Keywords: Plant-Wide Optimization
Abstract: To enhance safety, reliability, and productivity of industrial processes and to accelerate materials discovery, artificial intelligence (AI) and machine learning techniques have been widely used in process industries for many decades. In the current Industry 4.0 and digital era, AI propels advancement in a wide range of applications including image analytics, Natural language process (NLP), deep learning, reinforcement learning, hybrid modeling, and real-time analytics. As more AI successes are demonstrated in process industries, there is a growing misconception that AI is to replace human decision. The talk will highlight the importance of Responsible/Trustworthy AI and show industrial examples on how humans and AI must be working in the loop. One aspect is to understand how to incorporate AI methods to assist humans to accelerate discovery in research and to make well-informed decisions in manufacturing operations. The other aspect is to allow humans to incorporate engineering and science domain knowledge to make AI methods smarter. This talk will conclude by highlighting future research direction, workforce development need, and how academia, vendors, and industries must collaborate to unlock more value with AI.

Keywords: Machine Learning Assisted Modeling, Identification Methods, Data-Driven Optimization
Abstract: Causality is a fundamental relationship in the physical world, around which almost all activities of human life revolve. Causal inference refers to the process of determining whether an event or action caused a specific outcome, which involves the evaluation of cause-and-effect relationships in data. This paper presents a new approach to discover latent causal representations of crucial variables in easy-to-obtain data. The proposed method takes a form of trade-off between compression of input data and the causality between the learnt latent variables and critical variables, thereby removing the irrelevant information contained in input data and obtaining the decoupled, strongest causal factors. By introducing variational bounds and specific configurations, the optimisation objective is relaxed to a tractable problem. The approach compacts causal discovery and inference into one model, which is flexible to downstream tasks and parsimonious in the parameters. A case study on an exhaust-emission dataset shows that the proposed method improves the predictive performance over the baseline model, which is a variational information bottleneck model with the same hyperparameters.

Keywords: Optimal Control, Optimization under Uncertainties, Experiment Design
Abstract: Gas-lift is a strategy to enhance oil production from oil wells by reducing the hydrostatic pressure of the fluid. Efficient modeling of this process is a key point for the oil and gas industry to maximize oil output while minimizing gas consumption. This study introduces a novel hybrid model for the gas-lift process using the Universal Differential Equation (UDE) approach. Neural networks replace algebraic equations, trained based on physical laws. The UDE method eliminates dependence on unmeasured variables, enhancing accuracy. Tested with simulated data, the hybrid model outperforms traditional models, demonstrating effective prediction of state variables and efficient handling of algebraic variables. This approach holds promise for gas-lift process modeling, control, and optimization.

Keywords: Machine Learning Assisted Modeling, Adaptive and Learning Systems, Experiment Design
Abstract: The development of human-centric platforms that are able to combine computational resources and advanced analytics with human judgment and qualitative processing ability is a key driver of the Industry 5.0 movement. In this setting, humans are not only active in the loop but also play a key role in the decision-making process. In this work, we propose the use of Preferential Bayesian Optimization (PBO) for human-in-the-loop controller tuning. PBO relies on pairwise comparisons and preference feedback (A is better than B) to search for the optimal trade-off between different performance criteria from the user’s perspective. The advantages of PBO are demonstrated in a simulated Proportional Integral (PI) controller tuning example with real user feedback under a reduced number of experiments. The results show that PBO leads to a greater emphasis on closed-loop responses closer to the user’s desired behavior when compared with multi-objective alternatives, while being straightforward to implement from the user’s perspective.

Keywords: Fault Detection, Machine Learning Assisted Modeling
Abstract: Automated fault detection and diagnosis (FDD) in modern chemical process systems are essential for ensuring safe and reliable operation. Deep learning methods for FDD are gaining traction due to their high fault classification performances, but these methods lack interpretability which hinders their practical use. In this work, an interpretable neural network model for FDD is proposed which classifies the fault based on the fault propagation path. In our approach, fault propagation paths are embedded into the model as directed graphs in a graph neural network. This framework enforces connections between hidden layer nodes, giving them and their activations a physical and interpretable meaning. We evaluate the performance and interpretability of the proposed model on the benchmark Tennessee Eastman Process where it achieves a 92.9% classification accuracy on select fault datasets.

Keywords: Machine Learning Assisted Modeling, Estimation and Robust Estimation, Model Predictive Control
Abstract: This study presents a strategy for multi-step-ahead identification of robust machine learning (ML). Hence, we focus on the disparity between standard single-step prediction models and the requirement for multi-step forecasting, which is crucial for Model Predictive and Optimization schemes. This work explores how the proposed muti-step-ahead strategy can diminish the prediction uncertainty compared to the traditional single-step approach. The paper evaluates the multi-step identification with uncertainty assessment in different model architectures, including those based on recursive neural networks. A key aspect of the analysis is the application of these models to a polymerization reactor, a standard benchmark in algorithm evaluation. The results reveal that multi-step recursive models significantly reduce prediction uncertainty compared to single-step models, particularly when feedback mechanisms are involved. This study highlights the advantages of multi-step models and their potential benefits for control and optimization schemes.

Keywords: Model Predictive Control, Data-Driven Optimization, Machine Learning Assisted Modeling
Abstract: In this work, two model-based controllers were developed, one based on a nonlinear model-based controller (NMPC) using a population balance model (PBM) and another using a machine learning approach based on a neural network inverse model-based controller (NNIMC). The performance of the two model-based controllers was compared for different scenarios to obtain optimal temperature policies for controlling the mass yield and crystals' size for the unseeded batch cooling crystallization of paracetamol. The results show that both strategies are effective for crystallization control, presenting comparable results for the controlled variables in different scenarios. The controllers were also tested by applying random noise in the state variables. In these cases, the NNIMC presented advantages in having a lower computational cost for optimum control action calculations and less control effort regarding the manipulated variable's variation to reach values for the control variables at the end of the batch close to the NMPC and the setpoints.

Keywords: Fault Detection, Resilient, Safe, and Cyber-Secure Systems, Power and Energy Systems
Abstract: This work focuses on developing a neurofuzzy detector capable of identifying a cyber attack of false data injection into the outlet temperature sensor of a Fresnel-type solar field which has a PI+FF controller to control the refered temperature. A digital twin of the Fresnel plant and its controller are used for simulation purposes. The digital twin is situated in the domain of behavior and rules, as it contains a set of models, including a partial differential equation (PDE) model and a neurofuzzy model. Results from simulation are shown using three different scenarios: (1) without fault, (2) a ramp and threhold with negative injection and (3) the last scenario with positive injection. The presented fault data injection detector has solid performance with more than 97% detection accuracy and precision.

Keywords: Supply Chain and Enterprise Integration, Optimization under Uncertainties, Production Planning
Abstract: Multiproduct pipelines are crucial for delivering substantial quantities of refined oil products from major supply centers to clients within a nearby geographical area. Despite the significant infrastructure investment, the associated transportation costs are markedly lower than those incurred with traditional delivery trucks. However, the scheduling of these systems presents a formidable challenge, requiring meticulous planning of pumping runs well in advance to meet the anticipated demands of clients. In this work, we enhance an existing literature model of a multiproduct pipeline system by introducing uncertainty in the customer demand. The problem is then addressed via a two-stage stochastic formulation. The typical drawback with stochastic formulations is the high computational burden required. To address this challenge, we adapt the so-called Similarity Index decomposition, resulting in a 28-fold improvement in CPU time while achieving equivalent solutions compared to solving the full-space problem.

Keywords: Manufacturing Plant Control, Experiment Design, Power and Energy Systems
Abstract: Shrink tunnels are machines composed of an industrial oven and a conveyor belt; they are widely used in manufacturing applications for polymeric packaging. Manufacturing products are wrapped in a thin plastic film and inserted into the oven via the conveyor belt. The heat shrinks the plastic around the products, creating the pack. This paper presents a model-based temperature control architecture that tackles numerous goals: setpoint tracking in the presence of manual-automatic transitions, demanding disturbance rejection requirements, energy saving, and actuator limitations. The performances of the control architecture are experimentally validated on a workbench, highlighting its effectiveness in satisfying the specifications.

Keywords: Optimal Control, Optimization under Uncertainties, Experiment Design
Abstract: This work highlights the importance of dehydration in natural gas production, especially in offshore units, and tackles the challenges associated with adsorption processes. The main contribution is implementing digital solutions to monitor a Brazilian operational offshore natural gas dehydration unit. Bayesian inference and robust regression are utilized to determine the Remaining Useful Life (RUL) of the adsorbent in fixed beds. Furthermore, a mass balance in fixed beds provides valuable process insights, such as the adsorbed volume of water, a crucial variable for assessing the fixed bed's performance. Bayesian inference and the logistic function yielded the most accurate predictions for the end of the fixed bed's useful life. The proposed methodologies have been successfully integrated into a real-time monitoring dashboard at a Brazilian oil and gas plant.

Keywords: Power and Energy Systems
Abstract: In this work, we propose a simplified model for DSG parabolic trough collector solar plants consisting of a system coupled partial differential equations (PDEs) and ordinary differential equations (ODEs). Particularly, the PDEs represent the behavior of the two-phase gas-liquid flow in the collector field, which was obtained assuming a quasi-equilibrium on the mixture momentum balance of the classical transient homogeneous equilibrium model. The gas-liquid separator is given by ODEs obtained from the mass and energy balance laws. This formulation makes the model attractive for optimization and control applications since it is simpler than other approaches in the literature. Finally, in view of the obtained simulation results, the main uses and applications of the developed model are drawn describing a simulation example that proves how the closed-loop operation allows for obtaining higher production rates.

Keywords: Production Planning, Large-Scale and Networked Systems, Plant-Wide Optimization
Abstract: This paper focuses on the optimization of crude oil operations scheduling in a refinery that is supplied with crude oil by ship. One of the main challenges associated with the crude oil operations scheduling problem is the management of crude storage in tanks. Since storage capacity is limited and there are several types of crude oil depending on their composition, it is necessary to store mixtures of crude oil in tanks. This feature makes necessary the inclusion of nonlinear, non-convex constraints, which complicates the resolution of mathematical programming models. To address this problem, we have developed a mathematical programming model based on a continuous-time formulation using time slots, along with a strategy based on piecewise McCormick relaxation that allows us to efficiently handle the nonlinear constraints generated by blending crude oils in tanks.

Keywords: Machine Learning Assisted Modeling, Manufacturing Modeling for Management and Control, Estimation and Robust Estimation
Abstract: Deep learning models have been widely employed in various domains, yet they have certain limitations when it comes to industrial process applications. The two main challenges are their inability to effectively handle long-sequence predictions and the complexity of their internal structure, which makes it difficult to explain the output of the model. This work aims to build accurate and interpretable soft sensors for industrial processes. The Informer model is used to build accurate soft sensors due to its proficiency in long sequences. Additionally, an interpretable machine learning algorithm, SHapley Additive exPlanations (SHAP), is used to infer the global and local contributions of each feature to the predictions. The effectiveness of the proposed algorithms is validated on real industrial fluid catalytic cracker unit data, and the results show that the Informer model has higher accuracy and better long-sequence data prediction ability. Furthermore, the SHAP analysis enhances the model's utility by providing clear insights into the influence of individual features on the predictions, thereby increasing its transparency and trustworthiness in industrial settings.

Keywords: Machine Learning Assisted Modeling
Abstract: Real-time monitoring and accurate prediction of key variables are indispensable to ensure industrial production activities proceed as expected. With the increase in measurement data volume and the improvement of hardware computing power, the Transformer and its variants, due to their excellent capability in extracting global dependencies, are playing an increasingly important role among deep learning-based multidimensional time series prediction models. In addition, from the perspective of causality, cause variables contain parts of information in effect variables and can reduce the uncertainty of effect variables, which is beneficial for prediction. However, there has been relatively limited research on combining the Transformer and causal feature analysis. To fully use both advantages, this paper introduces the Causal-Transformer (CT) model, which utilizes semi-orthogonal projection to extract causal features from multiple input variables. A multi-head spatial-temporal causal attention mechanism is designed in the encoder block based on the classical Transformer model to simultaneously reduce feature dimensions and extract implicit causal features in both the temporal and spatial dimensions. The CT also utilizes the Granger causality analysis to select the causal teaching indicators of target variables to provide stable assistance by injecting explicit causality into the inputs of the decoder block. By leveraging more condensed and independent causal features, the CT possesses inherent advantages in predicting time series variables. Case study results show that the CT model outperforms the other models on the diesel refinery dataset, especially with a reduction of 46.0% and 30.4% in MSE towards the classic Transformer and informer in five-step prediction.

Keywords: Manufacturing Plant Control, Manufacturing Modeling for Management and Control
Abstract: Monitoring and predicting the key variables is significant in industrial processes. However, it is not effective in extracting temporal features of variables and predicting them accurately due to the high dimensionality and the long term of the input sequence. Therefore, this paper proposes an input-output driven cross-attention for the transformer network (IDCA-Former), which takes historical labels into account as a part of the input sequence. Then, cross-attention is conducted to compute the similarity between historical labels and original input data to capture more potential information. Moreover, sliding windows are designed by setting the input length of historical labels. The proposed IDCA-Former is applied for light naphtha prediction in the hydrocracking process. Extensive experiments show that IDCA-Former performs better in time series forecasting compared to other methods.

Keywords: Estimation and Robust Estimation, Wastewater Treatment
Abstract: Soft sensors experience an increased interest in recent years, as they can replace expensive hardware meters and the required embedded computing hardware has become cheap and powerful. Despite numerous academic results from laboratory or field tests, the widespread implementation of Softsensors has still not happened, due to a gap between laboratory tests and the real-world application. We address the problem of plant model mismatch and outline a possible modeling approach to solve the underlying problem of varying or uncertain hardware components.

Keywords: Machine Learning Assisted Modeling, Estimation and Robust Estimation
Abstract: Industrial soft sensing has gained widespread use in industrial processes due to its advantages in terms of low cost and easy maintenance. However, as industrial processes become increasingly complex, characterized by high dimensionality, coupling, and nonlinearity in the data, traditional data-driven soft sensing models often fall short of achieving the required level of accuracy. In this paper, a novel and enhanced variant of the Broad Learning System (BLS) called the Distributed Broad Seasonal Trend Learning System (DBSTLS) is proposed for the development of industrial soft sensing with improved accuracy. In the proposed DBSTLS, a distributed structure based on Seasonal-Trend Decomposition Procedure Based on LOESS (STL) is established. Through STL, dynamic process data can be mainly separated into two distinct components: the trend feature and the season feature. The distributed structure is built separately for the trend feature and the season feature. Subsequently, a fast learning strategy, based on BLS, is applied to both the distributed trend feature and the season feature. This integrated approach culminates in the development of the DBSTLS for industrial soft sensing. To validate the effectiveness of the proposed DBSTLS-based industrial soft sensing, the process data collected from the Pure Terephthalic Acid production process are used. Simulation results confirm that the DBSTLS-based industrial soft sensing outperforms other models in terms of accuracy.

Keywords: Fault Detection, Signal Processing, Estimation and Robust Estimation
Abstract: The predictive performance of soft sensors deteriorates over time which is called the performance change of a soft sensor. These changes occur due to differences between the current characteristics of the process or plant and the soft sensor model. The deviation is a type of plant-model mismatch (PMM). Initially, this mismatch may be acceptable. However, over time, the PMM can become so large that it affects the prediction quality of the soft sensor and may become unacceptable. This paper develops a new method to evaluate the impact of PMM on closed-loops with soft sensors. Using coprime factorisation and small-gain theory, a performance-change index is developed to characterise the PMM-induced performance degradation. Then, a performance-based online PMM detection method is proposed using this performance-change index. To validate the effectiveness of the proposed algorithm, we use a numerical example and a continuous stirred tank reactor (CSTR). It is shown that that the proposed index can detect the change of the PMM.

Keywords: Machine Learning Assisted Modeling, Model Predictive Control
Abstract: A novel big data-predictive control approach for nonlinear multi-timescale processes is presented in this paper. Multiple Dynamical Latent Variable Autoencoders (DLVAEs) are employed to approximate multi-timescale dynamics, utilizing timescale-based low-pass filtering and resampling of historical input-output data. The encoder in each DLVAE projects the nonlinear physical variable space onto a linear latent variable space, represented by a kernel space in behavioral system theory. During training, we not only impose kernel spaces and reconstruct data but also establish connections among latent variables from different DLVAEs at matching time-steps. Collectively, these multi-level latent variables span a wide prediction time horizon with limited (non-uniformly spaced) steps encompassing the current, near, and distant future. In online tracking control, we guide the latent variables from each DLVAE to their respective setpoints (derived from physical variable setpoints) while maintaining consistent physical variable values at matching time-steps, all within a linear framework.

Keywords: Machine Learning Assisted Modeling, Identification Methods, Experiment Design
Abstract: Bonnet polishing is a process that can achieve surface finishes down to sub-nanometer texture and form accuracies down to 5-10nm RMS, usually limited by metrology. While the polishing is conducted by computer numerical control (CNC) machines, the process remains imperfectly deterministic, requiring multiple iterations to converge on the desired surface quality. Key parameters are the axial force exerted by the bonnet tool on the workpiece, and the lateral components of frictional coupling. As direct real-time measurement of material removal is impractical, a force table equipped with loadcells has been used to estimate the forces at the tool-surface interface. In this work, a bespoke 3-axis force-table using six loadcells has been built and deployed on the horizontal work-piece table of a Zeeko IRP600 CNC polishing machine. Machine learning was then used to calibrate the force table and to account for any biases or cross-talk between axes.

Keywords: Machine Learning Assisted Modeling, Estimation and Robust Estimation
Abstract: For a non-linear system, sample-based state estimators have proven to be a better alternative to their derivative-based counterparts, due to their superior ability to obtain transformed probability densities. However, the computation relating to propagation of the numerous samples make these methods relatively expensive. In this work, a novel ML-based state estimation method is proposed to reduce computation time for a specific form of sample-based state estimator namely, constrained unscented transform based filters. The key feature of the proposed approach is that a trained neural network (NN) model is used as a map between initial posterior samples (or sigma points), manipulated inputs and new measurements to obtain filtered posterior sigma points. The efficacy of the proposed approach is demonstrated on a non-isothermal continuous stirred tank reactor and the Williams Otto reactor. It is shown that the proposed approach outperforms unscented transform-based state estimator in terms of computation time, while matching its performance.

Keywords: Machine Learning Assisted Modeling, Experiment Design, Identification Methods
Abstract: New products must be formulated rapidly to succeed in the global formulated product market; however, key product indicators (KPIs) can be complex, poorly understood functions of the chemical composition and processing history. Consequently, process scale-up must currently undergo expensive trial-and-error campaigns. To accelerate process flow diagram (PFD) optimization and knowledge discovery, this work proposes a novel digital framework to automatically quantify process mechanisms by integrating symbolic regression (SR) within model-based design of experiments (MBDoE). Each iteration, SR proposed a Pareto front of interpretable mechanistic expressions, and then MBDoE designs a new experiment to discriminate between them while automatically balancing the objective of PFD optimization. To investigate the framework’s performance, a new process model capable of simulating general formulated product synthesis was constructed to generate in-silico data for different case studies. The framework could effectively discover ground-truth process mechanisms within a few iterations, indicating its great potential within the general chemical industry for digital manufacturing and product innovation. Keywords: knowledge discovery, symbolic regression, model-based design of experiments, interpretable machine learning, process flow diagram optimization.

Keywords: Machine Learning Assisted Modeling, Optimal Control, Experiment Design
Abstract: Multiscale hybrid modelling of biosystems utilises advantageous aspects of several modelling approaches, from the physical interpretations of kinetic modelling to the power of a data-driven Artificial Neural Network (ANN). This study implements multiscale modelling to gain insight into the production of Trastuzumab (Herceptin) from Chinese Hamster Ovary (CHO) cells under challenging dynamics. A reduced metabolic network is subject to enzyme constraints with a Dynamic Metabolic Flux Analysis (ecDMFA) approach and integrated within a macro-scale hybrid kinetic model. The model can simulate fed-batch processes with optimized feed control, as well as providing insight into the control gained by alteration to the cell culture media. On the intracellular level, the influence from extracellular perturbations can be observed, in addition to giving an estimated production rate of unmeasured by-products. Overall, this model can be used as a reliable digital twin to estimate the underlying fed-batch process dynamics for future model predictive control and process optimisation.

Keywords: Machine Learning Assisted Modeling, Power and Energy Systems, Estimation and Robust Estimation
Abstract: Battery degradation poses a significant challenge for the usage of Lithium-ion batteries, making accurate capacity estimation crucial for efficient operation. Data-driven approaches hold promise for addressing this task, yet their complex structures often lead to overfitting and obscure the decision-making process. The objective of this work is to build a robust and interpretable model for capacity estimation. We propose the utilization of a robust decision tree-based ensemble model, extremely randomized trees (ERT), to accurately estimate battery capacity based on the features extracted from the partial charging curve. The random splits in the tree construction process enhance the model's generalization ability. Given that the combination of multiple decision trees reduces interpretability, we further employ SHAP to interpret the contributions of each feature to the ERT model's predictions. The effectiveness of the proposed method is validated on a large cycling dataset of Lithium-ion batteries.

Keywords: Environment and Agriculture, Optimal Control, Learning-Based Control
Abstract: This study proposes a Semi-centralized Multi-agent Reinforcement Learning (SCMARL) approach for irrigation scheduling in agricultural fields, which are characterized by spatial variability and therefore delineated into management zones. The SCMARL framework is hierarchical in nature, with a coordinator agent at the top level and local agents at the second/lower level. The coordinator agent makes daily ‘yes/no’ irrigation decisions based on field-wide observations from all the management zones, which are then communicated to local agents. These local agents are tasked with determining the optimal daily irrigation depths for specific management zones, utilizing both the coordinator agent’s decision and local observations. A comparison between the SCMARL method and a Fully Decentralized Multi-agent Reinforcement Learning approach is presented, highlighting the superior performance of the SCMARL approach in terms of water savings and improved irrigation water-use efficiency.

Keywords: Optimal Control, Wastewater Treatment, Environment and Agriculture
Abstract: Exploiting natural symbioses to enhance productivity of bioprocesses is an emerging trend. For optimizing such complex associations of microorganisms, a model of symbiotic interactions is vital. This challenging task has attracted much attention. Here, a reduced metabolic model describing a symbiotic interaction between bacteria E. coli, overproducing vitamin biotin (B7), and microalgae Chlorella is developed. The symbiosis involves B7 exchange, impacting lipid synthesis regulation in microalgae. Our model shows a trade-off between light availability and biotin production, leading to an optimization problem for lipid production. We numerically determine the optimal conditions, demonstrating the feasibility of this strategy to enhance microalgae cultivation.

Keywords: Optimal Control, Environment and Agriculture
Abstract: The industrial cultivation of microalgae has increased substantially over the past two decades. These microorganisms have the ability to adapt their photosynthetic pigments in response to the amount of light they experience. Herein, we investigate a dynamic model that describes pigment adaptation and its effect on microalgal productivity in a photobioreactor where light is shone onto the surface and attenuated as it traverses the culture medium. We consider two controls -- the light irradiance and the dilution rate of the photobioreactor under continuous operation and constant volume -- and analyze strategies for maximal production of microalgal biomass using Pontryagin's maximum principle. We also conduct a numerical investigation of turnpike properties in this context and discuss how self-shading within the culture could be exploited to increase productivity.

Keywords: Environment and Agriculture, Estimation and Robust Estimation, Machine Learning Assisted Modeling
Abstract: In this study, we introduce a triggered model reduction method for soil moisture and hydraulic parameter estimation. The approach employs cluster-based unsupervised learning to extract simplified models capturing essential dynamics of the complex original nonlinear system. To handle model mismatch over time, a sequential triggering method with an event trigger followed by a performance trigger is proposed for model identification. Further, an adaptive extended Kalman filter (EKF) that can take advantage of the adaptively reduced models is developed to estimate soil moisture and the associated hydraulic parameters. The performance of the proposed method is illustrated based on a large-scale agricultural field.

Keywords: Data-Driven Optimization, Machine Learning Assisted Modeling, Power and Energy Systems
Abstract: This article explores the use of the twin delayed deep deterministic policy gradient (TD3), a deep reinforcement learning algorithm, to reduce the cumulative greenhouse gas (GHG) emissions from the sailing trips of a hybrid-powered roll-on roll-off liquefied natural gas ferry. The objective of the algorithm is to optimally control the power load distribution between the ferry’s engines and battery to achieve a reduction in GHG emissions. Results from this study show that the TD3 agent achieved an average reduction in cumulative GHG emissions by 5% against actual operations for the sailing trips that were analyzed. The performance of the TD3 agent was compared to a rule-based energy management strategy (EMS) in which the ferry's battery was used to operate the ferry completely at low load operations and provide surplus power when the power demand was greater than the engine rating. The rule-based EMS failed to achieve GHG emissions reductions against actual operations thereby indicating the efficacy of the TD3 agent in achieving emissions reductions.

Keywords: Environment and Agriculture, Estimation and Robust Estimation, Signal Processing
Abstract: Developing a precise irrigation control strategy is essential for improving water use efficiency, and this requires accurate soil moisture information. However, certain challenges associated with state estimation must be addressed when dealing with large-scale fields. For instance, a vast farmland may be composed of different types of soil, making it challenging to obtain accurate parameters. Consequently, model mismatch becomes inevitable for agro-hydrological systems. In this study, we focus on addressing the issue of state estimation under such circumstance. A high dimensional nonlinear system is obtained by discretizing a 3D polar Richards equation that characterizes water movement dynamics. The proposed approach represents model mismatch as unknown inputs (UIs) relative to the state equations. To reduce computational complexity, a recursive expectation-maximization (EM) approach is modified from the existing batch EM algorithm to identify the UIs. The extended Kalman filter (EKF) is applied to calculate the posterior expectation of the states. Furthermore, an appropriate set of sensors is chosen to ensure complete observability of the system. The simulation results demonstrate the efficacy of the proposed estimation method.

Keywords: Estimation and Robust Estimation, Identification Methods, Machine Learning Assisted Modeling
Abstract: Chain-shuttling polymerization is widely used to synthesize specialized polymer materials with customized properties. The significance of modeling in chemical process simulation lies in accurately describing and analyzing complex chemical systems through mathematical and computational models, thereby enhancing the efficiency and reliability of process design. Due to limited and noisy measurements and the complex model structure, parameter identification for chain-shuttling polymerization has been a long-standing problem. To address this issue, in this work, we propose to first describe the dynamic process with a set of ordinary differential equations (ODEs) based on the method of moments. This method characterizes the dynamic variations of the average chain length. After that, we introduce Physics-Informed Neural Networks (PINNs) to estimate the unknown parameters in the ODEs. Since PINNs can incorporate the ODEs constraints during the training process, they can effectively integrate the polymerization mechanism with the observed process data, thereby reducing the amount of data needed for training. A comparative analysis between parameters estimated using PINNs and the ground truth values demonstrates high accuracy and efficiency, even with sparse and limited observations. This showcases the potential value of PINNs in chemical process identification.

Keywords: Estimation and Robust Estimation, Signal Processing
Abstract: This paper proposes a novel approach to tackle uncertainties in model parameters and noise statistics for state estimation. The proposed method leverages transfer learning to combine the strengths of the unbiased finite impulse response (UFIR) filter and the Kalman filter (KF), with UFIR serving as the source domain filter and KF as the target domain filter. To bolster the robustness of state estimation within the target domain, the proposed method transfers the predicted state probability density functions (pdfs) from UFIR and fine-tunes the error covariance of the KF filter to achieve seamless integration. Unlike conventional fusion techniques, this method avoids the need for UFIR's error covariance, thus mitigating its adverse impact on estimation accuracy. We demonstrate the competitiveness of this transfer state estimator in handling parameter uncertainties through moving target tracking, showing superior performance compared to existing fusion methods for state estimation.

Keywords: Estimation and Robust Estimation
Abstract: The implementation of an in-line control strategy for primary drying relies on the availability of the product temperature and sublimation front position. Such measurements are often inaccessible or difficult to obtain directly without interfering with the drying trajectory, thus motivating the design of in-line estimators. This article specifically addresses this issue, taking advantage of a global mass loss measurement and spatial characterization of vials along the chamber. The estimator proposed uses a least square algorithm and simplified models to re-calibrate the heat transfer coefficient related to different vial locations. Results allow comparing the quality of the estimations with regard to the product temperature and cycle time predictions under parametric disturbances and plant/model mismatch.

Keywords: Estimation and Robust Estimation, Identification Methods
Abstract: This paper presents the development and the validation of two macroscopic growth models with coordinated uptake of glucose and ammonium for the yeast Saccharomyces cerevisiae. The two models differ in the structure chosen for the reaction kinetics: one employs generalized kinetic laws, the other employs extended Haldane laws. The predictions of both models are in agreement with experimental data, however, a slightly better accuracy can be noticed for the generalized kinetics formalism. This study emphasizes that an accurate, robust model captures the biological phenomena shown by the experimental data regardless of its structure and of the formalism used for its development.

Keywords: Estimation and Robust Estimation, Identification Methods, Power and Energy Systems
Abstract: This paper revisits two parameter estimation techniques, based on the gradient descent (GD) and the least-square (LS) methods, to propose two novel estimators usable for estimating kinetic parameters in reaction systems. Typically, the activation energies of reactions appear in exponential terms of reaction rates, thereby resulting in non-separable nonlinearities while making the available techniques possibly worse. In this work, an overparameterized linear regression equation (LRE) is first derived, where reaction rates are computed from the vessel extents of reactions. On the one hand, we apply the dynamic regressor extension and mixing (DREM) procedure with a simple first-order differential operator to the LRE and then obtain the first estimator, called GD+D estimator. On the other hand, we adopt the technique, developed in Ortega et al. (2022), to design the second estimator, called LS+D estimator. Interestingly, the proposed estimators can estimate simultaneously all kinetic parameters and their convergence is ensured under the interval exciting condition that is more relaxed than the persistency of excitation. Simulations for the Van de Vusse reaction system illustrate the proposed estimators.

Keywords: Estimation and Robust Estimation, Machine Learning Assisted Modeling
Abstract: The ability to certify feasibility in process design and process operations is crucial in many applications. This includes quality-by-design in pharmaceutical manufacturing, where a key element is the characterization of the design space to better understand the links between materials, processes and products. Sampling-based approaches are versatile but they are cursed by dimensionality, which currently limits their application to problems in a few process variables only. We propose a decomposition approach that enables feasibility characterization for nominal settings of uncertain parameters in acyclic muti-unit processes by sampling. Our methodology leverages problem structure to decompose unit-wise, and deploys surrogate models to couple the resultant subproblems. We demonstrate it on a serial, batch chemical reactor network. In future research, we will extend this framework to cyclic multi-unit processes and the presence of parameter uncertainty.

Keywords: Fault Detection, Adaptive and Learning Systems, Estimation and Robust Estimation
Abstract: A large number of academic contributions on the topic of machine-learning-based control and monitoring have been published in the last five years. How is the process industry reacting to these developments? In this industrial forum, we will discuss the use of novel machine learning approaches to solve control-related challenges in various process industry sectors. The panel will discuss (1) to what extent are machine learning methods used in process industry today; (2) what are the gaps between machine-learning-based and model-based approaches to control solutions for industry implementation; (3) what types of challenges are more suitable for machine-learning-based solutions and what is the role of human supervision and intervention in such solutions and (4) what does it take to move a prototypic machine learning solution to production and to confirm its business case? The panelists are renowned experts in industrial process automation, chemicals, and mining sectors. The panelists are renowned experts in industrial process automation, chemicals, and mining sectors.

Panelists:

Lidia Auret (Stone Three and Stellenbosch University)

Elbert van der Bijl (Yokogawa Corporation of America)

Ivan Castillo (The Dow Chemical Company)

Alf Isaksson (ABB Corporate Research)

Keywords: Model Predictive Control, Estimation and Robust Estimation, Optimal Control
Abstract: Health-aware control (HAC) consists of computing the control action considering the degradation state of the plant components. The equipment degradation typically happens on a slower time scale, while the control and optimization of the economic performance on a much faster time scale. As such, HAC leads to a hierarchical control structure based on time-scale separation. We explore the different layers of this problem using a gas-lift network with choke valve degradation as an example. The main contributions of this work are to propose a new HAC system hierarchy and to show how regulatory layers can also reduce equipment degradation.

Keywords: Model Predictive Control, Optimal Control, Estimation and Robust Estimation
Abstract: The present work deals with the identification of a thermodynamic model for a one–stage spray drying tower. Motivated by the underlying time-scale separation for thermodynamic (slow) and dried powder specific (fast) states, in this first step the focus lies on the description of all relevant thermodynamic mechanisms which determine the resulting Particle Size Distribution (PSD) of the dried powder. Different to models discussed in the literature for similar drying processes, the model explicitly takes into account changes in the flow rates and densities due to evaporation, and proposes a simple monotonic dependency of the evaporation rate motivated by Monod kinetics from bioprocess modeling. The model parameters are partially taken from the literature and partially identified using a least squares procedure on the basis of experimental data. The experiments performed are parameter combinations of the spray tower configuration, which result in different particle size distributions and provide important information about the potential for future steps toward process monitoring and feedforward–feedback control to achieve desired PSDs.

Keywords: Environment and Agriculture, Plant-Wide Optimization, Power and Energy Systems
Abstract: Microalgae production in raceway photobioreactors is a very attractive process for biomass production because of its high sustainability. This paper presents a complete methodology for optimizing production in raceway photobioreactors. A first principles model has been developed that describes the dynamics and steady state of the system and has been used as the core of a real-time optimization to maximize the economic profit of the process, which handles the references of different variables of the system. For that, a steady-state real-time optimization approach complemented with PID controllers is presented in this work. The results obtained demonstrate the potential of steady-state real-time optimization for such systems, as well as the benefits of employing optimization techniques during process operation.

Keywords: Machine Learning Assisted Modeling, Estimation and Robust Estimation, Data-Driven Optimization
Abstract: The accurate prediction of the Gas-Oil Ratio (GOR) is crucial in optimizing oil production and ensuring the longevity of oil wells. This study evaluates the effectiveness of Neural Networks in predicting GOR for oil wells using gas lift production. We demonstrate that Neural Networks, particularly when trained with a mix of real and simulated data, show great promise for precise and rapid GOR predictions. To validate our approach, we analyzed data from real-world oil wells, employing Neural Networks for prediction. The results reveal that this method can predict GOR accurately and within a reasonable time frame, provided sufficient training data is available. Our findings offer valuable insights for oil well operators and engineers aiming to enhance their GOR prediction strategies

Keywords: Estimation and Robust Estimation
Abstract: Over the years, hundreds of applications have proved the effectiveness of constraint-based methods to validate the definition of metabolic networks, determine the robustness of metabolic models, and analyze the flow of metabolites through a network. However, stoichiometric models do not include information on flux capacity via enzymatic activity. Methods combining biological data from genome to metabolome have been developed to obtain improved flux predictions and constrain the range of possible flux distributions. Yet, these models still lack relevant information to design de novo metabolic pathways. Expressing the exogenous enzymes induces a cell burden due to competition for cell resources between the exogenous genes and the endogenous host ones, compromising the performance of the designed pathway. Thus, optimal selection of the expression strength of the pathway enzymes is still a challenge. Host-aware models have been developed to tackle cell burden in the context of designing increasingly complex synthetic genetic circuits in synthetic biology. This paper suggests a method to integrate host-aware gene expression models with constraint-based modeling to maximize the flux through an exogenous pathway by optimizing promoter and ribosome binding site strengths, crucial parameters that define the required transcription and translation strengths of the pathway enzymes’ genes. This study considers the formation of p-coumaric acid, shows promising results, and paves the way for further investigations.

Keywords: Model Predictive Control, Identification Methods
Abstract: This paper derives practical dynamic models for the glass industrial manufacturing process to be then included in model-based control solutions. In particular, the first section of the plant, that is, the glass melting furnace is investigated: silica sand and recovered glass are used as raw materials, and through methane and oxygen combustion melt glass is obtained which is then sent to the condition and final processing sections. Routine input-output data are employed to identify models of the furnace, including the loading machine, the fan, and the gas burners. Models of the various control valves are also identified, and finally, the parameters for the existing PI/PID controllers are estimated. A decentralized scheme comprised of SISO controllers and a centralized architecture with a model predictive controller (MPC) are designed and compared in a simulation scenario. The MPC solution guarantees higher performance with respect to the decentralized scheme by reaching a good trade-off between velocity of response and reduced oscillations.

Keywords: Fault Detection, Machine Learning Assisted Modeling
Abstract: This paper presents a hybrid model-based fault detection strategy for heating, ventilation, and air conditioning (HVAC) systems, focusing on air handling units (AHUs). Addressing the substantial energy inefficiencies in commercial buildings due to undetected HVAC faults, this research combines first-principles knowledge with data-driven techniques to enhance fault detection accuracy. First-principles based residuals (differences between expected and observed behaviors) are integrated with data (temperature measurements in different locations of AHU) to perform principal component analysis (PCA) (pre-processing step). Pre-processed data (principal component scores) are then utilized to perform clustering analysis using K-means and DBSCAN approaches. The proposed approach is tested against two common faults in AHUs and its performance is evaluated compared to a purely data-driven method. The results indicate that the hybrid method, which synergizes residual knowledge from first-principles models with data, significantly outperforms the purely data-driven approach. This is demonstrated through performance analysis using metrics like the adjusted rand index (ARI) and normalized mutual information (NMI). The research underscores the potential of the hybrid method in improving fault diagnosis of HVAC systems, helping to conserve energy by ensuring efficient and reliable operation.

Keywords: Machine Learning Assisted Modeling, Estimation and Robust Estimation
Abstract: A safe and reliable operation of electric arc furnaces (EAFs) is crucial for the mining and mineral industries. The lack of continuous measurements of critical process variables, such as the bath size of the molten phase, makes this operation challenging. Additionally, operator support decision tools able to predict the evolution of key process variables such as furnace sidewall temperatures would help to maintain safe operations. The present work proposes a data-driven (DD) modelling procedure to develop (1) a predictive model of the sidewall temperature and, (2) an online bath size estimator. Both sidewall temperature predictor and bath size estimator are based on long short-term memory (LSTM) networks. The preliminary developed models are validated on datasets collected on an industrial pilot-scale EAF and show good performance.

Keywords: Machine Learning Assisted Modeling, Identification Methods
Abstract: In the heat treatment processes, offline utilization of first-principles models is well-established. These models tend to be complex, computationally demanding, and rely heavily on empirical relations. The fidelity of these models degrades over time due to changes in the process resulting in plant-model mismatch, which is typically attributed to an incorrect constitutive relation of a physical mechanism in the model (i.e. fouling in the heat exchangers). In this paper, we propose two hybrid modeling approaches, namely Sparse Identification of Nonlinear Dynamics with Control and least square estimation, to learn the dynamics of the discrepancy between the measurement data and the simulation model. The hybrid modeling approach is implemented on a heat exchanger network (HEN) example and it is shown that the accuracy of the first principles dynamic model is improved.

Keywords: Identification Methods, Power and Energy Systems
Abstract: This paper examines the construction of a parameter-dependent voltage prediction model for a primary Zinc-air cell prototype, focusing on its response time when subjected to multiple step-wise discharge current levels. Laboratory tests have revealed that the dynamic response’s time constant varies with discharge current, a phenomenon not adequately addressed in previous analyses. The current research aims to contribute to the existing knowledge by employing a piecewise current profile during a single cell discharge and conducting an identification-type analysis of the relationships between the system’s time constants and other state and input variables. The findings presented in this paper hold significant potential for integration into Battery Management Systems and, in the long term, for addressing the inverse problem of State-of-Charge estimation.

Keywords: Manufacturing Modeling for Management and Control, Identification Methods
Abstract: We present a novel dynamic model of a flash clay calciner. The model consists of thermophysical properties, reaction kinetics and stoichiometry, transport, mass and energy balances, and algebraic constraints. This gives rise to a system of partial differential-algebraic equations (PDAE). Spatial discretization is performed to convert the PDAEs into a system of differential-algebraic equations (DAE). The model can be used, for example, to perform dynamic simulations with changing inputs, and process design and optimization. Moreover, it can be used to develop model-based control, which is relevant for flexible operation of a clay calcination plant for green cement production.

Keywords: Machine Learning Assisted Modeling, Identification Methods, Estimation and Robust Estimation
Abstract: Galerkin-projection and non-intrusive neural network reduced order models (ROMs) for a two-dimensional centrifugal pump model are presented and compared with respect to their suitability for control purposes. Singular value decomposition (SVD) is applied to reduce the dimensionality of the model and to extract a set of reduced basis functions. In Galerkin-projection, the temporal evolution of the reduced coefficients is expressed in terms of a small set of ordinary differential equations (in the order of ten). To improve the performance of this method, we fit the operators obtained by the projection step to the original data. On the other hand, the regression step of the coefficients is performed using a deep recurrent neural network (RNN) in the non-intrusive method. We compare the two methods with respect to the computational time required to build and evaluate each model, relative prediction error, long-term stability and sensitivity to initial conditions. Our results show that the projection-based ROM is slightly more accurate than the non-intrusive ROM, but it requires more time to be built. However, the non-intrusive ROM is more stable and less sensitive to initial condition variation.

Keywords: Fault Detection, Estimation and Robust Estimation
Abstract: Fault diagnosis and identification (FDI) is a critical aspect of process performance monitoring. In this work, statistical properties of decision variables of unconstrained Moving horizon estimation (MHE) are derived and further used for FDI. Once a fault is isolated, the fault magnitude refinement is carried out only for the isolated fault. Further, a hypothesis test is developed to terminate fault magnitude refinement when the fault magnitude saturates. When a sensor fault is isolated, the fault magnitude information is used for on-line compensation of measurements sent to the controller. The proposed approach is able to isolate and compensate for multiple single faults occurring sequentially in time and has embedded intelligence to carry out fault identification only when required. The efficacy of the proposed approach is demonstrated by simulating a non-isothermal CSTR system. Analysis of the simulation results underscore the effectiveness of the MHE-FDI scheme in correctly identifying faults in disturbance, actuator, and concentration measurements.

Keywords: Machine Learning Assisted Modeling, Identification Methods, Estimation and Robust Estimation
Abstract: Decision-making flexibility can be a key challenge in optimizing oil well production through a gas lift process. In this work, we introduce a multi-objective optimization strategy facilitated by deep neural networks (DNNs) as surrogate models to lessen the computational burden. Together with a likelihood test, we build a feasible operating region (FOR) using points from particle swarm optimization. Thus providing a tool for the refined process operation. We also subdivide the pareto region into constraint-compliant sub-regions, amplifying operational flexibility and identifying optimal settings. An optimality analysis is included to validate the results and assure the reliability of the surrogate-based optimization. This framework generates an operational map that can be instrumental for real-time process monitoring. Importantly, these computational tools can support the quality of real-time decisions in system operation by providing nuanced, data-driven insights into trade-offs and optimal conditions.

Keywords: Machine Learning Assisted Modeling, Model Predictive Control, Learning-Based Control
Abstract: Patients with chronic kidney disease suffer from renal anemia due to inadequate erythropoietin (EPO) secretion. Determining the optimal EPO dosage and frequency is complex and requires decision-support technologies. Model predictive control (MPC) is an effective decision-making technique that requires a prediction model of the controlled process. In this work, it was discovered that Physics-Informed Neural Networks for Control (PINNC), which integrates physiological model with data-driven methodology, were capable of predicting the patient hemoglobin level with good accuracy and computational efficiency. Based on this prediction model, we developed a zone MPC framework to optimize the dosing strategy. Simulation results show that the proposed control method can serve as an effective tool for determining the optimal EPO dosages for renal anemia patients.

Keywords: Model Predictive Control, Optimization under Uncertainties, Power and Energy Systems
Abstract: This paper presents a hierarchical MPC-based control framework for a real microgrid including solar panels and batteries, that considers the uncertainty from the point of view of faults and risks (F&R) mitigation. While fault management is applied during plant operation, risk management considers external factors that can change microgrid planning in the mediumlong term. Due to their different time-scales to apply, a two-layer control scheme is proposed using Model Predictive Control (MPC) at both levels. At the bottom layer, the fault-tolerant predictive controller optimizes the operation by manipulating inputs to follow microgrid setpoints. A reconfiguration strategy is implemented using structured residuals and stochastic thresholds. On the other hand, the upper layer develops an optimal mitigation strategy, also based on MPC, to reduce the effects of risks obtained from external information, i.e., unexpected changes in demands, maintenance costs, or deviations in generation. The decision variables of this layer are the selection of mitigation actions to be undertaken, which minimise a proposed multicriteria objective function. Different simulations have been carried out to show the goodness of this methodology in a F&R scenario from a stochastic point of view.

Keywords: Fault Detection, Signal Processing
Abstract: In this paper, we present a data-based monitoring approach designed for industrial data classification, aiming to minimize misclassifications of normal operations and to maximize the detection of anomalies and outliers. We make use of moving-horizon approaches and regression methods. Through evaluation of various algorithms on an industrial dataset, we showcase the effectiveness of the classification. As per our findings, effective detection can only be realized in conjunction of moving-horizon estimator with a regression model trained on historical measurements. The best prediction models consistently achieve accurate detection within the approved process tolerance, highlighting the efficacy of the proposed approach.

Keywords: Machine Learning Assisted Modeling
Abstract: For training the Physics-Informed Neural Networks (PINNs), the allocation of collocation points in the geometric domain plays a pivotal role in determining the model’s performance. We present a novel sampling method tailored for optimal point allocation in PINNs. The method involves an initial meshing of the domain, followed by a calculation of the sensitivity matrix relating the losses for each mesh element to local changes in the locations of the training points. Subsequently, based on the principles of A-optimal experimental design, the sampling probability is dynamically redistributed over the domain. In this way, areas of high sensitivity and corresponding physical significance receive further representation in the training data. Preliminary results illustrate the effectiveness of the proposed method when applied to the problem of developing flow between two parallel plates. This sensitivity-based sampling (SBS) is shown to increase the overall precision of PINNs since it can specifically capture sharp gradients in critical points within the geometric domain.

Keywords: Model Predictive Control
Abstract: Fast quantum control helps reduce the influence of unavoided disturbances and hence plays a vital role in practical quantum technology and chemical reactions. Instead of optimizing the terminal cost like standard optimal quantum control methods, this paper formulates the problem as a trajectory optimization problem, and implements the sequential quadratic programming algorithm to search for short control fields. The core idea is to minimize the cumulative intermediate error to incentivize early achievement of the designed gate. The numerical result on the Toffoli gate demonstrates the effectiveness of the proposed method.