Anomaly Detection
Parts 5 and 6 rest on two assumptions that require verification in the operation of a distribution network: that the readings of the meters are correct, and that each transformer violates its limits in the manner that its violation pattern predicts. Meters are subject to faults in which readings freeze, drop out, drift or are under-recorded, and some meters are tampered with. Inverters are subject to faults that produce no alarm at the meter, in which export is interrupted into short episodes or shifted in time. A transformer may also depart from its violation pattern as the adoption of Photovoltaic (PV) and Electric Vehicle (EV)s increases. Each of these conditions invalidates one step of the analysis of the previous parts. A faulty meter corrupts the clusters of Part 5 and the case base from which Part 6 retrieves, a violation outside its violation pattern receives an unsuitable mitigation measure, and a cluster model fitted before a change of behaviour assigns customers to the wrong strata.
Anomaly detection provides the validation of these inputs. Its methods originate in statistics and machine learning, in which outlier detection, novelty detection and change-point detection have been studied for several decades. This part adapts them to the Low Voltage (LV) network at three levels, the meter, the transformer and the violation event, and evaluates every detector against anomalies whose presence is known, either because they were injected into measured readings or because the power-flow simulation records the violations that occurred.
Chapter 1 states the theory in the notation of the book and illustrates each method on a synthetic metered population whose anomalies are known. It treats the kinds of anomaly, the assumptions of each family of detectors, the representation of a day by its readings or by its events, the conversion of an anomaly score into a decision with a guaranteed false-alarm rate, the explanation of a flag, the distinction between an anomaly and a change of normal behaviour, and the evaluation of a detector. Chapter 2 applies these methods to three measured populations, the AusNet households, the households of the Low Carbon London trial and the meters of the GoiEner cooperative, and to problems that a distribution network operator faces with their readings. It treats the validation of meter readings, thresholds per demand cluster of Part 5, the detection of PV and of EV charging behind an import meter, and the verification of the response of customers to a price event, using the labels these datasets hold where they exist. Chapter 3 treats non-technical loss through the energy balance between the meters and the transformer, inverter faults detected from the timing of export, the effect of faulty readings on the segments of Part 5 and the decisions of Part 6, and an indicator of when the models of Parts 5 and 6 require refitting. Chapter 4, in preparation, addresses the network: coincident export, which is not detectable at any single meter, violation events that depart from their violation pattern, the change of violation pattern with adoption, and the anomaly score as a feature for the ranking of live violations.