Unsupervised, Federated and Privacy-Preserving Detection of Anomalous Electricity Consumption in Real-World Scenarios

Being able to detect anomalous electricity consumption events in residential buildings is one of the main steps when reducing unnecessary energy usage. In this paper, we present an unsupervised approach for detecting such kind of events locally at individual households, while leveraging the knowledge from other households and ensuring the privacy of the consumption. This is realized by learning a model locally on the households’ edge devices for on-the-edge anomaly detection, using federated learning to leverage the knowledge from other households and ensure the privacy of the training data, and using a secure aggregation approach to handle model weight updates, protecting against a malicious central server and avoiding the risk of model inversion attacks. The approach has as advantages that it respects privacy, does not require annotated data, circumvents the cold-start problem, and is flexible with respect to detecting different types of occurring anomalies. We demonstrate its effectiveness on a real-world, highly granular electricity consumption dataset comprising 14 Belgian households during a period of 36 months, and show effectiveness by installing malfunctioning devices in a lab environment.