Beyond Observed Extremes: Can Hybrid Deep Learning Models Improve Flood Prediction?

Abstract. Predicting unprecedented floods is essential for disaster risk reduction and climate adaptation but remains a challenge for both hydrological and deep learning models. This study evaluates three hydrological models, a Long Short-Term Memory (LSTM) network, and three hybrid models in simulating extreme floods in more than 400 catchments in Central Europe. The hybrid models integrate hydrological process variables with meteorological inputs to enhance runoff simulations. Results show that the LSTM model outperforms traditional hydrological models, while hybrid models further reduce runoff simulation errors. However, all models tend to underestimate peak discharges, with over 50 % underestimation for unprecedented floods. LSTM-based models exhibit extrapolation limits, likely due to structural and statistical constraints. To improve extrapolation to rare events, future work should integrate physical principles into deep learning, including differentiable hydrological models, physics-guided loss functions, and synthetic extreme event generation. Additionally, regional modeling approaches, such as entity-aware LSTMs, could improve predictions by leveraging spatial hydrological similarities. Combining data-driven learning with physical reasoning will be key to improving flood simulations beyond observed extremes.