Artificial Neural Network Models for Solar Radiation Estimation Based on Meteorological Data

Abstract

The presence of solar energy in a particular area is closely related to meteorological parameters in that region. In this study, solar radiation estimation was carried out by using meteorological data recorded in different time series. Artificial neural networks (ANN) models were developed to determine the most effective parameters for solar radiation estimation. During the training and testing of ANN, site-specific meteorological data recorded by a meteorological station established in Hakkâri, Turkey, which has difficult climatic conditions, were used. To estimate solar radiation, basic input variables such as ambient temperature (T), wind speed (w), relative humidity (H), and atmospheric pressure (P), were modified by keeping the time series constant. To obtain the best estimation result, the number of input parameters of the input layer was applied with different possible input combinations, and the hidden layer neuron was changed to be multiples of the input layer (n, 2n, n²). The performance of all models was analyzed using statistical tools. ANN model, which has all possible combinations of input variables and determines the number of neurons in the hidden layer by framing the number of input variables, yielded the best estimation result. The performance indicator showed the mean square error (MSE) as the lowest value of 2.56 with all data entries and modeling the number of neurons in the hidden layer as n2. The mean absolute percentage error (MAPE) and relative root mean square error (rRMSE) values were obtained within the limits of high estimation accuracy in the network combination of T, P and H parameters as 1.99% and 1.91%, respectively. This study has revealed that increasing the variety and number of meteorological parameters affects solar radiation estimation success, but only basic meteorological parameters achieve very high estimation results.