Solar irradiance site adaptation

Overview

A satellite-based model describes the long-term solar resource over a wide area and a long history, but with relatively higher uncertainty at any single point. Ground measurements describe a single point accurately and with low uncertainty, but over a short period. Site adaptation combines the two, transferring the local accuracy of the measurements onto the long, area-representative satellite record. The result is a multi-year dataset that keeps the long history of the satellite-based model together with the low uncertainty and local representativeness of the ground measurements.

Table 1: Characteristics of the input data sources and the site-adapted result.

Site adaptation pursues several objectives:

• removal of the systematic bias between the modeled and measured data,

• reducing seasonal differences, improving the representativeness of model data for all seasons,

• improvement of the fit between the frequency distributions of the two datasets,

• reducing uncertainty of the site-adapted model data.

The methods adapt satellite-based GHI (Global Horizontal Irradiance) and DNI (Direct Normal Irradiance), and the parameters derived from them, to local conditions that the satellite and atmospheric inputs cannot resolve. Applied incorrectly, to non-systematic deviations, or with low-quality ground data, they degrade the dataset instead of improving it.

The site adaptation process

Site adaptation follows a fixed sequence. Quality-controlled ground measurements form the site reference which is correlated with the Solargis satellite-based model data over the period in which the two overlap. Correction parameters are derived from this overlap and applied to the full satellite record, producing a site-adapted time series and a report.

The correction parameters assume that the systematic differences between the model and the measurements are stable over a one- to two-year period. To remove the effect of differing time resolution and the conceptual difference between a point and a pixel, all metrics are computed on hourly-aggregated data. Solargis satellite data is available in 15-minute time steps, while ground measurements typically range from 1-minute to 60-minute.

Figure 1: The Solargis site adaptation process.

Site adaptation challenges

Spatial mismatch between satellite and ground data

A satellite radiometer integrates the signal over an area of several square kilometers, while a ground station measures at a single point. In complex regions, such as narrow valleys with frequent fog or coastal areas, one pixel can represent a mix of conditions that differ from what the station experiences. This mismatch, known as the nugget effect, accounts for nearly half of the hourly RMSD (Root Mean Square Deviation) for GHI and DNI, and is strongest during intermittent cloud cover and changing aerosol conditions.

DNI and DIF sensitivity

DNI is highly sensitive to cloud cover, aerosols, water vapor, and terrain shading, and the relationship between GHI and DNI uncertainty is nonlinear: a small error in GHI can correspond to a much larger error in DNI. DIF (Diffuse Horizontal Irradiance) is not modeled independently. It is derived from GHI and DNI through the closure relationship , where is the solar zenith angle (see Solar irradiance components and Validation of solar irradiance data). Because DIF is the residual of two larger quantities, a small absolute change in GHI or DNI produces a proportionally much larger relative change in DIF. The effect is strongest in arid, high-DNI climates, where the diffuse fraction is typically 15-30% of GHI.

Building a good reference

The site-adapted dataset can only be as accurate as the reference built from the ground measurements. Two conditions determine whether that reference improves the satellite data or adds noise to it: the uncertainty and length of the measurements, and their quality control.

The uncertainty of the ground measurements must be lower than the uncertainty of the satellite-based model, otherwise adaptation cannot improve the dataset. Even a well-installed and maintained Class A pyranometer has a measurement uncertainty of approximately 3%, which sets the practical floor for the uncertainty achievable after adaptation.

The length of the ground measurements is also an important factor determining the achievable reduction in uncertainty of the site-adapted dataset. The longer the measurements, the more information about the site they contain, which directly reduces the uncertainty about the solar radiation at the site.

Even well-maintained sensors produce erroneous readings, and only quality-controlled data can be used in site adaptation. Readings left uncorrected in the reference propagate directly into the site-adapted dataset. Solargis applies a multi-level quality control framework before adaptation - the procedures are described in Solar irradiance ground measurements and Harmonization of ground-measured solar data.

Table 2: Suitability of different measurement periods.

Figure 2: Indicative ranges of achievable uncertainty of site-adapted GHI and DNI data in relation to the length of ground measurements.

Note: SPN1, a non-rotating shadowband radiometer, cannot be used as the adaptation reference. Its measured DIF is underestimated by design, and its real measurement uncertainty, approximately 10 to 20% for DIF, is too high to improve the satellite data. Sensor characteristics are discussed in detail in the Solar irradiance ground measurements article.

Adaptation methods

Site adaptation uses a several methods of increasing sophistication, from statistical correction of the satellite output values to adjustment of the satellite-based solar resource model itself. In practice the methods are combined, and the choice depends on the quality of the reference data and on the understanding of the site and of solar resource physics. Because the value of the produced energy varies at the subhourly level, the goal is accuracy across all time scales, not only in the annual average, which usually requires applying the methods separately by season.

Statistical bias correction

The ratio method adjusts the long-term monthly and annual averages of GHI and DNI. Ratios between the ground-measured and satellite-derived solar radiation are computed over the overlap period and applied to recalibrate the long-term satellite dataset. The method is simple, but it corrects only the mean bias: it does not align the frequency distributions and does not use the full information in the measurements. It is the basis of many Measure-Correlate-Predict (MCP) approaches.

Distribution adaptation

Fitting the cumulative distribution function (CDF) aligns the CDFs of the satellite and ground data and matches their averages, improving the representation of both typical and extreme values. It usually reduces RMSD in addition to bias. Applying it correctly requires an understanding of the site's solar resource, and it is normally performed season by season.

Adaptation of the model

Expert methods adjust the satellite-based model itself and recompute all three components together, which keeps GHI, DNI, and DIF mutually consistent. They require operating the solar resource model, so that its inputs can be modified and the model can be re-run.

Adaptation of the clearness index uses the clearness index ( ), the ratio of surface radiation to top-of-atmosphere radiation) as the key model input. It is derived from the ground measurements with a solar radiation model, calibrated by least-squares regression to align hourly or daily differences between the ground-measured and satellite-based parameters, and used to recalculate GHI, DIF, and DNI. It reduces mean bias, RMSD, and KSI (Kolmogorov-Smirnov test Integral), but its use of a single parameter limits its capacity to handle complex atmospheric effects.

Adaptation of the model inputs replaces the clearness index with more detailed parameters, such as Aerosol Optical Depth (AOD) or the cloud index, and recomputes the entire model. By addressing seasonal and regional inaccuracies in the aerosol and cloud description, it reduces the mean bias to within the expected uncertainty of the measurement instruments while minimizing RMSD and KSI. It corrects errors at the subhourly level and is particularly valuable in semi-arid and desert regions, where aerosols and clouds dominate the variability.

Treatment of DIF

Site adaptation is anchored on the validated primary outputs, GHI and DNI. In the basic and advanced value-based methods, DIF is back-calculated as the residual so that the closure equation is always satisfied; in the expert methods, the model recomputes all three components together. Because GHI and DNI are adjusted independently, the diffuse fraction () can change after adaptation, and the relative change in DIF is normally several times larger than the change in GHI. Measured DIF is not used as an independent adaptation target: it is obtained with instruments of higher, and often systematic, uncertainty, and Solargis uses it only as a diagnostic to verify the internal consistency of the adapted dataset.

Note: A change in DIF and in the diffuse fraction (D2G) after site adaptation is expected, and the relative change in DIF is normally larger than in GHI or DNI. This follows from DIF being a derived residual rather than an independently adapted parameter. The direction and size of the change over a specific period depend on the site and the measurement period.

Common pitfalls

Applied without the necessary expertise, site adaptation can degrade the dataset while appearing to succeed. The typical failure modes are:

• Data manipulation without physical modeling can break the relationship between GHI, DNI, and DIF, most visibly as negative DIF values.

• Correcting the mean bias alone produces an accurate annual figure but leaves large errors at the daily, hourly, and subhourly levels.

• Rare events in the reference period, such as major dust storms, wildfires, or volcanic ash, skew the long-term result if they are not excluded before adaptation.

• A reference too short to cover all seasons leads to overcorrection of the observed season and misrepresentation of the others.

Assessing the results

Three metrics quantify the improvement, computed from all hourly daytime data pairs over the overlap period:

• Mean Bias, in absolute and relative form (relative to the daytime average GHI), describing the systematic deviation of the satellite data from the measurements.

• RMSD, in absolute and relative form, describing the magnitude of the fluctuations between the two datasets.

• KSI, describing how well the two frequency distributions match, particularly at the extreme values.

Bias and uncertainty are not the same. A site adaptation can reduce the bias to zero while the uncertainty remains at the level set by the reference data (approximately 3% at best) and by the site characteristics. Reporting bias alone describes the annual average but says nothing about accuracy at finer time scales.

Figure 3: Cumulative distribution graphs representing satellite-based vs measured GHI and DNI values, before and after site adaptation.

Further reading

• "Site adaptation of satellite-based DNI and GHI time series: Overview and SolarGIS approach": Cebecauer, T.; Suri, M. AIP Conf. Proc. 1734, 150002 (2016).

• "Analysis of different comparison parameters applied to solar radiation data from satellite and German radiometric stations": Espinar, B.; Ramírez, L.; Drews, A.; Beyer, H.G.; Zarzalejo, L.F.; Polo, J.; Martín, L. Solar Energy, 83(1), 118–125 (2009).

• "Effective accuracy of satellite-derived hourly irradiances": Zelenka, A.; Perez, R.; Seals, R.; Renne, D. Theoretical and Applied Climatology, 62, 199–207 (1999).