Cloud Motion Vector (CMV)

The foundational computer vision technique that partitions consecutive sky images into pixel sub-regions and identifies the displacement vector that maximizes the cross-correlation coefficient between frames. Normalized Cross-Correlation (NCC) is the standard similarity metric, computed as the dot product of normalized pixel intensities within a search window. The algorithm searches a bounded neighborhood in the subsequent frame to find the best match, generating a sparse velocity field. Multi-scale pyramidal implementations refine coarse displacements at reduced resolutions before fine-tuning at full resolution, dramatically reducing computational cost while capturing both small cumulus and large stratus advection.

A dense motion estimation alternative to block matching that computes a velocity vector for every pixel by solving the brightness constancy constraint equation. The Horn-Schunck method imposes a global smoothness regularization term, while the Lucas-Kanade method assumes constant flow within a local neighborhood and solves an overdetermined least-squares system. Modern deep optical flow architectures like RAFT (Recurrent All-Pairs Field Transforms) use iterative residual refinement with correlation volumes to produce state-of-the-art dense flow fields. Optical flow captures non-rigid cloud deformation that rigid block matching misses, improving advection fidelity for shearing and convective cloud fields.

Once the 2D motion field is derived, the current cloud map is propagated forward in time using a semi-Lagrangian advection scheme. Each pixel in the forecast frame is computed by tracing its trajectory backward along the velocity field to sample the source image at fractional coordinates, typically using bilinear interpolation. The core assumption is Lagrangian persistence—that cloud brightness patterns are conserved as they translate, ignoring formation, dissipation, and vertical development. This assumption holds for lead times of 0–60 minutes before convective evolution dominates. The advected cloud map is then converted to a Clear Sky Index field by applying a transmittance model, which is multiplied by the theoretical clear-sky irradiance to produce the final GHI forecast.

Ground-based CMV systems rely on Total Sky Imagers (TSI) or All-Sky Cameras equipped with fisheye lenses and hemispherical mirrors to capture the full sky dome in a single frame. Modern systems use high-dynamic-range (HDR) sensors to prevent saturation near the solar disk and spectral filters (typically 650 nm red band) to maximize cloud-sky contrast. The raw circular image undergoes geometric calibration to map pixel coordinates to zenith and azimuth angles, correcting for lens distortion. Paired with a shadowband or occulting disk to block direct sunlight, these imagers capture images every 15–30 seconds, providing the high temporal resolution necessary to track cloud motion before features decorrelate.

Geostationary satellites like GOES-16/17 and Meteosat provide CMV inputs at continental scales by tracking cloud features across sequential multispectral images. The Advanced Baseline Imager (ABI) captures full-disk scans every 10 minutes across 16 spectral bands, enabling motion vector derivation for multiple atmospheric levels. Atmospheric Motion Vectors (AMVs) are operationally produced by numerical weather prediction centers by tracking water vapor and infrared channel features. These satellite-derived vectors are assimilated into NWP models but can also drive standalone advection forecasts for utility-scale solar plants, offering a spatially continuous alternative to ground-based sky imagers without the need for on-site hardware deployment.

Operational forecasting systems combine CMV advection with Numerical Weather Prediction (NWP) using a nowcasting-blending framework. CMV dominates the 0–2 hour horizon where its high-resolution cloud tracking outperforms NWP's delayed spin-up. NWP takes over at 2–6 hours as its physics-based convection schemes capture cloud formation and dissipation that pure advection misses. The transition uses an exponential weighting function that smoothly interpolates between the two forecasts based on lead time. Advanced implementations use a dynamic weighting approach where the blending factor is a function of the local convective available potential energy (CAPE), giving more weight to NWP during unstable conditions when cloud evolution is rapid.

The rate of change of solar irradiance over time, typically measured in W/m² per minute. This metric quantifies sudden power fluctuations caused by moving clouds that threaten grid stability. CMV is specifically designed to predict these ramp events by tracking cloud edge positions and velocities. Key characteristics:

• Positive ramps: Cloud clearing, sudden irradiance increase
• Negative ramps: Cloud arrival, sudden irradiance decrease
• Critical threshold: Changes exceeding 50% of clear sky index within 5 minutes Grid operators use ramp forecasts to schedule fast-ramping reserves and prevent frequency excursions.