To support the emerging 5G applications and the 5G bearer networks, optical networks, as the critical infrastructure, are continuously evolving to be more dynamic and automatic. The vision of future autonomous networks with low link margins requires precise estimation/prediction of the quality of transmission (QoT) of optical links. Machine learning (ML) technologies provide promising solutions to predict QoT of unestablished links. In this paper, we investigated hybrid modelling and transfer learning to address the key issues for deployment of ML applications in optical networks. The proposed approach for multiple-channel prediction reduces the training data requirement by 80% while obtaining the same MSE of 0.267dB compared with the model without transfer learning. The approach facilitates a streamlined ML life-cycle for data collection, training, and deployment.