The Room Is Empty. Why Is the HVAC Still Running?
Research preparation and architectural analysis: the ArchUp research team.
Imagine a meeting room after the last participant leaves. Its chairs are empty, but air continues to move through the ceiling. This is an illustrative scenario, not a documented site observation. It opens a question that a photograph of an apparently smart building cannot answer: does the building know that the room is empty, and can it use that knowledge well?
Room occupancy recognition promises to connect building operation with actual use. Yet recognizing a person, predicting an arrival and reducing energy consumption are different tasks. A sensor can succeed at the first while the building fails at the last. The architectural question is therefore broader than which device to install. It concerns the relationship between space, service zones, operating rules and the people who must live with an automated decision.
Reading scope and evidence
This article develops the supplied literature review into an architectural analysis. It draws on selected references from that review, with targeted checks of publisher and research institution records for references [1], [4], [5], [6] and [7], including the accessible abstract and reported limitations of [5]. It is not a systematic review or an independent reading of every cited paper in full. Other findings remain attributed to the supplied secondary review. The proposals below are editorial deductions, not findings from an ArchUp field experiment.
The original brief also lists conference prototypes from 2025. Their headline savings are not reproduced here because short trials, differing baselines and incomplete independent verification do not provide a reliable basis for ranking products. The bibliography contains the studies actually used in this argument, rather than every item in the supplied list.
Presence, numbers and prediction are different design briefs
Occupancy can mean a binary distinction between presence and absence, an estimate of the number of people, or a prediction of future use. Reviews distinguish several sensing and modelling approaches rather than one universal solution [1,6]. These distinctions matter before the specification reaches a supplier. A room whose lighting only needs a reliable presence signal presents a different problem from a ventilation zone where population varies substantially.
More detailed information is not automatically more useful. Knowing that five people occupy a room has little operational value if the associated system can only switch an entire floor between two modes. Conversely, an otherwise capable control system may be constrained by a signal that cannot distinguish one occupant from a crowded meeting. The design brief should name the decision the information will support, the time available to make it and the consequence of getting it wrong.
Context adds another layer. A booking, the time of day and an occupied neighbouring zone may help anticipate demand, but none proves that someone has entered the room. The supplied review describes context dependent prediction in Adamopoulou and colleagues’ work [2]. Its architectural implication is a need to distinguish expectations from observations. An unused reservation should not indefinitely preserve an occupied setting, while an unbooked visitor should still receive an appropriate environment.
The room boundary must meet the service boundary
A floor plan often represents rooms more precisely than the mechanical system can control them. Partitions may divide a shared service zone, or a sensor may observe a circulation route while the controller serves several enclosed rooms. Under those conditions, improved classification alone cannot resolve the mismatch. The project needs a drawing that connects each observation area to the equipment and operating decision it can actually influence.
Consider two adjacent meeting rooms served together. If one empties and the other remains occupied, an instruction based only on the first room can inconvenience the second. This example is a design hypothesis, not a report of measured performance. It shows why the definition of a control zone belongs in architectural coordination. Furniture, doors, partitions and service distribution help determine whether occupancy information describes a meaningful unit of action.
Changes after handover also matter. A meeting room becomes a shared office; a corridor gains a waiting area; furniture blocks a sensor’s view. A model trained before these changes may continue producing plausible results while describing the wrong use pattern. Recommissioning should therefore follow consequential spatial changes, not only equipment replacement.
A convincing recognition score is not a measured energy saving
Mena-Martinez and colleagues studied environmental sensing in an office of approximately 230 square metres, using five devices over 99 days [5]. The publisher reports feature fusion and semi-supervised learning using temperature, humidity and pressure. Selected feature fusion achieved an average accuracy of 0.99 and F1 of 0.95; a semi-supervised self-training model reported 0.97 and 0.90 respectively. These are classification results, not a measured reduction in building energy use.
The paper’s accessible limitations also identify missing records for one occupancy class and a restricted ground truth collection period [5]. This makes the study valuable as a demonstration of reduced labelling effort, while limiting any claim that it proves universal performance. A successful method in one office should become a candidate for local testing, not an automatic procurement guarantee.
The distinction creates two acceptance tests. The first asks whether the system recognizes the relevant occupancy states, including a person sitting still. The second asks whether the resulting operating decisions reduce measured consumption while maintaining acceptable conditions. Reporting only the first leaves the central building performance claim unanswered.
The baseline determines what the saving means
A savings percentage requires a reference condition. Comparing a new controller with continuous operation produces a different question from comparing it with a well maintained schedule. The supplied review of Nikdel and colleagues illustrates how the chosen comparison changes the apparent value of occupancy control [3]. A published percentage cannot responsibly become a promise for another project without understanding that comparison.
Pang and colleagues investigated office prototypes through EnergyPlus simulations across different climates, code versions and sensing scenarios [4]. The institutional abstract reports that counting sensors provided only marginal additional benefit over presence sensing in the simulated scenarios, and that heating dominated climates generally offered stronger savings. This is simulation evidence under specified assumptions, not a universal verdict against counting sensors or a measured guarantee for a cooling dominated building.
A project evaluation should record the reference schedule, weather conditions, occupancy, service levels and relevant equipment changes. If a quieter month coincides with commissioning, the reduced bill cannot simply be assigned to the new sensor. If only lighting is controlled, whole building energy should not be presented as though every load were responsive. The measurement boundary must match the claim.
Being still must not mean being absent
A technically convenient error can become a poor experience. A reader, a person concentrating at a desk or someone resting may remain still for a long period. The supplied reviews identify limitations in individual sensing methods and discuss combining signals [1,6]. The appropriate architectural response is to test the actual activities the room supports, rather than assuming that constant movement represents normal use.
Errors are also unequal. A false occupied signal may extend operation unnecessarily. A false vacant signal may withdraw a service while a person is present. Their costs should be assessed separately. A high overall score can conceal poor recognition of rare but consequential conditions, especially where empty periods dominate the dataset.
Manual override, understandable feedback and a suitable fallback mode should therefore be part of the brief. These are proposed requirements, not a claim that one universal setting suits all systems. The appropriate delays and operating limits need to be developed with the responsible engineers and tested in the actual building.
Comfort and air quality belong in the same assessment
Zhang, Wu and Calautit reviewed 160 studies on occupancy prediction and its relationship to energy, air quality and thermal comfort [7]. Their framing supports a broader evaluation than prediction accuracy alone. An efficient classification pipeline does not by itself establish that occupants experience an acceptable environment.
Vacancy should not be treated as a universal instruction to stop every service. Temperature recovery, humidity, residual pollutants and the needs of equipment can continue after people leave. Those considerations require project specific operating sequences. An energy target should be evaluated alongside the environmental conditions and occupancy periods it affects.
For an architect, this changes the conversation at handover. Instead of asking whether the automation is installed, ask whether the rooms still support their intended activities when it acts. Complaints and overrides are useful evidence of interaction with the system, not merely behaviour to be suppressed.
Data science and IoT contribute at different points
The data science contribution concerns inference under imperfect evidence: missing labels, changing occupancy patterns and several sensor streams that disagree. The IoT contribution concerns reliable acquisition and delivery of those observations, plus the connection to a controllable service. Neither is sufficient on its own. A sophisticated model receiving stale data can be less useful than a modest model supported by dependable sensing.
The supplied review describes federated learning research by Khan and colleagues, in which occupancy prediction is distributed across devices [8]. Keeping raw information closer to its origin may support a privacy strategy, but the architecture of the model does not alone settle retention, permissions or security. A brief still needs to state what information is collected, who can access it and why it is necessary.
This suggests a practical limit on data collection: require each additional variable to justify the operational decision it enables. Identity recognition, for example, should not become a default requirement where anonymous presence would serve the design objective. That is an editorial principle of proportionality, not a performance result attributed to the reviewed experiments.
A research agenda that begins with the building
The most useful next experiment would compare sensing arrangements within a clearly defined service zone. It would test occupied and vacant periods, different activities, faults and changes in layout. The evaluation would record energy, environmental conditions, overrides, maintenance effort and the cost of establishing reliable occupancy labels. Where prediction is introduced, it would be compared with a simpler responsive strategy under equivalent conditions.
Such a study would help answer a question that algorithm rankings often leave open: how much is another layer of information worth here? Extra sensors bring purchase, installation, maintenance and replacement obligations. Their value should be demonstrated through a better decision, not inferred from a longer equipment list. A small improvement in recognition may matter greatly in one setting and add little in another.
The research gaps are consequently connected. Transfer between buildings depends on their physical and operational differences. Sensor placement depends on activities and visibility. Labelling cost depends on the practical ability to establish who was present. Comfort depends on the service response. These are opportunities for collaboration between architecture, building services and data science.
Architecture after the occupancy signal
The empty room is a useful beginning, but it is not a complete definition of efficiency. A building saves energy responsibly when its knowledge of use can become a suitable, verifiable operating decision. That requires aligned spatial and service boundaries, a credible baseline and a way to recognize when automation is wrong.
The architectural contribution is therefore to make the relationship between people, rooms and services explicit. Occupancy sensing becomes worthwhile when it helps the building serve actual use with less unnecessary consumption, while preserving the conditions people need. The decisive question is not how intelligently a room can count its occupants. It is whether the building can act on that knowledge well.
Sources
- Rueda L, Agbossou K, Cardenas A, Henao N, Kelouwani S. A comprehensive review of approaches to building occupancy detection. Building and Environment. 2020;180:106966. DOI: 10.1016/j.buildenv.2020.106966
- Adamopoulou AA, Tryferidis AM, Tzovaras DK. A context-aware method for building occupancy prediction. Energy and Buildings. 2016;110:229–244. DOI: 10.1016/j.enbuild.2015.10.003
- Nikdel L, Janoyan K, Bird SD, Powers SE. Multiple perspectives of the value of occupancy-based HVAC control systems. Building and Environment. 2018;129:15–25. DOI: 10.1016/j.buildenv.2017.11.039
- Pang Z, Chen Y, Zhang J, O’Neill Z, Cheng H, Dong B. Nationwide HVAC energy-saving potential quantification for office buildings with occupant-centric controls in various climates. Applied Energy. 2020;279:115727. DOI: 10.1016/j.apenergy.2020.115727
- Mena-Martinez A, Alvarado-Uribe J, Molino-Minero-Re E, Ceballos HG. Indoor occupancy monitoring using environmental feature fusion and semi-supervised machine learning models. Journal of Building Performance Simulation. 2024;17(6):695–717. DOI: 10.1080/19401493.2024.2399053
- Chen Z, Jiang C, Xie L. Building occupancy estimation and detection: A review. Energy and Buildings. 2018;169:260–270. DOI: 10.1016/j.enbuild.2018.03.084
- Zhang W, Wu Y, Calautit JK. A review on occupancy prediction through machine learning for enhancing energy efficiency, air quality and thermal comfort in the built environment. Renewable and Sustainable Energy Reviews. 2022;167:112704. DOI: 10.1016/j.rser.2022.112704
- Khan I, Delicato FC, Greco E, Guarascio M, Guerrieri A, Spezzano G. Occupancy Prediction in Multi-Occupant IoT Environments Leveraging Federated Learning. IEEE DASC/PiCom/CBDCom/CyberSciTech. 2023:0036–0043. DOI: 10.1109/dasc/picom/cbdcom/cy59711.2023.10361520
