A novel model to predict significant wave height based on long short-term memory network

Abstract A long short-term memory (LSTM) network is proposed for the quick prediction of significant wave height with higher accuracy than conventional neural network. The LSTM network is used for 1-h and 6-h predictions at ten stations with different environmental conditions. Using the wind speed of the past 4 h and the wave height and wind direction of the past 1 h as input parameters, the LSTM prediction results were obtained, and compared with results from a back propagation neural network, extreme learning machine, support vector machine, residual network, and random forest algorithm. Five statistical indicators were used to evaluate the results comprehensively. The minimum mean absolute error percentage of the 1-h and 6-h forecasts was 5.14% and 5.24%, respectively. The results demonstrate that the LSTM can achieve stable prediction effects, with accurate 1-h predictions and satisfactory 6-h predictions. In addition, predictions for four time spans, namely 12 h, 1 day, 2 days, and 3 days, were determined for Station 41008. The results show the powerful ability of LSTM to perform long-term prediction. The simulating waves nearshore-LSTM (SWAN-LSTM) model was proposed to make a single-point prediction, and it outperformed the standard SWAN model with an improvement in accuracy of over 65%.