A hybrid multi-step storm surge forecasting model using multiple feature selection, deep learning neural network and transfer learning

A real-time and accurate storm surge prediction model is of great scientific value and practical significance in reducing human casualties and economic losses in coastal areas. For this purpose, a novel storm surge multi-step forecasting framework integrating time-varying filtered empirical modal decomposition (TVF-EMD), fast Fourier transform (FFT), phase space reconstruction, convolutional neural network (CNN), and long short-term memory neural network (LSTM) is proposed in this study. Among the supplementary strategies, the TVF-EMD is used to extract the fluctuation features of the storm surge data and decompose the storm surge time series into a number of IMFs; the FFT is employed to calculate the frequency values of each IMF, and the subsequences with similar frequency values are combined and reconstructed. Meanwhile, CNN is adopted to predict the preprocessed high-frequency components, while the low-frequency is predicted by LSTM. Subsequently, the ultimate prediction results of the raw storm surge are calculated by superimposing the predicted values of all components. Three datasets collected from southeastern coastal region of China and five relevant comparison models are carried out to evaluate the proposed approach, where the corresponding results demonstrate that: (1) data preprocessing strategy applying TVF-EMD and FFT can significantly improve forecasting performance; (2) the TVF-EMD decomposition method is more effective under the influence of low sampling rate and noise; (3) by observing the characteristics of the subsequence, the prediction by modules can achieve better results. In addition, in order to apply the model to engineering, the proposed model is transferred to the small data domain as a pre-trained model using a transfer learning approach. According to the prediction results of Wenzhou station in the 7821-storm surge event, it can be found the proposed model still has good robustness and generalization ability even though the new sample data is small. This also proves that the model has strong practical value in coastal storm surge warning as well as disaster prevention and mitigation. Overall, the storm surge prediction framework proposed in this study has higher prediction accuracy and transferability, and can provide scientific and reasonable theoretical guidance for the emergency management to develop disaster prevention strategies.