IET Renewable Power Generation Research Article Improvement global solar radiation estimation ISSN 1752-1416 Received on 18th February 2017 Revised 4th April 2017 Accepted on 26th April 2017 doi: 10.1049/iet-rpg.2017.0118 www.ietdl.org Ismail Loghmari 1 , Youssef Timoumi 1 1 Mechanical Engineering Laboratory, National School of Engineering of Monastir, University of Monastir, Ibn Aljazzar Street, 5019 Monastir, Tunisia E-mail: yesine@live.fr Abstract: Energy yields of photovoltaic power plants are directly related to the availability of the global solar radiation (GHI). An accurate performance analysis of these power plants depends strongly on the quality and the reliability of the solar resource assessment. This study proposed to improve the accuracy of the GHI database provided by satellites. Two quality improvement methods have been proposed and evaluated in this study. The first developed method consists in combining a GHI satellite- derived database with the best ground station models, while the second one consists in performing a linear correction of a satellite database. The purpose of this study is to investigate the quality improvement method that gives more accurate GHI prediction. The comparison between the two developed methods shows that the resulting combination method database achieves higher GHI prediction accuracy than the linearly corrected satellite database. This combination reduces the uncertainty of the original satellite database by 1.95%, with a resulting relative root-mean-square error (rmse%) reaching 4.74%. 1௑Introduction Global solar radiation (GHI) is the fuel of all photovoltaic solar systems. In order to effectively estimate the technical and the financial potential of such solar projects, the project developer should accurately predict the long-term solar radiation availability at a targeted site. Due to their availability for the earth's entire surface and the long-term temporal coverage, historical satellite solar resources such as MACC-RAD [1], SolarGIS [2], Nasa SSE [3], SOLEMI [4], Helioclim [5] and EnMetSol [6] are commonly used. However, these databases may not be sufficient to meet the project developer's needs when making investment decisions because of the high uncertainty of such databases. Indeed, in the solar radiation market, the typical uncertainty of a satellite data resource can be largely higher than 10% [7]. To tackle this problem, one approach that could be used is to improve the accuracy of the satellite-derived database by applying empirical corrections and transformations. These corrections consist in carrying out calibration or adaptation of a long-term satellite- derived database by making empirical correlation with high-quality short-term measurement databases. Gueymard and Wilcox [8] performed corrections of long-term (15–30 years) databases (such as satellite-derived databases) using short-term (1–2 years) in-situ measurement database. Their method consists in calculating the ratio between these two databases for each period of time (monthly or daily) and then applied these ratios to the remaining years of the long-term time series satellite database. Beyer et al. [9] used a first-order polynomial relationship between measurements and satellite-derived database in order to perform correction on the satellite-derived database and to reduce its uncertainty. Mieslinger et al. [10] corrected four long-term satellite databases using a third-order polynomial relationship with short- term ground measurement databases. The correction method was tested in Plataforma Solar de Almería (southern Spain) and Tamanrasset (Algeria) and showed good results, i.e. the relative mean bias recorded between the corrected satellite database and the measurement database does not generally exceed ±2%. Another approach that could be adopted in order to reduce the uncertainty of solar energy prediction is by combining two or more databases by weighting each. A satellite-derived database could be combined with other solar radiation databases such as measurements or another satellite-derived database. Meyer et al. [11] suggested that the weight of a database is equal to the inverse of its uncertainty. He combined two satellite-derived databases and a short-term in-situ measurement database and he claimed that his method could be successfully applied only when the databases are independent and presented with sufficient uncertainty. These satellite-derived databases improvement approaches showed good results as regards increasing the accuracy of GHI. However, they usually require measurements of the solar radiation in a targeted site in order to be applied. Unfortunately, in the vast majority of the world's regions, solar radiation measurements are scarce, and the distances separating two measuring stations often exceed a 100 km [12]. This paper proposes to use the surrounding measuring stations (instead of the in-situ measurement) and extrapolate the measurements to the targeted site, using ground station models, in order to improve the accuracy of the satellite- derived databases. A wide range of GHI ground station models have been developed in the literature, ranging from simple extrapolation models such as nearby station model or inverse distance weighting (IDW) models [13] to more complex empirical models like Angstrom models (ÅMs) [14, 15], and recently, to artificial intelligence-based techniques and computational intelligence techniques such as artificial neural networks (ANNs) [16], genetic algorithm [17] and adaptive neuro-fuzzy methodology [18]. The IDW models use the measured GHI in the neighbouring region and the relative distance in order to perform prediction in a desired region. Many IDW models have been presented in the literature; the main difference between them is in the expression of the effective relative distance. Lefevre et al. [13] developed an IDW model by introducing new parameters in the effective distance expression. He applied his model to predict the monthly GHI over 586 European stations. The results of modelling were compared to measurements and revealed a root-mean-square error (rmse%) equal to 8%. Besharat et al. [19] collected 78 empirical models and classified them into four categories: sunshine duration (SD)-based models, cloud-based models, temperature-based models and other meteorological parameters-based models. Beshard et al. have applied these models to predict the monthly GHI in Yazd city, Iran. He showed that ELMetwally-Angstrom SD-based model [20] is the most accurate model with an rmse equal to 0.5385 MJ/m 2 /day. IET Renew. Power Gener. © The Institution of Engineering and Technology 2017 1