Mapping-finding input-parameter refinement paradigm for a dynamic multiband optical network digital twin: the Raman amplifier modeling case

Accurate quality-of-transmission (QoT) estimation tools are crucial to building digital twins (DTs) for optical networks. However, the input-parameter inaccuracy deteriorates the estimation accuracy of the physical models. To address this problem, an input-parameter refinement (IR) paradigm aiming at finding the mappings from uncertain parameters to their corresponding true values is proposed. The primary advantage of the IR paradigm, as demonstrated in this study for Raman amplifiers (RAs), lies in its applicability to dynamic optical networks, where system parameters such as loading conditions and optical device configurations are subject to frequent variations. The use of the proposed paradigm to refine the model of RAs is discussed in detail, while its applicability to other types of devices requires further investigation. The inaccuracy of fiber parameters, signal power, and pump power are taken into account. The particle swarm optimization (PSO) algorithm is utilized to address the problem of the coupling of these parameter inaccuracies. Experiments over a C + L band are conducted. In a single-span scenario, results show the proposed IR scheme can lower the physics-based RA model’s mean prediction error from
${\sim}{0.92}\;{\rm dB}$
to
${\sim}{0.20}\;{\rm dB}$
and lower the maximum absolute error (MAE) from
${\sim}{3.09}\;{\rm dB}$
to
${\sim}{1.12}\;{\rm dB}$
. The proposed IR scheme also exhibits high precision when applied in a two-span scenario, indicating its scalability to multi-span optical multiplexing section (OMS) scenarios. Furthermore, we demonstrate that the proposed IR scheme can also effectively enhance the accuracy of machine learning (ML) models. An IR-aided ML-based model training scheme is proposed. It offers significant advantages in scenarios where data collection from real systems is limited. With the proposed IR paradigm, the practical application of both physics-based models and ML-based models can be facilitated in future dynamic multiband optical networks.