Sky Camera Calibration and Cloud Motion Vector Extraction

Sky camera systems use wide-angle or fisheye lenses to capture the entire hemisphere. Without calibration, raw pixels have no physical meaning—like reading a map without knowing which direction is north or how far a centimeter represents. Calibration establishes the mathematical relationship between each pixel and its corresponding point in the sky, using either physical targets (e.g., a precisely surveyed dome grid) or natural references (sun position, star catalogs). This yields a projection model—most commonly zenithal equidistant or stereographic—that preserves angular distances critical for cloud tracking.

Beyond geometry, radiometric calibration corrects for lens vignetting, sensor nonlinearity, and atmospheric path radiance, ensuring that pixel intensity maps linearly to broadband irradiance. CMV extraction then relies on temporal consistency: optical flow algorithms assume brightness constancy, but clouds change reflectivity rapidly. Hence, robust pipelines use gradient-based methods (e.g., Farnebäck) combined with feature tracking (e.g., ORB + RANSAC) to separate true motion from albedo artifacts. Motion vectors are finally projected onto the ground plane using digital elevation models and plant-specific GIS footprints—so a vector pointing toward the southwest doesn’t just mean ‘cloud moving SW’ but ‘will obscure Block 3 PV array in 2.7 min’.

At scale, auditable CMV pipelines require version-controlled calibration parameters (stored as YAML with SHA-256 hashes), traceable to NIST-traceable irradiance standards (e.g., NREL’s SRRL reference cells), and embedded uncertainty propagation: each CMV carries a covariance matrix derived from reprojection error, temporal baseline jitter, and feature matching confidence. This satisfies ISO 15927-3 Annex B requirements for ‘quantified forecast confidence intervals’, enabling dispatch decisions under CAISO’s Real-Time Market rules where forecast error penalties begin at ±5% GHI deviation over 5-min windows.