Method for monitoring or controlling a chemical production process

Abstract

The invention is in the field of monitoring or controlling a chemical production process. It relates to computer-implemented method for monitoring and / or controlling a chemical production process comprising (a) receiving sensor data related to the chemical production process, (b) determining an operational instruction related to the chemical production process by providing the sensor data to a trained model, wherein the trained model contains a first partial model representing a first part of the chemical production process and a second partial model representing a second part of the chemical production process, wherein the first partial model is trained with historical data related to the first part of the chemical production process and the second partial model is trained with historical data related to the second part of the chemical production process and wherein the model containing the first trained partial model and the second trained partial model is retrained with historical data related to the chemical production process, and (c) outputting the operational instruction.

Description

[0001] Method for Monitoring or Controlling a Chemical Production Process

[0002] The invention is in the field of monitoring or controlling a chemical production process. The invention relates to a method for monitoring and / or controlling a chemical production process, a non-transitory computer-readable data medium storing a computer program including instructions for executing steps of the method for monitoring and / or controlling a chemical production process, the use of an operational instruction obtained by the method for monitoring and / or controlling a chemical production process, and a monitoring and / or controlling system.

[0003] Background

[0004] Modern chemical production plants are optimized to produce a maximum of product with the minimum of raw materials and energy. In addition to the experience of the human plant operators, computer-based models are used to keep the production processes close to their optimum. For simple production processes it is possible to use a single model which is trained with data relating to the overall process. However, for more complex processes, such an approach often does not yield sufficiently accurate results.

[0005] M. Bubel et al. disclose in Chemie Ingenieur Technik, volume 93 (2021), pages 1987-1997 a modular approach by using partial models each related to a part of a process. These partial models are separately trained and subsequently interconnected. This approach works well for not too complicated production processes, for example the pressure swing distillation shown by M. Bubel et al. However, if the production process becomes more complicated, the approach does not yield sufficiently accurate results.

[0006] WO 2021 123 385 A1 discloses a method for modelling a production plant involving sub-mod- els. However, the sub-models are separately trained. For complex production processes, insufficiently accurate results may be obtained.

[0007] Summary

[0008] It was therefore the objective of the present invention to provide a method for monitoring and / or controlling a chemical production process which is applicable to complex production processes and still renders results of high accuracy. The method was aimed at being suitable for a broad variety of production processes, even for those including multiple cycles or even nested cycles. The method was aimed at being applicable to production processes for which a limited physical chemical understanding is available. The method should further require a small amount of historical data.

[0009] These objectives were achieved by the present invention. In one aspect, the present invention relates to a computer-implemented method for monitoring and / or controlling a chemical production process comprising

[0010] (a) receiving sensor data related to the chemical production process,

[0011] (b) determining an operational instruction related to the chemical production process by providing the sensor data to a trained model, wherein the trained model contains a first partial model representing a first part of the chemical production process and a second partial model representing a second part of the chemical production process, wherein the first partial model is trained with historical data related to the first part of the chemical production process and the second partial model is trained with historical data related to the second part of the chemical production process and wherein the model containing the first trained partial model and the second trained partial model is retrained with historical data related to the chemical production process, and

[0012] (c) outputting the operational instruction.

[0013] In another aspect the present invention relates to a method for training a model suitable for monitoring and / or controlling a chemical production process containing a first partial model representing a first part of the chemical production process and a second partial model representing a second part of the chemical production process comprising

[0014] (a) training the first partial model with historical data related to the first part of the chemical production process,

[0015] (b) training the second partial model with historical data related to the second part of the chemical production process and

[0016] (c) retraining the model containing the first trained partial model and the second trained partial model with historical data related to the chemical production process.

[0017] In another aspect the present invention relates to a non-transitory computer-readable data medium storing a computer program including instructions for executing steps of the method according to the present invention.

[0018] In another aspect the present invention relates to a use of the operational instruction obtained by the method according to any one of the preceding claims for monitoring and / or controlling a chemical production process. In another aspect the present invention relates to a monitoring and / or controlling system comprising

[0019] (a) an input for receiving sensor data related to the chemical production process,

[0020] (b) a processor for determining an operational instruction related to the chemical production process, wherein the processor is adapted to provide the sensor data to a trained model, wherein the trained model contains a first partial model representing a first part of the chemical production process and a second partial model representing a second part of the chemical production process, wherein the first partial model is trained with training data related to the first part of the chemical production process and the second partial model is trained with historical data related to the second part of the chemical production process and wherein the model containing the first trained partial model and the second trained partial model is retrained with historical data related to the chemical production process, and

[0021] (c) an output for outputting the operational instruction.

[0022] The invention allows an improved option to monitor and / or control a chemical production process. In particular for more complex production processes, it has been observed that retraining the overall model yields results which closer resemble the overall production process in comparison to a model which contains partial models which are each trained for the part of the production process they relate to. A retraining of the overall model will lead to deviations for each partial model, i.e. each partial model does not fit as well to its partial process as before. Surprisingly however, these deviations make the overall model yield results which are closer to the real measurement values and hence enable more accurate and reliable monitoring and / or controlling of the chemical production process. It has been observed that this effect is particularly pronounced if the chemical production process contains loops, e.g. feeding back unreacted reagents which have been separated from the product. This effect is even more pronounced if the chemical process contains nested loops, i.e. a loop within a loop. Highly efficient chemical production processes have a quite considerable number of such loops as any side product including process heat, may be fed back to avoid energy and material loss and thus to increase process efficiency. Consequently, the present invention is applicable and favorable for a large number of efficient chemical production processes.

[0023] The term "monitoring" may refer to the observation and recording of any state of operation of the chemical production process in a plant. The state of operation includes internal parameters, such as those parameters which are solely relevant within the plant such as reactor temperature, pressure, electricity consumption, input or output material flows, rotational speeds of stirrers, states of valves, concentrations of vapors in the air within the plant, number of people inside the plant. The state of operation also includes external parameters, such as parameters which relate to any exchange with the environment of the plant, such as emission of chemical vapors, heat, sound, vibrations, light. Recording can mean storing the raw data onto a permanent data storage device or preparing documents in a format which are required by the company or by authorities.

[0024] The term "controlling" may refer to taking any actions to change the state of operation of the chemical production process. The actions can be direct, for example by changing the state of a valve, changing the temperature by additional heating or increasing the cooling. The actions can also be indirect, for example by prompting an operator to take actions, for example exchanging a filter or adjusting throughput.

[0025] The term “chemical production process” may refer to a set of actions which are performed in order to obtain a substance or a composition containing more than one substance. A chemical production process may contain a physical process and / or a chemical reaction. In many cases, a chemical production process contains more than one physical-chemical processes and more than one chemical reaction.

[0026] The term "physical process" refers to any process which involves the handling or modification of at least one substance, such as a chemical compound or a composition, without changing it chemical identity. Physical processes include purifications such as distillation, crystallization, filtration, centrifugation, decantation, floatation; formulations such mixing, spray drying, co-extru- sion, coating; or shape-changing process such as grinding, molding, agglomeration, extrusion.

[0027] The term "chemical reaction" may refer to a process involving the chemical transformation of one set of chemical species to another. A chemical reaction can in principle contain one elementary reaction. However, in practice, most chemical reactions contain more than one elementary reaction. The chemical reaction can contain elementary reactions in series or in parallel or both. An example for a chemical reaction containing a series of elementary reactions is a condensation reaction in which firstly a nucleophilic species adds to an electrophilic species as a first elementary reaction followed by the elimination of a small species, like water, as a second elementary reaction. An example for a chemical reaction containing several elementary reactions in parallel is a combustion reaction in which a chemical species reacts with oxygen forming various different partially oxidized species. Chemical reactions can be operated in a homogeneous or heterogeneous way. Homogeneous chemical reactions involve one phase, for example in a gas phase or in a liquid phase, such as a solution. Heterogeneous chemical reactions involve at least two phases. The at least two phases can be of different state of matter, for example one phase is solid and one phase is liquid, or one phase is solid and the other phase is gaseous, or one phase is liquid and the other phase is gaseous. The at least two phases can be of the same state of matter it they are immiscible, for example two immiscible liquid phases or two immiscible solid phases.

[0028] Chemical reactions can be operated in a continuous or discontinuous way, sometimes also referred to as batch chemical reactions. In a continuous chemical reaction, the reagents are continuously fed into a reactor where the reaction takes place and at the same time the products are continuously output from the reactor. In a discontinuous chemical reaction, a reactor is charged with the reagents, then the reaction takes place and after that the products are collected from the reactor. The reactor may be cleaned and is then again charged with new reagents.

[0029] Chemical production process are typically operated in a chemical plant. The term "chemical plant" may refer to a technical infrastructure that is used for an industrial purpose of manufacturing, producing or processing of one or more chemical products, i.e., run chemical reactions to produce chemical compounds, produce formulations by blending chemical compounds, increase the purity of a chemical compound, obtain chemical compounds by recycling waste, bring chemical compounds in a different form, or package chemical compounds or formulations containing chemical compounds.

[0030] The infrastructure of a chemical plant may comprise equipment or process units such as any one or more of a heat exchanger, a column such as a fractionating column, a furnace, a reaction chamber, a cracking unit, a storage tank, an extruder, a pelletizer, a precipitator, a blender, a mixer, a cutter, a curing tube, a vaporizer, a filter, a sieve, a pipeline, a stack, a filter, a valve, an actuator, a mill, a transformer, a conveying system, a circuit breaker, a machinery e.g., a heavy duty rotating equipment such as a turbine, a generator, a pulverizer, a compressor, an industrial fan, a pump, a transport element such as a conveyor system, a motor, etc.

[0031] Further, a chemical plant typically comprises a plurality of sensors and at least one control system for controlling at least one parameter related to the process, or process parameter, in the plant. Such control functions are usually performed by the 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").

[0032] Thus, at least some of the equipment or process units of the chemical plant may be monitored and / or controlled for producing one or more of the products. The monitoring and / or controlling may even be done for optimizing the production of the one or more products. The equipment or process units may be monitored and / or controlled via a controller, such as DCS, in response to one or more signals from one or more sensors. In addition, the plant may even comprise at least one PLC for controlling some of the processes. The chemical plant may typically comprise a plurality of sensors which may be distributed in the chemical plant for monitoring and / or controlling purposes. Such sensors may generate a large amount of data. The sensors may or may not be considered a part of the equipment. As such, production, such as chemical and / or service production, can be a data heavy environment. Accordingly, each chemical plant may produce a large amount of process related data.

[0033] Those skilled in the art will appreciate that the chemical plant usually comprises instrumentation that can include different types of sensors. Sensors may be used for measuring one or more process parameters and / or for measuring equipment operating conditions or parameters related to the equipment or the process units. For example, sensors may be used for measuring a process parameter such as a flowrate within a pipeline, a level inside a tank, a temperature of a furnace, a chemical composition of a gas, etc., and some sensors can be used for measuring vibration of a pulverizer, a speed of a fan, an opening of a valve, a corrosion of a pipeline, a voltage across a transformer, etc. The difference between these sensors cannot only be based on the parameter that they sense, but it may even be the sensing principle that the respective sensor uses. Some examples of sensors based on the parameter that they sense may comprise: temperature sensors, pressure sensors, radiation sensors such as light sensors, flow sensors, vibration sensors, displacement sensors and chemical sensors, such as those for detecting a specific matter such as a gas. Examples of sensors that differ in terms of the sensing principle that they employ may for example be: piezoelectric sensors, piezoresistive sensors, thermocouples, impedance sensors such as capacitive sensors and resistive sensors, and so forth.

[0034] A plurality of chemical plants may form a larger production unit. The term “plurality of chemical plants” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning.

[0035] The term specifically may refer, without limitation, to a compound of at least two chemical plants having at least one common industrial purpose. Specifically, the plurality of chemical plants may comprise at least two, at least five, at least ten or even more chemical plants being physically and / or chemically coupled. The plurality of chemical plants may be coupled such that the chemical plants forming the plurality of chemical plants may share one or more of their value chains, educts and / or products. The plurality of chemical plants may also be referred to as a compound, a compound site, a Verbund or a Verbund site. Further, the value chain production of the plurality of chemical plants via various intermediate products to an end product may be decentralized in various locations, such as in various chemical plants, or integrated in the Verbund site or a chemical park. Such Verbund sites or chemical parks may be or may comprise one or more chemical plants, where products manufactured in the at least one chemical plant can serve as a feedstock for another chemical plant.

[0036] The term "sensor data" may refer to any data which represents the operational state of the chemical production process or parts of it as measured by sensors attached a piece of equipment used for the chemical production process. The sensor data may be received directly from the sensors. Typically, the sensor data are collected by a digital signal controller or programmable logic controller and further transmitted from there. The sensor data may be adjusted, for example by a calibration system before being transmitted. The sensor data from the reactor may also be stored on a storage medium, for example a database on a hard drive or in a cloud system. Hence, the sensor data may be obtained from such storage medium for the purpose of the present invention.

[0037] The sensor data can comprise any measurable physical-chemical value, such as temperature, pressure, pH, concentration or partial pressure of a compound such as oxygen or water content, mass flow rate of reagents, of the reaction mixture in the reactor or of the products after the reactor, heat flow rate, stirrer speed, viscosity, turbidity. Usually, the sensor data also comprises a physical-chemical value associated with an identifier identifying the sensor which has measured the value. The identification of the sensor may comprise the type of sensor, for example a thermometer, and the location of the sensor. The latter is particularly useful if more than one sensor of the same type measures at different locations of the equipment. A typical example is a pressure sensor at the inlet of a reactor and another pressure sensor at the outlet of a reactor. The sensor data can also comprise time information, i.e. the time at which the sensor has collected the physical-chemical value, sometimes referred to as time stamp. The sensor data may contain only one value from one sensor or it may contain more than one value from one sensor, for example a time series of values. Hence, the sensor data may contain multiple values measured by a sensor at different points in time. The sensor data may contain a time series of values measured by a sensor, wherein a value is measured after a predefined time period after the other, for example one value every second. Sensor data related to the chemical production process is received. The term “related” has to be understood in a broad way, namely any information of a sensor which has an influence on the chemical production process or correlates to the state of the chemical production process. The information of a sensor may be the value the sensor outputs, for example the temperature value of a thermometer, or it may be derived value, for example a viscosity value derived from a pressure sensor and a flow rate from a flow meter.

[0038] The sensor data may be received directly from sensors of the first reactor, or it may be received from a data storage medium. The sensor data on the data storage medium may be recorded sensor data or manipulated sensor data. A reason for manipulating sensor data may be to simulate deviations and analyze the impact on the chemical production process with the goal to control the chemical production process in case such situation happens in reality. Another reason may be that a change of the chemical production process can be foreseen, for example a different grade of reagents is going to be employed which shall be taken into account as early as possible.

[0039] According to the present invention, an operational instruction related to the chemical production process is determined. The term “operational instruction” may refer to any data which can be used to control the emissions of the chemical production process by any means. The operational instruction may include the instruction to leave all settings as they are in case the emissions are in an acceptable range, for example below a given threshold. The operational instruction may also include the instruction to adjust one or more than one setting related to the chemical production process. An adjustment may refer to an increase or decrease of the temperature, pressure or throughput in one of the pieces of equipment used for the chemical production process.

[0040] The operational instruction may include a time indication indicating the time at which the operation shall be executed. The sensor data related to the chemical production process may contain sensor measurement values at a certain point in time while the operational instruction needs to be executed at a later point in time, for example 10 seconds after the sensor measurement. By using this information, an action can be taken at the right point in time, i.e. immediately when required, so no delays due to signal latencies or calculation time compromise the control of the chemical production process.

[0041] The determination of the operational instruction is executed using a trained model which receives the sensor data as input and which outputs the operational instruction. The term "model may refer to a mathematical description a chemical production process. The model can be a purely data-driven model or a hybrid model containing both a mechanistic model and a data- driven model. A hybrid model has the advantage that it can in parts strictly follow physicalchemical laws to the extend known and at the same time take into account historical data for parts which are less well understood. Hybrid models require less historical data and are at the same time less susceptible to overfitting. However, sometimes the physical-chemical laws are not well understood, so a purely data-driven model may be advantageous.

[0042] The term "mechanistic model" as used herein generally refers to a model which is based on the fundamental laws of natural sciences, for example any one or more of physical, chemical, biochemical principles, heat and mass balancing. Such models thus represent these principles using equations. A mechanistic model can comprise linear or nonlinear ordinary differential equations, linear or nonlinear partial differential equations, linear or nonlinear algebraic equations, or linear or non-linear differential algebraic equations. Such equations relate to a physical-chemical process.

[0043] A typical example for a mechanistic model is a chemical kinetic model. Essentially, such a model is composed of ordinary differential equations or differential algebraic equations describing the dynamics of chemical species that are being consumed or produced by a set of chemical reactions. The system of ordinary differential equations or differential algebraic equations are usually composed of rate laws that are algebraic equations describing the speed at which chemical species are consumed or produced in reactions. Such an algebraic equation typically depends on the concentrations of the chemical species, temperature in the given reaction and constants, which are usually temperature dependent. Furthermore, certain invariances, such as conservation of mass, can also be represented in such a mechanistic model as algebraic equations.

[0044] "Data-driven model" refers to a mathematical model that is parametrized according to a historical data set to reflect a chemical production process such as reaction kinetics in the first and / or second reactor. In contrast to a mechanistic model that is purely derived using physical-chemical laws, a data-driven model can allow describing relations that are difficult or even impossible to be modelled by physical-chemical laws. Data-driven models are set up without reflecting any underlying physical laws of nature. These are taken into account solely by using the correlations in the data. The data-driven model is preferably a data-driven machine learning model. The data-driven model can be a linear or polynomial regression, a decision tree, a random forest model, a Bayesian network, support-vector machine or, preferably an artificial neural network. The term "historical data" may refer to data sets including at least sensor data and physicalchemical values, wherein each data set is associated with a single chemical production process run. Hence each data set includes data associated with the chemical production process run in a predefined time period. For a batch process such predefined time period may be the beginning to the end of one batch run. For a continuous process, a characteristic period may be chosen, for example the time from charging a reactor with a catalyst until it needs to be replaced by new catalyst. Historical data can be obtained from an already existing plant in which the chemical production process shall be monitored and / or controlled. However, it can also originate from a laboratory, a pilot plant or a similar plant. Sometimes historical data from more than one of these are available.

[0045] The model contains a first partial model representing a first part of the chemical production process and a second partial model representing a second part of the chemical production process. Each partial model may represent one or more than one physical process and / or one or more than one chemical reaction. Each physical process and or chemical reaction contained in the chemical production process is represented by one partial model, hence all physical process and or chemical reaction contained in the chemical production process are represented by one partial model, but not more than one. A partial model may represent all physical processes and chemical reactions taking place in one piece of equipment or in more than one piece of equipment, for example one partial model represents a reactor and a pump for supplying reagents into the reactor. It is also possible that a partial model represents only parts of the physical processes and / or chemical reactions in a piece of equipment, for example in a distillation column, one partial model may represent the upper part for the volatile components and one partial model may represent the lower part for the compounds with high boiling point.

[0046] Training the partial models is typically achieved by adjusting the parameterization according to the historical data. Adjusting the parameterization in this context means varying the parameters in the data-driven or hybrid model such that the output of the model most closely resembles the chemical production process parameters of the training set. Depending on the type of data- driven or hybrid model, various methods of doing so are known and well described in the literature. A typical example is the calculation of a loss function which indicates the difference between the output of the model and the corresponding values of the historical data, determining a gradient of the loss function adjusting the model parameters according to the gradient in order to minimize the loss function.

[0047] After training a partial model may be analyzed by a method quantifying how much the output of the partial model depends on each of its inputs. One partial model may be analyzed or a selection of partial models or all partial models. A useful method for the case that the partial model is a neural network is described by S. Bach et al. in PLoS ONE 10 (7): e0130140 for images. This method may be applied for chemical production processes analogously by using the input variables of the partial model instead of the pixel values of the image. Further analysis methods including interpretable local surrogates, occlusion analysis, integrated gradi- ents / smoothgrad, or layerwise relevance propagation are described by W. Samek in Proceedings of the IEEE, volume 109 (2021), page 247-278. The outcome of such analysis may be used for selection of the input parameters, for example only those are selected which mostly influence the output of the partial model. Alternatively or a additionally, the analysis may be used to verify that the partial model correctly represents its part of the chemical production process, for example by comparing the outcome of the analysis with physical relations.

[0048] The model containing the first and the second partial models is retrained with historical data related to the chemical production process. The chemical production process may comprise both the first part of the chemical production process and the second part of the chemical production process. Hence, the historical data related to the chemical production process may comprise values of the first part of the chemical production process and the second part of the chemical production process. The historical data related to the chemical production process may comprise all or parts of the historical data related to the first part of the chemical production process and all or parts of the historical data related to the second part of the chemical production process. “Retraining” may mean that the already trained partial models undergo a second training. In the initial training of the partial models, each partial model may be trained individually with historical data containing input values and output values related to the part of the chemical production process which is represented by the partial model to be trained. In contrast, for retraining the model containing the partial models may be trained with historical data containing only those values which are input and output to the model, but not those values which are output by one partial model and input to another model. Nevertheless, these values may be used for optimizations, for example more accurate initial guesses for cycles. Upon retraining, the partial models may be changed such that each partial models may not fit as well to the part of the chemical production process it represents as before, but the overall model yields better results for the chemical production process.

[0049] The vales of the historical data used for the retraining may be contained in the historical data used for training the partial model or the values of the historical data used for the retraining may be different to the values of the historical data used for training the partial model, for example because sensor data of different runs of the chemical production process are used for the retraining of the model containing the partial models and for the individual training of the partial models. The retraining may involve calculating and minimizing a loss function, wherein the loss function is indicative of the deviation of the output of the model from the corresponding values in the historical data. After retraining the model containing the first and the second partial model the trained model is obtained which is able to determine the operational parameter from the sensor data.

[0050] The chemical production process may involve cycles in which at least parts of a product of a part of the chemical production process is fed back as reagent in the same or a different part of the chemical production process. The chemical production process may involve one or more than one cycles, like two, three or four. The chemical production process may involve a cycle in a cycle, sometimes referred to as a nested cycle. Retraining a model representing chemical production process containing a cycle may involve a first partial model which directly or indirectly uses an output of a second partial models and the second partial model directly or indirectly uses the output of the first partial model. “Indirectly” may mean that in between the output of the first partial model and the input of the second partial model, there are one or more other partial models in the data flow. Retraining may involve converting the cycle of partial models into a linear arrangement of partial models. The term “linear” may refer to an arrangement of partial models with an order from a first partial model to a last partial model, wherein any partial model may only receive as input the output of a partial model which is in the order before the former, but not after the former. A linear arrangement of partial models may sometimes also be referred to as feed-forward arrangement.

[0051] A linear arrangement may represent the cycle of partial models, wherein the output of the last partial model is no longer used as input for the first partial model. In other words, the linear arrangement may be formed by opening the circle between a first and a last partial model by no longer using the output of the last partial model as input for the first partial model. The input of the first partial model in the linear arrangement may be an initial guess instead of the output of the last partial model. Converting the cycle of partial models into a linear arrangement may involve duplicating the sequence of partial models and appending the duplicated sequence to the linear arrangement of partial models. This may involve using an initial guess as input for the first partial model and using the output of the last partial model as input for the duplicate of the first partial model. Duplicating the sequence of partial models may be done once or more than once, for example twice, three times, five times or ten times. Converting the cycle of partial models into a linear arrangement of partial models may for the case of nested cycles mean a recursive linearization, i.e. the outer cycle containing an inner cycle is first converted into a linear arrangement followed by converting the inner cycle into a linear arrangement. If both the outer and the inner cycle are duplicated, the linear arrangement may contain in each copy of the outer cycle a certain number of copies of the inner cycle. Hence, if the linear arrangement of the production process contains n copies of the outer cycle and the inner cycle is linearized by m copies, the linear arrangement contains n times m copies of the inner cycle.

[0052] Retraining of the model containing a converted cycle of partial models may involve determining a loss function for the entire sequence of partial models, determining a derivative of the loss function and backpropagating the derivative along the sequence of partial models in order to minimize the loss function.

[0053] Retraining a model representing a chemical production process containing a cycle may involve a solving method for an output of one partial model which is input to a different partial model in a cycle. In this way, the cycle may be brought more quickly into a steady state, i.e. each time the cycle of partial models is inferred it reproduces the same or essentially the same output values. Various solving methods are available, for example direct substitution, Wegstein method, Newton method and the Broyden-Fletcher-Goldfarb-Shanno (BFGS) method.

[0054] Retraining may comprise limiting changes to the model. This may ensure that retraining will not completely overwrite the training of the partial models. Limiting changes can be achieved in various ways. Changes may have absolute or relative boundaries for changes. If a loss function is used for the training, the changes may be part of the loss function such that large changes lead to large values for the loss function, so minimizing the loss function will tend to change the models only slightly.

[0055] Upon completion of retraining, the trained model is obtained. The trained model comprises trained partial models which have been adjusted by retraining. An operational instruction can be inferred from the trained model by providing sensor data as input.

[0056] The trained model may be analyzed by a method quantifying how much the output of the trained model depends on each of its inputs. Suitable methods are described above for the analysis of the partial model. The outcome of such analysis may be used for selection of the input parameters, for example only those are selected which mostly influence the output of the trained model.

[0057] Alternatively or a additionally, the analysis may be used to verify that the trained model correctly represents the chemical production process, for example by comparing the outcome of the analysis with physical relations. The method of the present invention comprises outputting the operational instruction obtained from the trained model. Outputting can mean writing the operational instruction on a non-transi- tory data storage medium, for example into a monitoring file or a control file, display it on a user interface, for example a screen, or both. It is also possible to output the operational instruction through an interface to a control system. Such control system may receive the operational instruction and based on such operational instruction change settings of equipment in the chemical production process.

[0058] In another aspect the present invention further relates to the use of the operation instruction for monitoring and / or controlling a chemical production process. Unless explicitly described differently in the following, the description including preferred embodiments for the method applies to the use. The operational instruction may be used for monitoring the present state of the chemical production process, to predict a future state of the chemical production process, to improve the operation of a chemical production process, for example by changing settings of equipment or replacing parts of equipment or complete pieces of equipment, planning a new plant for the chemical production process or predicting changes to the chemical production process when different reagents are used, for example a reagent from a different supplier which as a different purity level.

[0059] In another aspect the present invention relates to a non-transitory computer-readable data medium storing a computer program including instructions for executing steps of the method according to the present invention. “Computer-readable data medium” refers to any suitable data storage device or computer readable memory on which is stored one or more sets of instructions (for example software) embodying any one or more of the methodologies or functions described herein. The instructions may also reside, completely or at least partially, within the main memory and / or within the processor during execution thereof by the computer, main memory, and processing device, which may constitute computer-readable storage media. The instructions may further be transmitted or received over a network via a network interface device. Computer-readable data medium include hard drives, for example on a server, USB storage device, CD, DVD or Blue-ray discs. The computer program may contain all functionalities and data required for execution of the method according to the present invention or it may provide interfaces to have parts of the method processed on remote systems, for example on a cloud system.

[0060] In another aspect the present invention relates to a monitoring and / or controlling system. The monitoring and / or controlling system may be configured to execute the method of the present invention. Unless explicitly described differently in the following, the description including preferred embodiments for the method applies to the system.

[0061] The monitoring and / or controlling comprises an input configured to receive sensor data related to the chemical reaction in a first reactor. Such 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 sensors, or via a programmable logic controller, a distributed control system, or a storage medium including a cloud service. It is even possible that the system is part of a distributed control system.

[0062] The monitoring and / or controlling further comprises a processor configured to determine at least one physical-chemical parameter. The processor may be a local processor comprising a central processing unit (CPU) and / or a graphics 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.

[0063] The monitoring and / or controlling further comprises an output for outputting the operational instruction. Such output may comprise an interface for outputting the operational instruction. The output may send the operational instruction locally or remotely, for example via an interface to a telecommunication system, such as the internet. The output may send the operational instruction to a programmable logic controller, a distributed control system, or a storage medium including a cloud service. It is even possible that the system is part of a distributed control system.

[0064] Brief Description of the Figures

[0065] Figure 1 depicts an embodiment of the model used in the process of the present invention.

[0066] Figure 2 depicts an embodiment of the process of the present invention.

[0067] Figure 3 depicts an example for a production process the invention can be used for.

[0068] Figure 4 and 5 depict an example of how the model can be trained. Figure 6 depicts an exemplary production process of cumene the process of the invention can be used for.

[0069] Description of Embodiments

[0070] Figure 1 illustrates one embodiment for a model used in the process of the present invention. A chemical production process may have a first part 122 and a second part 124. A reagent 110 may be fed into the first part 122. A reagent 110 may be a single chemical compound, for example in case the chemical production process is a purification procedure for such chemical compound, a mixture of chemical compounds, for example in case the chemical production process is a separation process, or more than one chemical compounds separately, for example in the case the chemical production process is a synthetic process in which multiple reagents react with each other to produce one or more than one chemical compounds. The first part 122 of the production process may be a physical process, for example filtration or melting, or a chemical process, for example a chemical reaction. The first part 122 may yield an intermediate which is transferred to the second part 124. The second part 124 may be a physical process, for example filtration or melting, or a chemical process, for example a chemical reaction. The second part 124 yields a product 130. The product 130 may be a single chemical compound, for example in case the chemical production process is a purification procedure for such chemical compound, a mixture of chemical compounds, for example in case the chemical production process is a formulation process, or more than one chemical compound separately, for example in the case the chemical production process is a synthetic process in which multiple products chemical compounds are obtained and separated from each other. Typically, the second part 124 is a different process to the first part 122. However, in some cases, it can be the same process, for example each part is a distillation, wherein the reagent 110 is a non-pure substance purified in the first distillation yielding the compound with higher purity which is distilled once more to yield a product 130. In many cases, the first part 122 and the second part 124 relate to a process in one piece of equipment each, for example the first part 122 relates to a chemical reaction in a reactor and the second part 124 relates to a distillation of the efflux of the reactor.

[0071] The chemical production process is represented by a model 120. The model 120 contains a first partial model 121 representing the first part 122 of the chemical production process and a second partial model 123 representing a second part 124 of the chemical production process. The partial models 121 , 123 may be a data-driven model, for example a neural network, or a hybrid model, for example a differential equation containing one or more than one parameter which are the output of a data-driven model such as a polynomial regression. The model 120 represents the entire chemical production process, therefore all steps of the chemical production process are represented by one partial model. The model 120 may contain two partial models, but depending on the chemical production process, it may contain more than two partial models, for example three, five, ten or 20.

[0072] The model may be trained in a way as depicted in figure 2. Untrained partial model 211 representing a first part of the production process may be trained with historical data 212 in a training 213. The training 213 can use typical methods, for example calculating a loss function, calculating the derivative of the loss function with regard to the variables of the model and adjusting the variables according to the derivative of the loss function to minimize the loss function. The training 213 yields a trained partial model 214. Analogously, untrained partial model 221 representing a second part of the production process may be trained with historical data 222 in a training 223 to yield a trained partial model 224. The trained partial models 214 and 224 are combined 230 to yield a pretrained model 240 representing the entire production process. Combining 230 typically means that output of the trained partial model 213 is used as input for the trained partial model 223. The pretrained model 240 may already be useful for some simple chemical production processes. However, by retraining 250, the model 240 is further refined, i.e. slightly adjusted, to yield the trained model 260. Retraining 250 may also be referred to as end-to-end training which means that historical data essentially contains data relating to the input for the first partial model and the output of the last partial model. Retraining 250 involves historical data 251 which may be the combined historical data 212, 222 without those values which are output by one partial model and input into another partial model. It is also possible that historical data 251 contains all values of the combined historical data 212, 222, wherein the values which are output by one partial model and input into another partial model may be included in the retraining, for example to more efficiently stabilize cycles by picking initial guesses which closely fit to the final values. Trained model 260 which can be used efficiently even for complex chemical production processes.

[0073] Figure 3 shows a more complex example for a production process and how it may be represented. A reagent 310 is compressed by a pump 321 and fed into a reactor 322. In this reactor, the reagent 310 may be converted into an intermediate 330, for example by exposing it to a catalyst which is fixed in the reactor 322. The intermediate 330 is separated from unreacted reagent 324 in a distillation column 323. The unreacted reagent 324 is fed back into the reactor 322 by the pump 321 . The intermediate 330 is compressed by another pump 341 and fed into a reactor 342. This reactor may also contain a catalyst to convert it into product 350, which is purified in distillation column 343. This chemical production process may be represented by a model 300 containing partial model 320 and 340. Partial model 320 may represent the part of the chemical production process containing compression of reagent 310 in pump 321 , reaction in reactor 322, distillation in distillation column 323, back feeding of unreacted reagent 324 and outputting intermediate 330. The partial model 340 may represent the part of the chemical production process including compression of intermediate 330 in pump 341 , reaction in reactor 342, distillation in distillation column 343 and outputting product 350. Alternatively, it is also possible that model 340 contains more partial models which represent smaller parts of the chemical production process, for example one partial model for each piece of equipment, i.e. one model for each pump, reactor and distillation column.

[0074] Figure 4 illustrates the data flow between the partial models for the case that the model represents a process with contains a loop. As can be seen in Figure 4 A), reagent 410 is fed into a part of the chemical production process represented by partial model 421. Partial model 421 uses input values SR,I relating to the input of the reagent 410 as well as 83,1 related to the feedback of the part of the chemical production process represented by partial model 423. 83,1 is thus the output of partial model 423. The output of partial model 421 , Si ,2, is used as input for partial model 422. Partial model 422 has two outputs, namely S2,p relating to the output of product 430 and 82,3 relating to the a side product, for example unreacted reagent or heat. Partial model 423 receives 82,3 as input and outputs 83,1 which is used as input for partial model 421.

[0075] The values SR,I , 81,2, 82,3, 83,1 and S2,p may be scalars, vectors or matrices. They may contain mass flux values, temperatures, pressures, compositions, energy fluxes etc. In Figure 4 B) each partial model is expanded to illustrate its inputs and outputs. Each partial model may be trained separately by using historic values reflecting the inputs and outputs. The retraining of the model containing all partial models may only involve SR,I and S2,p, the remaining values are provided and used by the model. This may work for simple models, but can be improved by the method illustrated in Figure 5.

[0076] The loop formed by partial models 421 , 422, 423 is transformed into a linear arrangement by repeating partial models 421 and 422. The partial model 421 first receives an initial guess of 83,1 , yields the output 81,2 which is received by partial model 422 which yields the output 82,3 which is received by partial model 423. Its output 83,1 may be directly used as input for the next copy of partial model 421. Alternatively, the output 83,1 may be adjusted by solving 501 to more quickly find a steady state, i.e. find a value 83,1 which when input into partial model 421 is close to the output of partial model 423. This can be achieved, for example by direct substitution, Wegstein method, Newton method and the Broyden-Fletcher-Goldfarb-Shanno (BFGS) method. After solving, 83,1 may be passed to the copy of partial model 421 which yields a new Si,2 passed on to a copy of partial model 422 which may yield the output S2,p. It may be necessary to insert one or more than one copy of the partial models and repeat the calculation until S2,p converges. The value S2,p can be used to calculate a loss function 502 based on historical data. Starting from the loss function 502, a derivative can be calculated which can be backpropagated as shown in the lower part of figure 5. In this way, a new initial value 83,1 is obtained and the process can be repeated until the loss function is minimized.

[0077] Figure 6 shows a chemical production process for producing cumene out of benzene and propene. Propene from propene supply 601 , which contains 5 mol-% propane, and benzene from benzene supply 602 are mixed in a mixer 603. Benzene is provided both freshly as well as recycled from distillation column 608 which separates unreacted benzene from the reaction product and feeds it back to the benzene supply 602. The mixture of propene and benzene is evaporated in an evaporator 604 yielding a vapor at 209°C and 25 bar. The vapors are then preheated using a heat exchanger 605 which uses the heat of the effluent from reactor 606 to heat up the mixture containing propene and benzene to reach 360°C. This mixture is then transferred to reactor 606 which is a tubular reactor with 342 tubes, with 76.3 mm inside diameter, 2 mm wall thickness and 6 m of length. It is filled with a solid catalyst of density 2000 kg / m with a void fraction of 0.5 m3 / m3. After the effluent of reactor 606 is used to pre-heat the reactor’s feed in heat exchanger 605, whereupon its pressure is reduced yielding a liquid-vapor mixture at a pressure of 1 .75 bar and a temperature of 90 °C. From this liquid-vapor mixture, the liquid part is separated in a flash tank 607 and fed into distillation column 608. It has 15 stages, and the feed is located in stage 8, the average column pressure was specified to 1.75 bar, with a reflux ratio of 0.44 mol / mol. At the top of distillation column 608 the more volatile benzene is collected and feed back to the benzene supply 602. At the bottom of distillation column 608, the liquid is removed and feed into distillation column 609 which has 20 stages and operates at an average pressure of 1 bar. Pure cumene 610 is collected at the top of distillation column 609. The liquid collected at the bottom of distillation column 609 primarily contains the side product diisopropylbenzene 611.

[0078] A model may contain neural networks as partial models for each piece of equipment. The neural networks may contain two hidden layers with 50 neurons each. A skip connection in the form of a dense layer without activation function from the input to the output may be used. The activation function of the hidden layers may be the Softplus function. Before forwarding through a neural network, the input may be normalized by subtracting the data mean and dividing by the standard deviation. The output of each neural network may be reverse-normalized again with the mean and standard deviation of the output stream. This may enable the neural network to work with and predict values in reasonable data ranges and proved crucial for obtaining a good performance.

[0079] The partial models may be trained with historical data which may contain for each partial model the total mass flowrate in kg / s, temperature in K, pressure in bar and the chemical composition fractions in kg / kg for each of the compounds benzene, cumene, diisopropylbenzene, propane and propene. If two phases, i.e. liquid and vapor phase are present in a piece of equipment, said 8 parameters may be used for each phase, so 16 parameters may be used. For the heat exchanger 605 the heat power in kW may be used in addition. For the reactor 606 the conversion rate in kg / kg for both benzene and propene may be used in addition. For the distillation columns 608 and 609 the specified reflux ratio and rebound ratio in kg / kg may be used in addition.

[0080] The production process contains two cycles: an inner cycle includes the heat exchanger 605 and reactor 606 in which heat is fed back, and an outer cycle includes benzene supply 602, the mixer 603, the evaporator 604, the heat exchanger 605, the reactor 606, the flash tank 607 and the distillation column 608. The cycles contain values which are at the same time input for one partial model and output of another partial model, here the benzene mass flowrate from distillation column 608 back to benzene supply 602. For the individual training of the partial models, these values can be part of the training data. However, the overall model has to determine these values itself in order to reflect changes of the other input values. The model containing the partial models is converted into a chain of partial models, wherein the cycles are opened and duplicated twice, so each partial model which is part of the outer cycle is present three times, each partial model of the inner cycle is present nine times in total. For each cycle, the output of the last model in the chain which is fed back into the first partial model of the chain is subject to a solving method. The linearized model is subsequently retrained by calculating a loss function for the entire model and backpropagating to adjust the partial models such that the loss function for the model is minimized.

[0081] Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims.

[0082] For the processes and methods disclosed herein, the operations performed in the processes and methods may be implemented in differing order. Furthermore, the outlined operations are only provided as examples, and some of the operations may be optional, combined into fewer steps and operations, supplemented with further operations, or expanded into additional operations without detracting from the essence of the disclosed embodiments. In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude a plurality. A single unit or device may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutu- ally different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

Claims

Claims1. A computer-implemented method for monitoring and / or controlling a chemical production process comprising(a) receiving sensor data related to the chemical production process,(b) determining an operational instruction related to the chemical production process by providing the sensor data to a trained model, wherein the trained model contains a first partial model representing a first part of the chemical production process and a second partial model representing a second part of the chemical production process, wherein the first partial model is trained with historical data related to the first part of the chemical production process and the second partial model is trained with historical data related to the second part of the chemical production process and wherein the model containing the first trained partial model and the second trained partial model is retrained with historical data related to the chemical production process, and(c) outputting the operational instruction.

2. The method according to claim 1, wherein the chemical production process involves a cycle in which at least parts of a product of a part of the chemical production process is fed back as reagent in the same or a different part of the chemical production process.

3. The method according to claim 2, wherein the chemical production process involves a cycle within the cycle.

4. The method according to claim 2 or 3, wherein retraining involves converting the cycle of partial models into a linear arrangement of partial models, wherein an initial guess is used as input for the first partial model in the linear arrangement instead of the output of the last partial model.

5. The method according to claim 4, wherein converting the cycle of partial models into a linear arrangement further includes duplicating the partial models of the cycle.

6. The method according to claims 4 or 5, wherein retraining of the model containing a converted cycle of partial models involves determining a loss function for the entire sequence of partial models, determining a derivative of the of the loss function and backpropagating the derivative along the sequence of partial models in order to minimize the loss function.

7. The method according to any of the claims 2 to 6, wherein retraining the model involves a solving method for an output of one partial model which is input to a different partial model in a cycle.

8. The method according to any of the claims 1 to 7, wherein the model contains constraints limiting the changes of the model by training.

9. The method according to any of the claims 1 to 8, wherein the model contains a neural network.

10. The method according to any of the claims 1 to 9, wherein the chemical production process contains continuous product flow.

11. A method for training a model suitable for monitoring and / or controlling a chemical production process containing a first partial model representing a first part of the chemical production process and a second partial model representing a second part of the chemical production process comprising(a) training the first partial model with historical data related to the first part of the chemical production process,(b) training the second partial model with historical data related to the second part of the chemical production process and(c) retraining the model containing the first trained partial model and the second trained partial model with historical data related to the chemical production process.

12. The method according to claim 11 , wherein retraining involves a backpropagation algorithm.

13. A non-transitory computer-readable data medium storing a computer program including instructions for executing steps of the method according to any one of the preceding claims.

14. Use of the operational instruction obtained by the method according to any one of the preceding claims for monitoring and / or controlling a chemical production process.

15. A monitoring and / or controlling system comprising(a) an input for receiving sensor data related to the chemical production process,(b) a processor for determining an operational instruction related to the chemicalproduction process, wherein the processor is adapted to provide the sensor data to a trained model, wherein the trained model contains a first partial model representing a first part of the chemical production process and a second partial model representing a second part of the chemical production process, wherein the first partial model is trained with training data related to the first part of the chemical production process and the second partial model is trained with historical data related to the second part of the chemical production process and wherein the model containing the first trained partial model and the second trained partial model is retrained with historical data related to the chemical production process, and(c) an output for outputting the operational instruction.

Images