Neural network based method for conversion of solar radiation data

• Published in: Energy conversion and management Vol. 67; pp. 117 - 124
• Main Authors: Celik, Ali N., Muneer, Tariq
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier Ltd 01.03.2013; Elsevier
• Subjects: Applied sciences; Conversion; Energy; Exact sciences and technology; Horizontal; Irradiation; Isotropic sky model; Mathematical analysis; Muneer anisotropic model; Natural energy; Neural networks; Optimization; Regression; Solar energy; Solar energy conversion; Solar radiation; Isotropic sky model; Muneer anisotropic model; Solar radiation; Neural networks; Conversion; Models; Neural network; Modeling
• ISSN: 0196-8904; 1879-2227
• DOI: 10.1016/j.enconman.2012.11.010

► Generalized regression neural network is used to predict the solar radiation on tilted surfaces. ► The above network, amongst many such as multilayer perceptron, is the most successful one. ► The present neural network returns a relative mean absolute error value of 9.1%. ► The present model leads to a mean absolute error value of estimate of 14.9Wh/m2. The receiving ends of the solar energy conversion systems that generate heat or electricity from radiation is usually tilted at an optimum angle to increase the solar incident on the surface. Solar irradiation data measured on horizontal surfaces is readily available for many locations where such solar energy conversion systems are installed. Various equations have been developed to convert solar irradiation data measured on horizontal surface to that on tilted one. These equations constitute the conventional approach. In this article, an alternative approach, generalized regression type of neural network, is used to predict the solar irradiation on tilted surfaces, using the minimum number of variables involved in the physical process, namely the global solar irradiation on horizontal surface, declination and hour angles. Artificial neural networks have been successfully used in recent years for optimization, prediction and modeling in energy systems as alternative to conventional modeling approaches. To show the merit of the presently developed neural network, the solar irradiation data predicted from the novel model was compared to that from the conventional approach (isotropic and anisotropic models), with strict reference to the irradiation data measured in the same location. The present neural network model was found to provide closer solar irradiation values to the measured than the conventional approach, with a mean absolute error value of 14.9Wh/m2. The other statistical values of coefficient of determination and relative mean absolute error also indicate the advantage of the neural network approach over the conventional one. In terms of the coefficient of determination, the neural network model results in a value of 0.987 whereas the isotropic and anisotropic approaches result in values of 0.959 and 0.966, respectively. On the other hand, the isotropic and anisotropic approaches give relative mean absolute error values of 11.4% and 11.5%, respectively, while that of the neural network model is 9.1%.