Research On Multi-scale Network Traffic Prediction Method Based On Improved DESN

Accurate network traffic prediction is not only helpful to allocating network bandwidth resources and improving network service quality,but also helpful to ensuring network communication security.Therefore,network traffic prediction has always been a research hotspot in the network field.Echo State Network(ESN)has strong nonlinear processing ability and short-term memory ability,and it has been widely used in regression tasks and classification tasks.However,there are still some problems when applying ESN to network traffic prediction.First,network traffic has many characteristics such as multi-scale,non-linearity,and scale-dependence.These complex characteristics need to be considered to improve the prediction performance of network traffic.Second,the edge equipment has limited computing resources while Deep ESN(DESN)with good prediction performance has the large model size and computing requirements.When designing the model,the equipment resources and model cost need to be considered to make the model can be deployed on edge equipment.In view of the above problems,this thesis studies the multi-scale network traffic prediction method based on improved DESN.The details are as follows:1.To improve the prediction performance of network traffic,this thesis proposes a multi-scale network traffic prediction method based on DESN for the multi-scale,non-linearity,and scaledependence of network traffic.First,a multi-scale network traffic prediction model based on DESN(MNTPM-DESN)is constructed.Specifically,the model is designed with a multi-scale parallel hierarchical structure,each layer of which contains a scale prediction sub-model,and each scale prediction sub-model realizes network traffic feature extraction and network traffic prediction on the corresponding scale.Second,a training algorithm of MNTPM-DESN is proposed to train each scale prediction sub-model and ensemble weight.Finally,the proposed method is applied to three actual network traffic datasets.The simulation results demonstrate that compared with the state-of-the-art network traffic prediction methods,the proposed method increases a little running time,but significantly improves the prediction performance of network traffic.2.To solve the problem of high computing cost of MNTPM-DESN,this thesis proposes an optimization method for network traffic prediction model based on knowledge distillation.First,an optimization framework for network traffic prediction model based on knowledge distillation is constructed to optimize the network structure parameters and hyperparameters of MNTPM-DESN.Specifically,policy gradient is used to select the network structure parameters of MNTPM-DESN,and then the student model is generated.In addition,knowledge distillation is used to further optimize the hyperparameters of the student model.Second,the optimization process of network traffic prediction model based on knowledge distillation is summarized,which realizes the optimization of MNTPM-DESN in two stages.Finally,the optimized MNTPM-DESN is also applied to three actual network traffic datasets.The simulation results demonstrate that although the proposed method slightly loses the prediction performance,it significantly reduces the computing cost of the model.