A Deep Learning-Based CAE-LSTM Model for Enhanced Long-Term Prediction of Flood Wave Propagation

Predicting the dynamics of flood processes is paramount for effective disaster prevention and mitigation. Recently, Physics-Informed Neural Networks (PINNs) have been employed for flood dynamic prediction, demonstrating commendable performance in wave propagation forecasting. However, PINNs, which rely on traditional fully connected neural networks, exhibit certain limitations. Notably, their capacity for learning long-term wave propagation processes remains insufficient, and they struggle to generalize across diverse, previously untrained scenarios.In this study, we propose an innovative model that integrates a Convolutional Autoencoder (CAE) with a Long Short-Term Memory network (LSTM) to overcome these challenges. Drawing inspiration from the finite-difference method employed to solve the Shallow Water Equations (SWE), the CAE-LSTM model adeptly captures and predicts flow characteristics from both spatial and temporal dimensions. The CAE harnesses the power of convolutional neural networks to extract spatial features and generate compact latent representations, thereby reducing the complexity inherent in the physical system. Meanwhile, the LSTM captures the temporal dependencies within the latent feature space, enabling the prediction of the dynamic process based on time-series data.The efficacy of this model was validated through three classical two-dimensional dam-break scenarios. In the 60-second rolling prediction case, the accuracy of CAE-LSTM surpassed that of PINNs by approximately 60%, while its computational efficiency was enhanced by a factor of approximately 100. These results underscore the potential of CAE-LSTM to effectively capture the intricate dynamic behaviors of fluids, thereby offering a robust tool for predicting flood dynamics.