Federated learning (FL) is a distributed learning framework, which can be widely applied into wireless networks to improve network intelligence without disclosing users' private data. However, due to the scarce wireless resources, it is inefficient for massive user devices to participate in FL training simultaneously. To address this challenge, in this work, we introduce the client clustering scheme and decompose the training process into two stages. Firstly, we divide user devices into different clusters based on Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and Low Energy Adaptive Clustering Hierarchy (LEACH) algorithms. Then a scoring mechanism considering user devices' locations, velocities, link retention time, and training latency is proposed to determine the cluster head among the devices in each cluster. Finally, simulation results validate the performance of the proposed client clustering method in terms of model accuracy and training latency.