Research And Implementation Of Network Traffic Prediction Model Based On Spatio-Temporal Series

With the development of computer and mobile communication technology,it is more and more important for real-time prediction and alarm of operator network.In order to predict the changes of network traffic in a timely and accurate manner and reduce unnecessary resource consumption,this paper uses deep learning technology to research and practice the network traffic prediction model,including the following three aspects:Use normalization and other data processing methods to process a large amount of network traffic data into NodeNetFlow dataset.For network topologies of various scales,collect the traffic data in each network node in real time,divide the data into short-term,medium-term and long-term network traffic data according to the length of the time slice,normalize the data and build a network topology adjacency matrix for the training of network traffic models and verification of effects.A spatiotemporal series model based on attention mechanism is proposed to improve the prediction accuracy of the model.In the actual operator backbone network topology,there are primary nodes and secondary nodes.In order to increase the attention of the model to the primary nodes,a spatial attention mechanism is added.By increasing the attention weight of the core nodes in the network topology,the MAE index of the model on the NodeNetFlow dataset is optimized by about 2%;At the same time,there are also critical time points and non critical time points in the whole time series.In order to increase the attention of the model to the critical time,the time attention mechanism is added.By increasing the attention weight of the later part of the series,the forgetting loss caused by the increase of the sequence length in the series is reduced.The MAE index of the model on the NodeNetFlow dataset is optimized by about 1%.A spatio-temporal sequence model based on multi task training is proposed to improve the generalization ability of the model.Aiming at the problem that the model has high prediction accuracy only under a specific time slice length,a novel multi task training method is proposed,which can use traffic data with different time slice lengths for mixed training.The model can use one model to complete short-term,medium-term and longterm prediction tasks,which improves the efficiency to a certain extent,and ensures that the prediction accuracy is acceptable.