Prediction of Daily Solar Radiation Using Anns for Selected Provinces in Turkey

In this study, hourly solar radiation was predicted using ANNs. Standard backpropagation and backpropagation with momentum were used for training neural network. In order to train the neural network, meteorological data along the years (1998-2008) belonging to the six cities (Antalya, Konya, Mersin, Mugla, Sanliurfa and Sivas) in Turkey were used as training (five stations) and testing (one station) data. Input parameters are chosen as latitude, longitude, altitude, day of the year and mean temperature while the output parameter is the solar radiation. The various networks designs were tested and then the most successful network was found as tree layer network with 10 neurons in hidden layer. The logistic sigmoid activation function was used for both hidden and output layers. To estimate difference between measured and estimating values, mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination, R 2 were determined. The iteration numbers for the standard backpropagation and backpropagation algorithm with momentum were respectively obtained as 15000 and 7500. The RMSE, MAE and R 2 results for the standard backpropagation are respectively calculated as 0.0663, 0.0570 and 0.9870 while those for the backpropagation algorithm with momentum are computed as 0.0765, 0.0650 and 0.9821, respectively. As a conclusion, the standard backpropagation provides better approach with slower calculation speed than the backpropagation algorithm with momentum.