3D cloud detection and tracking system for solar forecast using multiple sky imagers. (August 2015)

• Record Type: Journal Article
• Authors: Peng, Zhenzhou; Yu, Dantong; Huang, Dong; Heiser, John; Yoo, Shinjae; Kalb, Paul
• Is Part Of: Solar energy. Volume 118(2015)
• Page Start: 496
• Page End: 519
• Publication Date: 2015-08
• Subjects: Sky imagery -- Cloud detecting -- Cloud tracking -- Short-term forecast; Solar energy -- Periodicals; Solar engines -- Periodicals; 621.47
• DOI: 10.1016/j.solener.2015.05.037
• Languages: English
• ISSNs: 0038-092X

Abstract

Highlights: A 3D cloud detection and tracking system based on multiple sky imagers is proposed. A supervised classifier and multi-source correction is used to identify cloud pixels on sky image. We propose to use clustering and aggregation to recognize multi-layer clouds based on spatial and temporal correlation. We investigate regression-based models for short-term irradiance forecast based on predicted image features. Abstract: We propose a system for forecasting short-term solar irradiance based on multiple total sky imagers (TSIs). The system utilizes a novel method of identifying and tracking clouds in three-dimensional space and an innovative pipeline for forecasting surface solar irradiance based on the image features of clouds. First, we develop a supervised classifier to detect clouds at the pixel level and output cloud mask. In the next step, we design intelligent algorithms to estimate the block-wise base height and motion of each cloud layer based on images from multiple TSIs. This information is then applied to stitch images together into larger views, which are then used for solar forecasting. We examine the system's ability to track clouds under various cloud conditions and investigate different irradiance forecast models at various sites. We confirm that this system can (1) robustly detect clouds and track layers, and (2) extract the significant global and local features for obtaining stable irradiance forecasts with short forecast horizons from the obtained images. Finally, we vet our forecasting system at the 32-megawatt Long Island Solar Farm (LISF). Compared with the persistent model, our system achieves at least a 26% improvement for all irradiance forecasts between one and fifteen minutes.