In recent years, battery-backed energy storage systems have received significant attention due to the proliferation of renewable energy. Monitoring energy storage systems and identifying abnormal batteries is essential for their usability and dependability. Due to the scarcity of anomalous measurements and privacy concerns, we collectively train a global autoencoder to identify anomalous batteries over multiple battery-backed energy storage systems. However, the repetitive model weights exchanges during federated learning induce large communication overhead. To address this issue, this poster presents an efficient communication protocol to reduce the number of bits to represent each model parameter, decreasing the transmission delay. Experimental results show the effectiveness of the proposed communication protocol.
Article ID: 2023GL3
Publisher: Canadian Artificial Intelligence Association