A Machine Learning Evaluation of Maintenance Records for Common Failure Modes in PV Inverters

Abstract

Inverters are a leading source of hardware failures and contribute to significant energy losses at photovoltaic (PV) sites. An understanding of failure modes within inverters requires evaluation of a dataset that captures insights from multiple characterization techniques (including field diagnostics, production data analysis, and current-voltage curves). One readily available dataset that can be leveraged to support such an evaluation are maintenance records, which are used to log all site-related technician activities, but vary in structuring of information. Using machine learning, this analysis evaluated a database of 55,000 maintenance records across 800+ sites to identify inverter-related records and consistently categorize them to gain insight into common failure modes within this critical asset. Communications, ground faults, heat management systems, and insulated gate bipolar transistors emerge as the most frequently discussed inverter subsystems. Further evaluation of these failure modes identified distinct variations in failure frequencies over time and across inverter types, with communication failures occurring more frequently in early years. Increased understanding of these failure patterns can inform ongoing PV system reliability activities, including simulation analyses, spare parts inventory management, cost estimates for operations and maintenance, and development of standards for inverter testing. Advanced implementations of machine learning techniques coupled with standardization of asset labels and descriptions can extend these insights into actionable information that can support development of algorithms for condition-based maintenance, which could further reduce failures and associated energy losses at PV sites.