Data-driven fault detection and isolation in DC microgrids without prior fault data: A transfer learning approach

Abstract

The lack of fault data is the major constraint on data-driven fault detection and isolation schemes for DC microgrids. To solve this problem, this paper develops an adversarial-based deep transfer learning model that can detect and classify short-circuit faults in DC microgrids without using historical fault data. In this transfer learning framework, the knowledge of faults is extracted from the transient features of line currents during normal operating disturbances, which is adversarially augmented and then transferred to a target domain as the labels of faults. With the transferred knowledge, a deep learning model combining convolutional neural network and attention-based bidirectional long short-term memory is trained, which is strengthened by attention and soft-voting ensemble mechanisms. In verification tests, this model reaches a high accuracy of over 90% in classifying various short-circuit faults in a multi-terminal DC microgrid model within a short response time of less than 1 ms. Moreover, it is robust against measurement noises and adaptive to system configuration changes. The test results prove the effectiveness of the proposed method in the protection of DC microgrids without prior knowledge of faults.