OPTIMIZATION AND SENSITIVITY ANALYSIS OF TRAINING DATA ON YORIMAWARI WAVE PREDICTION BY NEURAL NETWORK

Abstract

 In this study, we investigated the predictability and accuracy of high wave estimation at Toyama Bay using the neural network (NN) with the Japanese 55-year Reanalysis data (JRA-55). A hierarchical NN that outputs significant wave height or period from the atmospheric data compiled from multiple 10 days’ events including Yorimawari-wave is developed. The NN model is optimized by the error of peak wave height and the correlation coefficient of the time series of wave height. In addition, the process of high wave generation is extracted and discussed from the results of the sensitivity coefficient method to the NN model optimized to reproduce Yorimawari-wave on February 23, 2008. As the results of analysis, the optimum learning conditions of Yorimawari-wave have 5 units at hidden layer, 4 to 6 events (a total of 160 to 240 hours of weather data), 13 hours lead time, and teacher data including the cyclones passing through the Japanese Islands. Optimized NN model reproduces the significant wave height and period of the large Yorimawari-wave on February 23, 2008.