Research And Implementation Of Network Traffic Prediction System Based On Deep Learning

With the development of society,network has a wide range of concepts.Although different networks have different business meanings,such as communication networks,traffic networks and so on.They can be ed as data structures in the form of grid or graph.These forms are easy to carry out related research on the network.In recent years,due to the rapid development of science and technology,people's demand for business is also increasing year by year.Network traffic prediction is an important and popular direction in network research.Timely and accurate prediction of network traffic can facilitate people's daily life,reduce unnecessary loss of time and resources,and also help managers' command and scheduling.However,the complex spatial-temporal relationship in the network brings challenges to this research.Therefore,in order to model the dependencies of temporal dimension and spatial dimension at the same time,realize the integration and interaction of spatial-temporal data,this thesis carries out the research and implementation of network traffic prediction system based on deep learning.This thesis proposes a Multi-Channel Spatial Temporal(MCST)model to predict the traffic value based on the grid data.MCST is composed of multi-time convolution neural network and long short-term memory network.Multi-time convolution neural network uses multiple aligned convolution neural networks to realize spatial dependencies modeling in the same layer and complete proximity and periodicity modeling between layers,and also carry out the data smoothing operation.LSTM receives the output of convolutional neural network and models the dependencies of time dimension.Experiments are carried out on Milan communication network traffic dataset.The experimental results show that the proposed method performs best among all baseline methods over different proportion of train set.This thesis proposes a Spatial-Temporal Convolution Network(STCNet)to predict the traffic value based on the graph data.STCNet is composed of multi-receptive field temporal block and global-aware spatial block.The temporal block is responsible for adjusting the size of receptive field through the dilation rate,so as to model long-term periodic dependencies and short-term proximity dependencies.The spatial block uses graph convolution operation to realize the spatial node position relationship modeling through neighbor aggregation.Experiments are carried out on PeMS-BAY,and the proposed method has achieved excellent performance in all evaluation metrics.In addition,in the performance comparison of different days,the robustness of the method in the time dimension is verified.Meanwhile,the parameters,components and other experimental analysis are also carried out.Finally,in order to realize the visualization of traffic data,this thesis constructs a network traffic prediction system,which can realize the query of historical traffic and the prediction of future traffic.It provides a clear interactive interface and is convenient to use.