In-situ Atmospheric Intelligence for Hybrid Power Grids: Volume 6 (Convolutional Neural Networks for Whole Sky Imager Data Analysis)
Abstract
Automated analysis of whole sky imager (WSI) images can provide atmospheric intelligence for near-future solar power generation. Atmospheric intelligence in the form of input to solar radiation models can subsequently inform optimization algorithms of an energy management system, enabling efficient and effective power distribution. In this work, we developed three separate convolutional neural network models to assess WSI images including image segmentation by cloud type, estimation of thin/thick cloud percentage, and inference of current solar radiation. A cloud segmentation model was trained with limited cloud example data and was moderately successful, except when identifying a bright clear sky. A cloud quantification model was trained with hand-labeled percentages and was accurate enough for use in energy management system simulations. Finally, a solar radiation model was trained on a year-round data set from the National Renewable Energy Laboratory. This model was able to infer the total solar radiation within 50 W/sq m using a camera sensor limited to only visible wavelengths. Future studies will include model accuracy improvements with more training data and new prediction models of future sky conditions based on past WSI image sequences.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2022
- Accession Number
- AD1171285
Entities
People
- Gail T. Vaucher
- Michael S. D'arcy
- Michael S. Lee
Organizations
- United States Army