Cloud-shadow nowcasting for solar-farm output ramp prediction
High-cadence geostationary imagery from MSG SEVIRI, GOES-16 ABI and Himawari-9 AHI can track cloud shadows across solar farms every 5–15 minutes, giving grid operators a short warning window before output ramp events. Accuracy is honest: useful to roughly 30 minutes ahead, degrading sharply beyond that.
Sensors
- MSG SEVIRI (Meteosat Second Generation): 15-minute full-disc repeat, 3 km spatial sampling at nadir (degrading to ~5 km at mid-latitudes), 12 spectral channels including 0.6 µm visible and 10.8 µm thermal IR. Primary source for European and African solar-farm shadow tracking.
- GOES-16 ABI (Advanced Baseline Imager): Full-disc scan every 10 minutes, mesoscale sector every 30–60 seconds, 0.5 km visible channel (0.64 µm) and 2 km IR channels. Best available cadence for the Americas; the mesoscale mode is the key asset for ramp-event nowcasting.
- Himawari-9 AHI (Advanced Himawari Imager): 10-minute full-disc repeat, 0.5 km visible at nadir, 16 spectral bands. Covers East Asia, Southeast Asia and Australasia. Functionally similar to GOES-16 ABI in architecture and cadence.
- MTG-I SEVIRI / FCI (Meteosat Third Generation Imager): MTG-I1 launched December 2022, commissioning ongoing. The Flexible Combined Imager will deliver 2.5 km resolution at nadir with a 10-minute repeat, improving on MSG SEVIRI's spatial resolution by roughly 20% and adding richer spectral content for cloud-top height retrieval.
Why ramp events are the grid operator's real problem
A solar farm producing 200 MW at noon can drop to 40 MW in under three minutes when a large cumulus cell crosses the array. That 160 MW ramp requires spinning reserve to be already online, because thermal plant cannot respond fast enough from cold. The cost of holding that reserve is real. If an operator can see the shadow coming 10 to 20 minutes ahead, they can pre-position reserve more efficiently, reducing both cost and the risk of frequency deviation.
Satellite nowcasting does not replace numerical weather prediction for the day-ahead market. It fills a specific gap: the 0-to-30-minute window where NWP is too coarse and ground pyranometers see only what has already arrived. Geostationary imagery sits squarely in that window.
The parallax problem: cloud tops are not where their shadows are
A cloud at 3 km altitude with a solar zenith angle of 45° casts its shadow roughly 3 km horizontally displaced from the point directly below the cloud top. At 8 km altitude the displacement exceeds 8 km. If you track cloud motion using raw geostationary pixel coordinates and project forward without correcting for this, your shadow-arrival prediction can be wrong by several kilometres and several minutes, which is the entire value of the forecast.
The correction requires three inputs: cloud-top height (retrieved from the 10.8 µm brightness temperature channel against a modelled temperature profile, or from stereoscopic parallax between two viewing angles), the solar azimuth and zenith angle at the time of interest, and the satellite viewing geometry. EUMETSAT's cloud-top height product for MSG SEVIRI is publicly documented and provides the height field operationally. The shadow footprint is then a straightforward trigonometric projection. The residual error in that projection is dominated by uncertainty in cloud-top height, which can be ±500 m to ±1 km for convective cells, translating to positional errors of roughly 500 m to 1 km in the shadow footprint at moderate solar zenith angles.
Cloud motion vectors, also produced operationally by EUMETSAT and NOAA from successive geostationary frames, supply the advection component. Combining the corrected shadow footprint with a cloud motion vector gives a shadow-arrival time estimate. At 10-minute lead times the method is genuinely useful. At 30 minutes, cumulus cells have often grown, split or dissipated in ways that no advection model captures, and the prediction skill drops sharply.
Thin cirrus: the detection problem that honest vendors admit
Thick cumulus shadows are easy. A cumulonimbus anvil at 3 km visible reflectance saturates the sensor and casts a sharp, deep shadow that cuts surface irradiance by 70–90%. Thin cirrus is the opposite problem. At 3 km pixel resolution, a cirrus veil with an optical depth of 0.1 to 0.3 may reduce surface irradiance by 10–25% without producing a visible pixel signature that stands out cleanly from a clear-sky background. The 10.8 µm channel helps, because cirrus is detectable in the thermal IR, but the irradiance reduction is ambiguous: the same brightness temperature anomaly can correspond to a range of optical depths depending on ice crystal habit.
This matters because a 15–20% irradiance reduction across a large array is still a significant ramp. Operators should understand that geostationary-based nowcasting has a detection floor for thin high cloud, and that this floor is not a software problem but a physics and resolution constraint. MTG-I's improved spatial resolution and additional spectral bands will help at the margins, but will not eliminate the ambiguity.
From pixel motion to a grid-ready ramp alert
The operational pipeline has four steps. First, ingest the latest geostationary visible and thermal IR frames as they are published, typically with a latency of 5–8 minutes from observation time for MSG SEVIRI products via EUMETSAT's data services. Second, run cloud detection and cloud-top height retrieval. Third, apply the parallax correction and compute the shadow footprint polygon for each identified cloud cell. Fourth, intersect the advected shadow footprint with the solar-farm boundary polygon and compute the fraction of installed capacity that will be shaded at each forecast timestep.
The output for a grid operator is not a map. It is a time series: expected AC output as a percentage of clear-sky capacity, updated every 10–15 minutes, with a confidence interval that widens explicitly with forecast horizon. Coupling this with the farm's DC-to-AC clipping curve and a clear-sky irradiance model (such as the McClear model derived from CAMS aerosol data) gives a physically consistent ramp forecast rather than a statistical anomaly flag.
Satellize runs this pipeline on open geostationary data streams and can deliver the output as a JSON feed or a GIS-compatible time-series layer. The same parallax-correction and shadow-advection logic underpins the crop-estimation work done for the Kingdom of Tonga, where cloud-masking accuracy directly affects the quality of optical vegetation indices.
Honest performance envelope
At 0–15 minutes lead time, shadow-boundary position errors are typically under 2 km for well-organised cloud systems with stable motion vectors, translating to timing errors of roughly 1–4 minutes depending on cloud speed. For a 500 m wide cloud shadow moving at 10 m/s, a 2 km positional error means a timing error of about 3 minutes. That is operationally useful.
Beyond 30 minutes, convective initiation and dissipation dominate over advection, and the method adds little beyond what a good NWP ensemble already provides. The correct architecture is to use satellite nowcasting for the 0–30 minute window and hand off to NWP for everything beyond that. Treating geostationary shadow tracking as a standalone 2-hour forecast tool will disappoint. Treating it as the short-range complement to NWP will not.
Typical figures
| Visible channel spatial resolution (nadir) | MSG SEVIRI: 3 km; GOES-16 ABI: 0.5 km (0.64 µm band); Himawari-9 AHI: 0.5 km; MTG-I FCI: 1 km (VIS 0.6 µm) |
| Full-disc repeat cadence | MSG SEVIRI: 15 min; GOES-16 ABI: 10 min full-disc, ~60 s mesoscale sector; Himawari-9 AHI: 10 min; MTG-I FCI: 10 min (target) |
| Data latency (observation to available product) | Typically 5–8 minutes for MSG SEVIRI via EUMETSAT; 3–5 minutes for GOES-16 ABI via NOAA CLASS / AWS |
| Cloud-top height retrieval accuracy | ±500 m to ±1 km for convective cells using 10.8 µm brightness temperature method; better with stereoscopic retrieval where available |
| Shadow-position error at 10-min lead time | Typically under 2 km for organised cloud systems with stable motion vectors |
| Useful forecast horizon | 0–30 minutes; skill degrades sharply beyond 30 minutes for convective cloud |
| Thin cirrus detection floor | Optical depth roughly 0.1–0.3 is ambiguous at 3 km resolution; associated irradiance reduction of 10–25% may not be reliably flagged |
| Geographic coverage | MSG SEVIRI: Europe, Africa, Middle East; GOES-16 ABI: Americas; Himawari-9 AHI: East/Southeast Asia, Australasia |
| Archive depth | MSG SEVIRI: operational since 2004; GOES-16 ABI: since late 2017; both accessible via EUMETSAT and NOAA archives |
| Delivery formats | JSON time-series feed, GeoTIFF shadow-footprint raster, GIS polygon layer (GeoPackage or Shapefile), operator dashboard API |
Analytics Satellize can run
| Shadow-arrival time series per farm | Cloud motion vector advection with parallax-corrected shadow footprint projection (trigonometric, using EUMETSAT/NOAA operational cloud-top height and CMV products) | JSON feed updated every 10–15 minutes, giving expected shaded fraction of installed capacity at each timestep over the next 30 minutes |
| AC output ramp forecast | Shaded-fraction time series convolved with farm DC-to-AC clipping curve and McClear clear-sky irradiance model (CAMS-based) | Ramp-event alert with magnitude (MW), onset time (±3–5 min confidence), and duration estimate; delivered via API or email alert |
| Cloud-top height field | 10.8 µm brightness temperature retrieval against ERA5 or operational NWP temperature profile; parallax correction applied | Gridded GeoTIFF at geostationary native resolution, updated each scan cycle |
| Shadow footprint polygon sequence | Cloud mask thresholded on visible reflectance and thermal IR, shadow projected to ground using solar geometry and cloud-top height | GeoPackage or Shapefile sequence, one polygon per scan cycle, suitable for overlay in farm SCADA GIS |
| Forecast skill assessment by cloud regime | Retrospective comparison of shadow-arrival predictions against ground pyranometer or inverter output records over a historical period | PDF report quantifying mean absolute timing error and bias by cloud type (cumulus, stratocumulus, cirrus) for a specific farm location |
| Thin-cirrus irradiance-reduction flag | Combined visible reflectance and 10.8 µm / 12.0 µm split-window cirrus detection; flagged as uncertain where optical depth estimate falls below 0.3 | Supplementary uncertainty band on the ramp forecast feed, with explicit cirrus-flag field per timestep |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.