ResFCN-3DVar-based nuclear power plant main steam system modeling parameter inversion optimization method

Abstract

The invention discloses a ResFCN-3DVar-based nuclear power plant main steam system modeling parameter inversion optimization method, and the method comprises the steps: defining main steam system modeling parameter data, and generating a data sample; a residual error full connection network (ResFCN) is constructed, and optimization training is carried out on the residual error full connection network (ResFCN); a data assimilation model is constructed based on a three-dimensional variational method (3DVar), an objective function is optimized by using an iterative method, and finally modeling parameters of the main steam system are inverted. And evaluating the data assimilation performance, and verifying the noise interference resistance effect of the data assimilation result by comparing the analysis value, the observation value and the true value. According to the method, deep learning and data assimilation technologies are effectively fused, and remarkable advantages are shown in the aspects of model modeling parameter inversion and simulation precision optimization.

Description

Technical Field

[0001] The present invention belongs to the technical field of data analysis of the main steam system of nuclear power plants, and particularly relates to an inversion optimization method for modeling parameters of the main steam system of nuclear power plants based on ResFCN-3DVar. Background Art

[0002] As a low-carbon and clean energy, nuclear energy plays a crucial role in the global energy supply and climate change mitigation, accounting for about 9% of the global electricity. The application of nuclear energy is extensive and is mainly realized through nuclear power plants (NPPs), whose systems consist of two parts: the primary loop and the secondary loop. The primary loop includes a nuclear reactor, a main pump, a pressurizer, and a coolant. The secondary loop includes a steam generator, a steam turbine, a condenser, etc. The main steam system in the secondary loop plays a key role in nuclear power plants in transporting high-temperature and high-pressure saturated steam from the steam generator to the steam turbine unit and other steam-consuming equipment.

[0003] The main steam system is a key link connecting the reactor core and the steam turbine, and its operating state is of great significance to the safety and economy of nuclear power plants. Operators usually monitor the operating state of the main steam system, which is mainly obtained through numerical simulation to detect abnormalities and take measures. However, due to some inevitable factors, for example, the unknown internal roughness of the pipeline and the simplification of the pipeline curvature information, there are uncertainties in the modeling parameters, so there is still a difference between the numerical simulation and the actual state. These modeling parameters are defined by boundary values according to engineering rules and vary in the parameter space according to the current operating conditions, which become the main reason for restricting the accuracy of the simulation model. Therefore, calibrating the model through observed data is crucial for improving the simulation accuracy, thereby enhancing the safety and efficiency of nuclear power plants, and data assimilation (DA) is gradually becoming a potential method to achieve this goal.

[0004] In the field of engineering science, data assimilation (DA) technology has been widely applied, which combines measured values with simulated values to improve the accuracy of key parameter simulation. Multiple data assimilation methods, including the Kalman filter (KF), three-dimensional variational method (3DVar), four-dimensional variational method (4DVar), and ensemble Kalman filter (EnKF), have been developed and applied. Such as in the fields of numerical weather prediction, ocean reanalysis, hydrological research, and remote sensing, and have also been widely applied in nuclear engineering. Among many data assimilation methods, the three-dimensional variational method (3DVar) is particularly prominent due to its real-time performance and ability to handle nonlinear problems. For example, two-dimensional radial reflection calculations for optimizing power distribution, constructing the physical field of nuclear reactors, calibrating burnup distributions, and establishing the relationship between burnup and power distribution by combining neural networks.

[0005] The above shows that data analysis technology has good application prospects in the field of nuclear engineering. However, data analysis methods are mainly applied to the reconstruction of the core physical field of nuclear power plants, and the research at the system level of nuclear power units is relatively limited. At the same time, the system-level model of a nuclear power plant usually consists of multiple subsystems and components, forming a complex dynamic system with high coupling and nonlinearity.

[0006] The main steam system is responsible for transporting high-temperature and high-pressure saturated steam from the steam generator to the steam turbine and other steam-consuming devices, which is crucial for the safety of nuclear power plants. Predicting its operating parameters is of great significance for monitoring system performance and improving energy utilization efficiency. Since the damping coefficient of the steam pipeline cannot be directly measured, it is usually estimated through numerical simulation. However, due to model simplification, the simulation results often deviate from the actual situation. Therefore, it is urgent to adopt the three-dimensional variational method (3DVar) to solve the data assimilation (DA) problem of the modeling parameters in the main steam system of nuclear power plants. Summary of the Invention

[0007] The present invention provides an inversion optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar. This method effectively integrates deep learning and data assimilation technologies, showing significant advantages in the inversion of model modeling parameters and the optimization of simulation accuracy, providing a new idea for the modeling and optimization of complex systems in nuclear power plants.

[0008] The technical solution adopted by the present invention is as follows:

[0009] An inversion optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar, comprising the following steps:

[0010] Step 1: Define the modeling parameters of the main steam system and generate data samples;

[0011] Step 2: Construct a residual fully connected network (ResFCN) and optimize and train the residual fully connected network (ResFCN);

[0012] Step 3: Based on the three-dimensional variational method (3DVar), construct a data assimilation model, use an iterative method to optimize the objective function, and finally invert the modeling parameters of the main steam system.

[0013] Step 4: Evaluate the data assimilation performance, and verify the effect of the data assimilation result in resisting noise interference by comparing the analysis value, the observed value, and the true value.

[0014] The said Step 1 includes the following steps:

[0015] Step 1.1: Define the modeling parameters and the observed quantities:

[0016] The modeling parameter is the damping coefficient of the main steam system pipeline, including pipelines such as high-pressure cylinders, low-pressure cylinders, intermediate-pressure cylinders, and steam-water separators and reheaters; the observed variables include steam flow rate, pressure, temperature, etc., such as the extraction steam pressure of the third stage of the high-pressure cylinder, the main steam flow rate, etc.

[0017] Step 1.2: Generate parameter-observed variable data pairs:

[0018] Use Latin Hypercube Sampling (LHS) to generate 20,000 sets of random parameter samples in the damping coefficient parameter space. LHS is a method of uniformly distributed sampling in a multi-dimensional parameter space, ensuring stratified uniform sampling in each dimension, reducing the redundancy of traditional Monte Carlo sampling, and improving the efficiency of numerical calculations. The specific steps of Latin Hypercube Sampling (LHS) are as follows:

[0019] S1. Divide the interval: For each parameter x i , i = 1, 2,..., d, where d represents the dimension of the parameter space, that is, the number of parameters. Divide it into N intervals equally within its value range. In the present invention, N = 20,000.

[0020] S2. Random sampling: Randomly select a point within each interval to ensure that each interval has and only has one sample.

[0021] S3. Random sorting: Randomly permute the sample points of all dimensions to avoid the correlation between sampling points.

[0022] S4. Generate a sampling matrix: Finally, form a parameter matrix X of N×d dimensions, where each row represents a set of parameter samples.

[0023] Use the full-scope simulator of the nuclear power plant to simulate the observed variables corresponding to each group of parameters and generate parameter-observed variable data pairs. Specifically: Input 20,000 sets of parameter samples into the NPP full-scope simulator, calculate the corresponding observed variables, including steam flow rate, pressure, and temperature. Perform data matching, where each set of parameter samples x corresponds to an output y of the simulator, and finally form 20,000 sets of parameter-observed variable data pairs (X, Y).

[0024] Step 1.3: Split the parameter-observed variable data pairs:

[0025] Divide the generated parameter-observed variable data pair samples into: a training set (70%), a validation set (20%), and a test set (10%).

[0026] Step 2 includes the following steps:

[0027] Step 2.1: Design the structure of the Residual Fully Connected Network (ResFCN). ResFCN is a seven-layer fully connected network (FCN), as Figure 1 shown, including:

[0028] The input layer contains 10 neurons, corresponding to the damping coefficient parameters of the steam pipeline;

[0029] The hidden layer has a total of 5 layers, and the number of neurons is 64, 128, 128, 64, and 32 in sequence;

[0030] The third layer generates a 128-dimensional feature vector, which is added element-wise to the output of the fourth layer through a skip connection to ensure dimensional consistency;

[0031] The output layer consists of 6 neurons and uses a linear activation function, which is suitable for continuous value prediction.

[0032] All fully connected layers use ReLU (Rectified Linear Unit) as the activation function and introduce L2 regularization to prevent overfitting. The input of the model is the damping coefficient of the steam pipeline, and the output corresponds to the simulation observables. The ReLU activation function is defined as follows:

[0033] ReLU(x) = max(0, x)

[0034] Among them, when x > 0, ReLU directly outputs x, avoiding the problem of gradient disappearance such as that of other activation functions like sigmoid or tanh. When x ≤ 0, ReLU outputs 0, which can reduce the computational amount during parameter update and improve computational efficiency.

[0035] Step 2.2, Residual connection:

[0036] A residual connection is introduced between the third layer and the fourth layer, as Figure 1 shown, to enhance the feature representation ability and alleviate the problem of gradient disappearance. This makes ResFCN more stable and efficient when modeling the relationship between the main steam system parameters and observables, ensuring accurate modeling of complex non-linear relationships.

[0037] Step 2.3, Training and optimization:

[0038] Let the output of the residual fully connected network (ResFCN) represent the observable values predicted by ResFCN, corresponding to physical quantities such as steam flow rate, pressure, and temperature of the full-range simulator; ResFCN is optimized using the RMSprop optimizer, and the learning rate is set to 0.001 to ensure robust convergence. The loss function used is the Huber loss function, which is defined as:

[0039]

[0040] where L δ (a) represents the Huber loss function, which combines the sensitivity of the mean squared error (MSE) and the robustness of the mean absolute error (MAE); represents the residual between the predicted value and the true value; δ is the threshold, which is set to δ = 0.01 in the present invention. This loss function combines the sensitivity of the mean square error (MSE) and the robustness of the mean absolute error (MAE), and is suitable for handling outliers in the data.

[0041] Early stopping mechanism: When there is no improvement in the validation loss for 30 consecutive training epochs, stop training to prevent overfitting.

[0042] Step 2.4, Model validation:

[0043] Evaluate the performance of ResFCN by calculating the Huber loss and prediction error of the validation set. The formula for the prediction error is

[0044]

[0045] where r represents the relative error, which is used to measure the deviation between the predicted value of ResFCN and the true value, and is expressed as a percentage; m represents the size of the output vector y, that is, the total number of samples used to calculate the error in the dataset; represents the predicted value of the i-th sample by ResFCN; y i represents the true value of the i-th sample in the dataset;

[0046] The prediction error formula is used to verify whether the residual fully connected network (ResFCN) can accurately predict the observed quantity according to the modeling parameters.

[0047] The said step 3 includes the following steps:

[0048] Step 3.1, Using the 3DVar assimilation framework, introduce the prediction gradient formula into the cost function, which avoids the calculations in the process of constructing the background error covariance matrix and inverting the matrix. The cost function after introducing the prediction gradient formula is:

[0049]

[0050] where: J ξ (x) represents the objective cost function; ξ is the Tikhonov regularization factor, which is used to balance the two terms in the equation; |·|2 represents the Euclidean norm; represents the squared norm of the gradient difference between the state variable x and the background state x b , which is used to avoid directly calculating the background error covariance matrix B -1 ; T represents the transpose operation; R is the error covariance matrix of the observation; H is the observation operator, which is used to map the state variable to the observation space; the estimated state quantity x is estimated by using multi-source information through the data assimilation algorithm. Usually, the prior state information required by the data assimilation algorithm is divided into two parts: the initial state estimate x b(Also known as the background state) and the observation vector Y0.

[0051] Use the gradient operator The discrete predicted gradient term is expressed as:

[0052]

[0053] Where: D(x - x b ) represents the second-order finite-difference matrix transformation of the state variable x with respect to the background state x b . It can be used to measure the change of the state variable with spatial position and is used to replace the inverse matrix of the background error covariance matrix in the predicted gradient formula to improve the calculation efficiency. D ∈ R n×n is a second-order finite-difference matrix with Neumann boundary conditions; n represents the dimension of the system state variable, that is, the number of main steam system modeling parameters. In the invention, n represents the total number of key parameters such as the damping coefficient to be optimized.

[0054]

[0055] Step 3.2. Optimization and solution:

[0056] Use Tikhonov regularization to stabilize the optimization process. Tikhonov regularization is a technique for stabilizing the optimization of inverse problems. It adds a regularization term to the cost function to prevent the model from overfitting and enhance the stability of the solution. Specifically, the regularization factor ξ balances between the model fitting and generalization ability by controlling the smoothness of the solution, which helps to balance the model complexity and fitting error and avoid the instability of the solution.

[0057] Use the predicted gradient and ResFCN calculation to calculate the gradient of the cost function; in the 3DVar framework, the predicted gradient method is used to optimize the cost function to reduce the dependence on the background error covariance matrix B. Its discrete expression is as follows:

[0058]

[0059] Where, x and x b represent the estimated state quantity and the background state quantity respectively, and represent the gradient operations on these two quantities respectively. D is a second-order finite-difference matrix, approximately the inverse matrix of the background error covariance matrix. In ResFCN-3DVar, use ResFCN to model the observation operator H and calculate the gradient:

[0060]

[0061] Where, represents the gradient of J ξ(x) Calculate the gradient; ξ is the Tikhonov regularization factor; R represents the error covariance matrix of the observed data; Y0 represents the observed data vector. This gradient is used to guide the optimization process to make the parameter update converge towards the optimal solution.

[0062] Iteratively update the modeling parameters; during the data assimilation process, the modeling parameter x is optimized by the gradient descent method:

[0063]

[0064] where, x (k) and x (k+1) represent the values of the estimated state variables at the k-th and (k + 1)-th iterations respectively; α is the learning rate. In this invention, the RMSprop optimizer is adopted and the learning rate is set to 0.001 to ensure the convergence stability. This method can effectively optimize the modeling parameters of the main steam system, improve the simulation accuracy, and make the numerical simulation results closer to the real observed data.

[0065] In step 4 described above,

[0066] 4.1: Evaluate the data assimilation performance:

[0067] During the training process of the ResFCN model, the relative error between the ResFCN prediction result and the true observed value is used as the performance evaluation index r, and its expression is:

[0068]

[0069] where, is the i-th sample predicted by ResFCN, y i is the i-th sample in the dataset, and m is the size of the output vector y. The root mean square error (RMSE) is used to evaluate the performance of ResFCN-3DVar on the simulated data, and its expression is:

[0070]

[0071] where, n is the number of modeling parameters of the main steam system, x i is the i-th true value, is the i-th analysis value.

[0072] 4.2: Compare the analysis value with the observed value:

[0073] By comparing the analysis value, the observed value and the true value, verify the effect of the data assimilation result in resisting noise interference.

[0074] For a method for inverse optimization of modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar in this invention, the technical effects are as follows:

[0075] 1). Improvement in modeling accuracy:

[0076] By constructing the relationship between the modeling parameters and the observables of the main steam system through ResFCN, it can effectively capture complex non - linear mapping relationships, significantly improving the accuracy of model prediction. The relative deviation of the validation set samples is mainly concentrated within 0.5%.

[0077] 2). Excellent data assimilation effect:

[0078] Using 3DVar combined with the prediction gradient formula eliminates the need to estimate the background error covariance matrix and simplifies the computational complexity. The data assimilation results show that the root - mean - square error (RMSE) of the analysis value is always lower than 0.01, and the average relative deviation is only 0.03%, which can effectively correct the modeling parameters and improve the reliability of the simulation results.

[0079] 3). Strong noise resistance:

[0080] The experimental results show that when this method processes the observed data with added Gaussian noise, the analysis value can still be highly consistent with the true value, showing strong noise resistance and robustness.

[0081] 4). Efficiency and generalization:

[0082] ResFCN stabilizes the gradient flow through the residual structure, reduces the training difficulty, and at the same time adopts an early - stopping mechanism to avoid overfitting, with good training efficiency and generalization performance. Description of the drawings

[0083] The present invention will be further described below in conjunction with the drawings and examples;

[0084] Figure 1 It is a schematic diagram of the structure of the residual fully - connected network (ResFCN).

[0085] Figure 2 It is the flow chart of the method of the present invention.

[0086] Figure 3 It is a graph showing the manifestation of the Huber loss on the training set and the validation set.

[0087] Figure 4 It is for T IP,1 The ResFCN regression prediction graph of;

[0088] Figure 5 It is the relative deviation histogram of all samples in the validation set.

[0089] Figure 6 It is the root - mean - square error (RMSE) graph between the assimilation analysis result and the true value.

[0090] Figure 7(a) shows the estimation of modeling parameters using ResFCN-3DVar Figure 1 ;

[0091] Figure 7(b) shows the estimation of modeling parameters using ResFCN-3DVar Figure 2 .

[0092] Figure 8 Comparison chart of the output of the full-scale simulator of a nuclear power plant after data assimilation and the actual data. DETAILED DESCRIPTION

[0093] This paper proposes a ResFCN-3DVar method to calibrate the modeling parameters of the system through observables. This method uses the residual fully connected network (ResFCN) to construct an observation operator to capture the relationship between the modeling parameters and the observables, and uses the three-dimensional variational (3DVar) method combined with the prediction gradient formula for parameter estimation. In the data set experiment based on the full-range simulator of a nuclear power plant, the results show that the root mean square error (RMSE) after assimilation is less than 1%, verifying the high accuracy and effectiveness of the method.

[0094] The inverse optimization method of modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar includes the following steps:

[0095] like Figure 2 As shown in the figure, the data assimilation of the main steam system of a nuclear power plant based on three-dimensional variational and Gaussian weighted least squares method includes the following steps:

[0096] Step S1: Data preparation and sample generation:

[0097] Step S1.1: Definition of modeling parameters and observations:

[0098] Define the damping coefficient of the pipeline in the main steam system, including the damping coefficient of the high-pressure cylinder, low-pressure cylinder, medium-pressure cylinder, steam-water separator reheater and other equipment. Select key physical quantities such as pressure, temperature, steam flow rate, such as high-pressure cylinder extraction pressure, main steam flow rate, etc.

[0099] Step S1.2: Sample generation:

[0100] Latin Hypercube Sampling (LHS) is used to generate 20,000 sets of random parameter samples in the damping coefficient parameter space. The observations corresponding to each set of parameters are simulated using a full-range simulator to generate parameter-observation data pairs.

[0101] Step S1.3: Data segmentation:

[0102] The generated samples are divided into training set (70%), validation set (20%) and test set (10%).

[0103] Step S2: Construct a Residual Fully Connected Network (ResFCN):

[0104] Step S2.1: Network structure design:

[0105] Input layer: 10 neurons, corresponding to the damping coefficient parameters.

[0106] Hidden layer: 5 layers, with the number of neurons being 64, 128, 128, 64, and 32 in sequence. The ReLU activation function is used, and L2 regularization is added.

[0107] Output layer: 6 neurons, linearly activated, predicting continuous observable values.

[0108] Step S2.2: Residual connection:

[0109] A residual connection is introduced between the 3rd and 4th layers to enhance the feature representation ability and alleviate the problem of gradient disappearance.

[0110] Step S2.3: Training and optimization:

[0111] The Huber loss is adopted, which combines the advantages of MSE and MAE and adapts to the influence of outliers. The specific variation of the training loss with iterations is shown in Figure 3 , indicating that the losses of the validation set and the training set converge well. The RMSprop optimizer is used, and the initial learning rate is 0.001. The early stopping condition is set, and the training stops when the validation set loss shows no improvement after 30 training epochs to prevent overfitting.

[0112] Step S2.4: Model validation:

[0113] Calculate the Huber loss and relative error on the validation set to verify whether the ResFCN can accurately establish the relationship between the modeling parameters and the observable values. The comparison between the predicted values and the true values is shown in Figure 4 , where the blue curve represents the true value, the green curve represents the predicted value, and the red curve represents the relative error. It shows that the prediction effect is good.

[0114] Step S3: Construct a data assimilation model:

[0115] Step S3.1: Introduction of prediction gradient:

[0116] Using the 3DVar assimilation framework, the prediction gradient formula is introduced into the cost function, reducing the computational cost in the process of constructing and inverting the background error covariance matrix.

[0117] Step S3.2: Parameter inversion:

[0118] Taking the trained ResFCN as the observation operator, an assimilation model is constructed using the simulation data of the NPP full-range simulator. An iterative method is used to optimize the objective function, and finally the modeling parameters of the main steam system are inverted. The error distribution of the assimilation results is shown in Figure 5 , indicating that the errors are concentrated in a lower range and the inversion accuracy is relatively high.

[0119] Step S4: Verification and evaluation:

[0120] Step S4.1: Verification metrics:

[0121] The root mean square error (RMSE) and relative error of the assimilated modeling parameters are calculated using the test set to evaluate the optimization effect. The verification results are shown in Figure 6 , indicating that the RMSE of the assimilation results remains below 0.01.

[0122] Step S4.2: Comparison between the analyzed value and the observed value:

[0123] The fitting situation between the simulated output after assimilation and the true value is verified. The RMSE is less than 1%, indicating that the assimilation process significantly improves the accuracy of the simulation results. The modeling parameters assimilated by the ResFCN-3DVar method are shown in Figures 7(a) and 7(b). In Figures 7(a) and 7(b), the blue line represents the actual value, and the red line represents the assimilated analyzed value. The parameter inversion results are close to the actual value, indicating that the ResFCN-3DVar method can effectively calibrate the modeling parameters of the main steam system.

[0124] Step S4.3: Analysis of noise resistance:

[0125] The assimilation results before and after adding Gaussian noise are compared to verify the robustness of the ResFCN-3DVar method under noise interference. As shown in Figure 8 , the assimilation results can effectively resist the influence of noise. The red line represents the output after data assimilation, the blue line represents the actual value, and the green dots represent the observed data points generated by adding noise to the simulated values. ResFCN-3DVar can effectively approximate the simulator output to the actual value and has strong robustness to noise interference.

Claims

1. The inverse optimization method of modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar is characterized by The following steps are involved: Step 1: Define the main steam system modeling parameter data and generate data samples; Step 2: Construct a residual fully connected network ResFCN and optimize the training of the residual fully connected network ResFCN; Step 3: Build a data assimilation model based on the three-dimensional variational method 3DVar, use an iterative method to optimize the objective function, and finally invert the modeling parameters of the main steam system.

2. According to claim 1, the inverse optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar is characterized in that : It also includes step 4: evaluating the data assimilation performance, and verifying the effect of data assimilation results in resisting noise interference by comparing the analysis value, observation value and true value.

3. According to claim 1, the inverse optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar is characterized in that : The step 1 comprises the following steps: Step 1.1: Define modeling parameters and observations: The modeling parameters are the damping coefficient of the main steam system pipeline, and the observed quantities include steam flow, pressure, and temperature; Step 1.2: Generate parameter-observation data pairs: Use Latin hypercube sampling LHS to generate 20,000 sets of random parameter samples in the damping coefficient parameter space; The observations corresponding to each set of parameters are simulated using a full-range simulator of a nuclear power plant to generate parameter-observation data pairs; Step 1.3, parameter-observation data pair segmentation: The generated parameter-observation data pair samples are divided into: training set, validation set and test set.

4. The inverse optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar according to claim 3 is characterized in that : In step 1.2, the specific steps of Latin hypercube sampling LHS sampling are as follows: S1. Divide the interval: for each parameter x i , i = 1, 2, ..., d, d represents the dimension of the parameter space, that is, the number of parameters; its value range is equally divided into N intervals, and in the present invention, N = 20,000; S2. Random sampling: randomly select a point in each interval to ensure that each interval has only one sample; S3. Random sorting: Randomly arrange the sample points of all dimensions to avoid correlation between the sample points; S4. Generate sampling matrix: Finally, an N×d-dimensional parameter matrix X is formed, where each row represents a set of parameter samples.

5. The inverse optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar according to claim 4 is characterized in that: Generate parameter-observation data pairs, specifically: input 20,000 sets of parameter samples into the NPP full-range simulator, calculate their corresponding observations, including steam flow, pressure, and temperature; perform data matching, each set of parameter samples x corresponds to a simulator output y, and finally form 20,000 sets of parameter-observation data pairs (X, Y).

6. The inverse optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar according to claim 1 is characterized in that: The step 2 comprises the following steps: Step 2.1: Design the residual fully connected network ResFCN structure, including: The input layer contains 10 neurons, corresponding to the damping coefficient parameters of the steam pipe; There are 5 hidden layers, and the number of neurons is 64, 128, 128, 64, and 32 respectively; The third layer generates a 128-dimensional feature vector and adds it element-by-element to the output of the fourth layer through skip connections to ensure dimensional consistency; The output layer consists of 6 neurons and uses a linear activation function, which is suitable for continuous value prediction; Step 2.2, residual connection: The residual connection is introduced between the third and fourth layers, making ResFCN more stable and efficient in modeling the relationship between the main steam system parameters and the observed quantities; Step 2.3, training and optimization: Assume the output of the residual fully connected network ResFCN Represents the observed value predicted by ResFCN, corresponding to the physical quantities such as steam flow, pressure, and temperature of the full-range simulator; ResFCN uses the RMSprop optimizer for optimization, and the loss function used is the Huber loss function, which is defined as: Among them, L δ (a) represents the Huber loss function, which combines the sensitivity of mean square error (MSE) and the robustness of absolute error (MAE); Represents the residual between the predicted value and the true value; δ is the threshold; Step 2.4, model verification: The performance of ResFCN is evaluated by calculating the Huber loss and prediction error of the validation set; the formula for the prediction error is Among them, r represents the relative error, which is used to measure the deviation between the ResFCN predicted value and the true value and is expressed in percentage; m represents the size of the output vector y, that is, the total number of samples in the dataset used to calculate the error; Represents the i-th sample value predicted by ResFCN; y i Represents the i-th true sample value in the data set.

7. The inverse optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar according to claim 6 is characterized in that: In step 2.1, all fully connected layers use ReLU as the activation function, and L2 regularization is introduced to prevent overfitting; the input of the model is the damping coefficient of the steam pipe, and the output corresponds to the simulation observable; the ReLU activation function is defined as follows: ReLU(x)=max(0,x) Among them, when x>0, ReLU directly outputs x, and when x≤0, ReLU outputs 0.

8. The inverse optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar according to claim 6 is characterized in that: In step 2.3, an early stopping mechanism is also included: when the validation loss does not improve in 30 consecutive training rounds, training is stopped to prevent overfitting.

9. The inverse optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar according to claim 1 is characterized in that: The step 3 comprises the following steps: Step 3.1: Using the 3DVar assimilation framework, the prediction gradient formula is introduced into the cost function. The cost function after the prediction gradient formula is introduced is: Among them: J ξ (x) represents the objective cost function; ξ is the Tikhonov regularization factor, which is used to balance the two terms in the equation; |·|2 represents the Euclidean norm; Represents the state variable x and the background state x b The square norm of the gradient difference is used to avoid directly calculating the background error covariance matrix B -1 ; T represents the transposition operation; R is the observed error covariance matrix; H is the observation operator, which is used to map the state variable to the observation space; the estimated state quantity x is obtained by estimating the multi-source information through the data assimilation algorithm; The prior state information required by the data assimilation algorithm is divided into two parts: the initial state estimate x b and observation vector Y0; Using the gradient operator The discrete prediction gradient term is expressed as: Where: D(xx b ) represents the state variable x with respect to the background state x b The second-order finite difference matrix transformation can be used to measure the change of state variables with spatial position and is used to replace the inverse matrix of the background error covariance matrix in the prediction gradient formula; D∈R n×n is a second-order finite difference matrix with Neumann boundary conditions; n represents the total number of key parameters such as damping coefficients that need to be optimized; Step 3.2, optimization solution: The Tikhonov regularization is used to stabilize the optimization process, and the predicted gradient and ResFCN are used to calculate the gradient of the cost function. Under the 3DVar framework, the predicted gradient method is used to optimize the cost function, and its discrete expression is as follows: Among them, x and x b denote the estimated state and background state respectively, and Respectively represent the gradient operation of these two quantities; D is the second-order finite difference matrix, which is approximately the inverse matrix of the background error covariance matrix; in ResFCN-3DVar, ResFCN is used to model the observation operator H and calculate the gradient: in, Express J ξ (x) Find the gradient; ξ is the Tikhonov regularization factor; R represents the error covariance matrix of the observed data; Y0 represents the observed data vector; the gradient is used to guide the optimization process so that the parameter update converges towards the optimal solution; Iteratively update the modeling parameters; during the data assimilation process, the modeling parameters x are optimized by the gradient descent method: Among them, x (k) and x (k+1) They respectively indicate that iterations to the kth round and the k+1th round are the values ​​of the estimated state quantity; α is the learning rate.

10. The inverse optimization method for modeling parameters of the main steam system of a nuclear power plant based on ResFCN-3DVar according to claim 1, characterized in that: In step 4, 4.1: Data assimilation performance evaluation: In the training process of the ResFCN model, the relative error between the ResFCN prediction result and the true observation value is used as the performance evaluation index r, and its expression is: in, is the sample predicted by the i-th ResFCN, y i is the i-th sample in the dataset, and m is the size of the output vector y. The root mean square error (RMSE) is used to evaluate the performance of ResFCN-3DVar on simulated data, and its expression is: Where n is the number of modeling parameters of the main steam system, x i is the ith true value, is the i-th analysis value; 4.2: Comparison between analysis value and observation value: By comparing the analysis values, observation values ​​and true values, the effect of data assimilation results in resisting noise interference is verified.

Images