Combined rigorous and data-based modeling of an industrial process

Abstract

The invention relates to a method for modeling, during operation, an output variable of an industrial process (1) which can be used in a process plant, comprising: a) carrying out the industrial process (1), b) detecting, by means of suitable detection means, process values (y) arising from the industrial process (1), c) processing the detected process values (y) by means of computer-implemented rigorous modeling (2), the detected process values (y) being input values for the computer-implemented rigorous modeling (2), d) using process values of at least one internal variable (x) of the computer-implemented rigorous modeling (2) which were determined by means of the computer-implemented rigorous modeling (2) as input variables for additional computer-implemented modeling (3) which is data-based, e) determining output values of the output variable (zd) by means of the additional computer-implemented, data-based modeling (3).

Description

[0001] Description

[0002] Combined rigorous and data-based modeling of an industrial process

[0003] The invention relates to a method for the operational modeling of an output variable of an industrial process having the features of the preamble of claim 1. Furthermore, the invention relates to a device for controlling and / or monitoring an industrial system having the features of claim 10.

[0004] Process models are required for numerous applications. These include prediction tasks, process optimization, as well as classification and anomaly detection. Indeed, many process engineering processes cannot be modeled rigorously with sufficient accuracy, or this is only possible with excessive effort. In these cases, particularly in recent years, there has been a widespread shift to data-driven modeling using machine learning (ML) or deep learning (DL) methods. However, even here, sufficient accuracy can often only be achieved with considerable effort. Furthermore, method experts (e.g., data scientists) sometimes lack a deeper understanding of the process, which further complicates application.

[0005] The structure of, for example, process engineering processes is often not taken into account in data-driven modeling. However, if the structure of the process to be modeled is known (e.g., process engineering basic and process flow diagrams, possibly also P&I diagrams), this knowledge should also be taken into account in the structure of the data-driven model. The method according to the invention also considerably simplifies the often necessary online adaptation of the modeling (particularly due to changing processes) and significantly reduces the requirements regarding the quantity and quality of the data required for online adaptation. In most cases, only data-based modeling is used. However, "forgetting" the rigorous model and switching completely to data-based modeling is unsatisfactory and wastes what has already been achieved.Furthermore, even the data-based model may not achieve the desired accuracy with reasonable effort. Furthermore, many data-based models are not suitable for providing transparency about their predictions. This dampens user confidence and, in cases of doubt, prevents their use. Furthermore, trust in rigorous models tends to be greater.

[0006] It is therefore advisable to further examine the rigorous model and combine it with ML and DL methods. ML methods are often used to optimize model parameters using data to better adapt the rigorous model to actual behavior. Combining both methods is also advantageous for certification, particularly with regard to the EU AI Act.

[0007] Alternatively, the rigorous model can be used to generate simulation data, which can be used for pre-training the data-based model or for parallel training. However, the model error must be appropriately accounted for. Especially when only certain parts of the rigorous model are inadequate, data-based models are used to replace only those parts.

[0008] To create a dynamic data-based model, the structure of the data-driven model may need to be extensively adapted (e.g., multiple layers, etc.). This often requires considerable experience and, if necessary, several iterations of the model structure to account for process and measurement instrument dynamics in the data-driven modeling.

[0009] Furthermore, data-driven dynamic modeling requires significantly more data to reliably enable dynamic modeling. Furthermore, modeling is further complicated if the plant being modeled is operated with closed-loop control. Furthermore, adapting the data-driven model is necessary when plant behavior changes, which places significantly higher demands on the scope and quality of the process data.

[0010] Processes, for example, in process engineering, can have quite complex structures. It is, of course, possible to neglect these structures in data-driven modeling. However, neglecting the process structure results in significantly increased demands on the scope and quality of the process data.

[0011] In addition, several approaches are known (e.g., https: / / agupubs.onlinelibrary.wiley.com / doi / full / 10.1029 / 2018WR023333) to learn a correction model for the rigorous model using additional data. However, this additional data is only used during training, not in subsequent application, which distinguishes this approach from the inventive method presented below.

[0012] Another well-known approach is to generate additional meaningful features from the data using simple, mostly algebraic equations. This begins with the application of functions to individual quantities (e.g., quadrature) and also includes combined evaluations of multiple quantities, as described at https: / / turbulentflux.com / combining-physics-and-machine-1-earning-to-understand-multiphase-translent-flow /

[0013] There are also known applications where the rigorous and the data-based model are operated in parallel and the more reliable one is selected by means of weighted linear combination (https: / / towardsdatascience.com / combining-physics-and-deep-learning-54eac4afel46) , which can, however, be done using Markov models (https : / / www . sciencedirect . com / science / article / ab s / pii / S 18755 10021005436 ).

[0014] The invention is based on the object of specifying a method and an associated device which enables the operational modelling of output variables of an industrial process in a resource-saving but at the same time efficient manner.

[0015] The above-stated object is achieved by a method for the operational modeling of an output variable of an industrial process having the features of claim 1. Furthermore, the above-stated object is achieved by a device for controlling and / or monitoring an industrial system according to claim 10. Advantageous further developments are the subject of the dependent claims.

[0016] A method for the operational modeling of an output variable of an industrial process, which can be used in a process plant, comprises the following steps: a) carrying out the industrial process, b) recording process values ​​originating from the industrial process by suitable recording means, c) processing the recorded process values ​​by computer-implemented rigorous modeling, wherein the recorded process values ​​represent input values ​​of the computer-implemented rigorous modeling, d) using process values ​​of at least one internal variable of the rigorous modeling, which were determined by the rigorous modeling, as input values ​​of a further computer-implemented modeling which is data-based, e) determining output values ​​of the output variable by the computer-implemented data-based modeling.

[0017] In particular, rigorous modeling includes mathematical / physical relationships in the form of (partial) differential equations or differential-algebraic equations in order to determine output variables from the input variables or input parameters. In doing so, at least one internal variable is determined within the framework of rigorous modeling. The internal variable is characterized by the fact that it is neither an input variable nor an output variable of the rigorous modeling. The internal variable is therefore determined as a by-product in the course of rigorous modeling. In other words, the internal variable is defined in such a way that only part of the complete rigorous modeling is necessary to determine the internal variable (or variables).

[0018] The internal variable is used as the input variable or as input values ​​of a second modeling stage, which, in contrast to rigorous modeling, is data-based. In contrast to purely data-based modeling of the output values ​​of the output variable, data-based modeling within the framework of the method according to the invention can access significantly more extensive and precise information (i.e., the internal variable determined by rigorous modeling).

[0019] A feature of the invention is that both rigorous and data-based modeling are performed during operation. As explained in the introduction, this was unknown in the prior art. Rigorous modeling generally represents a reliable data source for data-based modeling in the case of deterministic technical systems.

[0020] The method is particularly advantageous because the actual output variable of the rigorous modeling does not necessarily have to be calculated. This can result in a significant reduction in the required (computing) resources. Advantageously, the internal variable is chosen so that it can be determined by the rigorous modeling with as little effort as possible. It is particularly advantageous that a modeling user is presented with the option of selecting which internal variable is particularly suitable as an input variable for the subsequent data-based modeling based on certain criteria such as modeling accuracy or computational effort. For this purpose, machine (self-) learning can be provided, which determines the most suitable internal variable for further modeling in an automated or semi-automated manner, for example using an assessment by the user.

[0021] Since the additional use of the internal variable results in a significant gain in information, the data-based model needs to depict fewer complex relationships, allowing for simpler model structures. Especially with regard to dynamics, it may be possible to completely dispense with dynamic components in the data-based model by considering dynamic changes in variables in the rigorous modeling. This can significantly reduce complexity.

[0022] Another major advantage is that by determining and querying the internal variable, easily interpretable intermediate results of the rigorous modeling are available. Thus, by comparing the initial values ​​of the output variable determined by the rigorous modeling with the initial values ​​of the output variable determined by the data-based modeling, the behavior of the data-based modeling can be better interpreted. A supplementary upper bound for the difference between the two variables could also increase user acceptance.

[0023] The collaboration between the modelers of the rigorous modeling and the data analysts (i.e., the modelers of the data-based modeling) also has the advantage that the rigorous modeling continues to be actively used, and the work of the rigorous modelers thus remains recognizable. Subsequent improvements to the rigorous modeling can also be easily incorporated. The aforementioned comparison of the initial values ​​of the output variable determined by the rigorous modeling and the initial values ​​of the output variable determined by the data-based modeling can also be jointly evaluated by the modelers to determine in which areas the rigorous modeling may still contain deviations and uncertainties.

[0024] The possibility of switching to rigorous modeling as a "fall-back" strategy at any time is also an advantage of the method according to the invention. If the data-based modeling is not trusted in a specific situation, the data-based modeling can be easily bypassed. Even if it is clear, e.g. through upstream anomaly detection, that the current situation differs so much from the training data that the data-based modeling would have to extrapolate, such a switch can be effective. To implement the "fall-back" strategy, a query can be made after the modeling to determine whether the data-based modeling should be dispensed with.

[0025] Rigorous modeling advantageously allows for dynamic changes in (process) measuring instruments and variables to be considered, whereas data-based modeling advantageously considers only stationary processes. Thus, the dynamic behavior of the recording devices can be considered within the framework of computer-implemented rigorous modeling. This reduces the load on the computing units involved in data-based modeling.

[0026] Within the framework of data-based modeling, the data can be structured, which can be based on the structuring of the process flow diagrams or the piping and instrumentation flow diagram (P&ID). This structured modeling can potentially allow subprocesses of the industrial process to be optimized even more efficiently.

[0027] Within the scope of an advantageous development of the invention, additional input values ​​are used for the rigorous modeling and / or the further, data-based modeling, which input values ​​include environmental conditions during the implementation of the industrial process or parameters of elements used during the implementation of the industrial process. For example, an ambient temperature or air pressure as an environmental condition, or a material composition of an element used during the implementation of the industrial process, can be used as input values.

[0028] The output variable to be modeled can be a concentration of nitrogen oxide, carbon monoxide, formaldehyde, or sulfur oxide. One goal of operational modeling is to determine the respective proportion of each substance in the total substance quantity.

[0029] The industrial process can involve the combustion of a fuel in a gas turbine, a burner or an engine.

[0030] Preferably, the recorded process values ​​are used as additional input values ​​for further computer-implemented, data-based modeling. This can further increase the accuracy of the data-based modeling.

[0031] The further data-based modeling can be designed as a neural network, a regression or a self-organizing map.

[0032] Within the scope of a preferred development of the invention, output values ​​determined by the computer-implemented rigorous modeling are used as additional input values ​​for the further computer-implemented, data-based modeling.

[0033] Most preferably, the initial values ​​of the output variable modelled during operation are used to optimise the operation of the industrial process and / or to predict maintenance requirements during the implementation of the industrial process.

[0034] The above-stated problem is also solved by a technical module designed to provide and transmit information, as explained above, to a technical system, as explained above. The technical module is preferably designed to offer at least one service as a technical service.

[0035] The previously formulated object is also achieved by a device for controlling and / or monitoring an industrial system, which is designed to carry out a method as explained above.

[0036] The above-described properties, features and advantages of this invention, as well as the manner in which they are achieved, will become clearer and more clearly understood in connection with the following description of the embodiment, which is explained in more detail in connection with the drawings.

[0037] FIG. 1 shows a schematic representation of a method according to the invention. An industrial process 1, such as combustion in a gas turbine, is carried out. In parallel to the implementation of the industrial process 1, i.e. during operation, a concentration of a nitrogen oxide in the gas turbine is to be determined as an output variable Zd within the framework of a modeling exercise. For this purpose, in a first stage, process values ​​and environmental conditions such as an ambient temperature of the industrial process 1 are recorded as input values ​​y by recording devices. These are further processed within the framework of a first computer-implemented rigorous modeling exercise 2. The rigorous modeling exercise 2 determines from the input values ​​y using (differential

[0038] ) Equations that characterize the technical process 1, the concentration of nitrogen oxide in the gas turbine as target value z r .

[0039] In the context of rigorous modeling 2, at least one internal variable x is determined. The internal variable x represents neither an input variable y nor a target variable z r It is a kind of intermediate result or product that results from the implementation of a subset of the rigorous modeling 2 and that can be accessed externally as a variable. For example, it could be the temperature in a gas turbine combustion chamber.

[0040] Both the input values ​​y from the industrial process 1 and the process values ​​of the internal variable x are input into a second, subsequent stage of a further computer-implemented, data-based modeling 3. In contrast to the rigorous, equation-based modeling 2, the further modeling 3 exclusively involves data-based modeling steps. The data-based modeling 3 can, for example, be a neural network or a self-organizing map.

[0041] As a result of the further, data-based modeling 3, the nitrogen oxide concentration is available as the output variable Zd to be determined. This output variable Z, modeled during operation, can be used in the context of an optimization of the industrial process 1 or to predict maintenance requirements during the implementation of the industrial process 1.

[0042] The computing means for the computer implementations of the models 2, 3 can, for example, be implemented in a cloud-based environment and, for example, comprise a personal computer (PC). The computing means can be connected to image means, for example via a bidirectional, wireless connection (NFC, WiFi, Bluetooth, etc.). The image means can be designed as data glasses. The computing means are designed to generate visualization information for an application in an augmented reality and to transmit this to the image means, which then generate a corresponding visual representation.

[0043] An essentially identical procedure is shown schematically in FIG 2. In contrast to the embodiment according to FIG 1, the target value z determined by the rigorous modeling 2 ras another input value into the further data-based modeling 3 . This allows the accuracy of the modeling of the output variable z r be further improved by data-based modeling 3 . However, it should be noted that this will result in the target value z r must actually be determined by the rigorous modeling 2 . This can lead to increased modeling effort . Nevertheless, the improvements achieved through the increased accuracy can compensate for the disadvantage in certain case constellations . It is preferable for a user of the method to have a choice regarding the determination of the target variable z r and the associated inclusion in data-based modelling 3 .

[0044] The invention is not limited to application in the implementation of a single technical process. Rather, it can also be used, for example, in a multi-stage process. For example, the method can be used to determine a nitrogen oxide concentration in a gas turbine with a downstream catalyst. The gas turbine and its nitrogen oxide concentration in the exhaust gas can be easily modeled as stationary (since it has very high dynamics, < 2 seconds). However, the measuring instrument dynamics have a dead time of approximately 2 to 5 minutes due to the measuring method, which can be determined from the physical conditions and introduced as rigorous modeling 2. The downstream catalyst (series connection of gas turbine and catalyst), on the other hand, has significant process dynamics (injection of NH3 into the catalyst), namely approximately 15 to 20 minutes until it settles into a new state.These catalyst dynamics and the instrument dynamics of the (downstream) nitrogen oxide concentration measurement can also be taken into account in the rigorous modeling 2. This structured modeling results in an improved prediction of nitrogen oxide emissions.

[0045] Although the invention has been illustrated and described in detail by the preferred embodiment and the figures, the invention is not limited by the disclosed examples and other variations can be derived therefrom by those skilled in the art without departing from the scope of the invention.

Claims

Patent claims 1. A method for the operational modeling of an output variable of an industrial process (1) which can be used in a process plant, comprising: a) carrying out the industrial process (1), b) acquiring process values ​​(y) originating from the industrial process (1) by suitable acquisition means, c) processing the acquired process values ​​(y) by a computer-implemented rigorous model (2), wherein the acquired process values ​​(y) represent input values ​​of the computer-implemented rigorous model (2), d) using process values ​​of at least one internal variable (x) of the computer-implemented rigorous model (2), which were determined by the computer-implemented rigorous model (2), as input values ​​of a further computer-implemented model (3), which is data-based, e) determining output values ​​of the output variable (Zd) by the further computer-implemented model,data-based modeling (3) ., 2. The method of claim 1, wherein additional input values ​​(y) are used for the computer-implemented rigorous modeling (2) and / or the further computer-implemented, data-based modeling (3), which input values ​​(y) comprise environmental conditions during the execution of the industrial process (1) or parameters of elements used in the execution of the industrial process (1).

3. Method according to claim 1 or 2, wherein the output variable to be modeled (Zd) is a volume flow, a concentration of an inorganic gaseous chlorine compound, an inorganic gaseous fluorine compound, a dust, a nitrogen oxide, a carbon monoxide, a nitrogen dioxide, a nitrogen monoxide, a dinitrogen monoxide, ammonia, a mercury, a methane, ei- represents oxygen, moisture, carbon dioxide, formaldehyde, or sulfur dioxide.

4. The method of claim 3, wherein the industrial process (1) comprises the combustion of waste or fuel in a gas turbine, burner or engine.

5. Method according to one of the preceding claims, wherein the recorded process values ​​(y) are used as additional input values ​​for the further computer-implemented, data-based modeling (3).

6. Method according to any of the preceding claims, wherein the further computer-implemented, data-based modeling (3) is designed as a neural network, a regression or a self-organizing map.

7. Method according to any of the preceding claims, wherein determined initial values ​​(e.g. r) of the computer-implemented rigorous modeling as additional input values ​​for the further computer-implemented, data-based modeling (3).

8. Method according to one of the preceding claims, wherein, within the framework of the computer-implemented rigorous modeling (2), a dynamic behavior of the detection means is taken into account.

9. Method according to one of the preceding claims, wherein the operationally modeled output values ​​of the output variable (Zd) are used for an optimization of the operation of the industrial process (1) and / or for a prediction of a maintenance requirement during the execution of the industrial process (1).

10. Device for controlling and / or monitoring an industrial system, configured to perform a method according to any one of claims 1 to 9.

11. Device according to claim 10, wherein the industrial The system comprises a gas turbine, a burner or an engine, in particular with a downstream catalyst.

12. Device according to claim 11, wherein the output variable (Zd) to be modeled is a volume flow, a concentration of an inorganic gaseous chlorine compound, an inorganic gaseous fluorine compound, a dust, a nitrogen oxide, a carbon monoxide, a nitrogen dioxide, a nitrogen monoxide, a dinitrogen monoxide, ammonia, a mercury, a methane, an oxygen, moisture, a carbon dioxide, a formaldehyde or a sulfur dioxide.

Images