Occupancy detection and prediction with sensors and online machine learning : Case study of the Elmia exhibition building in Jönköping

Sammanfattning: Buildings account for approximately 40% of energy consumption in the EU, and despite drastic increasing measures for energy efficient operation, the net value of energy consumption is expected to continue increasing because ofincreasing urbanization.While there has been increasing work done on managing external loads such as adapting to unpredictable climates, the research on internal loading, such as occupants, is still nascent.The specific case study here is on the Elmiaexhibition building, located in Jönköping. Elmia has highlyvariable event attendancesin the range of tens of thousandsthroughout the building occurring over multiple days.Theventilation is pre-programmed to supply air and heat for the maximum capacity at all events suchthat it is likely overventilated. Therefore, there is a need to better understand the real-time occupancy patternduring events to identify how to reduce overventilation. Therefore, this study focuses on one hall, Hall A.Among the variety of sensors out there,CO2sensorsarehighly effective indicatorsof occupancyin buildings and are already commonly used indemand-control ventilation(DCV). At Elmia, the spaces are substantially largerthan typical DCV-applied buildings, so the first part of the study explores how CO2sensormeasurementsvary by location within anexhibition hall.From initial results the CO2sensordata has slight differences based on locationdue to likely different occupant locationsduring events. The degree of difference from one multi-day event between the sensors varied by as high as 359ppm at peak occupancieswithin the hallwith median difference of 70 ppm. With real-time representation of the occupancy via sensors, the second part of the study soughtto establish predictive capabilities from sensor data. A single, average streaming CO2valuefor the hall was used for inputinto a machine learning prediction model, which representedan event with evenly distributed occupancy. For development purposes, data was obtainedfrom both simulation and real sensordata. Using the recently developed River Python Library for online machine learning, the model uses input of thesingle infinite stream of CO2concentrationand forecaststhis one hour ahead. This has useful implications for model-predictive controls because it can be used to adjust various possible parameters of the HVAC one hour ahead of timesuch as duringtransitions between occupied and non-occupied timeson event days, when the CO2differencesare most significant. Future analysis of the data from more events would confirm if multiple CO2datastreams would be needed torepresent occupancy more accurately.