FORECASTING WAVE HEIGHT AND WAVE PERIOD USING LONG SHORT-TERM MEMORY AND GATED RECURRENT UNIT NEURAL NETWORKS

Waves are the driving force that shapes the coastline. Precise prediction of wave height and period is of great importance in many coastal studies. Accurate wave parameters are required to execute coastal activities such as merchant vessel routing, offshore drilling, coastal protection works, naval operations more efficiently and safely. This study has employed Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks to predict past wave parameters such as wave height and wave period for five different data sets. The research aims to check model accuracy and compare results based on size and data type. Furthermore, the study attempts to provide details about model prediction accuracy for wave period parameters, which most research ignores. The wave data used to train networks consisted of that collected from the global model and buoy measurement. The results show that both models have a high capability to produce long-term forecasting. GRU model produces a slightly better result for wave height where data is less non-linear with RMSE ranging between 0.05 to 0.15. In contrast, LSTM model results slightly outperform the GRU model results for wave period where data is more non-linear with RMSE ranging between 0.31 to 1.62. Additionally, the performance of both models was compared against the Multi-Layer Perceptron (MLP) network for all five data sets. Results show that LSTM and GRU models outperformed the MLP model. Finally, it was established that the model performance increases with increasing training data size, the number of neurons in the hidden layer, and time steps.