ANALYSIS OF YORIMAWARI WAVES USING SENSITIVITY ANALYSIS OF NEURAL NETWORK

Abstract

 In this study, we constructed neural network (NN) models to predict Yorimawari waves at Toyama Bay, and conducted accuracy verification and sensitivity analysis for high wave over 5 meters that occurred after February 2008. From the analysis results, the influence of input-output relation was evaluated quantitatively. Using 3 types of sensitivity analysis methods were examined the applicability of the sensitivity analysis to the large networks to predict of extreme phenomena. As a result of analysis, it is shown that sensitivity analysis is suitable method that can consider time series, and the optimal learning condition of NN and influence of input-output relation differ according to the route of the bomb cyclone. It has become clear that it is necessary to learn by changing the composition of the model in the route developed around the Japan Sea or the Japanese archipelago and the route developed in the Pacific Ocean.