Research On Load Forecasting Of Data Center Based On Three-way Decision

Cloud computing technology has developed rapidly with its advantages of high scalability,ultra-large-scale,and high reliability,its essence is an Internet-based computing method that provides users with various software and hardware resources through pay-as-you-go services.In recent years,with the rapid expansion of cloud data center scale,traditional static resource management technology has to led to low resource utilization and high energy consumption.Therefore,data center dynamic resource management technology has become one of the important technologies in the current field of cloud computing research.The resource demand of the data center is in a state of constant change.By accurately predicting the resource demand,resources can be effectively managed to achieve the purpose of improving resource utilization.It has important theoretical significance and practical value to use load forecasting technology to predict the resource demand in the future and dynamically adjust the number of allocated resources.This paper focuses on the load prediction technology of data center in cloud environment.As the traditional prediction method is difficult to adapt to the dynamic change of resource load demand,we have carried out a series of studies on the related factors affecting the change of resource load demand.Firstly,feature vectors are constructed for resource load sequence,SOM neural network is used to classify the load sequence and serve as the input of prediction model,and attention mechanism is introduced into LSTM to predict resource load characteristics at future time.Secondly,modeling is carried out for the characteristics of resource load,and the characteristics of data center resource load are defined as stable period,volatility period and jitter period,and the implementation strategies are targeted according to the characteristics of the three periods.The load prediction technology enables data centers to predict the load state of resources in the future so as to allocate resources efficiently.(1)Since data center resource load characteristics usually show different fluctuation intensities,a cloud load characteristic prediction research based on an attention mechanism is proposed.Firstly,we use data analysis,feature extraction,and other methods to construct the feature vector of resource load requirements.Secondly,the SOM clustering algorithm is used for this feature vector to obtain higher internal similarity and serve as the input of the prediction model.Finally,by introducing the Attention mechanism to the LSTM network,the model can measure the importance of different information features to more effectively predict resource load volatility.(2)Aiming at the high dynamics of data center resource load changes,proposed a combined prediction method for energy consumption optimization.Based on the basic idea of three-way decision,three divisions are made according to the change of load characteristics to set up three different periods of cloud resource requirements and apply targeted strategies.Firstly,according to the three decisions,the load characteristics are defined as three periods: the stable period,the volatility period,and the jitter period.The simulated annealing algorithm is used to learn the thresholds required by the three-way decision models from the historical load sequence,and according to the threshold,the three branches of the data center load are divided.Secondly,with the help of the basic idea of three-way decision,a prediction model is constructed for the load sequences in the three periods respectively,to obtain a joint prediction of multiple load sequences.Finally,the prediction accuracy is further improved by prior error prediction.The prediction model proposed in this paper has been evaluated experimentally.We have been carried out the prediction model proposed in this paper,the experimental results show that the proposed method can effectively improve the prediction accuracy.