Comparative Assessment of Temperature Based ANN and Angstrom Type Models for Predicting Global Solar Radiation

Abstract

In this study, temperature based artificial neural network (ANN) models and Angstrom type models for predicting global solar radiation were developed for selected locations in Nigeria.The ANN models were standard multi-layered feed forward, back-propagation neural networks trained with the Levenberg Marquardt algorithm using seventeen years data collected from Nigerian Meteorological Agency (NIMET), Abuja, Nigeria and tested with twenty-two years monthly averaged data downloaded from National Aeronautical Space Administration (NASA) online database. The network inputs were latitude, longitude, elevation, month, maximum ambient temperature (T max ) and minimum ambient temperature (T min ), while monthly average global solar radiation was the network output.The Angstrom type empirical models correlated global solar radiation with minimum and maximum ambient temperatures. The performance of the models were evaluated using statistical performance indicators, namely RMSE, MBE, R 2 and rank score. The coefficients of determination (R 2) of the ANN models were always greater than 99% for all the selected locations while the highest coefficient of determination for the empirical models was 89%. The temperature-based ANN models were thus shown to deliver superior and more reliable outcomes in comparison with the empirical models.