Design and initialization algorithm based on modulus maxima of wavelet transform for wavelet neural network

Wavelet neural network (WNN) has many advantages and it receives wide attention in power system. Based on the problems encountered in application, the paper investigates the defects of WNN. It points out that the characteristics of activation functions of the continuous WNN and the back-propagation (BP) neural network have great differences, while the current continuous WNN uses the random initialization and back-propagation algorithm of BP network. Thus the continuous WNN has poor convergence performance. Referring to the design method of discrete WNN and signal edge detection and reconstruction theory, the paper proposes a design and initialization algorithm for single-input single-output continuous WNN. The novel algorithm utilizes the known data to search the modulus maxima of wavelet transform, then the number of hidden nodes and the initial parameters of continuous WNN can be obtained, which speeds up the training process as well as improves the convergence performance. Case study validates the effectiveness of the proposed method.