Control system, machining device and method for operating machining device

Abstract

The present invention relates to a method for operating a processing apparatus wherein the processing apparatus is configured to perform at least one of substance modification and substance conversion, preferably at least one of material modification and substance conversion, with the aid of at least one of a physical effect, a chemical effect, a biological effect and a nuclear effect. The invention relates to a method for separating air into nitrogen and / or oxygen and / or argon and / or other components included in the air, said method comprising the following steps executed by a control system comprising a model predictive controller having a control model (320): wherein the model predictive controller is a non-linear model predictive controller, and wherein the control model (320) is a hybrid model, the hybrid model includes at least one data-based component (324, 326, 424) and at least one dynamic component (322): initializing a control model; determining (230) one or more setpoints of one or more parameters of one or more process components and / or units of the machining apparatus based on the initialized control model; and providing the one or more setpoints to control the processing apparatus.

Description

[0001] The present invention relates to a method for operating a processing device, a control system and a processing device, such as a processing device including an air separation unit. Background Art

[0002] A processing device (also referred to as a chemical plant or chemical engineering plant) is generally understood to be a device for modifying and / or transforming substances with the aid of purposeful physical effects and / or chemical effects and / or biological effects and / or nuclear effects. Such modifications and transformations typically include comminution, screening, mixing, heat transfer, re-distillation, crystallization, drying, cooling, filling, mass transfer between different forming stages, and superimposed substance conversions, such as chemical reactions, biological reactions or nuclear reactions.

[0003] A processing device may include several different components, such as different heat exchangers (e.g., plate-fin heat exchangers (PFHE) or coil-wound heat exchangers (CWHE)), towers (e.g., distillation towers and / or adsorption towers and / or scrubbing towers), machines (such as compressors, boosters, turbines, pumps, cold boxes, etc.). For example, a processing device may be a device for producing a specific gas, i.e., the device may be or include an air separation unit, a natural gas device, an ethylene device, a hydrogen device, an adsorption device, etc.

[0004] One way to control the operation of such a processing device is to use so-called model predictive control or a model predictive controller (MPC). Linear models are generally widely used, but their scope of application is limited. However, using non-linear model predictive control can be difficult. In view of this, the object of the present invention is to provide an improved way of operating a processing device, preferably overcoming the difficulties of using non-linear module predictive control. Summary of the Invention

[0005] This object is achieved by providing a method for operating a processing device, a control system and a processing device according to the independent claims. Advantageous further embodiments form the subject matter of the dependent claims and the subject matter described hereinafter.

[0006] The present invention relates to an operating processing device. The processing device is configured to perform at least one of material modification and material transformation with the aid of at least one of physical effects, chemical effects, biological effects, and nuclear effects. Such modifications and transformations include, for example, at least one of the following: comminution, screening, mixing, heat transfer, redistillation, crystallization, drying, cooling, filling, and superimposed material conversions such as chemical reactions, biological reactions, or nuclear reactions. The processing device may include several different components such as different heat exchangers (e.g., plate fin heat exchangers (PFHE) or shell and tube heat exchangers (CWHE)), towers (e.g., distillation towers and / or adsorption towers and / or scrubbing towers), machines (such as compressors, boosters, turbines, pumps, cold boxes, etc.). For example, the processing device may be a device for producing a specific gas, i.e., the device may be or include an air separation unit, a natural gas device, an ethylene device, a hydrogen device, an adsorption device, etc.

[0007] Hereinafter, the present invention will be described with a processing device that is or includes one or more air separation units (ASU). However, it should be noted that the present invention is applicable not only to the processing device in the above sense but also to additional and / or other types of processing devices other than air separation units; the present invention is applicable to, for example, general industrial chemical processes. For example, the processing device may be configured to perform one or more processes related to at least one of the following: natural gas, syngas, petrochemicals, hydrogen, downstream processes of electrolysis (e.g., water electrolysis using PEM, AEL, AEM, SOEC, and combinations thereof, ammonia circuits and methanol circuits based on hydrogen from electrolysis). In one embodiment, the processing device includes or is at least one of the following: an air separation unit, a natural gas device, an ethylene device, a hydrogen device, and an adsorption device.

[0008] An air separation unit (ASU) is typically a highly integrated system that poses advanced requirements for the optimization of product recovery and load change capabilities. These methods should be able to respond quickly to changes in customer demand while ensuring the highest possible product yield at a specified purity. Depending on the device topology and / or component design and / or operating strategy, the load change time can vary, for example, from 0.5% per minute to 10% per minute. For example, if the air separation unit includes a column system for argon separation, controlling the transfer flow from a two-column to an argon system is challenging and limits the load change speed. For example, the typical impurity limit for nitrogen product is 1 ppm oxygen (O2). The oxygen product typically includes 99.5% to 99.9% oxygen. The argon product is typically contaminated with 1 ppm oxygen (O2) and 1 ppm nitrogen (N2).

[0009] For example, the variable availability and price pattern of electrical energy may require flexible (e.g., load-flexible and / or production rate-flexible) operation of the air separation process. From a control perspective, an air separation unit is a strongly coupled multivariable system. Therefore, it is necessary to optimize the setpoints of the base layer controller according to the current state of the plant, and this can be achieved by model predictive control (MPC) as described above. If model predictive control is applied, i.e., if the operation of the processing plant is based on model predictive control, a linear model predictive controller (LMPC) is typically used to control the process or industrial process (such as the air separation process).

[0010] Model predictive control is an advanced control strategy. A model predictive controller includes a model of the controlled process, such as a dynamic model. The dynamic model of the application process provides a prediction of the future process behavior of the processing plant or the controlled process. Mathematical algorithms can be used to calculate the optimal control action.

[0011] The dynamic model can be based on, for example, physical equations and / or design information (in which case the model is called a white box model). The dynamic model can also be based on, for example, a mathematical model of the input and / or output correlations for data information (in which case the model is called a black box model), such as a non-linear autoregressive exogenous (NARX) model. The dynamic model can also be based on, for example, model structures that are physically and / or mathematically motivated, which are (partially or fully) interpretable, combined with design information, and combined with empirical correlations from data information (in which case the model is called a grey box model).

[0012] White box models are based on physical principles and design information; such models are also called first principle models. Therefore, these models are valid and applicable within a wide operating range of the processing plant. The disadvantage is that for complex processes, these models may limit the speed of optimization when used as control models in plant operation. In many cases, grey box models can be optimized faster than white box models, but are more limited in terms of their validity range.

[0013] Depending on the modeling strategy, the control model (i.e., the model for MPC) can be linear or non-linear. A linear model gives rise to a linear model predictive controller LMPC, while a non-linear model gives rise to a non-linear model predictive controller NMPC.

[0014] LMPC is typically based on identifying a linear model from the recorded plant data. For example, plant data is obtained from a step test based on one or more operating points. In this case, it can be assumed that the dynamic system of the plant behaves linearly (at that operating point). However, air separation and other industrial processes typically exhibit non-linear dynamics. The assumption of linear dynamic behavior limits the reliable operating range of the generated LMPC. Therefore, the further the operation of the process is from the reference operating point (for which linearity is approximately valid), the worse the performance of the controller becomes.

[0015] Compared with LMPC, applying NMPC and specifically including the natural non-linearity of the process in the control model for industrial process control enables innovative automation solutions that go beyond LMPC. This allows, for example, the expansion of the controller operating range. This also allows, for example, reliable and safe exploration of new operating regions when including physical and / or mathematical knowledge about the system in the control model (white-box and grey-box models). More specifically, this feature allows the application of the model in operating regions outside the identification range of the model. This also enables, for example, faster load changes for load-flexible production. This allows, for example, further optimization of the operating efficiency.

[0016] However, operating or controlling a processing plant using or applying NMPC has some challenges. LMPC is a well-established automation solution in the industry, while the practical application of NMPC is still rare due to i) challenges in control model development, ii) real-time optimization of the control model, iii) online state initialization of complex non-linear models, and iv) lack of available commercial frameworks.

[0017] One way to achieve ii) real-time capabilities is to apply model reduction to the detailed process model. Model reduction refers to the process in which the size and / or complexity of the detailed model are reduced such that the properties of the reduced model (e.g., the number of model variables, the number and / or type of mathematical operations) allow less effort to be expended in terms of the purpose and / or use of the reduced model, such as mathematical optimization. However, model reduction has some drawbacks and challenges. Selecting an appropriate reduction method and performing model reduction and / or implementing the reduced model can be time-consuming, non-intuitive, and thus challenging for control engineers. Model reduction does not guarantee the real-time capabilities of the controller using the reduced model. The reduced model may still be computationally demanding. Model reduction may result in a significant offset between the reduced model and the detailed model or the real processing plant. If the offset is too large, it needs to be compensated. Large offsets affect the NMPC performance. Model reduction may be accompanied by a limitation of the validity range of the reduced model (compared to the more detailed model). If the validity range of the reduced control model is limited, the reliable operating range of NMPC is restricted.

[0018] A method for model simplification and simplified modeling is to use a so-called hybrid model. The hybrid model can combine at least one data-based component with at least one dynamic component (e.g., a low-dimensional dynamic component). In these general simplified models, the data-based component (e.g., an artificial neural network or a wave equation) represents the static or quasi-static non-linear characteristics and / or patterns of the process (i.e., "the general shape of the device"). On the other hand, the low-order dynamic model part describes the dynamic response of these patterns to external excitations (i.e., "the general time response of this shape").

[0019] In the present invention, the operation of the processing device is performed by means of a control system (or a controller, i.e., a computing or processing unit). The control system includes (or implements) a model predictive controller having a control model, i.e., using model predictive control. An operation method includes a step of initializing the control model. The method includes another step of determining one or more setpoints of one or more process components and / or units (or the processes of these process components and / or units) of the processing device based on the initialized control model. The one or more parameters may include one or more of the following parameters: valve position, medium flow rate, medium pressure, medium temperature, medium component concentration, and the throughput of a pump, compressor, or turbine. This step may include, for example, optimization or optimization calculations typical for model predictive control, which may include, for example, predicting different future values and selecting the best value. The method includes another step of providing one or more setpoints to control the processing device (especially actual control).

[0020] The model predictive controller used is a non-linear model predictive controller. In addition, the control model is a hybrid model (or a grey-box model); the control model includes at least one data-based component and at least one dynamic component. The data-based component may be or be based on, for example, an artificial neural network or a polynomial.

[0021] The hybrid model combines at least one data-based component (e.g., an artificial neural network or a wave equation) with at least one physical information component (e.g., a dynamic model structure, physical correlation, or the geometric structure and / or topological structure of the components of the processing device). The hybrid model does not necessarily have to involve physical equations, but can also be defined based on the physical equation structure (specifying the structure of the simplified model, but not necessarily the explicit equation).

[0022] (Physics) A dynamic model - or at least one dynamic component of a hybrid model - can be based on at least one of the following: thermal-fluid correlations, thermodynamic correlations, phenomenological (i.e., empirical) correlations, mechanical equations (of the states and physical properties of the device and the materials and components used therein). Additionally, the dynamic model can be based on the geometric structure and / or topological structure of the components of the processing device. Such (physics) models are also referred to as first-principles models because they are based on physical properties such as heat and mass balances, etc. In contrast, data-based models do not reflect the actual physical structure or characteristics of the processing device, but rather provide output data based on input data, like a black box.

[0023] The dynamic component can be or be based on a low-dimensional dynamic component of a mathematically simple type. In particular, the dynamic component can have a linear or bilinear structure, i.e., the component is a linear or bilinear dynamic model. Such a linear or bilinear structure reflects the modeled part of a processing device (e.g., a certain process) with linear behavior. This allows for efficient and fast optimization calculations. However, due to the additional data-based component, the model predictive controller (and the entire hybrid model itself) can be non-linear.

[0024] A (dynamic) hybrid model can be used, which (explicitly) has the following property, i.e., its dynamic component is linear / bilinear and its non-linearity is represented precisely and / or exclusively by the (data-driven) component. Its particular property is that such a model can be determined during the model simplification process (rather than via empirical input / output identification). The specific reason for using such a model is because such a hybrid model is particularly suitable for optimization, as the model structure can be used in a very specific way here (better than, for example, an empirical NARMAX model).

[0025] The arrangement of at least one data-based component and at least one dynamic component is sequential. In this case, the dynamic component makes (abstract) predictions in a timely manner, and the data-based non-linear component "converts" these predictions into available information, such as temperature, pressure, mass fraction, etc. (as required by the corresponding set points).

[0026] Such a sequential structure can be used very effectively for optimization because the time evolution of the (abstract) predictions does not depend on the data-based non-linear component. Therefore, the data-based non-linear component is not an explicit part of the optimization problem; rather, the data-based non-linear component is only (partially) implicitly evaluated as part of the operating constraints and cost function; the optimization is based on or performed using at least one dynamic component, i.e., only at least one dynamic component is used. This means that the optimization problem (including safety constraints) can be formulated with respect to a smaller variable space of the dynamic variables; thus, the optimization problem has a smaller dimension and can be solved very effectively.

[0027] Such hybrid models allow for efficient online optimization by a computer. The hybrid models used have a special structure that has been shown to be computationally beneficial for NMPC. Such hybrid models are characterized in that the computation and / or evaluation of the data-based component only depends on the dynamic component (and possibly external signals). NMPC can utilize this feature, resulting in a considerable computational speedup. Especially for sequential structures, the mathematical programs solved by the controller can only be formulated based on the dynamic components (variables) of the hybrid model. Therefore, NMPC compresses the complex optimization of the hybrid model into a very compact optimization. Compared with using a general optimization solver, this makes the computer solution of the control problem very efficient.

[0028] NMPC typically relies on the repeated optimization of a hybrid model with suitable objective functions and constraints. In an embodiment of the present invention, numerical algorithms (such as full discretization, direct single shooting method, or direct multiple shooting method) and underlying non-linear and / or linear programming methods (e.g., sequential quadratic programming method and / or interior point method) can be used to perform the optimization. The single shooting and multiple shooting optimization methods additionally include integration solvers, e.g., ordinary differential equation solvers or differential-algebraic equation solvers. In addition, numerical optimization methods may include structure detection methods, sparse utilization, and / or decomposition schemes.

[0029] In an example, the arrangement of at least one data-based component and at least one dynamic component can be integrated, i.e., these components are integrated with each other. Such an integrated structure cannot be utilized as effectively as a sequential structure, but in some cases, this structure leads to a more accurate model. Again, due to the simple structure, the optimization of such models is generally more effective (much more effective) than the optimization of (pure) physical models.

[0030] In one embodiment, it is preferable to optimize the control model online. Such (online) optimization can also be supplemented by the online improvement of the hybrid model, where the data-based component is adapted online to improve the controller performance. Adaptation means, for example, the online training or modification of the data-based component (such as an artificial neural network), where online data (such as the measurement results of the device) is used to improve the control model.

[0031] In an embodiment of the present invention, the control model is obtained by model simplification from a digital twin. For the generation of the simplified model, for example, a step-by-step process can be used, which clearly distinguishes the following tasks: i) digital twin modeling and simulation (data sampling), ii) model simplification (model structure specification and training of the data-based component), and iii) controller design, testing, and deployment.

[0032] Thereby, structured steps are provided, and gates and / or deliverables can be established during the modeling and control processes. Different steps can be performed by different experts in their respective fields.

[0033] In one embodiment of the present invention, at least one data-based component is trained based on at least one of simulation data and process data. The model structure selection can be part of a hybrid model building process. Hybrid model building is generally associated with a significant amount of work in model (structure) selection and model implementation. Now, the framework can provide a general formula library for hybrid models and allow for the simplification of the modeling (model building) of a specific process in a flow diagram manner, including simplifying the process structure and other design information. Simulation data and / or process data can be incorporated into the model building process, where the data-based component is specified (e.g., trained). The model building of the simplified model can start from a given digital twin model but can also be independent of the digital twin and built from scratch in a flow diagram manner. Note that such a framework can also be used for white box models and black box models (not only gray box models or hybrid models).

[0034] In one embodiment of the present invention, as described above, the control model can be improved (optimized) online. Since the hybrid model has at least one data-based component, the model can be set up offline and updated online. Therefore, the hybrid structure can not only enable computer-efficient online applications but also enable the continuous improvement of the model.

[0035] Generally, a digital twin is a digital representation of an intended or actual real-world physical product, system, or process (physical twin) that is used as a virtually indistinguishable digital counterpart of the physical twin for practical purposes such as simulation, integration, testing, system monitoring, and maintenance. The digital twin of an existing entity can but does not necessarily need to be used in real time and is synchronized with the corresponding physical system periodically. In the present invention, the digital twin is a digital representation of a real-world (physical) processing device and / or the process performed by the processing device. Therefore, no interference or online connection is required between the digital twin and the processing device. However, within the present invention, the digital twin can be used to provide an initial control model as a result of the model simplification process of the control model of the MPC for operating the (real-world) processing device.

[0036] The digital twin can be based on (or be) a (detailed) dynamic model of the processing device. This dynamic model can be based on at least one of the following: thermal-fluid correlations, thermodynamic correlations, phenomenological (i.e., empirical) correlations, mechanical equations (of the state and physical properties of the device and the materials and components used therein). Additionally, the dynamic model can be based on the geometric structure and / or topological structure of the components of the processing device. In one embodiment, the dynamic model can be configured to represent the transition from one state to another state of the processing device, particularly the separation column of the device.

[0037] Such (physical) models are also known as first - principle models because they are based on physical properties such as heat and mass balance. This model can represent the geometric and topological structures of equipment or components at the same level of detail used in process and / or equipment design. That is, a heat exchanger is modeled, for example, in one spatial dimension, while a distillation column is modeled as a series of theoretical trays. This approach enables the use of design correlations for pressure drop, heat transfer coefficient, fluid holdup, etc. However, unlike equipment design tools, this dynamic model specifically balances the holdup of gas and liquid as well as energy, and thus allows the modeling of the instantaneous state of the plant. This dynamic model also specifically covers the entire operating range of the plant. This entire operating range can include startup, normal operation, shutdown, and fault management. Such models can also be called pressure - driven models (but not necessarily must be pressure - driven models).

[0038] Such dynamic models used as or for digital twins are different from the control models used in MPC for operating a processing plant. Generally, the dynamic models used for digital twins are more detailed and complex than the control models. However, the control models used for MPC can be obtained, for example, from the dynamic models of digital twins. This is also known as model reduction, as described above.

[0039] The present invention enables the practical implementation and application of reduced models and NMPC. So far, model reduction has been performed as a problem - specific task and involves a large amount of modeling work. The practical application of NMPC is still rare in the industry because it is very difficult to implement. The present invention implements routines for generating reduced control models with little effort using a toolbox. Due to the combination of the customized reduced - model structure and the exploitation of the structure of the hybrid reduced model in the optimization step, it is also able to optimize NMPC in real time. Implementing NMPC instead of LMPC allows for faster load changes, an increased controller operating range, and further optimization of plant efficiency (e.g., based on the air intake during load changes and the steady - state full - load of the ASU without argon production, less power consumption of the MAC (main air compressor) is required), resulting in lower overall energy consumption of the processing plant.

[0040] One embodiment of the present invention relates to a control system that is preferably configured to execute the method according to the present invention by means of a computer program (stored on the control system). For example, such a control system can be or include a processing system or a processor, or be a part of a processing system or a processor.

[0041] Other aspects of the invention are a computer program and a computer-readable data carrier, the computer program having program code means for causing a control system or a processing system to execute the method according to the invention, and such a computer program being stored on the computer-readable data carrier. This allows for particularly low costs, especially when the computing unit being executed is still used for other tasks and thus exists anyway. Suitable media for providing the computer program are in particular floppy disks, hard disks, flash memories, EEPROMs, CD-ROMs, DVDs, etc. Downloading the program over a computer network (Internet, intranet, cloud applications, etc.) is also possible.

[0042] Further advantages and embodiments of the invention will become apparent from the description and the drawings.

[0043] It should be noted that, without departing from the scope of the invention, the features mentioned above and the features to be described further below can be used not only in the indicated combinations, but also in other combinations or alone.

[0044] The invention will now be described further with reference to the drawings showing preferred embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A processing apparatus according to an embodiment of the invention is schematically shown;

[0046] Figure 2 A method according to an embodiment of the invention is schematically shown;

[0047] Figure 3 A control model according to an embodiment of the invention is schematically shown;

[0048] Figure 4 A method according to an embodiment of the invention is schematically shown;

[0049] Figure 5 A control model according to another embodiment of the invention is schematically shown;

[0050] Figure 6 A method according to another embodiment of the invention is schematically shown; and

[0051] Figure 7 A method according to another embodiment of the invention is schematically shown. DETAILED DESCRIPTION

[0052] Figure 1Schematically shown is a processing apparatus 100 according to an embodiment of the present invention. By way of example, the processing apparatus 100 includes or is an air separation unit (ASU) having nitrogen product, oxygen product, and argon product. The processing apparatus 100 includes other components such as a main air supply source 1, a main heat exchanger 5, an expansion turbine 6, other heat exchangers, a pump 8, a high-pressure column 11, and a low-pressure column 12. The high-pressure column 11 and the low-pressure column 12 are part of a typical distillation column system.

[0053] In the air separation unit, an input air stream supplied via the main air supply source 1 is typically compressed, pre-cooled, and cleaned. Additionally, the air stream can be split into two streams, where one stream is fully cooled in the main heat exchanger 5 and the other stream is only partially cooled in the main heat exchanger 5. The partially cooled stream can then be expanded by means of the expansion turbine 6. The fully cooled air stream can be supplied to the high-pressure column 12 via a heat exchanger, while the partially cooled air stream can be supplied to the low-pressure column 11. Note that the air separation unit shown here is an example. Other types of ASUs are also possible.

[0054] Furthermore, a control system 200 is shown, which implements a model predictive controller 210 with a control model 220. In one embodiment, the processing apparatus includes the control system 200.

[0055] The control system 200 is configured to operate the processing apparatus 100, for example, where one or more setpoints of one or more parameters of one or more processes (or process components / units) of the processing apparatus 100 are determined and provided. By way of example, different parameters that can be controlled are shown by means of several controllers shown at specific locations within the processing apparatus 100. Such controllers include a flow controller FIC, a pressure controller PIC, a level controller LIC, and a temperature controller TIC. The control system can be used to specify the setpoints of one or more of these controllers, or replace one or more of these controllers LIC, PIC, FIC, TIC (or AIC, which is not shown Figure 1 here).

[0056] Figure 2 Schematically shown is a method according to an embodiment of the present invention. The method can be performed using, for example Figure 1 the control system 200 shown as well as the model predictive controller 210 with the control model 220 running on the control system 200. In Figure 2 it, only the control model 220 is shown. Preferably, the model predictive controller is a non-linear model predictive controller, that is, the control model 220 is a non-linear model.

[0057] Step 215 shows the initialization of the control model 220. This can include, for example, state estimation based on, for example, a non-linear extension of a Kalman filter. As mentioned above, many parameters of the processing device or its process cannot be measured or are not measured, but are to be used for operating the processing device.

[0058] In step or process 230, one or more setpoints for one or more parameters of one or more processes of the processing device are determined (typically in optimization). By way of example, two setpoints 252, 254 are shown. As with Figure 1 that which is mentioned, different setpoints for different parameters can be determined. The determination of these setpoints is based on the initialized control model 220, typically in optimization. Additionally, further requirements or values can be considered in such optimizations, for example, process objective 240 (e.g., desired medium flow rate), process constraint 242 (e.g., product purity), and / or tuning parameter 244.

[0059] In step 250, these setpoints are provided to operate the processing device 100, i.e., these setpoints are set in the real processing device 100 for, for example, the corresponding valves or medium flow rates, etc. Then measurements 260 can be made at the processing device 100. This can include measuring the values of the parameters for which the setpoints have been determined and / or other parameters. The measurements (or measurement values) can additionally be used within step 230 for determining the setpoints and within step 215 for initializing the control model, or rather, the new setpoints are used at a later point in time.

[0060] Figure 1 and Figure 2 shows how the control model is used within a model predictive controller in this way, while Figure 3 more particularly shows a control model 320 according to an embodiment of the present invention. For example, the control model 320 can correspond to the control model 220.

[0061] The control model 320 is a hybrid model including a data-based component 324 and a dynamic component 322. More than one data-based component and / or more than one dynamic component can also be used. The control model 320 has a sequential structure, i.e., the dynamic component 322 and the data-based component 324 are arranged sequentially as Figure 3 shown. The dynamic component 322 makes (abstract) predictions in a timely manner, and the data-based component "converts" these predictions into available information, such as temperature, pressure, mass fraction, etc. (as required according to the corresponding setpoints).

[0062] The dynamic component 322 can be or be based on a low-dimensional dynamic component of a mathematically simple type. For example, the dynamic component 322 can have a linear or bilinear structure, i.e., the component is a linear or bilinear dynamic model. Such a linear or bilinear structure reflects a modeling part of a processing device (e.g., a certain process) with linear behavior. This allows for efficient and fast optimization calculations. The data-based component 324 can be based on, for example, one or more artificial neural networks.

[0063] In addition, the initialization control model 320 (see also Figure 2 step 215 in) (especially the dynamic component 322 of the control model 320) can also be based on a data-based component (e.g., the data-based component 326). Similar to the data-based component 324, the data-based component 326 can be or be based on, for example, one or more artificial neural networks. Initializing the control model includes: determining initial or starting conditions of, for example, internal states before the optimization process starts. The reason for high efficiency is that the data component can take over steps that were originally (or usually) done by additional algorithms (such as Kalman etc. as mentioned above).

[0064] Such a sequential structure can be used for optimization very effectively because the (abstract) predicted time evolution does not depend on data-based non-linear components. Therefore, the data-based non-linear components are not an explicit part of the optimization problem; instead, the data-based non-linear components are only (partially) implicitly evaluated as part of the operating constraints and cost function, or the optimization is based on or performed using at least one dynamic component, i.e., only at least one dynamic component is used.

[0065] For example, this is shown in Figure 4 which shows the control model 320 with the data-based component 324 and the dynamic component 322 twice by way of example. The optimization 330 (which can be similar to Figure 2 step 230 shown) only utilizes the control variables and the variables of the dynamic component 322.

[0066] This also means that, for example, the safety constraints 340 that must be considered can be "transferred" to a smaller variable space of the dynamic component 322 "through" the data-based component 324, i.e., the safety constraints directly affect the dynamic component. Therefore, the optimization problem has a smaller dimension and can be solved very effectively.

[0067] Figure 5More particularly, the control model 420 according to the example is shown. For example, the control model 420 can be used in place of or instead of the control model 220. The control model 420 is a hybrid model including a data-based component 424 and a dynamic component 422. More than one data-based component and / or more than one dynamic component can also be used. The control model 420 has an integrated structure, i.e., the dynamic component 422 and the data-based component 424 are integrated with each other, as Figure 5 shown.

[0068] Such an integrated structure cannot be utilized as effectively as a sequential structure, but in some cases, this structure results in a more accurate model. Again, due to the simple structure, the optimization of such models is generally more effective (much more effective) than the optimization of (pure) physical models. Integration particularly means that in the model, information flows from the dynamic component to the data-based component and vice versa. Therefore, these components are integrated and dependent on each other (different from the sequential structure, in which the prediction of the dynamic component does not depend on the data-based component).

[0069] Figure 6 A method according to another embodiment of the present invention is schematically shown, particularly a method related to obtaining a control model. This can be divided into three tasks or groups of steps. For simplified model generation, for example, a step-by-step process can be used, which clearly distinguishes the following tasks: task 600, digital twin modeling and simulation (data sampling); task 610, model simplification (model structure specification and training of the data-based component); and task 620, controller design, testing, and deployment. Thereby, structured steps are provided, and gates and / or deliverables can be established during the modeling and control processes. Different steps can be performed by different experts in their respective fields.

[0070] Hereinafter, reference is made to Figure 6 to describe an exemplary method. In step 601, design information (process information, i.e., information related to the design of the process) can be provided. In step 602, a material data model and a detailed model for each process can be provided. Based on the information provided in steps 601 and 602, model building and verification - step 603 can be performed. The result of step 603 is the digital twin 604. In addition, simulation and data sampling - step 605 can be performed to obtain relevant data and a complex model 611 for the digital twin.

[0071] As described above, the digital twin can be based on (or be) a (detailed) dynamic model of the processing device. This dynamic model can be based on at least one of the following: thermal-fluid correlations, thermodynamics correlations, phenomenological (i.e., empirical) correlations, mechanical equations (of the state and physical properties of the device and the materials and components used therein). Additionally, the dynamic model can be based on the geometric structure and / or topology of the components of the processing device. In one embodiment, the dynamic model can be configured to represent the transition from one state to another state of the processing device, particularly the separation column of the device.

[0072] Such (physical) models are also referred to as first-principles models because they are based on physical properties such as heat and mass balances, etc. The model can represent the geometric structure and topology of the device or component at the same level of detail used in process and / or equipment design. That is, a heat exchanger is modeled, for example, in one spatial dimension, while a distillation column is modeled as a series of theoretical trays. This approach enables the use of design correlations for pressure drop, heat transfer coefficient, fluid holdup, etc. However, different from equipment design tools, this dynamic model particularly balances the holdup of gas and liquid as well as energy, and thus allows the instantaneous state of the device to be modeled. The dynamic model also particularly covers the entire operating range of the device. This entire operating range can include startup, normal operation, shutdown, and fault management. Such models can also be pressure-driven models (but not necessarily so).

[0073] Task 600 involves building or generating a (complex) model for the digital twin, while task 610 involves model simplification for obtaining a control model (such as the control models 220, 320, 420 described above).

[0074] In step 612, data analysis is performed (based on model 611). This is followed by or includes selecting a simplified model structure - step 613. Based on the selected simplified model structure, a control model is built (model building), which includes training of the model - step 614. At step 615, the built and trained model is validated. Depending on the validation process and its results, further or different selection of the simplified model structure may be required (see step 613). Steps 613 to 615 can be repeated until the validation step 615 yields a desired or sufficiently simplified model 621.

[0075] The following task 620 involves controller design, testing, and deployment. In step 622, the simplified model 621 is exported to create a control model for NMPC - step 623. Then, in step 624, the control model is tuned, in step 625, the control model is tested, and in step 626, the control model is finally deployed.

[0076] Figure 7A method according to another embodiment of the present invention is schematically shown, particularly relating to selecting a model structure as part of a simplified model building process.

[0077] A library 700 including general simplified models or model components may be provided. By way of example, models 701, 702, 703, 704 are included in the library 700. Using this library and its models, step 710 of model building may be performed. This step 710 may correspond to, for example Figure 6 step 613. For the model building step 711, simulation data and process data may be considered. In addition, the structure and its characteristics of the process to be modeled may be considered - step 712.

[0078] The result of the model building is a simplified model 720 for a specific process (see also Figure 6 model 621 in; or Figure 5 420 in; or Figure 3 320 in; or Figure 2 220 in; or Figure 1 220 in). By way of example, three simplified model components or blocks 721, 722, 723 are included in the simplified model 720; arrows from or to such simplified model blocks or between them illustrate information flow, material flow, or energy flow.

[0079] In summary, such a library of general models or formulas for simplified models allows the modeling (model building) of a specific process to be simplified in a flowchart manner, including simplifying the process structure and other design information. Simulation data and / or process data (see steps 711, 712) may be incorporated into the model building process, where components based on data are specified (e.g., trained). The model building of the simplified model may start from a given digital twin model, but may also be independent of the digital twin and build the simplified model from scratch in a flowchart manner.

Claims

1. A method for operating a processing device (100), wherein, The processing device (100) is configured to perform at least one of material modification and material conversion with the aid of at least one of a physical effect, a chemical effect, a biological effect, and a nuclear effect. Preferably, air is separated into nitrogen and / or oxygen and / or argon and / or other components comprised in air. The method comprises the following steps performed by a control system (200) comprising a model predictive controller (210) comprising a control model (220, 320, 420), wherein the model predictive controller is a non-linear model predictive controller, and wherein the control model (220, 320, 420) is a hybrid model, wherein the hybrid model comprises at least one data-based component (324, 326, 424) and at least one dynamic component (322, 422). Initialize (215) the control model. Determine (230) one or more setpoints (252, 254) of one or more parameters of one or more process components and / or units of the processing device based on the initialized control model. Provide (250) the one or more setpoints to control the processing device (100). wherein the at least one data-based component (324) and the at least one dynamic component (322) of the hybrid model (320) are arranged sequentially. wherein determining (230) the one or more setpoints comprises optimization (330), wherein the optimization is performed using only the at least one dynamic component (322) variables.

2. The method according to claim 1, wherein, The at least one dynamic component (322, 422) has a linear or bilinear structure.

3. The method according to claim 1 or 2, wherein Initializing (215) the control model comprises: initializing the at least one dynamic component based on the at least one data-based component or based on at least one other data-based component.

4. The method according to any one of claims 1 to 3, wherein Determining (230) the one or more setpoints comprises: determining a prediction based on the at least one dynamic component (322), and determining the one or more setpoints based on the prediction and the at least one data-based component (324).

5. The method according to any one of the preceding claims, wherein, The at least one data-based component (324, 326, 424) comprises or is based on an artificial neural network.

6. The method according to any one of the preceding claims, the method further comprising: The control model (220, 320, 420) is adapted by adapting the at least one data-based component (324, 424) and / or the at least one dynamic component (422).

7. The method according to any one of the preceding claims, wherein, The control model is obtained by model reduction from a digital twin (604).

8. The method according to any one of the preceding claims, wherein, The at least one data-based component is trained based on at least one of simulation data and process data.

9. The method according to any one of the preceding claims, wherein, The one or more parameters comprise one or more of the following parameters: valve position, medium flow rate, medium pressure, medium temperature, medium component concentration, and throughput of a pump, compressor, or turbine.

10. The method according to any one of the preceding claims, wherein, The processing device (100) comprises or is at least one of the following: an air separation unit, a natural gas plant, an ethylene plant, a hydrogen plant, and an adsorption plant.

11. A control system (200), which is preferably configured to execute the method according to any one of the preceding claims by means of a computer program.

12. A processing device (100), which is configured to perform at least one of material modification and material conversion with the aid of at least one of physical effects, chemical effects, biological effects and nuclear effects, wherein the processing device comprises the control system (200) according to claim 11.

Images