Estimation of Solar Radiation for Gaziantep, Antakya and Kahramanmaraş Using Artificial Neural Network and Angström-Prescott Equations

In this study, we estimated monthly average global solar radiation on a horizontal surface for selected regions of Gaziantep (37°.06N, 37°.35E, 750m), Antakya (36°.15N, 36°.08E, 100m), and Kahramanmaraş (37°.35N, 36°.55E, 572m) from the east of the Mediterranean region. For this purpose, an artificial neural network (ANN) model and Angström-Prescott type equations related to sunshine hosurs were applied using the data measured. Firstly, a multi-layer feed-forward back-propagation model containing two hidden layers with tangent sigmoid (tansig) as the transfer function and one output layer with utilized a linear transfer function for the best ANN model was used for the modelling. Levenberg-Marquardt back propagation training algorithm (trainlm) was chosen as the training algorithm in the ANN model. A period of fifteen years (1993-2007) meteorological data taken from the Turkish State Meteorological Service were used for training (eleven years) and testing (four years) the network. Secondly, five Angström-Prescott type regression models (M1-5) were also used for estimating the monthly annual global solar radiation using parameters such as monthly average sunshine duration (hour), monthly average temperature (°C), relative humidity and solar declination angle (). Estimated data from ANN and Angström-Prescott type equations were compared with measured data using four different statistical methods such as R2, RMSE, MAPE and MSE. For the ANN model, R2, RMSE, MAPE and MSE statically indicators were found to be 0.990, 0.586, 4.105 and 0.343 for Gaziantep, 0.997, 0.287, 2.584, and 0.083 for Antakya, and 0.997, 0.414, 2.445 and 0.171 for Kahramanmaraş, respectively. For five different Angström-Prescott models (M1-M5) models, M3 model is the best performance for Gaziantep and Kahramanmaraş, while M5 model is the best performing for Antakya, according R2, RMSE and MSE MSE performance indicators. As can be seen from the statistical error results, the estimated global solar radiation data from both ANN and Angström-Prescott type models are in reasonable agreement with the actual meteorological values. We suggest that the developed both ANN and Angström-Prescott type models can also be used to predict solar radiation another location.