Machine learning versus ray-tracing to forecast irradiance for an edge-computing SkyImager

Increasing penetration of photovoltaics (PV) into the electric grid necessitates accurate intra-hour solar irra-diance forecasting. The authors have previously developed a low cost all-sky imaging system at the University of Texas at San Antonio. The SkyImager hardware is designed around a Raspberry Pi single board computer (SBC) with a fully programmable, high resolution Pi Camera, housed in an all-weather enclosure. The software to process the images and to make minutes-ahead forecasts is written in Python 2.7 and utilizes the open source computer vision package OpenCV. A two-step approach for the forecasts uses optical flow to track low-level cumulus clouds and ray-tracing to predict the location of cloud shadows. This paper proposes to replace the ray-tracing approach which is traditionally known to be an ill-posed inverse problem with a machine learning strategy that utilizes a novel multi-layer perceptron (MLP) to classify cloud-cover in sub-images of the circumsolar region. The developed SkyImager was deployed at the National Renewable Energy Laboratory (NREL) in 2015, where it successfully collected months of all-sky image data. In 2016 a second SkyImager was integrated into a microgrid management system at Joint Base San Antonio. Results are presented to validate the proposed methodology.