Reliability Evaluation of GSR Prediction Using Neural Networks with Variant Atmospheric Parameters

Global Solar Radiation (GSR) has fluctuations in its measured values. This occurs by actions of several factors including clouds, dust, reflections, and others. The ambiguity associated with its prospective values forms a challenge for many engineering applications and manufacturers of solar-based systems. The intermittent nature of global solar radiation conflicts with the necessity to find correct and reliable values in advance. The neural network-based prediction has been adopted to fulfill a prior knowledge about these values for being highly efficient compared to the stochastic and statistic approaches. Despite that, the reliability of those networks is considered variant for being largely dependent on different inputs. This work evaluates the reliability of different neural networks that specifically use atmospheric parameters, considering them as single inputs and combinations of two and three parameters. The results appeared that the network that uses (Zenith Angle, Air Temperature, and Relative Humidity) is the most reliable one with 0.997 recorded for the correlation coefficient. Oppositely, the network of only (Air Temperature) is the network of the lowest reliability according to the 0.603 that is found for the correlation coefficient.