The development of occupancy monitoring for removing uncertainty within building energy management systems

Abstract

This paper provides an overview of methods and models used for the localised detection of building occupants by combining data from a range of sensor types, and the prediction of future occupancy rates based on past data. The occupancy detection proposed here is designed to be implemented as part of a real-time responsive building energy management system, catering building energy use directly to occupant needs. The initial stages of testing used sensor data collected in a small office building in Nottingham, UK. A Neural Network model was trained using data from local environmental sensors, including CO2 level, motion detection, temperature, window status and the detection of personal mobile devices through Wi-Fi and Bluetooth connections. A predictive neural network model was also trained using simulated occupancy rates, with consideration for the level of uncertainty in the model outputs. The results of the study show that the combination of a select group of sensors can provide a lower error in the estimated number of occupants per zone than any single sensor type alone. However, when limited training data is available, it is not viable to include all sensors in the model, as this leads to overfitting. The sensors that give the greatest information gain were found to be best identified by comparing the occupancy estimation made by each sensor individually. It was found that the predictive model outperformed simpler occupancy prediction heuristics, especially when occupant behaviours differ from typical patterns.