Estimation of daily global solar radiation using deep learning model
Introduction
Detailed knowledge about temporal and spatial variability of solar radiation (SR) is necessary in order to make plans for the present and the future because it directly or indirectly affects the life on our planet [[1], [2], [3], [4], [5]]. Many studies have indicated that the long-time database of SR is very important for detecting and the understanding climatic variations and global warming [[6], [7], [8], [9], [10], [11]]. SR data are also extensively used, especially in meteorology, agriculture, industry, engineering, health and even in tourism applications and researches [[12], [13], [14], [15], [16], [17], [18], [19]]. Although SR is one of the most commonly monitored meteorological data, the numbers of measurement stations are still sparse, particularly in developing and underdeveloped countries. Moreover, some measurements could be unreliable or questionable basically due to maintenance and calibration problems of used instruments. Direct measurement of SR is classically achieved by using pyronometers and/or actinographs. Measurements that are more accurate could be done by using with new generation instruments having automatic sensors, but the dense networks of this type instrument are available only in a few countries mainly due to its cost. In most of the countries, still actinographs have been carried out for this purpose. For example, 192 stations are equipped with Siap, Muller, and Fuess actinographs in Turkey but there are only 11 stations equipped with pyronometers [20]. Total of 51 stations installed with actinographs all over Chile [21]. Ten solar stations were installed with actinographs of the Robitzch type, distributed in Colombia [22]. On the other side, Aksoy [23] observed that the measurements obtained from the actinographs exhibited 15% and 42% error rates, respectively, in annual and monthly averages. Joaquín and Carlos [22] showed that the most frequent error percent for the actinograph was from 10% to 20% and it corresponds to the 56% of the days of the year. They also found that the greatest error frequency is estimated in the SR range between 3001 and 4000 Wh/m2, which corresponds to the 21% of the days of the year. These facts implies that the global solar radiation (GSR) has to be correctly estimated over any area where the direct measurements not possible or reliable.
Up to now, researches have shown that there are three distinct approaches for estimating the GSR [24]. The first and the simplest methodology deals with empirical relationships between the ratio of GSR on a horizontal surface on the ground and at the top of atmosphere and relative sunshine hours or cloud cover or temperature or their combinations (hybrid forms) with other related parameters such as precipitation, humidity, soil temperature, evaporation and geographical parameters [[25], [26], [27], [28], [29], [30], [31], [32], [33]]. Such type of approach has been extensively used due to its simplicity and accuracy, thus, hundreds of studies have been reported in literature. The second one take into account physical interactions of the SR with the atmospheric constituents, which is known as complex radiative transfer models [[34], [35], [36], [37]]. The algorithms of these models are generally too complex and require exhaustive collection of meteorological data (such as cloud type, cloud top height, cloud coverage, temperature, water vapor, ozone, aerosols etc.) which could be obtained rarely. Therefore, this type of approach has not been widely used and limited numbers of studies are found in literature. The third and more recent approach uses machine learning models in which information passes through from inputs to output [[38], [39], [40], [41], [42], [43]]. Inputs of these models are generally selected from most relevant meteorological parameters such as sunshine duration, cloud cover, temperature, humidity, precipitation, wind speed, evaporation and astronomical values such as extraterrestrial radiation, theoretical daily sunshine duration, latitude, longitude, day of a year, month of a year. A detailed overview of machine learning methods for SR prediction is discussed in the study of Voyant et al. [44]. In that study various machine-learning methodologies such as artificial neural networks (ANN), support vector machine (SVM), decision tree and many others, except deep learning, are analyzed and the results corresponding to many single and combining algorithms are presented and compared. It was also reported that the ANN, SVM and decision trees approaches are much more preferred than other methods. The reader can also find an extensive review about ANN based techniques in the study of Yadav and Chandel [45].
Recently an effective classification machine learning method, which can be thought as an extension of ANN method, known as deep learning (DL) method is introduced. It is shown that in some researches on classification or object detection competitions, deep networks obtain better results than those using SVM or other methods [46]. DL is generally used in medical imaging, speech recognition, natural language processing, autonomous driving and computer vision and it is now intended to be applied to other areas for forecasting studies [[47], [48], [49], [50], [51]]. It uses neural networks structures to represent the data. Using multiple hidden layers gives the opportunity to learn the most complex relations over the data.
To our knowledge, there is only one study that a DL model is employed for estimating the solar irradiance forecasting for a seasonal basis [52]. In that study the model was trained, using self-regulated particle swarm optimization (SPSO) algorithm. The purpose of current study is to evaluate the success of DL approach for estimating daily GSR using the daily meteorological and astronomical data covering the years between 2001 and 2007. We have used one astronomical parameter, extraterrestrial radiation (H0), and four measured meteorological parameters sunshine duration (SD), cloud cover (CC), maximum temperature (Tmax) and minimum temperature (Tmin) to get the GSR as output. Totally, 16 different combinations of inputs are constructed to find the optimum combination using the data belonging to 34 stations, which are almost uniformly distributed in Turkey (see Fig. 1). The simulated results corresponding to the best combination are compared with the ground-measured values and previous studies done for Turkey and other regions.
Section snippets
Data and study area
Turkey located at 26°–45° East and 36°–42° North, having area of 783,562 km2, is selected as the general study area for this study (see Fig. 1). The average altitude of country is 1130 m above sea level and the annual mean temperature changes from 3.6 °C to 20.1 °C [53]. The annual average rainfall is about 648 mm with an annual variation ranging from 295 to 2220 mm. Turkey is typically divided into seven different climate regions: Marmara, Black Sea, Aegean, Central Anatolian, Southeastern
Results and discussion
In this study, DL models were trained by using the best related and most widely used daily astronomical and meteorological parameters, namely, H0, SD, Tmin, Tmax and CC in order to estimate the daily GSR. In order to produce the best possible simulated results and see the effect of input attributes on simulated results 16 different combinations were constructed as shown in Table 2. Since H0 is the daily extraterrestrial SR and can be calculated for any area without needing any measurement it is
Conclusions
In this study, a simple DNN model was developed to estimate daily GSR for 34 stations, which are homogeneously distributed and represent all possible climatic conditions in Turkey. In order to get the daily GSR as output 16 different combinations were constructed using inputs H0, SD, Tmin, Tmax and CC. Our results indicated following outcomes;
- ➢
SD is the most important input attribute because combinations that include SD produced relatively good simulated results and lower statistical errors than
Conflicts of interest
The authors declare that there are no conflicts of interest regarding the publication of this paper.
Acknowledgments
The authors would like to thank the all reviewers for useful comments and suggestions. We are also grateful to Turkish State Meteorological Service for providing the meteorological data.
References (85)
- et al.
Comparison of estimates of daily solar radiation from air temperature range for application in crop simulations
Agric For Meteorol
(2008) - et al.
A geometric solar radiation model with applications in agriculture and forestry
Comput Electron Agric
(2002) - et al.
Statistical analysis for the characterization of solar energy utilization and inter-comparison of solar radiation at two sites in Cyprus
Appl Energy
(2017) - et al.
Global dimming: a review of the evidence for a widespread and significant reduction in global radiation with discussion of its probable causes and possible agricultural consequences
Agric For Meteorol
(2001) - et al.
Global solar radiation measurements in Maceió, Brazil
Renew Energy
(2005) - et al.
Estimation of solar radiation for use in crop modelling
Agric For Meteorol
(1998) Modeling global solar radiation using Particle Swarm Optimization (PSO)
Sol Energy
(2012)- et al.
Improving estimation of hourly, daily, and monthly solar radiation by importing global data sets
Agric For Meteorol
(2006) Comparison of artificial intelligence techniques via empirical equations for prediction of solar radiation
Comput Electron Agric
(2015)Estimated monthly average global radiation for Turkey and its comparison with observations
Renew Energy
(1997)