Flood Process Simulation in Macau's Inner Harbor Area Based on CNN-LSTM

This study aims to develop a CNN-LSTM hybrid network model integrated with a coupled self-attention mechanism, based on deep learning techniques, to simulate flood processes in the Inner Harbor area of Macau. With global climate change and accelerated urbanization, Macau, a low-lying coastal city, frequently experiences urban flooding due to typhoons and heavy rainfall. While traditional hydrological and hydrodynamic models can accurately predict flooding processes, their computational intensity and lack of real-time responsiveness make them unsuitable for emergency disaster warnings. To address these limitations, this paper proposes a convolutional long short-term memory (ConvLSTM) model enhanced with a coupled self-attention mechanism. The model leverages an encoder-decoder structure to predict the evolution of flood processes under 4–10 hours of heavy rainfall in the Inner Harbor area of Macau.

The model integrates CNN components for extracting spatial features, LSTM components for capturing temporal features, and a coupled self-attention mechanism to dynamically reweight spatial-temporal representations, improving the model's sensitivity to key flood patterns. The encoder encodes input sequences into fixed-length vectors, while the decoder translates these vectors into target sequences. The self-attention mechanism ensures the model focuses on critical spatial and temporal regions, further enhancing prediction accuracy and robustness.

The training and testing datasets were constructed from simulation data generated by hydrological-hydrodynamic models and static geographical information data, following preprocessing and normalization. Evaluation metrics, including mean squared error (MSE), Nash-Sutcliffe efficiency coefficient (NSE), and relative error, were used to assess model performance. Results demonstrate that the proposed hybrid model, augmented by the coupled self-attention mechanism, effectively simulates maximum water depth distribution and flood evolution processes, achieving high consistency with hydrodynamic simulation data while providing improved predictive performance.