Robust and data-Efficient Learning for Industrial Control

Informacje na temat projektu

RELIC

Identyfikator umowy o grant: 101063948

DOI

10.3030/101063948

Realizacja projektu zakończyła się 31 Sierpnia 2024

Data podpisania przez KE 9 Czerwca 2022

Data rozpoczęcia 5 Czerwca 2023

Data zakończenia 4 Czerwca 2025

Finansowanie w ramach

• Marie Skłodowska-Curie Actions (MSCA)

Koszt całkowity

Brak danych

Wkład UE

€ 210 911,04

Koordynowany przez

NORGES TEKNISK-NATURVITENSKAPELIGE UNIVERSITET NTNU
Norway

Periodic Reporting for period 1 - RELIC (Robust and data-Efficient Learning for Industrial Control)

Okres sprawozdawczy: 2023-06-05 do 2025-06-04

Podsumowanie kontekstu i ogólnych celów projektu

Increasing energy and resource efficiency in industrial systems is key to decrease harmful emissions by 90% by 2050. Reaching the environmental targets requires a holistic approach to how resources and energy are delivered to the industry by means of distribution networks, such as heat networks, electricity networks, or gas transport networks. Existing control frameworks are usually application specific and have limited use in large-scale systems. In the project, I wanted to advance theory in data analytics and optimisation and build on my industrial experience to develop operating strategies for distribution networks that will enable safe implementation and reaching the environmental targets.

The goal of the project was to address the gaps by designing a framework to solve control problems using minimal information about the system and minimal computational power. The objectives of the project were:
A. Development of a learning control system using knowledge of physics and new information available in real time;
B. Development of an efficient learning control algorithm to satisfy long-term environmental targets and safety requirements;
C. Development of a numerically robust control algorithm considering limited computational power available.

The first step of the project was to develop an efficient control framework for large-scale systems combining the knowledge about the physics with measured data. To enable efficient use of data in control, the next step was to analyse how the quality of data from monitoring systems affects learning control frameworks. Implementation in distribution networks requires considering limitations in how industrial equipment, such as pumps or compressors, can be safely operated. To achieve this goal, I worked on improving numerical implementation of safe-learning algorithms.

Prace wykonane od początku projektu do końca okresu sprawozdawczego oraz najważniejsze dotychczasowe rezultaty

Work Package 1
The first step of the project was to develop an efficient control framework for large-scale systems combining the knowledge about the physics with measured data. To this end, I worked in three areas:
1. A review of data-driven methods of degradation modelling (T1.1) in collaboration with AGH Krakow. My contributions consisted in supervision of a PhD student, writing, revisions. As a follow up, there is a planned submission.
2. Physics-aware Bayesian optimization (T1.1) in collaboration with Imperial College London. My contribution consisted in mathematical formulation of physics-aware Bayesian optimization problem as a bi-level optimization. Published at CCTA 2024:
• L. Dong, M. Zagorowska, T. Liu, A. Durkin, and M. Mercangöz. Pi-cof: A bilevel optimization framework for solving active learning problems using physics-information. In Proceedings of the 8th IEEE Conference on Control Technology and Applications (CCTA) 2024, 2023. August 21-23, 2024, Newcastle upon Tyne, UK. Preprint: arxiv.org/abs/2402.13588 available 6 Sep 2024
3. Online Feedback Optimization (T2.1 T3.2) in collaboration with ZHAW Centre for Artificial Intelligence, Eastern Switzerland University of Applied Sciences, and Imperial College London. My contribution here consisted in developing the idea, getting feedback from collaborators, subsequent implementation, writing, and revisions. Published at ADCHEM 2024:
• M. Zagorowska, L. Ortmann, A. Rupenyan, M. Mercangöz, and L. Imsland. Tuning of online feedback optimization for setpoint tracking in centrifugal compressors. In Proceedings of the 12th IFAC Symposium on Advanced Control of Chemical Processes (ADCHEM 2024), 2023. July 14-17, 2024, Toronto, Canada. Preprint: arxiv.org/abs/2312.01996 available 8 Dec 2023
4. Long-term Bayesian Optimization (T1.3) in collaboration with ZHAW Centre for Artificial Intelligence and ETH Zürich. My contribution here consisted in supervision of the work, providing a case study, writing, and revisions. Accepted to NeurIPS 2024:
• J. Li, M. Zagorowska, G. De Pasquale, A. Rupenyan, and J. Lygeros. Safe Time-Varying Optimization based on Gaussian Processes with Spatio-Temporal Kernel. In The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS 2024), 2024. December 10-15, 2024, Vancouver, Canada. Preprint: arxiv.org/abs/2409.18000 available 11 Oct 2024

Work Package 2
The objective of this task was to analyse how the quality of data from monitoring systems affects learning control frameworks. To achieve this, I worked on:
1. Modelling of degradation of electric motors (T1.2) in collaboration with AGH Krakow. My contributions consisted in writing and revisions. Accepted to IECON 2024.
• W. Bauer, M. Zagorowska, J. Baranowski. Detection of Electric Motor Damage Through Analysis of Sound Signals Using Bayesian Neural Networks. In Proceedings of the 50th Annual Conference of the IEEE Industrial Electronics Society (IECON 2024). November 3-6, 2024, Chicago, Illinois. Preprint: https://arxiv.org/abs/2409.08309 available 11 Oct 2024
2. Sensitivity of Online Feedback Optimization, especially for automatic tuning (T2.2 T3.3). My contributions include problem formulation and implementation. As a follow up, there is a planned submission to CCTA 2025.
3. Online Feedback Optimization with incomplete information, in collaboration with an MSc student of Lars Imsland. As a follow-up, there is a planned submission to CCTA 2025.

Work Package 3
The objective of this task was to Implementation in distribution networks requires considering limitations in how industrial equipment, such as pumps or compressors, can be safely operated. To achieve this, I worked on:
1. Numerically efficient methods for safe learning, in collaboration with ETH Zürich, ZHAW Centre for Artificial Intelligence, inspire AG. My contributions include problem formulation, writing, supervision, revisions (T3.1).
• M. Zagorowska, C. König, H. Yu, E. C. Balta, A. Rupenyan, and J. Lygeros. Efficient safe learning for controller tuning with experimental validation. Engineering Applications of Artificial Intelligence, 2023. Under review, manuscript number: EAAI-23-9169, preprint: arxiv.org/abs/2310.17431 available 21 Nov 2023
4. Use of limited data for Bayesian modelling (T3.3) in collaboration with AGH Krakow. My contributions consisted in problem formulation, supervision of a PhD student, writing, revisions. As a follow up, there is a planned submission.

Innowacyjność oraz oczekiwany potencjalny wpływ (w tym dotychczasowe znaczenie społeczno-gospodarcze i szersze implikacje społeczne projektu)

Overall research impact: My results are published in peer-reviewed conference proceedings to reach scientific audience and end users. I also used conference presentations to disseminate the results among peers and industrial practitioners (CCTA 2024, ADCHEM 2024, IECON 2024). In September 2023, I also presented an overview of my work to the group of Lars Imsland. I also gave a presentation at the Department of Engineering Cybernetics at NTNU as a part of “Cybernetic quarter” on 17.06.2024 with the title “Optimization algorithms as controllers, and how to tune them”. In particular, the algorithm proposed in the paper "Safe Time-Varying Optimization based on Gaussian Processes with Spatio-Temporal Kernel" extends the state of the art in Time-Varying Safe Bayesian Optimization by enabling solving optimization problems with time-varying reward and constraints without pre-defining the time changes that can be compensated. As such, the algorithm can be used at the design stage of operating strategies for safety-critical systems, such as medical dosage design and controller design in robotics, or during online operation of chemical plants or autonomous racing. To facilitate further uptake of results, it is necessary to validate the algorithm in a real-life implementation or a realistic simulator.

As a part of knowledge transfer to NTNU, I conducted a lecture on the introduction to Julia (open-source programming language) during the course Mixed integer optimization in energy and oil and gas systems led by Lars Imsland. As a part of my career development, I collaborated with one Master’s student from Lars Imsland on application Online Feedback Optimization with incomplete model information.