Presentation Details
| DC Loss Prediction for PV Systems from Measured Energy Time Series and Numerical Weather Predictions Thomas Haley, Emily Tansey, Kyle Seymour. Clean Power Research, Bellevue, WA, USA |
Abstract
Time series of PV DC system losses (snow, soiling, panel degradation, and string failure) were simulated at 14 SURFRAD/SOLRAD sites. PV power including the simulated losses was modeled from measured irradiance. An estimate of PV power at each site derived from only NWP sources was compared with modeled power from measured irradiance to infer total DC loss. Two supervised learning ML models were trained to predict total DC loss, and the contribution of each loss factor. The first ML model was trained with NWP features and the inferred total DC loss (“With Inferred Loss”). The second model was trained with only NWP features (“Without Inferred Loss”). Accuracy of the two ML loss segmentation models were compared with leave-one-out cross-validation by location. The “With Inferred Loss” model outperformed the “Without Inferred Loss” model for all loss types. The relative difference in accuracy (skill score) was larger for high-impact losses (snow and string failure) and smaller for soiling and panel degradation loss types.
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.