91ÇàÇà²Ý

Title of the thesis: Scalable and Robust Machine Learning Solutions for Adaptive Building Operations

Thesis defender: Davor Stjelja 
Opponent: Prof. Bo Norregaard Jorgensen, University of Southern Denmark, Denmark
Opponent: Prof. Ivo Martinac, KTH Royal Institute of Technology, Sweden
Custos: Prof. Risto Kosonen,Aalto University School of Engineering

Scalable and Robust Machine Learning Solutions for Adaptive Building Operations

The doctoral thesis by Davor Stjelja titled "Scalable and Robust Machine Learning Solutions for Adaptive Building Operations" explores the use of machine learning (ML) techniques to increase the value of building operational data by predicting energy consumption, estimating occupancy, and detecting operational anomalies. The research aimed to develop ML solutions that are both scalable, meaning they can be effectively applied to multiple buildings without extensive data collection, and robust, capable of maintaining accuracy under changing operational conditions.

The thesis addresses significant barriers in the practical implementation of ML in building management, specifically the heavy reliance on extensive ground truth data and the challenges posed by dynamic building environments. By offering solutions to these issues, the research contributes substantially to the broader field of sustainable building operations and energy management.

The thesis successfully utilized unsupervised learning to infer building occupancy from sub-metered energy consumption without labeled data, significantly reducing the dependence on costly ground truth. Transfer learning approaches developed in the study allowed accurate occupancy predictions using minimal labeled data (only a few days), enhancing scalability across diverse environments and creating models capable of generalizing across different building conditions. The research introduced a probabilistic model for energy consumption forecasting and employed a domain-informed approach to filter repeatable anomalies, achieving robust performance even when operational conditions, such as ventilation strategies, changed.

The developed methods can be directly applied to real-world building management scenarios, improving energy efficiency, reducing operational costs, optimizing space utilization, and enhancing occupant comfort. Facility managers can utilize these scalable ML solutions to automate occupancy-based control systems, predict energy demand reliably, and promptly detect operational anomalies.

The thesis demonstrates that it is possible to achieve scalability and robustness in ML-based solutions for adaptive building operations, while significantly enhancing the value extracted from routine operational data. The developed approaches advance the potential for more sustainable and resilient energy management across diverse building portfolios.

Keywords: Machine Learning, Smart Buildings, Building Analytics, Data-Driven Building Management, Occupancy Prediction, Anomaly Detection

Thesis available for public display 10 days prior to the defence at . 

Contact information: 
E-mail: davor.stjelja@aalto.fi