A Flexible Distributed Building Simulator for Federated Reinforcement Learning

Recently, researches on building control using reinforcement learning are underway to realize a sustainable society. Such building control methods typically use data acquired from sensors and other sources in the building, but such data may contain sensitive information such as human flow, and collecting these in the Cloud for model training can be problematic in terms of privacy and security. With federated learning, it is possible to share and train models for large numbers of buildings while protecting such data. However, it is almost impossible to conduct a model training experiment with federated learning while actually running real buildings. There are several studies of building simulators, but none of them are designed for simultaneous operation of large numbers of buildings, as is the case with federated learning. Therefore, we propose a distributed building simulator for federated reinforcement learning. This simulator is both scalable and flexible, allowing users to conduct experiments scaled to multiple machines without writing complex code for distributed processing. In this paper, we describe the design and implementation of this simulator and demonstrate that it can perform experiments with a large number of buildings, at least 1024 buildings, in a scalable manner.