Research On Hybrid Network Traffic Prediction Model Based On Mode Decomposition And Neural Networks

In recent years,people have increasing demands for such online services as instant messaging,search engines,social entertainment,telecommuting,online transactions,and public services,leading to explosive growth in the scale of online services.Technological progress and user needs have made the network more diversified.However,due to limited network resources,the continuously increasing network demands will inevitably lead to network congestion and reduce the service quality.Therefore,it is necessary to comprehensively monitor the status of the network so as to make network management timely and effective,thereby improving network performance.Since network traffic can reflect the overall or partial status of a network,network traffic prediction is helpful for network maintenance,network optimization,routing strategy design,load balancing,protocol design,and anomaly detection.Therefore,the research on network traffic prediction has attracted extensive attention from domestic and foreign research scholars and industry.However,the self-similarity,periodicity,chaos,and other characteristics of modern network traffic make it challenging to predict network traffic.In this dissertation,we mainly investigated the model construction of high-accuracy and low-time complexity network traffic prediction.The main content and innovations are as follows:Firstly,most of the existing traffic prediction models only focus on the self-similarity and burstiness of traffic,and only use the network traffic values as input,thus lacking a more comprehensive description of network traffic characteristics.Therefore,using periodic attribute information to assist the prediction of network traffic values can undoubtedly improve the accuracy of the prediction models.In addition,most of the existing prediction algorithms require real and effective traffic data as input.However,due to the complexity of the network topologies,the resource limitations of network equipment,and the high cost of monitoring high-speed networks,it is impractical to collect all real traffic data in the actual network,which also affects the robustness and stability of the prediction model.Therefore,reconstructing the network traffic data to make the input data to the prediction model more similar to the actual network traffic is helpful in improving the prediction accuracy.Taking the above factors into consideration,on the basis of the model RNN-VTD(Recurrent Neural Network that uses traffic value,timestamp,and day-of-the-week as input)proposed in the comparative literature,a network traffic prediction model based on Empirical Mode Decomposition(EMD)and Gated Recurrent Unit(GRU)neural network combined with data reconstruction is proposed in this work.This model is called GRU-VTD-RC-EMD,where RC means reconstruction.The model reconstructs traffic data through normalization,missing point completion,and outlier elimination and then decomposes the network traffic sequence into several components through the EMD algorithm.Then these components are used to train the corresponding GRU-VTD neural networks,and the final prediction result is obtained by combining the outputs of all GRU-VTD neural networks.The simulation results show that,compared with the RNN-VTD model,GRU-VTD-RC-EMD can effectively reduce the prediction errors.Secondly,compared with single prediction models,the hybrid prediction models have higher prediction accuracy,but their time complexities are also higher.To solve this problem,a network traffic prediction model based on Variational Mode Decomposition(VMD)and GRU neural network combined with data reconstruction is proposed in this dissertation,which is abbreviated as GRU-VTD-RC-VMD.Moreover,the prediction accuracy and time complexity of the GRU-VTD-RC-VMD model and the GRU-VTD-RC-EMD model are compared in the data sets with different scales and different data sets with the same scale to verify the feasibility of the proposed model in practical applications.The experimental results show that,compared to the GRU-VTD-RC-EMD model,the GRU-VTD-RC-VMD model can obtain higher prediction accuracy and lower time complexity in small-scale data sets when the number of components is small.