The invention discloses an optimal step length and dynamic model selection based ultrahigh arch dam response prediction method which comprises the steps of: preprocessing data and determining error values and missing values in the data; using the optimal step length and an interpolation method to predict the data, selecting an optimal interpolation model according to absolute errors, and using sample data of the optimal step length and the optimal interpolation method to predict a to-be-predicted stress value reflecting a working state of a ultrahigh arch dam, wherein optimal parameters include the step length and the interpolation method; using a wavelet neural network method to predict the data, classifying original data into a training sample and a test sample, selecting the node number and the iterative frequency of an input layer, using the training sample to train the wavelet neural network, and then using the trained wavelet neural network to predict the test sample; and analyzing a prediction result obtained by an experiment, evaluating a stress prediction model based on the optimal step length and the interpolation method as well as a stress prediction result based on the wavelet neutral network, and determining a stress safety threshold value.