I'm not quite sure if it would matter how the player numbers are represented.
I am assuming that the Neural Network would be trained on the time, court, surface, rank etc. and not on the player number, so the Player ID would likely be independent of the algorithm.
It sounds like you have a data structure that contains the players and their historical performance and rank, as well as a timetable of matches. So it is likely that the player and timetable information would be fed in as parameters of the Neural Network, but the Player ID would be an external parameter that governs the evaluation process. If this is true, then how this is implemented depends on your evaluation algorithm and not so much the Neural Network.