Method and computer readable storage medium for automatically generating an optimization model

Abstract

A computer-implemented method for automatically generating an optimization model for site-wide plant optimization includes mapping a process flow graph of a plant process to a graph comprising nodes and edges, where the nodes represent processes and the edges represent flows between the processes. Behaviors of each node of the graph are learned based at least on historical data of the plant process. One or more regression functions are modeled for each node to predict an output of each of the processes, where the one or more regression functions are simulated based on the learned behaviors of each node.

Description

Technical Field

[0001] This disclosure generally relates to computer systems for plant optimization and methods and software used therewith, and more specifically, to methods and systems for automatically generating optimization models for system-wide plant optimization to generate recommendations for operating setpoints of a processing plant. Background Technology

[0002] The manufacturing and process industry consists of a site-wide network of complex processes, each with its own set of inputs and outputs. Within a unit process, transient relationships frequently exist between setpoints, production volumes, desired output quality, and waste streams. The variability of input streams, operational requirements, maintenance, failures, production schedule changes, and more makes production processes dynamic. Plant-wide management requires the ability to predict dynamic process behavior and modify any controls to adhere to production plans as closely as possible. Manual optimization model generation is time-consuming, challenging, and requires domain and optimization experts.

[0003] Current industry practices include the very limited use of artificial intelligence (AI) to design operational strategies. For example, in the event of an unplanned process failure, plant managers or plant managers (PMs) can use experience-based or heuristic methods to determine a set point that provides optimal productivity under the operational constraints imposed by the failure.

[0004] A common characteristic of traditional processes is the use of nonlinear physics first-principles (e.g., thermodynamic) models for each process, coupled through flow and mass balance equations. Existing work has previously combined physical and data-driven models. The development of alternative statistical models for complex processes has begun, but primarily as a design of experiments, using detailed process simulators to estimate response surface models. Some traditional processes have used alternative models to optimize production processes, but this is typically done at the design stage to optimize flowsheet structure or in the context of optimizing the setpoint of a single process. Without using a plant simulator, traditional processes cannot rely solely on historical plant sensor data.

[0005] Traditional end-to-end learning methods consider only a single process and, due to the need to solve two-stage stochastic programming problems, can only handle specific categories of problems. Typical process assumptions posit that the optimization problem is convex with respect to control variables. This condition does not apply to many common regression functions, such as decision tree regression, deep neural networks, multivariate adaptive regression splines (MARS), and random forests. Some traditional end-to-end learning methods require the decision-centric learning framework to be represented as linear programming or submodule maximization. Because sensor data from a plant comes at different time resolutions, it is impossible to build end-to-end solutions for site-wide optimization problems.

[0006] This disclosure proposes a framework for automatically generating optimization models for multi-cycle setpoint control problems in multi-plant process industries. Summary of the Invention

[0007] According to various embodiments, a computer-implemented method is provided for automatically generating optimization models for site-wide plant optimization. The method includes mapping a process flow diagram of a plant process to a graph comprising nodes and edges, where nodes represent processes and edges represent flows between processes. The behavior of each node in the graph is learned, at least based on historical data of the plant process. One or more regression functions are modeled for each node to predict the output of each in the process, wherein one or more regression functions are modeled based on the learned behavior of each node.

[0008] In some embodiments, the graph is a directed acyclic graph of the plant.

[0009] In some embodiments, the inputs to each node include controlled and uncontrolled variables.

[0010] In some embodiments, the method further includes encoding the network topology by generating a graph adjacency matrix that identifies interconnected nodes.

[0011] In some embodiments, the method further includes using one or more matrices and the positions of nodes and edges from the graph to decode the network topology to draw a directed acyclic graph for input validation.

[0012] In some embodiments, the method further includes generating an optimization model from the regression function and adjacency matrix of each node, the optimization model providing optimized setpoints for one or more parameters for each node.

[0013] In some embodiments, the optimization model outputs a continuous optimization model with a function value estimator and a gradient estimator. In some embodiments, the optimization model outputs a mixed-integer linear procedure for modeling one or more regression functions.

[0014] In some embodiments, the method further includes providing different key performance indicators as input and generating options for a setpoint to achieve the input key performance indicators.

[0015] In some embodiments, the method further includes automatically determining whether the input to the regression function is correct.

[0016] In some embodiments, the graph is a two-level fully connected feedforward network.

[0017] In some embodiments, the regression function includes piecewise linear and nonlinear regression models.

[0018] According to various embodiments, a computer-implemented method and a non-transitory computer-readable storage medium tangibly contain computer-readable program code having computer-readable instructions, which, when executed, causes a computer device to perform a method for improving the computational efficiency of the computing device, for automatically generating an optimization model for site-wide plant optimization. The method includes defining a graphical representation of a process flow graph of a plant process, comprising nodes and edges, wherein nodes in the graphical representation represent processes, and edges in the graphical representation represent flows between plant processes. The network topology of the graphical representation is encoded by generating an adjacency matrix for the graph, the adjacency matrix identifying interconnected nodes. A set of equations defining the network topology is automatically generated. One or more regression functions are modeled using a machine learning platform to predict the output of each process in the plant process based on inputs received at each process. An optimization model is generated from one or more regression functions and the adjacency matrix for each node, providing a setpoint for each process in the plant.

[0019] Establishing advisory support for PMs presents a significant opportunity to explore and evaluate a broader set of options, thereby calculating optimal strategies for continuous decision-making at a plant-wide scale. First, to build such an AI-based advisor, the system is trained to have an expert-level understanding of the behavioral relationships between inputs and outputs. Second, this behavioral representation, along with other operational constraints, is used to determine the optimal setpoints for all control variables. The aspects of this disclosure provide solutions to these opportunities.

[0020] Using the concepts discussed herein, a tool is provided that improves the method for generating optimization models for site-wide plant optimization. As shown in the example below, the optimization model according to aspects of this disclosure can provide the optimal objective value with a significantly shorter average computation time than conventional regression methods.

[0021] These and other features will become apparent from the following detailed description of illustrative embodiments thereof, which is taken in conjunction with the accompanying drawings. Attached Figure Description

[0022] The accompanying drawings are illustrative embodiments. They do not illustrate all embodiments. Other embodiments may be used additionally or alternatively. Details that may be obvious or unnecessary may be omitted to save space or for more efficient illustration. Some embodiments may be implemented with additional components or steps and / or with all components or steps not shown. When the same numbers appear in different drawings, they represent the same or similar components or steps.

[0023] Figure 1 This is a schematic diagram of a time-indexed graphical representation of a network of multiple plants for optimization using an optimization model, consistent with the illustrative embodiment.

[0024] Figure 2 This is a flowchart illustrating the estimation of each process via a regression function, consistent with the illustrative embodiments.

[0025] Figure 3 This is a flowchart illustrating a generalized network representation of an applied regression function consistent with the illustrative embodiment.

[0026] Figure 4A and Figure 4B The use of a model generator to generate an optimized model with minimal input and optimization background is shown, consistent with the illustrative embodiment.

[0027] Figure 5 This is a schematic diagram of an end-to-end system for optimizing a solution, consistent with the illustrative embodiments.

[0028] Figure 6 This is a flowchart illustrating an automated model generation system consistent with the illustrative embodiment.

[0029] Figure 7A and Figure 7B The illustration shows a process flow factory modeled as a two-level fully connected feedforward network without the assumption of skip layers, according to an illustrative embodiment.

[0030] Figure 8 The preparation of the adjacency matrix for directed acyclic graph generation of the entire system is shown, consistent with the illustrative embodiment.

[0031] Figure 9 The preparation of the adjacency matrix for directed acyclic graph generation is shown, consistent with the illustrative embodiment, when a factory has more than one product.

[0032] Figure 10 A decoding network topology for input validation, consistent with the illustrative embodiment, is shown.

[0033] Figure 11 A continuous model of the regression function, consistent with the illustrative embodiment, is shown. Detailed Implementation

[0034] In the following detailed description, numerous specific details are illustrated by way of example to provide a thorough understanding of the relevant teachings. However, it will be apparent that the invention can be practiced without these details. In other instances, well-known methods, processes, components, and / or circuits are described at a relatively high level without detailed description to avoid unnecessarily obscuring aspects of the teachings.

[0035] Unless otherwise specifically stated and apparent from the following description and claims, it should be understood that throughout the description, the use of terms such as “processing,” “computing,” “operation,” “determining,” etc., refers to the actions and / or processes of a computer or computing system or similar electronic computing device that manipulate and / or convert data represented as physical quantities (such as electronic quantities) in the registers and / or memory of the computing system into other data similarly represented as physical quantities in the memory, registers, or other such information storage, transmission, or display devices of the computing system.

[0036] As discussed in more detail below, this disclosure generally relates to systems and computerized methods for automatically generating optimization models for site-wide plant optimization. These systems and computerized methods provide technological improvements in the efficiency and accuracy of computerized systems by leveraging data-driven machine learning and optimization techniques to improve plant production under various asset capacity constraints and different production strategies.

[0037] As used herein, the term "site-wide optimization" refers to a product manufacturing plant, such as an oil sands processing plant. Given the current state of all process variables and inventory, and known constraints on available capacity within the planning scope, site-wide optimization seeks to compute the optimal production schedule over time that maximizes the flow rate of the final product. This is typically expressed in terms of optimal flow rates and various quantities of inventory changing over time, and we seek to do so while considering various operational and economic considerations, such as the smoothness of the production schedule, maintaining acceptable inventory levels in each period, and meeting economic production or demand objectives.

[0038] In process industries, planning departments develop production plans for a planned scope based on market forecasts, demand forecasts, maintenance and resource scheduling, operating conditions, and field logistics. Given the inherent uncertainties in daily operations, production plans are revised daily or per shift. Product managers (PMs) focus on ensuring that key performance indicators (KPIs), such as throughput, align with the production plan as closely as possible. They collaborate with process engineers to coordinate production levels and targets across multiple processes. In the event of unexpected disruptions, PMs are responsible for minimizing recovery time and deviations from the production plan. Under normal operating conditions, the production plan helps identify anomalous deviations that drive process corrections. However, PMs cannot manually provide optimization parameters for the entire plant. Typically, PMs are assigned to individual processes and have their own KPIs, which may or may not align with those of PMs in downstream processes.

[0039] Typical process flow diagrams can be used for site-wide process optimization. Setpoints, such as temperature, pressure, and flow rate, are used to control the behavior of processes and the plant. Setpoints always have an impact on both upstream and downstream processes. Therefore, in a processing plant with thousands of controllable setpoints, it is likely that a single critical setpoint can affect the total throughput or output, and how far they are from the baseline production plan.

[0040] refer to Figure 1 and Figure 2 One step in developing a site-wide optimization model is creating a time-indexed graphical representation 100 (also known as a directed acyclic graph 100) of various processes, P1, P2, P3, P4, and P5, collectively referred to as processes 102. In a typical time-indexed graphical representation 100, each process 102 has a set of independent inputs 104 and outputs 106, where the output from an upstream process becomes the input of a downstream process. Figure 1 In the example, storage containers T1, T2, T3, and T4 can be used to identify the output of one process and the input of another. For example, the output of process P1 is in container T1, and the output of process P2 is in container T2. Container T3 is a combination of containers T1 and T2 and provides the input for process P3.

[0041] Each node in Graph 100 represents the state of a specific process (e.g., P1) or storage tank (e.g., T1), and each edge represents the corresponding material flow rate. More generally, each plant itself can be modeled using multiple lower-level process nodes connected together to form the plant. Employing a discrete-time model is useful, i.e., a set of time periods at a selected resolution, such as hourly intervals. Plant variables can be defined corresponding to time periods (e.g., flow rate, setpoint for the next hour) or to points in time, i.e., the end of any time period (e.g., tank level at the end of the next hour). Each node or edge is also equipped with a data structure to capture various parameters, such as the current setpoint, the state of other process variables in the plant (e.g., mass, temperature, density), static capacity limits, flow rate limits, etc. While the above network representation is necessarily a simplification of a real-world process flow graph, it can be used to implement predictive optimization frameworks.

[0042] Input 104 may include controlled variables, such as voltage and temperature, and uncontrolled setpoints, such as feed flow. Each process 102 can be observed, and observations may include sensor readings during a specific process. For each process 102, an output 106 can be predicted based on a regression function, as discussed in more detail below. Historical data for each process 102, based on the actual values ​​of input 104 and output 106, can be used to provide data for an optimization model to make such predictions. Therefore, a further action of the process according to this disclosure is to learn the behavior of each process. While historical data can be used initially, real-time data can be used during plant operation to update the machine learning, adjusting the output predictions as needed based on that data.

[0043] Another step in the site-wide optimization process is to address optimization issues. Figure 3 An overview of the operation of the optimization model disclosed herein in solving optimization problems is provided. Figure 3 In this context, x represents the controlled input variable, and f represents the predictive model applied to x to provide the output y. The variable z represents the uncontrolled variable, such as the flow. Therefore, reading from left to right... Figure 3 For a given process, given a first set of controlled variables x1, prediction model f1 can predict the output y1. This product, stored in g1, can be used as input to a second process, where z1 represents the stream from g1, x2 represents the controlled variable, and f2 is the prediction model used to predict the output y2. Similarly, y2 provides input to a further downstream process via g2, where z2 is an uncontrolled variable, such as the stream from g2, x3 is a controlled variable, and prediction model f2 can determine the output y3. The optimization model of this invention can, given measurements from sensors along various processes, determine the optimal values ​​of the controlled variables x1, x2, and x3 to optimize the desired output, which may be an output stream, product purity, etc.

[0044] Figure 4A and Figure 4B An exemplary process flow diagram is shown. As discussed in more detail below, an aspect of the invention is to simplify the process of generating a model from the process flow diagram into a model that can be used as an optimization model, where the number of inputs can be minimized. As mentioned above, manual optimization model generation is typically challenging and requires domain experts and optimization specialists. Furthermore, manual optimization models can only focus on each process and lack the capability to perform site-wide optimization model generation, as this would require considering not only hundreds or thousands of setpoints but also the flows between each process that change due to variations in these setpoints.

[0045] Figure 5 An end-to-end system for optimizing a solution is illustrated. Aspects of this disclosure focus on model generation 200 based on input from a programmer 206 with domain understanding, and the development of an optimization model 202 based on a KPI 204 identified by the factory. For example, the optimization model 202 can be developed based on KPI 204, which is maximizing product output.

[0046] Now for reference Figure 6 The diagram illustrates an automated optimization model generation system 260. In summary, aspects of this disclosure provide a method for automatically generating an optimization model 262 for site-wide optimization. The optimization model 262 can be generated from each node (such as process P1, see...) Figure 1 The regression function 264 was obtained, including control, observation, and environmental variables, as well as from the directed acyclic graph 100 of the entire plant (see Figure 1 The adjacency matrix 266 is obtained from the model. The output of the optimization model may include a continuous optimization model 208 that can provide function values, gradients, etc., and a mixed-integer linear programming model 270 that provides a mixed-integer programming (MIP) model 274. Users can experiment with different configurations and KPIs to generate a list of options for setpoints and expected target values, whereby the user can select the best option. A validation factory network 278 can be provided to verify the correctness of the inputs and display the network flow 280.

[0047] refer to Figures 7A to 11 The above general description of aspects of this disclosure is expressed in more detail below.

[0048] As described above, one action of this disclosure is to define a graphical representation of the process flow diagram for a specific plant. A two-level fully connected feedforward network (FFN) can be used, as described in more detail below.

[0049] For details, please refer to the following: Figure 4B , Figure 7A and Figure 7BThe process flow is represented as a directed graph, where input, control, and output variables are represented as arcs, and each unit operation is represented by a node. Each node can represent a unit process or a system of unit processes, and they are modeled together. The standard form of the single-cycle optimization problem is... Figure 4B As shown in the figure, where f l It is the set of regression functions in the l-th process layer, and linearly constrained. Capture network flow constraints that represent quality balance and inventory levels.

[0050] The objective function is defined based on some KPIs (e.g., throughput), and the formula is treated as a maximization or minimization problem. In this formula, each node is represented by a predictive model, which is expressed as a set of constraints as shown above. For deployment purposes, the system can solve a multi-period model for long-term planning.

[0051] This disclosure considers the problem of optimizing control variables over T time periods, where there are some coupling constraints between the periods. The multi-period optimization problem can be rewritten as follows:

[0052] Make:

[0053] ,

[0054]

[0055]

[0056]

[0057]

[0058] ,

[0059] For some ,in, and It is the storage limit of the can, and It has Figure 4B Let the constraint set of the form in the t-th period be... This represents a component-based product. A usage time index t∈{1,...,T} represents the t-th period model.

[0060] Optimization methods can effectively solve various types of regression models. For regression models such as linear regression and decision trees, optimization methods can reformulate the problem as a scalable mixed-integer linear programming (MILP). Relaxation methods, such as McCormick envelope and Sherali-Adams' reconstruction linearization techniques, can be used to handle specific bilinear forms arising from pooling constraints. Nonlinear regression functions, such as those in general deep neural networks or general ensemble models, can lead to nonlinear constraints or even black-box constraints. Furthermore, depending on the type of predictive model learned, nonlinear optimization algorithms can be used to develop special structures for the problem. For example, for single-period models, such as... Figure 4B As shown, when f l When a problem is highly nonlinear, cannot be linearized, and its gradient is available, the Augmented Lagrangian method (ALM) can be used to solve it. Figure 4B Another set of linear constraints captures the process flow representing quality balance and inventory levels (i.e., network flow constraints).

[0061] Transforming real-world problems into solvable mathematical models typically requires significant time and optimization skills. To effectively provide optimized model solutions that scale well to numerous use cases and scenarios, aspects of this disclosure create and deploy optimized models that would otherwise be impossible for domain and optimization experts to develop. Often, once deployed, updating the model, such as adding new constraints, can again require substantial time and expertise. However, aspects of this disclosure provide a framework for obtaining input that requires minimal knowledge from domain subject matter experts and optimization specialists to automatically generate optimized models.

[0062] This solution can use a two-stage feedforward network (FFN) as a graphical representation of the process flow graph, such as... Figure 7B As shown. The feedforward network structure is used to model each storage tank subsystem between the two process layers, resulting in a two-level network for the entire system, as follows. Figure 7B As shown. Users can define DAG 100 using the following rules (see...). Figure 1 Prepare the adjacency matrix: 1) Traverse column by column from left to right for indexing, 2) Label the i-th process with Pi and the i-th tank with Ti, and 3) Create the connection matrix A for the two-level feedforward network. Figure 8 An example of a process flow graph that can generate a connectivity matrix is ​​provided. An exemplary connectivity / adjacency matrix is ​​shown below:

[0063]

[0064] For processes that produce more than one product, the entries in a connection can be not only zero or one, but also two, three, or more. For example, as... Figure 9 As shown, if a process can produce two products, the connection matrix can include 2. Figure 9 In this example, for P1, there are two products, so the connection matrix can be (P1, T1) = 1, (P1, T2) = 2, (P1, T3) = 1 and (P1, T4) = 2. By following these rules, we note that any two-stage feedforward network has a unique connection matrix representation, and vice versa.

[0065] To decode the network topology, refer to Figure 10 Given a connection matrix A, detailed information about the second-level network can be reconstructed, such as the layer structure and location of each process and tank. Let [label = P1, … , P V1, T1, …, T M1 , P V1+1 , … ,P V1+V2 , … , T M1 + … +M L-2 +1, …,T M1 + … +M L-1 , … , P V1 + … +V L-1 +1, …, P V1+ …+ VL ] represents the column and row labels of A, which can generate A l y l +B l z l + C l s l = d l Linear constraints of type, where

[0066] V l It is the number of processes in the l-th processing layer.

[0067] M l It is the number of tanks in the l-th processing layer.

[0068] S l It is the number of tank layers between process layer l and l+1.

[0069] K s l It is the number of tanks in the s-th tank layer located between the l-th and (l+1)-th process layers.

[0070] W s l It is the connection matrix between the l-th and (l+1)-th process layers of the tank system, and the connection matrix between the s-th and (s+1)-th tank layers.

[0071] v t l,sIt is the flow between the s-th and (s+1)-th tank layers located between the l-th and (l+1)-th processing layers.

[0072] The flow balance constraint at the tank can be determined by the following formula.

[0073] (1)

[0075] in,

[0076] The flow balance constraints at the plant are linked to the output of the regression function: l = 2, … , L; p = 1, …, V s Determined by the following formula

[0077] (2)

[0078] Where i is the sequence number position of the product in the p-th process at the s-th process level.

[0079] Inputs linked to the regression function: l = 2, … , L; p = 1, … , V s

[0080] (3)

[0081] The zerotrim(v) function is used to trim all zero elements in the following vector:

[0082] Regression constraints can be defined by the following formula:

[0083] (4).

[0084] An example decoding algorithm is shown below.

[0085]

[0086] Translation:

[0087]

[0088] refer to Figure 11 The diagram illustrates a continuous model of the regression function. For piecewise linear partitioning models, such as decision trees (DT), multivariate adaptive regression splines (MARS), and FFN, the leaf nodes can be traced and the gradients computed. The function values ​​can then be calculated from the regression function.

[0089] Optimization Algorithm

[0090] This disclosure provides three approaches for three different types of regression functions: piecewise linear regressors, gradient-based regressors, and black-box regressors.

[0091] Mixed-integer formula for piecewise linear regression

[0092] For partitioned regression based on piecewise linear methods, including linear regression and decision trees, the regression function can be formulated as a mixed-integer linear procedure. The main idea is to apply this to each regression function f. l Use precise mixed-integer linear representation.

[0093] Decision tree regression y = h t (x) includes leaf nodes and branch nodes. L is represented by a set of leaf nodes, and B is represented by a set of branch nodes. For each leaf node... The form is r(x) = w l T x + c l The linear model has been learned from the training data based on the points assigned to the leaf nodes. Branch nodes From hyperplane a l T x + b l This means that if a l T x + b l If the eigenvector is less than 0, then point x will follow the left branch from that node; otherwise, it splits into the right branch. Because the topology of the tree is fixed, for each eigenvector x, there exists a path from the root to the leaf node of the tree. The only path. Let N be. L (l) represents the ancestor node of l, where the left branch follows on the path from the root to l. Let N be the ancestor node of l. R (l) represents the set of ancestors of the right branch on this path. (Binary variable) The instruction indicates that if x is assigned to a leaf node l, then e l =1.

[0094] Choose exactly one leaf node for data point x

[0095] (6)

[0096] To determine the unique path route to the leaf node, in the indicator variable e l With the help of [the relevant authority], the following constraints are applied to the split modeling at the branch nodes.

[0097] (7)

[0098] Here, M1 and M2 are very large numbers. The decision tree regression y = h(x) can be represented as a mixed-integer bilinear model.

[0099] (8)

[0100] This process can be further linearized by the bilinear term c as follows: l (w l T x + c l For some constants y l L , y l U Assume y l L <= w l T x + c l <=_y l U This assumption is reasonable because x is usually bounded; for example, normalizing the training data to the range of 0 - 1, w l and c l These are fixed model parameters. Therefore, the process has y l = c l (w l T x + c l ) is equivalent to

[0101] (9)

[0102] Therefore, decision tree regression y = h t The mixed-integer linear representation of (x) is as follows

[0103] (10)

[0104] When these linear-based representations are inserted into single-period and multi-period models, the MILP formula is obtained, which can be efficiently solved to the optimum by the MILP solver.

[0105] When the regression function is nonlinear, this disclosure proposes to calculate the function value and gradient of the regression quantity. Single-period and multi-period problems are solved using gradient-based algorithms, such as the augmented Lagrange algorithm or the alternating direction multiplier (ADMM) algorithm.

[0106] For illustrative purposes, various embodiments of this teaching have been described, but these descriptions are not intended to be exhaustive or limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein has been chosen to best explain the principles of the embodiments, their practical application, or technical improvements to technologies found in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

[0107] While what is considered the best state and / or other examples have been described above, it should be understood that various modifications may be made therein, and the subject matter disclosed herein can be implemented in various forms and examples, and the teachings can be applied to many applications, only some of which are described here. The following claims are intended to claim any and all applications, modifications, and variations that fall within the true scope of this teaching.

[0108] The components, steps, features, purposes, benefits, and advantages discussed herein are merely illustrative. They, and the discussions related to them, are not intended to limit the scope of protection. While various advantages have been discussed herein, it should be understood that not all embodiments are necessarily required to include all advantages. Unless otherwise stated, all measurements, numerical values, grades, positions, sizes, and other specifications set forth in this specification (including the appended claims) are approximate, not precise. They are intended to have a reasonable range consistent with the functions they address and the conventions of the art to which they belong.

[0109] Many other embodiments have also been considered. These include embodiments with fewer, additional, and / or different components, steps, features, purposes, benefits, and advantages. These also include embodiments in which components and / or steps are arranged and / or ordered in different ways.

[0110] This document describes aspects of the disclosure with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0111] These computer-readable program instructions may be provided to a processor of a suitably configured computer, special-purpose computer, or other programmable data processing apparatus to produce a machine such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / actions specified in one or more blocks of a flowchart and / or block diagram. These computer-readable program instructions may also be stored in a computer-readable storage medium that can instruct a computer, programmable data processing apparatus, and / or other device to operate in a certain manner, such that the computer-readable storage medium in which the instructions are stored includes an article of manufacture comprising instructions for implementing aspects of the functions / actions specified in one or more blocks of a flowchart and / or block diagram.

[0112] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device, thereby producing a computer-implemented process, such that the instructions, which execute on the computer, other programmable apparatus or other device, perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0113] The call flows, flowcharts, and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, fragment, or portion of instructions, including one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a non-consecutive order. For example, two blocks shown consecutively may actually execute substantially simultaneously, or these blocks may sometimes execute in reverse order, depending on the functions involved. It will also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented by a system based on dedicated hardware that performs the specified function or action or executes a combination of dedicated hardware and computer instructions.

[0114] While the foregoing has been described in conjunction with exemplary embodiments, it should be understood that the term "exemplary" is merely an example and not the best or optimal. Nothing other than that stated or described above is intended or should not be construed as offering any component, step, feature, object, benefit, advantage, or equivalent to the public, whether or not it is stated in the claims.

[0115] It should be understood that the terms and expressions used herein have the general meaning consistent with their respective fields of investigation and research, unless otherwise specified herein. Relational terms such as "first" and "second" may be used merely to distinguish one entity or action from another, without necessarily requiring or implying any actual such relationship or order between these entities or actions. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element beginning with "a" or "an" does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes that element.

[0116] An abstract of this disclosure is provided to allow the reader to quickly determine the nature of the technical disclosure. It is submitted on the understanding that it will not be used to interpret or limit the scope or meaning of the claims. Furthermore, as can be seen from the foregoing detailed description, various features have been combined in various embodiments for the purpose of simplifying this disclosure. This method of disclosure should not be construed as reflecting an intention that the claimed embodiments have more features than expressly recited in each claim. Rather, as reflected in the following claims, the inventive subject matter is less than all the features of a single disclosed embodiment. Therefore, the following claims are thus incorporated into the detailed description, each claim on its own as a separately claimed subject matter.

Claims

1. A computer-implemented method, comprising: Map the process flow graph of the factory process to a graph that includes nodes and edges, where nodes represent processes and edges represent flows between processes; The behavior of each node in the graph is learned based at least on historical data of the plant process, including sensor measurements of flow rate, temperature, or pressure in the plant. Model one or more regression functions for each node to predict the output of each in the factory process, wherein the one or more regression functions are modeled based on the learned behavior of each node, and the output includes output stream or product purity; The network topology is encoded by generating an adjacency matrix for the graph, the adjacency matrix identifying interconnected nodes; as well as An optimization model is generated based on the regression function and the adjacency matrix of each node. The optimization model provides a setpoint for each node to control the plant process, wherein the setpoint includes temperature, pressure, or flow rate.

2. The computer-implemented method according to claim 1, wherein the input of each node includes controlled variables and uncontrolled variables.

3. The computer-implemented method of claim 1 further includes using one or more matrices and the positions of nodes and edges from the graph to decode the network topology to draw a directed acyclic graph for input verification.

4. The computer-implemented method according to claim 1, wherein the optimization model provides optimized setpoints for one or more parameters for each node.

5. The computer-implemented method according to claim 4, wherein the model generator outputs a continuous optimization model having a function value estimator and a gradient estimator.

6. The computer-implemented method of claim 4, wherein the model generator outputs a mixed-integer linear program for modeling the one or more regression functions.

7. The computer-implemented method according to claim 4 further includes the option of providing different key performance indicators as input and generating setpoints to achieve the input key performance indicators.

8. The computer-implemented method according to claim 1 further includes automatically determining whether the input of the regression function is correct.

9. The computer-implemented method according to claim 1, wherein the graph is a two-level fully connected feedforward network.

10. The computer-implemented method according to claim 1, wherein the regression function includes piecewise linear and nonlinear regression models.

11. A computer-implemented method for automatically generating optimization models for site-wide optimization of a plant, comprising: Define a graphical representation of the process flow graph of the plant, the graphical representation including nodes and edges, wherein the nodes represent processes and the edges represent flows between processes; The network topology of the graphical representation is encoded to generate an adjacency matrix of the graphical representation, the adjacency matrix identifying interconnected nodes; Automatically generate a set of equations that define the network topology; A machine learning platform is used to model one or more regression functions at each node to predict the output of each process in the plant based on inputs received at each process. This modeling utilizes historical data of the plant's processes, including sensor measurements of flow rate, temperature, or pressure. The inputs include controlled and uncontrolled variables, and the outputs include output stream or product purity. An optimization model is generated based on the one or more regression functions of each node and the adjacency matrix. The optimization model provides setpoints for controlling each process of the plant, wherein the setpoints include temperature, pressure, or flow rate.

12. The computer-implemented method of claim 11, wherein the graphical representation is a two-level fully connected feedforward network without the assumption of skip layers.

13. The computer-implemented method of claim 11, wherein the encoding of the network topology is performed with fewer inputs than the inputs of the factory process.

14. The computer-implemented method of claim 11, further comprising using a matrix and the positions of nodes and edges from the graph representation to decode the network topology to draw a directed acyclic graph for input verification.

15. The computer-implemented method of claim 11, wherein the model generator outputs a continuously optimized model having a function value estimator and a gradient estimator.

16. The computer-implemented method of claim 11, wherein the model generator outputs a mixed-integer linear program for modeling the one or more regression functions.

17. A non-transitory computer-readable storage medium comprising computer-readable program code having computer-readable instructions, wherein, when the computer-readable program code is executed, a computer device performs a method for improving the computational efficiency of the computing device, for automatically generating an optimization model for site-wide plant optimization, the method comprising: A graphical representation of a process flow graph for a factory process is defined, comprising nodes and edges, wherein the nodes of the graphical representation represent processes, and the edges of the graphical representation represent flows between factory processes; The network topology represented by the graph is encoded by generating an adjacency matrix for the graph, the adjacency matrix identifying interconnected nodes; Automatically generate a set of equations that define the network topology; Modeling one or more regression functions at each node using a machine learning platform to predict the output of each process in the plant based on inputs received at each process, the modeling utilizing historical data of the plant's processes, including sensor measurements of flow rate, temperature, or pressure; the inputs include controlled and uncontrolled variables; and the outputs include output stream or product purity. An optimization model is generated from the one or more regression functions of each node and the adjacency matrix, the optimization model providing setpoints for controlling each process of the plant, wherein the setpoints include temperature, pressure or flow rate.

18. The non-transitory computer-readable storage medium of claim 17, wherein executing the code by a processor also configures the computing device to perform actions including decoding the network topology using a matrix and the positions of nodes and edges from the graph to draw a directed acyclic graph for input validation.

19. The non-transitory computer-readable storage medium of claim 17, wherein the model generator outputs a continuously optimized model having a function value estimator and a gradient estimator.

20. The non-transitory computer-readable storage medium of claim 17, wherein the model generator outputs a mixed-integer linear program for modeling the one or more regression functions.

Images