Spatiotemporal traffic matrix prediction: A deep learning approach with wavelet multiscale analysis

Network traffic analysis has always been a key technique for operating and managing a network. However, due to some (non) technical issues, it is not trivial to directly obtain network-wide traffic data. Although a large number of traffic matrix (TM) prediction methods have been used to obtain future network-wide traffic, they achieve somewhat limited accuracy due to neglecting spatiotemporal evolution features of TM series at different time scales. In order to improve the performance of TM prediction, we propose a novel end-to-end deep neural network based on wavelet multiscale analysis, called WSTNet. In this network, the original TM series is first decomposed into multilevel time-frequency TM subseries at different time scales by using discrete wavelet decomposition, and then the convolutional neural networkwithout pooling is used to extract the spatial patterns among traffic flows, and finally, the long short-term memory neural network with a self-attention mechanism by relating different positions of input sequences across entire time steps is employed to explore the temporal evolution features within TM series. To investigate the performance of our proposed model, extensive experiments are conducted on two real network traffic data sets from the Abilene and GEANT backbone networks. The results show that WSTNet is significantly better than the other four state-of-the-art deep learning methods.