Improving solar radiation estimation in China based on regional optimal combination of meteorological factors with machine learning methods

Highlights

• Correlations were analyzed between meteorological factors and solar radiation.

• Some meteorological factors were exchangeable in solar radiation estimation.

• Machine learning methods could improve the accuracy in radiation estimation.

• The locally optimal combinations of input meteorological factors were determined.

• The extreme learning machine is more suitable for radiation estimation in China.

Abstract

The values of global solar radiation are important fundamental data for potential evapotranspiration estimation, solar energy utilization, climate change study, crop growth model, and etc. This research tried to explore the optimal combination of input meteorological factors and the machine learning methods for the estimation of daily solar radiation under different climatic conditions so as to improve the estimation accuracy. Based on the correlation between meteorological factors, different meteorological factor input combinations were established and the support vector machine method was used to estimate global solar radiation at 80 weather stations in four climatic regions of China mainland. The results showed that, the optimal combinations of input meteorological factors were different in the four different climatic zones in China mainland. Three meteorological factors of sunshine hours, extraterrestrial radiation, and air temperature had greater impacts on the solar radiation estimation. Adding the factor of precipitation could obviously improve the estimation accuracy in humid regions, but not remarkably in arid regions. Wind speed had very little influence on solar radiation estimation. The accuracies of machine learning methods were better than the Angstrom-Prescott formula and the multiple linear regression method. Among them, support vector machine and extreme learning machine were more appropriate. In some sites, the root mean square error of support vector machine method was even 20% less than that of the Angstrom-Prescott formula. In general, reasonable division of the areas and establishment of appropriate input combinations of meteorological factors according to the climatic conditions, combined with machine learning methods, can effectively improve the accuracy of solar radiation estimation.

Introduction

Global solar radiation is the main source of energy on the earth, as well as the basic driving force for various physical and biological processes on the earth surface [1]. Many natural phenomena on the earth are mainly caused by the difference, transformation, and transportation of solar radiation energy. Global solar radiation is of great importance to many research fields, such as reference evapotranspiration estimation [2], solar energy utilization [3], climate change [4], and crop growth models [5]. However, radiation observation equipment is usually expensive to construct and maintain, which makes solar radiation observation not as easy as sunshine hours, temperature, precipitation, and etc [6]. At present, among the more than 2000 national meteorological stations in China mainland, only about 100 stations have continuous observations of solar radiation. Thus, the limited number of existing observation stations of solar radiation can hardly meet the needs of scientific research and production [7]. Solar energy as a clean energy has been given a full attention [8]. The mainland of China has abundant solar energy resources [9] and Chinese government has also formulated a series of energy policies. Thus, the accurate estimation of global solar radiation is helpful for the development of new energy-related industries in China [10].

In order to solve the problem of insufficient observations of solar radiation, previous researchers usually used empirical models [11], machine learning models [12], and satellite-based methods [13] to estimate global solar radiation. The empirical model and machine learning model are more commonly used in practice because of their low cost and high estimation accuracy [14]. Over the past several decades, scientists in various countries have established different empirical models to estimate global solar radiation, including sunshine-hour-based models, temperature-based models, and models combining various meteorological factors [15]. According to previous studies, empirical models based on sunshine duration were generally better than the models based on temperature or other single meteorological factors [16]. The Angstrom-Prescott formula, which links the relative sunshine hours with the clear sky index, is the most widely used estimation method in the world. Thereafter, the later empirical models of solar radiation estimation were more or less based on the transformation of the Angstrom-Prescott model or the introduction of other meteorological factors [17]. In addition, there were many studies to calibrate and validate the Angstrom-Prescott model in different parts of the world [18]. However, traditional empirical models were not able to deal with the complex non-linear relationship between variables and other abnormal conditions [19]. In recent years, machine learning methods have been widely used in many fields with the development of computer technology [20]. In terms of solar radiation estimation, empirical models were not able to completely meet the different needs since meteorological data were always incomplete and unavailable in targeted regions. Research by Tymvios et al. [21] showed that the accuracy of the artificial neural network (ANN) method was better than the Angstrom formula. Chen et al. [22] compared the support vector machine (SVM) method with other empirical models and found that the error of the SVM method was smaller in solar radiation estimation based on temperature data. Thus, machine learning methods have become a promising way for solar radiation estimation due to its high accuracy and flexible combination of input variables.

When estimating solar radiation based on other common meteorological data and machine learning methods, it was necessary to select several relevant meteorological factors as model inputs. Chen and Li [23] used SVM to estimate solar radiation with the inputs of sunshine hours, temperature, relative humidity, and vapor pressure. It was found that the combination of sunshine hours and temperature had the highest estimation accuracy, while the input combination without sunshine hours had poor accuracy. The results of the research by Fan et al. [24] showed that the estimation accuracy of solar radiation by the input of sunshine duration, temperature, and precipitation was better than the results of using only sunshine duration. Meenal and Selvakumar [25] identified month, latitude, maximum temperature and sunshine hours as the most influential and relative humidity as the least influential input parameters in solar radiation estimation. In addition, some relevant studies showed that the inclusion of precipitation could help improve the estimation accuracy of solar radiation. Both the amount of rainfall (mm) and the binary form of rainfall event (1 for rainfall and 0 for no rainfall) were widely used [26]. The above research showed that different input combinations of meteorological factors might have great impacts on the estimation results. Yadav and Chandel [27] proposed that the prediction accuracy of neural network model depended on input parameter combination, training algorithm and architecture configuration, which also illustrated the importance of selecting appropriate input meteorological factors. Due to the different climate conditions in different regions, the correlation between local meteorological factors and global solar radiation was different. The input combinations of the best meteorological factors may also differ between regions, and the accuracy obtained with the same method of radiation estimation was also different. Alizamir et al. [28] used six machine learning models to estimate solar radiation. With the same method, the estimation errors of the Turkish site were larger than that of the US site. China has a vast area and complex internal climate [29]. However, previous studies usually focused on a specific region in China [30] or generally took the whole country as a single region [31], which did not reflect the differences within China mainland, So it is necessary to divide China mainland into several different regions based on the climatic conditions and then conduct the relevant research.

Current research focused mainly on the improvement of radiation estimation methods, while few studies concentrated on the selection of meteorological factors required for the models [32]. If there was only one set of input combinations, it obviously could not meet the estimation requirements under different meteorological conditions. It is still unclear what are the optimal combinations of input meteorological factors and the best machine learning methods for solar radiation estimation in different climatic regions of China. Based on climatic characteristics, this study divided China mainland into four different climatic zones, and applied different methods to estimate daily solar radiation for each of them. The objectives were to (1) explore the correlations among different meteorological factors in different climatic zones of China mainland; (2) to obtain the optimal combinations of input meteorological factors required for the solar radiation estimation based on machine learning methods in different climatic regions of China mainland; and (3) to assess the estimation accuracies of different machine learning methods based on the optimal combinations of input meteorological factors determined above in China mainland. Finally, the estimation accuracy of daily global solar radiation will be improved in China mainland.