Server Load Prediction Based on Wavelet Packet and Support Vector Regression

Abstract

Wavelet packet theory and support vector regression (SVR) were introduced into server load prediction. A novel prediction algorithm called wavelet packet-SVR was proposed. Firstly, the algorithm decomposed and reconstructed the load time series into several signal branches by wavelet packet analysis. Secondly, SVR prediction models were constructed respectively to these branches and finally their predicted results were combined into final load value. Theory analysis and experiments show that wavelet packet transform is the extension of wavelet theory and has better frequency resolution. So it can decompose the original load series into several time series that have simpler frequency components and are easier to be forecasted; support vector regression has greater generation ability and guarantees global minima for given training data, it performs well for non-stationary time series prediction. So the proposed method is superior to wavelet based approach