Optimization and Application Research of Wavelet Neural Network

Abstract

In allusion to the problems that the conventional wavelet neural network has disadvantages of training slowly, convergence to the local minimum easily and poor approximation performance, two aspects including initial parameters selection and network training methods were selected to be optimized after analyzing its approximation performance. A kind of self-adaptive method to get the number of hidden layer nodes was put forward. And the WNN model based on SCG optimization algorithm was constructed, combining with SCG algorithm and the method of setting the initial parameters based on self-correlation. The model has been used to predict the settlement of high-rise building foundation under complicated geological conditions, and the results showed that the model not only solved the problems of approximation performance very well, but also is better than both of the BP neural network and the conventional WNN based on BP algorithm.