Research On Network Traffic Prediction Method Based On ST-LSTM Neural Network

With the rapid development of Internet technology,the network scale is increasing,network applications are growing dramatically,and network behavior is becoming increasingly complex,which poses new challenges for network management and maintenance.Accurate prediction of network traffic in future periods can help operators allocate network resources more rationally and provide better quality of service,and also help assess the carrying capacity of the network and analyze the health status of the network.Real-time anomaly detection can help operators discover abnormal data in the network in time and quickly find the root cause of the problem to avoid unnecessary losses.Therefore,in the face of the increasingly complex network environment,it is of great importance and practical significance for cloud service providers to establish effective traffic prediction models and detect network anomalies in real time.In order to help operators better solve network management issues such as resource allocation and network security,this paper proposes ST-LSTM,an integrated traffic prediction method based on combining Savitzky-Golay(SG),temporal convolutional network(TCN)and long short-term memory network(LSTM),and SR-CNN,an anomaly detection method based on combining spectral residual(SR)and convolutional neural network(CNN).The data used in this study is derived from all page views data under all wiki projects provided by the Wikimedia Foundation.The data is extracted by aggregating all page views under the same wiki project in hourly intervals,and then connecting them in chronological order to form a time series for prediction and analysis.The research contributions of this paper are two main dimensions as follows:Firstly,this work presents network traffic prediction method based on ST-LSTM neural network.The method effectively combines SG filter,TCN and LSTM.Firstly,SG filter can be applied to smooth the time series,which can reduce the noise and anomalies in the series while ensuring the overall trend of the series.Secondly,TCN is a special convolutional neural network that can be used to learn short-term local dependencies and extract high and low frequency information in the sequence.Finally,LSTM is inherently adept at modelling sequence problems and can effectively capture the long-time dependencies in sequences.Meanwhile,for the purpose of accessing the performance of the proposed model in this paper,multiple baseline models,including autogressive integrated moving average(ARIMA),backpropagation(BP),LSTM and TCN,are built for comparison experiments.Secondly,this work proposes anomaly detection method based on SR-CNN.Based on the aforementioned prediction study,this paper further proposes network traffic anomaly detection method based on SR and CNN.The goal of this phase is to detect the collective anomalies in the traffic timing data and alert the operators in time.To address the problem of low robustness of the SR method of manually setting thresholds,this paper manually injects anomalies into the SR-transformed saliency map data and finally generates new synthetic data.The CNN is then used to perform supervised learning on the newly generated synthetic data to learn a decision criterion,which is much simpler than learning directly on the original problem.The SR-CNN method transforms the unsupervised problem into a supervised problem that can adapt to the distribution of the data,greatly improving the accuracy of anomaly detection and reducing the false alarm rate.