A comparative study of machine learning methods for identifying the 15 CIE standard skies

For energy-efficient building designs, the solar irradiance and daylight illuminance derived from the CIE standard skies are useful. Over time, the sky luminance distributions have been used to identify these standard skies, but these are sparingly measured. Thus, the use of available climatic variables has become a viable alternative. Nevertheless, it is necessary to determine if these climatic variables could correctly identify these skies. This study addresses the lack of luminance distribution measurement by classifying the standard skies using measured climatic data in Hong Kong. The classification approach was improved by using the machine learning (ML) method. For comparative analysis, five popular ML classification algorithms i.e., decision tree (DT), k-nearest neigbhour (KNN), light gradient boosting machine (LGBM), random forest (RF) and support vector machines (SVM) were used. The findings show that accuracies of 68.1, 73.1, 74.3, 74.5, and 75.4% were obtained for the DT, KNN, SVM, LGBM, and RF models, respectively. Similarly, the F1 scores were 66.6, 70.2, 71.8, 72.1 and 72.9%, for the DT, KNN, SVM, LGBM, and RF models. The result shows that the RF model gave the best performance while DT performed the least. Also, the obtained accuracies and F1 scores show that all models would classify the standard skies with reasonable accuracy. Furthermore, feature importance was done, and it was found that K d , T v , K t , α, sun, and cld are the most important input parameters for sky classification. Lastly, vertical solar irradiance ( G VT ) and illuminance ( G VL ) were estimated using the skies predicted by the proposed models. Upon predictions, it was observed that the G VT ranged from 14.7 to 24.6% while the G VL from 13.8 to 19.9%. Generally, most of the predictions were less than 20%, which shows good predictions were obtained from the models.