Significant wave height forecasting using long short-term memory neural network in Indonesian waters

Significant wave height (SWH) plays an important role in supporting marine operational and maritime activities, such as shipping, construction, and monitoring. Forecasting of significant wave height has been studied numerically using various ocean wave models. This numerical approach needs to cover quite a large domain to get better result prediction. Moreover, this kind of computation can be costly if we consider acquiring higher resolutions. In this study, we propose a novel modeling approach based on long short-term memory (LSTM) neural network model with SWH observation data set as the only input data. The LSTM model is used in predicting SWH in several conditions of Indonesian waters, which cover areas of the open sea, straits, nearshore, and inner sea. Based on previous SWH input data, single-step predictions were carried out, as well as multi-step with lead times of 12-, 24-, and 48-h to come with a recursive scheme. Accurate results are obtained for single-step predictions with RMSE ranging from 5.53 cm (nearshore area) to 27.13 cm (open sea). Different results are obtained when predicting in a multi-step scheme, the predicted values are still not consistent in capturing the upward, downward trend, and the maximum and minimum conditions from SWH data pattern. In this study, it was found that the length of the data had a significant effect on the performance of the LSTM model in predicting SWH in a single-step. Meanwhile, in predicting multi-step, the model's performance was influenced by fluctuations and data complexity.