For reliable and efficient network planning and operation, accurate estimation of Quality of Transmission (QoT) before establishing or reconfiguring the connection is necessary. In optical networks, a design margin is generally included in a QoT estimation tool (Qtool) to account for modeling and parameter inaccuracies, ensuring the acceptable performance. In this work, we use monitoring information from an operating network combined with supervised machine learning (ML) techniques to understand the network conditions. In particular, we model the penalties generated due to i.) Erbium Doped Fiber Amplifier (EDFA) gain ripple effect, and ii.) filter spectral shape uncertainties at Reconfigurable Optical Add and Drop Multiplexer (ROADM) nodes. Enhancing the Qtool with the proposed ML regression models yields estimates for new or reconfigured connections that account for these two effects, resulting in more accurate QoT estimation and a reduced design margin. We initially propose two supervised ML regression models, implemented with Support Vector Machine Regression (SVMR), to estimate the individual penalties of the two effects and then a combined model. On Deutsche Telekom (DT) network topology with 12 nodes and 40 bidirectional links, we achieve a design margin reduction of ~1dB for new connection requests.

No abstract

Wavelength dependent EDFA gain ripple has an impact on connection's OSNR performance. We propose a machine learning regression model to estimate the end to end gain ripple penalty and to increase QoT estimation accuracy.

Optical network optimization involves an algorithm and a Physical Layer Model (PLM) to estimate the Quality of Transmission (QoT) of connections while examining candidate optimization operations. In particular, the algorithm typically calculates intermediate solutions until it reaches the optimum which is then configured to the network. If it uses a PLM that was aligned once to reflect the starting network configuration, then the algorithm within its intermediate calculations can project the network into states where the PLM suffers from low accuracy, resulting in a suboptimal optimization. In this paper, we propose to solve dynamic multivariable optimization problems with an iterative closed control loop process, where after certain algorithm steps we configure the intermediate solution so that we monitor and realign/retrain the PLM to follow the projected network states. The PLM is used as a Digital Twin (DT), a digital representation of the real system which is realigned during the dynamic optimization process. Specifically, we study the dynamic launch power optimization problem, where we have a set of established connections and we optimize their launch powers while the network operates. We observed substantial improvements in the sum and the lowest margin when optimizing the launch powers with the proposed approach over optimization using a one-time trained PLM. The proposed approach achieved near to optimum solutions as found by optimizing and continuously probing and monitoring the network, but with a substantial lower optimization time.

We propose physical layer model extensions that capture the performance variations of multivendor transponders and use those with appropriate algorithms to optimize connections' launch powers. We estimated potential SNR improvements upto ~35% for 4 different vendors in a network of 27 nodes.