Abstract: In autonomic database system，workload characterization is a key part.In workload characterization，workload should be classified，then anticipate workload performance.Workload classification uses cluster algorithm.And in cluster algorithm，the typical is the K-means cluster algorithm.But in the K-means cluster algorithm，k should be defined and not changed.This paper makes an improvement in the algorithm，so the k is changed if needed.The result of the experiment shows that the veracity using Cluster algorithm Based on Feature Vectors（CFV） making of forecasting workload runtime is improved.