Designing Smart Buildings with Humans in Mind: A conversation on machine learning, personal comfort and intelligent control in the built environment

Interview with Dr. Zeynep Duygu Tekler by Lada Hensen Centnerová

Dr. Zeynep Duygu Tekler is a Senior Research Associate in the Department of Engineering Science at the University of Oxford. Her research focuses on occupant-centric smart building technologies and machine learning approaches to improve energy efficiency and occupant well-being in the built environment. Prior to Oxford, she was a Postdoctoral Research Fellow at the National University of Singapore and completed her PhD at the Singapore University of Technology and Design (SUTD) under the SUTD-MIT International Design Centre in 2021. 

LHC: You are a co-author of the paper titled “Hybrid system controls of natural ventilation and HVAC in mixed-mode buildings: A comprehensive review.” [1]  Could you clarify what do you mean by a ‘mixed-mode building’?

ZDT: A mixed-mode building combines both natural ventilation (such as operable windows, vents or other passive airflow strategies) and mechanical HVAC systems. The idea is to take advantage of natural ventilation when conditions are suitable, and switch to mechanical systems when needed. In this paper [1], we conducted a comprehensive review of control strategies in such buildings, focusing on how to effectively manage the interaction between the two systems to improve energy efficiency while maintaining occupant comfort. My previous team at the National University of Singapore, led by Prof. Adrian Chong, is pioneering research in this area.

LHC: I found a very interesting figure in that paper. Figure 1. presents a keyword analysis of the titles and abstracts of the reviewed publications, showing the averaged publication year per keyword. It looks like hybrid ventilation started with night ventilation.

ZDT: Yes, that’s correct. If we look at this Figure 1 closely, we can observe a clear evolution in research focus over time. Initially, around the early 2000s, studies on hybrid ventilation often centred around night ventilation, especially in connection with natural ventilation and energy saving strategies. After 2015, and increasingly after 2020, there is a shift towards adaptive control. The growing number of papers suggest that AI is gradually being introduced into our lives and implemented in building control systems. This shift reflects a rising interest in smarter and more responsive technologies for improving energy efficiency and occupant comfort.

Figure 1. Keyword analysis of the reviewed publications’ title and abstract.

LHC: The mentioned paper provides an overview of 5 control strategies. It starts with On-Off and PID control and ends with Fuzzy logic and data-driven control. Why do you need more complex controls than PID controllers?

ZDT: I think the complexity of control strategies depends on the building’s capabilities and requirements. Nowadays, in many existing buildings, the control systems still rely on very simple, rule-based approaches — such as fixed temperature setpoints. But as buildings become smarter and more integrated with IoT technologies, there is a greater potential to optimize the performance by learning from and responding to occupant behavior. So, the simple answer is that AI enables control systems to become more intelligent — evolving from basic On-Off and PID controllers to more adaptive and sophisticated systems which respond dynamically to how occupants use the space.

LHC: I would like to go back to mixed-mode buildings with hybrid ventilation. Such buildings with good control can save significant amount of energy for conditioning spaces and at the same time guarantee good thermal conditions, but what about noise and outdoor air pollution?

ZDT: They are equally important. There are many dimensions of comfort and thermal comfort is just one aspect. Among the 74 papers we reviewed, only one addressed control strategies that also considered outdoor noise and air pollution. That said, with the advancement of smart technologies, it is getting increasingly feasible to integrate noise and air quality monitoring into HVAC control systems using real-time data. Ideally, noise levels and outdoor air pollution should be continuously monitored and factored into the building’s control strategy to ensure a holistic approach to occupant comfort.

LHC: You mentioned real-time monitoring and optimization. Let’s speak about your other papers “Data-efficient comfort modeling: Active transfer learning for predicting personal thermal comfort using limited data” [2] and “Enhancing personalised thermal comfort models with Active Learning for improved HVAC controls”[3]. Can you explain what you are proposing in these papers?

ZDT: In these papers, we focused on two main goals: one is improving building energy efficiently, and the other one is improving comfort. Traditionally, thermal comfort research has concentrated on group level comfort, but there is a growing shift toward personalized comfort –models that can predict individual level comfort responses.

That is what we explored – developing models that can predict how hot, cold or comfortable occupants feel under changing room conditions on an individual level. One promising approach in this context is thermal preference-based HVAC control, which adjusts the heating and cooling setpoints based on each occupant’s thermal comfort preferences to enhance both energy efficiency and occupant comfort.

The biggest challenge with personal comfort models is that they typically require a lot of user feedback, which means frequent surveys. This process can be very labor intensive, time-consuming and expensive. To address this, we leveraged active learning, a machine learning technique that minimizes the amount of data needed by identifying the most informative data points for training, rather than using the entire dataset. In [2], we applied active transfer learning to develop data-efficient thermal comfort models. In [3], we extended this work by integrating these models into HVAC control systems in a more cost-effective and scalable way.

LHC: Well, I think another problem is that people don’t want to provide answers all the time.

ZDT: Yes, exactly, they don’t want to be constantly bothered. And that is what makes this work particularly valuable. We’re exploring how algorithmic approaches, including AI and its subbranches, can help reduce the amount of data needed to develop data-driven comfort models. The goal is to accurately predict individual level comfort without requiring extensive user input.

LHC: How far can you go to avoid collecting lots of users’ feedback data?

ZDT: That really depends on the complexity of the problem and the approach used. When we developed models to predict personal thermal preferences, we found that the need for users’ feedback could be reduced by up to 46% using active learning [4]. In our follow-up work [2], we combined transfer learning with active learning, and found that less than 10% of the user feedback was sufficient to achieve models as accurate as traditional ones that require 100% of the feedback. This is a significant reduction, making data-driven comfort models more scalable and practical for real-world applications.

LHC: This sounds very promising but still, first you need to collect feedback data. Can you use this approach in a new building?

ZDT: Yes, this can be done using a specific area in machine learning called transfer learning. The idea is to take models trained on data-rich buildings where we have lot of information and transfer that knowledge to a new building with limited data. In the paper I mentioned earlier [2], we explored how transfer learning can be combined with active learning through multiple experiments. Essentially, we start with a pre-trained model and fine-tune it using small amount of new data from the target building. This approach significantly reduces the amount of training data and time required, making it much more practical for real-world deployment in new buildings.

LHC: You did this study in Singapore. Could it be also used in Europe mild climate?

ZDT: Definitely. I see great potential for applying this approach in climates like the UK’s, or other mild European climates. In fact, one of our planned future works is to apply the framework across buildings in various climate zones. This will help us further test and ensure the generalizability of the findings beyond the context of a tropical climate like Singapore.

LHC: Is there something else from your research, we should talk about in this interview?

ZDT: My expertise lies in applying machine learning in the built environment, with a strong focus on occupant behavior. In addition to personal thermal comfort, I am very much interested in occupancy —specifically, how people move through and use spaces over time, and its impact on energy use in the buildings. In my research at the National University of Singapore, I focused mostly on institutional buildings, while at the University of Oxford my work centred on domestic buildings.

What I find especially exciting is the challenge of integrating multiple aspects of occupant behavior, such as diverse comfort preferences and movement patterns, into responsive and adaptive building systems. “How can we design buildings and buildings systems that are not only energy-efficient, but also truly human-centric—able to anticipate and respond to the diverse needs of their users in real-time?” I believe this is one of the most important and complex challenges in creating the next generation of smart and sustainable buildings.

LHC: What is your definition of smart or intelligent building? How do you explain it to your students?

ZDT It depends on whether the students have any background in AI or machine learning. If they are completely new to the topic, I start by explaining that a smart or intelligent building is one that can learn and make decisions, and it’s responsive to its environment using technology.

I often describe it a building that can sense what’s happening both inside and outside of the building through various sensors – taking measurements like temperature, occupancy, light levels and air quality. It can collect and exchange data via the Internet, which is where the Internet of Things (IoT) comes into play. Based on this data, it can also make automated decisions and that’s where AI becomes crucial.

For example, a smart building can automatically adjust HVAC systems, lighting or plug loads based on occupancy or time of the day – something I explored in my PhD research. So in short, a a smart building is one that uses data and intelligent systems to operate more efficiently and improve occupant comfort.

LHC: The last question. The use of AI and IoT for monitoring and optimizing buildings’ energy performance consume energy, do you think that achieved energy savings are higher than energy used for AI and IoT?

ZDT: That's a great question and an important one. Yes, AI and IoT systems do consume some energy for data processing, communication, and sensor operations. However, in most cases, the energy savings they enable far outweigh the energy they consume.

That said, it's important to design these systems efficiently, choosing lowpower sensors, optimizing data transmission, and leveraging edge computing where possible to reduce energy overhead. So, when implemented thoughtfully, the net energy impact is positive, contributing to both operational savings and sustainability goals.

Sources

[1]     Peng, Y., Lei, Y., Tekler, Z. D., Antanuri, N., Lau, S. K., & Chong, A. (2022). Hybrid system controls of natural ventilation and HVAC in mixed-mode buildings: A comprehensive review. Energy and Buildings, 276, 112509. doi: 10.1016/j.enbuild.2022.112509

[2]       Tekler, Z. D., Lei, Y., & Chong, A. (2024). Data-efficient comfort modeling: Active transfer learning for predicting personal thermal comfort using limited data. Energy and Buildings, 319, 114507. doi: 10.1016/j.enbuild.2024.114507.

[3]     Tekler, Z. D., Lei, Y., Dai, X., & Chong, A. (2023, November). Enhancing personalised thermal comfort models with Active Learning for improved HVAC controls. In Journal of Physics: Conference Series (Vol. 2600, No. 13, p. 132004). IOP Publishing. doi: 10.1088/1742-6596/2600/13/132004

[4]     Tekler, Z. D., Lei, Y., Peng, Y., Miller, C., & Chong, A. (2023). A hybrid active learning framework for personal thermal comfort models. Building and Environment, 234, 110148. doi: 10.1016/j.buildenv.2023.110148