In 5G mobile communication systems, millimeter Wave (mmWave) can provide greater bandwidth and higher transmission rates. The mmWave base station can transmit a high gain directional narrow beam through a large-scale antenna array to increase signal coverage. When mmWave base station are densely deployed, users have to search a large number of narrow beams which sent by multiple base stations to find the optimal beam. This process can take a lot of time and computational resources. Therefore, this paper designs a system-level simulation model in a multi-base station scenario, focusing on the layout of users and base stations, large-scale fading, beam-orientation gain, and characteristics of mmWave channels. Finally, a beam search algorithm based on machine learning is designed. Compared with the traditional exhaustive algorithm, this algorithm has lower latency and computational overhead.