Dear Systems & Control Division,
Please find the details of the winter 2022 webinar series and zoom links will be circulated by Prof. Nicolas Hudon.
Monday January 31st, 2022 3PM EST
- Soumya Ranjan Sahoo, University of Alberta
Title: Model Reduction and Remote Sensing for Precision Irrigation Applications
Abstract: Freshwater scarcity is one of the biggest global risks due to population growth, climate change and the increase in pollution. Since a large percentage of freshwater withdrawals are for agricultural irrigation, increasing water-use efficiency in irrigation becomes extremely important. Closed-loop irrigation is one of the precision irrigation techniques which has the potential to improve water-use efficiency than the traditional approaches. Due to the large-scale features of agricultural fields and significant uncertainty, there are significant challenges associated with closed-loop irrigation including sensor placement, data assimilation, and controller design. To achieve the objective, the following tasks are addressed: (1) the structure-preserving model reduction techniques are developed to handle the large-dimension of agricultural field model; (2) a systematic approach has been developed to find the minimum number and best location of the sensors using observability and degree of observability analysis; (3) the framework for the state estimation of the largescale field using the adaptive moving horizon estimation (MHE) is developed; (4) an optimization-based closed-loop scheduler for large agricultural fields is developed to provide optimal irrigation time and amount; (5) the algorithm for surface soil moisture estimation using the thermal and optical remote sensing method is developed using LSTM model. Some of the developed methods are applied to a real-agricultural farm located in Lethbridge, Canada.
- Judith Ogwuru, Queen’s Unviersity
Title: Decentralized Proportional-Integral Extremum Seeking Control for Heating, Ventilation and Air Conditioning (HVAC) Systems
Abstract: This work considers the application of decentralized extremum seeking control to heating, ventilation and air conditioning (HVAC) systems in residential, commercial, and industrial buildings. The HVAC system considered comprises two rooftop units that each provide cool air to two zones. The compressor, fan and expansion valve of each rooftop unit are controlled by three inner loop proportional-integral (PI) controllers to meet specified control requirements. The objective is to determine the optimal supply air temperature setpoint for each rooftop unit that minimizes the overall power consumption of the units. In addition, each setpoint must satisfy the control objectives of the three inner loop PI controllers. To tackle this problem, a decentralized proportional-integral extremum seeking control technique that avoids the need for communication between the units is employed. A simulation result is included to show the effectiveness of this technique.
Monday February 28th, 2022 3PM EST
- Oscar Palma-Flores, University of Waterloo
Title: Integration of Design and NMPC-based control for chemical processes under uncertainty: An MPCC-based framework
Abstract: The use of nonlinear predictive control (NMPC) for the integration of design and control remains as an open area of research. The incorporation of NMPC into the framework for design and control leads to a bilevel formulation. Conventionally, a bilevel problem formulation is solved with the implementation of systematic iterative methodologies. In this work, a classical KKT transformation strategy is implemented to transform the original bilevel problem for design and NMP-based control into a single level dynamic optimization problem. This single level formulation is known as a mathematical program with complementarity constraints (MPCC). The use of an MPCC-based formulation allows to fully incorporate the NMPC’s necessary conditions for optimality into the design problem as a set of algebraic constraints. Then, the MPCC-based formulation is solved as a conventional NLP, i.e., the use of decomposition or simplification methodologies is avoided for the solution of the bilevel problem. This strategy is illustrated in two case studies to demonstrate the application of MPCCs for design and NMPC-based control under process disturbances and uncertainty.
- Anjana Puliyanda, University of Alberta
Title: Machine learning-based monitoring of complex reactive systems
Abstract: Processing of complex feedstocks to produce value-added chemicals is industrially important. The lack of a priori knowledge of the innumerable species and the reaction pathways governing their conversion, has posed challenges to monitoring these processes. Although, data-driven models have been used, they are limited by their lack of interpretability and an end-to-end modeling framework. On the other hand, systems where the mechanistic knowledge is obtained from first-principles simulations, face computational challenges when deployed for process design. This talk focuses on two aspects: (i) developing inferential machine learning models to enhance the interpretability of data-driven models, and (ii) developing predictive machine learning models to limit the computational cost of first-principles simulations, in modeling chemical systems. The first aspect of developing inferential machine learning models focuses on the identification of species, reaction pathways, and kinetic parameter estimation from spectroscopic data of the system, with application to the visbreaking of bitumen. Spectroscopic curve resolution methods that are structure-preserving, interpretable, and jointly parse data from multiple sensors, to extract latent features for species identification have been presented with an increasing degree of sophistication as follows:(i) selfmodeling multivariate curve resolution (SMCR), (ii) joint non-negative matrix factorization (JNMF) as a data fusion analogue of SMCR, and (iii) joint non-negative tensor factorization (JNTF) as a structure-preserving higher order analogue of JNMF. Next, Bayesian structure learning among the extracted spectral features has been used to causally infer plausible reaction pathways that have been validated by automated mapping to domain knowledge. Finally, the latent factorization and causal inference models have been used as an engine to interpret the modes identified by training hidden semi-Markov models on spectra, for the realtime monitoring of reaction mechanism dynamics with changing operating conditions. Projections of spectroscopic data onto the temporal mode of data collection via latent factorization, interpreted as concentrations, are used to develop constrained kinetic models using chemical neural ODEs. The second aspect of training predictive machine learning models focuses on not only reducing the computational cost of the ab initio molecular dynamics (AIMD) simulations of chemical systems, but also the cost of developing such models. This has been demonstrated with application to the transglycosylation of cellobiose, to assess whether the solvent molecules reorganize significantly in going from the reactant to the product configurations. A self-supervised 3D convolutional neural network autoencoder is trained to extract features from the reactant and product simulation trajectories, the probability distributions across the difference between which is used to assess if the solvent reorganization is significant. Similarity between the reactant configuration features of other chemical systems with those extracted from that of the cellobiose systems is then used as a basis to inform the extent of reorganization in the product profiles, without having to explicitly run AIMD simulations for the same.
Monday March 28th, 2022 3PM EST
- Christian Euler, University of Toronto
Title: Far from equilibrium kinetics in metabolic analysis
Abstract: Metabolism, like all open systems, is fundamentally constrained by flows of mass, energy, and information. Mass balance has formed the basis of systems-level metabolic analysis in both FBA and kinetic modelling methodologies, and thermodynamic consistency is now readily considered at scale in metabolic modelling. However, informational constraints in metabolic networks remain unexplored, despite the fact that understanding such constraints in biological regulatory systems has proven useful. Sensitivity analysis is used here to show that thermodynamic constraints and saturation kinetics have the combined effect of causing information loss at metabolic reactions operating far from equilibrium. However, inhibition by the substrates of such reactions maximally transfers information, thus providing a mechanism by which information loss can be prevented. Genome-wide analysis of inhibitors in E coli is then used to show that they are indeed significantly more likely to be substrates of reactions which are constrained to operate far from equilibrium, suggesting that maximizing information transfer is a design goal of metabolic regulation. The elements required for the informational dynamics described here – a driving force flux, saturation kinetics, and thermodynamically constrained reactions – are all features of hypothesized chemical reaction networks which preceded metabolism, so these results may be applicable to the origin and early evolution of metabolism.
- Cecilia Pereira-Rodrigues, Universite Laval
Title: Modeling and Simulation of a Tablet Pan Coating Process for Control System Design
Abstract: Process modeling and simulation are key to control system design and evaluation, especially for processes such as pan coating of pharmaceutical tablets that present multiple inter-dependent variables. Models can describe phenomena from either a macro or a micro and their complexity, informative and predictive abilities can differ considerably. Previous efforts to model pan coating systems from a macro level have greatly centered towards technology transfer and scale-up, exploring what if scenarios (prediction), new process design and troubleshooting. Micro level models mostly aimed at studying film formation and water adsorption within the tablet cores. However useful, these models have not been often used towards the development of control strategies for coating processes. The objective of the current work is to develop a simplified dynamic process model for batch pan coating processes, based on mass and energy balances, to support control system design and validation. The model consists of a system of ODE’s and algebraic equations. Model parameters are estimated using pilot plant batch runs and simulated trajectories of the main process variables are compared to the experimental data.
Monday April 25th, 2022 3PM EST
- Emilie Thibault, Ecole Polytechnique de Montreal
Title: Operating Regime-Based Data Processing Framework for Process Troubleshooting and Decision-Making
Abstract: For many years now, pulp and paper mills have faced ongoing challenges including highly competitive and even declining markets, new environmental regulations – all while managing complex operations and daily operating issues with abnormal/faulty situations (deterioration of process energy performance, reduction in product quality…). As a result, mills are continuously seeking to improve operations to maintain their competitive position. At the same time, process data available in today’s mills are extensive, and contain the knowledge that can help mills address this.
Archived process signals from measurement sensors are impacted by many types and sources of signal disturbances that are not representative of the process operation, and an important goal of data cleaning is to distinguish between the true process measurement and these disturbances. The latter can be due to signal conversion and transmission, physical vibration of process elements, signal drift due to fouling or corrosion, incorrect calibration of measurement instruments, sensor quality, sensor installation and maintenance, abnormal events, sampling rate, etc., all of which can “contaminate” the signal.
Signal processing aims to identify and remove disturbances that affect a signal making it less representative of the true process operation. To improve decision-making based on information from plant sensors, process data need to be cleaned, and then interpreted. The data treatment methodology developed in this study was enabled by a characterization of process signal noise according to various criteria such as the noise frequency, amplitude, skewness, and kurtosis. This characterization of the process signal noise can help for their automatic processing to “clean” the data.
Once the outliers and the noise are removed from the process data, steady-state operating regimes are identified. In a process, there are some operating regimes that are fundamentally inherent to the process operations (it is not possible to avoid them). On the other hand, certain other operating regimes may be eliminated if they are less profitable. Once operating regimes have been identified, it is possible to assess key parameters for each operating regime including for example energy consumption, efficiency, performance, cost, yield, product quality. When the different operating regimes are not identified, the standard deviations are much wider, hence when focusing on fault detection, equipment failure, monitoring, or troubleshooting, it is easier to miss something.The last step in this data processing framework is data reconciliation to make sure process data fits with mass and energy balances.
- Daniel McClement and Nathan Lawrence, University of British Columbia
Title: Reinforcement Learning Applications to Process Control
Abstract: Deep reinforcement learning (RL) is an optimization-driven framework for producing control strategies for general dynamical systems without explicit reliance on process models. Although promising results have been reported in simulation, there are many the challenges in implementing deep RL algorithms physical systems. In this talk, we motivate the use of deep RL through the lens of proportional-integral-derivative (PID) control. Our simple strategy is shown to be effective on a physical two-tank system. However, online training can be time-consuming, making it difficult to scale the approach to numerous systems. On the other hand, meta-learning is a branch of machine learning which trains neural network models to synthesize a wide variety of data in order to rapidly solve new problems. In process control, many systems have similar and well-understood dynamics, which suggests it is feasible to create a generalizable controller through meta-learning. We formulate a meta reinforcement learning (meta-RL) control strategy that takes advantage of known, offline information for training, such as the system gain or time constant, yet efficiently controls novel systems in a completely model-free fashion. Our meta-RL agent has a recurrent structure that accumulates “context” for its current dynamics through a hidden state variable. This end-to-end architecture enables the agent to automatically adapt PID parameters to changes in the process dynamics. Moreover, the same agent can be deployed on systems with previously unseen nonlinearities and timescales.