Prediction of window opening behavior and its impact on HVAC energy consumption at a residential dormitory using Deep Neural Network

Individuals globally spend about 88-92% of their entire time in the indoor environment. The implementation of regularly scheduled systems operation is common in many commercial and residential building types. Occupant Behavior (OB) is highly stochastic, making it difficult to depict the human factor using simple schedules. In the present work, window-opening tendencies were found to be highest in summer and transition seasons and lowest during winter. The Air Changes per Hour (ACH) value obtained from blower door testing was compared with the average ACH value obtained from the mass balance equation. These ACH values were imported to the EnergyPlus model. A three-hidden layered deep neural network was developed to predict the window's state at 16 residential dorms. The data was collected from August 2021 to May 2022 to train and test the DNN models. 48 windows' operations were measured during this timeframe and 48 DNN models were created with an average accuracy of 96.71% and an average True Positive Rate (TPR) of 0.9136. Bedroom windows were found to be more time-dependent than living room windows. Incorporating time components into the bedroom window opening models can increase the TPR by 0 to 6%. The models were then integrated into the Energy Plus simulation engine using the newly introduced Python API to simulate the total HVAC energy consumption during the occupied heating season of 2021-2022. The DNN coupled model predicts 16,350 kWh energy is lost through windows from eight homes in an occupied heating season.