Synthetic Data in DC Microgrids: Label Creation for Ensemble Learning for Fault Isolation

The difficulty in acquiring fault label data is a major obstacle to the application of data-driven fault isolation in DC microgrids. To remove this barrier, this paper introduces an approach of generating synthetic data with the line currents measured during normal operation as the substitute for fault label data in training an ensemble model, which is aimed to isolate line short-circuit faults in DC microgrids. Based on a high-frequency model of DC microgrids, we prove that the line currents during the closing of unloaded DC lines have similar high-frequency features as those during a line short-circuit fault. On this basis, the synthetic data are obtained through performing discrete wavelet packet transform on the line currents measured during the circuit breaker operation. With normal operating data and synthetic data, an ensemble model is trained as the fault classifier. In the verification tests of different fault scenarios in a three-terminal DC microgrid model, the detection rates of the proposed ensemble model are over 90% while the false positive rates are below 0.5%. These results prove the effectiveness of using synthetic data as the fault labels in the ensemble learning-based fault isolation in DC microgrids.