Buildings and their heating, ventilation, and air conditioning (HVAC) systems account for a significant share of overall energy consumption. Ventilation systems, in particular, offer considerable energy-saving potential, but traditional rule-based control methods do not always respond optimally to continuously changing operating conditions.
Traditionally, air handling units are controlled using rule-based and schedule-driven methods. Suvanto’s thesis investigated how AI methods could be applied to control systems that learn from data and automatically adapt to different operating situations.
Learning-Based Control Strategy Using Real Building Data
The study utilized deep reinforcement learning, a method particularly well suited for complex optimization problems. One of its key advantages is the ability to learn directly from data and experience without requiring all control rules to be defined in advance.
A common challenge with reinforcement learning algorithms is their extensive need for data and training. As a result, training them directly in a real building would be both time-consuming and costly. Algorithms are therefore typically trained in simulation environments. However, simulated data may not fully capture the behavior of a real building and its complex nonlinear dynamics.
“Traditionally, reinforcement learning algorithms are trained using simulation environments. In this study, we wanted to utilize real building data to ensure that the models would reflect the behavior of an actual property as accurately as possible. The behavior of both the building and the air handling unit was modeled using an LSTM neural network, which served as the training environment for the control algorithm itself,” says Suvanto.
The Long Short-Term Memory (LSTM) neural network was trained using historical data from a real building to recognize the key characteristics of the system’s operation. The network then acted as the training environment for the control algorithm, allowing it to be trained for a specific building without learning directly in the real system.
The actual control algorithm was based on the Soft Actor-Critic (SAC) reinforcement learning method, which is known for its strong performance in solving continuous control problems.
Improved Energy Efficiency Without Compromising Indoor Air Quality
The results of the thesis were promising. The developed control strategy improved the energy efficiency of the air handling unit compared to a traditional control approach while maintaining the same level of indoor air quality.
“In practice, the results showed that the system consumed less energy while delivering the same indoor air quality as the conventional solution used as a benchmark,” says Suvanto.
This finding is particularly noteworthy because the benchmark rule-based control strategy already included demand-based features that adapted to indoor air quality conditions. Despite this, the AI-based solution was able to identify a more energy-efficient operating strategy.
Foundation for Future Development
Although the thesis focused on controlling a single air handling unit, the methodology offers opportunities for much broader applications.
“A single air handling unit was a natural starting point for the research. What surprised us was how clearly energy savings could already be achieved using relatively simple data and by controlling just one unit. The next step is to apply a similar approach to optimizing an entire building or multiple technical systems simultaneously, which could lead to even greater benefits,” Suvanto explains.
The thesis served as a foundation for further development being carried out as part of the AI Champion project. AI Champion is a Co-Innovation project coordinated by Tampere University and one of the pilot initiatives of the Finnish Ministry of Economic Affairs and Employment’s Data Economy Growth Program. The project aims to develop one hundred AI agents to improve automation and information flow across HVAC supply chains.
In Koja Group’s follow-up project, the models developed in the thesis have been further utilized to create a scalable, large language model-based approach for training machine learning control strategies.
The project has been carried out as part of the AI Champion researcher exchange program in collaboration with Koja Group at its Jalasjärvi factory. As part of the project, Researcher Rajratan Wankhade has built an AI Factory for testing air handling units. The AI Factory operates as a multi-agent system powered by multiple AI agents.
From Student to Part of Koja Group’s Development Work
Alina Suvanto’s journey with Koja Group began in the summer of 2025, when she was studying computer science at Aalto University and joined the company as a summer trainee. At the same time, the idea for a master’s thesis emerged, combining AI research with the development of energy-efficient buildings in a real business environment.
“Energy efficiency and building a more sustainable future are meaningful and inspiring topics. This work provided an opportunity to create concrete, practical solutions that can make a real impact,” says Suvanto.
Suvanto completed her master’s thesis in spring 2026 and graduated with a Master of Science in Technology later that summer. Today, she works as a Software Engineer in Koja Group’s Digitalization Team, focusing on software development for Chiller Oy.
“It was extremely rewarding to work on a thesis connected to real development work within a company. What motivated me most was seeing how research can help solve practical challenges while also laying the foundation for future innovation. Now I have the opportunity to continue working on these same themes as part of Koja Group’s Digitalization Team,” says Suvanto.
*****
Explore Alina Suvanto’s Master’s Thesis
Suvanto’s thesis, "Deep Reinforcement Learning for Air Handling Unit Control", has been published in Aalto University’s Aaltodoc repository.
*****
Learn More About the AI Champion Project and the AI Factory
A news article published on the Tampere University website provides further information on the AI Champion project, the operation of AI agents, and the AI Factory created at Koja’s Jalasjärvi facility.
