Türkiye genelinde Bahel modeli ve yapay sinir ağları ile güneş radyasyonu tahmini

Abstract

Bulunduğu coğrafi konum dolayısıyla, dünyada güneş enerjisi potansiyeli yüksek olan ülkeler arasında bulunan Türkiye'de güneş enerjisinden mümkün olduğunca etkin bir biçimde faydalanılması gereklidir.Güneş enerjisi üretiminde gelecekteki saatlik küresel güneş radyasyonunun doğru tahmin edilmesi büyük önem taşır. Bu tür tahminlere mevcut ölçüm verilerinin eksiksiz ve doğru olarak girilmesi, tahminin başarımını doğrudan etkileyen parametrelerin başında gelmektedir.Ancak saatlik küresel güneş radyasyonu ölçümleri ülke genelinde sınırlı sayıda meteoroloji istasyonunda gerçekleştirilmekte olup, ölçüm verilerinde sıklıkla veri kayıpları ve ölçüm hatalarına rastlanılmaktadır Bu durum saatlik bazda güneş radyasyonu modellerinin geliştirilmesi ihtiyacını doğurmaktadır. Bu sebeple saatlik küresel güneş radyasyonu verilerindeki eksik ve hataları giderebilmek adına Bahel (kübik Angstrom-Prescott) modeli Türkiye'deki her iklim bölgesini temsil edecek şekilde 8 istasyon için uygulanmıştır. Öncelikle yıllık kalibrasyon yöntemi denenmiş, ancak Angström-Prescott temelli modellerin genel sıkıntısı olan katsayıların sabitliği problemi nedeniyle farklı ölçekteki kalibrasyon yöntemlerinin de denenmesine karar verilmiştir. Buna göre tezin son halinde model uygulamasında aylık, mevsimlik ve yılık olacak şekilde 3 ayrı uygulama tekniği denenerek bunlar kendi içlerinde kıyaslanmıştır. Elde edilen sonuçlara göre mevsimlik kalibrasyon yöntemi uygulanan Bahel modeli en yüksek başarım değerine sahiptir. Bahel modellerinin başarımı doğrudan istasyonların bulunduğu bölgedeki iklimsel özelliklere de bağlıdır. Mevsimlik Bahel modeli için 2016 yılında tüm istasyon ortalamasında R2 değeri 0,9430, RMSE değeri 59,283 ve MBE değeri -3,661 olarak bulunmuştur. Ardından, seçilen mevsimlik Bahel modeli üzerinde yapay sinir ağı ile model çıktılarını iyileştirme işlemi uygulanmış ve iyileştirilmiş Bahel modeli ile başlangıç Bahel modeli sonuçları kıyaslanmıştır. Yaklaşık 500 yapay sinir ağı denemek ve eğitim seti olarak 2011-2015 yılları arası verileri yerleştirmek suretiyle yapay sinir ağının eğitimi gerçekleştirilmiştir. Yapay sinir ağlarıyla iyileştirilmiş 2016 yılı Bahel modelinin 0,9587 R2, 52,4117 RMSE ve -0,796 MBE değerlerine sahip olduğu belirlenmiştir.Elde edilen sonuçlar değerlendirildiğinde yapay sinir ağları uygulanarak iyileştirilmiş kübik Angström-Prescott modelinin, sadece kübik Angström-Prescott uygulamasına nazaran daha iyi sonuçlar verdiği tespit edilmiştir.2011-2016 yılları datası kullanılarak gerçekleştirilen bu çalışmada T.C. Orman ve Su İşleri Bakanlığı Meteoroloji Genel Müdürlüğü tarafından sağlanan veriler kullanılmıştır.

Energy has been among the most basic requirements of humankind for many years. Energy has many indispensable qualities in many fields such as heating, production and transportation. Moreover, energy has been generated from fossil and unsustainable energy resources such as coal, oil and natural gas in the past. However, after the industrial revolution, the rising uncontrolled use of fossil resources has led to the reduction of these resources and increase in environmental pollution.Today, with the increase of robotic technologies, the use of energy in production systems has reached its peak. Every activity to reduce environmental pollution, which is a direct result of industrial production, also has affected the rising of energy consumption. For example, electronic billing, which has become popular to reduce paper consumption, increased the computer use and the electricity consumption. Furthermore, energy is becoming indispensable for environmental pollution control because of the continuous process of garbage in solid waste storage areas throughout the world.Due to its geographical location, Turkey is among the countries with the highest solar energy potential in the world. Therefore it is important to utilize solar energy in Turkey as effectively as possible.In the production of solar energy, accurate forecasting of the hourly global solar radiation is of great importance. Complete and accurate entry of the current measurement data into such estimates is among the parameters directly affecting the estimation performance.However, hourly global solar radiation measurements are carried out in a limited number of meteorological stations nationwide, and data loss and measurement errors are frequently encountered in data measurement. This necessitates the development and improvement of solar radiation models on an hourly basis.Although Angstrom-Prescott-based solar radiation models have been implemented for many years, the models on the hourly scale are still limited. The relationship between the duration of sunshine and global solar radiation allowed the diversification of Angström-Prescott models and the formation of square, cubic, logarithmic versions. The greatest advantage of the Angström-Prescott model is that it only allows calculating using a single observation data (sunrise time). Models that use more than one measurement data (including temperature, humidity, etc. parameters) can affect the calculations because of the aggregation of unwanted measurement errors. With the use of models with single observation values, the errors caused by the measurements will be minimized as much as possible.In this study, a 6-year data period from January 2011 to December 2016 was selected and the data was obtained from T.C. Ministry of Forestry and Water Affairs, Meteorology General Directorate (MGM) Meteorological Data Information Sales and Presentation System (MEVBİS). It was acquired free of charge for scientific study purposes through Istanbul Technical University.Within the scope of the thesis study, the stations that have solar radiation data between the years 2011-2016 were examined and the stations with the minimum missing data and the representative features in terms of spatial characteristics were determined.Accordingly, a total of 8 stations were selected, two stations from the Black Sea region and the rest of the stations are from every other region, and the studies were carried out with data from these stations. These stations are Ağrı, Ankara, Artvin, Kastamonu, Isparta, Kırklareli, Mardin and Muğla stations.Consequently, it was decided to apply the Bahel model, also known as the cubic Angström-Prescott model, for the hourly global solar radiation calculations carried out within the scope of the thesis. Firstly, the annual calibration method was tried but it was decided to try calibration methods of different dimensions because of the problem of the invariability of the coefficients which is the general problem of Angström-Prescott based models. According to this, 3 different application techniques have been tried in the final application of the thesis as monthly, seasonal and yearly model application.The 2011 – 2016 data period for empirical calculations was initially divided into two groups, training and test set. Empirical calculations between 2011 and 2012 were selected as the training set, while between the years 2013 – 2016 were the test set.During the study, it was observed that the night hours with S / S0 sunshine index was equal to 0/0 constituted approximately one third of the total amount of data, and these values were not included in the training set because they caused deviation in the empirical formula.The Bahel model calculations for the selected time periods were carried out with the MATLAB® program and there were 17 coefficient groups (1 yearly, 4 seasonal and 12 monthly) for each station.At all stations, the annual Bahel model shows about 3 times higher results than actual measurements. This is caused by the stability of the model coefficients throughout the year. The seasonal and monthly models, where the coefficients are more variable during the year, give similar results and give slightly lower results than the actual observations.R2, RMSE and MBE values were used in error assessment of global radiation estimates. When the R2 values of the Bahel models are compared, it is seen that the annual R2 values are very low compared to the monthly and seasonal calibrations for each station and the average R2 value is 0.7958. R2 values of monthly and seasonal calibrations are close to each other, and the seasonal R2 value is slightly higher than monthly value. While the monthly average of all stations R2 value is 0.9363, seasonal average R2 value is as found 0.9430. When the RMSE values are examined, the annual calibration values give an average value of 365,218. In the monthly calibration, the RMSE value is 63,218, while the seasonal RMSE is 59,283, giving the lowest value among all calibrations. When the MBE values were examined, the annual calibration values were positive at all stations and the average value was 495,953. For monthly and seasonally calibrations, the mean values of MBE were calculated as -2,5574 and -3,661, respectively.According to the results obtained, the annual model is the worst model among the 3 models. Among the seasonal and monthly models, the seasonal Bahel model with an average value of 0.9430 R2, 59.283 RMSE and -3.661 MBE was the best model selected. It has been decided to implement Model Output Statistics (MOS) to this model.Model Output Statistics is essentially a compilation of statistical modeling techniques and has been used as a post - processing method in the 70's to improve the results of numerical weather prediction models. Linear, polynomial, logarithmic regression techniques can be used for a statistical method for creating Model Output Statistics, as well as various artificial intelligence approaches such as artificial neural networks, support vector machines can be used.In this thesis, it has been decided to improve the Angström-Prescott model outputs with artificial neural networks. Artificial Neural Networks (ANN) constructed similar to biological neural networks can be defined as systems that allow many simple processors to generate complex interconnection and perform concurrent processing. Learning occurs in artificial neural networks by the connections between neurons just like the biological neural networks.In order to improve the model outputs, artificial neural network application has been performed on the seasonal model which gives the best result among the Bahel models applied at different time scales within the thesis study. The codes for ANN are written in MATLAB® program and related calculations are done.Artificial neural networks applied within the scope of the thesis study are aimed to approximate the outputs of the seasonal Bahel model to the real solar radiation measurement values. For ANN application, 5 years period between 2011-2015 is selected as the training set and 2016 year test set is selected as the test set. For the purpose of training, Bahel model results and seasonal global solar radiation measurements obtained from the MGM are given to the artificial neural network for a 5-year period. Pure-linear function and Logarithmic Sigmoid function are used together as activation functions.Using the Levenberg Marquardt algorithm, the ANN tries to approximate the Bahel model results to real observations results. The Levenberg-Marquardt algorithm, one of the Quasi-Newtonian methods, is known to bring a different approach to the Hessian matrix and to offer rapid training to ANNs.The trainings were created by testing 500 ANN in the training of each ANN application, so it was ensured that ANN was tested in a sufficiently wide range and that the error was at minimum in the ANN detection. At the end of the training, the ANN structure with the best results was selected and thus the training was finalized. During the training, the parameters such as the number of nods in the input layer and the number of hidden layer neurons in ANN are changed. In the whole of the stations, the number of nodes is found as 2 in the input layer. This indicates that the internal dependency of the system is dependent on 2 input values, meaning 1 hour time period.Since ANN is operated on a seasonal basis for each station, a total of 4 ANN are operated in each station. Thus, a total of 32 ANNs were created for 8 stations.The artificial neural network giving the highest R2 value among the applied ANNs was selected and the results of the estimation of the selected ANN and the initial Bahel model results were compared.When R2 values are examined, it is seen that ANN application improves all seasonal data of all stations. ANN made the biggest improvements on the Kırklareli values in winter and increased the R2 value of 0.7968 to 0.8978. The least improvement was at the Mardin station in springtime and the R2 value of 0.9546 rose to 0.9828. R2 values are between 0.7743 and 0.9851 for the Bahel model and between 0.859 and 0.9937 for ANN. When the RMSE values are examined, ANN has made the biggest improvements on the Isparta station in summer and decreased the RMSE value of 86,0752 to 59,2318. The least improvement was in the Mardin station during the winter months and the RMSE value of 59.33 decreased to 59.0241. RMSE values are between 43,0597 and 94,6689 for Bahel model and between 32,247 and 75,879 for ANN. When the MBE values are examined, it is seen that the values for the seasonal Bahel model change between -24,8372 and 17,3405, while for ANN these values change between -14,9072 and 11,964.As a result of the ANN application, the mean values of all stations were 0.9587 R2, 52.4117 RMSE and -0.796 MBE. If all the results are evaluated, it can be said that ANN application gives better results than Bahel model. Hereby, the hybrid model which is the application of ANN post-processing for the cubic Angström-Prescott model outputs, provides much better results than just the Cubic Angström-Prescott model application.