Bratislav Svetozarevic

Dr. sc. ETH Bratislav Svetozarevic
Empa 
Swiss Federal Laboratories for Materials Science and Technology
Überlandstrasse 129
CH-8600 Dübendorf

Email: 

Office tel: + 41 58 765 65 35

 

Google ScholarResearch GateLinkedIn

About

Bratislav Svetozarevic is a post-doctoral researcher at the Urban Energy Systems laboratory at Empa, working within the ehub group. He joined Empa in November 2018.

His current work is related to data-driven, learning-based control algorithms for processes and systems in buildings. The goal of self-optimizing control algorithms is to improve the operational efficiency of buildings and offer (close-to) optimal control performance over the lifespan of a building without a substantial or any need for human interventions. Therefore, besides energy savings, these algorithms reduce the need for maintenance in buildings and allow for continuous commissioning, i.e. automatic recommissioning after a part of the system was replaced, thus reducing the overall buildings' operational costs. Bratislav's particular interests lie in the transferability and large-scale applicability of these algorithms within building systems and districts with integrated renewable energy generation, conversion, and storage systems, as well as connected electric vehicles with uni- and bi-directional charging. 

Bratislav Svetozarevic has a highly interdisciplinary background and his research is bridging the fields of automatic control, machine learning, building systems, and energy management. Before moving to EMPA, he was part of several single- and interdisciplinary teams in academia (Chair of Architecture and Building Systems, ETH Zurich  – dynamic photovoltaic building envelopes, 2014 - presentAutomatic Control Laboratory, ETH Zurich – fault detection and isolation for wind turbines, 2010-2012, and ETF Robotics, University of Belgrade – control algorithms for compliant, tendon-driven robotic mechanisms, 2009-2010and industry (spin-off Siltectra GmbH (2013), acquired in 2018 – R&D engineer for the automated waste-free splitting of crystal materials). He has led and co-authored publications in engineering (IEEE, Applied Energy, Energy and Buildings) and science (Nature) in the domain of energy management, smart building systems, and robotics. He is a leading co-author of a Nature Energy paper on dynamic BIPV facades, which was also featured on the cover page of the August issue 2019.

Bratislav Svetozarevic received a PhD at the Swiss Federal Institute of Technology (ETH) in Zurich. He worked at the Chair of Architecture and Building Systems under the supervision of Prof. Arno Schlueter and co-supervision of Prof. Zoltan Nagy (the University of Texas at Austin). In his PhD work, he developed a hybrid soft-hard-material actuator and applied it within a multi-element dynamic photovoltaic façade. He is a leading co-author of a patent that was awarded a prestigious Spark award for Top 20 innovations at ETH Zurich. 

Ongoing research projects

Data-driven, self-tuning RB and PID controllers for large-scale deployment

More than 90% of industrial controllers are still based on rule-based (RB) control algorithms, such as proportional-integral-derivative (PID), as no other algorithms match the simplicity, clear functionality, and applicability of these two types. However, the manual tuning of PID and RB controllers, which requires optimization of a few parameters, is a time-consuming task and requires high-skilled personal. During the lifetime of a system, due to ageing or changing of some components of the system, the overall process conditions typically change, leading to sub-optimal performance of the system. This calls for manual re-tuning of the controllers and incurs additional maintenance (personal costs) and operational costs, due to the downtime of the system during the re-tuning. 

In this project, we aim to develop new data-driven self-adaptive control strategies and compare them to the state-of-the-art autotuning RB control strategies, as well as to the classical PID and RB controllers. In this, we will particularly look at the scalability potentials of the algorithms, i.e. in the algorithms that do not require substantial, if any at all, engineering or expert skills, to apply them to another system within the same class of systems or even from another class of systems. Our current efforts involve the development of controllers for heat pump systems and room temperature control algorithms.  

Funding body: internal

Partners: ETH Zurich, Industry

 

Deep reinforcement learning for building control

Both conventional, rule-based (RB) and advanced, model-based controllers, such as Model Predictive Control (MPC) show limitations to control modern buildings. RB controllers need manual tuning and are suitable only for single control loops, thus cannot provide system wide optimisation. On the other hand, MPC controllers can provide system wide optimal performance, but require a model of the building, which is expensive to obtain. Recent advancements in reinforcement learning (RL), and in particular deep RL (DRL), have attracted growing research interest among building control engineers and demonstrated the potential to enhance the building performance while addressing the limitations of RB and MPC control techniques. We have developed a first prototype of a fully black-box, data-driven pipeline to obtain building control policies based on DRL. We have successfully implemented it and tested at different units of NEST demonstration building at our campus and for different purposes, from room temperature control to EV charging. We are currently expanding the efforts in this domain with the focus on the following topics:

  • Physics-inspired, data-efficient DRL methods (PhD student Loris Di Natale)
  • Transfer learning for DRL-based building control
  • Extraction of simple rules from the DRL control policies for facilitated industry adoption of these algorithms
Funding body: internal

Partners: ETH Zurich, EPFL, Industry

 

Energy and comfort optimization in living and working environments through a user-centred predictive control

Abstract: User need for comfort is one of the most critical barriers for energy-efficient living and working. Recent advancements in automatic optimization of room climate successfully foster energy efficiency, but usually, ignore comfort needs. In this project, we will introduce an approach that aims at maximizing energy efficiency and user comfort at the same time. For this purpose, Empa NEST is used as a testbed. State of the art energy optimization algorithms based on reinforcement learning are combined with real-time user feedback. Our previous work suggests an energy-saving potential of around 20%.

Funding body: Swiss Federal Office of Energy  
Project title: Energy and comfort optimization in living and working environments through a user-centred predictive control (Starts in 2021.)
Partners: Industry
Student supervision

PhD students:

  • 2021 - ongoing: Xu Wenjie, MSc Chinese University of Hong Kong, project title: Tuning Digital Twins via Differentiable ComputingCo-supervised with Prof. Colin Jones (EPFL) and Philipp Heer (Empa)
  • 2020 - ongoing: Loris Di Natale, MSc EPFL, project title: Physics-inspired deep reinforcement learning for building controlCo-supervised with Prof. Colin Jones (EPFL) and Philipp Heer (Empa)
    Abstract:  Existing works on Deep Learning (DL) and Deep Reinforcement Learning (DRL) have largely overlooked the vast existing knowledge we have of most dynamical systems, mostly from the laws of physics. As a consequence, a new field of study has emerged on physics-based neural networks (NNs), and more generally on the various possibilities to introduce prior knowledge in DL schemes. We conjecture that incorporating physics-inspired intuition in the model architectures or learning procedures will be a stepping stone towards generally applicable - i.e. safe, scalable and data-efficient - DRL approaches. In this work, we will thus explore various solutions to introduce prior knowledge in (building) models and DRL agents, design algorithms to apply them, and deploy the resulting controllers on existing case studies to assess their advantages. 

 

Master students:

2021

  • MSc thesis, Kosta Jovanovic, BSc University of Belgrade, project title: Transfer learning for deep reinforcement learning with applications to building control, Co-supervised with Prof. Predrag Tadic (University of Belgrade) and Philipp Heer (Empa)
  • MSc thesis, Marko Skakun, BSc University of Belgrade, project title: Extracting simple rules from deep reinforcement learning building control policies, Co-supervised with Prof. Predrag Tadic (University of Belgrade) and Philipp Heer (Empa)
  • MSc thesis, Ana Dodig, BSc University of Belgrade, project title: Deep reinforcement learning-based occupant centred control, Co-supervised with Prof. Predrag Tadic (University of Belgrade) and Philipp Heer (Empa)

2020

  • MSc thesis, Luca Funk, BSc ETH Zurich, project title: The impact of bidirectional charging electric vehicles on the Swiss electricity system. Co-supervised with Prof. Giovanni Sansavini (ETH Zurich) and Dr. Martin Rüdisüli (Empa)

2019

  • MSc thesis, Christian Baumann, BSc ETH Zurich, project title: Joint control of bidirectional electric vehicle charging and home energy scheduling using reinforcement learning. Co-supervised with Prof. Melanie Zeilinger (ETH Zurich)
  • Internship, Angelos MikelisMSc NTUA, project title: PID tuning using extremum seeking: Application to a heat pump systemCo-supervised with Prof. Miroslav Krstic (UC San Diego)
  • MSc thesis, Jeffrey LungthokBSc ETH Zurich, project title: Safe Bayesian optimization with improved convergence speed for online autotuning of PI controllers: Simulation and experimental resultsCo-supervised with Prof. Andreas Krause (ETH Zurich)
Publications

 

Google ScholarResearch Gate

News articles:

  • Darum braucht dieses Haus 25 Prozent weniger Strom, 20min.ch, April 2021 (link)
  • AI for electricity distribution: Energy house-keeping, Empa media release, March 2021 (link)
  • Energy house-keepingEmpa Quarterly Magazine, January 2021 (pdf)

Journal and conference papers in preparation / submitted:

  • X. Wenjie, C. Jones, B. Svetozarevic, C. Laughman, A. Chakrabarty, "VABO: Violation-aware Bayesian optimization for closed-loop control performance optimization with unmodeled constraints", (submitted) 
  • B. Svetozarevic, C. Baumann, S. Muntwiler S., L. Di Natale, P. Heer, M. Zeilinger, "Data-driven MIMO control of room temperature and bidirectional EV charging using deep reinforcement learning: simulation and experiments", (submitted; preprint: https://arxiv.org/abs/2103.01886
  • L. Di Natale, B. Svetozarevic, P. Heer, C. Jones, "Deep Reinforcement Learning for room temperature control: a black-box pipeline from data to policies", (CISBAT 2021, to appear)
  • L. Di Natale, F. Luca, M. Rüdisüli, B. Svetozarevic, B. Pareschi, P. Heer, G. Sansavini, "The Potential of Vehicle-to-Grid to Support the Energy Transition: A Case Study on Switzerland", Energies, 2021; 14(16):4812. https://doi.org/10.3390/en14164812
  • B. SvetozarevicJ. Lungthok,  P. Heer, A. Krause, "Safe Bayesian optimization with guaranteed convergence speed for online autotuning of feedback controllers: simulation and experiments", (in preparation)
  • B. SvetozarevicL. Di Natale,  P. Heer, C. Jones, "Deep reinforcement learning for building control: Towards scalable and transferable control policies", (in preparation)
  • L. Di Natale, B. Svetozarevic, P. Heer, C. Jones, "Physically consistent neural network predictions from data: an application to room temperature modeling", (in preparation)

 

Journal articles (selected):

  • B. Svetozarevic, M. Begle, P. Jayathissa, Z. Nagy, S. Caranovic, R. Shepherd, I. Hischier, J. Hofer,  A. Schlueter, "Dynamic photovoltaic building envelopes for adaptive energy and comfort management", Nature Energy (4), 671-682, 2019. https://doi.org/10.1038/s41560-019-0424-0 (also cover page)
  • B. Svetozarevic, M. Begle, S. Caranovic, Z. Nagy, A Schlueter, "Quick-cast: A method for fast and precise scalable production of fluid-driven elastomeric soft actuators", Extreme Mechanics Letters, in press. https://arxiv.org/abs/1810.05920
  • Z. Nagy, B. Svetozarevic, P. Jayathissa, M. Begle, J. Hofer, G. Lydon, A. Willmann, A. Schlueter, "The Adaptive Solar Facade: From concept to prototypes", Frontiers of Architectural Research 5 (2), 143-156, 2016. https://doi.org/10.1016/j.foar.2016.03.002

Conference papers (selected):

  • B. Svetozarevic, Z. Nagy, J. Hofer, D. Jacob, M. Begle, E. Chatzi, and A. Schlueter, "SoRo-Track: A Two-Axis Soft Robotic Platform for Solar Tracking and Building-Integrated Photovoltaic Applications", IEEE International Conference on Robotics and Automation (ICRA), 2016. https://doi.org/10.1109/ICRA.2016.7487700
  • B. Svetozarevic, Z. Nagy, D. Rossi, A. Schlueter, "Experimental Characterization of a 2-DOF Soft Robotic Platform for Architectural Applications", Proceedings of the Robotics: Science and Systems, Workshop on Advances on Soft Robotics, Berkeley, CA, USA, 2014. https://www.research-collection.ethz.ch/handle/20.500.11850/92114
  • B. Svetozarevic, P. M. Esfahani, M. Kamgarpour, J. Lygeros, "A Robust Fault Detection and Isolation Filter for a Horizontal Axis Variable Speed Wind Turbine", Proceedings of the American Control Conference, 4453 – 4458, 2013. https://doi.org/10.1109/ACC.2013.6580526

Patents:

  • B. Svetozarevic, A. Schlueter, M. Begle, J. Prageeth, and S. Caranovic, Hybrid hard-soft-material pneumatic actuator with adjustable mechanical impedance, Patent No. WO/2019/096642 (link)
Awards
  • Spark award for Top 20 innovations at ETH Zurich, 2018.
Talks
  • Deep reinforcement learning for building control: From energy and comfort management to scalability, Empa internal, UES Lab, Topical meeting, April 2021
  • Deep reinforcement learning for building control, Machine learning workshop, Empa, January 2021 (150+ participants)
  • PID autotuning via two-stage safe Bayesian optimization, Applied Machine Learning Days, 2020 (link to video