Solar Energy Potential Forecasting and Optimization Using Artificial Neural Network: South Africa Case Study

Abstract

Explored in this work is the comparison of different artificial neural network (ANN) training functions for modeling the daily average solar irradiance (DASI) forecaster. The aim of this work is to establish the comparison studies with reference to the performances of four training functions namely, the Resilient Backpropagation (TRAIN-RP), Levenberg-Marquardt (TRAIN-LM), Scaled Conjugate Gradient (TRAIN-SCG), and Quasi-Newton (TRAIN-BFGS) implemented in the training of the ANN for forecasting the DASI. Using the Levenberg-Marquardt's learning rule the feed-forward network with back propagation architecture is implemented in the approach of the neural network design. The comparison is done by looking at the Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of correlation (R) of the models. The TRAIN-LM performed better than the other models and achieved a 99.742% coefficient of correlation with a mean absolute percentage error (MAPE) of 2.394%. The training, testing, and validation of the neural network is done by utilizing the actual data from six different locations in the North West cities, and these data are mined from the NASA database. The findings in this paper will be useful in the optimizing the micro-grid and its performance.