ML-Assisted Optimal Power and GSNR Estimation in Multi-band Elastic Optical Networks

A significant challenge in next-generation intelligent and autonomous optical networks is the rapid estimation of quality of transmission (QoT), particularly in multi-band and low-margin systems. These systems pose additional complexities, primarily stemming from inter-channel stimulated Raman scattering, which affects the gain-loss power and generalized signal-to-noise ratio (GSNR) profiles. GSNR profile calculation depends on various factors, including the loading state of the links, modulation format, launch power, and channels’ bandwidth. These complexities contribute to the time-consuming nature of estimating QoT in such systems. This study employs a semi-closed form model (CFM) based on the (enhanced) Gaussian noise (GN/EGN) analytical equations to estimate the GSNR, generating datasets. Machine learning (ML) models are then trained using these datasets to accurately estimate optical power and GSNR profiles, achieving errors below 0.04 dB for power and 0.1 dB for GSNR in $99 \%$ of cases. The use of ML models is justified by their computational efficiency, aiding in online network management. We adopt the ML models in a power optimization algorithm that maximizes network total capacity. The power optimization using the ML models take 25-50 times lower time, while resulting in a maximum of 0.1 dBm error versus the analytical semi-CFM.