Robust and fast detection of small power losses in large-scale PV systems

Abstract

Due to the fast growth in global installed photovoltaic (PV) capacity, performance monitoring for large-scale PV systems is an increasingly relevant and important topic. A large volume of research exists in this field, but there is a need for comparison of different methods and their performance toward relevant metrics, a broad discussion of the different choices involved, and subsequent consolidation. In this article, we focus on the detection of small power losses on string-level. We discuss the different choices involved in building a robust string performance monitoring scheme. We suggest the following approach: 1) identify bad data quality and do data-filtering; 2) calculate the daily specific yield on string level; 3) calculate the relative difference in specific yield between the strings (relative yield); 4) identify historical faults; 5) correct for seasonal variations; and 6) apply control charts to detect performance losses in new data and issue alarms/report to the system operators. Based on data from a utility scale PV power plant we compare different control charts in terms of detection time and sensitivity. We show that the cumulative sum (CUSUM) median and the Tukey-CUSUM charts are the most promising fault detection methods of the ones we have tested. We can robustly detect faults causing a performance loss of about 1% within 35 days of the drop in performance.