Methodology for Incident Detection in Alarm Management of Power Systems

Grid operators are confronted regularly with an increased rate of incoming events of the alarm management system, called an alarm flood, decreasing their situational awareness. Therefore, intelligent alarm management is necessary to support grid operators in analyzing the alarm events. In this paper, we present a data-driven methodology to detect incidents based on incoming alarm events. The basis for our data-driven methodology are historical alarm event logs and incident reports from a German distribution system operator. We introduce an incident event log, in which the alarm event log is annotated with the incident cases from the reports by location and time. Our detection metric is based around the relative frequency of event messages, calculated specifically for each location and incident type. With this detection metric, we developed a system which is able to detect current incidents at the execution time based on the sum of the metric for each location and incident type. The system can be configured with a detection threshold and an expiration time, where only events not expired contribute to the detection sum. To analyze the system behaviour, we conducted a sensitivity analysis with 2541 parameter. Comparing the detection results to the benchmark incidents, the pareto optimal parameter set, using a threshhold of 1.2 and an expiration time of 36 seconds, resulted in one false negative and 26 false positives.