COMMITTEE CHAIR: Dr. Daniel Doe
COMMITTEE CO-CHAIR: Dr. Anthony Hill

TITLE: A HYBRID PHYSICS-GUIDED AI FRAMEWORK FOR SOURCE-AWARE ANOMALY DETECTION AND DIAGNOSIS IN THREE-PHASE PHOTOVOLTAIC SYSTEMS

ABSTRACT: As photovoltaic (PV) systems progressively integrate into modern power architecture, reliable monitoring and fault detection methods are essential to ensure efficient and stable operation. Customary monitoring systems constantly rely on simple data visualization or threshold-based alarm, which sometimes fail to detect subtle abnormalities in the system’s behavior. This research proposes a physics informed ground truth framework for anomaly detection in PV power systems. It will exemplify a system designed to integrate statistical analysis with electrical power system constraints. The proposed methodology uses operational data collected from a three-phase photovoltaic system. The specific features utilized are measurements of phase voltages, currents, power, solar irradiance, and cell temperature. This data analysis is performed in Python, where statistical anomaly detection techniques are applied to find unusual operating conditions within the time-series dataset. To make detection more reliable, the statistical results are further confirmed using Physics-based ground-truth equations derived from three-phase power system principles. Beyond detection, the framework includes a source-aware classification stratum that distinguishes whether the anomalies originate from environmental factor, like fluctuations in irradiance, variations in temperature, and electrical faults within the system. Results are attained by analyzing the relationships between environmental inputs and electrical outputs to identify physically consistent patterns associated with each anomaly type. To further evaluate the effectiveness of the proposed system architecture, PSIM simulations are used to model photovoltaic system behavior under both normal and abnormal operating conditions. The simulations tested are scenarios varying in environmental disturbances and electrical faults. Anomalies identified are cross-referenced with simulated system responses to validate the framework’s ability to correctly identify both the presence and source of abnormal behavior. The results show that combining machine learning with the validity of Physics-based equations and source-aware classification can significantly improve the accuracy and clarity of anomaly detection in photovoltaic systems. The proposed structure provides a reliable way for identifying abnormal system conditions. It also offers a scalable foundation for developing more intelligent monitoring of three-phase solar energy systems.

Keywords: Anomaly Detection, machine learning, photovoltaic systems (PV), renewable energy systems, time-series analysis, physics-based validation

Room Location: Electrical Engineering Conference Room 315D