Abstract: The prediction accuracy of server load was improved with a novel prediction method based on wavelet. The non-stationary load series was decomposed and reconstructed into one low frequency signal and several high frequency signals by wavelet. Then the approximate stationary low frequency signal was predicted using ARMA model; the high frequency signals were forecasted respectively using neural networks that had different parameters. After one-step-ahead prediction, the predicted results of these signals were combined into the final predicted result of the original load series. Experiments results show that this method can predict non-stationary server load series efficiently and has higher prediction accuracy than traditional methods.

Key words: Neural network, ARMA model, Load prediction, Wavelet, Server load

摘要： 为提高服务器负载预测的精度，提出一种新的基于小波的预测方法。该方法首先对具有非平稳特征的服务器负载序列进行小波分解与重构，得到一个低频信号和多个不同尺度的高频信号；对具有近似平稳特征的低频信号建立ARMA预测模型；对变化较多的各高频信号分别建立神经网络预测模型；然后分别对各信号进行一步预测并组合预测结果，获得原始负载的最终预测。实验表明：该方法能够有效预测非平稳的服务器负载序列，预测精度明显高于传统预测方法。

关键词: 神经网络, ARMA模型, 负载预测, 小波, 服务器负载