Input data selection for solar radiation estimation

Abstract

AbstractModel input data selection is a complicated process, especially for non‐linear dynamic systems. The questions on which inputs should be used and how long the training data should be for model development have been hard to solve in practice. Despite the importance of this subject, there have been insufficient reports in the published literature about inter‐comparison between different model input data selection techniques. In this study, several methods (i.e. the Gamma test, entropy theory, AIC (Akaike's information criterion)/BIC (Bayesian information criterion) have been explored with the aid of non‐linear models of LLR (local linear regression) and ANN (artificial neural networks). The methodology is tested in estimation of solar radiation in the Brue Catchment of England. It has been found that the conventional model selection tools such as AIC/BIC failed to demonstrate their functionality. Although the entropy theory is quite powerful and efficient to compute, it failed to pick up the best input combinations. On the other hand, it is very encouraging to find that the new Gamma test was able to choose the best input selection. However, it is surprising to note that the Gamma test significantly underestimated the required training data while the entropy theory did a better job in this aspect. This is the first study to compare the performance of those techniques for model input selections and still there are many unsolved puzzles. Copyright © 2009 John Wiley & Sons, Ltd.