Agnostic Prediction of Multiple Types of Time-Varying Traffic in Optical Networks

Relentless competition among communications ser-vice providers, increasing expectations of users, and escalating variety of new applications and services trigger the need to develop advanced solutions that can support the optimization and management of communication networks. Prediction of network traffic is one of the possible solutions providing additional data analytics information. In this paper, we consider an application-aware optical network that transmits various types of time-varying traffic. We analyze traffic prediction under two scenarios. As a reference scenario, we assume that the system is aware of multiple traffic types, and various prediction models are developed and trained for each traffic type separately. The second scenario - agnostic prediction - assumes that a single prediction model agnostic of the traffic types is created and trained. We develop several models for each of the analyzed scenarios using various regression methods. Next, we run extensive numerical experiments on real and semi-synthetic datasets to verify the performance of the proposed regression methods and compare both analyzed scenarios. The obtained results demonstrate that the proposed prediction model agnostic to the forecasted type of traffic provides excellent results; in many cases, it outperforms the reference scenario with dedicated prediction models for each traffic type. Moreover, we evaluate the proposed model's adaptability to predict unseen-before traffic types, showing that the quality loss is negligible. Finally, we test the proposed framework in a multilayer network with time-varying traffic and show how using an aggregated model does not lead to bandwidth blocking increase compared to dedicated prediction models.