The real time forecasting of wave features is very important to the ocean operations. The traditional way for forecasting wave height is based on physical process, it needs a lot of fundamental data to establish the model, which is the biggest obstacle to the model's accuracy. To solve this problem, we used a long short-term memory network model to forecast coastal wave features in different time series. Furthermore, compared it to linear model and support vector regression model with various kernel functions. Root Mean Square Error (RMSE) and correlation coefficient (R) were used for evaluating the model. The results showed that compared to other 4 models long short-term memory network model shows a higher accuracy in forecasting significant wave height; the correlation coefficient is 0.93. With the extension though of the forecasting leading time, the accuracy decreased. The correlation coefficient of forecasting next 3 hours’ wave height equals to that of forecasting 1 hour's wave height, and it is 0.98. In summary, LSTM model has engineering value in small-scale wave forecasting for trailing suction dredger constructing.
Waterway transportation, dredging, reclamation and other projects are highly sensitive operations to the marine environment. During the process of operations, real-time marine environment should be mastered to ensure that the projects are carried out safely and efficiently.
The existing research of forecasting coastal wave height is based mostly on numerical models, which require comprehensive climatic and geographical data to build and consume significant computational power and time (Thomas and Dwarakish, 2015). However, it is not applicable for most of the construction projects. Therefore, it is necessary to build a wave prediction model that is computationally efficient and can be applied across different wave condition.
With the development of machine learning techniques, it has been widely used for wave height prediction. Feed-forward network had been proven that it has a more general and adaptable capability in real-time wave height prediction (Deo, Sridhar and Naidu, 1998). Then, studies have shown that recurrent neural network (RNN) performed better in significant wave height prediction (Mandal and Prabaharan, 2006; Sadeghifar et al., 2017). RNN excels at extracting features from the input data of long sequences but still suffers from vanishing and exploding gradient problem. Besides, RNN updates the network with the latest information, it still performs poorly for sequence predictions that are dependent on long-term dependence. To solve this problem, an idea of increasing networked storage arose. LSTM (long short-term-memory) networks with special implicit units were first proposed, and their natural behavior is to save input in a long term (Hochreiter and Schmidhuber, 1997). LSTM is widely used in sequence studies with strong temporal correlation, such as speech recognition, handwriting recognition, and image information recognition (Graves and Jaitly, 2014; Graves et al., 2009; Aggarwal, 2018). Also, it has performed well in meteorological fields such as sea surface temperature prediction (Xiao et al., 2019) and weather forecasting (Salman et al., 2018).
