A hybrid method design that considers different insolation conditions for solar radiation forecasting

The world's ever-increasing population and developing technology cause the demand for energy to increase continuously. The usage of clean energy sources has become a necessity in order to satisfy both environmental sensitivity and increasing energy demands. Although solar energy is one of the most important of these sources, the amount of energy produced also varies because of the variable and intermittent nature of the radiation. As a consequence, in order to determine the amount of energy to be produced from the sun and plan production, the solar radiation value must first be determined correctly. In this study, a hybrid method for predicting solar radiation is proposed, wherein the prediction model is determined based on the clearness index. The study used two-year (2013-2014) solar radiation data of the province of Mardin obtained from the Turkish State Meteorological Service (TSMS). The data were divided into two parts: training and testing, with solar radiation data from 2013 used for training and 2014 utilized for testing. As predictors, Artificial neural networks, NARX networks, and Ridge regression methods were used, and the training data were modeled with all three approaches in the first stage of the study. While modeling the irradiance value one hour later, the current irradiance value, the irradiance value one hour ago, and the extraterrestrial irradiance value one hour later were used as inputs. The clearness index was determined into three ranges; slightly cloudy, cloudy, and mostly cloudy. The training data were modeled with three methods used as estimators, and the success of each method was examined in each defined clearness index range. After determining the more successful methods in the selected intervals, the clearness index time series was modeled using artificial neural networks. As a result, in the hybrid prediction algorithm, the clearness index is first estimated using artificial neural networks, and then the future solar radiation value is predicted by using the most successful model within the predicted clearness index range. Experimental results show that more successful predictions are made with the proposed hybrid method than when models are used individually.