Unsupervised Event Detection, Clustering, and Use Case Exposition in Micro-PMU Measurements

GraphPMU: Event Clustering via Graph Representation Learning Using Locationally-Scarce Distribution-Level Fundamental and Harmonic PMU Measurements

“…In [12], an unsupervised clustering method is proposed for some specific faults; such as single-line-to-ground versus lineto-line faults. In [7], an unsupervised event detection and clustering method is proposed, which requires solving a mixed integer program. Importantly, the study in [7] is limited to the analysis of phasor measurements from only one D-PMU.…”

“…In [7], an unsupervised event detection and clustering method is proposed, which requires solving a mixed integer program. Importantly, the study in [7] is limited to the analysis of phasor measurements from only one D-PMU. It is inherently unrelated to the idea of taking into account the topological information of D-PMUs in the clustering task.…”

“…None of the unsupervised learning methods in [7], [11], [12] takes into account the information on the topology of the power distribution network. Furthermore, none of them takes into account the availability of harmonic phasor measurements in addition to the fundamental phasor measurements.…”

“…The case where D-PMUs also act as H-PMUs to measure harmonic phasors will be discussed later in Section V. When an event occurs, its impact is simultanously captured by all the D-PMUs at the buses in set M. Let X j i denote the time series of the phasor measurements that are captured during event i by the D-PMU at bus j, where j ∈ M. Similar to [7], such time series is assumed to be a window of the following measurements at a given D-PMU: the perphase magnitude of voltage V φ , the per-phase magnitude of current I φ , and the per-phase power factor PF φ , where φ is the given phase, i.e., φ ∈ {A, B, C}. The reason for using these measurements is to remove the impact of off-nominal frequencies in the phase angle measurements, see [1, pp.…”

“…Hence, although the available D-PMUs provide synchronized phasor measurements at high resolution, such high resolution data availability is limited to only a few locations on the power distribution feeder. 2) Limited Prior Knowledge about Event Signatures: Analysis of events in power distribution systems is difficult, because: 1) such events are diverse and many of them are infrequent; 2) most events are inherently unscheduled; therefore, we do not know when to expect them; 3) the cause of many events and hence the characteristics of their signatures are not known in advance [7]. The key to address the second issue is to conduct unsupervised learning to identify the type and cause of the events.…”