Adaptive deep learning predictive control method and device for process industry

Abstract

The application provides a kind of adaptive deep learning predictive control method and device suitable for process industry.It includes: according to the historical control information of process industry, based on deep learning technology, the full-condition dynamic model of target control variable is constructed, based on the adaptive segmented algorithm of threshold, the change gain curve of full-condition dynamic model is segmented, and the segmented linear gain model is obtained;Determine the segmented linear variable constraint of segmented linear gain model, and determine the optimization goal of target control variable based on segmented linear variable constraint;Based on segmented linear gain model, the step response sequence of target control variable at the current working point is obtained, and the future control execution sequence of target control variable is predicted according to step response sequence and optimization goal.Solve the problem that the mapping relationship between control variable and controlled variable is usually constructed based on linear model in existing predictive control method, which cannot accurately reflect the nonlinear characteristics of control variable, and further leads to the problem of low accuracy of predictive control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of adaptive deep learning predictive control of process industry, in particular to an adaptive deep learning predictive control method and device suitable for process industry, a computer readable storage medium and an electronic device. BACKGROUND

[0002] Model predictive control (MPC) is an efficient control method, which is widely used in the field of industrial processes, especially in the process industry such as petrochemical and chemical industry, and has fully demonstrated its effectiveness. By constructing a prediction model of the controlled object, combining real-time state feedback and rolling optimization in the future time domain, the dynamic balance between temperature, pressure, flow and other variables can be effectively coordinated under the premise of ensuring safety constraints, greatly improving the operation efficiency and stability of the device. However, like many advanced technologies, MPC is not perfect and has some limitations in actual application. The production devices in process control industry (such as chemical, petroleum, refining, pharmaceutical, etc.) usually have complex physical, chemical and thermodynamic processes. For example, the Arrhenius equation of reaction rate in a chemical reactor presents an exponential dependence on temperature nonlinear relationship; the separation efficiency of a distillation column presents a strong coupling nonlinear relationship between feed components and reflux ratio. In addition, different production indicators, such as production load and market demand, may cause the on-site device to present multiple working condition switching. This makes it difficult to achieve precise optimization control for the prediction control relying on the model, and the modeling of the mechanism-based model is complex and lacks test data. Therefore, in the actual construction process, one or more local linear models are often selected for approximation. Therefore, some improved predictive control methods aiming at handling working conditions are widely proposed. Their goal is to establish multiple models for each operating mode and design a model switching strategy. However, in actual industrial processes, due to the complexity of the environment, it is difficult to accurately distinguish each mode from offline data, which further reduces the control performance and robustness; and the adaptive predictive control method in the prior art usually constructs a mapping relationship between the operating variables (independent variables) and the controlled variables (dependent variables) based on a linear model, which will result in that the model cannot accurately reflect the nonlinear characteristics of the process, and thus the problem of low accuracy of predictive control. SUMMARY

[0003] The main purpose of the present application is to provide an adaptive deep learning predictive control method and device suitable for process industry, a computer readable storage medium and an electronic device, to at least solve the problem that the existing predictive control method usually constructs a mapping relationship between the operating variables and the controlled variables based on a linear model, which cannot accurately reflect the nonlinear characteristics of the control variables, and thus leads to the problem of low accuracy of predictive control.

[0004] In order to achieve the above object, according to one aspect of the present application, there is provided an adaptive deep learning predictive control method suitable for process industry, comprising: constructing a full-condition dynamic model of a target control variable based on deep learning technology according to historical control information of the process industry, wherein the target control variable is composed of independent variables and dependent variables, and the independent variables and the dependent variables are in a nonlinear coupling relationship; segmenting a change gain curve of the full-condition dynamic model based on a threshold adaptive segmentation algorithm to obtain a segmented linear gain model; determining segmented linear variable constraints of the segmented linear gain model, and determining an optimization target of the target control variable based on the segmented linear variable constraints, wherein the optimization target is composed of independent variable optimization values and dependent variable optimization values; obtaining a step response sequence of the target control variable at a current working condition point based on the segmented linear gain model, and predicting a future control execution sequence of the target control variable according to the step response sequence and the optimization target, wherein the future control execution sequence is used to control the independent variables of the target control variable.

[0005] According to another aspect of the present application, there is provided an adaptive deep learning predictive control device suitable for process industry, comprising: a first construction unit configured to construct a full-condition dynamic model of a target control variable based on deep learning technology according to historical control information of the process industry, wherein the target control variable is composed of independent variables and dependent variables, and the independent variables and the dependent variables are in a nonlinear coupling relationship; a segmentation processing unit configured to segment a change gain curve of the full-condition dynamic model based on a threshold adaptive segmentation algorithm to obtain a segmented linear gain model; a determination unit configured to determine segmented linear variable constraints of the segmented linear gain model, and determine an optimization target of the target control variable based on the segmented linear variable constraints, wherein the optimization target is composed of independent variable optimization values and dependent variable optimization values; and a prediction unit configured to obtain a step response sequence of the target control variable at a current working condition point based on the segmented linear gain model, and predict a future control execution sequence of the target control variable according to the step response sequence and the optimization target, wherein the future control execution sequence is used to control the independent variables of the target control variable.

[0006] According to still another aspect of the present application, there is provided a computer readable storage medium comprising a stored program, wherein the computer readable storage medium controls a device where the computer readable storage medium is located to execute any one of the adaptive deep learning predictive control methods suitable for process industry when the program runs.

[0007] According to a further aspect of the present application, an electronic device is provided, comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs comprise a program for performing any one of the adaptive deep learning predictive control methods suitable for process industry.

[0008] By applying the technical solution of the present application, the target control variable full-condition dynamic model constructed by deep learning technology can accurately capture the complex nonlinear coupling relationship between independent variables and dependent variables, effectively covering the dynamic characteristics under various conditions such as raw material fluctuation, load adjustment, and equipment aging in process industry. Combined with the adaptive segmentation algorithm based on threshold, the nonlinear gain curve is converted into a segmented linear model, which not only simplifies the calculation complexity in model predictive control, but also ensures the applicability and accuracy of the model under different conditions. By determining the segmented linear variable constraint, the independent variables and dependent variables can be optimized specifically, achieving dual improvement of economic benefit and operation safety. In particular, the predictive control framework is constructed based on the real-time step response sequence, which can dynamically adjust the future control execution sequence to ensure that the independent variables of the target control variable always move towards the optimization target, even when facing sudden condition changes. The problem of low prediction control accuracy caused by the fact that existing predictive control methods usually construct the mapping relationship between the operation variable and the controlled variable based on linear models, which cannot accurately reflect the nonlinear characteristics of the control variable, is solved. BRIEF DESCRIPTION OF DRAWINGS

[0009] The accompanying drawings, which form a part of the present description, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of the present application, and their

[0010] Figure 1 A hardware structure block diagram of a mobile terminal for performing an adaptive deep learning predictive control method suitable for process industry is shown according to an embodiment of the present application;

[0011] Figure 2 A flowchart of an adaptive deep learning predictive control method suitable for process industry is shown according to an embodiment of the present application;

[0012] Figure 3 A flowchart of an adaptive deep learning predictive control method suitable for process industry is shown according to an embodiment of the present application;

[0013] Figure 4 A step response curve of a deep learning model under different conditions is shown according to an embodiment of the present application;

[0014] Figure 5 A flowchart illustrating the steady-state optimization algorithm provided according to an embodiment of this application is shown;

[0015] Figure 6 A flowchart of a dynamic control algorithm provided according to an embodiment of this application is shown;

[0016] Figure 7 A structural block diagram of an adaptive deep learning predictive control device suitable for process industries, provided according to an embodiment of this application, is shown. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] For ease of description, the following explains some of the nouns or terms used in the embodiments of this application:

[0021] MPC, or Model Predictive Control, is a special type of control. Its current control action is obtained at each sampling instant by solving a finite-time open-loop optimal control problem.

[0022] Deep learning specifically refers to machine learning based on deep neural network models and methods. It has evolved from statistical machine learning, artificial neural networks, and other algorithmic models, combined with the development of modern big data and high-performance computing. The most important technical characteristic of deep learning is its ability to automatically extract features.

[0023] Step response sequence: refers to the zero-state response of a system after receiving a unit step function input. The zero-state response means that the system is in its initial state before receiving the specified input, that is, the response change of the system is entirely due to the specified input (in this case, a unit step input).

[0024] As described in the background section, existing predictive control methods typically construct the mapping relationship between control variables and controlled variables based on linear models, which cannot accurately reflect the nonlinear characteristics of control variables, resulting in low predictive control accuracy. To address the problem that existing predictive control methods typically construct the mapping relationship between control variables and controlled variables based on linear models, which cannot accurately reflect the nonlinear characteristics of control variables and thus result in low predictive control accuracy, embodiments of this application provide an adaptive deep learning predictive control method, apparatus, computer-readable storage medium, and electronic device suitable for the process industry.

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0026] The methods and embodiments provided in this application can be executed on a mobile terminal, a computer terminal, or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 This is a hardware structure block diagram of a mobile terminal for an adaptive deep learning predictive control method applicable to the process industry, according to an embodiment of the present invention. Figure 1 As shown, a mobile terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. The mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the mobile terminal described above. For example, the mobile terminal may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0027] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the adaptive deep learning predictive control method for process industries in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the mobile terminal via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The transmission device 106 is used to receive or send data via a network. Specific examples of the aforementioned networks may include wireless networks provided by the mobile terminal's communication provider. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can be connected to other network devices via a base station to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0028] This embodiment provides an adaptive deep learning predictive control method suitable for the process industry, which runs on a mobile terminal, computer terminal, or similar computing device. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0029] Figure 2 This is a flowchart of an adaptive deep learning predictive control method applicable to the process industry, according to an embodiment of this application. Figure 2 As shown, the method includes the following steps:

[0030] Step S201: Based on the historical control information of the process industry, construct a full-condition dynamic model of the target control variables using deep learning technology. The target control variables consist of independent variables and dependent variables, and the independent variables and dependent variables are non-linearly coupled.

[0031] Process industries, including chemical, petroleum, and power industries, often involve a large number of complex control variables in their production processes. These variables have nonlinear coupling relationships, which traditional linear models struggle to accurately describe. Deep learning technologies, including but not limited to neural networks, can learn from historical control information, such as the relationships between independent variables like temperature, pressure, and flow rate, and dependent variables like product quality and energy consumption, to construct more accurate dynamic models covering all operating conditions.

[0032] Step S202: Based on the threshold-based adaptive segmentation algorithm, the change gain curve of the above full-condition dynamic model is segmented to obtain a segmented linear gain model.

[0033] Specifically, employing a threshold-based adaptive piecewise algorithm to linearize the change gain curve of the full-condition dynamic model piecewise can significantly improve the model's applicability and the control system's performance. The main advantage of this technique lies in its ability to decompose complex nonlinear gain curves into multiple simple linear segments. Each linear segment accurately describes the system behavior within a specific operating range, thus simplifying the model, reducing computational complexity and response time, and improving the model's accuracy and adaptability under different operating conditions. This enables the control system to respond more flexibly and efficiently to dynamic changes in the industrial environment, reducing resource waste, improving the overall control strategy optimization level, and ultimately enhancing system stability and production efficiency.

[0034] Step S203: Determine the piecewise linear variable constraints of the above piecewise linear gain model, and determine the optimization objective of the above target control variable based on the above piecewise linear variable constraints, wherein the above optimization objective consists of the optimization values ​​of the independent variable and the optimization values ​​of the dependent variable;

[0035] Specifically, by precisely setting constraints on piecewise linear variables, such as the range of each linear interval and slope limits, we can ensure that the model maintains reasonable linearity while avoiding overfitting under different operating conditions. The goal of this optimization is to maximize control efficiency and minimize control error while meeting system stability and performance requirements. Technically, this method allows control systems to more intelligently adjust their operating parameters to respond to rapidly changing industrial environments. For example, in chemical production, when faced with fluctuations in key parameters such as temperature and pressure, the piecewise linear gain model can instantly identify trends and automatically adjust heater power or valve openings based on the optimization objective. This not only maintains stable process operation but also optimizes energy consumption, achieving energy conservation, emission reduction, and cost control. Furthermore, this automatic adjustment based on optimization objectives effectively reduces human intervention, avoids operational errors, and further improves the safety and automation level of industrial production.

[0036] Step S204: Based on the piecewise linear gain model, obtain the step response sequence of the target control variable at the current operating point, and predict the future control execution sequence of the target control variable according to the step response sequence and the optimization target, wherein the future control execution sequence is used to control the independent variable of the target control variable.

[0037] In this embodiment, by applying steps S201, S202, S203, and S204, the target control variable dynamic model constructed using deep learning technology can accurately capture the complex nonlinear coupling relationship between independent and dependent variables, effectively covering the dynamic characteristics under various changing conditions in the process industry, such as raw material fluctuations, load adjustments, and equipment aging. Combined with a threshold-based adaptive piecewise algorithm, the nonlinear gain curve is transformed into a piecewise linear model, which not only simplifies the computational complexity in model predictive control but also ensures the applicability and accuracy of the model under different operating conditions. By determining the piecewise linear variable constraints, the independent and dependent variables can be optimized in a targeted manner, achieving a dual improvement in economic efficiency and operational safety. In particular, the predictive control framework constructed based on real-time step response sequences can dynamically adjust future control execution sequences, ensuring that the independent variable of the target control variable always moves towards the optimization objective, and can respond quickly even when facing sudden changes in operating conditions. This solves the problem that existing predictive control methods typically construct the mapping relationship between the operated variable and the controlled variable based on linear models, which cannot accurately reflect the nonlinear characteristics of the control variable, thus leading to low predictive control accuracy.

[0038] In the specific implementation process, an adaptive segmentation algorithm based on a threshold is used to segment the changing gain curve of the above-mentioned full-condition dynamic model to obtain a piecewise linear gain model. This includes: obtaining N segmented intervals by uniformly sampling the operating points of the changing gain curve; determining the gain value of each segmented interval; and determining the mean gain of the changing gain curve using the gain values ​​of each segmented interval, where N > 1; calculating the approximate deviation of each segmented interval based on the mean square error, using the mean gain and the gain values ​​of each segmented interval; and dividing the segments with approximate deviations less than a deviation threshold. The segment interval is marked as a valid segment and stored in the segment set; the cutting step is as follows: the target segment interval is divided into a left sub-interval and a right sub-interval. The target segment interval is cut by a genetic optimization algorithm with the goal of minimizing the total approximate deviation of the left sub-interval and the right sub-interval, resulting in two sub-segment intervals. The target segment interval represents the segment interval or sub-segment interval where the approximate deviation is greater than or equal to the deviation threshold; the cutting step is repeated until the approximate deviation of all the sub-segment intervals is less than the deviation threshold, thus obtaining the piecewise linear gain model.

[0039] This method employs a threshold-based adaptive piecewise algorithm, which effectively segments the changing gain curve to obtain a more realistic piecewise linear gain model, further improving the optimization effect of the control strategy. In process industries, the changing gain curve of control variables is often not linear but changes with the operating point. By uniformly sampling the operating points of the changing gain curve to obtain N segmented intervals, the complex nonlinear changing gain curve can be decomposed into multiple linear segments, thus simplifying the calculation of the control strategy. A genetic optimization algorithm is used to cut the target segmented intervals to achieve the goal of minimizing the total approximation deviation, which not only improves the accuracy of the model but also enhances its robustness. Through the combination of various technical features, namely the combination of the adaptive piecewise algorithm and the genetic optimization algorithm, the technical problem of accurately modeling the nonlinear changes of the changing gain curve of control variables in process industries is solved, improving the optimization effect and robustness of the control strategy.

[0040] More specifically, determining the piecewise linear variable constraints for the above piecewise linear gain model includes:

[0041] According to the constraint formula:

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049] Determine the piecewise linear variable constraints for the above piecewise linear gain model, where, The optimization action quantity for the above independent variables. , Let these be the minimum and maximum changes of the independent variables mentioned above. , These are the upper and lower limits of the actions of the above independent variable. , Let M be the minimum and maximum single change of the dependent variable, and i be the effect on the i-th dependent variable. This represents the gain of the k-th operating condition on the i-th dependent variable. Let the optimization cost be the cost of the i-th dependent variable. Optimize the increment for the independent variable. For the action quantity of the above independent variable, This refers to the single change in the dependent variable mentioned above. , This represents the upper and lower limits of the dependent variable's operations. To optimize the initial point, This is a segmentation flag. This represents the gain relationship between the independent and dependent variables. For the k-th segmentation point, For segmented optimization variables.

[0050] This method, which defines the piecewise linear variable constraints of the piecewise linear gain model, is a crucial step in ensuring the feasibility of the control strategy in practical applications. In process industries, constraints on control variables often include, but are not limited to, the minimum and maximum changes in the independent variable, and the single minimum and maximum changes in the dependent variable. Setting these constraints prevents the control strategy from exceeding the actual control range of the equipment during execution, thus ensuring its feasibility. Through the interaction of various technical features—the combination of the piecewise linear gain model and the piecewise linear variable constraints—the technical problem of control strategies potentially exceeding the equipment control range in practical applications in process industries is solved, ensuring the feasibility and stability of the control strategy.

[0051] Furthermore, based on the aforementioned step response sequence and the aforementioned optimization objective, the future control execution sequence of the aforementioned target control variable is predicted, including:

[0052] According to the formula:

[0053]

[0054] The above future control execution sequence is determined, wherein, The residuals of the above dependent variable and reference trajectory, For the weights of the dependent variable, This represents the algorithm's prediction of the future value of the dependent variable. For the current time, Let j represent the j-th dependent variable mentioned above, Ncv be the number of dependent variables, P be the number of steps to be predicted, and r be the future time. j This represents the reference trajectory for the expected value of the j-th dependent variable mentioned above. For the motion quantity of step i The penalty is given by M, where M is the control time domain and Nmv is the number of the aforementioned independent variables. The action penalty weight is the independent variable. For the above future control execution sequence, For the future action sequence that needs to be optimized, For the moment of prediction of the future, For open-loop prediction sequences, Let be the step response sequence between the i-th independent variable and the j-th dependent variable at operating point k. For the minimum increment of the action of the independent variable, The maximum increment of the action of the independent variable. To The future prediction value at time, The amount of motion to be applied at a future time l. The lower limit of the independent variable. The upper limit of the independent variable, This is the lower limit of the operation on the dependent variable. This represents the upper limit of operations on the dependent variable.

[0055] This method enables optimized control of control variables in process industries by predicting the future control execution sequence of the target control variable. The step response sequence reflects the dynamic response characteristics of the independent variable to the dependent variable, while the optimization objective includes the expected control values ​​for both the dependent and independent variables. By predicting the future control execution sequence, the control actions of the independent variable can be adjusted in advance to achieve the optimization objective, such as reducing energy consumption or improving product quality. In process industries, the optimization of control strategies often requires considering the combined effects of multiple dependent variables and the penalty costs of independent variable control actions; therefore, the calculation of the predicted control execution sequence needs to comprehensively consider these factors. Through the interaction of various technical characteristics—namely, the prediction of the step response sequence, the optimization objective, and the future control execution sequence—this method solves the technical problem of the difficulty in adjusting control strategies in advance to achieve optimization objectives in process industries, thereby improving the optimization effect and efficiency of control strategies.

[0056] Furthermore, after predicting the future control execution sequence of the target control variable based on the step response sequence and the optimization objective, the method further includes: obtaining real-time control information of the target control variable and updating the real-time control information into the full-condition dynamic model.

[0057] This method updates the full-condition dynamic model in real time, enabling the control strategy to quickly adapt to the dynamic changes in the process industry, thus enhancing the robustness and stability of the control system. In practical applications, the production processes in the process industry often undergo dynamic changes, such as equipment aging and raw material changes, which affect the relationships between control variables. By acquiring real-time control information of the control variables and updating it to the full-condition dynamic model, these changes can be reflected in a timely manner, thereby adjusting the control strategy and ensuring control effectiveness. Through the combination of various technical features—namely, the acquisition of real-time control information and the updating of the full-condition dynamic model—the technical problem of the difficulty in quickly adapting control strategies to the dynamic changes in the production process in the process industry is solved, enhancing the robustness and stability of the control system.

[0058] Specifically, in the process of constructing a full-condition dynamic model of the target control variable based on the historical control information of the process industry and using deep learning technology, the above method also includes: constructing a full-condition linear model of other control variables based on the historical control information of the process industry and using an identification algorithm, wherein the independent variables and dependent variables of the other control variables have a linear coupling relationship, and the identification algorithm includes the FIR algorithm and the subspace algorithm.

[0059] This method constructs a full-condition linear model of other control variables, which can further refine the control strategy and improve control effectiveness. In process industries, in addition to the target control variable, there are other control variables, and the relationship between these variables and the dependent variable may be linearly coupled. By constructing a full-condition linear model of these control variables based on the identification algorithm, the control process in the process industry can be described more comprehensively, thereby improving the optimization effect of the control strategy. The FIR algorithm and the subspace algorithm can construct a linear model using historical control information, such as the input-output data between independent and dependent variables. Through the interaction of various technical features, namely the construction of the full-condition linear model by the identification algorithm, the technical problem of accurately modeling the linear coupling relationship between other control variables and the dependent variable in the process industry is solved, thus refining the control strategy and improving control effectiveness.

[0060] More specifically, after constructing a full-condition linear model of other control variables based on the identification algorithm, the above method further includes: integrating the above full-condition linear model and the above full-condition dynamic model into the controller model as sub-models with control variables as units, to form a hybrid controller model, wherein the above control variables include the above target control variables and other control variables.

[0061] This method integrates a full-condition linear model and a full-condition dynamic model to form a hybrid controller model, enabling optimized control of complex control processes in process industries. In practical applications, process industry control processes often involve both linear and nonlinear coupling relationships. By integrating the full-condition linear and dynamic models into the controller model, both linear and nonlinear relationships can be considered simultaneously, thereby improving the optimization effect of the control strategy. Furthermore, the construction of the hybrid controller model enhances the versatility and adaptability of the control strategy, allowing it to be applied to a wider range of control scenarios. Through the interaction of various technical features—the integration of the full-condition linear model, the full-condition dynamic model, and the hybrid controller model—the technical problem of control processes in process industries involving both linear and nonlinear coupling relationships is solved, achieving optimized control of complex processes and improving the versatility and adaptability of the control strategy.

[0062] Furthermore, this embodiment also includes: introducing a meta-learning mechanism to enhance the model's generalization ability. In the training phase of the deep learning prediction model, in addition to traditional supervised learning methods, a meta-learning mechanism is introduced. Meta-learning is a learning method that enables a model to quickly adapt to new tasks with limited data. Specifically, a model-agnostic meta-learning (MAML) framework can be used. By adjusting the initialization method of the model parameters, it can achieve good prediction results by updating through fast gradient descent when receiving a small amount of new task data.

[0063] During the pre-training phase, a large amount of known operating condition data is used to find a "meta-initialization point" through the MAML method. This initialization point allows the model to be efficiently fine-tuned with only a small amount of data when facing unknown operating conditions. In actual deployment, when encountering new operating conditions, a portion of the real-time data is extracted as a "rapid adaptation" dataset for meta-learning. Through several fine-tuning iterations, the model can respond quickly and accurately to new operating conditions.

[0064] Suppose this method is applied to the control system of a distillation column in a chemical plant. The separation efficiency of the distillation column is affected not only by temperature and pressure, but also closely related to the feed composition, which frequently changes during plant operation, causing uncertainty in operating conditions. Through a meta-learning mechanism, our deep learning model, when faced with suddenly changing feed compositions, quickly adjusts its prediction model with only a few days of data, keeping the prediction accuracy of the control system within ±3%, far superior to the weeks or even months of time and resources required to retrain the model.

[0065] This embodiment reduces the need for extensive data collection and processing required to adapt to new operating conditions, improving the flexibility of responding to emergencies in industrial settings. It also effectively avoids delays and errors in control decisions.

[0066] This embodiment also includes a priority-based dynamic update strategy for control variables. During the dynamic predictive control phase, it is considered that the real-time updates of certain control variables (such as the opening degree of safety valves) are crucial to overall operational efficiency and safety. We have developed a priority-based dynamic update strategy that categorizes control variables into different priorities based on their importance and response speed. For high-priority variables, a shorter prediction window and a higher update frequency are used; while for low-priority variables, a longer prediction window and a lower update frequency are used.

[0067] This method improves the response speed of key control variables by dynamically adjusting the parameter configuration of predictive control, while reducing the consumption of computing resources. This strategy is particularly important when dealing with large-scale industrial processes.

[0068] Specific Implementation Plan: In a catalytic cracking unit of an oil refinery, temperature control is crucial for ensuring chemical reaction efficiency and product quality, while pressure regulation is relatively secondary. By applying a priority-based dynamic update strategy for control variables, the prediction window for temperature control was set to 10 time steps, while the prediction window for pressure control was extended to 20 time steps. This strategy reduced the response time of temperature control by 50%, maintaining process parameters within the target range even when feedstock composition and market conditions change rapidly, significantly improving product consistency and unit safety.

[0069] The rapid response of the key control variables in this embodiment effectively prevents potential production failures and safety incidents. By rationally allocating computing resources, the computational burden of non-critical variables is reduced, the overall execution efficiency of the control algorithm is improved, and unnecessary over-regulation is reduced, thereby lowering energy consumption and maintenance costs.

[0070] To enable those skilled in the art to better understand the technical solution of this application, the implementation process of the adaptive deep learning predictive control method applicable to the process industry will be described in detail below with reference to specific embodiments.

[0071] To address the model mismatch problem faced by traditional MPC in complex process industries due to its multi-condition and time-varying characteristics, especially the shortcomings of multi-model MPC such as ambiguous mode division and rigid switching strategies caused by offline data limitations.

[0072] This embodiment relates to a specific adaptive deep learning predictive control method applicable to process industries. Its core objective is to overcome the limitations of fixed or locally linear models by combining deep fusion of data-driven modeling and model predictive control to achieve full-condition coverage and real-time adaptive control of the dynamic characteristics of complex industrial processes. For example... Figure 3 As shown, it specifically includes the following:

[0073] Step 1: Controller Modeling: Hybrid Modeling Strategy;

[0074] A hybrid model architecture incorporating linear and deep learning models is constructed. A deep learning prediction model is used to characterize the nonlinearity, time-varying nature, and dynamic characteristics across multiple variables under various operating conditions. The specific steps are as follows:

[0075] The dynamic processes in general process industries can be represented by the following formula: ;

[0076] Where F is a linear or nonlinear function, k represents discrete time, N represents the model length of the dynamic process, and Y0 represents the baseline measurement value of the controlled variable. U and Y represent the manipulated variable (independent variable) and the controlled variable (dependent variable), respectively.

[0077] First, based on the control scheme and prior knowledge, if the dynamic response of a certain CV variable (the controlled variable, which is the dependent variable) over the entire operating range is known to be characterized by a single linear model, the linear model can be identified and obtained using traditional identification algorithms such as FIR and subspace methods. The structure of the linear model is as follows: A(t) is the step response sequence.

[0078] If the dynamic response of the CV variable across the entire operating condition range cannot be represented by a single linear model, a deep learning model can be used to build a dynamic model covering the entire operating condition range. This model is obtained by inputting discrete data samples. Specifically, offline collected data can be sampled over an appropriate period T. s History window length T past and prediction window length T fut The related input variables u(t) and output variables y(t) are organized in a multiple-input single-output (MISO) format to form a time-series dataset. This dataset is then input into a deep learning-based system step response identification system. Simultaneously, the unit step response at each operating point, output by its unit step response prediction submodule, can be obtained. Figure 4 As shown.

[0079] ,in, For deep learning models, This represents the operating point at time t. This represents the controlled variable corresponding to the operating condition point at time t.

[0080] Finally, the traditional linear model and the deep learning model are integrated into the controller model as sub-models, with the controlled variable as the unit, to form a hybrid controller model. : Where i represents the i-th CV variable, For traditional linear models, It is a deep learning model.

[0081] Step 2: Piecewise linearization of global gain:

[0082] In the process industry, predictive control often employs a two-layer structure: the upper layer is for steady-state target calculation or steady-state optimization, and the lower layer is for dynamic control. Two-layer predictive control, compared to traditional predictive control, better ensures system stability and safety benefits.

[0083] The control section of this embodiment will adopt a two-layer control architecture of steady-state optimization and dynamic control, while traditional two-layer predictive control algorithms only support linear model solutions. To support efficient optimization of the hybrid model at the steady-state layer, an adaptive piecewise linear method needs to be designed for the dynamically changing gain curves of the entire operating condition represented by the deep learning model. This method converts the dynamically changing steady-state gain relationship of the entire operating condition into an analytical piecewise linear model, while retaining the steady-state gain change characteristics of key operating points. The specific process is as follows:

[0084] A full-condition step response sequence is established based on a deep learning model, and its steady-state gain is calculated for each operating point. : ;

[0085] Where Us and Ys represent the actual operating conditions, and based on the steady-state gain distribution across all operating conditions, assuming the process is linear within a threshold range or confidence interval, the global gain curve of the deep learning model under all operating conditions can be represented as a linear piecewise model. Based on this idea, a threshold-based adaptive piecewise algorithm is designed.

[0086] First, define the entire operating condition range as... The initial set of segments is .

[0087] Then, M operating points are uniformly sampled within the interval. The average gain within the interval is obtained. : The approximate deviation is calculated using the mean square error (MSE). ;

[0088] If the MSE is less than the set threshold, the current interval is marked as a valid segment and stored in the segment set; if the deviation exceeds the threshold, a candidate segmentation point l is selected within the interval to divide the interval into a left sub-interval and a right sub-interval; the optimal cutting point that minimizes the total MSE at both ends after a single cut can be quickly searched using a genetic optimization algorithm. :

[0089] ,in, Let be the mean square error of the left sub-interval. The mean square error of the right sub-interval;

[0090] This continues until the deviations in all intervals are within the threshold or the preset maximum number of segments is reached. A piecewise linear gain model can be obtained through linear partitioning: ;

[0091] Indicates the interval indicator function (when Take 1 at a time. Let be the gain of the k-th segment.

[0092] Step 3: Steady-state optimization of the hybrid model, such as... Figure 5 As shown, it specifically includes the following:

[0093] In the two-layer predictive control architecture of the process industry, the steady-state optimization layer mainly utilizes the process steady-state model to solve for the steady-state optimal solution that satisfies process constraints and control objectives, thereby providing targets for subsequent dynamic control and meeting some of the user's benefit requirements. In this embodiment, the steady-state gain curve of the deep model is approximated using a piecewise linear model using the method in step two before being used in the steady-state optimization calculation.

[0094] First, due to the introduction of a piecewise linear model, binary variables need to be defined in the optimization problem. This indicates whether the current operating condition is in segment k. Simultaneously, the following independent constraints must be satisfied:

[0095] , representing the kth segment of the segmented interval where the current operating point is located;

[0096] For each segment k, linearize the effect of the input on the controlled variable: Where i represents the impact on the i-th controlled variable. This represents the gain of the k-th operating condition on the i-th controlled variable, and then the interval constraints are further refined: ;

[0097] M is a sufficiently large constant. This constraint, combined with the uniqueness constraint, indicates that only the unique interval constraint will actually take effect. Furthermore, piecewise linear variable constraints and linear variable constraints are integrated:

[0098] This indicates that the economic objective is to minimize the amount of action required for operations.

[0099] , which represents the increment of the operated variable;

[0100] , which represents the positional constraint of the operand;

[0101] , which represents the increment of the controlled variable;

[0102] This is represented as a positional constraint on the controlled variable;

[0103] , which is a unique constraint on the piecewise linear range of values;

[0104] ;

[0105] in, Optimize the amount of motion for the amount of operation; and These represent the minimum and maximum changes in the amount of work performed. and These are the upper and lower limits of the operational quantity; and This is the maximum single change of the controlled variable, and is generally not set; among which... This represents the influence of the nonlinear model on the controlled variable, while This represents the influence of the linear model on the controlled variable. By solving the above steady-state optimization problem, the actual optimization objective (U, Y) can be obtained.

[0106] Step 4: Dynamic predictive control, such as Figure 6 As shown, it specifically includes the following:

[0107] After obtaining the global steady-state optimization objective, it is necessary to further formulate dynamic control sequences for the manipulated variables and controlled variables to reach the optimal operating point. The predictive control framework used in this invention is dynamic matrix control, which requires predictive control based on step response sequences. This algorithm can well meet the requirements of computational efficiency and control accuracy in industrial settings. To achieve predictive control calculations for the hybrid model, it is necessary to construct dynamic matrix calculations by locally linearizing the deep model. The specific content includes the following:

[0108] Obtain the step response sequence from the current steady-state point to the optimization target. The linear model has a fixed step response sequence, which can be directly obtained from the identification algorithm in step one. However, the step response process of the deep prediction model, due to its multi-condition and nonlinear characteristics, depends not only on the current steady-state position but also on the subsequent direction of movement and the steady-state target. Therefore, after obtaining the optimization target, it is necessary to obtain the step response sequence in real time based on the deep prediction model. The specific method is as follows: First, based on real historical data, construct the future control variable U for the standard step. fut for:

[0109] .

[0110] in, This involves calculating the required action quantities for the given scenario. The constructed future control quantities are input into the deep prediction model for feedforward prediction, yielding a set of predicted quantities. After further processing, the current dynamic sequence can be inferred. ;

[0111] In the formula, A (k) M represents the step response sequence of the controlled variable at the current operating point. deep This is the deep learning model trained in step one. Then, based on the real-time acquired step response sequence, a predictive control dynamic programming proposition can be constructed:

[0112] ;

[0113] The optimization objective consists of two parts. Let be the residual between the controlled variable and the reference trajectory, where j represents the j-th controlled variable, Ncv is the number of controlled variables, P represents the number of steps to be predicted, and r j This represents the reference trajectory expected by the j-th controlled variable; For the motion quantity of step l The penalty is denoted by M, where M is the control time domain and Nmv is the number of operands. Within the constraints, For open-loop prediction sequences, real-time prediction using a deep prediction model can be obtained. (k) ij Let be the step response sequence between the operated variable i and the controlled variable j at operating point k. The desired future input sequence increment is defined by the following three constraints: upper and lower limits of the manipulated variable, the rate of change, and the upper and lower limits of the controlled variable. By solving the above optimization problem, a set of future control variables can be obtained. This refers to the control execution sequence of the control variables for the next P cycles. The first step in the algorithm's issuance of the control quantity sequence is executed.

[0114] Step 5: Online Incremental Update: Although deep learning models have achieved dynamic characteristic coverage of all operating conditions through large-scale offline data, in actual industrial scenarios, sudden changes in raw material properties (such as coal ash content exceeding thresholds in coal chemical raw materials), equipment state drift (such as catalyst activity decay), or extreme operating conditions (such as sudden load fluctuations) may still exceed the training data distribution, causing the model to deviate. Therefore, a real-time online incremental model update mechanism is needed to enhance the algorithm's adaptive capability. Specifically, a sliding window mechanism is used to store and clean real-time data. Once the data volume reaches the training requirements, the real-time data is input into the deep time series prediction model. It is important to note that an elastic weight regularization term is introduced to balance new and old knowledge and avoid sudden changes in model parameters. The model accuracy is then fine-tuned using online real-time data.

[0115] , where L new For the new loss function, This indicates the parameters that need to be updated. The weights are for the regularization terms. Fi() is an example of the regularization expression. Steps two through five are then repeated. During the control process, data is collected in real time to update the deep prediction model. Based on the deep prediction model, the model information relied upon in the control process is corrected and optimized, thereby achieving stable control across all operating conditions in the process industry.

[0116] This embodiment employs a deep time-series model at the upper layer to capture the complex dynamic characteristics of process industries, including multivariate nonlinear coupling, time-varying operating condition transitions, and long-period lag effects. The lower layer constructs a predictive control model using a locally linearized model with real-time inference to generate the optimal control action sequence. Furthermore, this embodiment proposes a model hybrid strategy that integrates the deep time-series predictive model and the linear model. Variables with stable dynamic response characteristics across all operating conditions are handled using the FIR / subspace model method to obtain an FIR model with a constant step response. Variables with dynamic characteristics that significantly change with operating points are handled using a deep predictive model to establish a nonlinear mapping of the input-output relationship. This embodiment also proposes a predictive control algorithm adapted to the deep time-series predictive model. The steady-state optimization layer considers the gain across all operating conditions to generate the optimal operating objective for all conditions; the dynamic layer infers the locally linear sequence in real-time and solves for the optimal control increment that satisfies the constraints in the rolling time domain.

[0117] This embodiment specifically achieves the following technical effects:

[0118] 1. A deep learning-based two-layer prediction-control coupling architecture is proposed: the upper layer employs a deep time-series prediction model, which can accurately characterize the nonlinearity, time-varying nature, and dynamic characteristics of multiple variables under all operating conditions; the lower layer extracts the step response model of the current operating point through real-time inference and calculates the optimal control action. Compared with traditional single linear models, this embodiment effectively solves the model mismatch problem caused by changes in system dynamic characteristics after operating condition switching in industrial processes, significantly improving control robustness under multiple operating conditions. Furthermore, compared with directly using deep learning models for control calculations, the output is more interpretable, safety is guaranteed, and computational efficiency is greatly improved.

[0119] 2. Integration of Linear and Deep Learning Models: The stable dynamic response characteristics of variables across all operating conditions are modeled as a single step response, while variables exhibiting significant dynamic changes are modeled using deep neural networks. This hybrid strategy inherits the computational efficiency of traditional linear models while leveraging deep learning to overcome its limitations in dynamically changing scenarios, balancing model accuracy and real-time performance requirements. Furthermore, this integration strategy is adaptable to a wider range of industrial processes, providing a foundation for large-scale deployment.

[0120] 3. Enhanced Control Accuracy and Adaptability through Full-Condition Nonlinear Optimization: Steady-State Optimization: Based on the global nonlinear gain curve of the deep model, a nonlinear programming solver is constructed. Compared with traditional linearized steady-state optimization methods, it can more accurately capture the coupling relationship between variables, and the optimized output results are more accurate. Dynamic Control: By acquiring the step response sequence of the current operating point in real time, the predictive model parameters are dynamically corrected, enabling the controller to adapt to complex scenarios such as operating condition drift and time-varying parameters, significantly improving the adaptive capability and anti-interference performance of multi-step predictive control.

[0121] This application also provides an adaptive deep learning predictive control device suitable for process industries. It should be noted that this adaptive deep learning predictive control device for process industries can be used to execute the adaptive deep learning predictive control method for process industries provided in this application. This device is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0122] The following describes the adaptive deep learning predictive control device for the process industry provided in the embodiments of this application.

[0123] Figure 7 This is a schematic diagram of an adaptive deep learning predictive control device suitable for the process industry, according to an embodiment of this application. Figure 7 As shown, the device includes:

[0124] The first building unit 71 is used to build a full-condition dynamic model of the target control variables based on the historical control information of the process industry and deep learning technology. The target control variables are composed of independent variables and dependent variables, and the independent variables and dependent variables have a non-linear coupling relationship.

[0125] The segmented processing unit 72 is used for a threshold-based adaptive segmentation algorithm to segment the change gain curve of the above-mentioned full-condition dynamic model to obtain a segmented linear gain model.

[0126] The determining unit 73 is used to determine the piecewise linear variable constraints of the above piecewise linear gain model, and to determine the optimization objective of the above target control variable based on the above piecewise linear variable constraints, wherein the above optimization objective consists of the optimization values ​​of the independent variable and the optimization values ​​of the dependent variable;

[0127] The prediction unit 74 is used to obtain the step response sequence of the target control variable at the current operating point based on the piecewise linear gain model, and to predict the future control execution sequence of the target control variable based on the step response sequence and the optimization target, wherein the future control execution sequence is used to control the independent variable of the target control variable.

[0128] In this embodiment, the first construction unit is used to construct a full-condition dynamic model of the target control variable based on historical control information of the process industry and deep learning technology. The target control variable consists of independent and dependent variables, which are non-linearly coupled. The piecewise processing unit is used to process the change gain curve of the full-condition dynamic model in pieces using an adaptive piecewise algorithm based on thresholds, resulting in a piecewise linear gain model. The determination unit is used to determine the piecewise linear variable constraints of the piecewise linear gain model and determine the optimization objective of the target control variable based on these constraints. The optimization objective consists of the optimized values ​​of the independent and dependent variables. The prediction unit is used to obtain the step response sequence of the target control variable at the current operating point based on the piecewise linear gain model, and predict the future control execution sequence of the target control variable based on the step response sequence and the optimization objective. The future control execution sequence is used to control the independent variable of the target control variable. The full-condition dynamic model of the target control variable constructed using deep learning technology can accurately capture the complex non-linear coupling relationship between the independent and dependent variables, effectively covering the dynamic characteristics under various changing conditions in the process industry, such as raw material fluctuations, load adjustments, and equipment aging. By combining a threshold-based adaptive piecewise algorithm, the nonlinear gain curve is transformed into a piecewise linear model. This not only simplifies the computational complexity of model predictive control but also ensures the model's applicability and accuracy under different operating conditions. By defining piecewise linear variable constraints, the independent and dependent variables can be optimized in a targeted manner, achieving a dual improvement in economic efficiency and operational safety. In particular, constructing a predictive control framework based on real-time step response sequences allows for dynamic adjustment of future control execution sequences, ensuring that the independent variable of the target control variable always moves towards the optimization objective, enabling rapid response even in the face of sudden changes in operating conditions. This solves the problem that existing predictive control methods typically construct the mapping relationship between the manipulated and controlled variables based on linear models, failing to accurately reflect the nonlinear characteristics of the control variables and thus leading to low predictive control accuracy.

[0129] As an optional scheme, the segmented processing unit includes a first determining module, a calculation module, a marking module, a segmentation module, and a repetitive execution module; the first determining module is used to obtain N segmented intervals by uniformly sampling the operating points of the aforementioned variable gain curve, determine the gain value of each segmented interval, and determine the average gain of the aforementioned variable gain curve through the gain values ​​of each segmented interval, wherein N>1; the calculation module is used to calculate the approximate deviation of each segmented interval based on the mean square error, according to the average gain and the gain values ​​of each segmented interval; the marking module is used to mark the segmented intervals whose approximate deviation is less than the deviation threshold. The segments are marked as valid and stored in the segment set. The cutting module is used to perform the cutting steps: the target segment interval is divided into a left sub-interval and a right sub-interval. The target segment interval is cut by a genetic optimization algorithm with the goal of minimizing the total approximate deviation of the left sub-interval and the right sub-interval, resulting in two sub-segment intervals. The target segment interval represents the segment interval or sub-segment interval where the approximate deviation is greater than or equal to the deviation threshold. The repeat execution module is used to repeat the cutting steps until the approximate deviation of all the sub-segment intervals is less than the deviation threshold, thus obtaining the piecewise linear gain model.

[0130] In one alternative approach, the determining unit includes a second determining module, used to determine the constraint formula:

[0131]

[0132]

[0133]

[0134]

[0135]

[0136]

[0137]

[0138] Determine the piecewise linear variable constraints for the above piecewise linear gain model, where, The optimization action quantity for the above independent variables. , Let these be the minimum and maximum changes of the independent variables mentioned above. , These are the upper and lower limits of the actions of the above independent variable. , Let M be the minimum and maximum single change of the dependent variable, and i be the effect on the i-th dependent variable. This represents the gain of the k-th operating condition on the i-th dependent variable. Let the optimization cost be the cost of the i-th dependent variable. Optimize the increment for the independent variable. For the action quantity of the above independent variable, This refers to the single change in the dependent variable mentioned above. , This represents the upper and lower limits of the dependent variable's operations. To optimize the initial point, This is a segmentation flag. This represents the gain relationship between the independent and dependent variables. For the k-th segmentation point, For segmented optimization variables.

[0139] In one alternative approach, the prediction unit includes a third determining module for determining the formula:

[0140]

[0141] The above future control execution sequence is determined, wherein, The residuals of the above dependent variable and reference trajectory, For the weights of the dependent variable, This represents the algorithm's prediction of the future value of the dependent variable. For the current time, Let j represent the j-th dependent variable mentioned above, Ncv be the number of dependent variables, P be the number of steps to be predicted, and r be the future time. j This represents the reference trajectory for the expected value of the j-th dependent variable mentioned above. For the motion quantity of step i The penalty is given by M, where M is the control time domain and Nmv is the number of the aforementioned independent variables. The action penalty weight is the independent variable. For the above future control execution sequence, For the future action sequence that needs to be optimized, For the moment of prediction of the future, For open-loop prediction sequences, Let be the step response sequence between the i-th independent variable and the j-th dependent variable at operating point k. For the minimum increment of the action of the independent variable, The maximum increment of the action of the independent variable. To The future prediction value at time, The amount of motion to be applied at a future time l. The lower limit of the independent variable. The upper limit of the independent variable, This is the lower limit of the operation on the dependent variable. This represents the upper limit of operations on the dependent variable.

[0142] In an optional embodiment, the apparatus further includes an acquisition unit, configured to acquire real-time control information of the target control variable after predicting the future control execution sequence of the target control variable based on the step response sequence and the optimization objective, and update the real-time control information into the full-condition dynamic model.

[0143] In an optional embodiment, the apparatus further includes a second building unit, used to construct a full-condition dynamic model of the target control variable based on deep learning technology and historical control information of the process industry, and to construct a full-condition linear model of other control variables based on an identification algorithm based on the aforementioned historical control information of the process industry, wherein the independent and dependent variables of the other control variables have a linear coupling relationship, and the identification algorithm includes an FIR algorithm and a subspace algorithm.

[0144] In an alternative embodiment, the apparatus further includes an integration unit, which, after constructing a full-condition linear model of other control variables based on the identification algorithm, integrates the full-condition linear model and the full-condition dynamic model into the controller model as sub-models based on control variables, forming a hybrid controller model, wherein the control variables include the target control variable and other control variables.

[0145] The aforementioned adaptive deep learning predictive control device for process industries includes a processor and a memory. The first construction unit, segmented processing unit, determination unit, and prediction unit are all stored as program units in the memory. The processor executes these program units stored in the memory to achieve the corresponding functions. All of the above modules reside in the same processor; alternatively, the modules may be located in different processors in any combination.

[0146] The processor contains a kernel, which retrieves the corresponding program unit from memory. One or more kernels can be configured. By adjusting kernel parameters, the problem of low predictive control accuracy can be addressed. Existing predictive control methods typically rely on linear models to construct the mapping relationship between control and controlled variables, failing to accurately reflect the nonlinear characteristics of the control variables.

[0147] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0148] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the adaptive deep learning predictive control method applicable to the process industry.

[0149] This invention provides a processor for running a program, wherein the program executes the adaptive deep learning predictive control method applicable to process industries.

[0150] This invention provides an electronic device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements at least the steps of the adaptive deep learning predictive control method applicable to the process industry described above. The device described herein may be a server, PC, PAD, mobile phone, etc.

[0151] This application also provides a computer program product that, when executed on a data processing device, is adapted to perform a program that initializes at least the steps described above for an adaptive deep learning predictive control method applicable to process industries.

[0152] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0153] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0154] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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 program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, 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, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0155] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0156] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0157] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0158] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0159] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0160] It should also be noted that 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 elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0161] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An adaptive deep learning predictive control method suitable for process industries, characterized in that, include: Based on historical control information from the process industry, a dynamic model of the target control variables under all operating conditions is constructed using deep learning technology. The target control variables consist of independent variables and dependent variables, and the independent variables and dependent variables are nonlinearly coupled. An adaptive segmentation algorithm based on thresholds is used to segment the changing gain curve of the dynamic model under all operating conditions to obtain a segmented linear gain model. Determine the piecewise linear variable constraints of the piecewise linear gain model, and determine the optimization objective of the target control variable based on the piecewise linear variable constraints, wherein the optimization objective consists of the optimized values ​​of the independent variables and the optimized values ​​of the dependent variables; The step response sequence of the target control variable at the current operating point is obtained based on the piecewise linear gain model. The future control execution sequence of the target control variable is predicted based on the step response sequence and the optimization target. The future control execution sequence is used to control the independent variable of the target control variable. An adaptive piecewise algorithm based on a threshold is used to segment the changing gain curve of the full-condition dynamic model to obtain a piecewise linear gain model, including: By uniformly sampling the operating points of the variable gain curve, N segmented intervals are obtained, the gain value of each segmented interval is determined, and the average gain of the variable gain curve is determined by the gain value of each segmented interval, where N > 1. Based on the mean square error, the approximate deviation of each segment interval is calculated according to the mean gain and the gain value of each segment interval; The segmented intervals whose approximate deviation is less than the deviation threshold are marked as valid segments and stored in the segment set; Cutting steps: Divide the target segmented interval into a left sub-interval and a right sub-interval. Use a genetic optimization algorithm to cut the target segmented interval with the goal of minimizing the total approximate deviation between the left sub-interval and the right sub-interval, to obtain two sub-segmented intervals. The target segmented interval represents the segmented interval or sub-segmented interval where the approximate deviation is greater than or equal to the deviation threshold. Repeat the cutting step until the approximate deviation of all the sub-segment intervals is less than the deviation threshold, and then obtain the piecewise linear gain model.

2. The method according to claim 1, characterized in that, Determining the piecewise linear variable constraints of the piecewise linear gain model includes: According to the constraint formula: ； ； ； Determine the piecewise linear variable constraints for the piecewise linear gain model, wherein, This means minimizing the amount of action required for the economic objective, which is a non-linear independent variable. The optimized action quantity is for a nonlinear independent variable, and the optimized action quantity lies between the minimum and maximum changes. The sum of the optimized action quantity and the action quantity of the nonlinear independent variable lies between the upper and lower limits of the action quantity of the nonlinear independent variable. , Let i represent the minimum and maximum single change of the dependent variable, and let i represent the effect on the i-th nonlinear independent variable. Let represent the gain of the k-th operating condition with respect to the i-th nonlinear independent variable. Let be the optimization cost of the i-th nonlinear independent variable. Optimize the increment for linear independent variables. The single change of the dependent variable. , This represents the upper and lower limits of the dependent variable's operations. To optimize the initial point, This is a segmentation flag.

3. The method according to claim 1, characterized in that, Predicting the future control execution sequence of the target control variable based on the step response sequence and the optimization objective includes: According to the formula: ； Determine the future control execution sequence, wherein, The residuals between the dependent variable and the reference trajectory, For the weights of the dependent variable, This represents the algorithm's prediction of the future value of the dependent variable. For the current time, Let j represent the j-th dependent variable, Ncv be the number of dependent variables, P be the number of steps to be predicted, and r be the number of future moments. j This represents the reference trajectory of the expected value of the j-th dependent variable. For the motion quantity of step i The penalty, M is the control time domain, and Nmv is the number of independent variables. The action penalty weight is the independent variable. For the future control execution sequence, For the future action sequence that needs to be optimized, A moment for predicting the future.

4. The method according to claim 1, characterized in that, After predicting the future control execution sequence of the target control variable based on the step response sequence and the optimization objective, the method further includes: Obtain real-time control information of the target control variable and update the real-time control information into the full-condition dynamic model.

5. The method according to claim 1, characterized in that, In the process of constructing a full-condition dynamic model of the target control variables based on historical control information from the process industry and using deep learning technology, the method further includes: Based on the historical control information of the process industry, a full-condition linear model of other control variables is constructed based on the identification algorithm, wherein the independent and dependent variables of the other control variables are linearly coupled. The identification algorithm includes the FIR algorithm and the subspace algorithm.

6. The method according to claim 5, characterized in that, After constructing a full-condition linear model of other control variables based on the identification algorithm, the method further includes: The full-condition linear model and the full-condition dynamic model are integrated into the controller model as sub-models with control variables as the unit, forming a hybrid controller model, wherein the control variables include the target control variable and other control variables.

7. An adaptive deep learning predictive control device suitable for process industries, characterized in that, include: The first construction unit is used to construct a full-condition dynamic model of the target control variables based on the historical control information of the process industry and deep learning technology. The target control variables consist of independent variables and dependent variables, and the independent variables and dependent variables have a non-linear coupling relationship. The segmented processing unit is used to segment the changing gain curve of the full-condition dynamic model into segments using an adaptive segmentation algorithm based on a threshold, thereby obtaining a segmented linear gain model. A determining unit is used to determine the piecewise linear variable constraints of the piecewise linear gain model, and to determine the optimization objective of the target control variable based on the piecewise linear variable constraints, wherein the optimization objective consists of the optimized values ​​of the independent variables and the optimized values ​​of the dependent variables; The prediction unit is used to obtain the step response sequence of the target control variable at the current operating point based on the piecewise linear gain model, and predict the future control execution sequence of the target control variable based on the step response sequence and the optimization target, wherein the future control execution sequence is used to control the independent variable of the target control variable; The segmented processing unit includes a first determining module, a calculation module, a marking module, a cutting module, and a repetitive execution module. The first determining module is used to obtain N segmented intervals by uniformly sampling the operating points of the changing gain curve, determine the gain value of each segmented interval, and determine the mean gain of the changing gain curve using the gain values ​​of each segmented interval, where N > 1. The calculation module is used to calculate the approximate deviation of each segmented interval based on the mean square error, according to the mean gain and the gain values ​​of each segmented interval. The marking module is used to mark the segmented intervals whose approximate deviation is less than a deviation threshold as... The effective segmentation is stored in a segment set; the cutting module is used to perform the cutting steps: dividing the target segment interval into a left sub-interval and a right sub-interval, and cutting the target segment interval using a genetic optimization algorithm with the goal of minimizing the total approximate deviation of the left sub-interval and the right sub-interval, to obtain two sub-segment intervals, wherein the target segment interval represents the segment interval or sub-segment interval where the approximate deviation is greater than or equal to the deviation threshold; the repeat execution module is used to repeatedly execute the cutting steps until the approximate deviation of all the sub-segment intervals is less than the deviation threshold, thus obtaining the piecewise linear gain model.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the adaptive deep learning predictive control method for process industries as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including methods for performing the adaptive deep learning predictive control method for process industries as described in any one of claims 1 to 6.

Images