Prediction of global solar radiation on tilted surfaces using Artificial Neural networks in Kampala, Uganda.

Abstract

In this study, we developed a model using artificial neural networks to predict average daily global solar radiation on tilted surfaces based on sunshine hours, minimum temperature, maximum temperature and relative humidity as the input parameters. We used the measured data for Kampala, Uganda for five (5) years i.e, from 04th April 2011 to 26 th December 2016 (excluding the year 2013 whose data was very little and could not match the data of other years). The study was conducted at the Department of Physics, Makerere University located at 0.35◦ N and 32.58◦ E for surfaces tilted at 15◦ , 22.5◦ and 30 ◦. Feed Forward-Back Propagation network was used for this study model because it has a high prediction accuracy. The network was trained using the Steepest-Descent training algorithm. We used the trial and error method to arrive at the best model architecture for each angle and twenty eight (28) neurons were appropriate for training of the model. The tan-sigmoid, rectified linear unit and log-sigmoid activation functions were tested in the hidden layer and the output layer utilised the linear transfer function. The tan-sigmoid showed a better performance for all angles of tilt, when predicted and measured values of global solar radiation were compared during the training and testing process. The results from the testing process gave high positive correlation coefficients of 0.95, 0.94 and 0.92 for angle 15◦ , 22.5◦ and 30◦ respectively. The respective mean bias errors were 0.01 M J m−2 day −1 , 0.02 M J m−2 day −1 and 0.02 M J m−2 day −1 , and the respective root mean square error of 1.63 M J m−2 day −1 , 1.79 M J m−2 day −1 and 2.02 M J m−2day −1 were obtained. Comparison was made between the developed ANN model and the hybrid empirical model proposed by Kirya et al. (2016) for prediction of global solar radiation at the same angles of tilt on the same study site. The results from the empirical model proposed by Kirya et al. (2016) gave correlation coefficients of 0.55, 0.53 and 0.52 for angle 15◦ , 22.5◦ and 30◦ respectively. The respective mean bias errors were 2.27 M J m−2 day −1 , 2.57 M J m−2 day −1 and 3.33 M J m−2 day −1 , and the respective root mean square error of 5.89 M J m−2 day −1 , 6.48 M J m−2 day −1 and 7.13 M J m−2 day −1 were obtained. This comparison emphasized the superiority of the developed ANN prediction model over the empirical model.