Benchmarks for solar energy data, methods


In the ever-evolving world of solar energy applications, access to accurate and reliable modeled irradiance data is crucial. Modeled irradiance data based on satellite products and numerical weather prediction models are frequently used. Many of such sources of data are offered by institutional or commercial providers. However, it is difficult and time consuming for users to independently identify the best provider for their specific application and location.

Benchmarking Solar Data

In our Task 16 report “Worldwide Benchmark of Modelled Solar Irradiance Data” we address this challenge. The report presents a benchmark of model-derived direct normal irradiance (DNI) as well as global horizontal irradiance (GHI) data that considers 129 globally distributed sites where ground-based radiation measurement stations are or have been installed. DNI and GHI estimates are compared against high-quality observations from these stations. The performance of the modeled data is analyzed with respect to different regions and climate zones. This study helps the solar industry make better informed decisions about solar resource assessments.

Building a reference database

Big efforts were made to build the reference database. We finally used data from 25 different providers with 129 stations during 2015–2020. Only quality-assured data has been considered in this benchmark through a comprehensive set of best practices and newly implemented quality-control procedures. These include automatic as well as manual data quality-control tests carried out by a team of experts led by CSP Services GmbH for all stations and result in flags describing the quality for each time stamp. The bulk of the quality-controlled data, covering all continents and many climate zones, has been published within this benchmark.

Global representation

One of the strengths of this benchmark is its global reach. The 129 ground-based stations are strategically distributed worldwide, encompassing diverse regions and climate zones. This global representation includes 31 stations in Africa, 31 in Asia, 27 in North America, 20 in Europe, 13 in Australia, 5 in South America, and even 2 in Antarctica. By spanning such a wide geographical scope, this benchmark provides insights into the performance of model-derived data in various environmental conditions, from arid deserts to polar extremes

Assessing model-derived data

Ten different models were tested, although not all provide estimates for all stations. Amongst other statistical performance parameters, the mean bias deviation, root mean square deviation, and standard deviation are calculated for each year and per station. The results for the relative mean bias deviation affecting GHI are shown in Figure 1.

Benchmark findings

Based on the results of the statistical analysis, the most appropriate dataset might depend on site, climate, or continent of interest. The model errors and the differences between the various modeled datasets are much higher for DNI than for GHI.

Popular content

Figure 1: Relative mean bias deviation (rMDB) for GHI and all stations and years. Magenta color indicates results out of the color bar range. The point size corresponds to the total number of datapoints in the tested data from 2015 to 2020.

Solar irradiance measurements and time series play a decisive role in supporting solar resource assessments, especially for medium and large PV installations. Not only they represent the foundation of solar resource assessment and forecasting, but they also drive prospective PV yield studies, and can be used as a calibration reference when using satellite data, evaluating PV systems’ performance, or developing forecasting algorithms.

Challenge: data gaps

However, such datasets inevitably have gaps – periods with missing data – as a result of defaults during data-logging, sensor failures or quality check procedures that can compromise their applicability and value. An additional issue is that data gaps can be further enlarged when computing temporal aggregations, notably for intra-daily to daily, daily to monthly and yearly averages, thus further degrading the dataset.

This has raised the need for gap-filling methods that can post‑process either static historical datasets or more dynamic real-time data streams. Each case is characterized by different constraints, such as the access to data that follows a given data gap or the acceptable time lag for generating the replacement synthetic data.

Benchmark for GHI gap-filling methods

In our report “Framework for Benchmarking of GHI Gap-Filling Methods”, led by Mines Paris PSL, we propose a gap-filling benchmark framework and evaluate a set of possible baseline algorithms for intra-hourly and daily sums of GHI time series. Five methods have been compared for different lengths of gaps: Nearest neighbors, linear interpolation, two machine learning approaches and use of satellite data. For short gaps linear interpolation works best, as for longer gaps the use of satellite data is suggested.


With access to comprehensive statistical performance parameters and insights into model errors, analysts can make informed decisions about which data providers and models align with their needs. Our recently published reports serve as a guiding light, making the uncertainty of data and models comparable for solar industry professionals and researchers, ultimately advancing the field and promoting sustainable energy solutions worldwide.

This article is part of a monthly column by the IEA PVPS programme. It was contributed by IEA PVPS Task 16 – Solar Resource for High Penetration and Large Scale Applications. Further information can be found in Task 16’s recent reports:

By Jan Remund, Meteotest AG, Switzerland

The views and opinions expressed in this article are the author’s own, and do not necessarily reflect those held by pv magazine.

This content is protected by copyright and may not be reused. If you want to cooperate with us and would like to reuse some of our content, please contact: