Optical Power Spectrum Prediction Using Cascaded Learning With Uncertainty Propagating Noisy Input Gaussian Processes

We present an uncertainty propagating cascaded learning approach to multi-span optical link power modeling using noisy input Gaussian processes (NIGP). This is done by defining each individual Erbium-doped fiber amplifier (EDFA) as an distinct NIGP model, allowing the uncertainty output by one model to be input into the subsequent model. The theory behind standard and noisy input GPs are outlined in addition to the application of this model in a cascaded setting. We then experimentally validate the approach on a 5-span link with 6 EDFAs where the standard GP model are shown to outperform the artificial neural network models in both cascaded and end-to-end scenarios. Additionally, our uncertainty propagating cascaded model using the improved NIGP model is shown to outperform the standard GP approach to achieve a mean absolute error of 0.12 dB. Moreover, our uncertainty propagating approach requires 86% less training data than standard GP while providing improved prediction confidence by 22% for partially-loaded configurations.