Quantification of occupant response to influencing factors of window adjustment behavior using explainable AI.

Window adjustment is one of the most common ways employed by occupants to control the indoor environment, and window adjustment behavior (WAB) is known to be a crucial factor for predicting building energy consumption. In recent decades, many studies have attempted to develop a reliable WAB model with the following three issues. First, most current studies have treated WAB models in an "average occupant" fashion. This average WAB model is prone to ignoring the variability of individual preferences. The second issue is the analysis of the impact of these factors, which still considers how occupants respond to environmental and non-environmental factors. Finally, regarding the prediction models, there are diverse modeling approaches such as probabilistic and machine learning approaches. Each modeling method has advantages and disadvantages in terms of adaptability, complexity, and explainability. This study focuses on explainable artificial intelligence (XAI) to presents quantitative analysis on the aforementioned three issues. For this purpose, occupant data (occupant presence and window state) and indoor/outdoor environmental data (temperature, humidity, illuminance, CO 2 concentration, PM 2.5 concentration, and time of day) were collected from 12 households for one year using multiple sensors installed in each household. Three XAI approaches, namely, logistic regression, XGBoost, and Shapley additive explanations (SHAP), were introduced to analyze the interaction between environmental and non-environmental factors and occupant behavioral patterns in terms of feature influence. The results showed that (1) occupant personal preferences on WAB significantly vary from household to household; (2) occupant personal preferences on WAB cannot be defined only with a single environmental parameter; (3) there are unmeasurable unknown factors such as psychological and social factors; and (4) similar to logistic regression models, the current complex black-box models can also be described by applying XAI techniques in terms of feature influence. [ABSTRACT FROM AUTHOR]