Relationship Between Training Data and Error in Tsunami Arrival Time Prediction Using Machine Learning

It is important to predict a tsunami immediately after an earthquake has occurred, and it is also important to know the prediction error in advance. In this study, a neural network (NN) capable of instantaneous prediction by learning the input and output relationship in advance was used to predict the tsunami arrival time from the initial water level immediately after an earthquake has taken place. A total of 2,400 data sets were examined, and NN models were created for each earthquake scale considering the variation of the data. As a result, the errors were different for each earthquake scale and number of data, and the overestimation and underestimation differed in the three target regions considered. For the proposed model, variation in the output of the training data has a large impact on the prediction results. Thus, when the variation doubles, more than twice training data are necessary to achieve the same accuracy.