Artificial neural network-based multi-input multi-output model for short-term storm surge prediction on the southeast coast of China

In recent years, to reduce social and economic losses, timely and accurate storm surge forecasts have been attracting growing attention from coastal engineers. Although a host of studies have demonstrated the feasibility of artificial neural networks (ANNs) in predicting storm surges, few elaborated parametric studies have been performed to investigate the optimal sliding window sizes of input variables of ANN, and the effect of the selection of training data, particularly concerning typhoon intensity and tracks, on model performance remained less understood. This work proposes a multi-input and multi-output (MIMO) neural network to forecast storm surge time series along the southeast coast of China (SCC). More specifically, we explore whether simple ANNs are capable of learning to predict storm surge time series using only historical observations. The ANN models were independently trained with long-term observational data of storm surges and typhoon parameters collected at Xiamen, Dongshan, and Shantou stations from 1950 to 2000. Then the models were employed to forecast storm surges under multiple typhoon scenarios with various lead times. The results suggest that the forecast skills of the present models are affected by the station locations, and the amplitudes and shapes of storm surge time series, excluding typhoon landfall locations. The optimal window sizes for typhoon parameters and previous surge levels (SLs) are different. Previous 1-h or 2-h typhoon information is sufficient, whereas a larger window size of SLs is needed to make more accurate predictions. The optimal values also differ across stations, indicating that a systematic parametric analysis is necessary for the implementation of ANN at a specific station. Furthermore, despite a slight underestimate of peak values and temporal shifts observed in some typhoon cases, the results highlight the accuracy of ANN in short-term forecasting for mild and moderate storm surges, especially those with a cnoidal profile. Our study also demonstrated the importance of the selection of training samples. It is expected that introducing additional extreme typhoon surge scenarios and using a more state-of-the-art model can reduce the generalization errors, particularly in forecasting extreme situations.