Artificial intelligence model for operating plant

Abstract

To provide a machine-learned method and system for use upon operating an industrial plant.SOLUTION: A digital twin of an industrial plant is configured to simulate plant operations based on operating variables from a data store. A machine-learned model includes a stabilizing agent and a disrupting agent. The stabilizing agent modifies the operating variables within the digital twin to perform a stabilizing action for limiting a degree of shutdown and the disrupting agent modifies the operating variables within the digital twin to perform a disruptive action for increasing the degree of shutdown. A composite action reward is configured to reward the machine-learned model for reducing the degree of shutdown from an initial state of the digital twin to a post-action-state of the digital twin.SELECTED DRAWING: None

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to Indian Patent Application No. 202311062877, filed on September 19, 2023, and U.S. Patent Application No. 18 / 511,560, filed on November 16, 2023, which claims priority to Indian Patent Application No. 202311062877, filed on September 19, 2023.

[0002] Field The present disclosure relates generally to machine learned models for use in industrial plants. [Background technology]

[0003] background Industrial plants may experience upsets and shutdowns of processes within the plant. Process upsets and shutdowns may adversely affect the performance and safety of the plant. Therefore, it is important to mitigate the upsets and shutdowns as quickly as possible to reduce the impact of the upsets and shutdowns on the plant.

[0004] Current mitigation actions taken to stabilize processes and prevent hazards within an industrial plant are highly dependent on the knowledge base and experience level of the plant operator. In some cases, it may take the plant operator hours to days to determine an optimal mitigation strategy to stabilize the plant, which then requires additional time for implementation. In other instances, the plant operator may need to act immediately to stabilize the plant without having time to determine an optimal mitigation strategy. Either case may result in additional hazards to the plant.

[0005] Therefore, there is a need for improved systems and methods for determining and implementing optimal mitigation strategies for stabilizing a plant during a process upset or plant shutdown. Summary of the Invention

[0006] Aspects of the present disclosure enable the building and utilization of artificial intelligence (AI) models for plant operations. In an exemplary implementation, the AI ​​models are trained using reinforcement learning based on historical process disturbance data and extrapolated scenarios using a digital twin environment.

[0007] One aspect of the disclosure includes a computer-implemented method for training a machine-learned model for use in an industrial plant. The method includes obtaining training variable inputs from a data store. The data store is configured to store a plurality of operational variables associated with plant operation in the industrial plant. The plant operation is simulated within a digital twin of the industrial plant. The training variable inputs are processed with the machine-learned model. An initial shutdown degree is evaluated based on a standard operating condition criterion to analyze an initial state of the digital twin of the industrial plant. At least one of a stabilizing action and an inhibiting action is performed to modify one or more operational variables in the digital twin. A subsequent shutdown degree is evaluated based on a standard operating condition criterion to analyze a post-operation state of the digital twin. The subsequent shutdown degree is compared to the initial shutdown degree to determine a change in the shutdown degree in the digital twin. A composite operation reward is obtained based on at least the change in the shutdown degree in the digital twin. The composite operation reward is configured to reward the machine-learned model for reducing the subsequent shutdown degree relative to the initial shutdown degree. A prediction is generated based on the composite operation reward. The prediction is configured to optimize the shutdown degree in the plant operation in the industrial plant.

[0008] In another aspect, a system for operating an industrial plant includes a data store of a plurality of operational variables associated with plant operations in the industrial plant. A digital twin of the industrial plant is configured to simulate plant operations based on the operational variables from the data store. The machine-learned model includes a stabilization agent configured to modify one or more of the operational variables in the digital twin to perform at least one stabilization action in the digital twin. The stabilization action is configured to limit a degree of shutdown, the degree of shutdown representing a likelihood of a shutdown of the plant operation occurring based on a current value of the operational variables for a given state. The inhibition agent is configured to modify one or more of the operational variables in the digital twin to perform at least one inhibition action in the digital twin. The inhibition action is configured to increase the degree of shutdown. The action scheduler is configured to schedule either the stabilization agent to perform the stabilization action or the inhibition agent to perform the inhibition action. The composite action reward is configured to reward the machine-learned model for reducing a degree of shutdown from an initial state of the digital twin to a post-operation state of the digital twin. The composite action reward is further configured to penalize the machine-learned model for increasing a degree of shutdown from an initial state of the digital twin to a post-operation state of the digital twin. The prediction generator is configured to generate a prediction based on the composite operation reward. The operation command generator is configured to generate operation commands based on the prediction to optimize plant operation.

[0009] Other objects and features will be in part apparent and in part pointed out hereinafter. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is a block diagram illustrating a system for operating an industrial plant according to one embodiment. [Diagram 2] FIG. 13 illustrates an example of determining a composite action reward according to one embodiment. [Diagram 3] 1 is a flow diagram illustrating an example process for a machine-learned model of the system in accordance with one embodiment.

[0011] Corresponding reference numbers indicate corresponding parts throughout the drawings. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] A system for operating an industrial plant 8 is generally indicated at 10. The system 10 includes a data store 12, a digital twin 14, and a machine-learned model 16 having an action scheduler 18, a stabilization agent 20, an inhibition agent 22, and a composite action reward 24, a prediction generator 26, and an action command generator 28. The system 10 is configured to monitor operational variables related to plant operation in the industrial plant 8 to obtain training variable inputs. The digital twin 14 is configured to simulate plant operation based on the training variable inputs. The machine-learned model 16 is configured to process the training variable inputs using the digital twin 14 as a processing environment. The machine-learned model 16 evaluates an initial degree of shutdown in the digital twin 14. The action scheduler 18 schedules either the stabilization agent 20 or the inhibition agent 22 to operate in the digital twin 14. The machine-learned model 16 evaluates the subsequent shutdown degree to analyze the post-operation state of the digital twin 14. Suitably, the machine-learned model 16 obtains a composite action reward 24 based on the change in the shutdown degree in the digital twin 14 from the initial state to the post-operation state. The prediction generator 26 generates a prediction based on the composite action reward 24 optimizing the shutdown degree to provide a mitigation strategy to execute in the industrial plant 8. The action command generator 28 generates an action command to execute in the industrial plant 8 based on the prediction. A plant operator may use the system 10 to monitor plant operations (including process upsets and plant shutdowns) in the plant 8. As described in further detail below, the plant operator may use the machine-learned model 16 to efficiently determine and implement optimal mitigation strategies to stabilize the industrial plant 8 during a process upset or plant shutdown.

[0013] The industrial plant 8 of the present disclosure may refer to any equipment used in connection with or as part of any process or system for industrial production. A process upset may include any instance within the industrial plant 8 that causes an upset in process flow, quality, and throughput. A plant shutdown may include any instance within the plant that completely stops process flow. As shown in FIG. 1, the system 10 is configured to operate the entire industrial plant 8, although it will be understood that the system 10 may be configured to operate individual components of the industrial plant 8, such as equipment, processes, and sub-processes.

[0014] The data store 12 is configured to monitor and store a plurality of operational variables related to plant operations within the industrial plant 8. The operational variables include, for example, at least one of operator operation data, alarm data, process behavior data, safety data, equipment data, and constraint data, although it will be apparent to one of ordinary skill in the art that the operational variables can include other data related to plant operations without departing from the scope of the present disclosure. The data store 12 can further include a pre-processing component configured to pre-process the operational variables based on a scenario including one or more industrial processes to be considered for training. The data store is configured to send the operational variables as training variables to be input to the digital twin 14 and the machine-learned model 16.

[0015] The digital twin 14 is configured to simulate the industrial plant 8 and to simulate plant operation based at least in part on the training variables input from the data store 12. The digital twin 14 provides a processing environment in which the machine-learned models can be used to process the training variables input from the data store 12.

[0016] The machine-learned model 16 is configured to receive training variable inputs from the data store 12 and process the training variable inputs to provide optimal mitigation strategies for stabilizing the plant 8 during a process upset or shutdown. The machine-learned model 16 utilizes the shutdown degree to assess process health, since the shutdown degree determines the likelihood of a shutdown occurring based on the operational variable values ​​and standard operational condition criteria. The standard operational condition criteria defines regions of operational conditions (i.e., low damage region, high damage region, safe region) within the industrial plant 8. The machine-learned model 16 includes an action scheduler 18 configured to schedule either the stabilization agent 20 or the inhibition agent 22. The action scheduler 18 determines which agent to schedule based on at least one of the initial shutdown degree and historical operational data from the plant 8. In one embodiment, the action scheduler 18 is further configured to determine a time frame based on hyperparameters during which each action must be executed. The hyperparameters are configured to be modified based on at least one training requirement.

[0017] The stabilization agent 20 is configured to perform stabilization actions in the digital twin 14 by modifying one or more operating variables of the digital twin 14 to maintain or reduce the shutdown degree. The stabilization actions are based on operator actions taken in the industrial plant 8 to stabilize the plant as well as mitigation strategies from the engineering documentation. Alternatively, the inhibition agent 22 is configured to perform inhibition actions in the digital twin 14 by modifying one or more operating variables in the digital twin to increase the shutdown degree. The inhibition actions are based on past process upset and plant shutdown data as well as estimated operating scenarios.

[0018] The machine-learned model 16 is further configured to obtain a composite action reward 24 based on at least a change in the degree of shutdown from the initial state of the digital twin for the post-operation state of the digital twin 14. The composite action reward 24 is configured for a parameter independent of the degree of shutdown, in one example, by using the absolute value of the parameter and configuring a desired value. The composite action reward 24 includes a state reward, a state change reward, and a directional reward. The state reward determines whether the degree of shutdown of the initial state of the digital twin 14 and the degree of shutdown of the post-operation state of the digital twin are the same. The state change reward determines whether the degree of shutdown of the initial state and the degree of shutdown of the post-operation state are different. The directional award determines whether the degree of shutdown of the post-operation state is closer to or further away from the degree of shutdown of the desired (i.e., safe) operational state than the degree of shutdown of the initial state. The composite action reward 24 rewards or penalizes the machine-learned model 16 based on the state reward, the state change reward, and the directional reward, as described in more detail below.

[0019] In the example in Figure 2, Tank level is 100 m Composite reward = a1*state reward + b1*state change reward + c1*directional reward State Reward: EX: The digital twin state is in the L-LL (i.e. 20-30) region and the machine learned model is penalized for the time it spent in that region. State Change Reward: EX: The digital twin state moves into the safe area (30-70) and the machine learning model is rewarded for the movement. Directional Reward: EX: While the digital twin state is in L-LL (20-30), the machine-learned model is penalized for being in that state, but directional reward provides a positive reward for moving towards the safe region (25 → 27). a1, b1, c1 are constants that can be adjusted based on the scenario type.

[0020] Based on the composite action reward 24, the prediction generator 26 of the machine-learned model 16 is configured to generate a prediction for optimizing the degree of shutdown, plant performance, and safety in the industrial plant 8. The prediction includes one or more stabilization actions performed in the digital twin 14 to optimize the degree of shutdown and a predicted degree of shutdown for each stabilization action. Accordingly, the action instruction generator 28 of the machine-learned model is configured to generate an action instruction based on the prediction. The action instruction generator 28 organizes the one or more stabilization actions of the prediction into an operator instruction for execution at the industrial plant 8. The system 10 is configured to communicate the prediction and the action instruction as a recommendation output to a plant operator via a display. The plant operator can use the recommendation output to determine how to stabilize the industrial plant 8 during a process upset or plant shutdown. Furthermore, the plant operator can execute an operator action within the industrial plant 8 based on the action instruction to stabilize the industrial plant 8. In a suitable embodiment, the system 10 can include a controller configured to automatically execute the action instruction within the industrial plant 8 to stabilize the plant.

[0021] As described above, the system 10 includes the machine-learned model 16 and is configured to operate an industrial plant. A computer-implemented method for training the machine-learned model for use with the industrial plant is further described below. Initially, the data store 12 stores operational variables related to plant operation. The data store 12 transmits the operational variables as training variables that are input to the digital twin 14 and the machine-learned model 16. The digital twin 14 simulates the plant operation of the industrial plant 8 based on the training variable inputs.

[0022] Now referring to FIG. 3 , the machine-learned model 16 processes the training variable input. First, the machine-learned model 16 evaluates an initial shutdown degree based on a standard operating condition criterion to analyze the initial state of the digital twin 14. The action scheduler 18 determines whether to schedule the stabilization agent 20 or the inhibition agent 22 to perform either a stabilization action or an inhibition action in the digital twin 14 based on the initial shutdown degree and at least one of the plant operation history data. Furthermore, the action scheduler 18 determines a time frame in which either the stabilization agent 20 or the inhibition agent 22 must perform either a stabilization action or an inhibition action. The scheduled agent performs at least one of the stabilization action or the inhibition action to modify one or more operation variables in the digital twin 14. The machine-learned model 16 evaluates a subsequent shutdown degree based on the same standard operating condition criterion to analyze the post-operation state of the digital twin 14. The machine-learned model 16 compares the subsequent shutdown degree with the initial shutdown degree to determine a change in the shutdown degree in the digital twin 14. The machine learned model 16 determines and obtains a composite action reward 24 based on at least the change in shutdown degree within the digital twin 14.

[0023] The machine learned model 16 processes the training variable inputs until a composite action reward 24 is optimized such that the degree of shutdown in the digital twin 14 is within a safe region. A prediction generator 26 of the machine learned model 16 generates a prediction that optimizes the degree of shutdown of the industrial plant 8 by generating a prediction based on the composite action reward 24. Further, an action instruction generator 28 generates an action instruction having an operator action for execution within the industrial plant 8 based on the prediction. In one embodiment, the system 10 automatically executes the operator action within the industrial plant 8 via a controller. In another embodiment, the system 10 communicates the action instructions and predictions via a recommendation output to a plant operator via a display for the plant operator to execute at the industrial plant 8.

[0024] In an exemplary embodiment, the method further includes validating and retraining the machine-learned model 16. Suitably, an initial shutdown degree of the initial state of the industrial plant 8 may be evaluated. An operator action based on the operation instruction of the digital twin 14 is performed and monitored in the industrial plant 8. A subsequent shutdown degree of a post-operator action state of the industrial plant 8 may be evaluated. The initial shutdown degree of the initial state of the industrial plant 8 is compared with a subsequent shutdown degree of the post-operator action state to determine a change in the shutdown degree in the industrial plant 8. A comparison between the change in the shutdown degree in the digital twin 14 and the industrial plant 8 is evaluated to evaluate a performance of the machine-learned model 16. Based on the comparison, one or more parameters of the machine-learned model 16 may be modified to retrain the machine-learned model.

[0025] To use the machine learned model 16 to predict future shutdown severity and develop future mitigation strategies in the industrial plant 8, the method according to an embodiment further comprises creating extrapolated training variable inputs based on the extrapolated scenario of the industrial plant. The extrapolated training variable inputs may be sent to the digital twin 14 to simulate the extrapolated scenario in the digital twin. Suitably, the machine learned model 16 processes the extrapolated training variable inputs to provide a recommended output for the extrapolated scenario.

[0026] In one embodiment, the method is used to build and utilize a machine-learned model 16 for plant operation. The method generally includes capturing operator actions during operation of the industrial plant 8, and more specifically, operator actions performed in response to process upsets and shutdowns in the industrial plant. The machine-learned model 16 is trained based at least in part on operator actions and other operational variables associated with the operation of the plant 8 and its associated equipment, processes, and sub-processes. The machine-learned model 16 utilizes agents 20, 22 trained using reinforcement learning on past process disturbances and extrapolated scenarios by using the digital twin 14 as a processing environment. The system 10 is configured to detect and predict future process upsets and shutdowns in the industrial plant 8. The machine-learned model 16 is trained using a composite operational reward 24 such that the machine-learned model is rewarded for keeping operational variables in a safe state based on the degree of shutdowns in the industrial plant 8.

[0027] Advantageously, the present disclosure provides a system and method for determining and implementing optimal mitigation strategies for stabilizing an industrial plant during a process upset or plant shutdown. The system continuously monitors operating variables in the plant to determine a shutdown degree in the plant. This allows a plant operator to quickly detect process upsets and shutdowns. Furthermore, the machine-learned model is configured to maintain the optimized shutdown degree in the digital twin by performing stabilization actions. The system communicates the stabilization actions to the plant operator via a recommended output. The recommended output provides a mitigation strategy for stabilizing the plant that optimizes plant performance and safety. The system achieves a plant stabilization process that eliminates process variations due to operator experience levels and knowledge bases, reducing the time plant operators spend determining mitigation strategies to implement in the plant.

[0028] Although embodiments of the present disclosure include special purpose computers including various computer hardware, as described in more detail herein, and are described in connection with an exemplary computing system environment, they are operable in other special purpose computing system environments or configurations. The computing system environment is not intended to suggest any limitation as to the scope of use or functionality of any aspect of the present invention. Moreover, the computing system environment should not be interpreted as having any dependency or requirement relating to any one or combination of components illustrated in the exemplary operating environment. Examples of computing systems, environments, and / or configurations that may be suitable for use with aspects of the present disclosure include, but are not limited to, personal computers, server computers, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable appliances, mobile phones, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like.

[0029] Aspects of the disclosure may be described in the general context of data and / or processor-executable instructions, such as program modules, stored in one or more tangible, non-transitory storage media and executed by one or more processors or other devices. Generally, program modules include, but are not limited to, routines, programs, objects, components, and data structures that perform particular tasks or implement particular abstract data types. Aspects of the disclosure may also be practiced in distributed computing environments where tasks are performed by remote processing devices linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote storage media, including memory storage devices. For purposes of illustration, programs and other executable program components may be depicted as separate blocks. However, it will be appreciated that such programs and components may reside at various times in different storage components of a computing device and be executed by the device's data processor(s).

[0030] In operation, a processor, computer, and / or server may execute processor-executable instructions (e.g., software, firmware, and / or hardware) as illustrated herein to implement aspects of the present invention. The processor-executable instructions may be organized into one or more processor-executable components or modules on a tangible, processor-readable storage medium. Also, embodiments may be implemented with any number and configuration of such components or modules. For example, aspects of the present disclosure are not limited to the specific processor-executable instructions or to the specific components or modules illustrated in the figures and described herein. Other embodiments may include different processor-executable instructions or components having more or less functionality than illustrated and described herein.

[0031] The execution or order of execution of operations in accordance with aspects of the disclosure illustrated and described herein is not required unless specifically specified. That is, operations may be performed in any order unless specifically specified, and embodiments may include additional or fewer operations than those disclosed herein. For example, performing or executing a particular operation before, contemporaneously with, or after another operation is considered to be within the scope of the disclosure.

[0032] Not all of the components shown or described are required. Furthermore, some embodiments and implementations may include additional components. Changes in the arrangement and type of components may be made without departing from the spirit or scope of the claims set forth herein. Additional, different, or fewer components may be provided, and components may be combined. Alternatively or additionally, a component may be implemented by several components.

[0033] Having described the invention in detail, it will be apparent that modifications and variations are possible without departing from the scope of the invention as defined in the appended claims.

[0034] When introducing elements of the invention or preferred embodiments thereof, the articles "a", "an", "the" and "said" are intended to mean that there are one or more elements. The terms "comprising", "including" and "having" are intended to be inclusive and mean that there may be additional elements other than the listed elements.

[0035] In view of the above, it will be seen that the several objects of the invention are achieved and other advantageous results attained.

[0036] Since various changes may be made in the above-described products without departing from the scope of the invention, it is intended that all matter contained in the above description and shown in the accompanying drawings shall be interpreted as illustrative and not in a limiting sense.

[0037] The Abstract and Overview are provided to aid the reader quickly ascertain the nature of the technical disclosure. They are submitted with the understanding that they will not be used to interpret or limit the scope or meaning of the claims. The Overview is provided to introduce in a simplified form a selection of concepts that are further described in the Detailed Description. The Overview is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the claimed subject matter.

Claims

1. 1. A computer-implemented method for training a machine-learned model for use in an industrial plant, comprising: obtaining training variables input from a data store, the data store configured to store a plurality of operational variables related to plant operations in the industrial plant; simulating the plant operation using a digital twin of the industrial plant; processing the training variable inputs with the machine-learned model, evaluating an initial shutdown degree based on a standard operating condition criterion to analyze an initial state of the digital twin of the industrial plant; performing at least one of a stabilization action and an inhibition action to modify one or more operational variables in the digital twin; evaluating a subsequent shutdown degree based on the standard operating condition criteria to analyze a post-operation state of the digital twin; comparing the subsequent shutdown degree to the initial shutdown degree to determine a change in the shutdown degree within the digital twin; obtaining a composite action reward based on at least a change in the degree of shutdown in the digital twin, the composite action reward being configured to reward the machine-learned model for reducing the degree of subsequent shutdowns relative to the degree of initial shutdown; generating a prediction based on the composite operation reward, the prediction configured to optimize the degree of shutdowns of plant operations in the industrial plant.

2. The method of claim 1 , wherein the plurality of operational variables includes operator action data, alarm data, process behavior data, safety data, and equipment design and constraint data.

3. The method of claim 1 or 2, further comprising the step of pre-processing the operational variables in the data store based on an industrial process.

4. The method of claim 1 , wherein assessing the likelihood of a shutdown comprises determining the likelihood that a shutdown will occur based on values ​​of operational variables.

5. 5. The method of claim 1, further comprising determining whether to perform the stabilizing action or the inhibiting action based on at least one of an incipient shutdown of the digital twin and historical plant operation data from the industrial plant.

6. 6. The method of claim 5, wherein performing at least one of the stabilizing and inhibiting actions to modify one or more operational variables in the digital twin further comprises determining a time frame based on hyperparameters within which the respective actions should be performed, the hyperparameters being configured to be modified based on at least one or more training requirements.

7. 7. The method of claim 1, wherein performing the stabilization action comprises modifying one or more of the operational variables in the digital twin to maintain or reduce the degree of shutdown in the digital twin.

8. 8. The method of claim 1, wherein performing the inhibiting action comprises modifying one or more of the operational variables in the digital twin to increase the degree of shutdown in the digital twin.

9. 9. The method of claim 1, wherein the step of determining the composite action reward comprises the steps of: determining a state reward for evaluating whether the shutdown degree of the initial state and the shutdown degree of the post-action state are the same; determining a state change reward for evaluating whether the shutdown degree of the initial state and the shutdown degree of the post-action state are different; and determining a directional reward for evaluating whether the shutdown degree of the post-action state is closer to or farther from the shutdown degree of a desired action state than the shutdown degree of the initial state.

10. The method of claim 1 , further comprising generating operation instructions for execution at the industrial plant based on the prediction.

11. The method of claim 10 , further comprising automatically executing the operation instructions in the industrial plant via a controller.

12. The method of claim 10 or 11, further comprising the step of communicating the operation instructions to an operator for execution by the operator at the industrial plant.

13. 13. The method of claim 10, further comprising the steps of: assessing an initial shutdown degree of an initial state of the industrial plant; monitoring an operator action performed on the industrial plant; assessing a subsequent shutdown degree of a post-operator action state of the industrial plant; and comparing the initial shutdown degree to the subsequent shutdown degree to determine a change in shutdown degree within the industrial plant.

14. 14. The method of claim 13, further comprising: evaluating a comparison between a change in shutdown degree in the digital twin and a change in shutdown degree in the industrial plant; and modifying one or more parameters of the machine-learned model based at least in part on the comparison to retrain the machine-learned model.

15. 15. The method of claim 1, further comprising: processing extrapolated training variable inputs based on an extrapolated scenario in the industrial plant using the machine-learned model; and simulating the extrapolated scenario in the digital twin to predict a future shutdown severity in the industrial plant and determine a mitigation strategy for the extrapolated scenario.

16. 1. A system for operating an industrial plant, comprising: a data store including a plurality of operational variables associated with plant operations within the industrial plant; a digital twin of the industrial plant configured to simulate plant operations based on the operational variables from the data store; and A machine-learned model, a stabilization agent configured to modify one or more of the operating variables in the digital twin to perform at least one stabilization action in the digital twin, the stabilization action being configured to limit a shutdown degree, the shutdown degree representing a likelihood of a shutdown of the plant operation occurring based on a current value of the operating variables at a given state; an interference agent configured to modify one or more of the operational variables in the digital twin to perform at least one interference action in the digital twin, the interference action being configured to increase the shutdown degree; an action scheduler configured to schedule either the stabilization agent to perform the stabilization action or the inhibition agent to perform the inhibition action; a composite operation reward configured to reward the machine-learned model for reducing the degree of shutdown from the digital twin's initial state to the digital twin's post-operation state, the composite operation reward being further configured to penalize the machine-learned model for increasing the degree of shutdown from the digital twin's initial state to the digital twin's post-operation state; a prediction generator configured to generate a prediction based on the composite action reward; an operation command generator configured to generate operation commands based on the predictions to optimize the plant operation; A system including:

17. The system of claim 16 , wherein the disruptive action is based on process upsets history data.

18. 18. The system of claim 17, wherein the composite action reward includes a state reward for evaluating whether a shutdown degree of the initial state and a shutdown degree of the post-action state are the same, a state change reward for determining whether a shutdown degree of the initial state and a shutdown degree of the post-action state are different, and a directional reward for determining whether a shutdown degree of the post-action state is closer to a shutdown degree of a desired operating state or farther from the shutdown degree of the desired operating state than a shutdown degree of the initial state.

19. 19. The system of claims 16 to 18, further comprising a controller configured to automatically execute the operation instructions within the industrial plant.

Images