DYN Season's Greetings 2021

Season's Greetings

Dear co-workers, project partners, colleagues, friends, and former members of the dyn and pas groups!

We hope that this message reaches you in good health and good spirits, despite the risks and restrictions of another Corona winter! Sometimes one would wish that every decision maker had to take (and pass!) a compulsory short course in fundamentals of feedback control, just some basics: What is runaway, why it is better to use indicators which react fast rather than indicators that provide information weeks later when the problems have accumulated, and that when controlling a system the dynamics of which you do not fully understand, you better use small probing moves and react fast to the outcomes than doing bang-bang control...

Looking back on 2021, we believe that our groups managed the fully online semesters very well with good outcomes, thanks to the enormous efforts of all group members. The only issue for which we did not come up with solutions that were received favourably by all was how to offer online exams assuring fairness, but also flexibility.

This semester we switched back to classroom teaching, at least until now, which the students seem to enjoy very much. We have been using all channels of communication from the start of the semester, live lectures, live broadcasting via Zoom, and providing the recordings for asynchronous and multiple use. We believe in the maturity of the students to use these opportunities in the way that fits them best. After the experience of the online semesters, the interaction in the lectures so far was livelier than ever, and certainly our teaching materials also have improved.

With the new lecture “Machine Learning Methods for Engineers” the PAS group introduced a new course entirely dedicated to machine learning which has been very well received by students.

Regarding research, the DFG Transregio InPROMPT will end in mid-2022 after 12.5 successful years. The achievements are collected in an impressive book. The KEEN project continues to provide a stimulating collective learning experience about the potential and limitations of machine learning when applied to real-world examples and data, both for the partners from industry and academia, and has already triggered interesting methodological work.

In the new EU project Circular Foam that started in October 2021, our groups will contribute to the development of large-scale chemical recycling solutions for polyurethane foam from refrigerators and construction waste, in particular to system-wide modelling, simulation and optimization.

The pas group successfully applied for funding in the framework of a new Priority Program (SPP) of DFG, dealing with machine learning in chemical engineering. The project will explore new ideas to achieve safe reinforcement learning for the optimal startup and operation of complex chemical processes.

Just today, we got the news that the successor project of SIMPLIFY, SIMPLI-DEMO will receive funding from the EU for 4 years, providing a jump-start for the PAS group in the application domain of the production of particles and highly viscous materials.

You find a lot more information on research, people, publications etc. on our Seasons’s Greetings web page.

Unfortunately, due to the pandemic, we were not able to plan a live event in 2021, we hope for an opportunity where we can meet many of you in person in 2022!

We would like to thank all group members, project partners and colleagues for the pleasant and rewarding collaborations in 2021, and we wish you enjoyable holidays and a successful and happy year 2022 in good health!

Sebastian Engell and Sergio Lucia

EU-Project Circular Foam Started

Closing the materials cycle for rigid polyurethane foams: This is the ambitious goal of the new pan-European "CIRCULAR FOAM" project. The EU-funded lighthouse project brings together 22 partners from 9 countries from industry, academia and society. Within four years, the consortium will jointly establish a complete circular value chain for raw materials for rigid polyurethane foams used as insulation material in refrigerators and the construction industry. Once implemented across Europe, the system could help to save 1 million tons of waste, 2.9 million tons of CO2 emissions and 150 million euros in incineration costs annually, starting in 2040.

The CIRCULAR FOAM project aims at bringing multiple improvements to the existing material cycle and building a new sustainable circular e cosystem for rigid polyurethane foam. Besides the development of two novel chemical recycling routes for end-of-life materials, waste collection systems and dismantling and sorting solutions and logistic solutions will be set-up and demonstrated. The different elements will be combined into an optimized systemic solution, based on integrated system modelling and simulation.

In CIRCULAR FOAM, we will develop an integrated system-wide modelling, simulation and optimization framework into which models of the different elements, collection and separation of waste, logistics, chemical processing and separation and feed-in into industrial sites can be embedded. We will also contribute to the design of the chemical processing and separation steps, in particular with respect to robustness and flexibility in terms of the throughput and the composition of the feed streams.

More information: www.circular-foam.eu

InPROMPT Project

After more than 12 years, the SFB Transreagio 63 ”Integrated Chemical Processes in Liquid Multiphase Systems”. will end in 2022. The Transregio currently consists of 14 research projects with researchers from TU Berlin, TU Dortmund, OVGU Magdeburg, HTW Berlin, TU Darmstadt, KIT, MPI Magdeburg and HS Anhalt. The goal is understand and develop processes based on reactive multiphase system. All levels of the process, from the molecular level to the design and operation of the processes in miniplants are considered.

The results of the research projects within InPROMPT will be published in a book “Integrated chemical processes in liquid multiphase systems”, by De Gruyter. The book covers the fundamentals of the thermodynamics of multiphase systems, kinetic modeling and modelling of mass transfer in multiphase systems. It is discussed how three types of phase systems, thermomorphic multiphase systems, microemulsion systems and pickering emulsions can be characterized and which aspects have to be considered for process design. Tools for process systems engineering including modelling and simulation, process optimization and model-based monitoring are presented and in the final chapter, the integration of the tools into a comprehensive design methodology is presented.

The dyn group contributed sections on surrogate modeling of thermodynamic equilibria, optimization under uncertainty in process development, iterative real-time optimization applied to the hydroformylation of 1-dodecene in a TMS-system on miniplant scale, and on the integrated model-based process design methodology.

A final colloquium of the Transregio is planned for March 31 – April 1, 2022 at DECHEMA, Frankfurt. We would like to thank all colleagues from the Transrigio SFB for the pleasant and productive collaboration!

Image of the miniplant with highlighted process units.

M.Jokiel et al., "Miniplant-Scale Evaluation of a Semibatch-Continuous Tandem Reactor System for the Hydroformylation of Long-Chain Olefins", Industrial & Engineering Chemistry Research, 58(7), pp. 2471-2480, 2019.

SIMPLIFY - Sonication and microwave processing of material feedstock

The SIMPLIFY project is an EU Innovation Action aiming at the electrification of the chemical industry. Until April 2023, a consortium of 11 European organizations including the dyn group will focus on the development of flexible electrified continuous processes for three industrial case studies.

One of the three case studies is the production of polyurethanes which are used as rheology modifiers in water-based paints. These rheology modifiers are conventionally produced in batches of several cubic meters and directly formulated with water within the same vessel. This production process takes several hours per batch. Within the SIMPLIFY project, the transition to a continuous production of these paint thickeners by reactive extrusion on a twin-screw extruder is investigated, which offers numerous advantages. This transition significantly reduces the time and costs associated with cleaning and enables a fully electrified production using renewable energy sources.

The dyn group is contributing to the project by our expertise in the fields of process modelling, automation and control. A twin-screw extruder model has been extended to account for the reactive extrusion process and is combined with a new model of the chemical system based on the experimental work carried out by the project partners. Model based optimization is used for the decision making process of the optimal screw design and the operating region. The experimental validation of the results and process demonstration is carried out atFraunhofer ICT in Pfinztal on an 18mm Leistritz Maxx extruder as shown in Figure 1. In collaboration with Fraunhofer ICT a process automation and measurement concept for this extruder was implemented and validated. This concept includes the control and recording of all process values centrally with a soft PLC using OPC-UA. With this automation system it was possible to demonstrate a stable production of 4kg/h of product over a duration of 8 hours this year. Furthermore, the technical requirements are now met to implement advanced process control in the coming year. Recent theoretical investigations of the application of MAWQA to reactive extrusion processes showed that this control method is well suitable for the reactive extrusion process and offers major economic and ecologic benefits.

Figure 1: Picture of the reactive extrusion setup with temperature, pressure and viscosity measurements (left) and the ultrasound sonotrode (right) realized at Fraunhofer ICT. © Fraunhofer ICT.

Publications

M. Cegla and S. Engell, 2021, "Reliable Modelling of Twin-screw Extruders by Integrating the Backflow Cell methodology into a Mechanistic Model". Computer Aided Chemical Engineering 48, 175-180.

M. Cegla and S. Engell, 2021, "Application of Model Predictive Control to the reactive extrusion of e-Caprolactone in a twin-screw extruder". IFAC-PapersOnLine 54 (3), 225–230.

Cegla,M. and Engell,S. 2022, "Application of Real-Time Optimization with Modifier Adaptation to the Reactive Extrusion of Hydrophobically Modified Ethoxylated Urethanes". In Review.

KEEN-TP7: Self-optimizing plants - Dynamic gray-box modeling of fermentation processes

KEEN connects 20 partners, industrial end users, solution providers and scientific institutions with the objective to introduce artificial intelligence (AI) technologies and methods in the process industry and to evaluate and realize their technical, economical, and social potential.

The goal of the work on self-optimizing plants in KEEN is to improve process operations using machine learning (ML) models within advisory systems or in closed-loop control. In this context, a it is crucial to ensure the reliability of the model predictions, especially for applications in feedback control.

So-called gray-box models are a promising direction to increase reliability and interpretability of ML-models. Here, mechanistic model parts are combined with ML-models. Among others, we consider models, where the dynamics are described by mechanistic relations, but some embedded variables are represented by ML-models. An example could be a (bio-) chemical reaction system, where the underlying reactions are known, but the dependency of a reaction rate on temperature or concentration is unknown and modelled using ML-models. Because the ML-model is here embedded into the dynamic equations, it is challenging to choose an appropriate ML-model structure along with its parameters. This is because the dynamic behavior of the system has to be simulated, to evaluate the fitness of one set of model structure and parameters.

We approach this problem in several steps by first estimating what values the ML-models should output to accurately describe the experimental data. This is done by using continuous piece-wise linear functions of time instead of the complex ML-model. This model depends only on the values at the knot points. These knot point values are a lot easier to fit than e.g. a neural net. In the next step, the ML-submodels can be trained using any ML-toolbox, using the values of the piece-wise linear functions as training data. Finally, a full parameter estimation is performed on the basis of a dynamic simulation of the complete model. This procedure is illustrated in Fig. 1.

Figure 1: Steps of decomposed parameter estimation problem for parameterizing dynamic gray-box models with embedded machine learning models

We have investigated different algorithmic options and obtained promising results for the fermentation of a sporulating bacterium [1, 2]. The methodology is used on real-world experimental data obtained in pilot-plant scale at Evonik in Hanau.

Figure 2: Picture of the investigated pilot plant. Image by Air Liquide

In one of the investigated use-cases, the dyn group collaborates with Air Liquide on the optimal control of a pre-reforming process. Air Liquide provides data from a pilot plant located at their “Innovation Campus” in Frankfurt where ML-based control algorithms can also be tested. In an adiabatic pre-reformer, higher hydrocarbons are catalytically converted into H2/Syngas. The pre-reforming reactor is installed in front of the main reformer, improving the overall process in terms of efficiency, catalyst lifetime and feedstock flexibility. The goal in KEEN is to increase the energy and material efficiency and to reduce the green-house emissions by improved process control. By the dyn group, data-based (ML) models of the pilot plant are developed based on plant data of the controlled plant supplied by Air Liquide. The developed models will be used in a model predictive controller. Additionally, a mechanistic model for the process has been designed to be used as a benchmark. The solutions will be applied to the pilot plant in the coming years

Publications

[1] Winz, J., & Engell, S. (2021). A methodology for gray-box modeling of nonlinear ODE systems. Submitted to ESCAPE32.

[2] Winz, J., & Engell, S. (2021). A methodology for reliable dynamic nonlinear gray-box modeling. Submitted to DYCOPS2022.

OptiProd project

Make-and-pack process plants make up a significant slice of the total of production plants in the process industry. Their competitiveness, productivity, and resource efficiency largely depend on the quality of the production scheduling. Today, production scheduling is often done manually, which is tedious, expensive, and inefficient. For those reasons, there is an increasing industrial drive towards automated and optimal scheduling solutions.

A schematic representation of a two-stage formulation and filling plant that is investigated as a case study in the OptiProd.NRW project, is shown in Figure 1. It consists of a formulation stage, which is decoupled from the downstream filling stage by a set of buffer tanks that are connected through the transfer panel. In addition, the logistics of raw materials and final products as well as the shift schedules of the operators must be considered.

The difficulty of such scheduling problems is the tremendous degree of detail that must be considered to guarantee the feasibility of the plans on the shop floor level. To capture the complex interactions precisely, we make use of the commercial simulation software from INOSIM. It provides a powerful and flexible formalism to capture the features of industrial production processes and can be configured via an easy-to-use interface that enables the design of models without deep knowledge on simulation. The INOSIM software can simulate simple scheduling rules, but an optimization of schedules is not offered.

Figure 1: Schematic representation of the industrial formulation plant.

During the OptiProd.NRW project, an optimization framework [1] is developed to generate schedules automatically. The framework is based upon customizable Evolutionary Algorithms that propose schedules which are evaluated by the INOSIM simulation model and which are iteratively improved based on the simulation results.

The simulation-optimization approach under development will generate schedules that are validated by a very detailed simulation model and optimize the operation of the plant. In 2022, a testing phase with the industrial partner is planned. INOSIM intends to commercialize the project results thereafter.

Publications

[1] Klanke, C., Bleidorn, D., Koslowski, C., Sonntag, C., Engell, S., 2021a. Simulation-based scheduling of a Large-scale Industrial Formulation Plant Using a heuristics-assisted Genetic Algorithm, in: GECCO’21: Proceedings of the Genetic and Evolutionary Computation Conference Companion, Association for Computing Machinery. pp. 1587–1595.

HyPro Project

Motivation Blast furnace-based steel-making is the dominant route for the worldwide production of steel which accounts for 70% of the final product. An industrial blast furnace is a huge gas-solid reactor that typically produces around 4 million tons of liquid iron per year with an annual total CO2 emission of more than 7 million tons. The stable, economically optimal, and environmental-friendly operation of blast furnaces is a challenge due to the complexity of the multi-phase and multi-scale physical and chemical phenomena, the lack of direct measurements of key inner variables, and the occurrence of a wide range of unknown disturbances. Blast furnaces are operated in a semi-automated manner and the quality of the control depends on the skills and dedication of the operators. The steel industry is strongly interested in better support for the operators or full automation to achieve a more energy-efficient and stable operation of blast furnaces.

Goal

A main objective of the HYPRO project is to develop new control strategies that achieve an improved energetic efficiency. In collaboration with partners such as VDEh-Betriebsforschungsinstitut GmbH (BFI) and thyssenkrupp Steel Europe AG, we developed a hybrid dynamic model-based control scheme for achieving the desired operational objectives of a blast furnace.

Outcome

Within the hybrid model, we integrated a first-principles-based model of medium complexity with data-based dynamic neural net model in order to combine the advantages of these two types of models. The hybrid dynamic model provides insights into the blast furnace operation status in terms of thermal status, safety, productivity, and efficiency, by predicting quantities at the furnace boundaries such as the molten iron and slag quality indices and the off-gas analysis parameters (temperature (TG), pressure drop (DP), and efficiency factor (η)). Free run simulation results of the multi-step ahead predictions of the hot metal silicon content ([Si]) and slag basicity (SB) are shown in Fig. 2. Using this model, an optimizing model predictive controller (MPC) was developed for controlling the hot metal silicon content and slag basicity at their desired set-points, subject to operational constraints. This controller regulates the gas phase variables to counteract the process disturbances that are caused by variations in the solid feed. Low values of [Si] and SB lead to improved energy efficiency of the blast furnace process. The simulation results of the control scheme are shown in Fig. 3.

Fig. 2. Multi-step ahead prediction of silicon content ([Si]) and slag basicity (SB) by the hybrid model. Fig. 3. Simulation results of the MPC scheme.

Publications

Azadi, P., Klock, R. and Engell, S., "Model Predictive Control of Molten Iron and Slag Quality Indices in a Large-scale Ironmaking Blast Furnace using a Hybrid Dynamic Model", submitted to IFAC MMM2022.

Azadi, P., Winz, J., Leo, E., Klock, R., and Engell, S., "A hybrid dynamic model for the prediction of molten iron and slag quality indices of a large-scale blast furnace", 2022, Computers & Chemical Engineering, 156, 107573.

Azadi, P., Klock, R., and Engell, S., "Efficient Utilization of Active Carbon in a Blast Furnace through a Black-Box Model-Based Optimizing Control Scheme", IFAC-PapersOnLine, 2021, 54(3), 128-133.

Azadi, P., Minaabad, S. A., Bartusch, H., Klock, R., and Engell, S., "Nonlinear Prediction Model of Blast Furnace Operation Status", In Computer Aided Chemical Engineering (Vol. 48, pp. 217-222), 2020.

Left Joined

dyn Members 2021

We all would like to thank our partners, colleagues, students and alumni for their support and the fruitful collaboration all throughout 2021. We wish you a bright, happy and successful year 2021!

Farewell to our colleagues

Fabian Schweers

Fabian Schweers started his research career with the dyn group in 2015. He worked in the EU project CONSENS. Since November 2021, Fabian is employed at BP Gelsenkirchen.

Afaq Ahmad

Afaq did his BSc in electrical and electronic engineering, focusing on automation and control. Later, he completed his master's in automation & robotics in Dortmund. During his master's studies, he focused on process automation. He then started his research career with the dyn group in May 2016 as a Research Associate, where he focused on Real-Time Optimization. He worked in DACH, DAAD-PPP, CoPro, and KEEN projects. Since November 2021, Afaq is employed at the BASF SE in large capital projects.

Jesus David Hernandez Ortiz

Jesus joined the dyn group in early 2017 as a Marie-Curie fellow. His research focused on the modeling and optimization of the electric arc furnace steelmaking process. As a Marie-Curie Early Stage Researcher, Jesus spent most of his time at Acciai Speciali Terni in Italy. Since August 2020, he is employed at BASF SE as E&I engineer.

New dyn Members

Marion Fiebiger

Marion started working as Prof. Engell´s administrative Assistant in April 2021. Getting to know the administrative work for the various research projects of the dyn and PAS group and managing the finances for the chair, she was able to get a position at the Department of Educational Sciences and Psychology of the TU Dortmund. We will miss her and she will keep in good memory the collegial work atmosphere.

Engelbert Pasieka

Engelbert Pasieka studied Chemical Engineering at the TU Dortmund completing both the Bachelor and the Master program. He conducted his master thesis with the title “Metaheuristic Optimal Batch Production Scheduling with Direct and Heuristics-based Encodings” in cooperation with the INOSIM Software GmbH. He joined the group in August 2021 and is currently involved in the OptiProd.NRW Project.

Latest news People Events Publications Master/Bachelor Theses

New funded DFG Project

A new project entitled „Safe Reinforcement Learning for Start-up and Operation of Chemical Processes“ has been awarded by the DFG this year. The project is part of the Priority Program 2331 with the title “Machine Learning in Chemical Engineering. Knowledge Meets Data: Interpretability, Extrapolation, Reliability, Trust”. Our project will explore how promising approaches from the field of reinforcement learning can be combined with ideas from robust predictive control to achieve algorithms that can be achieve a safe operation. The algorithms will be demonstrated for the start-up and operation of a complex distillation column, together with our project partners at TU Berlin (Prof. Repke)

Machine learning methods for engineers

During the summer semester of 2021, we created a new course called “Machine learning methods for engineers”. This lecture is the first one at the Department that is completely focused on machine learning. It covers important basics on probability and optimization as well as efficient implementations and current limitations of machine learning. The course was very well received by the students and will be continued as an elective at the master level.

PAS Members

Lukas Lüken and Moritz Heinlein joined the newly founded Laboratory for Process Automation Systems (PAS) at TU Dortmund this year.

Lukas Lüken

Lukas Lüken received his M.Sc. in Automation Engineering from RWTH Aachen University in January 2021. His master thesis considered the nonlinear model predictive control (NMPC) of a large-scale air separation unit using model reduction techniques and surrogate modeling with neural networks. He started working as a research associate at PAS in February 2021. His research interests are in the conjunction of control engineering, optimization and machine learning. His main focus is on directly incorporating optimization problems such as MPC into end-to-end learning frameworks for efficient and robust learning and control in complex environments. He sincerely thanks his colleagues at DYN and PAS for the warm welcome and looks forward to great learning opportunities

Moritz Heinlein

Moritz Heinlein studied Chemical Engineering in Dortmund from 2015 to 2021 completing both the Bachelor and the Master program. His Bachelor thesis is titled “Theoretische Betrachtung kinetischer Modelle für die Methanchlorierung und Pyrolyse”. He completed his master thesis titled „Comparison of Robust Subspace Predictive Control Methods for non-linear Systems” at the PAS chair. In November 2021, he joined the PAS group as a research associate. Currently, his research focusses on reducing the number of branches in the scenario tree for nonlinear multi-stage MPC by exploiting certain system properties.

Workshop on Robust Model Predictive Control at Uni Freiburg

DFG Workshop Robust Model Predictive Control (2021). The Laboratory of Process Automation Systems (PAS) is meeting the Systems Control and Optimization Laboratory at Universität Freiburg.

Felix Fiedler, Benjamin Karg, Lukas Lüken and Sergio Lucia from the Laboratory of Process Automation Systems visit the Systems Control and Optimization (Syscop) Laboratory (Prof. Moritz Diehl) at University of Freiburg for a Workshop on Robust Model Predictive Control.
The Workshop is part of our ongoing collaborations with Syscop in the context of the DFG project Robust MPC with high-dimensional uncertainty. During the three day event (21.-23.06.2021) all participants had the opportunity to present their current research in the field of Robust Model Predictive Control.

do-mpc developer conference at TU Dortmund

Prof. Dr.-Ing. Sergio Lucia during the presentation.

The first do-mpc developer conference was hosted from 13.09.-15.09.2021 at TU Dortmund. Over the course of three days the Laboratory of Process Automation Systems hosted multiple online and offline events for developers, users and supporters do-mpc. The conference was a great success and we are already looking forward to the next iteration.

pas @ Process Control Conference 2021

M.Sc. Benjamin Karg during the presentation.

From June 1st to 4th, Benjamin Karg represened the pas group at the 23rd International Conference on Process Control in Bratislava, Slovakia. The conference was held in a fully virtual format. pas contributed the following paper to the conference:

Benjamin Karg and Sergio Lucia, "Reinforced approximate robust nonlinear model predictive control".

pas @ ECC 2021

Fekix Fiedler represented the pas group at the European Control Conference from June 29th to July 2nd. Originally the conference was planned to take place in Rotterdam, Netherlands, but it took place in a virtual form due to the Covid-19 pandemic. He presented the paper:

Felix Fiedler and Sergio Lucia, "On the relationship between data-enabled predictive control and subspace predictive control".

Publications

Journal Articles 2021

Karg, B., Alamo, T., Lucia, S.:

Probabilistic performance validation of deep learning-based robust NMPC controllers

International Journal of Robust and Nonlinear Control, vol. 31, no. 18, pp. 8855-8876, Full Paper

Solving nonlinear model predictive control problems in real time is still an important challenge despite of recent advances in computing hardware, optimization algorithms and tailored implementations. This challenge is even greater when uncertainty is present due to disturbances, unknown parameters or measurement and estimation errors. To enable the application of advanced control schemes to fast systems and on low-cost embedded hardware, we propose to approximate a robust nonlinear model controller using deep learning and to verify its quality using probabilistic validation techniques. We propose a probabilistic validation technique based on finite families, combined with the idea of generalized maximum and constraint backoff to enable statistically valid conclusions related to general performance indicators. The potential of the proposed approach is demonstrated with simulation results of an uncertain nonlinear system.

Karg, B., Lucia, S.:

Approximate moving horizon estimation and robust nonlinear model predictive control via deep learning

Computers & Chemical Engineering, vol. 148, pp. 107266, Full Paper

Karg, B., Lucia, S.:

Model Predictive Control for the Internet of Things

Recent Advances in Model Predictive Control: Theory, Algorithms, and Applications, pp. 165-189, Full Paper

In this chapter, we argue that model predictive control (MPC) can be a very powerful technique to mitigate some of the challenges that arise when designing and deploying control algorithms in the context of the internet of things. The development of new low-power communication technologies and the widespread availability of sensing and computing capabilities, which are a characteristic of the internet of things, enables the consideration of a large amount of interconnection and feedback loops. However, this also introduces important challenges such as the very limited communication capabilities of low-power wide area networks or the limited computational resources of low-cost micro-controllers. We argue that, as a predictive control scheme, MPC is a powerful technique to deal with limited communication capabilities and can be naturally extended to the context of distributed control, for cases where all the sensing information cannot be centralized. We also present an approach to approximate the solution of MPC problems using deep neural networks, rendering the embedded implementation of complex controllers possible even on very limited hardware. The ideas are illustrated with an example of distributed model predictive control of a smart building.

Conference Articles 2021

Fiedler, F., Lucia, S.:

On the relationship between data-enabled predictive control and subspace predictive control

ECC 2021: European Control Conference

Data-enabled predictive control (DeePC) is a recently proposed approach that combines system identification, estimation and control in a single optimization problem, for which only recorded input/output data of the examined system is required. The same premise holds for the subspace predictive control (SPC) method in which a multi-step prediction model is identified from the same data as required for DeePC. This model is then used to formulate a similar optimal control problem. In this work we investigate the relationship between DeePC and SPC. Our primary contribution is to show that SPC is equivalent to DeePC in the deterministic case. We also show the equivalence of both methods in a special case for the non-deterministic formulation. We investigate the advantages and shortcomings of DeePC as opposed to SPC with and without measurement noise and illustrate them with a simulation example.

Karg, B., Lucia, S.:

Reinforced approximate robust nonlinear model predictive control

PC 2021: 23rd International Conference on Process Control, doi.org/10.1109/PC52310.2021.9447448

Model predictive control (MPC) has established itself as the standard advanced process control method. However, solving the resulting optimization problems in real-time can be challenging, especially when uncertainty is explicitly considered in a robust nonlinear predictive control approach. An increasingly popular alternative to avoid the online solution of the resulting optimization problems is to approximate their solution using neural networks. The networks are trained using many solutions of the MPC problem for different system states and therefore this approach is often called imitation learning. Controllers obtained via imitation learning have important drawbacks, since it is difficult to learn behaviors that are not well represented in the data and they must be redesigned from scratch when the control task changes. In this work, we show that these two drawbacks can be mitigated by combining imitation learning and concepts from reinforcement learning. The central idea is to use imitation learning as a very good initialization of a control policy that is iteratively updated using reinforcement learning, taking advantage of the fact that an explicit and differentiable expression of the approximate MPC controller is available. The efficacy of the combination of the two learning paradigms is highlighted via simulations of a semi-batch industrial polymerization reactor.

Theses

Master Theses 2021

Rohit Panindra:

Efficient Data Sampling for Learning-based Controllers

supervised by: Benjamin Karg

Model predictive control (MPC) is a popular optimisation-based control approach which allows direct consideration of multivariate, nonlinear systems and arbitrary constraints. But solving the underlying optimal control problem can be prohibitive, especially when the considered system needs a very fast sampling time or when the computational requirements for solving the optimization problem cannot be met. Various methods to reduce the computational complexity have been proposed, amongst which deep neural networks (DNN) were established as promising candidates due to their ability to learn complex functions based on data. The quality and amount of data necessary for obtaining the desired approximation quality is crucial for the application of DNN as imitators of MPC. Since generating data means solving many complex optimization problems, the goal of this work is to reduce the amount of data necessary to obtain a desired performance of the learningbased controller w.r.t. to commonly used approaches in literature. In this work, random sampling, quasi-random sampling, sampling from closed-loop trajectories, and active learning are the methods investigated to infer the most efficient method. Quasi-random sampling involves (scrambled) Halton and (randomised) Sobol sequences, and the batch Bayesian optimisation via local penalization (BBO-LP) method is adopted for active learning. A modification for the BBO-LP method is presented to overcome the problem of sampling infeasible states. The methods are analysed for a 2-dimensional nonlinear CSTR system with one control input and a 14-dimensional spring-mass system with 4 control inputs.

Moritz Heinlein:

Comparison of Robust Subspace Predictive Control Methods for non-linear Systems

supervised by: Felix Fiedler

Subspace predictive control (SPC) is a linear predictive control approach, that creates a linear multi-step ahead prediction model through projections of the input/output data of a system. It is thus easy to implement and can even be explicitly solved under certain assumptions. Due to the linear nature of the model, the applicability to nonlinear systems is restricted. However, related linear predictive control algorithms like the recently proposed data enabled predictive control showed good results in nonlinear case studies. In this work, the performance of SPC is compared to other linear predictive control algorithms by the means of case studies with nonlinear systems. In these it achieves partly similar performance to nonlinear model predictive control with the full system model. The main contribution in this work are two new approaches to further robustify the performance of SPC in terms of closed loop convergence and constraint satisfaction. Regularized SPC uses a ridge regression in the parameter estimation step to better condition the parameters. Better convergence rates can be achieved, as presented in the application to a nonlinear Lotka-Volterra model. Multicluster SPC uses within-cluster-regression to sort the input/output data after their proximity and linear prediction accuracy into clusters. This way the dynamics of different linearizations of the nonlinear system should be captured. These multiple models are implemented similar to multi-stage-MPC. Although no stability guarantees can be given, this leads to better constraint satisfactions and convergence rates, when the number of clusters can be chosen high, as shown by the means of the same Lotka-Volterra model.

Journal Articles Conference Papers Conference Presentations Master Theses Bachelor Theses

Journal Articles 2021

Azadi, P., Winz, J., Leo, E., Klock, R., Engell, S.:

A hybrid dynamic model for the prediction of molten iron and slag quality indices of a large-scale blast furnace

Computers & Chemical Engineering, vol. 156, pp. 107573, Full Paper

The stable, economically optimal, and environmental-friendly operation of blast furnaces is still a challenge. Blast furnaces consume huge amounts of energy and are among the biggest sources of CO2 in the metal industry. The operation of industrial blast furnaces is challenging because of their sheer size, multi-phase and multi-scale physics and chemistry, slow dynamics with response times of 8 hours and more, and the lack of direct measurements of most of the important inner variables. Model-based schemes are prime candidates for providing the missing information and improving the operation. However, only recently, such schemes have been applied successfully, and there is still a lot of room for improvements. The spatial extension, the lack of precise mechanistic knowledge about the chemical and physical phenomena, and the presence of unmeasured disturbances make the application of first-principle models to process operations extremely challenging. In this work, a hybrid dynamic model is developed for the prediction of the hot metal silicon content and the slag basicity in the blast furnace process. These two variables are the key indicators of the internal process conditions, and the ultimate goal of our work is to control them by a model-based scheme. The core relationships between the process variables are imposed by a first-principles-based steady-state model, and a parallel data-based model represents the process dynamics and compensates for the deficiencies of the mechanistic model. Validation results for real plant measurements of a world-scale blast furnace show that the hybrid model is more accurate than the rigorous model and a stand-alone data-based model in long-term predictions of the dynamic behavior of the process.

Klanke, C., Yfantis, V., Corominas, F., Engell, S.:

Short-term scheduling of make-and-pack processes in the consumer goods industry using discrete-time and precedence-based MILP models

Computers & Chemical Engineering, vol. 154, pp. 107453, Full Paper

This work deals with the short-term scheduling of a two-stage continuous make-and-pack process with finite intermediate buffer and sequence-dependent changeovers from the consumer goods industry. In the present coupled layout of the plant under consideration, the two stages, product formulation and packaging, are directly coupled, i.e. the products of the formulation stage go directly to their dedicated unit in the packaging stage. As for different products either the formulation or the packaging stage can be the bottleneck due to a stage and product dependent processing rate. A gain in productivity can be obtained if the two stages are decoupled by a buffer so that the formulation lines and the packaging lines can both run at full capacity. To evaluate the benefit of introducing a buffer between the stages, a rigorous discrete-time mixed-integer linear programming (MILP) model was developed. As the results of the discrete-time model were unsatisfactory with respect to total plant downtime due to changeovers and idle times, a two-step solution strategy including a second immediate precedence-based MILP model was developed. As the problem is intractable for the planning horizons of interest, an order-decomposition strategy for both models that is enhanced by several heuristics, was incorporated in the solution strategy. It is demonstrated that the redesign of the production plant yields significant productivity improvements that can be realized using the proposed scheduling approach. The computational results on several real cases show that the increased modeling and development effort of the two-step solution strategy pays off in terms of solution quality and computation times.

Leo, E., Engell, S.:

Condition-based maintenance optimization via stochastic programming with endogenous uncertainty

Computers & Chemical Engineering, vol. 156, pp. 107550, Full Paper

In this work we address the challenge of integrating production planning and maintenance optimization for a process plant. We consider uncertain predictions of the equipment degradation by adopting a stochastic programming formulation with decision-dependent uncertainty. The probability of the uncertain parameters, in this work the remaining useful time of the plant, depends on the operating conditions of the plant which is modeled by embedding a prognosis model, the Cox model, into the optimization problem. A separation of the variables is suggested to decompose the MINLP formulation via two different primal decomposition algorithms. We provide computational results and compare the performance of the proposed decompositions with the global solver BARON enhanced with a custom branching priority strategy.

Leo, E., Dalle Ave, G., Harjunkoski, I., Engell, S.:

Stochastic short-term integrated electricity procurement and production scheduling for a large consumer

Computers & Chemical Engineering, vol. 145, pp. 107191, Full Paper

This paper addresses the problem faced by large electricity consumers to simultaneously determine the optimal day-ahead electricity procurement and the optimal energy-aware production schedule. The inherent uncertainty of the problem, due to the bidding process in the day-ahead market, is dealt with by means of the stochastic programming modeling framework. In particular, a two-stage problem is formulated with the aim of establishing the optimal bidding strategy and the optimal production schedule hedging against price uncertainty. The optimal integrated solution is defined to minimize the overall cost and to control the risk of high cost scenarios due to uncertain price peaks. The stochastic model is solved with a scenario-decomposition approach. Extensive numerical experiments have been carried out to assess the performance of the proposed decision approach. The results collected when considering an industrial relevant case-study show the superiority of the proposed methodology in comparison with a deterministic approach.

Gottu Mukkula, A.R., Mateáš, M., Fikar, M., Paulen, R.:

Robust multi-stage model-based design of optimal experiments for nonlinear estimation

Computers & Chemical Engineering, vol. 155, pp. 107499, Full Paper

We study approaches to the robust model-based design of experiments in the context of maximum-likelihood estimation. These approaches provide robustification of model-based methodologies for the design of optimal experiments by accounting for the effect of the parametric uncertainty. We study the problem of robust optimal design of experiments in the framework of nonlinear least-squares parameter estimation using linearized confidence regions. We investigate several well-known robustification frameworks in this respect and propose a novel methodology based on multi-stage robust optimization. The proposed methodology aims at problems, where the experiments are designed sequentially with a possibility of re-estimation in-between the experiments. The multi-stage formalism aids in identifying experiments that are better conducted in the early phase of experimentation, where parameter knowledge is poor. We demonstrate the findings and effectiveness of the proposed methodology using four case studies of varying complexity.

Gottu Mukkula, A.R., Engell, S.:

Handling measurement delay in iterative real-time optimization methods

Processes, vol. 9, no. 10, pp. 1800, Full Paper

This paper is concerned with the real-time optimization (RTO) of chemical plants, i.e., the optimization of the steady-state operating points during operation, based on inaccurate models. Specifically, modifier adaptation is employed to cope with the plant-model mismatch, which corrects the plant model and the constraint functions by bias and gradient correction terms that are computed from measured variables at the steady-states of the plant. This implies that the sampling time of the iterative RTO scheme is lower-bounded by the time to reach a new steady-state after the previously computed inputs were applied. If analytical process measurements (PAT technology) are used to obtain the steady-state responses, time delays occur due to the measurement delay of the PAT device and due to the transportation delay if the samples are transported to the instrument via pipes. This situation is quite common because the PAT devices can often only be installed at a certain distance from the measurement location. The presence of these time delays slows down the iterative real-time optimization, as the time from the application of a new set of inputs to receiving the steady-state information increases further. In this paper, a proactive perturbation scheme is proposed to efficiently utilize the idle time by intelligently scheduling the process inputs taking into account the time delays to obtain the steady-state process measurements. The performance of the proposed proactive perturbation scheme is demonstrated for two examples, the Williams–Otto reactor benchmark and a lithiation process. The simulation results show that the proposed proactive perturbation scheme can speed up the convergence to the true plant optimum significantly.

Rahimi-Adli, K., Leo, E., Beisheim, B., Engell, S.:

Optimisation of the operation of an industrial power plant under steam demand uncertainty

Processes, vol. 14, no. 21, pp. 7213, Full Paper

The operation of on-site power plants in the chemical industry is typically determined by the steam demand of the production plants. This demand is uncertain due to deviations from the production plan and fluctuations in the operation of the plants. The steam demand uncertainty can result in an inefficient operation of the power plant due to a surplus or deficiency of steam that is needed to balance the steam network. In this contribution, it is proposed to use two-stage stochastic programming on a moving horizon to cope with the uncertainty. In each iteration of the moving horizon scheme, the model parameters are updated according to the new information acquired from the plants and the optimisation is re-executed. Hedging against steam demand uncertainty results in a reduction of the fuel consumption and a more economic generation of electric power, which can result in significant savings in the operating cost of the power plant. Moreover, unplanned load reductions due to lack of steam can be avoided. The application of the new approach is demonstrated for the on-site power plant of INEOS in Köln, and significant savings are reported in exemplary simulations.

Janus, T., Engell, S.:

Iterative Process Design with Surrogate-Assisted Global Flowsheet Optimization

Chemie Ingenieur Technik, vol. 93, no. 12, pp. 2019-2028, Full Paper

Flowsheet optimization is an important part of process design where commercial process simulators are widely used, due to their extensive library of models and ease of use. However, the application of a framework for global flowsheet optimization upon them is computationally expensive. Based on machine learning methods, we added mechanisms for rejection and generation of candidates to a framework for global flowsheet optimization. These extensions halve the amount of time needed for optimization such that the integration of the framework in a workflow for iterative process design becomes applicable.

Winz, J., Nentwich, C., Engell, S.:

Surrogate Modeling of Thermodynamic Equilibria: Applications, Sampling and Optimization

Chemie Ingenieur Technik, vol. 93, no. 12, pp. 1898-1906, Full Paper

Models based on first principles are an effective way to model chemical processes. The quality of these depends critically on the accurate description of thermodynamic equilibria. This is provided by modern thermodynamic models, e.g., PC-SAFT, but they come with a high computational cost, which makes process optimization challenging. This can be addressed by using surrogate models to approximate the equilibrium calculations. A high accuracy of the surrogate model can be achieved by carefully choosing the points at which the original function is evaluated to create data for the training of the surrogate models, called sampling. Using a case study, different approaches to sampling are discussed and evaluated with a focus on new approaches to adaptive sampling.

Subramanian, S., Lucia, S., Paulen, R., Engell, S.:

Tube-enhanced multi-stage model predictive control for flexible robust control of constrained linear systems with additive and parametric uncertainties

International Journal of Robust and Nonlinear Control, vol. 31, no. 9, pp. 4458-4487, Full Paper

The trade-off between optimality and complexity has been one of the most important challenges in the field of robust model predictive control (MPC). To address the challenge, we propose a flexible robust MPC scheme by synergizing the multi-stage and tube-based MPC approaches. The key idea is to exploit the nonconservatism of the multi-stage MPC and the simplicity of the tube-based MPC. The proposed scheme provides two options for the user to determine the trade-off depending on the application: the choice of the robust horizon and the classification of the uncertainties. Beyond the robust horizon, the branching of the scenario-tree employed in multi-stage MPC is avoided with the help of tubes. The growth of the problem size with respect to the number of uncertainties is reduced by handling small uncertainties via an invariant tube that can be computed offline. This results in linear growth of the problem size beyond the robust horizon and no growth of the problem size concerning small magnitude uncertainties. The proposed approach helps to achieve a desired trade-off between optimality and complexity compared to existing robust MPC approaches. We show that the proposed approach is robustly asymptotically stable. Its advantages are demonstrated for a CSTR example.

Subramanian, S., Abdelsalam, Y., Lucia, S., Engell, S.:

Robust Tube-Enhanced Multi-Stage NMPC with Stability Guarantees

IEEE Control Systems Letters, vol. 6, pp. 1112-1117, Full Paper

We propose a robust Nonlinear Model Predictive Control (NMPC) scheme that provides an improved trade-off between optimality and complexity when compared to other available strategies. Two controllers are employed in the proposed framework: A multi-stage primary controller that optimizes a given objective in the presence of uncertainties with tightened constraints and a multi-stage ancillary controller that tracks the predicted tree of state and input trajectories of the primary controller. The primary controller optimizes the original objective by considering different realizations of the most significant uncertainties in the predictions. The ancillary controller provides robustness against other uncertainties by tracking the predicted tree of state and input trajectories of the primary controller. We establish sufficient conditions for closed-loop stability. The advantages of the scheme are demonstrated for a continuous stirred tank reactor (CSTR) example.

Hernández, J.D., Onofri, L., Engell, S.:

Numerical Estimation of the Geometry and Temperature of An Alternating Current Steelmaking Electric Arc

Steel Research International, vol. 92, no. 3, 2000386, Full Paper

A channel arc model (CAM) that predicts the temperature and the geometry of an electric arc from its voltage and impedance set-points is presented. The core of the model is a nonlinear programming (NLP) formulation that minimizes the entropy production of a plasma column, the physical and electrical properties of which satisfy the Elenbaas–Heller equation and Ohm's law. The radiative properties of the plasma are approximated utilizing the net emission coefficient (NEC), and the NLP is solved using a global numerical solver. The effects of the voltage and impedance set-points on the length of the electric arc are studied, and a linear formula that estimates the length of the arc in terms of its electrical set-points is deducted. The length of various electric arcs is measured in a fully operative electric arc furnace (EAF), and the results are used to validate the proposed models. The errors in the predictions of the models are 0.5 and 0.4 cm. In comparison, the existing empirical and Bowman formulae estimate the length of the experimental arcs with errors of 2.1 and 2.6 cm. A simplified formula to estimate the temperature of an electric arc in terms of its electrical set-points is also presented.

Conference Papers 2021

Klanke, C., Bleidorn, D., Koslowski, C., Sonntag, C., Engell, S.:

Simulation-based scheduling of a large-scale industrial formulation plant using a heuristics-assisted genetic algorithm

GECCO '21: Proceedings of the Genetic and Evolutionary Computation Conference Companion, pp. 1587-1595, Full Paper

Research has brought forth several promising approaches, e.g. Mixed-integer linear programming or Constraint Programming to represent industrial batch production plants and to optimize their production schedules. But still, after decades of research in the field, scheduling is done (semi)-manually in industry in almost all cases. Main reasons for this besides the intrinsic combinatorial complexity are that model development and maintenance require expert knowledge. This work tackles these challenges by using a simulation-based optimization approach in which a Genetic Algorithm provides high-level encodings of schedules and an industrial-strength discrete-event simulator that provides a detailed model of the plant and is used as a fitness evaluator. To enable this approach to solve problems of industrial complexity, the scheduling heuristics are used to reduce the search space to a reasonable size that still contains the important degrees of freedom so that still optimal or at least high-quality solution are obtained. The case study that is considered here is an industrial two-stage formulation plant which is modeled in detail, down to the level of shiftmodels of the operators and mass-balances from source to sink, thus ensuring that the computed schedules are directly applicable at the real-world plant.

Gerlich, S., Arab, H., Buchholz, M., Engell, S.:

Experimental application of individual column state and parameter estimation in SMB processes to an amino acid separation