Scalable and Robust Machine Learning Solutions for Adaptive Building Operations

Thumbnail Image

Files

isbn9789526425894.pdf (10.3 MB)

URL

Journal Title

Journal ISSN

Volume Title

School of Engineering | Doctoral thesis (article-based) | Defence date: 2025-06-13
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2025

Department

Energia- ja konetekniikan laitos
Department of Energy and Mechanical Engineering

Major/Subject

Mcode

Degree programme

Language

en

Pages

91 + app. 73

Series

Aalto University publication series Doctoral Theses, 113/2025

Abstract

Achieving a more efficient, resilient, and low-emission energy system is crucial for long-term sustainability. One of the primary paths for this transformation is the digitalization of energy systems and buildings, which has led to a surge in operational data. Machine learning (ML) methods can leverage this data to provide deeper insights, automate complex processes, and reveal new opportunities in building management. Specifically, ML methods are applied in this dissertation to predict energy consumption, estimate building- and room-level occupancy, and detect changes in consumption patterns. This dissertation addresses two main challenges in applying ML to these building operation applications: scalability and robustness. The first goal is scalability, ensuring that the developed ML solutions can be applied across various buildings, each with its own unique characteristics and operational dynamics. Achieving scalability means minimizing the need for extensive ground truth data collection, which can be costly and time-consuming. By reducing reliance on ground truth data, the dissertation tries to make these solutions more practical and transferable across different building environments. The second key goal is model robustness, the ability of machine learning systems to consistently perform well, even when faced with changing environments and varying operational conditions. This characteristic is fundamental for systems deployed in dynamic building environments. Furthermore, deviations from robustness can reveal underlying shifts in system behavior. This objective focuses not only on developing models that adapt to such changes, but also on utilizing robustness failures as signals to inform and enhance decision making in real-world scenarios. To meet these goals, the study utilizes building data on energy consumption, sub-metering, indoor temperature, and CO2 levels. Scalability was enhanced through several strategies: unsupervised learning (clustering) was applied to infer occupancy from sub-metered consumption withoutground truth data, while transfer learning enabled room-level occupancy prediction using only a few days of labeled CO2 and temperature data, significantly reducing the dependence on extensive ground truth. Additionally, domain knowledge was incorporated into a probabilistic energy consumption model, improving its ability to detect meaningful shifts and potential anomalies without explicit labels. Robustness was improved by employing transfer learning combined with sliding normalization, enabling model to adapt effectively to operational changes, such as adjustments in ventilation settings. Additionally, deviations in model performance were leveraged as signals to identify anomalies, utilizing domain-informed probabilistic modeling. Tests carried out with low to moderate noise in the training data confirmed that the proposed method maintained reliable performance. Overall, this research demonstrates the feasibility of scalable and robust ML solutions that capitalize on readily available building operations data. Accurate occupancy estimates, energy consumption predictions, and anomaly detection can be achieved with minimal reliance on extensive ground truth, paving the way for more data-driven, efficient, and robust building management. The work shows that it is possible to balance scalability and robustness by designing solutions that generalize across diverse buildings while remaining responsive to changing operational conditions.

Description

Supervising professor

Kosonen, Risto, Prof., Aalto University, Department of Energy and Mechanical Engineering, Finland

Thesis advisor

Jokisalo, Juha, Dr., Aalto University, Department of Energy and Mechanical Engineering, Finland

Keywords

machine learning, smart buildings, building analytics, data-driven building, management, occupancy prediction, anomaly detection

Other note

Parts

  • [Publication 1]: D. Stjelja, J. Jokisalo, R. Kosonen. From Electricity and Water Consumption Data to Information on Office Occupancy: A Supervised and Unsupervised Data Mining Approach. Applied Sciences, 10, 24, 9089, December 2020.
    Full text in Acris/Aaltodoc: https://urn.fi/URN:NBN:fi:aalto-2020123160445
    DOI: 10.3390/app10249089 View at publisher
  • [Publication 2]: D. Stjelja, J. Jokisalo, R. Kosonen. Scalable Room Occupancy Prediction with Deep Transfer Learning Using Indoor Climate Sensor. Energies, 15, 6, 2078, March 2022.
    Full text in Acris/Aaltodoc: https://urn.fi/URN:NBN:fi:aalto-202204062778
    DOI: 10.3390/en15062078 View at publisher
  • [Publication 3]: D. Stjelja,V. Kuzmanovski, J. Jokisalo, R. Kosonen. Building consumption anomaly detection: A comparative study of two probabilistic approaches. Energy & Buildings, 313, 114249, May 2024.
    Full text in Acris/Aaltodoc: https://urn.fi/URN:NBN:fi:aalto-202405293993
    DOI: 10.1016/j.enbuild.2024.114249 View at publisher

Citation

Permanent link to this item

https://urn.fi/URN:ISBN:978-952-64-2589-4

Collections

[diss] Insinööritieteiden korkeakoulu / ENG