Multi-region Modeling of Daily Global Solar Radiation with Artificial Intelligence Ensemble

Abstract

Solar radiation data are crucial for the design and evaluation of solar energy systems, climatic studies, water resources management, estimating crop productivity, etc. As so, for locations where direct measurements are not available, reliable models may be developed to estimate solar radiation from more readily available data. In this study, two artificial intelligence (AI)-based models including artificial neural network and adaptive neuro-fuzzy inference systems, three temperature-based empirical models including Meza–Varas, Hargreaves–Samani, and Chen, and a conventional multi-linear regression (MLR) model were employed for multi-region daily global solar radiation estimation for Iraq. To ensure appropriate selection of input variables, sensitivity analysis was conducted to determine the dominant parameters. Finally, two ensemble approaches, neural average ensemble and simple average ensemble, were applied to improve the performance of the single models. For this purpose, daily meteorological data of maximum temperature $$ \left({T_{\hbox{max} } } \right) $$ , minimum temperature $$ \left({T_{\hbox{min} } } \right) $$ , mean temperature $$ \left({T_{\text{mean}} } \right) $$ , relative humidity $$ \left({R_{\text{H}} } \right) $$ , and wind speed $$ \left({U_{2} } \right) $$ were obtained from January 2006 to December 2016 from four major cities in Iraq representing, north, west, south, and east regions. The results revealed that temperatures $$ \left({T_{\hbox{max} } , \;T_{\text{mean}} ,\; T_{\hbox{min} } } \right) $$ and relative humidity are the dominant parameters. While temperature-based empirical models and MLR model could be employed to achieve reliable results, AI-based models are superior in performance to other models. Also promising improvement in daily global solar radiation modeling could be achieved by model ensemble. The results of this study affirmed that the provided ensemble approaches can increase the performance of single models up to 19.19%, 7.59%, and 16.81% in training, validation, and testing steps, respectively.