Periodic Reporting for period 3 - ScalableControl (Scalable Control of Interconnected Systems)
Période du rapport: 2022-09-01 au 2024-02-29
One of the greatest technical challenges in our society is to efficiently, sustainably and safely control large infrastructure systems. Applications include networks for transportation, communication and energy supply, but also industrial production, logistics and health care. Rapidly growing access to sensors, actuators and computers has created a multitude of new opportunities for control technology in these areas. Structured and integrated methods for control are required to render them efficient and reliable. The EU-funded ScalableControl project aims to address this need, by creating rigorous and efficient methods for control synthesis, adaptation and verification of large-scale interconnected systems. It exploits theory for monotone and positive systems to develop scalable methods for control, as well as machine learning techniques for design of dynamic feedback systems. Collaboration with industry on district heating networks gives access to experimental data that can validate and inspire relevant theory and methodology.
Progress has been made along all main directions of the proposal:
1) Scalable methods for control synthesis have been developed by mathematically exploiting properties of symmetry and monotonicity that appear naturally in many infrastructure networks. A number of important special cases have been treated successfully, but we are still looking for a unifying framework that cover all the essential aspects.
2) Methods for learning and adaptation have been developed under the assumption of adversarial disturbances. This has lead to a remarkable breakthrough, where the standard Riccati equation for optimal robust control has been extended to give rigorous bounds for robustness and performance of adaptive feedback loops. The result was presented at the main venue of the field, the symposium L4DC (Learning for Dynamics and Control) 2021. Follow-up work in several directions is now carried out, including analysis of optimality, scalability, regret and state estimation.
3) The study of decomposable certificates for robustness verification has been carried out in synergy with a project funded by the European Space Agency, where machine learning based components for use in space missions need to be certified before the launch. New certificates based on Integral Quadratic Constraints have been developed for neural network components used in dynamic feedback systems. Results are promising, but still at an early stage.
4) Applications in district heating are being investigated in parallel to create synergies with the theoretical research. The main problem studied so far concerns networks operating close to their capacity limits. It has been observed that in such situations consumers far away from production sites are more likely to suffer from delivery losses due to insufficient capacity. In heating networks this leads to reduced temperatures in the buildings. To address this problem, we have developed updated control methodology that creates a fair distribution of the available capacity, resulting in drastically reduced temperature losses for the individual customers.
1) Scalable methods for control synthesis have been developed by mathematically exploiting properties of symmetry and monotonicity that appear naturally in many infrastructure networks. A number of important special cases have been treated successfully, but we are still looking for a unifying framework that cover all the essential aspects.
2) Methods for learning and adaptation have been developed under the assumption of adversarial disturbances. This has lead to a remarkable breakthrough, where the standard Riccati equation for optimal robust control has been extended to give rigorous bounds for robustness and performance of adaptive feedback loops. The result was presented at the main venue of the field, the symposium L4DC (Learning for Dynamics and Control) 2021. Follow-up work in several directions is now carried out, including analysis of optimality, scalability, regret and state estimation.
3) The study of decomposable certificates for robustness verification has been carried out in synergy with a project funded by the European Space Agency, where machine learning based components for use in space missions need to be certified before the launch. New certificates based on Integral Quadratic Constraints have been developed for neural network components used in dynamic feedback systems. Results are promising, but still at an early stage.
4) Applications in district heating are being investigated in parallel to create synergies with the theoretical research. The main problem studied so far concerns networks operating close to their capacity limits. It has been observed that in such situations consumers far away from production sites are more likely to suffer from delivery losses due to insufficient capacity. In heating networks this leads to reduced temperatures in the buildings. To address this problem, we have developed updated control methodology that creates a fair distribution of the available capacity, resulting in drastically reduced temperature losses for the individual customers.
As described above, the project has already delivered significant advancements beyond the state of art. Continued work along the main directions will consolidate the achievements and most likely also create significant synergies across the different directions. As a result, the project carries significant potential for research outcomes significantly exceeding the original expectations:
1) A unifying framework for scalable control synthesis applicable to monotone and differentially positive systems.
2) A theory for learning based control with optimal dynamic performance and robustness.
3) Numerical validation techniques for feedback systems involving neural network components.
4) Methodology for data-driven modelling and control of energy networks for heating and cooling.
1) A unifying framework for scalable control synthesis applicable to monotone and differentially positive systems.
2) A theory for learning based control with optimal dynamic performance and robustness.
3) Numerical validation techniques for feedback systems involving neural network components.
4) Methodology for data-driven modelling and control of energy networks for heating and cooling.