Operational solar-farm performance anomaly detection

Satellite-derived irradiance from SEVIRI and Sentinel-2 can be compared against SCADA generation data to flag underperforming sub-arrays, isolating soiling, shading, and inverter faults as distinct spatial signatures without a site visit.

Sensors

SEVIRI on Meteosat Second Generation: Provides the irradiance reference. Full-disc imagery every 15 minutes at roughly 3 km nadir resolution. The Meteosat SARAH-3 climate data record, published by EUMETSAT, gives surface solar irradiance with a typical hourly bias below 10 W/m² over flat terrain. This is the energy-available input against which SCADA output is compared.
Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands (bands 2–4, 8), 20 m in red-edge and shortwave infrared (bands 11–12). Revisit is 5 days at the equator with both satellites. Useful for detecting soiling extent and vegetation encroachment as reflectance changes, and for mapping shading geometry. Cannot resolve individual strings or panels.
Landsat 8/9 OLI+TIRS: 30 m multispectral, 100 m thermal infrared (resampled to 30 m in products). 16-day revisit per satellite; combined 8-day with both. TIRS band 10 (10.6–11.2 µm) provides coarse thermal context for large sub-array temperature anomalies, but spatial resolution limits diagnosis to blocks of many rows rather than individual strings.
Planet SuperDove: 3 m resolution, 8 spectral bands including red-edge and near-infrared, daily revisit over most land areas. Sufficient to distinguish soiled panel clusters from clean ones via reflectance contrast, and to detect partial shading patterns from adjacent structures or vegetation. Still insufficient for single-string fault isolation.
Airborne or drone-mounted thermal imagers: Typically 10–20 cm resolution at 100 m altitude. The only platform that resolves individual cells, strings, and bypass diodes. Required for fault-level diagnosis once satellite data has flagged a suspect sub-array. Not a satellite product, but the necessary second-stage instrument; satellite analysis defines where to task it.

The performance ratio gap and why it is hard to find from the ground

A solar farm's performance ratio, the fraction of theoretically available energy that reaches the meter, should sit between 0.75 and 0.85 for a well-maintained site. When it drops, the cause is rarely obvious. Soiling, partial shading from new obstructions, inverter degradation, and string disconnections all reduce output, but they produce similar signals in aggregated SCADA data. Distinguishing them usually means dispatching an inspection team, which is expensive and slow across a multi-hectare site.

Satellite data breaks the ambiguity spatially. Each cause leaves a different footprint. Soiling reduces reflectance uniformly across a dusty sector. Shading produces a geometric shadow pattern that moves predictably with sun angle. Inverter faults produce thermal anomalies concentrated at specific combiner-box zones. None of these patterns is visible in a single SCADA channel, but each is detectable, at the right resolution, from above.

Irradiance from orbit: what SEVIRI actually measures and what it cannot

SEVIRI's 15-minute cadence makes it the standard irradiance input for performance-ratio calculations at the farm scale. The EUMETSAT SARAH-3 product converts top-of-atmosphere radiance to surface solar irradiance using a Heliosat-derived algorithm, accounting for cloud cover and aerosol optical depth. Over flat, low-albedo terrain the hourly root-mean-square error is typically 20–30 W/m²; over complex terrain or during rapid cloud transitions it is higher. That uncertainty propagates directly into the performance ratio, so a farm showing a 3% shortfall against SEVIRI-derived expectation is within the noise. A 10% shortfall sustained over several clear-sky days is not.

The 3 km pixel footprint means SEVIRI cannot attribute irradiance variation within a farm. A cloud shadow crossing one corner of a 200-hectare site looks the same as clear sky in the SEVIRI record. For intra-site irradiance mapping, Sentinel-2 cloud masks applied to clear-sky composites give a better picture of persistent shading, though at the cost of temporal resolution.

What a satellite can see on the panels themselves

Soiling is the most tractable problem from orbit. Dust and bird fouling raise the reflectance of panel surfaces in visible wavelengths. Sentinel-2 band 4 (red, 665 nm) and band 11 (SWIR, 1610 nm) together provide a soiling-sensitive index because clean panels absorb strongly in the red and SWIR while soiled panels reflect more. At 10–20 m resolution this works at the sub-array level, perhaps 20–50 rows, not at the string level. Planet SuperDove at 3 m can push that down to clusters of a few rows.

Shading from new obstructions, a growing hedgerow, a new warehouse, a temporary structure, appears as a geometric shadow pattern in high-resolution imagery that correlates with sun-angle geometry. Comparing multi-date Sentinel-2 or Planet imagery against a shadow model built from a digital surface model identifies new shading sources without a site visit.

Thermal anomalies from inverter or string faults require spatial resolution below 1 m to isolate individual components. Landsat TIRS at 100 m and even airborne sensors at 1 m are too coarse for cell-level diagnosis. The practical workflow is: satellite flags a suspect sub-array zone; drone thermal survey is tasked to that zone. This staged approach reduces drone flight hours by an order of magnitude compared with blanket surveys.

Resolution is the binding constraint, and pretending otherwise wastes money

Sentinel-2 at 10 m covers a typical 500 kW string zone in roughly one pixel. That is useful for identifying which quadrant of a farm is underperforming, not which string has a failed bypass diode. Anyone selling string-level fault detection from Sentinel-2 alone is selling something the physics does not support.

The honest resolution hierarchy is: SEVIRI for farm-level performance ratio; Sentinel-2 or Planet for sub-array soiling and shading extent; Landsat TIRS for very coarse thermal context; airborne or drone thermal for string and cell diagnosis. Each layer narrows the search area for the next. The value of the satellite layers is not that they replace inspection, it is that they make inspection targeted. A 200-hectare farm inspected at random with a drone takes a full day. One where satellite analysis has already flagged two suspect sub-arrays takes two hours.

Running the comparison in practice

The core calculation is straightforward: take SEVIRI-derived plane-of-array irradiance for each 15-minute interval on clear-sky days, multiply by nameplate capacity and expected performance ratio, compare against SCADA-reported generation at the inverter or combiner-box level, and flag intervals where the gap exceeds a threshold. The difficulty is in the inputs. SEVIRI irradiance must be corrected to plane-of-array using tilt and azimuth data. Temperature coefficients must be applied using land-surface temperature estimates, which Landsat TIRS or ERA5 reanalysis can supply. Aerosol optical depth from MODIS or Copernicus Atmosphere Monitoring Service adjusts for dust events that reduce irradiance without appearing as clouds.

Once a persistent underperformance flag is confirmed across multiple clear-sky days, Sentinel-2 or Planet imagery over the same period is examined for reflectance anomalies. If soiling is the candidate, a cleaning schedule can be prioritised by sub-array rather than applied uniformly. If shading geometry matches a new structure, the asset manager has documented evidence for a planning or lease dispute. If neither explains the gap, the drone goes out, but only to the flagged zone.

Satellize runs this comparison workflow on open-constellation data for operational sites, combining SEVIRI irradiance records with client-supplied SCADA exports and Sentinel-2 or Planet tasking. The Tonga crop-estimation programme uses a similar multi-source anomaly detection logic, adapted here for panel reflectance rather than canopy indices.

Archive depth and what it reveals over time

Sentinel-2 archive runs from 2015; Landsat 8 from 2013. A farm commissioned in 2016 therefore has nearly a decade of satellite-observable history. Year-on-year comparison of clear-sky reflectance in SWIR bands tracks gradual panel soiling accumulation and, combined with SCADA records, can separate seasonal soiling cycles from progressive degradation. The published literature on panel degradation rates, typically 0.5–0.7% per year for crystalline silicon, gives a baseline against which satellite-derived performance trends can be checked. Deviations above that rate, sustained across multiple seasons, point to a physical problem rather than normal ageing.

Typical figures

Irradiance reference spatial resolution	~3 km (SEVIRI nadir); SARAH-3 product hourly RMSE ~20–30 W/m² over flat terrain
Irradiance reference temporal resolution	15-minute full-disc imagery from Meteosat SEVIRI
Multispectral anomaly detection resolution	10 m (Sentinel-2 visible/NIR), 20 m (SWIR); 3 m (Planet SuperDove)
Multispectral revisit	5 days (Sentinel-2A+B combined); daily (Planet SuperDove)
Thermal resolution available from satellite	100 m (Landsat TIRS band 10/11); insufficient for string-level diagnosis
Minimum detectable soiling zone	Approximately 20–50 panel rows at Sentinel-2 resolution; a few rows at Planet 3 m
Satellite archive depth	Sentinel-2 from 2015; Landsat 8 from 2013; Meteosat SARAH-3 from 1983
Latency from acquisition to analysis	Sentinel-2 Level-2A available within ~3 hours of overpass via Copernicus Data Space
SCADA integration format	Client-supplied CSV or time-series export; 15-minute or hourly intervals preferred
Deliverable formats	GeoTIFF anomaly maps, GeoJSON sub-array flags, PDF performance-ratio reports, CSV time series

Analytics Satellize can run

Clear-sky performance ratio deviation map	SEVIRI SARAH-3 irradiance vs SCADA generation comparison on cloud-screened days; plane-of-array correction applied using site tilt/azimuth metadata	Monthly PDF report with sub-inverter performance ratio heatmap and flagged zones
Soiling extent and severity index	Sentinel-2 or Planet multispectral reflectance ratio (red/SWIR) compared against clean-panel baseline from commissioning-period imagery	GeoTIFF raster and GeoJSON polygons showing soiling severity by sub-array, updated each clear-sky overpass
Shading source identification	Multi-date high-resolution imagery shadow pattern compared against solar geometry model; new obstructions flagged by change detection	GIS layer identifying shading source locations and estimated generation loss per affected zone
Thermal anomaly pre-screening zone	Landsat TIRS land-surface temperature anomaly relative to farm mean on clear days; zones exceeding threshold flagged for drone tasking	GeoJSON bounding boxes defining priority drone-survey areas, reducing survey area by estimated 80–90% vs blanket flight
Long-term degradation trend	Annual clear-sky performance ratio time series from Sentinel-2 reflectance and SEVIRI irradiance archive; compared against published crystalline silicon degradation benchmarks (0.5–0.7%/year)	Annual trend report with per-sub-array degradation curves and deviation-from-expected flags
Aerosol-corrected irradiance input	MODIS or Copernicus Atmosphere Monitoring Service aerosol optical depth applied to SEVIRI surface irradiance to correct for dust-event attenuation	Corrected irradiance time series CSV for client SCADA integration, with aerosol event log

Sources

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.