Solar inverter failure detection, classification and early warning across four plants
Models that detect production-loss events on inverters, classify them by failure type, and flag the intermittent power-loss pattern that precedes longer outages: 8% of production losses avoided in a live pilot.
Challenge
Inverter failures were the operator's main source of production loss. Its SCADA gave no insight into coming failures, could not separate production loss from communication loss, and maintenance was either scheduled or corrective after the event.
Solution
On data from four plants, Arundo built an anomalous-power-pattern model that cross-references irradiance against active power to detect production-loss events and label them by likely cause (IGBT, ground fault and others), and a power-loss monitoring model trained on five days of features to predict a power loss of more than an hour the next day.
Impact
The models detected and classified production-loss events across ten months where only a handful of corrective work orders had been logged, flagged intermittent power loss days before a three-day outage, and gave early indications of IGBT and ground-fault failures. On that basis the models went into a live pilot on the operator's plants, where 8% of production losses were avoided.