FloodSformer: A transformer-based data-driven model for predicting the 2-D dynamics of fluvial floods

Abstract

This paper presents the open-source FloodSformer (FS) model, which uses a transformer-based deep learning architecture to simulate real-time evolution of fluvial floods. A cross-attention mechanism captures spatiotemporal correlations between inundation maps and inflow discharges, while maps compression is obtained by an autoencoder neural network. Long-duration events are predicted using an autoregressive approach. Model performance is assessed considering two case studies: an urban flash flood at laboratory scale and real flood events along the Po River (Italy). Results show that prediction errors are within the range of uncertainties typical in hydraulic modelling. The FS model accurately predicts 2D inundation maps over time with negligible accumulation error and requires minimal computational time, making it suitable for real-time forecasting. These results demonstrate the model’s potential to improve flood prediction accuracy and responsiveness, supporting more effective flood management and resilience strategies.