An advanced metering infrastructure (AMI) system plays a key role in the smart grid (SG), but it is vulnerable to cyberattacks. Current detection methods for AMI cyberattacks mainly focus on the data center or a distributed independent node. On one hand, it is difficult to train an excellent detection intrusion model on a self-learning independent node. On the other hand, large amounts of data are shared over the network and uploaded to a central node for training. These processes may compromise data privacy, cause communication delay, and incur high communication costs. With these limitations, we propose an intrusion detection method for AMI system based on federated learning (FL). The intrusion detection system is deployed in the data concentrators for training, and only its model parameters are communicated to the data center. Furthermore, the data center distributes the learning to each data concentrator through aggregation and weight assignments for collaborative learning. An optimized deep neural network (DNN) is exploited for this proposed method, and extensive experiments based on the NSL-KDD dataset are carried out. From the results, this proposed method improves detection performance and reduces computation costs, communication delays, and communication overheads while guaranteeing data privacy.