Combined photonic time-delay reservoir computing and LSTM based on VCSEL for network traffic prediction

With the progress of Internet technology and the continuous expansion of network scales, the demand for intelligent network management has become increasingly critical. Consequently, accurate network traffic prediction is essential for enhancing network security and optimizing user service experience. Network traffic typically exhibits characteristics of non-stationarity, nonlinearity, and long-range dependence, making it challenging to develop prediction models that can accurately capture these features to achieve precise traffic forecasting. Existing approaches prove inadequate for addressing the complexities inherent in network traffic prediction tasks. To overcome these limitations, this study introduces a prediction model integrating time-delay reservoir computing (RC) with long short-term memory (LSTM) networks. The output layer of the delay RC is replaced by LSTM in this model. The time-delay RC based on a vertical cavity surface emitting laser (VCSEL) can map the input network traffic data to a high-dimensional space, and extract complex features from the data through nonlinear mapping of virtual nodes. An LSTM network can further exploit the temporal characteristics of data and improve the accuracy of prediction. The proposed model can provide high-dimensional feature extraction capability, enhancing the resistance to noise and outliers, while capturing dynamic changes in the time series data, making the model maintain good prediction performance under different network environments. Simulation results show that the normalized root mean square error (NRMSE) is reduced by 19% and 24%, respectively, for prediction of datasets collected in the United Kingdom Academic Network, compared to RC and LSTM. For datasets collected in the American Research and Education Networks, NRMSEs are decreased by 19% and 20%, respectively. In addition, the effects of the number of training rounds and the initial learning rate on the proposed model are also considered.