Real-time AI-based Anomaly Detection and Classification in Modern Power Systems

The conventional power grid is transforming into the Modern Power System to meet the electricity demands of the twenty-first century. Distributed Energy Resources and increased communication between nodes in the grid unlock new potential for the Modern Power System, but also new avenues for anomalous behavior including physical faults and cyber attacks. This work addresses these challenges by creating an artificial intelligence framework designed with deep-learning neural-networks to detect anomalies. Training data is collected though a comprehensive digital twin of a power system to safely mimic the system behavior during anomalies. Deep learning techniques, including long short-term memory and graph neural networks, are trained under various schemes to detect anomalies including short circuit faults, false data injection attacks, and line-line faults. These frameworks are combined with threshold detection techniques to optimize for real-time detection while preventing false positives. Finally, case studies support the detection of these faults quickly enough to trigger mitigation techniques and improve the resiliency of the Modern Power System.