流出解析用ニューラルネットワークにおけるリカレント構造の考察

Abstract

Recurrent type neural networks which are said to be good for time series are commonly used for runoff calculation. In usual recurrent neural networks with 3 layers, feedback from only the second layer (hidden layer) to the first layer (input layer) is set. In this paper, a recurrent structure that has feedback from the third layer (output layer) to the first layer in addition to the usual feedback is used. The neural network with this type of recurrent structure gives much better results in case of extrapolating calculation, that is, estimation of a big flood it has not experienced during the learning.