Robust Topology Identification in Distribution Networks Enabled by Latent Low-Rank and Sparse Embedding Feature Extraction

Due to the insufficiency of measurement devices, the topology of distribution networks is not monitored. Recently, the application of machine learning approaches has been examined for identifying the network topology based on the available information in the network, including measurements and pseudo-measurements. Most of these approaches, however, cannot efficiently handle numerous inputs. Consequently, some of the contributing variables have been excluded from the explanatory variables of the applied machine learning approach. In this paper, a latent low-rank and sparse embedding (LLRSE) model is applied to extract a set of representative features with a much lower dimension, from the contributing variables. The extracted features are then considered as the inputs to a deep neural network (DNN) applied as the topology identifier. This reduction in dimension allows the DNN to exploit all available information. The LLRSE model also effectively deals with the uncertainty in data and provides robustness against measurement errors. Results on the IEEE 123-node test feeder demonstrate that by employing the LLRSE model, the performance of the applied DNN significantly improves.