Prediction of solar radiation in China using different adaptive neuro‐fuzzy methods and M5 model tree. (5th May 2016)

• Record Type: Journal Article
• Authors: Wang, Lunche; Kisi, Ozgur; Zounemat‐Kermani, Mohammad; Zhu, Zhongmin; Gong, Wei; Niu, Zigeng; Liu, Hongfu; Liu, Zhengjia
• Is Part Of: International journal of climatology. Volume 37:Number 3(2017)
• Page Start: 1141
• Page End: 1155
• Publication Date: 2016-05-05
• Subjects: global solar radiation -- adaptive neuro‐fuzzy inference system -- grid partitioning -- substructive clustering -- M5Tree -- China; Climatology -- Periodicals; Climat -- Périodiques; Climatologie -- Périodiques; 551.605
• DOI: 10.1002/joc.4762
• Languages: English
• ISSNs: 0899-8418

Abstract

ABSTRACT: Solar radiation is one of the major factors for agricultural, meteorological and ecological applications. In this study, two different optimized adaptive neuro‐fuzzy inference systems (ANFIS), ANFIS with grid partition (ANFIS‐GP) and ANFIS with subtractive clustering (ANFIS‐SC), and M5Tree (M5Tree) methods are proposed for modelling daily global solar radiation (G). Daily meteorological variables at 21 stations in China are used for training and testing the applied models, which is evaluated through root mean square errors (RMSE), mean absolute errors (MAE) and determination coefficient ( R 2 ). Above models will be compared with a calibrated empirical Ångström model and the results indicate that the ANFIS models provide better accuracy than the M5Tree and empirical method, for example, the RMSE values for ANFIS‐SC, ANFIS‐GP, M5Tree and the Ångström model range 2.10–3.08, 2.07–3.08, 2.79–3.87 and 2.54–3.69 MJ m −2 day −1, respectively. The model performances also show some differences at different stations for each model, for example, the ANFIS models produce the most accurate estimations at station 58238, while M5Tree brings the best accuracy at the station 51777. Meanwhile, the models underestimate high radiation values for some stations, which may due to the differences in training and testing data ranges and distribution of the stations. Finally, the reasons for the differences in model performance are investigated in detail.