Presentation Details
| Tracker anomaly detection in PV fleet data via interpretable machine learning Corentin Servouze1, Bennet Meyers2, Matthew Muller2. 1Stanford University, Stanford, CA, USA.2NREL, Golden, CA, USA |
Abstract
We present an application of interpretable, convex-optimization–based anomaly detection to photovoltaic (PV) power plant data with single-axis/double-axis trackers and partially labeled operational failures. Building on previous work, we apply our previously published methodology to a new setting characterized by sparse observations, heterogeneous failure periods across systems, and labels associated with tracker malfunctions. In this manuscript, we present a method for adapting our model to partial observations and incomplete data. The trained model is then evaluated on its ability to recover known tracker failure days in the original dataset. Our method combines time-dependent quantile normalization, linear regression across correlated PV signals, and residual scoring to detect anomalous days. Artificial anomalies are generated to simulate tracker failures, using patterns informed by labeled faulty days. The trained model is then evaluated on its ability to recover known tracker failure days in the original dataset.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
Content Locked. Log into a registered attendee account to access this presentation.