A Hybrid Novel Deep Learning Approach Applied for Time-Series Typhoon-Induced Storm Surge Prediction: A Case Study on Taiwan Northeastern Coast

This study investigates the application of a one-dimensional convolutional neural network-long short-term memory (1D CNN-LSTM) hybrid model for typhoon-induced storm surge forecasting, examining the effects of including or excluding time-lagged inputs (specifically 3-hour lag, t-3, and 6-hour lag, t-6) on prediction performance. By combining the strength of CNNs in extracting local features with the ability of LSTMs to capture long-term temporal dependencies, this hybrid approach is particularly well-suited for extending forecast lead times and improving prediction accuracy in storm surge applications. Moreover, CNNs can effectively reduce the number of model parameters through weight sharing in convolutional layers. Compared to fully connected neural networks (FCNs), this design helps lower the risk of overfitting and enhances training efficiency. The Longdong tidal station on the northeastern coast of Taiwan was selected as the study area, using 10 historical typhoon events from 2005 to 2024 for model training and validation. The results show that under a one-hour lead time forecast, the 1D CNN-LSTM model achieved a correlation coefficient of 0.98 with observed data, highlighting its strong potential for short-term prediction. Even for longer lead times (24-hour ahead prediction), the model maintained a correlation coefficient above 0.90, demonstrating significant improvements over traditional shallow static networks (BPNN) and standalone dynamic networks (LSTM). Overall, this hybrid model is promising for early warning and emergency response planning for typhoon-induced storm surges.