predict future hydrological streamflow under scenarios of significant warming. In contrast, LSTM models, known for their capacity to learn from extensive sequences of data and capture temporal dependencies, present a promising alternative. This study is performed on 148 catchments, comparing four traditional hydrological models, each calibrated specifically on each catchment, against two LSTM models. The first LSTM model is trained regionally across the 148 catchments, while the second incorporates data from an additional 1000 catchments at the continental scale, many located in climate zones representative of the future climate within the study domain. The climate sensitivity of all six hydrological models is assessed using four simple climate scenarios (+3, +6 °C, −20 %, and +20 % mean annual precipitation) and an ensemble of 22 CMIP6 GCMs under the SSP5-8.5 scenario. Results indicate that LSTM-based models demonstrate a different climate sensitivity compared to traditional hydrological models. Moreover, analyses of precipitation elasticity to streamflow and multiple streamflow simulations on analogue catchments suggest that the continental LSTM model performs better and is therefore better suited for climate change impact studies – a conclusion that is also supported by theoretical arguments.