A data-driven approach for window opening predictions in non-air-conditioned buildings

In non-air-conditioned buildings, opening or closing of windows is one of the most common behaviours that occupants tend to carry out to restore their thermal comfort. As an alternative approach to studying the occupant behaviour, particularly when it is difficult to run extensive field studies or due to limits like privacy concerns, this work explores a data-driven method to predict the window openings based on thermal comfort evaluation. The Gradient Boosting Decision Trees (GBDT) algorithm is applied to investigate the importance of selected features, including weather and main building characteristics, to the indoor thermal comfort in non-air-conditioned buildings across whole China. The training set comprises the building simulation results of 95 main cities covering all the five climate regions in China and has 828,360 groups of data in total. The predictor achieves a high accuracy of approximately 95%, and therefore enables the users to estimate the likelihood of window opening based on outdoor weather conditions and local building characteristics. As an original contribution, the study shows that conditioned upon the availability of adequate simulation data, a machine learning predictor trained solely on simulation data can accurately predict realistic window opening behaviours, without relying on any indoor measurement.