Research On Network Traffic Forecast Based On Neural Network

In recent years,with the continuous development of communication network technology,the demand for network traffic is also increasing.Network traffic models and predictions play an important role in analyzing actual network traffic and network performance.With sufficiently accurate predictions,network utilization and performance can be improved.Network traffic is random and non-linear.The traditional linear model is difficult to characterize the randomness and non-linearity of network traffic,and its prediction results have large limitations.With the development of machine learning and deep learning technology,more prediction models can be applied to traffic prediction problems.However,due to the complex dynamic spatial and temporal correlation of traffic and the complexity of long-term traffic prediction over multiple hours,accurate prediction of network traffic is still very difficult.Therefore,howto perform single-step and multi-step accurate prediction of network traffic is the key issue of this thesis.For single-step network traffic prediction,the existing model only considers the time characteristics of the time series and ignores the spatial correlation between regions;or uses CNN to extract the spatial correlation,requiring that the dataset must be divided according to the standard grid,which limits the usage scenarios.In order to simultaneously obtain the complicated correlation between network traffic in space and time and make the model more widely applicable,this thesis proposes a STA-LSTM(Spatial Temporal Attention LSTM),a single-step prediction model for network traffic based on improved spatio-temporal multi-layer attention mechanism and LSTM.This thesis first uses a spatial attention mechanism to enhance the encoder's learning of the spatial correlation of geographic regions.Then the temporal attention mechanism is used to adaptively select relevant time steps for prediction.Experimental results on actual datasets show that STA-LSTM can efficiently obtain the spatio-temporal correlation of network traffic data and reduce prediction errors.At the same time,because the dataset does not need to be divided according to the standard grid form,STA-LSTM can be more widely applied to various traffic datasets.However,in reality,many situations need to predict the network in advance to scientifically distribute network resources.At this time,a multi-step traffic prediction model is needed,and how to reduce the cumulative error of the multi-step traffic prediction has always been a difficult point.This thesis analyzes the effect of existing methods on multi-step prediction of existing error accumulation.Finally,the training method of Scheduled Sampling is used to reduce the mismatch between the actual value used in training and the predicted value used in testing,which is effective to reduce the cumulative error.In addition,in order to solve the problem of insufficient long-term memory ability,this thesis helps the model to memorize the information of a longer period of time by improving the L STM structure by selectively adding past cell states to the current cell state to perform multi-step and long-term traffic prediction.The STAL-LSTM(Spatial Temporal Attention Long-term LSTM)model is finally formed.Experiments show that the STAL-LSTM model can reduce the error of multi-step prediction.The effect is more obvious when the number of prediction steps increases.To sum up,for the network traffic prediction problem,this thesis has studied and improved the single-step prediction and the multi-step prediction separately.The proposed model can improve the overall prediction accuracy.