A neuro-fuzzy model of the solar diffuse radiation with relevance vector machine

Solar radiation data are essential for most solar energy research and applications. However, the only measurements of solar radiation for which long-term records are available from a large number of locations are the measurements of the global solar radiation. Thus the diffuse solar radiation and the direct solar radiation must be estimated by various means from the global solar radiation. Although a number of diffuse radiation models are available nowadays, they have many drawbacks such as their dependence on geographic location, the type of data, weather conditions etc. The paper presents a different approach to modeling the diffuse solar radiation from the approaches presented in literature which is based both on fuzzy logic and artificial neural network techniques. The adopted neuro-fuzzy model employs a fuzzy inference system with the same structure as that of a Takagi-Sugeno fuzzy model but with a neural network learning mechanism called relevance vector machine. The number of fuzzy rules and parameter values of membership functions of the model are automatically generated through the relevance vector machine. The performance of the model is tested against independent measurement data and is compared to the performance of other models reported in literature. The obtained results show great effectiveness of the adopted neuro-fuzzy model, its main features being the small model dimension (fewer fuzzy rules) and a very good generalization.