Research On Workload Prediction Model Of Cloud Computing Center Based On Integrated Deep Learning

With the rapid development of the modern Internet and the rise of cloud computing technology,more and more individual and business users are choosing cloud computing to provide technical and operational support for their products.Because of the advantageous features of cloud computing,such as pay-as-you-go and elastic scaling,users can access computing resources on their own demand from any device at anytime and anywhere,and only pay for the amount they use.These features significantly lower the threshold for users to access general or high-performance computing power.In today's cloud computing centers,how to perform accurate and efficient resource scheduling and guarantee high availability of services for a long time is an important research topic in cloud computing.To this end,this paper hopes to improve forecasting accuracy by combining current cutting-edge methods and techniques in the computer field,starting from several aspects such as forecasting model and data preprocessing stages.In this paper,two workload forecasting algorithms are proposed for single-point and multi-point forecasting scenarios,respectively.The forecasting algorithms are guaranteed to achieve high forecasting accuracy without losing efficiency.The main research work of this paper is as follows.(1)This paper investigates the cloud load prediction algorithms at home and abroad and analyzes the advantages and shortcomings of each scheme by algorithm category.The characteristics and advantages of cloud computing technology are introduced.Explains three cloud computing service models: Iaa S,Paa S,and Saa S.The characteristics of virtual machines and containers are compared.The calculation principle of Recurrent Neural Network and Gated Recurrent Unit network are described in detail.(2)An ensemble deep learning forecasting method combining Variational Mode Decomposition(VMD)and R-Transformer is proposed,which is oriented to single-point forecasting scenarios.VMD is applied to the data preprocessing stage to decompose the original workload into sequences of different frequencies to reduce the instability and noise.The forecasting module consists of R-Transformer and autoregressive model,where RTransformer uses the Local RNN and multi-head attention mechanism to obtain the nonlinear correlation in the sequence.The autoregressive model is used to obtain linear correlations.Finally,a detailed comparison experiment is conducted to verify the effectiveness of the proposed method.(3)A forecasting method based on Stacked Auto-Encoder(SAE)and attention-based GRU Encoder-Decoder(GRU-ED)network is proposed,which is used for multi-point forecasting scenarios.The StackedAutoencoder compresses the workload sequence and extracts the important features in it;the GRU Encoder-Decoder network outputs multiple forecasting results of any future time length to improve the forecasting efficiency,and an attention mechanism is added to this network to make the network focus on the important historical time point information.Finally,the forecasting accuracy of the method is demonstrated by some comparison experiments.