Modeling method and device for turbine-boiler coordination system of coal-fired power generating unit

Abstract

The invention provides a modeling method and device for a machine-boiler coordination system of a coal-fired power generation unit. The method comprises the following steps: constructing a physical information neural network; taking the current moment control variable and the previous moment state variable as the input of the physical information neural network, and taking the current moment state variable and the current system output variable as the output; constructing a first loss function based on a physical law according to the current moment state variable and the current system output variable; constructing a second loss function according to the predicted output and the actual output, and obtaining a total loss function based on the first loss function and the second loss function; inputting a training set into the physical information neural network for training, updating parameters of the physical information neural network by adopting back propagation and gradient based on a total loss function, and introducing a neural tangent accounting method to dynamically update a physical information loss weight in a first loss function in each round of training; the method can improve the accuracy of the coordination control system model of the coal-fired power generating unit.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to a modeling method and apparatus for a coal-fired power generation unit boiler-generator coordination system. Background Technology

[0002] With the continuous increase in installed capacity of fluctuating renewable energy sources such as wind and solar power, the requirements for coal-fired power plants to ensure the consumption of new energy and grid stability are becoming increasingly stringent. Coal-fired power generating units (CFPPs) are complex systems involving multiple types of equipment, including fluid compression and transport, heat exchange, heat-work conversion, and mechanical transmission, and involving the coupling of various energy flow and conversion processes. Establishing a dynamic model of coal-fired power generating units and understanding the interrelationships between key variables are crucial for optimizing control strategies, tapping into flexible potential, and improving unit operating performance. However, the significant changes in characteristics brought about by deep-condition migration and the lack of clear understanding of the mechanisms during rapid dynamic adjustments increase the modeling difficulty of coal-fired power generating units.

[0003] Traditional coal-fired power plant modeling often employs mechanistic modeling methods. While this method offers advantages such as clear physical meaning and accurate depiction of dynamic laws, its modeling process heavily relies on a deep understanding of the unit's operating mechanisms and design parameters, and requires the introduction of numerous reasonable assumptions, resulting in significant challenges in model construction and parameter calibration. Furthermore, mechanistic models consist of a massive coupling of nonlinear algebraic and differential equations; the complex model structure and enormous computational load make them difficult to directly apply to engineering practices for unit operation optimization. The vast amounts of data generated during the full-condition operation of coal-fired power plants contain rich information reflecting the unit's dynamic characteristics and operating laws. Data-driven model identification methods can extract mapping relationships between key variables by mining data features, without relying on the complex physical mechanisms of the unit, providing a convenient and efficient technical path for coal-fired power plant model construction. For example, patent text CN121028599A discloses a coordinated modeling method for a gated cyclic unit coal-fired power plant that integrates a self-attention mechanism, including: 1. collecting operating data of the coal-fired power plant and preprocessing the data; 2. using the preprocessed data as input and the predicted values ​​of the power and main steam pressure of the coal-fired power plant as output to construct a GRU network model that integrates a self-attention mechanism; 3. training and optimizing the GRU network model that integrates a self-attention mechanism.

[0004] However, while these machine learning-based modeling methods demonstrate strong capabilities in fitting complex nonlinear models, they also suffer from inherent drawbacks such as weak model interpretability and insufficient generalization reliability. Furthermore, model performance is highly dependent on data quality and sample size, and model mismatch is prone to occur in scenarios where data is scarce or unevenly distributed. Summary of the Invention

[0005] This invention provides a modeling method and apparatus for the coordinated system of a coal-fired power generation unit, so as to improve the accuracy of the coordinated control system model of a coal-fired power generation unit.

[0006] A modeling method for a coal-fired power generating unit boiler-generator coordination system includes: Construct a physical information neural network for coordinated control of coal-fired power generating units and boilers; The current control variable and the previous state variable are used as inputs to the physical information neural network, and the current state variable and the current system output variable are used as outputs to the physical information neural network. A first loss function based on physical laws is constructed based on the current state variables and the current system output variables; a second data-driven loss function is constructed based on the predicted output and the actual output; and the total loss function is obtained based on the first and second loss functions. Acquire historical status data of coordinated operation of coal-fired power generating units and boilers, and construct a training set; The training set is input into the physical information neural network for training. Based on the total loss function, the parameters of the physical information neural network are updated using backpropagation and gradient. A neural tangent kernel algorithm is introduced to dynamically update the physical information loss weights in the first loss function in each round of training.

[0007] Furthermore, the physical information neural network includes an input module, a fully connected layer, and an output module; the input module is used to input the current control variable sequence and the previous state variable and perform feature extraction, input the extracted features to the fully connected layer, the fully connected layer performs calculations based on the input features, and outputs the current state variable and the system output variable through the output module.

[0008] Furthermore, the physical information neural network also includes an attention mechanism module, which is used to enhance the relevance of the features extracted by the input module to obtain a deep feature vector, and the fully connected layer performs calculations based on the deep feature vector.

[0009] Furthermore, the first loss function includes a differential equation loss function and an algebraic equation loss function; Construct a first loss function based on physical laws using the current state variables and the current system output variables, including: The physical mechanism equations for the coordinated operation of coal-fired power generating units and boilers are constructed. The physical mechanism equations include differential equations relating control variables, state variables and the rate of change of state variables, as well as algebraic equations relating control variables, state variables and system output variables. A differential equation loss function is constructed based on differential equations and current state variables to characterize the consistency between the dynamic performance and process mechanism of physical information neural networks. An algebraic equation loss function is constructed based on algebraic equations and current system output variables to characterize the consistency between the static performance and process mechanism of a physical information neural network.

[0010] Furthermore, the second loss function is the mean squared error loss function; The total loss function is the product of the differential equation loss function and the first physical information loss weight, the product of the algebraic equation loss function and the second physical information loss weight, and the sum of the second loss function.

[0011] Furthermore, a neural tangent kernel algorithm is introduced to dynamically update the physical information loss weights in the first loss function in each training round, including: Based on the differential equation loss function, the algebraic equation loss function, the second loss function, and the gradient vector of the current parameters, construct the neural tangential kernel matrix; Solve for the trace of the neural tangential kernel matrix to obtain the first intensity value of the contribution of the differential equation loss to the gradient of the current parameter, the second intensity value of the contribution of the algebraic equation loss to the gradient of the current parameter, and the third intensity value of the contribution of the second loss to the gradient of the current parameter. The first physical information loss weight and the second physical information loss weight are calculated based on the first intensity value, the second intensity value, and the third intensity value.

[0012] Further, based on the differential equation loss function, the algebraic equation loss function, the second loss function, and the gradient vector of the current parameters, a neural tangential kernel matrix is ​​constructed, including: Select the first and second samples from the current training data; Based on the gradient vectors of the first sample, the second sample, and the current parameters, as well as the differential equation loss function, calculate the first neural tangential kernel element vector corresponding to the differential equation loss; Based on the gradient vectors of the first sample, the second sample, and the current parameters, as well as the algebraic equation loss function, calculate the second neural tangential kernel element vector corresponding to the algebraic equation loss. Based on the gradient vectors of the first sample, the second sample, the current parameters, and the second loss function, calculate the third neural tangential kernel element corresponding to the second loss; The first neural tangential kernel element vector, the second neural tangential kernel element vector, and the third neural tangential kernel element constitute the neural tangential kernel matrix.

[0013] Furthermore, the first neural tangential kernel element vector is the inner product of the gradient vector of the differential equation loss with respect to the current parameters for the first sample and the transpose of the gradient vector of the differential equation loss with respect to the current parameters for the second sample. The second neural tangential kernel element vector is the inner product of the gradient vector of the algebraic equation loss for the first sample with respect to the current parameters and the transpose of the gradient vector of the algebraic equation loss for the second sample with respect to the current parameters. The third neural tangential kernel element vector is the inner product of the gradient vector of the second loss with respect to the current parameters for the first sample and the transpose of the gradient vector of the second loss with respect to the current parameters for the second sample.

[0014] Furthermore, the first physical information loss weight is the ratio of the third intensity value to the regularized first intensity value; The second physical information loss weight is the ratio of the third intensity value to the regularized second intensity value.

[0015] A modeling device for a coal-fired power generating unit boiler-generator coordination system includes: The network construction module is used to build a physical information neural network for the coordinated control of the boiler and generator units in coal-fired power generation units; The input / output definition module is used to take the current control variable and the previous state variable as the input of the physical information neural network, and take the current state variable and the current system output variable as the output of the physical information neural network. The loss function construction module is used to construct a first loss function based on physical laws based on the predicted value of the current state variable and the current system output variable; construct a data-driven second loss function based on the predicted output and the actual output; and obtain the total loss function based on the first loss function and the second loss function. The data processing module is used to acquire historical status data of the coordinated operation of coal-fired power generating units and boilers and to build a training set; The training module is used to input the training set into the physical information neural network for training. Based on the total loss function, it uses backpropagation and gradient to update the parameters of the physical information neural network, and introduces a neural tangent kernel algorithm to dynamically update the physical information loss weights in the first loss function in each round of training.

[0016] Furthermore, the physical information neural network includes an input module, a fully connected layer, and an output module; the input module is used to input the current control variable sequence and the previous state variable and perform feature extraction, input the extracted features to the fully connected layer, the fully connected layer performs calculations based on the input features, and outputs the predicted value of the current state variable and the system output variable through the output module.

[0017] Furthermore, the physical information neural network also includes an attention mechanism module, which is used to enhance the relevance of the features extracted by the input module to obtain a deep feature vector, and the fully connected layer performs calculations based on the deep feature vector.

[0018] Furthermore, the first loss function includes a differential equation loss function and an algebraic equation loss function; The loss function construction module constructs a first loss function based on physical laws according to the current state variables and the current system output variables, including: The physical mechanism equations for the coordinated operation of coal-fired power generating units are constructed. The physical mechanism equations include differential equations relating control variables, state variables and the rate of change of state variables, as well as algebraic equations relating control variables, state variables and system output variables. A differential equation loss function is constructed based on differential equations and current state variables to characterize the consistency between the dynamic performance and process mechanism of physical information neural networks. An algebraic equation loss function is constructed based on algebraic equations and current system output variables to characterize the consistency between the static performance and process mechanism of a physical information neural network.

[0019] Furthermore, the second loss function is the mean squared error loss function; The total loss function is the product of the differential equation loss function and the first physical information loss weight, the product of the algebraic equation loss function and the second physical information loss weight, and the sum of the second loss function.

[0020] Furthermore, the training module introduces a neural tangent kernel algorithm to dynamically update the physical information loss weights in the first loss function in each training round, including: Based on the differential equation loss function, the algebraic equation loss function, the second loss function, and the gradient vector of the current parameters, construct the neural tangential kernel matrix; Solve for the trace of the neural tangential kernel matrix to obtain the first intensity value of the contribution of the differential equation loss to the gradient of the current parameter, the second intensity value of the contribution of the algebraic equation loss to the gradient of the current parameter, and the third intensity value of the contribution of the second loss to the gradient of the current parameter. The first physical information loss weight and the second physical information loss weight are calculated based on the first intensity value, the second intensity value, and the third intensity value.

[0021] Furthermore, the training module constructs a neural tangential kernel matrix based on the differential equation loss function, the algebraic equation loss function, the second loss function, and the gradient vector of the current parameters, including: Select the first and second samples from the current training data; Based on the gradient vectors of the first sample, the second sample, and the current parameters, as well as the differential equation loss function, calculate the first neural tangential kernel element vector corresponding to the differential equation loss; Based on the gradient vectors of the first sample, the second sample, and the current parameters, as well as the algebraic equation loss function, calculate the second neural tangential kernel element vector corresponding to the algebraic equation loss. Based on the gradient vectors of the first sample, the second sample, the current parameters, and the second loss function, calculate the third neural tangential kernel element corresponding to the second loss; The first neural tangential kernel element vector, the second neural tangential kernel element vector, and the third neural tangential kernel element constitute the neural tangential kernel matrix.

[0022] Furthermore, the first neural tangential kernel element vector is the inner product of the gradient vector of the differential equation loss with respect to the current parameters for the first sample and the transpose of the gradient vector of the differential equation loss with respect to the current parameters for the second sample. The second neural tangential kernel element vector is the inner product of the gradient vector of the algebraic equation loss for the first sample with respect to the current parameters and the transpose of the gradient vector of the algebraic equation loss for the second sample with respect to the current parameters. The third neural tangential kernel element vector is the inner product of the gradient vector of the second loss with respect to the current parameters for the first sample and the transpose of the gradient vector of the second loss with respect to the current parameters for the second sample.

[0023] Furthermore, the first physical information loss weight is the ratio of the third intensity value to the regularized first intensity value; The second physical information loss weight is the ratio of the third intensity value to the regularized second intensity value.

[0024] The modeling method and apparatus for the boiler-generator coordination system of a coal-fired power generation unit provided by this invention have at least the following beneficial effects: The constructed physical information neural network architecture, combining physical mechanisms with deep data fusion modeling, achieves equivalent mapping and derivative alignment of the physical state equations of the boiler-turbine coordinated system, enhancing the physical interpretability of the model and introducing a self-attention mechanism. This enables dynamic decoupling of the multivariable strongly coupled system and improves the adaptive identification capability of the contribution weights of control variables. By aligning the output with the physical mechanism equations in real time, the model strictly adheres to physical conservation laws while fitting the data, ensuring the physical rationality of the predictions. The neural tangent kernel algorithm is used to dynamically balance the competition between data fitting and physical constraints. This provides a reliable and accurate prediction model for high-performance control and optimization of coal-fired power generating unit coordinated systems. Attached Figure Description

[0025] Figure 1This is a flowchart of an embodiment of the modeling method for the boiler-generator coordination system of a coal-fired power generation unit provided by the present invention.

[0026] Figure 2 This is a schematic diagram of the physical information neural network in one embodiment of the modeling method for the boiler-generator coordination system of a coal-fired power generation unit provided by the present invention.

[0027] Figure 3 This is a schematic diagram of an embodiment of model training in the modeling method for the boiler-generator coordination system of a coal-fired power generation unit provided by the present invention.

[0028] Figure 4 This is a schematic diagram of the prediction results in one application scenario of the modeling method for the boiler-generator coordination system of a coal-fired power generation unit provided by the present invention.

[0029] Figure 5 This is a schematic diagram of one embodiment of the modeling device for the boiler-generator coordination system of a coal-fired power generation unit provided by the present invention. Detailed Implementation

[0030] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0031] refer to Figure 1 In some embodiments, a modeling method for a coal-fired power generation unit boiler-generator coordination system is provided, including: S1. Construct a physical information neural network for coordinated control of coal-fired power generation units and boilers; S2. The current control variable and the previous state variable are used as inputs to the physical information neural network, and the current state variable and the current system output variable are used as outputs to the physical information neural network. S3. Construct a first loss function based on physical laws according to the current state variables and the current system output variables; construct a data-driven second loss function according to the predicted output and the actual output; and obtain the total loss function based on the first and second loss functions. S4. Obtain historical status data of coordinated operation of coal-fired power generating units and boilers and construct a training set; S5. Input the training set into the physical information neural network for training. Based on the total loss function, use backpropagation and gradient to update the parameters of the physical information neural network, and introduce a neural tangent kernel algorithm to dynamically update the physical information loss weights in the first loss function in each round of training.

[0032] Further, refer to Figure 2In step S1, the physical information neural network includes an input module 1, a fully connected layer 2, and an output module 3. The input module 1 is used to input the current control variable sequence and the previous state variable and perform feature extraction. The extracted features are input into the fully connected layer 2, which performs calculations based on the input features. The current state variable and the system output variable are output through the output module 3.

[0033] In some embodiments, the physical information neural network further includes an attention mechanism module 4, which is used to enhance the relevance of the features extracted by the input module to obtain a deep feature vector, and the fully connected layer performs calculations based on the deep feature vector.

[0034] Specifically, in step S2, in some embodiments, the control variables include coal feed rate, water flow rate, and turbine regulating valve opening; the state variables include furnace coal quantity, steam pressure and enthalpy at the steam-water separator outlet; and the output variables include power generation, main steam pressure, and steam-water separator outlet enthalpy. The input to the physical information neural network is defined as the control variables at the current moment and the state variables at the previous moment, and the predicted value of the state variables at the current moment and the current system output variables are used as the output of the physical information neural network.

[0035] In step S3, a loss function is constructed, the first loss function including a differential equation loss function and an algebraic equation loss function.

[0036] Construct a first loss function based on physical laws using the predicted values ​​of the current state variables and the current system output variables, including: S31. Construct the physical mechanism equations for the coordinated operation of the boiler and generator of a coal-fired power generation unit. The physical mechanism equations include differential equations relating control variables, state variables and the rate of change of state variables, as well as algebraic equations relating control variables, state variables and system output variables. S32. Construct a differential equation loss function based on differential equations and current state variables to characterize the consistency between the dynamic performance and process mechanism of physical information neural networks; S33. Construct an algebraic equation loss function based on algebraic equations and current system output variables to characterize the consistency between the static performance and process mechanism of a physical information neural network.

[0037] Specifically, in step S31, the current state variables and current system output variables output by output module 3 are used to align in real time with the preset physical mechanism equations of the furnace-machine coordination system. By calculating the deviation between the predicted trend and the physical laws, the neural network is forced to strictly adhere to the laws of energy conservation and mass conservation while fitting historical data.

[0038] The physical mechanism equation for the coordinated operation of coal-fired power generating units is as follows: (1) in, Let x represent the rate of change of the state variable over time, u represent the control variable, and y represent the system output variable. Let y = g(x, u) represent a differential equation, and let y = g(x, u) represent an algebraic equation.

[0039] In step S32, the loss function of the differential equation is as follows: (2) Among them, Loss phy1 Let N represent the loss function of the differential equation, and let N represent the number of samples involved in the loss calculation. This represents the predicted value of the state variable at the current moment. The residuals for the network approximation of differential equations characterize the consistency between the dynamic performance and process mechanism of the physical information neural network.

[0040] In step S33, the loss function of the algebraic equation is as follows: (3) Among them, Loss phy2 Let N represent the loss function of the algebraic equation, and let N represent the number of samples involved in the loss calculation. This represents the current system output variable. The residuals for network approximation of algebraic equations characterize the consistency between the static performance and process mechanism of physical information neural networks.

[0041] Furthermore, the second loss function is the mean squared error loss function, as shown below: (4) Among them, Loss data This represents the second loss function. 'o' represents the predicted output, and 'o' represents the actual output.

[0042] The total loss function is the sum of the product of the differential equation loss function and the weights of the first physical information loss, the product of the algebraic equation loss function and the weights of the second physical information loss, and the second loss function, as shown below: (5) Where Loss represents the total loss function, λ1 represents the weight of the first physical information loss, and λ2 represents the weight of the second physical information loss.

[0043] Further, in step S4, historical state data of the coordinated operation of the coal-fired power generating unit and boiler are obtained and a training set is constructed, including control variables, state variables and system output variables. The control variables include coal feed rate, water flow rate and turbine regulating valve opening. The state variables include furnace coal quantity, steam pressure and enthalpy at the steam-water separator outlet. The system output variables include power generation, main steam pressure and steam-water separator outlet enthalpy.

[0044] In step S5, the physical information neural network is trained, such as... Figure 3 As shown, the control variables at the current time and the state variables at the previous time are used as inputs to the physical information neural network. A self-attention mechanism is introduced, and by calculating the correlation score matrix between the input variables, the contribution weight of different control actions to the system output variables is automatically identified, thereby realizing the dynamic decoupling of heterogeneous input variables from specific output variables.

[0045] Specifically, a linear mapping is performed between the predicted values ​​of the state variables from the previous time step and the control variables from the current time step to construct a query matrix Q, a key matrix K, and a value matrix V: Q=W Q ·[x(t-1); u(t)];(6) K=W K ·[x(t-1); u(t)];(7) V=W V ·[x(t-1); u(t)];(8) In the formula, W Q W K W V Let x(t-1) be a learnable weight matrix, where x(t-1) represents the predicted value of the state variable at the previous time step, and u(t) represents the control variable at the current time step.

[0046] Scaled dot product attention is used to calculate the dynamic correlation score matrix A between input variables, which is used to quantify their contribution to the output. (9) In the formula, d k is the feature dimension scaling factor used to prevent gradient vanishing; each element in matrix A represents the decoupling strength between different input dimensions.

[0047] The dynamic correlation score matrix A and the value matrix V are weighted and aggregated to obtain the deep feature vector after correlation enhancement: Z = A * V; (10) In the formula, Z is the deep feature vector after correlation enhancement. This vector replaces the feature layer of the equal weighted input in the traditional FNN, realizing the automatic suppression of redundant information.

[0048] The deep feature vector Z is fed into the subsequent fully connected layer to compute the predicted values ​​of the state variables and the system output variables in parallel.

[0049] A neural tangent kernel algorithm is introduced to dynamically update the physical information loss weights in the first loss function in each training round, including: S51. Construct the neural tangential kernel matrix based on the differential equation loss function, the algebraic equation loss function, the second loss function, and the gradient vector of the current parameters; S52. Solve for the trace of the neural tangential kernel matrix to obtain the first intensity value of the contribution of the differential equation loss to the gradient of the current parameter, the second intensity value of the contribution of the algebraic equation loss to the gradient of the current parameter, and the third intensity value of the contribution of the second loss to the gradient of the current parameter. S53. Calculate the first physical information loss weight and the second physical information loss weight based on the first intensity value, the second intensity value and the third intensity value.

[0050] Further, in step S51, a neural tangential kernel matrix is ​​constructed based on the differential equation loss function, the algebraic equation loss function, the second loss function, and the gradient vector of the current parameters, including: Select the first and second samples from the current training data; Based on the gradient vectors of the first sample, the second sample, and the current parameters, as well as the differential equation loss function, calculate the first neural tangential kernel element vector corresponding to the differential equation loss; Based on the gradient vectors of the first sample, the second sample, and the current parameters, as well as the algebraic equation loss function, calculate the second neural tangential kernel element vector corresponding to the algebraic equation loss. Based on the gradient vectors of the first sample, the second sample, the current parameters, and the second loss function, calculate the third neural tangential kernel element corresponding to the second loss; The first neural tangential kernel element vector, the second neural tangential kernel element vector, and the third neural tangential kernel element constitute the neural tangential kernel matrix.

[0051] Wherein, the first neural tangential kernel element vector is the inner product of the gradient vector of the differential equation loss with respect to the current parameters for the first sample and the transpose of the gradient vector of the differential equation loss with respect to the current parameters for the second sample. The second neural tangential kernel element vector is the inner product of the gradient vector of the algebraic equation loss for the first sample with respect to the current parameters and the transpose of the gradient vector of the algebraic equation loss for the second sample with respect to the current parameters. The third neural tangential kernel element vector is the inner product of the gradient vector of the second loss with respect to the current parameters for the first sample and the transpose of the gradient vector of the second loss with respect to the current parameters for the second sample.

[0052] Specifically, a neural tangent kernel algorithm is introduced to address the competition between physical constraints and data fitting objectives during the training process. This algorithm can monitor the gradient flow field characteristics during model learning and automatically adjust the weight ratio of physical constraints in the total loss. This ensures that the model can maintain stable prediction accuracy and physical logic rationality even when the unit experiences significant peak shaving or drastic fluctuations in operating conditions.

[0053] The neural tangential nucleus matrix is ​​shown below: (11) (12) (13) Among them, z i Let z represent the first sample. j Indicates the second sample. This represents the gradient vector of the current parameter θ of the physical information neural network. This indicates that for the first sample z i Loss of differential equations phy1 The gradient vector of the current parameters, This indicates that for the second sample z j Loss of differential equations phy1 The transpose of the gradient vector of the current parameters. Indicates the inner product. Represents the vector of elements of the first neural tangential kernel; This indicates that for the first sample z i Algebraic equation loss phy2 The gradient vector of the current parameters, This indicates that for the second sample z j Algebraic equation loss phy2 The transpose of the gradient vector of the current parameters. Represents the element vector of the second neural tangential nucleus; This indicates that for the first sample z i The second loss data The gradient vector of the current parameters, This indicates that for the second sample z j The second loss data The transpose of the gradient vector of the current parameters. This represents the vector of elements in the tangential nucleus of the third neural network.

[0054] In step S52, the trace (i.e., the sum of the diagonal elements) of the neural tangential kernel matrix is ​​solved to obtain the first intensity value of the contribution of the differential equation loss to the gradient of the current parameters, the second intensity value of the contribution of the algebraic equation loss to the gradient of the current parameters, and the third intensity value of the contribution of the second loss to the gradient of the current parameters; the calculation formulas are as follows: (14) (15) (16) Among them, S phy1 S represents the first intensity value of the contribution of the differential equation loss to the gradient of the current parameters. phy2 S represents the second intensity value of the contribution of the algebraic equation loss to the gradient of the current parameters. data This represents the third strength value of the contribution of the second loss to the gradient of the current parameters.

[0055] Further, in step S53, the first physical information loss weight is the ratio of the third intensity value to the regularized first intensity value; The second physical information loss weight is the ratio of the third intensity value to the regularized second intensity value.

[0056] The specific calculation formula is as follows: (17) (18) in, This is a regularization term used to prevent the denominator from approaching 0.

[0057] The method provided in this embodiment will be further verified in the following specific application scenarios.

[0058] Model training was conducted to obtain a dynamic model of the coal-fired power plant-boiler coordination system. For example... Figure 4 As shown in the figure, this embodiment presents a dynamic model of a coal-fired power plant coordination system based on PINN (Physical Information Neural Network). The single-step prediction results for main steam pressure, steam enthalpy at the steam-water separator outlet, and power generation show that, compared with traditional neural network methods (FNN, NARXNN), PINN exhibits better prediction accuracy across all output variables. Especially during dynamic changes in operating conditions, it can accurately follow the inflection points of the actual curves with smaller prediction errors. Table 1 shows that the RMSE and MAE indices of the proposed PINN model are significantly reduced, particularly for power generation and steam enthalpy at the steam-water separator outlet, demonstrating the higher prediction accuracy and reliability of the PINN model.

[0059] Table 1

[0060] refer to Figure 5 In some embodiments, a modeling apparatus for a coal-fired power generator unit boiler-generator coordination system is provided, comprising: Network construction module 201 is used to construct a physical information neural network for coordinated control of coal-fired power generation units and boilers; The input / output definition module 202 is used to take the current control variable and the previous state variable as the input of the physical information neural network, and take the current state variable and the current system output variable as the output of the physical information neural network. The loss function construction module 203 is used to construct a first loss function based on physical laws according to the current state variables and the current system output variables; construct a data-driven second loss function according to the predicted output and the actual output; and obtain the total loss function based on the first loss function and the second loss function. Data processing module 204 is used to acquire historical status data of coordinated operation of coal-fired power generating units and boilers and construct training sets; The training module 205 is used to input the training set into the physical information neural network for training. Based on the total loss function, it uses backpropagation and gradient to update the parameters of the physical information neural network, and introduces a neural tangent kernel algorithm to dynamically update the physical information loss weights in the first loss function in each round of training.

[0061] Furthermore, the physical information neural network includes an input module, a fully connected layer, and an output module; the input module is used to input the current control variable sequence and the previous state variable and perform feature extraction, input the extracted features to the fully connected layer, the fully connected layer performs calculations based on the input features, and outputs the current state variable and the system output variable through the output module.

[0062] Furthermore, the physical information neural network also includes an attention mechanism module, which is used to enhance the relevance of the features extracted by the input module to obtain a deep feature vector, and the fully connected layer performs calculations based on the deep feature vector.

[0063] Furthermore, the first loss function includes a differential equation loss function and an algebraic equation loss function; Loss function construction module 203 constructs a first loss function based on physical laws according to the current state variables and the current system output variables, including: The physical mechanism equations for the coordinated operation of coal-fired power generating units and boilers are constructed. The physical mechanism equations include differential equations relating control variables, state variables and the rate of change of state variables, as well as algebraic equations relating control variables, state variables and system output variables. A differential equation loss function is constructed based on differential equations and current state variables to characterize the consistency between the dynamic performance and process mechanism of physical information neural networks. An algebraic equation loss function is constructed based on algebraic equations and current system output variables to characterize the consistency between the static performance and process mechanism of a physical information neural network.

[0064] Furthermore, the second loss function is the mean squared error loss function; The total loss function is the product of the differential equation loss function and the first physical information loss weight, the product of the algebraic equation loss function and the second physical information loss weight, and the sum of the second loss function.

[0065] Furthermore, the training module 205 introduces a neural tangent kernel algorithm to dynamically update the physical information loss weights in the first loss function in each round of training, including: Based on the differential equation loss function, the algebraic equation loss function, the second loss function, and the gradient vector of the current parameters, construct the neural tangential kernel matrix; Solve for the trace of the neural tangential kernel matrix to obtain the first intensity value of the contribution of the differential equation loss to the gradient of the current parameter, the second intensity value of the contribution of the algebraic equation loss to the gradient of the current parameter, and the third intensity value of the contribution of the second loss to the gradient of the current parameter. The first physical information loss weight and the second physical information loss weight are calculated based on the first intensity value, the second intensity value, and the third intensity value.

[0066] Further, the training module 205 constructs a neural tangential kernel matrix based on the differential equation loss function, the algebraic equation loss function, the second loss function, and the gradient vector of the current parameters, including: Select the first and second samples from the current training data; Based on the gradient vectors of the first sample, the second sample, and the current parameters, as well as the differential equation loss function, calculate the first neural tangential kernel element vector corresponding to the differential equation loss; Based on the gradient vectors of the first sample, the second sample, and the current parameters, as well as the algebraic equation loss function, calculate the second neural tangential kernel element vector corresponding to the algebraic equation loss. Based on the gradient vectors of the first sample, the second sample, the current parameters, and the second loss function, calculate the third neural tangential kernel element corresponding to the second loss; The first neural tangential kernel element vector, the second neural tangential kernel element vector, and the third neural tangential kernel element constitute the neural tangential kernel matrix.

[0067] Furthermore, the first neural tangential kernel element vector is the inner product of the gradient vector of the differential equation loss with respect to the current parameters for the first sample and the transpose of the gradient vector of the differential equation loss with respect to the current parameters for the second sample. The second neural tangential kernel element vector is the inner product of the gradient vector of the algebraic equation loss for the first sample with respect to the current parameters and the transpose of the gradient vector of the algebraic equation loss for the second sample with respect to the current parameters. The third neural tangential kernel element vector is the inner product of the gradient vector of the second loss with respect to the current parameters for the first sample and the transpose of the gradient vector of the second loss with respect to the current parameters for the second sample.

[0068] Furthermore, the first physical information loss weight is the ratio of the third intensity value to the regularized first intensity value; The second physical information loss weight is the ratio of the third intensity value to the regularized second intensity value.

[0069] The modeling method and apparatus for the boiler-generator coordination system of a coal-fired power generating unit provided in the above embodiments have at least the following beneficial effects: The constructed physical information neural network architecture, combining physical mechanisms with deep data fusion modeling, achieves equivalent mapping and derivative alignment of the physical state equations of the boiler-turbine coordinated system, enhancing the physical interpretability of the model and introducing a self-attention mechanism. This enables dynamic decoupling of the multivariable strongly coupled system and improves the adaptive identification capability of the contribution weights of control variables. By aligning the output with the physical mechanism equations in real time, the model strictly adheres to physical conservation laws while fitting the data, ensuring the physical rationality of the predictions. The neural tangent kernel algorithm is used to dynamically balance the competition between data fitting and physical constraints. This provides a reliable and accurate prediction model for high-performance control and optimization of coal-fired power generating unit coordinated systems.

[0070] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A modeling method for a coal-fired power generating unit boiler-generator coordination system, characterized in that, include: Construct a physical information neural network for coordinated control of coal-fired power generating units and boilers; The current control variable and the previous state variable are used as inputs to the physical information neural network, and the current state variable and the current system output variable are used as outputs to the physical information neural network. A first loss function based on physical laws is constructed based on the current state variables and the current system output variables; a second data-driven loss function is constructed based on the predicted output and the actual output; and the total loss function is obtained based on the first and second loss functions. Acquire historical status data of coordinated operation of coal-fired power generating units and boilers, and construct a training set; The training set is input into the physical information neural network for training. Based on the total loss function, the parameters of the physical information neural network are updated using backpropagation and gradient. A neural tangent kernel algorithm is introduced to dynamically update the physical information loss weights in the first loss function in each round of training.

2. The method according to claim 1, characterized in that, The physical information neural network includes an input module, a fully connected layer, and an output module. The input module is used to input the control variable at the current time and the state variable at the previous time and perform feature extraction. The extracted features are then input to the fully connected layer, which performs calculations based on the input features. The output module outputs the state variable at the current time and the system output variable.

3. The method according to claim 2, characterized in that, The physical information neural network also includes an attention mechanism module, which is used to enhance the relevance of the features extracted by the input module to obtain a deep feature vector. The fully connected layer performs calculations based on the deep feature vector.

4. The method according to claim 1, characterized in that, The first loss function includes a differential equation loss function and an algebraic equation loss function; Construct a first loss function based on physical laws using the current state variables and the current system output variables, including: The physical mechanism equations for the coordinated operation of coal-fired power generating units and boilers are constructed. The physical mechanism equations include differential equations relating control variables, state variables and the rate of change of state variables, as well as algebraic equations relating control variables, state variables and system output variables. A differential equation loss function is constructed based on differential equations and current state variables to characterize the consistency between the dynamic performance and process mechanism of physical information neural networks. An algebraic equation loss function is constructed based on algebraic equations and current system output variables to characterize the consistency between the static performance and process mechanism of a physical information neural network.

5. The method according to claim 4, characterized in that, The second loss function is the mean squared error loss function; The total loss function is the product of the differential equation loss function and the first physical information loss weight, the product of the algebraic equation loss function and the second physical information loss weight, and the sum of the second loss function.

6. The method according to claim 5, characterized in that, A neural tangent kernel algorithm is introduced to dynamically update the physical information loss weights in the first loss function in each training round, including: Based on the differential equation loss function, the algebraic equation loss function, the second loss function, and the gradient vector of the current parameters, construct the neural tangential kernel matrix; Solve for the trace of the neural tangential kernel matrix to obtain the first intensity value of the contribution of the differential equation loss to the gradient of the current parameter, the second intensity value of the contribution of the algebraic equation loss to the gradient of the current parameter, and the third intensity value of the contribution of the second loss to the gradient of the current parameter. The first physical information loss weight and the second physical information loss weight are calculated based on the first intensity value, the second intensity value, and the third intensity value.

7. The method according to claim 6, characterized in that, Based on the differential equation loss function, the algebraic equation loss function, the second loss function, and the gradient vector of the current parameters, a neural tangential kernel matrix is ​​constructed, including: Select the first and second samples from the current training data; Based on the gradient vectors of the first sample, the second sample, and the current parameters, as well as the differential equation loss function, calculate the first neural tangential kernel element vector corresponding to the differential equation loss; Based on the gradient vectors of the first sample, the second sample, and the current parameters, as well as the algebraic equation loss function, calculate the second neural tangential kernel element vector corresponding to the algebraic equation loss. Based on the gradient vectors of the first sample, the second sample, the current parameters, and the second loss function, calculate the third neural tangential kernel element corresponding to the second loss; The first neural tangential kernel element vector, the second neural tangential kernel element vector, and the third neural tangential kernel element constitute the neural tangential kernel matrix.

8. The method according to claim 7, characterized in that, The first neural tangential kernel element vector is the inner product of the gradient vector of the differential equation loss with respect to the current parameters for the first sample and the transpose of the gradient vector of the differential equation loss with respect to the current parameters for the second sample. The second neural tangential kernel element vector is the inner product of the gradient vector of the algebraic equation loss for the first sample with respect to the current parameters and the transpose of the gradient vector of the algebraic equation loss for the second sample with respect to the current parameters. The third neural tangential kernel element vector is the inner product of the gradient vector of the second loss with respect to the current parameters for the first sample and the transpose of the gradient vector of the second loss with respect to the current parameters for the second sample.

9. The method according to claim 6, characterized in that, The first physical information loss weight is the ratio of the third intensity value to the regularized first intensity value; The second physical information loss weight is the ratio of the third intensity value to the regularized second intensity value.

10. A modeling device for a coal-fired power generator unit boiler-generator coordination system, characterized in that, include: The network construction module is used to build a physical information neural network for the coordinated control of the boiler and generator units in coal-fired power generation units; The input / output definition module is used to take the current control variable and the previous state variable as the input of the physical information neural network, and take the current state variable and the current system output variable as the output of the physical information neural network. The loss function construction module is used to construct a first loss function based on physical laws based on the current state variables and the current system output variables; construct a data-driven second loss function based on the predicted output and the actual output; and obtain the total loss function based on the first and second loss functions. The data processing module is used to acquire historical status data of the coordinated operation of coal-fired power generating units and boilers and to build a training set; The training module is used to input the training set into the physical information neural network for training. Based on the total loss function, it uses backpropagation and gradient to update the parameters of the physical information neural network, and introduces a neural tangent kernel algorithm to dynamically update the physical information loss weights in the first loss function in each round of training.

Images