Construction and Application of Compacted Wavelet Neural Network Model

Zhike KUANG

Hunan City University, Hunan Yiyang 413000, China

Abstract: Because of several problems, such as randomness of neural network model in predicting the presence, lack of transparency in mechanism, difficulty to determine the initial parameters, the phenomenon of over-fitting and easiness to fall into local minima, this paper presents a compacted wavelet neural network model. The model will transplant the wavelet function to the hidden layer of neural network in place of sigmoid activation function, and uses a randomly determined state command to obtain certain predictions. Finally, on the gas emission wavelet packet-Wavelet network forecasting experimental results, it showed that: compacted wavelet neural network model is possessed of higher speed of training, easier operation, applicability to large quantities of data for training and processing, data adaptability and robustness. It is more convenientthan optimized algorithms such as genetic algorithms, particle swarm algorithm.

Keywords: Compactness; Convergence; Wavelet Function; Reconstruction