On-demand resource provisioning is with great challenge in cloud systems. The key problem is how to learn about the future workload in advance to help determine resource allocation. There are various prediction models developed to predict the future workload. The major problem of pre vious researches is that they assume that application workload has static pattern. In practice, so many application workloads have hybrid dynamic pattern overtime. To achieve high prediction accuracy, we find that it’s essential to detect both workload pattern stage and the changes in the model parameters. In this paper, we present a Pattern Sensitive Resource Provisioning Scheme, named PSRPS. It can recognize application workload patterns and choose suitable prediction models for prediction online. Besides, when there is maladjustment in prediction models, PSRPS can switch prediction models or adjust the parameters of the model by itself to adaptively to guarantee prediction accuracy.