Physics-Informed Closed-Loop Python–OpenDSS Co-Simulation for Fault Detection and Localization in Inverter-Dominated Microgrids

Increasing penetration of inverter-interfaced distributed energy resources (DERs) presents new challenges for fault diagnosis in distribution networks because DER operating conditions can modify the electrical signatures available to protection and diagnostic algorithms. This paper presents a physics-informed, closed-loop Python–OpenDSS co-simulation framework for machine-learning-based fault detection, fault-type classification, and fault localization. OpenDSS provides the physics-based network solution, while Python performs automated scenario orchestration, electrical-feature extraction, dataset construction, model development, and diagnostic evaluation. The framework was evaluated using the IEEE 13-node, IEEE 34-node, and IEEE 37-node benchmark feeders, with the IEEE 13-node feeder additionally serving as the detailed PV/BESS sensitivity and binary fault-detection case study. A scenario-level data-partitioning strategy was implemented to ensure separation of physical scenarios between model development and final held-out evaluation. Under the evaluated IEEE 13-node conditions, the binary fault detector achieved 100.00% held-out balanced accuracy and 100.00% macro-F1, although the healthy-state test population was limited. For fault-type classification, the supervised models achieved held-out accuracies ranging from 83.33% to 100.00% across the benchmark feeders, whereas the highest verified fault-localization accuracies were 66.67%, 83.70%, and 61.90% for the IEEE 13-node, IEEE 34-node, and IEEE 37-node feeders, respectively. Matched DER operating cases on the IEEE 13-node feeder further demonstrated measurable changes in network-level fault signatures, with mean absolute changes in maximum phase-current magnitude of 27.78 A for PV operation and 39.41 A for combined PV–BESS operation relative to the no-DER condition. The results demonstrate that diagnostic performance is task- and feeder-dependent and that DER operating conditions measurably influence the network-level fault signatures used by the diagnostic models. The conclusions are limited to the phasor-domain network behavior represented by OpenDSS and do not extend to detailed electromagnetic-transient inverter or converter-control dynamics.