# Forecasting of daily global solar radiation

**Contents**show

As solar radiation is the starting point for improving the use of renewable energy. In recent years, several methods have been used to predict daily solar radiation, similar to artificial intelligence and hybrid models. In the present study, a wavelet-coupled Gaussian Process Regression (GPR) (W-GPR) model was adopted to forecast daily total solar radiation received on horizontal surface components in Ghardaïa (Algeria). For this purpose, a statistical period of six years (2012-2012), five years (2012-2016) was considered to train the model and last year (2017) to test the model predicting the daily total solar radiation. Different combinations of input data and different types of wave mother have been evaluated. The results showed that the hybrid wavelet-GPR model is more accurate than the GPR model in the global based on the minimum air temperature, relative humidity and the duration of the sun gives the best results in the relative error in error box (rRMSE), medium term absolute bias error (MBE) RMSE error and correlation coefficient (r). The values obtained from these indicators are …..% and ….. MJ / m2 and …….%, respectively.

### Introduction

The solar field is a set of data that describes the evolution of solar radiation available in a particular place during a given period. It is used to simulate the potential operation of solar energy systems. The study of the solar field is the starting point for any solar energy investigation. The long-term detailed and detailed identification of available global solar radiation data (GSR) is necessary in various forms to design and implement a good solar system [10]. Inadequate number of meteorological stations in which global solar radiation is recorded [11] [1]. In addition to the lack of access to solar radiation measurement stations, researchers have encouraged the development of suitable solar radiation models. In this context, several models have been proposed for forecasting global solar radiation using different meteorological data [2].

In literature, many researchers proposed empirical formulas for the purpose of estimating solar radiation, based on available meteorological parameters such as relative humidity, air temperature, wind speed, duration of sunlight, etc. Forecast models based on the solar period are the most accurate models [2-12]. Angstrom-Prescott [13] proposed a reference formula for this category, in which he defined a simple relationship between global solar radiation and the duration of the sun\’s rays. The use of air temperature and humidity is also widely expected. Due to the inability of these models when high tests are needed and to overcome this deficiency, the requirement for precise methods even more important.[3]

Time series models such as ARIMA and SARIMA have been used to predict SOLAR by many researchers (C. CRAGGS et [4l. 1999; [Irmak et al. 2003 [5]; [Christophe Paoli. [6]; [Yang Dazhi. [7]. However, the use of high-resolution daily data is a limitation of the SARIMA model, the non-convergence of the model because the season is 365 days. Encourage researchers to use other modeling techniques, including statistical learning machines such as support vector machines (SVM), artificial neural networks (ANN), Gaussian process regression (GPR), adaptive neuro-fuzzy (ANFIS). Banguanem et al. [17]. (2004) used RBF neural networks to estimate the daily global solar radiation in Madinah (Saudi Arabia) and showed that the RBF was capable of predicting daily high-resolution global solar radiation. Senkal et al. [18] used MLP (Multiple Layer Perceptron) to forecast global solar radiation on horizontal surfaces in 12 regions in Turkey In their research, two types of delay (weekly and annual) were used, and their results showed that MLP accordingly showed that the advanced MLP model gave better predictions with the RMSE than 91W / m2, compared to physical methods (RMSE = 125W / m2( [germou]

Some researchers used machine learning algorithm called extreme learning machine (ELM), and has received considerable attention in the scientific field, because of its performance, rapid implementation and ease of training. Shamshirband et al] [26] Use KELM to model global daily solar radiation. Developed three models KELM on the basis of maximum air and minimum temperature. Many of the tests are made, and the results reveal that the KELM model achieves higher accuracy, especially when using Tmax and Tmax-Tmin inputs (MABE = 1.35MJ / m2, RMSE = 2.02MJ / m2, RRME = 11.25MJ / m 2, R2 = 90.57%) [GERMOU]. The family of Bayesian nonparametric methods has provided an excellent alternative to neural networks and SVMs for the estimation of many biophysical parameters.Of useful ways particularly in this context is the decline Gaussian process (GPR) [12]. One of the main roads in this context is the decline Gaussian process (GPR) [12].

The GPR algorithm has received considerable interest in the machine learning community for applications such as model multivariate regression, and classification problems, which have been successfully used in Earth sciences and remote sensing problems in recent years. [13] – [ 16]. GPR is simpler and generally more robust than other statistical regression tools. In addition to good computational performance and stability, GPR requires a relatively small training data set, which can adopt highly flexible kernel functions and provide better prediction areas [11]. Sancho et al. (2014) [10] Use model GRP applied to predict the global solar radiation daily. The findings show that the GPR model works better than conventional methods.

The ANFIS model was used by Rahoma et al. [28] with the aim of predicting daily global radiation in Egypt. He used data from ten years (1991-2000) to develop the model. This method is a combination of logical and neural network techniques. The results obtained show that the Fuzzy model gives better accuracy (R2 = 96%, RMSE K >1 were deleted ([60]). [13](2) If rew was a month with more than 5 days missing and the values of global solar radiation meter is reliable, it has been withdrawn this month completely from the data set. However, if the number of missing and incorrect values in the month, less than 5, were those values to replace the appropriate values based on interpolation ([60]). [13] clearness Index (KT) is the division of global solar radiation events on a horizontal surface (H) on extraterrestrial solar radiation on a horizontal surface (Ho).For clarification: Ho has a fixed value for each specific day in any geographic location However, solar attenuation occurs as radiation passes through the atmosphere due to some atmospheric phenomenon such as cloud extinction, aerosol extinction, , Rayleigh scattering and so on. Therefore, in the available data all values of H should be smaller than Ho, which means KT < 1.