CN-121997751-A - Modeling method and device for coal-fired generator set furnace coordination system

Abstract

The invention provides a modeling method and device of a coal-fired power generation unit machine furnace coordination system, the method comprises the steps of constructing a physical information neural network, taking a current moment control variable and a last moment state variable as inputs of the physical information neural network, taking the current moment state variable and a current system output variable as outputs, constructing a first loss function based on a physical rule according to the current moment state variable and the current system output variable, constructing a second loss function based on predicted output and actual output, 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, adopting back propagation and gradient to update parameters of the physical information neural network based on the total loss function, and introducing a neural tangent kernel algorithm to dynamically update physical information loss weight in the first loss function in each round of training.

Inventors

• XU YINGWEN
• PAN SHU
• WANG SHUHENG
• SHEN WEI
• JU LING
• WU RONG
• WU XIAO
• LI YANLIU
• WU JIAN
• WAN XIAOJIN
• LI LIUBIN
• SHI LIN
• ZHAO YULIN
• CHEN LINGXIAO
• LIU HUI

Assignees

• 国网江苏省电力有限公司泰州供电分公司
• 国网江苏省电力有限公司
• 东南大学

Dates

Publication Date

20260508

Application Date

20260128

Claims (10)

1. 1. A modeling method of a coal-fired power generation unit furnace coordination system, comprising the following steps: Constructing a physical information neural network for coordinated control of a coal-fired generator set and a furnace; taking a current moment control variable and a last moment state variable as the input of the physical information neural network, and taking a current moment state variable and a current system output variable as the output of the physical information neural network; Constructing a second loss function driven by data 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; acquiring historical state data of coordinated operation of a coal-fired generator set and a furnace, and constructing a training set; Inputting the training set into the physical information neural network for training, adopting back propagation and gradient to update parameters of the physical information neural network based on the total loss function, and introducing a neural tangent kernel algorithm to dynamically update physical information loss weight in the first loss function in each training round.
2. 2. The method according to claim 1, wherein the physical information neural network comprises an input module, a full-connection layer and an output module, the input module is used for inputting a current time control variable and a last time state variable and extracting features, the extracted features are input to the full-connection layer, the full-connection layer calculates based on the input features, and the current time state variable and the system output variable are output through the output module.
3. 3. The method of claim 2, wherein the physical information neural network further comprises an attention mechanism module for performing correlation enhancement on the features extracted by the input module to obtain depth feature vectors, and wherein the fully connected layer performs computation based on the depth feature vectors.
4. 4. The method of claim 1, wherein the first loss function comprises a differential equation loss function and an algebraic equation loss function; Constructing a first loss function based on a physical rule according to a current time state variable and a current system output variable, wherein the first loss function comprises the following components: Constructing a physical mechanism equation of coordinated operation of the coal-fired generator set and the furnace, wherein the physical mechanism equation comprises a differential equation of the relation among a control variable, a state variable and a state variable change rate and an algebraic equation of the relation among the control variable, the state variable and a system output variable; Constructing a differential equation loss function for representing the consistency of the dynamic performance of the physical information neural network and the process mechanism based on the differential equation and the state variable at the current moment; and constructing an algebraic equation loss function for representing the consistency of the static performance of the physical information neural network and the process mechanism based on the algebraic equation and the current system output variable.
5. 5. The method of claim 4, wherein the second loss function is a mean square error loss function; The total loss function is the sum of 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 second loss function.
6. 6. The method of claim 5, wherein introducing a neural tangent kernel algorithm dynamically updates physical information loss weights in the first loss function in each round of training, comprising: Constructing a neural tangential kernel matrix according to the differential equation loss function, the algebraic equation loss function, the second loss function and the gradient vector of the current parameter; Solving the trace of the neural tangential kernel matrix, and respectively obtaining a first intensity value of the contribution of the differential equation loss to the current parameter gradient, a second intensity value of the contribution of the algebraic equation loss to the current parameter gradient, and a third intensity value of the contribution of the second loss to the current parameter gradient; and calculating according to the first intensity value, the second intensity value and the third intensity value to obtain the first physical information loss weight and the second physical information loss weight.
7. 7. The method of claim 6, wherein constructing a neural tangential kernel matrix from the differential equation loss function, algebraic equation loss function, the second loss function, and gradient vectors of current parameters comprises: Selecting a first sample and a second sample in the current training data; Calculating a first neural tangential nuclear element vector corresponding to differential equation loss based on the first sample, the second sample, the gradient vector of the current parameter and the differential equation loss function; calculating a second neural tangential core element vector corresponding to algebraic equation loss based on the first sample, the second sample, the gradient vector of the current parameter and the algebraic equation loss function; calculating a third neural tangential nuclear element corresponding to the second loss based on the first sample, the second sample, the gradient vector of the current parameter and the second loss function; The first, second, and third neurotangential core element vectors constitute the neurotangential core matrix.
8. 8. The method of claim 7, wherein the first neural tangential nuclear element vector is an inner product of a differential equation loss for a first sample versus a gradient vector of a current parameter and a differential equation loss for a second sample versus a transpose of the gradient vector of the current parameter; The second neural tangential core element vector is an inner product of algebraic equation loss for the first sample and a transpose of algebraic equation loss for the second sample to the gradient vector of the current parameter; The third neural tangential nuclear element vector is an inner product of a gradient vector of a second penalty for the first sample to the current parameter and a transpose of the gradient vector of the second penalty for the second sample to the current parameter.
9. 9. The method of claim 6, wherein the first physical information loss weight is a ratio of a third intensity value to the regularized first intensity value; the second physical information loss weight is a ratio of a third intensity value to a regularized second intensity value.
10. 10. A modeling apparatus for a coal-fired power generation unit furnace coordination system, comprising: the network construction module is used for constructing a physical information neural network for coordinated control of the coal-fired generator set and the machine furnace; the input-output definition module is used for taking the current moment control variable and the last 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 of the physical information neural network; The system comprises a loss function construction module, a data driving second loss function construction module, a total loss function calculation module and a data driving module, wherein the loss function construction module is used for constructing a first loss function based on a physical rule according to a current moment state variable and a current system output variable; the data processing module is used for acquiring historical state data of coordinated operation of the coal-fired generator set and the machine furnace and constructing a training set; the training module is used for inputting the training set into the physical information neural network for training, adopting back propagation and gradient to update parameters of the physical information neural network based on the total loss function, and introducing a neural tangent kernel algorithm to dynamically update the physical information loss weight in the first loss function in each round of training.

Description

Modeling method and device for coal-fired generator set furnace coordination system Technical Field The invention relates to the field of artificial intelligence, in particular to a modeling method and device for a coal-fired power generation unit machine furnace coordination system. Background Along with the continuous increase of the installed capacity of renewable energy power generation by fluctuation of wind, light and the like, the requirements on the new energy consumption and the power grid stability of coal power guarantee are higher and higher. The coal-fired generating Set (CFPPs) is a complex system comprising multiple types of equipment such as fluid compression transportation, heat exchange, heat power conversion, mechanical transmission and the like, and relates to multiple energy flow and conversion process coupling. A dynamic model of the coal-fired power generation unit is established, and the mutual correlation among key variables is mastered as a key for optimizing control strategies, mining flexible potential and improving the running performance of the unit. However, the problem of unclear mechanism cognition in the fast dynamic process caused by the great change of characteristics due to the deep-tuning working condition migration of the unit increases the modeling difficulty of the coal-fired power generation unit. The traditional modeling of the coal-fired generator set mostly adopts a mechanism modeling method. Although the method has the advantages of clear physical meaning and accurate dynamic law depiction, the modeling process is highly dependent on depth mastering of unit operation mechanisms and design parameters, and a large number of reasonable assumptions are required to be introduced, so that the difficulty of model construction and parameter calibration is high. Meanwhile, the mechanism model is formed by coupling a mass nonlinear algebraic equation and a differential equation, and the complex model structure and huge calculated amount make the mechanism model difficult to be directly applied to engineering practice of unit operation optimization. The mass data generated in the whole-working-condition operation process of the coal-fired unit contains rich information capable of reflecting the dynamic characteristics and operation rules of the unit. The data-driven model identification method can extract the mapping relation among key variables by mining data characteristics, does not need to rely on complex physical mechanisms of units, and provides a convenient and efficient technical path for the construction of the coal-fired generator unit model. For example, patent document CN121028599A discloses a coordinated modeling method of a gating cycle unit coal-fired unit fused with a self-attention mechanism, which comprises the steps of 1, collecting working condition data of the coal-fired unit, preprocessing the data, 2, taking preprocessed data as input and predicted values of power and main steam pressure of the coal-fired unit as output, constructing a GRU network model fused with the self-attention mechanism, and 3, training and optimizing the GRU network model fused with the self-attention mechanism. However, the modeling method based on machine learning still has the inherent defects of weak model interpretability and insufficient generalization reliability while exhibiting strong complex nonlinear fitting capability, and the model performance is highly dependent on data quality and sample size, so that the problem of model mismatch is extremely easy to occur in the scene of data scarcity or uneven distribution. Disclosure of Invention The invention provides a modeling method and device for a coal-fired power generation unit machine furnace coordination system, which are used for improving the accuracy of a coal-fired power generation unit coordination control system model. A modeling method of a coal-fired power generation unit furnace coordination system, comprising: Constructing a physical information neural network for coordinated control of a coal-fired generator set and a furnace; taking a current moment control variable and a last moment state variable as the input of the physical information neural network, and taking a current moment state variable and a current system output variable as the output of the physical information neural network; Constructing a first loss function based on a physical rule according to a state variable at the current moment and an output variable of a current system, constructing a second loss function driven by data according to the predicted output and the actual output, acquiring a total loss function based on the first loss function and the second loss function, Acquiring historical state data of coordinated operation of a coal-fired generator set and a furnace, and constructing a training set; Inputting the training set into the physical information neural network for training, adopting back propagation and gradient to upd