A data-driven superheated steam temperature model predictive control method based on knowledge distillation

Abstract

This invention discloses a data-driven model predictive control method for superheated steam temperature based on knowledge distillation. The method first constructs a superheated steam temperature prediction model as the teacher model using a Physical Information Guided Neural Network (PINN). The model input includes control and disturbance variables, and it integrates the system's physical mechanisms and data characteristics. Subsequently, a multilayer perceptron (MLP) network is trained using knowledge distillation to construct a student model, which serves as the final predictive model for control. This student model takes the main control variables and historical outputs as inputs, learns the physical laws and disturbance response characteristics inherent in the teacher model, and achieves efficient fitting of the system behavior. Based on this, a model predictive controller (MPC) for superheated steam temperature regulation is designed using the student model, and the effectiveness and engineering feasibility of the proposed method are verified using actual operating data. Compared with traditional methods, this invention has advantages such as lightweight structure, high prediction accuracy, and strong physical consistency, and is suitable for modeling and intelligent control tasks of superheated steam temperature systems in thermal power units and other complex equipment in process industries.

Description

Technical Field

[0001] This invention belongs to the field of intelligent modeling and control of superheated steam temperature system in coal-fired power unit boilers. Specifically, it relates to a data-driven predictive control method for superheated steam temperature model based on knowledge distillation. Background Technology

[0002] During the operation of thermal power units, superheated steam temperature is a crucial parameter reflecting the safety and economy of the boiler thermal system. Its control quality directly affects the stability of main steam parameters and is key to ensuring the long-term safe and stable operation of the unit. However, the superheated steam temperature control process is affected by various factors, such as load changes, fuel quality fluctuations, and feedwater temperature disturbances. The system exhibits significant strong nonlinearity, large hysteresis, and multiple disturbance coupling characteristics, posing a significant challenge to modeling and control.

[0003] Traditional control strategies primarily rely on PID, cascade control, or fuzzy control. While these methods offer advantages in terms of simple design and ease of engineering implementation, they suffer from poor adaptability to changing operating conditions, often exhibiting response lag or severe overshoot under conditions of deep peak regulation or intense disturbances. Furthermore, these methods heavily depend on operational experience and on-site parameter tuning, hindering intelligent regulation. To overcome the shortcomings of traditional control, researchers have recently introduced data-driven methods for superheated steam temperature modeling and control. Machine learning models such as neural networks and support vector machines have been used to simulate the dynamic response of steam temperature to changes in control variables, improving model accuracy and adaptability to some extent. However, existing data-driven methods also suffer from issues such as physical interpretability, structural complexity, and low efficiency. Therefore, there is an urgent need for more efficient, robust, and deployable modeling and control technologies to meet the high-quality control requirements of modern thermal power units.

[0004] In view of this, the present invention is proposed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a neural network superheated steam temperature model prediction and control method based on knowledge distillation, so as to solve the problems of physical interpretability, complex structure and low efficiency of existing data-driven methods.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: In a first aspect, the present invention provides a data-driven predictive control method for superheated steam temperature based on knowledge distillation, comprising the following steps: S1. Collect relevant parameters of the superheated steam temperature system, including control variables, disturbance variables, and system state variables; S2. A superheated steam temperature prediction model is constructed based on physical information-guided neural network as a teacher model. The model input includes control and disturbance quantities. The teacher model integrates the physical mechanism and data characteristics of the system, and physical constraints are added during the training process. S3. A student model is constructed by training a multilayer perceptron network using knowledge distillation technology. This student model is used as the final predictive model for control. It takes the main control variables and historical outputs as inputs, learns the physical laws and disturbance response characteristics contained in the teacher model, and fits the system behavior. S4. Design a model predictive controller based on the student model, solve for the optimal control quantity through nonlinear optimization, and perform online distillation error correction based on the teacher model during operation; S5. Based on field data, the model's prediction and control effects are verified using preset evaluation indicators.

[0007] Specifically, in step S1: The relevant parameters are selected from desuperheating water flow rate, desuperheating water temperature, desuperheating water pressure, desuperheater inlet temperature, superheater inlet temperature, superheater outlet temperature, load, main steam pressure, and total coal quantity.

[0008] Specifically, step S2 includes: S201. Design a recurrent neural network structure to satisfy the state equations of a superheated steam temperature system. The recurrent neural network structure contains three physical driving gates, and their calculations are as follows. , , , The physical embedded network computation is as follows. , , in, Here is the weight matrix for each part. This represents the bias of each part. and These are physical drive gates based on input and state, respectively. For output-based physical drive gates, The dot product between matrices; S202. Train the model network from S201 to establish a superheated steam temperature prediction model, which will serve as the teacher model. Physical constraints will be incorporated during model training. The loss function is designed as follows. , The first term is the model prediction error loss, and the second term is the physical information loss.

[0009] Specifically, step S3 includes: S301. Design a student model, learn the teacher model based on knowledge distillation, and train the student model. The student model uses a multilayer perceptron. For a multilayer perceptron with N hidden layers, its calculation is as follows: , , in, The output of the nth layer of the MLP. ; , Here are the weight matrix and biases for the nth layer; For activation functions; S302. Select the feature information from the teacher model to be learned, and select the output hidden layer. and model output As a learning objective, the loss function for training the student model is designed accordingly, and its calculation is as follows. , in, For the output of the student model, The first term represents the prediction error loss between the student model and the teacher model, and the second term represents the intermediate layer feature loss between the student model and the teacher model.

[0010] Specifically, in step 4: The student prediction model performs multi-step predictions, calculated as follows: , in, To predict the time domain, For the student MLP prediction model being invoked; The input of the student prediction model for, , in, Let N be the system output at time t, and N be the time step. For time t, this is the control quantity; for a superheated steam temperature system, it is the water spray desuperheating quantity. A nonlinear optimizer is designed, with the objective function and constraint adjustments as follows. , , in, , These are the weighting coefficients. To control the change in quantity, To control the time domain; The objective function is optimized using the gradient descent method to obtain the optimal control quantity at the next time step, as calculated below. , During operation, the prediction model is distilled online based on the teacher model to achieve error correction. The objective function is as follows. , in For input The model predicts the output. This is the output of the teacher model, i.e., the output of the actual system.

[0011] Specifically, step S5 includes: Using root mean square error (RMSE) and coefficient of determination (R²) 2 The evaluation index is calculated using the following formula to assess the model's predictive performance: , , in, This is the actual value; This is a predicted value; The average of the actual values; The number of samples; The mean absolute error (MAE) is used to evaluate the control effect of the model, and its calculation is as follows. .

[0012] In a second aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the control method described above.

[0013] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, wherein the computer program / instructions, when executed by a processor, implement the control method described above.

[0014] Fourthly, the present invention provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the control method described above.

[0015] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: This invention provides a neural network-based predictive control method for superheated steam temperature based on knowledge distillation. A superheated steam temperature prediction model is constructed using PINN as the teacher model. The model input includes control and disturbance variables, and it integrates the system's physical mechanisms and data characteristics. Subsequently, a student model is constructed by training an MLP network using knowledge distillation technology, serving as the final predictive model for control. This student model takes the main control variables and historical outputs as input, learning the physical laws and disturbance response characteristics inherent in the teacher model, achieving efficient fitting of system behavior. Based on this, an MPC for superheated steam temperature regulation is designed using the student model, and the effectiveness and engineering feasibility of the proposed method are verified using actual operating data. This method significantly improves computational efficiency, control accuracy, and generalization ability, providing reliable technical support for real-time control optimization of superheated steam temperature systems. Attached Figure Description

[0016] The accompanying drawings, as part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation of the invention. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings: Figure 1 A control strategy framework diagram for a neural network-based superheated steam temperature model predictive control method based on knowledge distillation provided by this invention; Figure 2 The diagram illustrates the knowledge distillation principle of a neural network-based superheated steam temperature model predictive control method provided by this invention. Figure 3 This is a comparison diagram of the teacher model and the student model of the neural network superheated steam temperature model prediction and control method based on knowledge distillation of the present invention. Figure 4 The control effect diagram of the neural network superheated steam temperature model predictive control method based on knowledge distillation provided by the present invention is shown.

[0017] It should be noted that these accompanying drawings and textual descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art by referring to specific embodiments. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0019] like Figure 1As shown, an embodiment of the present invention provides a neural network-based superheated steam temperature model predictive control method based on knowledge distillation, comprising the following steps: S1: Collect relevant parameters of the superheated steam temperature system of the coal-fired unit. According to one embodiment of the present invention, the relevant parameters of the superheated steam temperature system include desuperheating water flow rate, desuperheating water temperature, desuperheating water pressure, desuperheater inlet temperature, superheater inlet temperature, superheater outlet temperature, load, main steam pressure, and total coal quantity.

[0020] S2: Design a recurrent neural network structure to meet the physical characteristics of the superheated steam temperature system.

[0021] According to one embodiment of the present invention, the recurrent neural network structure includes state variables and output variables. Based on the RNN network structure design, the classic expression of RNN is as follows:

[0022]

[0023] in, , This is the coefficient weight matrix; This is the output of the system at time t; This represents the state of the system at time t. The input of the system at time t, For bias, This is the activation function.

[0024] It should be noted that when establishing the model of the superheated steam temperature system, its system of differential-algebraic equations is as follows:

[0025] in, The working fluid storage capacity of the desuperheater header. , For the inlet and outlet flow rates of the working fluid; , The enthalpy values ​​of the working fluid at the inlet and outlet are... , To reduce the inlet flow rate and enthalpy of the heated water. , , The mass of the working fluid inside the heat exchanger, the mass of the metal tube wall, and the mass of the flue gas; , , , where is the inlet and outlet flow rate of the working fluid and the flue gas flow rate; , The enthalpy values ​​of the working fluid at the inlet and outlet are... , , The specific heat of the metal and the isobaric specific heat of the flue gas inlet and outlet, the flow rate and enthalpy of the desuperheating water inlet; , , These are the heat released by the flue gas, the heat absorbed by the working fluid, and the heat dissipation, respectively. , For the working fluid outlet temperature and pressure, , The temperature at the flue gas inlet and outlet is denoted as .

[0026] Furthermore, the loop network structure is determined based on the mechanistic equation. Since the superheater model differential equation includes output variables, the second network structure is selected, with a state variable of 3. The calculation of the second network structure is as follows: The three physical drive gates are calculated as follows:

[0027]

[0028]

[0029] The physical driving gate, as explained, serves to introduce higher-order nonlinear coupling terms for simulation. Physically strongly coupled terms can capture complex couplings between variables and enhance the nonlinear expressive power of the model.

[0030] Further physical embedded network computation is as follows:

[0031]

[0032] in, Here is the weight matrix for each part. This represents the bias of each part. and These are physical drive gates based on input and state, respectively. For output-based physical drive gates, This is the dot product between matrices.

[0033] It should be noted that introducing saturation constraints at the state variables serves two purposes: firstly, to simulate the saturation phenomenon in actual physical conditions, and secondly, to mitigate the gradient explosion phenomenon during subsequent training.

[0034] The model network was further trained to establish a superheated steam temperature prediction model, which served as the teacher model. Physical constraints were incorporated into the model training process, and the loss function was designed as follows:

[0035] It should be noted that the physical loss constraints added to the superheater steam temperature model are calculated as follows:

[0036]

[0037] The first term is the enthalpy constraint, and the second term is the working fluid energy conservation constraint.

[0038] The total loss is further calculated using the following formula:

[0039] in , Weights for physical loss The first term is the model prediction error loss, and the second term is the physical information loss, which depends on the specific application.

[0040] S3: Design a student model, learn the teacher model based on knowledge distillation, and train the student model.

[0041] According to one embodiment of the present invention, the student model employs a relatively simple multilayer perceptron (MLP).

[0042] It should be noted that for an MLP with N hidden layers, the calculation is as follows:

[0043]

[0044] in The output of the nth layer of the MLP. ; , Here are the weight matrix and biases for the nth layer; This is the activation function.

[0045] Further select the feature information from the teacher model that needs to be learned, and select the output hidden layer. and model output As a learning objective, the loss function for training the student model is designed accordingly, and its calculation is as follows:

[0046] in, For the output of the student model, These are the weighting coefficients. The first term represents the prediction error loss between the student model and the teacher model, and the second term represents the intermediate layer feature loss between the student model and the teacher model.

[0047] It should be noted that the inputs to the teacher model include historical outputs, control variables, and perturbation variables, while the inputs to the student model include historical outputs and control variables. Therefore, when the student model learns the hidden layers of the teacher model, it is actually learning the physical constraints and perturbation information in the teacher model.

[0048] S4: When designing a model prediction controller based on the distilled student prediction model, it should be noted that: Input to the student prediction model for:

[0049] in Let N be the system output at time t, and N be the time step. This is the control quantity at time t, and for the superheated steam temperature system, it is the water spray desuperheating quantity.

[0050] Further prediction models perform multi-step predictions as follows:

[0051] in To predict the time domain, The student MLP prediction model is invoked.

[0052] The nonlinear optimizer is further designed, with the objective function and constraint adjustment as shown below:

[0053]

[0054] in, , These are the weighting coefficients. To control the change in quantity, To control the time domain.

[0055] Furthermore, the objective function is optimized using gradient descent to obtain the optimal control quantity at the next time step, which is calculated as follows:

[0056] Furthermore, to ensure the accuracy of the prediction model, online distillation is performed on the prediction model based on the teacher model during operation to achieve error correction. The objective function is as follows:

[0057] in For input The model predicts the output. This is the output of the teacher model, i.e., the output of the actual system.

[0058] S5: Model and controller verification based on field data. It should be noted that: Using root mean square error (RMSE) and coefficient of determination (R²) 2 The evaluation index is calculated using the following formula to assess the model's predictive performance:

[0059]

[0060] in, This is the actual value; This is a predicted value; The average of the actual values; This represents the number of samples.

[0061] Furthermore, the mean absolute error is used to evaluate the model control effect, and its calculation is as follows:

[0062] The following specific embodiment of the present invention will be used to verify the beneficial effects of the knowledge distillation-based neural network superheated steam temperature model prediction and control method of the present invention.

[0063] The example test was conducted using relevant collected data on the wall temperature of the superheater of the rear screen of an actual coal-fired unit. In this embodiment, relevant parameters of the superheated steam temperature of the coal-fired unit were collected.

[0064] A controlled object network model of the superheated steam temperature system is established, with desuperheating water flow rate, desuperheating water temperature, desuperheating water pressure, desuperheater inlet temperature, load, main steam pressure, and total coal quantity as model inputs, and superheater outlet temperature as output. This model serves as the teacher model for subsequent knowledge distillation.

[0065] Based on the theory of knowledge distillation, student models are trained using teacher models, and the underlying principle is as follows: Figure 2 As shown, by learning the hidden features and output distribution of the teacher model, the prediction model can achieve the corresponding accuracy and dynamic characteristics.

[0066] A model predictive controller is designed based on the trained predictive model, and the gradient descent method is used for optimization to establish a complete closed-loop control model. Its overall process framework is as follows: Figure 1 As shown.

[0067] Finally, based on the trained network model, the performance of the model and controller was validated using field-collected data. Figure 3 The predictive performance of the teacher and student models is demonstrated. Figure 4The control effect after applying actual field data is shown. It can be seen that the student model's prediction performance has learned well from the teacher model, achieving a high level of prediction accuracy. The RMSE of the student model is 0.23℃, while the RMSE of the teacher model is 0.2℃. Furthermore, compared with field data, the average absolute error of the steam temperature control proposed in this invention is 1.04℃, while the average error in the actual field is 5.82℃, representing an 82% improvement in relative control accuracy. The knowledge distillation-based neural network superheated steam temperature model predictive control method proposed in this paper can achieve high-precision control while maintaining stable superheated steam temperature, ensuring the safe and stable operation of the unit.

[0068] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0069] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0070] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0071] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0072] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0073] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0074] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A data-driven predictive control method for superheated steam temperature based on knowledge distillation, characterized in that, Includes the following steps: S1. Collect relevant parameters of the superheated steam temperature system, including control variables, disturbance variables, and system state variables; S2. A superheated steam temperature prediction model is constructed based on physical information-guided neural network as a teacher model. The model input includes control and disturbance quantities. The teacher model integrates the physical mechanism and data characteristics of the system, and physical constraints are added during the training process. S3. A student model is constructed by training a multilayer perceptron network using knowledge distillation technology. This student model is used as the final predictive model for control. It takes the main control variables and historical outputs as inputs, learns the physical laws and disturbance response characteristics contained in the teacher model, and fits the system behavior. S4. Design a model predictive controller based on the student model, solve for the optimal control quantity through nonlinear optimization, and perform online distillation error correction based on the teacher model during operation; S5. Based on field data, the model's prediction and control effects are verified using preset evaluation indicators.

2. The control method as described in claim 1, characterized in that, In step S1: The relevant parameters are selected from desuperheating water flow rate, desuperheating water temperature, desuperheating water pressure, desuperheater inlet temperature, superheater inlet temperature, superheater outlet temperature, load, main steam pressure, and total coal quantity.

3. The control method as described in claim 1, characterized in that, Step S2 includes: S201. Design a recurrent neural network structure to satisfy the state equations of a superheated steam temperature system. The recurrent neural network structure contains three physical driving gates, and their calculations are as follows. ， ， ， The physical embedded network computation is as follows. ， ， in, Here is the weight matrix for each part. This represents the bias of each part. and These are physical drive gates based on input and state, respectively. For output-based physical drive gates, The dot product between matrices; S202. Train the model network from S201 to establish a superheated steam temperature prediction model, which will serve as the teacher model. Physical constraints will be incorporated during model training. The loss function is designed as follows. ， The first term is the model prediction error loss, and the second term is the physical information loss.

4. The control method as described in claim 1, characterized in that, Step S3 includes: S301. Design a student model, learn the teacher model based on knowledge distillation, and train the student model. The student model uses a multilayer perceptron. For a multilayer perceptron with N hidden layers, its calculation is as follows: ， ， in, The output of the nth layer of the MLP. ; , Here are the weight matrix and biases for the nth layer; For activation functions; S302. Select the feature information from the teacher model to be learned, and select the output hidden layer. and model output As a learning objective, the loss function for training the student model is designed accordingly, and its calculation is as follows. ， in, For the output of the student model, The first term represents the prediction error loss between the student model and the teacher model, and the second term represents the intermediate layer feature loss between the student model and the teacher model.

5. The control method as described in claim 1, characterized in that, In step 4: The student prediction model performs multi-step predictions, calculated as follows: ， in, To predict the time domain, For the student MLP prediction model being invoked; The input of the student prediction model for, ， in, Let N be the system output at time t, and N be the time step. For time t, this is the control quantity; for a superheated steam temperature system, it is the water spray desuperheating quantity. A nonlinear optimizer is designed, with the objective function and constraint adjustments as follows. ， ， in, , These are the weighting coefficients. To control the change in quantity, To control the time domain; The objective function is optimized using the gradient descent method to obtain the optimal control quantity at the next time step, as calculated below. ， During operation, the prediction model is distilled online based on the teacher model to achieve error correction. The objective function is as follows. ， in For input The model predicts the output. This is the output of the teacher model, i.e., the output of the actual system.

6. The control method as described in claim 1, characterized in that, Step S5 includes: Using root mean square error (RMSE) and coefficient of determination (R²) 2 The evaluation index is calculated using the following formula to assess the model's predictive performance: ， ， in, This is the actual value; This is a predicted value; The average of the actual values; The number of samples; The mean absolute error (MAE) is used to evaluate the control effect of the model, and its calculation is as follows. 。 7. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory, wherein the processor, when executing the computer program, implements the control method according to any one of claims 1-6.

8. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the control method according to any one of claims 1-6.

9. A computer program product, comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by the processor, they implement the control method according to any one of claims 1-6.

Images