Retentive Time Series: A Scalable Machine Learning Model for Traffic Prediction in Elastic Optical Networks

Accurate traffic prediction is crucial for elastic optical networks (EONs) as it improves network scalability, performance, and operational efficiency. High accuracy is fundamental for performing long-term traffic management and network planning, while time-sensitive predictions allow efficient real-time adaptability. In this study, we propose a new RNN-based model, retentive time series (Ret-TS), which overcomes some of the limitations of traditional RNNs in EONs scenarios. Compared to RNNs, which successively consider data and always suffer from problematic vanishing gradients, Ret-TS can handle long-term dependencies efficiently through parallel computation; hence, it is more suitable for large-scale and real-time applications. In contrast to the Transformer models, Ret-TS is computationally more efficient, with a lower time and memory complexity of O(L), making it more suitable for resource-limited devices. This research demonstrates that Ret-TS is robust for both short- and long-term traffic forecasting with better prediction accuracy and computational efficiency than traditional models based on RNNs and Transformers. Extensive simulations performed on traffic datasets have shown that Ret-TS reduces prediction errors and network blocking probabilities under different traffic loads in three network topologies: NSFNET, Janos-US, and US100. The results confirm Ret-TS’s robustness and scalability, making it an effective solution for real-world applications in modern optical networks, including both telecommunications and data center networks.