Towards Human-in-the-Loop Smart Buildings: Data-Driven Predictive Control and Occupant Modeling
Time: Fri 2025-10-10 10.00
Location: K1, Teknikringen 56, Stockholm
Video link: https://kth-se.zoom.us/s/63644880139
Language: English
Subject area: Energy Technology
Doctoral student: Mahsa Farjadnia , Tillämpad termodynamik och kylteknik
Opponent: Professor Alessandra Parisio, Department of Electrical and Electronic Engineering, University of Manchester
Supervisor: Docent Joachim Claesson, Tillämpad termodynamik och kylteknik; Professor Karl H. Johansson, Reglerteknik; Dr. Marco Molinari, Tillämpad termodynamik och kylteknik
Abstract
The building sector accounts for almost 40% of the European Union’s total energy consumption, and a large portion of this consumption is related to heating, ventilation, and air-conditioning (HVAC) systems. In addition to HVAC loads, occupant behavior plays a critical role in building energy use. However, traditional energy performance analyses typically model occupants as passive recipients of indoor conditions rather than as active participants who influence building performance. Such simplifications can lead to notable discrepancies between predicted and actual energy use. Motivated by this challenge, this thesis develops robust data-driven predictive control methods that can explicitly account for uncertainties. This method can be useful at a later stage for including the impact of occupant behaviors in smart building control. Moreover, occupant behavior models are developed by leveraging high-resolution measurements from the KTH Live-In Lab to quantify their impact on heating energy consumption. Finally, it is investigated how social interactions among occupants can improve sustainable behaviors and further reduce building energy use.
The first technical contribution in this thesis is to design a data-driven, robust tube-based zonotopic predictive-control (TZPC) approach for unknown discrete-time linear systems with bounded uncertainties, using input–state data. We prove the recursive feasibility, robust constraint satisfaction, and the robust exponential stability of the closed-loop system. This approach is then extended to unknown nonlinear systems by exploiting reachability analysis and designing a controller that relies solely on input–output data. We prove that the proposed nonlinear zonotopic predictive control (NZPC) approach satisfies the constraints under any admissible bounded uncertainties.
The thesis’s second contribution examines how physical environmental and categorical variables influence occupants’ window operation in a Swedish residential building at the KTH Live-In Lab, based on four years of winter data. Using a multiple logistic regression approach, twelve distinct behavior patterns are modeled. These models are integrated into a digital model of the building to quantify their effect on heating demand. Simulation results indicate that variations in window operation patterns can increase heating energy consumption by up to three times compared to a baseline scenario without window interaction.
Finally, this thesis includes a longitudinal experimental study with selected occupants at the KTH Live-In Lab, investigating the influence of social interactions on promoting sustainable behaviors and reducing energy consumption. The study highlights that digitalizing households into socially interconnected networks effectively improves sustainable lifestyle choices, such as optimized resource use and consumption.