OctoAI: The next generation of high-performance edge AI for smart buildings

Motivation and Research Question

The building stock in the European Union accounts for approximately 40% of final energy consumption and 36% of CO₂ emissions. Increasing urbanization, rising energy costs, and the growing demand for energy-efficient solutions require innovative approaches to reducing energy consumption and integrating renewable energy sources. At the same time, existing cloud-based systems pose challenges regarding data protection, latency, and availability.

The OctoAI research project explores the application of edge AI technologies for intelligent buildings to enable data-driven decisions efficiently and securely. The goal is to develop alternative computing models that improve both energy efficiency and comfort in buildings.

Initial Situation/Status Quo

Many current smart building applications rely on cloud computing, which presents significant challenges: high latency, dependence on a stable internet connection, and data protection and security risks. Edge computing offers an alternative where AI models are processed directly on end devices. This reduces response times, ensures data sovereignty for users, and increases reliability. Despite these advantages, edge AI adoption in buildings is still limited, and there is a lack of practical applications with validated use cases.

Project Content and Objectives

The OctoAI project focuses on the development and testing of edge AI technologies in intelligent buildings. Two key use cases have been identified that will be optimized using innovative machine learning techniques.

The first use case is occupancy detection. This examines how sensor data can be used to determine the presence of people in indoor spaces as accurately as possible. Data collected, including CO₂ concentration, noise levels, light intensity, and door opening status, were processed using various AI methods to enable precise room occupancy classification. The primary goal was to develop resource-efficient models that can run efficiently on edge devices without requiring cloud connectivity. In addition to improving building automation through intelligent heating and ventilation systems, the focus was also on minimizing energy consumption.

The second use case involves thermal comfort prediction. A model was developed to predict room temperature for the next 24 hours with high accuracy. By integrating weather forecasts and real-time sensor data, an adaptive climate control system was achieved. The model is based on the EN 16798-1 standard for comfort zones and provides a data-driven decision basis for efficient building heating and cooling management.

The project aimed to develop robust and scalable edge AI models that can be implemented in both new and existing buildings. Optimizing computing power, ensuring high data quality, and practically validating the developed models in real test environments were central to the project. Additionally, the project sought to contribute to the general acceptance and dissemination of edge AI technologies by making the benefits of this technology tangible for building operations.

Methodical Approach

The OctoAI project followed an iterative and data-driven approach that encompassed the development, implementation, and validation of AI-based models in various real test environments. Initially, a comprehensive requirements analysis was conducted to identify the key needs and challenges in building automation. Expert interviews and a large-scale stakeholder survey provided valuable insights into the status quo and the expectations of industry participants.

After defining the technical requirements, sensor data were collected and processed in pilot projects. Sensors for measuring air quality, noise levels, light intensity, and room temperature were installed in two different office buildings. These datasets formed the basis for training the AI models. Various machine learning techniques were tested and compared for occupancy detection, including neural networks, decision trees, and classical statistical methods. Special attention was given to evaluating model accuracy under real conditions and optimizing the algorithms for use on resource-constrained edge devices.

At the same time, a model for thermal comfort prediction was developed based on autoregressive techniques and external influencing factors such as weather data. The models were tested in real-time, and their prediction accuracy was checked through rolling validation. The goal was to create a reliable method that dynamically adapts to changing environmental conditions and provides practical recommendations for building managers.

A key component of the project was also the development of an interactive dashboard that visualizes sensor data and enables users to make informed decisions based on model predictions. Usability tests with various target groups helped optimize the dashboard's user-friendliness and improve the practical applicability of the developed solutions.

Results and Conclusions

The developed occupancy detection models achieved an accuracy of up to 87%, allowing precise control of heating and ventilation systems. The predictive models for thermal comfort assessment were able to forecast temperature trends with a mean error (RMSE) of 0.158°C, providing a solid foundation for future control strategies.

Additionally, it was demonstrated that edge AI is suitable for intelligent buildings and can deliver reliable results even with low-performance hardware. Users benefited from reduced latency times and higher data security.

Outlook

The results of the OctoAI project lay the foundation for further developments in the field of edge AI for buildings. Future research should focus on integrating the technology into existing smart building systems, expanding sensor networks, and scaling the technology for larger building complexes. Furthermore, combining edge AI with decentralized energy supply systems offers promising perspectives for energy-efficient and resilient buildings of the future.

Project management

TU Graz - Institut for Software Technology

Project or cooperation partners

• TU Graz - Institute for Building Physics, Building Technology and Structural Engineering
• DiLT Analytics GmbH

TU Graz
Institut for Software Technology
Gerald Schweiger
Inffeldgasse 16b
A-8010 Graz
E-mail: gerald.schweiger@tugraz.at
Web: www.tugraz.at/institute/ist/institute/