The Applied Research Of BP Network Combination Forecasting On Network Traffic Prediction

With the rapid development of the internet and its applications,network size is growing,network applications' complex is also increasingly.Since the Internet is a highly complex nonlinear system,in order to achieve reliable data transmission research and rational resource allocation,network control mechanisms and complex behaviors characteristic is particularly important.Network traffic prediction analysis is important for network performance evaluation.Critical network traffic prediction is normal behavior of network traffic through a description of the network traffic to analyze future trends in order to achieve predictable and alarm.The relatively mature network traffic prediction method features include time-based prediction,forecasting method based on support vector machine,model-based prediction methods seasonal,prediction method based on wavelet transform,each method has its own characteristics and limitations sex.For network traffic prediction,this paper presents an improved combination forecasting model,the main work is as follows:First,a single application of BP network prediction only suitable for solving stationary random time series prediction,BP neural network here will combine local relevance vector machine forecasting methods and support vector machine prediction methods to obtain a new combination forecasting model to broad its range of applications.BP neural network learning process includes forward propagation and reverse propagation,the two prediction method prediction value as the training sample,first calculate the output results and error,if the error is too large,then reverse the adjustment threshold and connection weights,thirdly recalculated,until the error reaches a certain standard.BP algorithm is a self-learning process,generally selected based on past experience learning rate,with the progress of BP learning algorithm,it is difficult to guarantee the effectiveness of proposed,so in this paper adjustment adaptive learning algorithm and adding momentum are proposed to select appropriate learning algorithm BP rate,in order to improve the performance of BP algorithm.Adaptive learning rate adjustment rule is: correction weight values detected whether to reduce the error,error increases or decreases,the learning rate will be reduced or increased by a certain multiple of,otherwise remain unchanged,until the learning process is stabilized.Secondly,in the BP network combination forecasting one of the input--SVM predictor,cuckoo search algorithm is modified to optimize SVM penalty factor and nuclear width.In order to solve the relationship between global optimization capability and accuracy between the different stages of the search results,the step size adaptive dynamic adjustment,and build predictive models.By contrast experiments,proving MCS-SVM algorithm is better than genetic algorithm,particle swarm optimization to predict the performance of the algorithm.Finally,the design of the network traffic data analysis model,based on the network NetFlow data collection and data collected for the prediction model for analysis.Experimental results show that compared to a single prediction L-RVM model and MCS-SVM model to predict performance based on non-linear combination of BP network can be effectively improved.