Cloud Detection and Sun Visibility Forecasting via Full-Sky Imaging and Computer Vision

Abstract

In 2012, renewable energy composed about 12.5% of total energy production in the United States, of which over half was generated by wind. More recently, solar power generation through photovoltaics has continued to grow at record levels as cost and scalability challenges have been overcome. The variability of solar power continues to be an obstacle for high-penetration of solar power. In particular, successful grid integration has a higher necessity for accurate intra-hour solar forecasting to cope with clouds, and as forecasting methods improve, microgrids will more safely be able to become more dependent on solar power generation. This research discusses the development of an intra-hour sun visibility system. Sun visibility is used as an approximate indicator for solar irradiance. The primary hardware aspect of the system is a camera with a 150° fisheye lens; a neutral density filter is utilized to attenuate the brightness of the sun. The system software contains three primary modules: hardware interfacing, feature detection, and visibility forecasting. Computer vision techniques are employed for sun detection, cloud detection, and cloud movement detection. Multiple series of test images representing different partly cloudy days were used to simulate forecasting. Up to 5-minute forecasts were producible, and 1-minute forecasts had the highest average accuracy of 67%. This points to some limitations of direct-camera sky imaging and light attenuation. It is concluded that many techniques implemented extract cloud features effectively; however, the shifts in cloud shape and speed continue to be a challenge to practically model and forecast intra-hour, in the sub-kilometer range.