Estimating Subhourly Inverter Clipping Loss From Satellite-Derived Irradiance Data

Kevin Anderson, Kirsten Perry

Research output: Contribution to conferencePaperpeer-review

9 Scopus Citations


Photovoltaic system production simulations are conventionally run using hourly weather datasets. Hourly simulations are sufficiently accurate to predict the majority of long-term system behavior but cannot resolve high-frequency effects like inverter clipping caused by short-duration irradiance variability. Direct modeling of this subhourly clipping error is only possible for the few locations with high-resolution irradiance datasets. This paper describes a method of predicting the magnitude of this error using a machine learning regressor ensemble model, comprised of a random forest and an XGBoost model, and 30-minute satellite irradiance data. The method predicts a correction for each 30-minute interval with the potential to roll up into 60-minute corrections to match an hourly energy model. The model is trained and validated at locations where the error can be directly simulated from 1-minute ground data. The validation shows low bias at most ground station locations. The model is also applied to gridded satellite irradiance to produce a heatmap of the estimated clipping error across the United States. Finally, the relative importance of each predictor satellite variable is retrieved from the model and discussed.

Original languageAmerican English
Number of pages6
StatePublished - 14 Jun 2020
Event47th IEEE Photovoltaic Specialists Conference, PVSC 2020 - Calgary, Canada
Duration: 15 Jun 202021 Aug 2020


Conference47th IEEE Photovoltaic Specialists Conference, PVSC 2020

Bibliographical note

See NREL/CP-5K00-76021 for preprint

NREL Publication Number

  • NREL/CP-5K00-79285


  • clipping
  • high-frequency
  • inverter
  • irradiance
  • modeling
  • photovoltaic
  • satellite
  • saturation
  • variability


Dive into the research topics of 'Estimating Subhourly Inverter Clipping Loss From Satellite-Derived Irradiance Data'. Together they form a unique fingerprint.

Cite this