Anomaly Detection in Similarly Behaving Solar Inverters

Inverters are the primary cause of ticket creation and energy loss in PV plants. Automatic detection of inverter malfunction can help PV operators optimize maintenance, thereby reducing overall downtime and maintenance costs. To this end, anomaly detection is used on performance data generated by a group of similarly behaving inverters. At any point in time, the first and third quartiles of the data for all the inverters serve as baseline for the range of values that would be generated by a normally operating inverter. Two measures of abnormality based on this principle are thereby devised. One tracking the differences between the distributions of the data generated by the inverter and baseline, while the other tracking the cumulated differences between inverter and baseline data. Finally, the two abnormality measures are combined into a single anomaly score ranging from 0 to 1 for increasing degree of abnormality. A threshold is then set on the anomaly score above which an alarm goes off warning of inverter malfunction. The methodology has been tested on a real PV plant for the specific task of detecting inverter faults yielding 80 and 25 percent detection and false alarm rates, respectively.