What is it about?
Buildings use a lot of energy, and a big part of that depends on how people actually live in them, when they are home, asleep, cooking, or away. However, many energy models rely on simple or unrealistic assumptions about daily routines, leading to inaccurate predictions. In this study, we developed a method that uses artificial intelligence to create realistic daily activity patterns for people living in homes. Instead of relying on fixed schedules, our model learns from real survey data about how people spend their time and generates many different, lifelike daily routines. We tested these generated routines and showed that they closely match real human behavior while still capturing natural variation from day to day. We then used them in building energy simulations and found that they can improve the accuracy of energy use estimates. They also help evaluate smarter control strategies, such as turning heating or cooling systems on and off based on whether people are home. Overall, this work helps make building energy models more realistic and useful. By better representing how people actually live, it can support the design of more energy-efficient homes and smarter energy management strategies.
Featured Image
Photo by Apartment Life on Unsplash
Why is it important?
Most building energy models assume simple and fixed daily routines, such as people being home at the same times every day. In reality, human behavior is much more varied and unpredictable. This gap can lead to inaccurate energy predictions and missed energy-saving opportunities. Our work is important because it introduces a new way to generate realistic daily activity patterns using artificial intelligence, based on real-world data. Unlike previous approaches, our method captures both the timing and variability of human behavior in fine detail throughout the day. This allows building simulations to better reflect how people actually live. This is especially timely as buildings become smarter and more responsive, with systems that adjust heating, cooling, and ventilation based on occupancy. Accurate behavior models are essential for these technologies to work effectively. By improving how occupant behavior is represented, our approach can lead to more reliable energy predictions, better design of energy-efficient buildings, and more effective control strategies. Ultimately, this can help reduce energy consumption, lower costs, and support broader sustainability goals.
Perspectives
Working on this paper was particularly meaningful to me because it sits at the intersection of human behavior and building systems, two areas that are often studied separately. One of the biggest challenges in building energy modeling is that people are unpredictable, yet most models treat them as if they follow simple, fixed routines. This disconnect has always felt like a missing piece in making simulations truly realistic. What I found most exciting about this work is that it shows how data-driven and AI-based approaches can begin to bridge that gap. Instead of simplifying human behavior, we embraced its variability and complexity and demonstrated that it can be modeled in practice. Seeing the generated activity patterns closely resemble real-life behavior, and then observing their impact on energy simulations, was especially rewarding. I also see this work as a step toward more human-centered building design. As buildings become smarter and more responsive, understanding how people actually live in them becomes increasingly important. I hope this research encourages others to think more deeply about the role of occupant behavior and inspires further work that connects human dynamics with engineering systems in a meaningful way.
Yunyang Ye
Colorado School of Mines
Read the Original
This page is a summary of: An autoregressive-based framework for simulating stochastic occupant behavior in residential buildings, Building Simulation, March 2026, Tsinghua University Press,
DOI: 10.1007/s12273-026-1406-3.
You can read the full text:
Contributors
The following have contributed to this page
