Prediction and combination method of host load based on non-stationary series

A novel index-concurrent connection number was proposed to measure host load. Based on the new index, the self-similarity and non-stationary attributes of host load are analyzed and a novel prediction algorithm is constructed. This algorithm decomposes and reconstructs the host load series by wavelet into one low frequency signal in the largest scale and several high frequency signals at different scales. The low frequency signal is predicted with the AR model; the high frequency signal at the smallest scale with the weighed moving average method and the others with the support vector regression (SVR) models. After one-step-ahead prediction, the predicted results of these signals are combined into the final value based on SVR. Theoretical analysis and experimental results showed that this new algorithm can improve the predictive precision obviously.