Few-Shot Flood Water Level Prediction Based on Mechanistic Models and Deep Learning: A Collaborative HEC-RAS-LSTM Framework

Abstract

Hydrological and hydraulic modelling are crucial for flood forecasting and disaster early warning. While integrating neural networks with hydrological and hydraulic models enhances flood prediction performance, the generalization capability of neural networks tends to degrade under conditions of limited sample availability. Mechanistic models like HEC-RAS exhibit limitations in complex hydrological scenarios or when key hydrological data are missing. In this study, a novel approach that combination of HEC-HMS and HEC-RAS with neural networks is proposed. High-quality hydrological data are generated through HEC-RAS to address the issue of missing observed data. And the simulation data are used to drive a neural network model (LSTM) for flood water level prediction. When the mechanism model has insufficient data or it is difficult to directly solve nonlinear dynamic processes, neural networks can compensate for the limitations of the mechanism model through data-driven modelling using high-quality data generated by simulation. The synergy between above approaches significantly enhances prediction reliability. Experimental results demonstrate that the LSTM model, enhanced by HEC-RAS-generated data, achieves excellent performance, with R2 of 0.950 and a MSE of 0.034 m. This provides a new paradigm of "physical mechanism-data driven"