Research On Network Traffic Prediction Model Based On Neural Networks In SDN

With the advent of the era of interconnection of all things,network equipment is increasing,network scale is expanding,network traffic management has become increasingly complex.How to improve the efficiency of network management has become the focus of many researchers.Network traffic prediction has a vital impact on network management and maintenance.Accurate traffic prediction results can greatly improve the efficiency of traffic scheduling,anomaly location and attack detection.This paper presents a new online traffic prediction model,LSTM-NW,which runs in SDN network environment.LSTM-NW model firstly decomposes the original traffic data into approximate data and multiple sets of detail data using wavelet transform,and then uses the decomposed approximate data and detail data as input to predict the flow value at the next moment through LSTM neural network.We divide the model into two stages: initial learning and online learning/prediction,and derive the calculation steps of Online learning in detail.The computational complexity of each updating model in the online learning process is calculated,which is low enough to allow the model to run in the SDN controller for a long time and continuously provide high accuracy traffic prediction data for other network management applications.The network traffic has the characteristics of random mutation,and the weight of the neural networks needs to be updated frequently in the online learning process.At this time,The random mutation of network traffic will cause weight oscillation,in order to suppress this problem,this paper also proposes a neural network weight optimization algorithm suitable for network traffic prediction,called CGD algorithm.As a part of LSTM-NW,CGD algorithm enables LSTM-NW model to restrain the oscillation of gradient and avoid the negative impact of network traffic sudden change on the model.At the same time,it can adapt to the changing trend of network traffic pattern and keep good prediction accuracy.In the experiment,we used two datasets,the British Academic Backbone and the European Urban Backbone,and conducted a number of experiments.Firstly,the necessity of using wavelet transform to decompose the original network traffic data before training with LSTM model is verified by experiments.Then,the LSTM-NW model proposed in this paper is divided into several versions under different conditions,which are compared in detail with DBNG model,the best network traffic prediction model at present.For the first time,we divide the test set into several intervals according to the time sequence,and compare the RMSE and MAPE of the prediction results with the real traffic data of various models in different time intervals,so as to judge whether the prediction accuracy of the model can be maintained for a long time.Finally,the experimental results verify the superiority of the LSTM-NW model in various network traffic prediction problems.