Towards more accurate very short-range flood forecasts with physics-enhanced machine learning models

ABSTRACT Graphical abstract illustrating a physics-enhanced LSTM framework for very short-range flood forecasting. The left panel shows a conventional data-drive n approach using fixed input windows, resulting in redundant and temporally misaligned upstream signals. The central panel presents the proposed physics-enhanced ap proach, where inputs are aligned according to inter-station lag times to ensure hydrological consistency. The right panel summarizes the resulting improvements, includ ing better flood peak prediction, reduced high-flow bias, and lower computational demand. Flood-related disasters have become increasingly frequent in recent years, as evidenced by severe events in Brazil, Spain, and Germany. The development of impact-based flood early warning systems (IBFWS) has been an essential tool for minimizing human and economic losses. Recent advances in data-driven models show strong potential for improving flood forecasting capabilities. More promising approach that has emerged is the use of physics-enhanced machine learning models. These models incorporate physical and hydrological concepts into data-driven frameworks, which enhance their interpretability and robustness. This paper proposes a physics-enhanced Long Short-Term Memory (LSTM) model to incorporate inter-station lag times into the model's feature selection and temporal configuration, improving flood forecasts. The framework is applied to a flood-prone urban basin using high-resolution (10-minute) rainfall and streamflow data, assessing both overall forecast skill and the accuracy of flood events, particularly the peak magnitude and timing errors. Results demonstrate that the physics-enhanced configuration consistently increases prediction accuracy by reducing redundancy among inputs. Moreover, it maintains the physical coherence of the hydrological processes, supporting the transition from black-box to grey-box modeling. The resulting architecture remained computationally efficient, highlighting the potential of physics-enhanced neural networks for operational and impact-based flood forecasting.