Edge-Enabled AI for Anomaly Detection in Distributed Energy Networks: A Federated Learning Approach

The increased penetration of Distributed Energy Resources (DERs) in the power grid has led to significant data deluge problems that complicate efficient data storage and management and, the likelihood of cyber threats occurring on inverters and other DER-related assets. Such attack vectors reduce the visibility of DER management systems (DERMS), thereby compromising the confidentiality, integrity, availability, and accountability (CIAA) properties of stakeholders. This paper explores the application of Federated Edge Learning (FEL) based AI-ready testbed (AI-TB), as a solution for anomaly detection in DERs to address these challenges. Multiple distinct DER anomalous signatures were generated and processed at the edge devices. The edge AI-TB comprises NVIDIA nano (client) and NVIDIA AGX (server) devices. Thus, raw data is processed at the dedicated local server, and DER inverter privacy is preserved. Specifically, a Long Short-Term Memory (LSTM) Autoencoder model was utilized to detect DER anomalies on frequency. The preliminary results indicate that the FEL approach yields improved accuracy on unseen datasets compared to the non-federated learning (NFL) method. The FEL method indicated a Mean Absolute Percentage Error (MAPE) of 0.022, while NFL methods achieved a MAPE of 0.023. Additionally, the maximum MAPE for FEL was 0.085, compared to 0.065 for NFL methods.