Using Deep Neural Network to Predict Natural Ventilation Behavior in Student Dorms

People spend about 87 - 92% of their overall time of day in some form of indoor environment. It is also well established that buildings consume about 40% of the total electricity generated. Studies have shown that operating windows during HVAC use can increase the heating and cooling load. Until now, fixed window opening and closing schedules have been used in Building Energy Simulation engines (BES) like Energy Plus. We studied the window opening behavior of students living in a residential dorm at Syracuse University for over a year. The window opening behavior changed depending on the location of the window and the time of the year. During summer, the bedroom windows were opened substantially for a longer time and living room windows were opened briefly. During transition seasons, all windows were opened very briefly. We created a specific set of neural network models that could predict the state of the window and the duration of the window opening. We modeled the open/close state change for all three windows in three different zones using Artificial Neural Network (ANN) machine learning architecture. The three model, when trained on individual datasets generated a True Positive Rate (TPR) of 0.97, 0.99, and 0.97. The newly released Python API incorporated the models into Energy Plus. The sensible heating/cooling energy for the living room, north-facing bedroom, and south-facing bedroom were calculated based on the window schedule predicted by the ANN model. Compared to the baseline model of complete mechanical ventilation, the heating energy was reduced by 3.96% and 24.78% for north and south-facing bedrooms, respectively, for October.