Method for monitoring and / or controlling a chemical plant using a hybrid model - Patents.com

Abstract

The present invention relates to a computer-implemented method for monitoring and / or controlling a chemical plant. In particular, the present invention relates to a computer-implemented method for monitoring and / or controlling a physico-chemical process in a chemical plant, comprising: (a) receiving sensor data related to the physico-chemical process; (b) determining at least one physico-chemical parameter by providing the sensor data to a plant model, the plant model comprising: a mechanistic model including at least two equations each representing a portion of the physico-chemical process; and a data-driven model associated with the mechanistic model, wherein a total number of scalars as output parameters from the data-driven model is less than a number of equations of the mechanistic model; and (c) outputting the at least one physico-chemical parameter determined by the plant model.

Description

[Technical field]

[0001] explanation The present invention relates to a computer-implemented method for monitoring and / or controlling a chemical plant. [Background technology]

[0002] Modern chemical plants are highly optimized to obtain maximum output and minimum by-products. This allows them to provide chemical products at a reasonable low price and generate a minimum of by-products that may affect the environment. Various physicochemical processes are carried out in chemical plants. For example, a particularly efficient method used in chemical plants is the continuous reactor for chemical reactions. Reagents are continuously fed into the reactor while the product is continuously output from the reactor. While some physicochemical values ​​such as temperature and pressure as well as the quality of the product can be relatively easily determined, for example with sensors, many other important physicochemical values ​​such as catalyst degradation cannot be measured directly. However, in order to maximize the efficiency of a chemical plant, it is important to have as detailed information as possible about all physicochemical values ​​in the chemical plant. Ideally, such information is obtainable in real time.

[0003] A very useful way to obtain such "hidden" physicochemical values ​​from chemical plants is to use available sensor data and subject them to physical models. A physical model may calculate physicochemical values ​​that are not easily measurable by applying the laws of physics or physicochemistry. Many such models have been developed. Nevertheless, the models have limitations because not all details in chemical plants are fully understood. To improve these models, it has been suggested to add data-driven models. These are also called black-box models because, in contrast to physical models, they do not easily reveal how they get from inputs to outputs.

[0004] GD Bellos et al. in Chemical Engineering and Processing, volume 44 (2005), pages 505-515, disclose modeling the performance of an industrial hydrodesulfurization reactor using a hybrid neural network approach. The neural network is applied to determine the kinetic parameters of the reaction as well as the reaction enthalpy and hydrogen consumption constant. Thus, the neural network outputs four parameters for one single reaction. This approach works well when sufficient historical data is available to train the neural network. The authors have data from three different plants that operate in the same way. However, in most cases, there is not a lot of historical data available. This is especially true for new plants or plants making specialty products. To make matters worse, in these cases, there is often not even a well-fitting physical model available, so the data-driven model needs to compensate for even more than the known process. Summary of the Invention [Problem to be solved by the invention]

[0005] It was therefore an object of the present invention to provide a method for monitoring and / or controlling a chemical plant, which allows an accurate determination of physicochemical values ​​with a minimum of historical data. It was an object to provide a method that can be easily applied to different chemical plants, even if their production processes are not very well understood mechanistically. The method should be easy to implement and provide highly accurate results while using a minimum of resources. If the operation of the plant starts to deviate from the optimum conditions to allow high product yields and minimal unwanted by-products and greenhouse gas emissions, the results of the method should be available within a short time to allow a rapid adjustment of the plant. [Means for solving the problem]

[0006] The object of the present invention is to provide a computer-implemented method for monitoring and / or controlling a physico-chemical process in a chemical plant, comprising the steps of: (a) receiving sensor data relating to a physicochemical process; (b) determining at least one physicochemical parameter by providing the sensor data to a plant model, the plant model comprising: - a mechanistic model containing at least two equations, each of which represents a part of a physicochemical process; and - a data-driven model associated with the mechanistic model, the data-driven model being trained using a training dataset based on a set of historical data including sensor data and physicochemical parameters related to a chemical reaction, wherein a total number of scalars as output parameters from the data-driven model is less than a number of equations of the mechanistic model; determining whether (c) outputting at least one physicochemical parameter determined by the plant model; and This is accomplished by a computer-implemented method comprising:

[0007] The invention further relates to a non-transitory computer-readable data medium storing a computer program comprising instructions for carrying out the steps of the method according to the invention.

[0008] The invention further relates to the use of the physico-chemical parameters obtained by the method according to the invention for monitoring and / or controlling a chemical plant.

[0009] The present invention relates to a production monitoring and / or control system for monitoring and / or controlling a physico-chemical process in a chemical plant, comprising: (a) an input configured to receive sensor data related to a physicochemical process; (b) a processor configured to determine at least one physicochemical parameter by providing the sensor data to a plant model, the plant model comprising: - a mechanistic model containing at least two equations, each of which represents a part of a physicochemical process; and at least one data-driven model associated with the mechanistic model, the at least one data-driven model being trained using a training dataset based on a set of historical data including sensor data and physicochemical parameters related to a chemical reaction, wherein a total number of scalars as output parameters from the at least one data-driven model is less than a number of equations of the mechanistic model; a processor including: (c) an output configured to output at least one physicochemical parameter determined by the plant model; and The present invention further relates to a production monitoring and / or control system comprising:

[0010] The present invention relates to a method for training a plant model suitable for determining at least one physico-chemical parameter from sensor data of a physico-chemical process in a chemical plant, the method comprising the steps of: (a) receiving a training dataset based on a set of historical data including sensor data and physico-chemical parameters related to a physico-chemical process; (b) training a plant model by adjusting a parameterization according to the training data set, the plant model comprising: - a mechanistic model containing at least two equations, each of which represents a part of a physicochemical process; and at least one data-driven model associated with the mechanical model, wherein a total number of scalars as output parameters from the at least one data-driven model is less than a number of equations of the mechanical model; and (c) outputting the trained plant model; The present invention further relates to a method comprising the steps of:

[0011] Brief description of some figures of the drawing To easily identify the discussion of any particular element or operation, the leftmost digit or digits of a reference number refer to the figure number in which that element is first introduced. [Brief description of the drawings]

[0012] [Figure 1] 1 illustrates a method and system of the present invention. [Diagram 2] 1 shows an example of a plant model. [Diagram 3] 13 illustrates another embodiment of a plant model. [Figure 4] 13 illustrates another embodiment of a plant model. [Diagram 5] An example of determining sensitivity for different plant model modifications is provided. [Figure 6] 1 shows an example of a production process in which the present invention can be used. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] The present invention relates to a method for monitoring and / or controlling a physico-chemical process in a chemical plant. Figure 1 shows a possible implementation of the invention. Sensor data from a plant 108 is received by an input 102. The sensor data is provided to a processor 104 programmed to execute a plant model. This plant model uses the sensor data as input and has physico-chemical parameters as output. The physico-chemical parameters are output by an output 106.

[0014] "Monitoring" refers to the observation and recording of any operating condition of a chemical plant. Operating conditions include internal parameters such as reactor temperature, pressure, power consumption, input or output material flow, agitator rotation speed, valve status, vapor concentration in the air within the plant, number of people inside the plant, and those parameters that are of concern only within the plant. Operating conditions also include external parameters such as parameters related to any exchange of the plant with its environment, such as chemical vapors, heat, sound, vibration, light radiation, etc. Recording can mean storing raw data on a permanent data storage device or preparing documentation in a format required by a company or government agency.

[0015] "Control" refers to taking any action to change the operating state of a chemical plant. Actions can be direct, such as by changing the state of a valve, changing the temperature by adding additional heating or cooling. Actions can also be indirect, such as by prompting an operator to take an action such as changing a filter or adjusting throughput.

[0016] "Physicochemical process" refers to any process involving the manipulation or modification of at least one substance, such as a compound or composition. Physicochemical processes include chemical reactions, purification such as distillation, crystallization, filtration, centrifugation, decantation, floatation, forming such as mixing, spray drying, coextrusion, coating, or shape modification processes such as grinding, molding, agglomeration, extrusion.

[0017] A "chemical reaction" refers to a physicochemical process involving the chemical transformation of one set of chemical species into another set. A chemical reaction may in principle include one elementary reaction. However, in practice, most chemical reactions include two or more elementary reactions. A chemical reaction may include sequential or parallel elementary reactions or both. An example of a chemical reaction involving a series of elementary reactions is a condensation reaction in which a nucleophilic species is first added to an electrophilic species as the first elementary reaction, followed by the removal of a smaller species such as water as the second elementary reaction. An example of a chemical reaction involving several elementary reactions in parallel is a combustion reaction in which a chemical species reacts with oxygen to form a variety of different partially oxidized species.

[0018] A chemical reaction can be operated homogeneously or heterogeneously. A homogeneous chemical reaction involves one phase, e.g. a gas phase or a liquid phase, e.g. a solution. A heterogeneous chemical reaction involves at least two phases. The at least two phases can be of different states of matter, e.g. one phase is solid and one phase is liquid, or one phase is solid and the other phase is gas, or one phase is liquid and the other phase is gas. The at least two phases can be of the same state of matter that is immiscible, e.g. two immiscible liquid phases or two immiscible solid phases.

[0019] Chemical reactions can be operated continuously or discontinuously and may be referred to as batch chemical reactions. In a continuous chemical reaction, reagents are continuously fed into a reactor where the reaction occurs and simultaneously the product is continuously output from the reactor. In a discontinuous chemical reaction, the reactor is charged with reagents, then the reaction occurs, after which the product is collected from the reactor. The reactor may be cleaned and then charged again with fresh reagents.

[0020] "Elementary reaction" refers to a chemical reaction in which one or more chemical species react directly to form a product in a single reaction step without intermediates that can be observed or further isolated. An elementary reaction can usually be described as a reaction that has a single transition state.

[0021] "Chemical plant" refers to any technological infrastructure used for industrial purposes for the manufacture, production or processing of one or more chemical products, i.e. for carrying out chemical reactions to produce chemical compounds, for producing preparations by mixing chemical compounds, for increasing the purity of chemical compounds, for obtaining chemical compounds by recycling waste products, for making chemical compounds into different forms or for packaging chemical compounds or preparations containing chemical compounds.

[0022] The infrastructure of a chemical plant may include equipment or process units such as any one or more of: heat exchangers, columns such as fractionation towers, furnaces, reaction chambers, cracking units, storage tanks, extruders, pelletizers, precipitators, blenders, mixers, cutters, curing tubes, vaporizers, filters, sieves, pipelines, stacks, filters, valves, actuators, mills, transformers, conveying systems, circuit breakers, machinery, e.g., large rotating equipment such as turbines, generators, grinders, compressors, industrial fans, pumps, transport elements such as conveyor systems, motors, and the like.

[0023] Additionally, chemical plants typically include a number of sensors and at least one control system for controlling at least one parameter related to a process within the plant, i.e., the process parameter. Such control functions are typically performed by a control system or controller in response to at least one measurement signal from at least one of the sensors. The controller or control system of the plant may be implemented as a distributed control system ("DCS") and / or a programmable logic controller ("PLC").

[0024] Thus, at least some of the equipment or process units of the chemical plant may be monitored and / or controlled to produce one or more of the industrial products. The monitoring and / or control may further be performed to optimize the production of the one or more products. The equipment or process units may be monitored and / or controlled via a controller, such as a DCS, in response to one or more signals from one or more sensors. In addition, the plant may further include at least one PLC for controlling some of the processes. A chemical plant may typically include multiple sensors that may be distributed within the chemical plant for monitoring and / or control purposes. Such sensors may generate large amounts of data. The sensors may or may not be considered part of the equipment. As such, production, such as the production of chemicals and / or services, may be a data-heavy environment. Thus, each chemical plant may result in a large amount of process-related data.

[0025] Those skilled in the art will appreciate that chemical plants typically include instrumentation that may include different types of sensors. Sensors may be used to measure one or more process parameters and / or equipment operating conditions or parameters associated with equipment or process units. For example, sensors may be used to measure process parameters such as flow rate in a pipeline, level in a tank, temperature in a furnace, chemical composition of gas, and some sensors may be used to measure vibration of a grinder, speed of a fan, opening of a valve, corrosion in a pipeline, voltage across a transformer, and the like. The differences between these sensors may not only be based on the parameter they sense, but also on the sensing principle that each sensor employs. Some examples of sensors based on the parameter they sense may include temperature sensors, pressure sensors, radiation sensors such as optical sensors, flow sensors, vibration sensors, displacement sensors, and chemical sensors such as those that detect specific substances such as gases. Examples of sensors that employ different sensing principles may include impedance sensors such as piezoelectric sensors, piezoresistive sensors, thermocouples, capacitance sensors, and resistance sensors, and the like.

[0026] The multiple chemical plants may form a larger production unit. The term "multiple chemical plants" as used herein is a broad term and should be given its ordinary and customary meaning to those skilled in the art and should not be limited to a special or customized meaning. The term may specifically refer to, but is not limited to, a complex of at least two chemical plants having at least one common industrial purpose. Specifically, the multiple chemical plants may include at least two, at least five, at least ten or even more chemical plants that are physically and / or chemically combined. The multiple chemical plants may be combined such that the chemical plants forming the multiple chemical plants may share one or more of the value chains, extracts and / or products. The multiple chemical plants may also be referred to as a complex, a complex site, a Verbund or a Verbund site. Furthermore, the value chain production of the multiple chemical plants through various intermediate products to the final product may be distributed in various locations, such as various chemical plants, or may be integrated into a Verbund site or chemical park. Such a Verbund site or chemical park may be or may include one or more chemical plants, where products produced in at least one chemical plant may serve as feedback to another chemical plant.

[0027] "Sensor data" refers to any data representative of the operating condition of a production plant or part thereof, measured by a sensor in a chemical plant. The sensor data may be received directly from the sensor. Typically, the sensor data is collected by a digital signal controller or a programmable logic controller in the chemical plant and further transmitted from there. The sensor data may be conditioned, for example by a calibration system, before being transmitted. The sensor data from the chemical plant may also be stored on a storage medium, for example on a hard drive or on a database in a cloud system. Thus, the sensor data may be retrieved from such a storage medium for the purposes of the present invention.

[0028] The sensor data may include any measurable physicochemical value such as temperature, pressure, pH, concentration or partial pressure of compounds such as oxygen or moisture, flow rate of reagents of the reaction mixture in the reactor or the product after the reactor, agitator speed, viscosity, turbidity, etc. Typically, the sensor data also includes the location of the sensor, especially if two or more sensors measure at different locations of the instrument. A typical example is a pressure sensor at the inlet of the reactor and a pressure sensor at the outlet of the reactor. The sensor data may also include time information, i.e. the time when the sensor collected the physicochemical information, which may also be called a timestamp.

[0029] The sensor data is related to the physicochemical process being monitored and / or controlled. The term "related" should be understood broadly, i.e. any information of the sensor that has an effect on or correlates to the state of the physicochemical process.

[0030] The sensor data may be received directly from sensors in the chemical plant or from a data storage medium. The sensor data on the data storage medium may be recorded sensor data or manipulated sensor data. The reason for manipulating the sensor data may be to simulate deviations and analyze the effects on physicochemical parameters with the aim of controlling the physicochemical process if such a situation actually occurs. An example may be a change in heat supply, where it may be desired to analyze whether such a heat supply can be compensated by an increase in pressure or a change in flow rate.

[0031] However, for controlling a chemical plant, the sensor data is preferably received in real time, preferably directly from sensors in the chemical plant. Real time relates to low latency, i.e. latency less than 10 seconds or even less than 1 second. Typically, the lower the latency, the higher the accuracy of controlling the chemical plant.

[0032] The method according to the invention includes (b) determining at least one physicochemical parameter by providing the sensor data to a plant model.

[0033] "Physicochemical parameters" refer to information that characterizes a physicochemical process that is theoretically measurable. In practice, however, this is usually not possible directly by a sensor, for example because suitable sensors do not exist, sensors cannot be placed in a position where the information can be detected, or such measurements are economically disadvantageous. Physicochemical parameters may refer to chemical species, for example chemical composition, concentration, pressure, purity, viscosity, turbidity. Physicochemical parameters, when used, may also refer to catalysts, for example chemical composition, concentration, pressure, purity, activity, age, surface area. Physicochemical parameters may also refer to equipment, for example total pressure, temperature, flow rate in a part, for example a pipe, pressure drop along a part, amount and / or rate of deposition of insoluble material on the wall, also called fouling, for example heat flow rate in a heat exchanger.

[0034] Typically, the physicochemical parameters are determined which are most relevant for monitoring and / or controlling the chemical reaction. Particularly relevant physicochemical parameters for monitoring and / or controlling are reaction yield, catalyst activity and equipment fouling. In some cases, it is sufficient to determine one physicochemical parameter. In other cases, it is beneficial to determine two or more reaction parameters, for example at least two, three, five or ten. In this way, a more detailed insight into the chemical reaction can be obtained, so that sufficiently suitable measurements can be utilized to control the chemical reaction. The determination is made by providing sensor data to a plant model.

[0035] "Plant model" refers to a model that mathematically describes a physico-chemical process or processes in a chemical plant. The plant model receives sensor data as input and outputs physico-chemical parameters. The plant model includes a mechanistic model and a data-driven model. Thus, the plant model may be referred to as a hybrid model.

[0036] "Mechanistic model" refers to a model based on any one or more of the fundamental laws of natural sciences, such as the principles of physics, chemistry, biochemistry, heat and mass balance. Such models therefore represent these principles using equations. Mechanistic models may include linear or nonlinear ordinary differential equations, linear or nonlinear partial differential equations, linear or nonlinear algebraic equations, or linear or nonlinear differential-algebraic equations. Such equations relate to physicochemical processes.

[0037] A typical example for a mechanistic model is a chemical kinetics model that models a physicochemical process. Essentially, such a model is composed of ordinary differential or differential algebraic equations that describe the dynamics of chemical species being consumed or produced by a set of chemical reactions. A system of ordinary differential or differential algebraic equations is usually composed of rate laws, which are algebraic equations that describe the rate at which chemical species are consumed or produced in a reaction. Such algebraic equations typically depend on the concentrations of chemical species, the temperature in a given reaction, and constants, which are usually temperature dependent. Furthermore, certain invariants, such as the law of conservation of mass, can also be expressed as algebraic equations in such mechanistic models.

[0038] It may be known which mechanistic model best fits a certain physicochemical process. In this case, the selection of an appropriate mechanistic model is straightforward. However, if it is not known which mechanistic model best fits a physicochemical process, a set of mechanistic models for similar physicochemical processes may be selected. Sometimes there may be no similar physicochemical process available, perhaps because the underlying mechanism is still unknown or suitable information is not available for different reasons. In this case, it may be sufficient to select an arbitrary mechanistic model from a model library that includes various mechanistic models for known physicochemical processes. Obviously, such an arbitrary mechanistic model will not fit a given physicochemical process very well. However, the result may be sufficient for undemanding purposes, since the associated data-driven model may compensate for at least part of the deviation. Alternatively, different mechanistic models may be randomly selected and tried in turn to select the mechanistic model that best fits the physicochemical process. Such a selection may be automated. Thus, a mechanistic model can be selected from a model library, for example, by a computer program randomly selecting several mechanistic models, applying them in turn to the physico-chemical process, determining how well the mechanistic models fit the physico-chemical process, and selecting the best-fitting mechanistic model.

[0039] "Data-driven model" refers to a mathematical model parameterized according to a training data set to reflect physicochemical processes such as reaction kinetics of a production plant. The training data set may include sensor data and physicochemical parameters obtained from experiments or previous production runs. In contrast to mechanistic models that are derived purely using physicochemical laws, data-driven models may make it possible to describe relationships that are difficult or even impossible to model by physicochemical laws. Data-driven models are set up without reflecting any underlying natural physical laws. These are taken into account only by using correlations in the data.

[0040] The data-driven model is preferably a data-driven machine learning model. The data-driven model may be a linear or polynomial regression, a decision tree, a random forest model, a Bayesian network, a support vector machine or preferably an artificial neural network.

[0041] According to the present invention, the plant model includes a mechanistic model including at least two equations each representing a part of a physicochemical process. In case of chemical reactions, each equation may represent an elementary reaction of the chemical reaction, or each equation may represent several elementary reactions, for example by approximating with one hypothetical elementary reaction. In case of distillation, each equation may represent the evaporation and condensation of one compound. The plant model further includes at least one data-driven model associated with the at least one mechanistic model. The term "associated" means that there is a data exchange between the mechanistic model and the data-driven model. For example, the output of the data-driven model may be used as an input for the equations of the mechanistic model, or the output of the equations of the mechanistic model may be used as an input for the data-driven model. It is possible that the output of the data-driven model is used in more than one equation of the mechanistic model. In this case, the data-driven model may output one scalar as an output parameter that is used as an input in more than one equation of the mechanistic model, for example two or three equations. It is further possible that one output parameter of the data-driven model is used in all equations of the mechanistic model. When the output of the mechanistic model is used as an input for a data-driven model, one output parameter of the mechanistic model can be used in one data-driven model or more than one, for example two, three data-driven models. It is further possible that one output parameter of the mechanistic model is used in all data-driven models. It is also possible that the mechanistic model has more than one scalar as an output parameter, each output parameter being used in a different data-driven model. It is also possible that the mechanistic model has more than one output parameter, some of which are used in more than one data-driven model, and some of which are used in only one data-driven model. It is also possible that the output of the mechanistic model is used as an input for a data-driven model, and the output of this data-driven model is used again as an input for the mechanistic model, thus forming a feedback loop. This can be useful for physicochemical processes, where some of the products are recycled by using it again as a reagent.

[0042] According to the present invention, the total number of scalars as output parameters from at least one data-driven model is less than the number of equations of the mechanism model. The term "total number" means the sum of all scalars as output parameters of all data-driven models. In this context, the output parameters may be scalars, so the number of scalars as output parameters is equal to the number of output parameters. The output parameters may be vectors or matrices. In this case, the total number of scalars as output parameters refers to the number of components or elements of the vector or matrix.

[0043] The plant model includes at least one data-driven model. The plant model may include one data-driven model, or it may include two or more, for example, two or three data-driven models. The two or more data-driven models may all be identical to each other or different, for example, the plant model may include polynomial regression and artificial neural network. When one data-driven model is used, the total number of scalars as output parameters is equal to the number of output physicochemical parameters of the data-driven model. When two or more data-driven models are used, the number of scalars as output parameters per data-driven model is added to arrive at the total number of scalars as output parameters. Figures 2, 3 and 4 show some examples of how the plant model may look when the physicochemical process is represented by a mechanistic model including three equations. For illustrative purposes, the rounded square represents a data-driven model with only one output parameter. For data-driven models with two or more output parameters, the respective number of rounded squares shall be displayed.

[0044] 2 shows a plant model 204 that includes a data-driven model 206 that receives sensor data 202 as an input. The output of the data-driven model 206 is used as an input for a mechanism model 208, which may use the sensor data 202 as an additional input. For example, the data-driven model 206 may output correction constants for equation 210. Equations 212 and 214 use only the sensor data 202 as an input. The plant model 204 outputs physicochemical parameters 216.

[0045] 3 shows another plant model 302 that includes two data-driven models 312 and 314 that receive as input the outputs of equations 306 and 310, respectively, and output a physicochemical parameter 316. For example, equations 306 and 310 may output a physicochemical parameter that is corrected by data-driven models 312 and 314. Equation 308 is not associated with a data-driven model.

[0046] 4 shows another plant model 404. It includes one data-driven model 406 that receives sensor data as input. Its output is used as input in all equations 410, 412, and 414. Equations 410, 412, and 414 output physicochemical parameters 416.

[0047] A data-driven model usually uses sensor data as input, sometimes in addition to the output of the equations of the mechanism model. In order to reduce the need for a large amount of historical data and at the same time have a highly accurate plant model, it is advantageous to reduce the input of the data-driven model to a minimum. As a result, only a part of the sensor data is used as an input for the data-driven model. The appropriate selection of the sensor data used as an input for the data-driven model may include one or more of the following options:

[0048] i) Subset selection by identifying a data-driven model using a subset of input parameters that has an accuracy close to that of the data-driven model using the full input parameters. Several techniques are known in the literature to efficiently identify such a subset.

[0049] ii) Regularization or shrinkage techniques, typically applicable to neural networks and linear regression based methods, in which the contributions of some of the input parameters are shrunk towards or set to zero. This is typically achieved by introducing a penalty into the loss function of the data-driven model.

[0050] iii) Dimensionality reduction is a projection method, e.g., principal component analysis, where input parameters are projected into a reduced dimensional space resulting in "derived input parameters" that are then used in data-driven models. An introduction to these techniques can be found in James, Gareth, et al. An introduction to statistical learning. Vol. 112. New York: springer, 2013.

[0051] Essentially, all these approaches reduce the complexity of the data-driven model by performing a selection of the input parameters associated with the data-driven model, either by eliminating less relevant variables or by finding a lower dimensional "derived input parameter" space.

[0052] The plant models used in the present invention are typically more accurate than the mechanistic models alone, but require less historical data to train the data-driven models compared to traditional hybrid models, especially when those output parameters of the data-driven models with the highest sensitivity are selected.

[0053] "Sensitivity" refers to the effect that an output parameter of a data-driven model has on a physicochemical parameter, i.e., the relative difference of the physicochemical parameter, when such an output parameter changes, i.e., increases or decreases. In some cases, it is sufficient to select only the output parameter with the highest sensitivity. In other cases, it may be necessary to use two or three output physicochemical parameters with the highest sensitivity. Usually, the number of output parameters selected is a trade-off between the available historical data and the required accuracy of the reaction parameters.

[0054] In some cases, an expert in reaction modeling may be able to perform a direct selection of the output parameters of the data-driven model. However, in many cases, due to the complexity of the situation, the sensitivities must be determined systematically before it is possible to select suitable output parameters. Figure 5 shows a method for determining the sensitivity of the output parameters. Starting from a physicochemical process scheme 522 containing details of the physicochemical process in a chemical plant, a plant model 502 is generated. This plant model 502 includes a mechanistic model that includes equations for each part (504, 506, 508) of the physicochemical process. Based on the plant model 502, derived plant models (510, 518, 526) are generated by associating the data-driven model with the mechanistic model, and in each plant model (510, 512, 514), the data-driven model is associated with the mechanistic model in a different way. In this example, the output of the data-driven model is used as an input for one of the equations of the mechanistic model. Obviously, more options are possible, such as using the output of the data-driven model for more than one equation of the mechanism model, or using a data-driven model that uses the output of one or more equations of the mechanism model as input. For each plant model (510, 512, 514), a data-driven model is trained using historical data. The validation data is then used to determine the outputs (524, 526, 528) for each plant model (510, 512, 514). The outputs of the data-driven model for each plant model (510, 512, 514) are changed and the change in the outputs is determined. The relative difference indicates the sensitivity. The plant model with the highest sensitivity found can be used for the method of the present invention. In this example, plant model 502 exhibits low sensitivity 530, plant model 512 exhibits high sensitivity 532, and plant model 514 exhibits medium sensitivity 534.

[0055] The plant model may further include an integrated model that integrates the output of the mechanistic model and / or the data-driven model into physicochemical parameters. This is particularly useful for chemical reactions that include a series of reaction steps, i.e. the product of one reaction step is a reagent of the next reaction step. The integrated model is usually based on boundary conditions that are evident from the laws of nature. A typical boundary condition is a mass balance. No mass is created or destroyed by a chemical reaction, only chemical species are converted into each other. Other boundary conditions can be minimum or maximum values ​​for certain parameters, e.g. concentrations cannot be negative or pressures cannot differ significantly in open connected volumes.

[0056] The plant model is trained using a training dataset based on a set of historical data including sensor data and physicochemical parameters related to chemical reactions. "Historical data" refers to a dataset including at least sensor data and physicochemical parameters, each dataset being associated with a single physicochemical process run. Thus, each dataset includes data associated with a physicochemical process run within a predefined time period. In the case of a batch process, such predefined time period may be from the start to the end of one batch run. In the case of a continuous process, a characteristic period may be selected, for example the time from when a reactor is loaded with catalyst until it needs to be replaced with new catalyst. The historical data may be obtained from an existing plant to be monitored or controlled. However, the historical data may also originate from a laboratory, a pilot plant, or a similar plant. Historical data from one or more of these may be available.

[0057] Training a plant model is typically done by adjusting the parameterization according to a training data set. Adjusting the parameterization in this context means varying the parameters in the data-driven model included in the plant model so that the output of the plant model most closely resembles the reaction parameters of the training set. Depending on the type of data-driven model, different ways of doing so are known and well documented in the literature.

[0058] The method according to the invention further comprises (c) outputting at least one physicochemical parameter determined by the plant model. Outputting may mean writing the physicochemical parameter on a non-transitory data storage medium, for example in a monitoring or control file, displaying it on a user interface, for example on a screen, or both. The method according to the invention may be called a soft sensor or virtual sensor that measures the physicochemical parameters indirectly by computationally deriving them from observables represented by the sensor data.

[0059] It is also possible to output the physicochemical parameters through the interface to a control system. Such a control system may receive the physicochemical parameters and, based on such physicochemical parameters, change the settings of the equipment in the chemical plant where the physicochemical process is performed. As an example, the plant model determines a reduction in catalyst activity of a certain value compared to the maximum catalyst activity. The control system may receive the catalyst activity and cause the input valve to reduce the flow rate of the reagents through the reactor. In this way, the reagents remain in the vicinity of the catalyst longer, thus compensating for the reduced catalyst activity. Thus, the catalyst may react completely to produce the desired product with high yield and good quality.

[0060] The present invention further relates to a non-transitory computer-readable data medium storing a computer program comprising instructions for performing the steps of the method according to the present invention. A "computer-readable data medium" refers to any suitable data storage device or computer-readable memory on which one or more sets of instructions (e.g., software) embodying any one or more of the methodologies or functions described herein are stored. The instructions may also reside completely or at least partially in the main memory and / or in the processor during their execution by the computer, the main memory and the processing device, which may constitute a computer-readable storage medium. The instructions may further be transmitted or received over a network via a network interface device. Computer-readable data media include, for example, a hard drive on a server, a USB storage device, a CD, a DVD or a Blu-ray disc. The computer program may include all the functionality and data necessary for the execution of the method according to the present invention, or it may provide an interface to allow parts of the method to be processed on a remote system, for example a cloud system.

[0061] The invention further relates to a production monitoring and / or control system for monitoring and / or controlling a physico-chemical process in a chemical plant. Such a system is configured to carry out the method according to the invention. Therefore, all definitions, examples and preferred embodiments given for the method also apply to the system.

[0062] The system according to the invention comprises an input configured to receive sensor data related to a physicochemical process. Such an input may comprise an interface for receiving the sensor data. The input may receive the sensor data locally or remotely, for example via an interface to a telecommunication system such as the Internet. The input may receive the sensor data directly from the sensor or via a storage medium including a programmable logic controller, a distributed control system or a cloud service. It is further possible that the system is part of a distributed control system.

[0063] The system further includes a processor configured to determine at least one physicochemical parameter. The processor may be a local processor including a central processing unit (CPU), and / or a graphic processing unit (GPU), and / or an application specific integrated circuit (ASIC), and / or a tensor processing unit (TPU), and / or a field programmable gate array (FPGA). The processor may also be an interface to a remote computer system, such as a cloud service. EXAMPLES

[0064] Working Example To further illustrate the invention, Figure 6 shows an example of a chemical reaction in a chemical plant producing phenol and acetone in two stages from benzene and propene. Benzene and propene are mixed from a benzene supply 602 and a propene supply 604 and injected into a tubular reactor 606 controlled by a valve 618. The tubular reactor 606 has a solids bed containing a Friedel-Crafts alkylation catalyst that converts the benzene to cumene 608. The cumene 608 is fed with oxygen 610 into a tubular reactor 612 controlled by a valve 620. The tubular reactor 612 also has a solids bed containing an oxidation catalyst that converts the cumene to phenol and acetone. The product flow is controlled by a valve 622. The phenol and acetone are collected and purified. The reactor is equipped with a temperature sensor 624 and a pressure sensor 626 that measure the temperature and pressure and transfer these values ​​to a distributed control system 628. Valves 618, 620 and 622 are equipped with sensors that measure the gas flow so that the partial pressure of each reagent can be determined. The corresponding values ​​are also transferred to the distributed control system 628.

[0065] Thus, the sensor data collected by the distributed control system 628 includes the total mass flow rate per unit area (Gz), the partial pressure of propene (p pr ), the partial pressure of benzene (p Bz ), partial pressure of oxygen (p O2), the partial pressure of cumene (p Cm ), the temperature in the tubular reactor 606 (T1), and the temperature in the tubular reactor 612 (T2). The distributed control system 628 transfers the sensor data to a processor 630 that executes a plant model. The plant model uses these sensor data and the reactor configuration as input parameters and the parameter f NN The neural network has one hidden layer. The parameter f NN is used as input for a mechanistic model that includes two equations:

[0066] The mechanistic model contains one equation for the rate constant of each elementary reaction. For the reaction of benzene and propene to cumene, the following equations are used, with k1 and E A1 is a constant found in the literature for this reaction.

[0067]

number

[0068] For the reaction in tubular reactor 612 to form phenol and acetone from cumene, the following equations are used, k2 and E A2 is a constant found in the literature for this reaction.

[0069]

number

[0070] The physicochemical parameters of phenol and acetone yields were determined by using an integrated model obtained from the mass balances separately for each reaction step. i is the mass fraction of component i, and ρ cat is the packing density of the catalyst, M w,i is the molecular mass of component i, and v i is the stoichiometric coefficient of component i.

[0071]

number

[0072] The yields of phenol and acetone are forwarded from processor 630 to distributed control system 628. If the yields decrease, indicating reduced catalyst activity, distributed control system 628 may reduce the gas flow rates of benzene and propylene by operating valve 618. Alternatively, distributed control system 628 may reduce the gas flow rates of cumene and oxygen by operating valve 620. In this way, the time that benzene and propylene or cumene and oxygen are in contact with the catalyst is increased, which may lead to higher conversions and restore product yields.

Claims

1. A computer-implemented method for monitoring and / or controlling a physico-chemical process in a chemical plant, comprising: (a) receiving sensor data related to the physico-chemical process; (b) determining at least one physico-chemical parameter by providing the sensor data to a plant model, wherein the plant model comprises: i. a mechanism model including at least two equations each representing a part of the physico-chemical process; and ii. a data-driven model associated with the mechanism model, trained using a training data set based on a set of historical data including sensor data related to chemical reactions and physico-chemical parameters, wherein the total number of scalar output parameters from the data-driven model is less than the number of equations of the mechanism model; (c) outputting the at least one physico-chemical parameter determined by the plant model. A computer-implemented method as claimed in claim 1.

2. The computer-implemented method as claimed in claim 1, wherein the at least one physico-chemical parameter includes at least one of reaction yield, catalyst activity or equipment fouling.

3. The computer-implemented method as claimed in claim 1, wherein the sensor data includes temperature, pressure and reagent flow rate.

4. The computer-implemented method as claimed in claim 1, wherein the output of the data-driven model is used as an input for at least one equation of the mechanism model.

5. The computer-implemented method as claimed in claim 1, wherein the data-driven model is an artificial neural network.

6. The computer-implemented method as claimed in claim 1, wherein the at least one physico-chemical parameter is output to a control system capable of changing the settings of equipment in the chemical plant in which the physico-chemical process is performed based on the physico-chemical parameter.

7. The computer-implemented method as claimed in claim 1, wherein the output parameters from the at least one data-driven model are selected based on the sensitivity of the output parameters, the sensitivity being the relative difference of the physico-chemical parameters when the output parameters of the data-driven model change.

8. ​ The computer-implemented method according to claim 1, wherein the data-driven model uses a part of the sensor data determined by one or more of subset selection, regularization, and dimensionality reduction.

9. A non-transitory computer-readable data medium storing a computer program including instructions for performing the steps of the method according to any one of claims 1 to 8.

10. Use of the physicochemical parameters obtained by the method according to any one of claims 1 to 8 for monitoring and / or controlling a chemical plant.

11. A production monitoring and / or control system for monitoring and / or controlling a physicochemical process in a chemical plant, (a) an input configured to receive sensor data related to the physicochemical process, and (b) a processor configured to determine at least one physicochemical parameter by providing the sensor data to a plant model, wherein the plant model i. a mechanism model including at least two equations each representing a part of the physicochemical process, and ii. at least one data-driven model associated with the mechanism model, trained using a training data set based on a set of historical data including sensor data and physicochemical parameters related to a chemical reaction, and the total number of scalars as output parameters from the at least one data-driven model is less than the number of equations of the mechanism model, at least one data-driven model including a processor, and (c) an output configured to output the at least one physicochemical parameter determined by the plant model including a production monitoring and / or control system.

12. The production monitoring and / or control system according to claim 11, which is part of a distributed control system of the chemical plant or is connected to the distributed control system of the chemical plant.

13. The production monitoring and / or control system according to claim 11 or 12, wherein the sensor data is received from sensors within the chemical plant.

14. A method of training a plant model suitable for determining at least one physicochemical parameter from sensor data of a physicochemical process in a chemical plant, receiving a training data set based on a set of historical data including sensor data and physicochemical parameters related to the physicochemical process; training a plant model by adjusting parameterization according to the training data set, the plant model comprising: i. a mechanism model including at least two equations each representing a part of the physicochemical process; and ii. at least one data-driven model associated with the mechanism model, wherein the total number of scalars as output parameters from the at least one data-driven model is less than the number of equations of the mechanism model; training, including; outputting the trained plant model; A method including.

Images