PREDICTING ONE-DAY AHEAD WAVE HEIGHTS USING A HYBRID MODEL OF LONG-SHORT TERM MEMORY AND ARTIFICIAL NEURAL NETWORK WITH RESAMPLING METHODS

In Japan, 1 m significant wave height (Hs) is treated as a threshold to make a decision of whether marine construction work for one week later can be carried out or not. Therefore, one-week-ahead wave predictions are critical. NOAA, JMA, and ECMWF forecast global wave information approximately one week ahead, while most marine construction work refers to nearshore wave observations (e.g., NOWPHAS) located along coasts. According to Hara et al. (2019), the accuracy of forecasted global wave heights is insufficient for making the decision for the nearshore marine construction. The present study challenges improving the accuracy of one-day ahead wave heights at Hitachinaka Port, Japan, by developing a framework of the hybrid model of LSTM and ANN with the resampling method to produce datasets for training and validating. In conclusion, the current model is able to predict 24 hrs-ahead wave heights with the normalized root mean square error (NRMSE) of 8.6 percent and the correlation coefficient (CC) of 0.96.