The Research Of Network Traffic Prediction Based On Improved Gated Recurrent Neural Network

Network traffic is an important indicator of how well the network is performing.According to the nature of network traffic,a reasonable model is established to predict network traffic,and corresponding measures are taken in advance to play a major role in network security,network management,and network resources.However,the nature of network traffic is very complex and diverse,the most important of which is long-term correlation,which makes the selection of predictive models extremely difficult.At present,the recurrent neural network is very suitable for the prediction of network traffic due to its powerful learning ability and unique memory function for nonlinear sequences,especially the long short-term memory network and gated recurrent neural network.Representatives,they introduced the door structure to control the flow of information,which can capture the long-term correlation of network traffic,so the prediction accuracy is very high.However,complex gating structures result in long training times.In addition,as deep learning models,they usually have a large number of layers,which also lengthens their training time.Based on these serious shortcomings of long short-term memory network and gated recurrent neural network,after careful analysis and in-depth research,this paper mainly made the following work:1)In this paper,the principles of long short-term memory unit and gate recurrent unit are deeply analyzed.The long short-term memory unit specifically uses a cell state to preserve long-term memory,and uses the forget gate to control the deletion of historical information,the input gate for new information addition,and the output gate to determine how much information from the cell state is used as output of the hidden layer;Gate recurrent unit uses a reset gate to control the loss of old information,and the update gate is used to control the addition of new information,so both of them can capture the long-term correlation of network traffic.2)Based on the above principle analysis and the nature of network traffic,a network predictive model called stacked bidirectional and unidirectional minimal gated recurrent networks is proposed for network traffic prediction.Its hidden unit is the latest gate recurrent unit.It only retains the update gate to simultaneously control the forgetting of old information and the addition of new information.It can maintain long-term memory and its internal structure is very simple,so the training is very fast.the model is confirmed by experiments on two network traffic data sets.Its training speed is much faster than the existing model.3)A new hidden unit is designed,which not only has the ability to maintain longterm memory,but also can get rid of the burden of the depth of the deep learning model on the training speed.In this paper an improved gated recurrent network for network traffic prediction is proposed.The network traffic prediction model is then carried out on two data sets.The experimental results show that its training speed is further improved than the training speed of the stacked bidirectional and unidirectional minimal gated recurrent networks,and the accuracy is slightly improved.