Building Occupancy Forecasting Using Transformers

Forecasting occupancy status is crucial for various real-time applications such as crowd management, evacuation, energy management, and others. Using machine learning can provide occupancy information and forecasting with high accuracy. Having abundant and convenient access to sensors and environmental data has allowed research in occupancy prediction to be done in less invasive ways. Deep learning models such as transformers can be used to predict future time steps of occupancy values using data from previous time steps. Environmental sensors such as CO2 and temperature, commonly found in smart homes, are a great way to be able to gather the data needed for occupancy prediction models. They are also much less invasive and intrusive as opposed to cameras and microphones. In this work, we present a transformer-based, machine-learning model, for future time-step prediction using the previous occupancy count. This work expands on our previous work where we estimated the current occupancy count based on the sensor data at that time step. Using the estimated occupancy count till time t, we were able to estimate the occupancy at t+1 with a high accuracy of about 97%.