Regulating valve prediction model construction method based on instantaneous multi-opening response

Abstract

The invention relates to a regulating valve prediction model construction method based on instantaneous multi-opening response. According to the method, a multi-opening-degree three-dimensional model is constructed based on actual structure parameters, and flow field data including outlet pressure, steam volume fraction and outlet speed are obtained through CFD transient simulation. The time, the opening degree and the inlet speed are used as input, the outlet parameters are used as output, a multi-layer feedforward neural network is adopted to establish an agent model, and a Levenberg-Marquardt algorithm is used for training and optimization. The established model can quickly predict the valve flow response under any opening degree and time conditions, has the characteristics of high efficiency, good precision, strong generalization and the like, and is suitable for real-time state prediction, intelligent control and optimization design of the regulating valve.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fluid machinery modeling, computational fluid dynamics (CFD) and artificial intelligence, and in particular to a method for constructing a proxy model suitable for predicting the flow response of a regulating valve at different opening degrees and transient states, which is widely applicable to intelligent regulating systems in the tobacco, chemical, energy and other industries. BACKGROUND

[0002] As one of the most commonly used and critical actuators in industrial process control systems, regulating valves play an important role in fluid transportation, flow regulation, pressure control and energy conversion in various engineering links. The change of its opening degree directly determines the flow rate, pressure and phase distribution of the medium flowing through the valve, and has a significant impact on the dynamic response and operational stability of the entire system. Especially in conditions such as tobacco re-humidification, steam control, chemical transportation, and energy fluid distribution, which require high precision regulation, accurate modeling and prediction of regulating valve performance become one of the key technologies to ensure stable and safe operation of the process.

[0003] Traditional methods for evaluating the performance of regulating valves mainly rely on experimental testing and computational fluid dynamics (CFD) simulation. Although experimental methods are intuitive and have a certain degree of authenticity, they are usually costly, time-consuming, and difficult to control parameters, and cannot meet the requirements of rapid and repetitive analysis under multiple variables, multiple opening degrees and multiple working conditions. In contrast, CFD technology has the advantages of flexibility, high precision and strong visualization, but its calculation process is complex, resource consumption is large, and simulation time is long, making it difficult to meet the actual engineering needs of real-time prediction, online control or multi-scheme optimization.

[0004] In addition, regulating valves have typical nonlinear dynamic characteristics during operation: the relationship between opening degree and pressure response is complex, especially in the small opening degree throttling area or the rapid regulation process, where phenomena such as turbulent flow evolution, flow field disturbance, and gas-liquid interface movement exhibit strong transient and coupling characteristics. Under this background, modeling methods based on traditional physical models are difficult to effectively characterize these nonlinear coupling behaviors, and the flow field laws under different opening degrees do not have obvious linear predictability, thereby increasing the difficulty of modeling and prediction.

[0005] In recent years, with the development of artificial intelligence and data-driven methods, the use of machine learning algorithms to construct "surrogate models" has become a new solution. This method collects a portion of representative CFD simulation data, uses statistical learning methods to extract the nonlinear mapping relationship between input and output, and realizes fast approximate prediction of the target output parameters without changing the original physical meaning. Among them, multi-layer feedforward neural networks (FNN) have strong expression ability and nonlinear fitting advantages, and have become one of the mainstream methods for constructing surrogate models.

[0006] However, existing research has largely focused on predictions under single operating conditions or steady-state conditions, lacking systematic modeling methods for problems involving "multiple opening degrees + instantaneous response." Especially during dynamic changes in opening degree, the flow field response at different time points is highly nonlinear, making it difficult to efficiently represent using traditional methods. How to construct a control valve proxy model that can encompass complex data from multiple operating conditions and possesses both prediction efficiency and generalization ability is a key technical challenge that urgently needs to be overcome in the fields of intelligent simulation and fluid control. Summary of the Invention

[0007] This invention aims to propose a method for constructing a predictive model for control valves based on instantaneous multi-opening response. By combining 3D modeling, CFD transient multiphase flow simulation, and deep neural network surrogate model technology, it enables rapid flow field response prediction of control valves under multiple operating conditions and time nodes. This breaks through the bottleneck of traditional CFD simulation calculation being time-consuming and difficult to apply in real time, and significantly improves the prediction efficiency and response accuracy of intelligent control systems for industrial control valves.

[0008] The technical solution of this invention is as follows: A method for constructing a predictive model for a control valve based on instantaneous multi-opening response includes the following steps: S1: Establish a three-dimensional model of the control valve and parameterize the opening degree; S2: Perform geometric simplification on the 3D model; S3: Configure CFD transient simulation; S4: Collect flow field data; S5: Data preprocessing; S6: Define input / output mapping; S7: Construct a neural network proxy model; S8: Train the neural network model; S9: Evaluate model performance; S10: Enables rapid prediction.

[0009] Furthermore, step S1 specifically includes: First, a three-dimensional solid model of the control valve, including key components such as the valve body, valve core, and flow channel, is constructed. Using the 3D modeling software SolidWorks, a complete 3D solid model is created based on the actual structural parameters. A mapping relationship between the valve core position and the valve opening degree is established through parametric driving, so as to generate geometric variants under different opening degrees in batches, ensuring the consistency of geometric features between the virtual model and the physical entity.

[0010] Furthermore, step S2 specifically includes: Secondly, a geometric simplification strategy based on the principle of physical equivalence is adopted to improve computational efficiency while retaining key flow characteristics. Specifically, this includes deleting or replacing non-critical small features (such as chamfers), using partitioned modeling or symmetrical boundary conditions for symmetrical structures, and replacing thin-walled and small curvature radii with simplified profiles, thereby reducing the number of meshes and computational resource consumption.

[0011] Furthermore, step S3 specifically includes: The multi-aperture 3D geometric model was imported into CFD software, the fluid domain and computational mesh were set up, and multiphase flow and turbulence models were configured. In Fluent software, the fluid domain was established and the mesh was generated. A multiphase flow model was selected to simulate the interaction between steam and air. A Realizable k-ε turbulence model was used, and boundary conditions such as inlet velocity, pressure, and temperature were set. The time step and total computation time were determined to capture unsteady-state characteristics.

[0012] Furthermore, step S4 specifically includes: Then, during the simulation, monitoring surfaces were arranged at the outlet section and key areas to record transient flow field parameters. Monitoring surfaces were deployed in the transient simulation to record parameters such as outlet pressure, outlet velocity, and steam volume fraction at each time step. Data from all opening degrees and time nodes were then compiled to construct a multi-dimensional time-series dataset.

[0013] Furthermore, step S5 specifically includes: The simulation data is cleaned and standardized to improve data quality and training stability. Noise reduction, outlier removal, and normalization are performed to reduce dimensional differences and remove outliers, resulting in a standardized training dataset.

[0014] Furthermore, step S6 specifically includes: The input and output variables of the surrogate model are clearly defined, and a complete mapping relationship reflecting the valve's fluid dynamic state is constructed. Using valve opening, time, and inlet velocity as input features, and outlet pressure, steam volume fraction, and flow rate as output targets, an input-output mapping function is defined to capture the time-varying and nonlinear characteristics of the flow process.

[0015] Furthermore, step S7 specifically includes: A multi-layer feedforward neural network is used as the core structure of the surrogate model to achieve high-precision fitting of nonlinear mappings. An FNN structure with two hidden layers is designed, using the tanh activation function in the hidden layer and a linear activation function in the output layer, to fully exploit the nonlinear relationships between features and ensure the continuity of the prediction results.

[0016] Furthermore, step S8 specifically includes: The Levenberg-Marquardt algorithm was used to train and optimize the surrogate model to improve its convergence speed and stability. After initializing the network parameters, the Jacobian matrix was calculated based on the error, the damping factor was dynamically adjusted, and the parameters were iteratively updated until the loss function reached a preset threshold or the maximum number of iterations. Finally, the model's generalization ability was tested using a validation set.

[0017] Furthermore, step S9 specifically includes: The accuracy and robustness of the surrogate model were systematically verified using a multi-index comprehensive evaluation method. Training, validation, and test sets were divided, and the coefficient of determination R0 was used. 2 The model is evaluated using metrics such as root mean square error (RMSE) and mean absolute percentage error (MAPE) to ensure it performs well under unknown operating conditions.

[0018] Furthermore, step S10 specifically includes: Based on a pre-trained surrogate model, the outlet parameters of the control valve at any opening degree and time point can be predicted in milliseconds, meeting the requirements of real-time control. This enables rapid prediction of outlet flow field parameters at any given opening degree and instantaneous time point, significantly improving the response speed and intelligence level of industrial control.

[0019] Compared with the prior art, the present invention has the following effects: 1. This invention effectively overcomes the shortcomings of traditional simulation models, such as low computational efficiency and poor real-time performance, by integrating three technologies: 3D modeling, CFD transient simulation, and deep learning surrogate model. This enables rapid and high-precision prediction of the instantaneous response of control valves at multiple opening degrees. Specifically: the 3D model highly replicates the valve's structural features, ensuring the physical authenticity of the data and the accuracy of the calculations; the transient CFD simulation meticulously captures the coupling characteristics of multiphase flow and turbulence, providing a solid physical foundation for the data-driven model; the neural network surrogate model has a reasonable structure, excellent training effect, and strong nonlinear fitting and generalization capabilities; and the prediction speed is fast, enabling real-time response to the control system's rapid assessment needs for valve status, significantly improving the intelligence level and operational efficiency of industrial automation systems.

[0020] 2. In step S1, all models maintain a unified structural reference plane and boundary naming convention to facilitate batch import into CFD software for automated simulation. Parametric drive control not only improves model generation efficiency but also provides interface support for CAD-CFD data integration and the definition of input variables for proxy models. In subsequent data modeling, the opening value... It will be used as one of the input features in the neural network algorithm for high-precision prediction training.

[0021] 3. In step S2, by deleting minor features (such as threads, small chamfers, and marking grooves), a large number of fine meshes are avoided in these areas. The overall mesh count is reduced by approximately 30%-60%, directly reducing computation time and storage usage. Using a symmetric model reduces the simulation area by half or more, shrinking the fluid computation domain and making it less prone to severe local gradients in the flow field, simplifying turbulence modeling and significantly reducing the number of computational units. Simulation time is shortened by approximately 40%-70%, and the number of iterations is reduced by more than 20% while maintaining the same residual convergence criteria.

[0022] 4. In step S3, by setting up a VOF multiphase flow model and a Realizable k-ε turbulence model, and combining transient solutions with local mesh refinement, this study successfully captured the steam-air flow characteristics of the control valve under different opening conditions. The simulation process employed a small time step and convergence residual control to ensure the stability and reliability of the numerical solution. Furthermore, by providing time-series outputs of pressure, velocity, and volume fraction, high-resolution data support was provided for the construction of the surrogate model. The simulation results demonstrate the model's efficient performance in response dynamic capture, turbulent structure reconstruction, and multiphase interface tracking.

[0023] 5. In step S7, the FNN can achieve high-precision nonlinear mapping. The first two hidden layers implement nonlinear feature mapping, predicting multiple output response variables from input features. The tanh activation function maps the input to the [-1,1] interval, improving training stability and alleviating the gradient vanishing problem. The network as a whole can generate corresponding predictions of outlet pressure, velocity, and steam volume fraction based on the instantaneous valve opening, time, and inlet velocity of the input, achieving multi-output regression.

[0024] 6. In step S8, the LM algorithm combines the advantages of gradient descent and the Gauss-Newton method, enabling rapid convergence in tasks with small samples, multiple outputs, and high nonlinearity, while ensuring training stability. During training, the iteration direction is controlled by automatically adjusting the damping factor μ, achieving high-precision fitting of outlet pressure, velocity, and steam volume fraction. The final surrogate model's MSE is less than 1e-4, indicating that the model has extremely strong fitting and predictive capabilities for simulation data. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the prior art and the implementation scheme will be briefly introduced below.

[0026] Figure 1 Flowchart for constructing an instantaneous agent model for a control valve.

[0027] Figure 2 This is a diagram of the neural network model structure.

[0028] Figure 3 This is a flowchart of the Levenberg–Marquardt algorithm.

[0029] Figure 4 Implement a case process for the proxy model. Detailed Implementation

[0030] To better understand the technical solution of this invention, the following case study details the construction process of a predictive surrogate model for a control valve based on instantaneous multi-opening response. This case study uses a pneumatic diaphragm control valve used in an industrial rehumidifier as the research object, and fully implements 3D modeling, Fluent CFD transient simulation, and predictive surrogate model construction.

[0031] A method for constructing a predictive model for a control valve based on instantaneous multi-opening response includes the following steps: S1: Establish a 3D model of the control valve and parameterize the opening degree. A complete 3D model of the control valve, including its main components such as valve body, valve core, and inlet / outlet flow channels, was created using SolidWorks software. A parametric driving method was used to establish a mapping relationship between the valve core position and the valve opening degree, and five typical opening degrees were selected:

[0032] Generate corresponding geometric variant models in batches.

[0033] S2: Geometric simplification of the 3D model To improve computational efficiency, the model was geometrically simplified based on the principle of physical equivalence. Specifically, this included removing small features such as local chamfers that had negligible impact on global flow; and, given the symmetry of the valve body structure, symmetrical boundary conditions will be used in subsequent simulations to replace full model calculations.

[0034] S3: Configure CFD transient simulation settings All geometric models of the valve openings were imported into ANSYS Fluent software in STEP format. The total model size was approximately 222mm × 128mm × 119mm, with a valve core diameter of 40mm and a flow channel diameter of DN40. A fluid domain was established and a computational mesh was generated. ANSYS Meshing was used for mesh generation, resulting in an overall mesh element count of approximately 1.8 × 10^6. Local refinement was applied to key areas such as the valve core and flow channel inlet / outlet. The physical model used was the VOF multiphase flow model to simulate the interaction between steam and air, and the turbulence model employed the Realizable k-ε model. Boundary conditions were set: the inlet was a velocity inlet boundary condition, with an inlet velocity of... Exports are pressure exports, export pressure. The wall surface is subject to no-slip boundary conditions. The total transient calculation time is set to 50 seconds, and the time step is [not specified]. The iteration steps are 2000, and the residual convergence criterion is 1×10^-4.

[0035] S4: Collect flow field data During transient simulation, a monitoring surface is established at the valve outlet section to record the outlet pressure at each time step. Data from 5 opening degrees, each with 100 time steps, were aggregated to construct a multidimensional time series dataset containing 500 samples.

[0036] Let the sample number be Each sample can be represented as: ,

[0037] This constitutes the input-output sample pair required for supervised learning:

[0038] Finally, a multidimensional time-series dataset for data-driven modeling was constructed.

[0039] S5: Data Preprocessing The collected raw simulation data is preprocessed. First, outlier identification and removal are performed, followed by processing of the input variable (opening degree). Instantaneous time Entrance speed and output variables (export pressure) Steam volume fraction Export speed The following steps are performed using linear normalization: To eliminate the influence of units of measurement, a standardized training dataset is formed. The preprocessed dataset exhibits good numerical stability and feature consistency, eliminating anomalous perturbations and variable scaling bias, thus providing a reliable data foundation for the subsequent training of neural network surrogate models.

[0040] S6: Define input / output mapping Define the input and output variables of the proxy model and construct a complete mapping relationship. Define the input vector as:

[0041] The output vector is:

[0042] This mapping aims to capture the complex relationship between valve flow response and time and nonlinear opening.

[0043] S7: Constructing a neural network proxy model: A multilayer feedforward neural network (FNN) is used as the surrogate model. The network structure is designed as follows: one input layer (3 neurons), two hidden layers (the first hidden layer has 20 neurons with the activation function tanh; the second hidden layer has 15 neurons with the activation function tanh), and one output layer (3 neurons with a linear activation function), to achieve high-precision nonlinear fitting from input to output. The network calculation formula is as follows:

[0044]

[0045]

[0046] S8: Training the neural network model: The Levenberg-Marquardt (LM) algorithm was used to train the FNN model described above. The mean squared error (MSE) was used as the loss function. After initializing the network weights, biases, and damping factors, the Jacobian matrix was iteratively calculated and the parameters were updated. The damping factor was dynamically adjusted to ensure rapid and stable convergence during training.

[0047] Loss function: Mean Squared Error (MSE)

[0048] Levenberg–Marquardt (LM) algorithm:

[0049] S9: Evaluate model performance: To evaluate the performance of the trained surrogate model, the 20% of data not used in training was used as the test set for validation. The coefficient of determination (R²) was used. 2 The model performance is comprehensively evaluated using metrics such as root mean square error (RMSE) and mean absolute percentage error (MAPE).

[0050] Calculate model performance metrics using independent test sets: Coefficient of determination :

[0051] Root Mean Square Error (RMSE):

[0052] Mean Absolute Percentage Error (MAPE):

[0053] in, These are the model's predicted values. This is for outputting the mean of the variable.

[0054] The coefficient of determination can be obtained through calculation. ), root mean square error ( ) and mean absolute percentage error ( The results show that the model has high accuracy and good generalization ability in predicting pressure, velocity, and steam volume fraction. Test results demonstrate that this surrogate model can accurately capture the transient multi-opening response of valves, achieving efficient prediction of complex nonlinear flow characteristics.

[0055] S10: Enables rapid prediction: The trained proxy model was deployed on a computing platform. Verification showed that the model can predict flow field parameters at any given opening and instantaneous time point within milliseconds, achieving a computational efficiency more than 99% higher than traditional CFD simulation, fully meeting the needs of industrial real-time control systems.

[0056] The above description is merely a preferred embodiment of the present invention. It should be noted that the present invention is not limited to the specific embodiments described above. Any simple modifications, equivalent changes, and alterations made by those skilled in the art to the above embodiments based on the technical essence of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a predictive model for a control valve based on instantaneous multi-opening response, characterized in that, Includes the following steps: S1: Establish a three-dimensional model of the control valve and parameterize the opening degree; S2: Perform geometric simplification on the 3D model; S3: Configure CFD transient simulation; S4: Collect flow field data; S5: Data preprocessing; S6: Define input / output mapping; S7: Construct a neural network proxy model; S8: Train the neural network model; S9: Evaluate model performance; S10: Enables rapid prediction.

2. The method according to claim 1, characterized in that, In step S1, a three-dimensional solid model of the regulating valve, including the valve body, valve core, and flow channel, is established using three-dimensional modeling software. The position of the valve core is associated with the valve opening degree through parametric driving, and geometric models under different opening degrees are generated in batches.

3. The method according to claim 1, characterized in that, In step S2, the model is geometrically simplified based on the principle of physical equivalence, including deleting non-critical features, using symmetric boundary conditions, or simplifying profile replacements to reduce the number of grids and computational resource consumption.

4. The method according to claim 1, characterized in that, In step S3, a multiphase flow model and a turbulent flow model are set up in the CFD software, and boundary conditions, time steps, and total computation time are configured to capture unsteady flow characteristics.

5. The method according to claim 1, characterized in that, In step S4, monitoring surfaces are arranged at the outlet section and key areas to record the outlet pressure, outlet velocity and steam volume fraction at each time step, thus constructing a multidimensional time series dataset.

6. The method according to claim 1, characterized in that, In step S5, the simulation data is cleaned, denoised, outlier removed, and normalized to form a standardized training dataset.

7. The method according to claim 1, characterized in that, In step S6, the input-output mapping relationship is constructed using the opening degree, time, and inlet velocity as input variables, and the outlet pressure, steam volume fraction, and outlet velocity as output variables.

8. The method according to claim 1, characterized in that, In step S7, a multi-layer feedforward neural network is used as a surrogate model, with the hidden layer using the tanh activation function and the output layer using the linear activation function.

9. The method according to claim 1, characterized in that, In step S8, the Levenberg-Marquardt algorithm is used to train the neural network, and the damping factor is dynamically adjusted to optimize the convergence process.

10. The method according to claim 1, characterized in that, In step S9, the coefficient of determination R is used. 2 The root mean square error (RMSE) and mean absolute percentage error (MAPE) are used to evaluate model performance.

Images