Accurate storm surge prediction using a multi-recurrent neural network structure

Abstract

This paper considers storm surge prediction using a neural network and considering multiple physical characteristics. Based on the factors that influence storm surges and historical observation data, we divide the input to the neural network into time features extracted from the prediction target and the auxiliary features that affect storm surges, and construct a feature gate within multiple recurrent neural network (RNN) cells. Historical hurricane data are used to assess the effectiveness and accuracy of the proposed model. Comparative analysis against a long short-term memory (LSTM) storm surge prediction model is conducted to verify the prediction performance of the proposed method. The comparison results show that the multi-RNN model is superior to the LSTM model in terms of four evaluation metrics and for all lead times. In particular, the multi-RNN model accurately predicts the maximum storm surge water level, and the prediction results are more consistent with the rise and fall of the water. A comparison of the storm surge forecasts using inputs from different time intervals under different evaluation indices confirms the generalization and stability of our proposed model. The experiments of storm surge prediction at six stations further confirm the wide applicability of the model.

A Review of Application of Machine Learning in Storm Surge Problems

“…In terms of timeliness, an operational forecast usually can give a longer forecast horizon of over 72 h. However, the forecast horizons of ML models are usually up to 12 h ahead [10,49,102,139,149]. Only a small part of studies extended the lead time to 24 h [44,83], and fewer studies attempted to output the forecast water level time series with a length of over 24 h [38,48,62,78,79].…”

“…For example, Bai and Xu [49] found that for two different types of hurricanes, different combinations of predictors could give the best prediction. By comparing different characteristic inputs, Feng and Xu [102] demonstrated that wind speed was the most important physical factor before the storm surge peaked, while pressure became the dominant feature influencing surge heights afterward. Rus et al [79] performed an ablation study to evaluate the importance of individual feature encoders, indicating that the removal of the atmospheric encoder gave rise to the most significant performance decline.…”

“…The cell state is the pathway for information transfer, allowing it to be passed on through the sequence. A host of studies have explored the potential of LSTM for predicting storm surges and have demonstrated that LSTM models can effectively capture temporal patterns and dependencies between sequential values [46,47,49,64,81,92,100,102,[128][129][130]. For instance, Wei and Nguyen [87] proposed an encoder-decoder LSTM neural network model for predicting time-varying storm surges, showcasing its versatility in predicting storm surges generated by bypassing and landfalling storms of different sizes and intensities.…”

“…The evaluation metrics are necessary to quantify the accuracy of ML model predictions of storm surges. Commonly used evaluation metrics include the mean squared error (MSE) [18,102], mean absolute error (MAE) [125,130], root mean squared error (RMSE) [100,130], coefficient of determination (R 2 ) [81,125], and correlation coefficient (CC) [45,48]. In particular, the Kling-Gupta Efficiency (KGE) [53] accounts for three different important aspects of the reconstructed time series, including correlation coefficient variability and bias ratio term, thus providing comprehensive insights into how well the storm surge time series are reproduced.…”

Storm surge forecasting based on physics-informed neural networks in the Bohai Sea

“…Machine learning has been applied in climate prediction [2], and many scholars have used machine learning to conduct research and find applications in the field of storm surges. Lei [3] used the recurrent neural network (RNN) to forecast storm surges and found that it can obtain better prediction results than the backpropagation (BP) neural network. Liu [4] used long short-term memory (LSTM) to establish a single-station storm surge nowcasting model, and the model using the wind speed and wind direction process data in the previous 2 hours and the tide level process data in the previous 7 hours has the smallest error in terms of predicting the astronomical tide and storm tide level in the next 1 to 3 hours.…”