Machine learning algorithm for inverter fault monitoring, prediction

Researchers at the University of Lisbon in Portugal have developed a machine learning algorithm that classifies and predicts inverter failures in utility scale PV plants.

The new algorithm monitors, in particular, the inverter subsystems and sends alarms when maximum and minimum values are reached. It analyzes data and categorizes variables according to historic values.

The scientists identified the types of failures according to the errors obtained in the inverters and the description of occurrences. Failures include grid faults, grid overvoltage, temporary grid overvoltage, grid undervoltage, low voltage, temporary AC overcurrent, grid overfrequency, grid underfrequency, grid power failure, excessive stray current, supply grid fault, 10-minute grid overvoltage, output overload, and unbalanced load of grid device fault.

The group tested the proposed approach on two ground-mounted PV systems with capacities of 140 kW and 590 kW. Both installations rely on inverters provided by German manufacturer SMA. “The variables of each inverter were analyzed, and the following types of failure were verified in the case of the variable yield, due to the inverter errors,” it explained.

The data were characterized via fine tree, medium tree, and coarse tree prediction models. In tree-based models, a set of splitting rules actively divides the feature space into multiple smaller, non-overlapping regions with similar response values.

The academics claim that the proposed algorithm is able to identify seasonal variations in inverter failures and that the results it provides can be used for reliability analysis. “The data-driven evaluation developed in this study indicates that inverters subsystems emerge for categorizing failure modes,” they stressed.

They also suggested protecting inverters from inrush and overcurrent automatically by using clamp circuits to the resonant capacitance in parallel. “High power-conversion efficiency can be achieved by regenerating the clamp current to the input voltage source,” they concluded.

The novel algorithm was presented in the study “Machine learning for monitoring and classification in inverters from solar photovoltaic energy plants,” published in Compass in Solar.