Federated Learning (FL) is a distributed machine learning type of processing that preserves the privacy of user data, sharing only the parameters of ML models with a common server. The processing of FL requires specific latency and bandwidth demands that must be fulfilled by the operation of the communication network. This article introduces two Dynamic Wavelength and Bandwidth Allocation algorithms for TWDM-PONs: one based on bandwidth reservation and the other on statistical multiplexing for the Quality of Service provisioning for FL traffic over 50 Gb/s Ethernet Passive Optical Networks.