Original Article

Analysis of fault detection and defect categorization in photovoltaic inverters for enhanced reliability and efficiency in large-scale solar energy systems

¹ Fraunhofer IMWS, Fraunhofer Institute for Microstructure of Materials and Systems, Walter-Hülse-Str. 1, Halle 06120, Germany
² DiSUN, Deutsche Solarservice GmbH, Mielestr. 2, Werder 14542, Germany
³ DENKweit GmbH, Blücherstr. 26, Halle 06120, Germany
⁴ Saferay holding GmbH, Rosenthaler Str. 34/35, Berlin 10178, Germany
⁵ Leipziger Energiegesellschaft mbH & Co. KG, Burgstraße 1, Leipzig 504109, Germany

^* e-mail: Stephanie.Malik@csp.fraunhofer.de

Received: 30 September 2024
Accepted: 27 March 2025
Published online: 27 May 2025

Abstract

This study presents a systematic approach for examining the performance and vulnerability of large-scale, grid-connected PV systems in relation to inverter faults − particularly those linked to insulated-gate bipolar transistor (IGBT) component. The focus is on an interdisciplinary approach, utilising methodologies from materials science, data analysis, statistics, and machine learning to investigate defect mechanisms, identify recurring issues, and analyse their impacts for a system portfolio of 64 MWp. A root cause analysis identified the failure pattern through material diagnostics of several power modules from inverters previously installed in the field. Prolonged exposure to high temperatures led to the degradation of the IGBT semiconductors, resulting in a breakthrough due to the short-term release of excessive heat. In parallel, an impact analysis was carried out based on historical monitoring data, that identified a faulty control behaviour of the inverter during curtailment. Due to the sharp increase in curtailment occurrences, a correlation of this observation was noted across nearly the entire portfolio. Finally, the study explored whether this randomly observed fault pattern that led to the inverter failure could have been detected in the data without prior knowledge of it. To achieve this, a method combining an artificial neural network and density-based clustering was proposed to automatically detect this recurring and propagating error pattern. This process was carried out in three steps: predicting the normal behaviour of the inverter, distinguishing between normal behaviour and anomalous behaviour, and differentiating the anomalous behaviour. The fault patterns were clearly assigned to four clusters. By introducing a scalable, data-driven fault diagnostics method, this study highlights how advanced materials science and data analytics can improve early fault detection and maintenance in PV portfolio monitoring, while also providing a deeper understanding of defect mechanisms. These combined approaches ultimately enhance inverter reliability and operational efficiency.