WAVE FORCASTING IN INNER BAY AREA USING NEURAL NETWORK BASED ON AMeDAS OBSERVATION DATA

Abstract

 The use of neural networks for wave forecasting has the advantage that it does not require deep domain knowledge and can be implemented at low cost for each region. However, reanalyzed Grid Point Value (GPV) data of weather fields are frequently used as explanatory variables for neural networks that hinder real-time wave prediction. In this study, we constructed a neural network using AMeDAS and NOWPHAS observation data for real-time wave prediction and predicted the wave height and period at Kobe Port. The combination of explanatory variables and SHAP values allowed us to explain, to a certain extent, the contribution of each explanatory variable to the prediction accuracy. The results showed that information on the wind at an appropriate distance on the upwind side was important to improve the prediction accuracy and that the contribution of historical information on waves (wave height and period) was extremely high, and in particular, the wave height could be predicted with sufficient accuracy for practical use.