Incipient Fault Identification in Power Distribution Systems based on Siamese Temporal Graph

In this paper, a method called Siamese Temporal Graph, which captures the temporal-spatial relationship between sampling points in waveform, is proposed for incipient fault detection. The procedure is explained as follows: The original waveform is re-structured into a temporal graph form and mapped by the defined Temporal Graph Convolution. An optimal mapping function is determined by minimizing the intra-class distance and maximizing the inter-class distance. Due to the transfer learning ability, this model trained with simulation data can be used to detect field data. Experiments show that it is much better than other classifiers, bringing a new insight for the incipient fault detection task in power distribution systems.