In 5G cellular networks, a lightweight, online, real-time, and precise user flow prediction can improve the scheduling of adaptive radio and core network resources. We compare six different methods for short-to-medium-time prediction that fulfill these requirements. In our scenario, we predict the flow of users in a small cell indoor network. We then compare the performance of each method to predict the number of users per access point for several discrete time intervals without any intermediate information. To benchmark the performance of each method in live networks, we collected data from an operational WLAN network at the university over a period of one semester. The results show that common Markov-based solutions perform well only on very short prediction horizons. Beyond that, they are even outperformed by naive predictors. Solely the machine learning approach based on a neural network outperforms all other methods for any prediction horizon. This method enables a real-time and precise user flow prediction at the edge nodes of the network, as the complex task of training can be performed centralized and offline.