Solar PV output prediction from video streams using convolutional neural networks

Solar photovoltaic (PV) installation is growing rapidly across the world, but the variability of solar power hinders its further penetration into the power grid. Part of the short-term variability stems from sudden changes in meteorological conditions, i.e. , change in cloud coverage, which can vary PV output significantly over timescales of minutes. Images of the sky provide information on current and future cloud coverage, and are potentially useful in inferring PV generation. This work uses convolutional neural networks (CNN) to correlate PV output to contemporaneous images of the sky (a "now-cast"). The CNN achieves test-set relative-root-mean-square error values (rRMSE) of 26.0% to 30.2% when applied to power outputs from two solar PV systems. We explore the sensitivity of model accuracy to a variety of CNN structures, with different widths, depths, and input image resolutions among other hyper-parameters. This success at "now-cast" prediction points to possible future uses for short-term forecasts. To forecast solar power's short-term fluctuation stemming from changes in cloud coverage, a convolutional neural networks (CNN) is used to correlate PV output to contemporaneous sky images.