Assessment of machine learning, time series, response surface methodology and empirical models in prediction of global solar radiation

Solar radiation (SR) knowledge plays a vital role in the design, modelling, and operation of solar energy conversion systems and future energy investment policies of the governments. However, these data are not measured for all regions due to the non-availability of SR measurement equipment at the weather stations. Therefore, SR has to be accurately predicted using various prediction models. In this research, four models from different classes are being used to predict monthly average daily global SR data. The models used in this study are based on a machine-learning algorithm (feed-forward neural network), empirical models (3 Angstrom-type models), time series (Holt-Winters), and mathematical model (RSM). As the prediction locations, four provinces (Ankara, Karaman, Kilis, and Sirnak) in Turkey are selected. The dataset including pressure, relative humidity, wind speed, ambient temperature, and sunshine duration is supplied from the Turkish State Meteorological Service and it covers the years 2008-2018. In the study, monthly average daily global SR data for the year 2018 is being predicted, and the performance success of the models is discussed in terms of the following benchmarks R-2, MBE, RMSE, MAPE, and t-stat. In the results, R-2 value for all models is varying between 0.952 and 0.993 and MAPE and RMSE value for all models is smaller than 10% and 2 MJ/m(2)-day, respectively. Evaluation in terms of t-stat value, no models exceed the t-critic limit. Considering all the models together, ANN has presented the best results with an average R-2, MBE, RMSE, MAPE, and t-stat of 0.9911, 0.1323 MJ/m(2)-day, 0.78 MJ/m(2)-day, 4.9263%, and 0.582, respectively. Then Holt-Winters, RSM, and empirical models closely followed it, respectively. (C) 2020 Elsevier Ltd. All rights reserved.