Conference Agenda

This paper focuses on the control and dynamics of an oil refinery process on an intermediate level - the flows, masses and compositions of and between units within the refining operation. It aims to elucidate optimal strategies for the routing of streams during upset events imposed on the process. A general flowsheet simulation technique including tunable controllers for flows, compositions, levels and reaction extents is incorporated in a Linear Programming model.

In this work the flow/composition nonlinearity is dealt with by updating stream compositions iteratively in a series of LP solutions until convergence. A standard node represents a mixed receiving tank, with exit streams which can be split, converted and separated. These nodes can be inter-connected arbitrarily in the flowsheet, even allowing recycling.

Many processing industries use intermediate storages as materials advance through the various units. This presents a type of supply-chain problem where the phasing of transitions at various points can be optimized to improve process efficiency and avoid unit shut-downs. In the present work, it was sought to take advantage of the robust features of Linear Programming, which efficiently handles the large number of variables involved. The model includes:

• Setpoint control of flows, storage masses & compositions, and reaction extents
• High and Low constraints for flows, storage masses & compositions
• All setpoints and setpoint weights, as well as constraints, can vary in time
• A weight (value) can be set on the total mass in a tank for the objective function
• Move suppression for steady control, by penalty weights on absolute flow changes in the objective function
• Rate-of-change (ramp) limits for flows
• Exit streams from a tank can undergo reaction, fractional component separation, or total flow splitting

The oil refining process is modelled with a crude distiller delivering four “straight run” streams (gasoline, naphtha, diesel and fuel oil) (Boucheikhchoukh et al, 2022). These split and combine through a flowsheet which includes catalytic reforming and catalytic cracking in order to maintain specifications for four different-valued products (premium gasoline, regular gasoline, diesel fuel and fuel oil). Intermediate nodes within the process may be set at various storage capacities to alleviate upsets.

Optimal strategies are presented for steady-state operation, crude supply and product demand steps, a crude change, and shut-downs of the catalytic cracker and catalytic reformer. In each case the dynamics and gross profit of the process are monitored over a 20-day period. The gross profit was based on the marginal values of initial and final stock, less operating costs. In the case of the reformer, a comparison is made between a planned shutdown as opposed to an unplanned shutdown.

The novelty of this work lies in the full optimisation on the dynamic changes over the defined horizon, with the process returning to a steady-state economic optimum. Product flows are maintained on specification by re-routing through the flowsheet. The iterative updating of stream compositions facilitates stoichiometric conversions, component separations and stream splitting.

Mineral flotation, a critical physiochemical process in the mineral industry, is one of the largest separation techniques in mineral processing [1, 2]. This method extracts desired minerals through a froth phase by leveraging differences in surface properties. Despite its conceptual simplicity, problems such as non-linear dynamics, complex interactions, and multiphase instabilities pose control challenges in these systems [2]. Achieving optimal control is crucial, as even marginal improvements in recovery can yield substantial economic benefits due to the process’s large-scale operation [3]. Model Predictive Control (MPC) is a prevalent strategy, capable of handling non-linear, multi-variable processes and implementing explicit constraints. An integral part of MPC strategies is the process model, which was replaced in this study by a Gaussian Process (GP) representation of the mineral froth flotation system to capture its complex, nonlinear dynamics and provide probabilistic state predictions.

In this study, a GP-MPC strategy is proposed. The motivation is to directly leverage available process data and build a data-driven model that can be used for control. The rationale behind using GPs as the model of choice is the uncertainty quantification and probabilistic attributes that it provides, which allows to better control the complex and non-linear dynamics of the mineral froth flotation process. Simultaneously, this approach, capitalises on the predictive capabilities and multivariable control advantages associated with MPC. The use of the Google machine learning framework ’JAX’ was undertaken to accelerate the training process of the GP through leveraging just-in-time compilation as well as autodifferentiation for gradient computation. Altogether, the GP model was trained, optimised with JAX, assessed with relevant metrics and implemented to be used alongside a MPC strategy as a surrogate. This surrogate would then subsequently provide a probabilistic process model to be queried by the MPC optimisation process for future state predictions.

Overall, the results of this GP-MPC strategy are promising. The GP demonstrates accurate predictions with low amounts of errors on testing datasets, with an average root mean squared percentage error of 0.47% and average mean absolute percentage error of 0.33%. Furthermore, implementing JAX reduced training runtime and improved the accuracy of state predictions during training. The MPC portrayed low amounts of error between control setpoint and actual control actuation aswell as producing a control trajectory which resulted in high amounts of recovery of the desired mineral.

Altogether, this GP-MPC approach highlights the potential for integrating data-driven methods into control strategies for mineral flotation. This work aims to encourage the broader adoption of such approaches within the mineral flotation community, demonstrating that data-driven control strategies are a viable and promising option worthy of further, in-depth investigation.

[1] P. Quintanilla, S. J. Neethling, and P. R. Brito-Parada, "Modelling for froth flotation control: A review," Minerals Engineering, vol. 162, 2021.

[2] B. A. Wills, Wills’ Mineral Processing Technology, 8th ed., Butterworth-Heinemann, 2015.

[3] J. P. Ferreira II and B. K. Loveday, "An Improved Model for Simulation of Flotation Circuits," 2000.

Solar Thermal Power plants (STP) are used for large-scale electricity production from solar energy. However, STP faces significant challenges in operations resulting from (i) Supply-side disturbance, namely diurnal and seasonal solar radiation variations and cloud-cover induced uncertainties, (ii) Demand-side disturbance, namely fluctuating prices of electricity and (iii) Operational challenges, namely highly dynamic operation, and need for frequent plant shutdown if adequate energy storage is unavailable.

An optimal operating strategy is preferred to operate the STP in such a dynamic scenario. In literature, revenue maximization has been used as an objective to obtain optimal operating conditions (Camacho et al. 2013). However, existing research relies on simplified steady-state models while generating the optimal operating conditions, which may not adequately capture dynamic variations of the plant. In contrast, in the current work, we propose a dynamic plant-wide model-based revenue optimization approach to obtain optimal operating conditions. The objective function representing revenue in the proposed approach accounts for changing electricity prices and power generated by the STP. The objective function is proposed to be maximized while accounting for plant dynamics in the presence of several operational constraints. Thus, the proposed approach accounts for solar radiation variability, changing electricity demand, and process dynamics, enhancing revenue optimization and operational reliability.

To demonstrate the approach, we conducted a simulation case study on a 1 MW hybrid solar thermal power plant in Gurgaon, India (Surrender et al. 2019). This plant includes two solar fields, namely a Parabolic Trough Collector (PTC) field and a Linear Fresnel Reflector (LFR) field, along with a High-temperature Tank (HT), Low-temperature Tank (LT), Super-Heater (SH), Steam-Generator (SG), Pre-Heater (PH), and Steam Drum (SD). A dynamic model of plant behavior is available in Surrender et al. (2019) and is used in the current work. For short-term revenue optimization (over 1-day operation), with nominal variation in solar radiation, we focus on optimizing the oil mass flow rate through the PTC. Variation of this flow rate determines the amount of heat gained at the PTC outlet, which in turn affects the steam production by the heat-exchanger assembly (PH, SH, SG). The mass flow rate is allowed to vary at discrete time instants during the operation and is held constant between these time instants. Thus, from an optimization perspective, the flow rates at the specified discrete times become the decision variables. These decision variables affect the revenue (to be maximized) via the plant dynamic model, resulting in a Nonlinear Linear Programming (NLP) problem. These decision variables are optimized in the presence of several operational constraints.

The results show the effectiveness of our optimization strategy in generating an operational profile that significantly boosts the revenue generated by the STP.

References:

1) E. F. Camacho and A. J. Gallego, “Optimal operation in solar trough plants: A case study,” Solar Energy, vol. 95, pp. 106–117, 2013

2) S. Kannaiyan, S. Bhartiya, and M. Bhushan, “Dynamic modeling and simulation of a hybrid solar thermal power plant,” Industrial & Engineering Chemistry Research, vol. 58, pp. 7531–7550, 2019.

Driven by increasing urbanization and industrialization, clean water demand has firmly become one of humanity's top problems. The restoration of wastewater quality and recovery of wastewater resources have emerged as topics of high interest for researchers and practitioners of the wastewater industry. As the typical wastewater plant (WWTP) accounts for 30% to 40% of the total urban energy consumption and greenhouse gas (GHG) of the WWTP sector accounts for 1% to 2 % of the global GHG emissions, the importance of economic and sustainable operation of the WWTP turns out to be very high. Modelling and process control have the remarkable potential to support the tough energy and GHG emissions restrictions and cope with continuous stricter regulations for clean water quality.

This study considered the WWTP's water line with an Anaerobic-Anoxic-Oxic design that operates according to the largest spread activated sludge technology. The developed model of the WWTP was calibrated with process data from the Cluj-Napoca municipal WWTP, characterized by a treatment capacity of around 420,000 PE.

The paper presents solutions for operating the WWTP based on advanced process control methods, merging the benefits of the cooperation between the lower-level decentralized control loops with the higher-level model predictive control strategy and setpoints optimization. The low-level control approach combines the feedback and feedforward control configurations with cascade control. They are used to control the nitrification in the aerated bioreactors and the denitrification in the anoxic reactor. The nitrate and nitrites concentration (NO) in the anoxic reactor is controlled by manipulating the internal recycle flowrate and the aeration is controlled either by directly controlling the Dissolved Oxygen (DO) or indirectly by controlling the ammonia (NH) concentration in the aerated reactors (in feedback or cascade configuration with DO control), using as manipulated variable the air flowrate. The traditional PI control of NO, DO, and NH is compared to the supervisory control having at the upper level either an optimization layer that provides the optimal values of the setpoints or the model predictive control layer implementing the traditional or the optimized setpoints emerged from an additional and uppermost optimization control layer. The optimized setpoints are computed based on a global optimization index consisting of a weighted sum of spent energy (aeration and pumping), effluent quality, and greenhouse gas (CO₂ and N₂O) emissions performance sub-indices.

Comparison of the various control strategies' performance, the inclusion of the GHG sub-index in the setpoints optimization-based computation and finding of the solution of the optimization task with the support of the specially trained artificial neural network, aimed to promptly predict the performance indices and support the real-time implementation, are the novel contributions of this work.

In industrial processes, balancing stringent safety requirements with operational efficiency is a complex challenge, especially when safety constraints conflict with performance objectives. This research introduces a theoretical framework that integrates safety and control objectives through a chance-constrained Explicit Model Predictive Control (eMPC) approach, designed to achieve an optimal trade-off between operational performance and safety assurance. The core innovation of this framework is the introduction of tolerance-based safety constraints, which permit controlled violations of safety-critical limits. This flexibility allows the system to maximize operational efficiency while maintaining a robust level of safety, addressing the inherent trade-off between the two competing goals. Central to this framework is the incorporation of Bayesian-based dynamic risk assessment in a receding horizon manner, enabling the system to continuously update and adjust safety constraints in response to real-time shifts in risk associated with safety-critical variables. This real-time adaptability ensures that safety margins evolve in tandem with the operational context, enhancing the system’s ability to respond to uncertain and fluctuating conditions. Moreover, the eMPC model is equipped with learning/adaptive capabilities, allowing it to retain the knowledge of historical safety incidents and fault data. This learning mechanism enables the control system to proactively mitigate risks in future operations, anticipating safety issues before they escalate into critical failures. As a result, the system progressively refines its decision-making over time, achieving a stronger balance between safety compliance and operational efficiency. The framework's validation is presented through a case study involving a Continuous Stirred Tank Reactor (CSTR) under thermal runaway conditions. The study illustrates how the learning control model forecasts system vulnerabilities, diagnoses emerging faults and proactively adjusts control strategies to preempt safety-critical failures.

Batch processes are still mainly operated via operation recipes, which are sequential procedures^[1] of consecutive operation steps conjoined by logical transition conditions. These recipes are typically derived from expert process knowledge. Their application is advantageous when the batch operation is optimized as it provides a parameterization of the (now restricted) control space and reduces the dimensionality of the optimization problem significantly, especially for nonlinear dynamic process models.

However, the use of heuristically determined recipes leads to an event-driven system, which complicates the implementation as an equation-based, smooth, dynamic optimization (DO) problem. Therefore, previous operation recipe implementations relied on imperative programming (e.g., while loops, if-else statements) to represent the recipe, requiring sequential solution methods.

In the presented novel recipe formulation, the information on the currently active recipe step is stored using auxiliary differential variables, which enables simultaneous optimization, including the sequence of the recipe itself. The decision variables determining active recipe steps are formulated using sigmoidal functions combined with auxiliary differential variables, thus avoiding the presence of binary variables in the DO problem^[2]. Although this optimization formulation requires auxiliary variables and equations, the problem size only increases linearly with the number of recipe steps.

The proposed formulation is first applied and tested on a model of moderate size and solved in Python using a collocation approach. The examined zero-dimensional dynamic model consists of balance equations, kinetic rate equations, and further constitutive equations. For the proposed optimization problem, the computational expense of sequential vs. simultaneous methods is compared using the case process model.

For the moderate problem size in our case study, the computational expense is lower for the applied sequential method than for a simultaneous method. However, our recipe formulation facilitates parallelization for the dynamic optimization of recipe-based batch processes, while its performance against alternative formulations, such as mixed-integer dynamic optimization models needs to be benchmarked. In the next step, our formulation will be applied to a more complex case study of in-situ reaction/separation in an ammonia synthesis-sorption unit.

Sources

[1] Brand Rihm, Gerardo & Esche, Erik & Repke, Jens-Uwe. (2023). Efficient dynamic sampling of batch processes through operation recipes. Computers & Chemical Engineering. 179. 108433. 10.1016/j.compchemeng.2023.108433.

[2] Torben Talis, Erik Esche, Jens-Uwe Repke. (2024). A Smooth and Pressure-Driven Rate-Based Model for Batch Distillation in Packed Columns Using Hold-Time Constraints for Bang-Bang Controllers. (Eds.:) Flavio Manenti, Gintaras V. Reklaitis , BoA of (ESCAPE34/PSE24), June 2-6, 2024, Florence, Italy.

Acknowledgements

Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them. Grant agreement No 101058643.