Network Traffic Prediction Based On Deep Learnin

Network traffic forecasting plays an important role in network planning and traffic management,and helps network providers effectively allocate network resources accordingly,avoiding overload or inefficient use of network resources,thereby improving overall network efficiency and reducing the risk of network congestion.In addition to the above significance,network traffic forecasting also provides an important basis for base station construction.However,the current methods proposed in the field of network traffic forecasting only consider the time factor in the forecasting process,the sensitivity of the model to data jumps is not enough,and the method that considers the space factor captures the time-space characteristics step by step.Based on the above background,this paper transforms the network traffic forecasting problem in the time domain into a combination of time domain and frequency domain for forecasting.The main research contents are as follows:1.Weighted connection prediction by CEEMDAN decomposition and ST-GCN model.First,CEEMDAN decomposes the data set to obtain the n-order eigenmode function and a residual item of the data set,and then confirms whether the two sequences are related according to the correlation between the residual items obtained after different sequence decompositions There is an influence relationship,if there is,establish an edge connection between the two,and then calculate the similarity of the intrinsic mode function on the two sequences with edge connection as the weight on the connection,through the ST-GCN model Make predictions.2.The combination of ConvLSTM model,self-attention mechanism and 3D convolution is used to capture spatio-temporal features and improve the sensitivity of the model to time series.After the operation described in 1,although the flow prediction problem in the time domain is combined with the frequency domain for processing,the spatial correlation in the physical sense between regions is also lost.Therefore,the ConvLSTM model is used here to collect spatio-temporal features,and then a self-attention mechanism is added to it to strengthen the model’s sensitivity to the self-dependence of the sequence in the time dimension,and at the same time,the 2D convolution is replaced with a 3D convolution to achieve synchronization.Capturing the role of spatiotemporal characteristics of data.3.This study also proposes an integrated model of the above two methods,which combines the advantages of the two methods.Experiments on the network traffic data set of Milan in the Telecom Italia Challenge show that the method proposed in this paper can effectively improve the prediction accuracy and the sensitivity to data jumps.