IFAC PapersOnLine 50-1 (2017) 14137–14142 ScienceDirect ScienceDirect Available online at www.sciencedirect.com 2405-8963 © 2017, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Peer review under responsibility of International Federation of Automatic Control. 10.1016/j.ifacol.2017.08.1856 © 2017, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Keywords: direct normal irradiance, fractional cloud cover, clear-sky index, concentrating solar power, adaptive network-based fuzzy inference system. 1. INTRODUCTION Forecasting the solar resource using a sky imager, due to the ability of such device to provide a detailed cartography of the sky, is gaining interest at a fast pace. However, for now, only a few studies are dedicated to forecasting Direct Normal Irradiance (DNI) using sky-imaging data. Keep in mind that DNI (I ), which can be defined as the direct irradiance received on a plane normal to the Sun, can be split into two terms, i.e. clear-sky DNI (I cs ) and the clear- sky index (k c ) (1): I = I cs · k c (1) I cs is about the solar power received at ground level per unit of area, at a specific location, when there is no cloud occulting the Sun. k c is derived from the attenuation of clear-sky DNI caused by clouds. These few studies have been conducted by a group of researchers belonging to the University of California in San Diego (UCSD). First, Marquez and Coimbra proposed in 2013 a deterministic forecasting model for DNI based on the cloud cover estimated in a region of the sky with potentially sun-blocking clouds (Marquez and Coimbra, 2013). In this model, clear-sky DNI is supposed to be constant (I cs = 900 W m -2 ) during the time interval of interest (i.e. from 10 a.m. to 2 p.m.). The forecast horizon is up to 15 minutes ahead. Cloud pixels are identified by using the hybrid algorithm developed by Li et al. (Li et al., 2011) whereas cloud motion is estimated using the PIV (Particle Image Velocimetry) method (Marquez and Coimbra, 2013). Clearly, performance is degraded by not taking changes in I cs into consideration. Following this first study, a new (deterministic) forecasting model for DNI has been developed by the Coimbra’s group (Chu et al., 2013). The forecast horizon is up to 10 minutes ahead. I cs is estimated using a 8th-order polynomial of the cosine of the Solar Zenith Angle (SZA). Note that a 3rd-order polynomial has proven to do the trick. Both the cloud map and cloud motion are estimated in the same way Marquez and Coimbra have done (Marquez and Coimbra, 2013). In addition, a hybrid forecasting model that makes use of two artificial neural networks has been developed: the first neural network is used in case of high variability whereas the second one is used when variability is low. The authors highlight that two artificial neural networks, instead of only one network dealing with all types of variability, allow DNI to be forecasted with much better accuracy. The cloud cover in regions of the sky having great potential for affecting ground DNI values as well as last-minute values of DNI are serving as model inputs. Only data that agrees with the following condition are considered: cos(SZA) < 0.6, that is to say SZA  53 ◦ . As a result, DNI is forecasted when the Sun is near the zenith. In that case, perspective effects are lower. A short time afterwards, a novel method for cloud tracking that makes use of TSI images to forecast DNI up to 20 minutes ahead is presented (Quesada-Ruiz et al., 2014). Cloud pixels are identified by using the hybrid algorithm developed by Li et al. (Li et al., 2011) and I cs is computed from a 8th-order polynomial of the cosine of the solar zenith angle. The authors introduce both a sector method to detect the motion of potentially sun- blocking clouds and an adjustable-ladder method, which is based on a size-adjustable set of grid elements that focus on sky regions of greatest potential for affecting ground DNI values. The model performs better than when ∗ PROMES-CNRS (UPR 8521), Tecnosud, Rambla de la thermodynamique, 66100 Perpignan, France ∗∗ University of Perpignan Via Domitia, 52 Avenue Paul Alduy, 66860 Perpignan, France Tel.: +33468682257; e-mail: stephane.grieu@promes.cnrs.fr Abstract: In a context of sustainable development, interest for Concentrating Solar Power (CSP) is growing rapidly. One of the most challenging topics is to improve solar resource assessment and forecasting in order to optimize power plant operation. Indeed, in CSP plants, electricity generation is directly impacted by both availability and variability of the solar resource and, more specifically, by Direct Normal Irradiance (DNI). Moreover, in the framework of the CSPIMP research project, PROMES-CNRS has developped a sky imager able to provide High Dymanic Range (HDR) images. As a result, the present paper deals with the short- term forecasting of DNI using sky-imaging data. Preliminary results highlight that models (in particular based on artificial intelligence tools) that make use of the fractional cloud cover have the potential to outperform persistence models in terms of forecasting accuracy. Julien Nou ∗ Rémi Chauvin ∗ Julien Eynard ∗,∗∗ Stéphane Thil ∗,∗∗ Stéphane Grieu ∗,∗∗ Towards the short-term forecasting of direct normal irradiance using a sky imager