SPLTRAK Abstract Submission
Applying unsupervised machine learning for the detection of shading on a portfolio of commercial roof-top power plants in Germany
Nicolas Holland1, Klaus Kiefer1, Christian Reise1, Eduardo Sarquis Filho2, Bernd Kollosch3, Bjoern Mueller4
1Fraunhofer Institute for Solar Energy Systems ISE, Freiburg, Germany
/2IDMEC, Instituto Superior Tecnico, Universidade de Lisboa, Lisbon, Portugal
/3Pohlen Solar GmbH, Geilenkirchen, Germany
/4Enmova GmbH, Freiburg, Germany

Obstacles that cast shading on commercially operated PV power plants can lead to a variety of issues like false alerts in failure detection systems or skewed performace ratios. The detection and monitoring of shading effects using on-site inspections can be challenging, especially when one handles a large portfolio of power plants over a period of many years, since shading behaviours can also change over time. We apply an unsupervised method for detecting shading directly from power measurements to create so called shading masks, which make binary statements over whether or not a power plant or subplant is subject to shading at a given time. The shading masks are compared with the results of on-site inspections and they are used to create loss estimates.