A Deep Learning Algorithm for Solar Radiation Time Series Forecasting: A Case Study of El Kelaa Des Sraghna City Mohammed Ali Jallal 1* , Abdessalam El Yassini 1 , Samira Chabaa 1,2 , Abdelouhab Zeroual 1 , Saida Ibnyaich 1 1 I2SP Research Team, Physics Department, Faculty of Sciences Semlalia, Cadi Ayyad University, Marrakesh 40000, Morocco 2 Industrial Engineering Department, National School of Applied Sciences, Ibn Zohr University, Agadir 80000, Morocco Corresponding Author Email: mohammedali.jallal@edu.uca.ac.ma https://doi.org/10.18280/ria.340505 ABSTRACT Received: 26 July 2020 Accepted: 12 October 2020 Nowadays, the studies that address solar radiation (SR) forecasting tend to focus on the implementation of conventional techniques. This provides good results, but researchers should focus on the creation of new methodologies that help us in going further and boost the prediction accuracy of SR data. The prime aim of this research study is to propose an efficient deep learning (DL) algorithm that can handle nonlinearities and dynamic behaviors of the meteorological data, and generate accurate real-time forecasting of hourly global solar radiation (GSR) data of the city of El Kelaa des Sraghna (32°2’53”N 7°24’30”W), Morocco. The proposed DL algorithm integrates the dynamic model named Elman neural network with a new input configuration-based autoregressive process in order to learn from the seasonal patterns of the historical SR measurements, and the actual measurements of air temperature. The attained performance proves the reliability and the accuracy of the proposed model to forecast the hourly GSR time series in case of missing values detection or pyranometer damage. Hence, electrical power engineers can adopt this forecasting tool to improve the integration of solar power resources into the power grid system. Keywords: artificial intelligence, global solar radiation, deep learning, Elman neural network, forecasting, time series 1. INTRODUCTION The power companies and decision-makers are still facing the challenging task of incorporating renewable power shares into electrical grids. Green power technology is a promising alternative for electricity production, particularly solar photovoltaic (PV) technology, since it is a clean source of electricity [1]. In the last years, by realizing the benefits of PV technology, large PV solar plants have been implemented worldwide. In order to commercialize and use this technology on a large scale, several issues have to be resolved. Most of the challenging issues to boost the integration of solar power production into the electrical grid are its dependency on weather conditions. Hence, an accurate prediction of solar radiation measurements is a mandatory step for controlling, sizing, and designing PV systems. Solar radiation estimation has been extensively considered by implementing various estimating approaches. Statistical techniques-based time series regression [2] and data driven- based artificial intelligence (AI) algorithms [3] are two commonly employed estimating tools. Among the most suitable techniques implemented to SR time series forecasts are AI techniques since these techniques do not necessitate a clear explanation of the sun's physical characteristics to carry out this type of prediction. The artificial neural networks (ANNs) have experienced a vital evolution during the last years and is extensively employed in different engineering areas, such as the green power sector [4, 5], medicine [6, 7], civil engineering [8], computer vision [9, 10], and so on. According to our literature review, several researches based on ANNs and other machine learning algorithms have been conducted to perform accurate solar radiation forecasts. Jallal et al. proposed a new ensemble technique based on artificial neural network to estimate the hourly global solar radiation time series, where the reliability is evaluated using 7 years of measurement of several meteorological data [11]. The same authors develop an artificial neuron combined with an autoregressive process to predict GSR parameter. The developed model performed accurate predictions during 6 years, which is reserved for the testing process [12]. Kisi et al. applied a dynamic neuro-fuzzy inference forecasting approach based on mono-variate AT structure to estimate GSR data [13]. Mousavi et al. have proposed a new training strategy based on simulated annealing optimizer to train the ANN’s parameters in order to generate precise daily GSR predictions [14]. Xue et al. compare the training performance of genetic and particle swarm optimization (PSO) algorithms to tune ANN’s weights and biases for forecasting accurately daily GSR measurements. The authors demonstrate the efficiency of adopting PSO optimizer than the genetic algorithm [15]. Garcia-Hinde et al. developed a smart learning approach based on reducing the dimensionality curse in the input layer for generating precise solar radiation time series estimation [16]. In the reviewed studies for SR time series prediction, we have faced the following issues: • The temporal features of solar radiation time series have not been considered by most of the reviewed Revue d'Intelligence Artificielle Vol. 34, No. 5, October, 2020, pp. 563-569 Journal homepage: http://iieta.org/journals/ria 563