Predicting open-plan office window operating behavior using the random forest algorithm

Understanding window operating behavior in offices is important in terms of its influence on reducing energy consumption and improving indoor comfort. Researchers have applied different mathematical methods to develop useful window operating behavior models, however, the applicable machine learning algorithms are still in their preliminary research stage, requiring additional development. In the work described here, the authors applied the random forest (RF) algorithm to predict window operating behavior in open-plan offices, using data from three such offices in Nanjing, Jiangsu Province, China. The three open-plan offices were different in terms of their areas, office types, numbers of occupants, and layouts. The importance of various elements influencing window operating behavior was determined through the RF method, and the resulting rankings were consistent with the occupants' subjective understanding, as determined using a questionnaire. The sensitivity of the RF model to the number of inputs was explored, and showed that, with four inputs, its accuracy could reach 80%. The RF model also showed high accuracy and stability in predicting window operating behavior in two application formats-namely, for different offices, and for the same office over different years. Meanwhile, the RF model was compared with the other two popular machine learning methods, namely SVM and XGBoost algorithms, which also proves the high accuracy of RF models. The results obtained in this study should provide insights into applying machine learning methods to window operating behavioral studies generally-and may also inspire their application to other behavioral study types.

Automated summary

Understanding window operating behavior in offices is essential for energy saving and indoor comfort improvement. Machine learning algorithms, particularly the random forest model, showed high accuracy and stability in predicting window operating behavior. The sensitivity of the random forest model to the number of inputs was explored, and it achieved 80% accuracy with four inputs.