CEEMDAN-SE based Multi-predictor hybrid model for cloud resource prediction

Effective prediction of cloud resource demand can provide decision support for cloud service providers’ resource allocation, which helps optimize resource utilization and improve service quality. In cloud data centers, the time series of cloud resource demand is affected by both explicit change trend and multi-dimensional recessive factors. The existing prediction models are difficult to meet the cloud resource prediction demand under complex change scenarios. In this paper, we propose a sample entropy-optimized Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN-SE) based multiple predictor hybrid model for cloud resource prediction. Firstly, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise is used to decompose time series into set of subsequences. Then the sample entropy is used to evaluate and reconstruct the subsequences set, and the original time series is divided into two parts: the low frequency trend component and the high frequency fluctuation component. Finally, we used multi-predictor hybrid model based on PatchTST and Informer to train and predict trend component and the fluctuation component respectively, and the prediction results were synthesized to obtain the final predicted value. We conducted comparative experiments on the open source dataset Google-Cluster-v2011, comparing it with popular advanced models. The results show that our model has better performance than other models in predicting the time series of cloud resources.