Content area
Abstract
As building systems such as heating, cooling, ventilation, and lighting continue to reduce their energy consumption, the energy use of miscellaneous plug loads becomes a growing concern. Though the efficiency of these devices should be addressed, it is also important to turn them off or place them into low- or no- power modes when no service is being provided by their operation. Current methods of determining vacancy require the time-consuming task of gathering ground truth in order to train an inference model. This study develops, tests, and evaluates a method of inferring vacancy in buildings that uses easily obtainable data during model training. The approach infers vacancy from any numerical building data having a suitable correlation with vacancy patterns using their cumulative distributions during times of expected vacancy. These times are easily extracted from general knowledge of building vacancy patterns. Decision-level sensor fusion allows the usage of one or many input data streams, where the use of multiple inputs can improve the quality of the vacancy inference. The proposed method was piloted at an office space in Davis, CA USA using the following data streams: electricity demand, room carbon dioxide levels, and the number of active connections to the office’s Wi-Fi network. Evaluation is performed by comparing model outputs against ground truth obtained from security camera footage. ROC analysis is performed, and a new comparison is proposed that compares false positive and false negative rates in an ROC-like manner, called the CMC curve. The proposed method of generating inference curves using root mean square for fusion shows an area under the ROC curve and an area under the CMC curve of 0.960 and 0.041 (where 1 and 0 are the best possible), respectively. This moderately outperforms logistic regression using any of the applied fusion methods, for which the best area under the ROC curve and area under the CMC curve are 0.955 and 0.045, respectively. This shows that using the proposed method allows for quality vacancy inference while reducing the up-front data requirement inherent to model training.