Implementation method of digital twin model of nuclear power plant based on physical-data hybrid drive

Abstract

A method for implementing a digital twin model of a nuclear power plant based on physical-data hybrid drive, constructing a feedforward-backward propagation fully connected neural network in the offline stage, embedding physical knowledge into the loss function of the network to construct a DPNN neural network, introducing residual connections into the DPNN neural network, and then constructing a ResDPNN neural network, using a preprocessing method to obtain a target data set to train the ResDPNN neural network after hyperparameter optimization; in the online stage, the trained ResDPNN neural network is used to construct a digital twin of the nuclear power plant. The present invention can effectively map the coolant temperature of the primary circuit of a nuclear reactor. The method can be used in the construction of the basic model of the digital twin system of a nuclear power plant, and is suitable for building any working conditions. The model can realize the mapping of a real physical system to a digital model in a virtual space, while achieving the purpose of reducing the difficulty of calculation, reducing the amount of calculation, and saving calculation time. It can realize the high-fidelity real-time mapping of a real nuclear power plant in a virtual space, and provide a reference for the construction of the basic model of the digital twin of a future nuclear power plant.

Description

Technical Field

[0001] The present invention relates to a technology in the field of nuclear power plant simulation, specifically a method for realizing a digital twin model of a nuclear power plant based on physical-data hybrid drive. Background Art

[0002] Digital twin models based on mechanism models have limitations such as simplified assumptions, cognitive limitations, modeling errors and numerical errors, which make it impossible to guarantee the calculation accuracy. In addition, due to the complexity of the nuclear reactor system, the use of simulation programs for calculations is time-consuming and costly. The disadvantage of data-based modeling is that data-driven modeling does not rely on system mechanisms, so the model has poor interpretability and its performance is highly dependent on the quantity and quality of input data. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention proposes a method for realizing a digital twin model of a nuclear power plant based on a physical-data hybrid drive, which can effectively map the coolant temperature of the primary circuit of a nuclear reactor. This method can be applied to the construction of the basic model of the digital twin system of a nuclear power plant, and is suitable for building any working conditions. This model can realize the mapping of a real physical system to a digital model in a virtual space, while achieving the purpose of reducing the difficulty of calculation, reducing the amount of calculation, and saving calculation time. It can realize the high-fidelity real-time mapping of a real nuclear power plant in a virtual space, and provide a reference for the construction of the basic model of the digital twin of a future nuclear power plant.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to a method for realizing a digital twin model of a nuclear power plant based on physical-data hybrid drive. In an offline stage, a feedforward-backward propagation fully connected neural network is constructed. After the loss function of the network is embedded in physical knowledge to construct a DPNN neural network, a residual connection is introduced into the DPNN neural network to construct a ResDPNN neural network. A preprocessing method is used to obtain a target data set to train the ResDPNN neural network after hyperparameter optimization. In an online stage, the trained ResDPNN neural network is used to construct a digital twin of a nuclear power plant.

[0006] The target data set is obtained by performing data preprocessing and selecting characteristic parameters using the actual operating conditions of a level 0 operation event in which a nuclear power plant loses its main external power supply.

[0007] The present invention relates to a construction system of a digital twin model of a nuclear power plant for realizing the above method, comprising: a data preprocessing unit, a physical-data hybrid driven digital twin model construction unit and a result evaluation and comparison unit, wherein: the data preprocessing unit cleans, extracts and integrates the real data of the nuclear power plant, selects characteristic parameters and searches for physical property parameters required in physical information according to specific goals, obtains a data set that can be input into the digital twin model and further divides it into a training set, a verification set and a test set; the digital twin model construction unit is based on a back propagation fully connected neural network, embeds physical knowledge into the loss function of the back propagation fully connected neural network, constructs a DPNN neural network, further introduces residual connection to the DPNN neural network, constructs a ResDPNN neural network, and optimizes and trains hyperparameters to obtain a digital twin model of a nuclear power plant; the result evaluation and comparison unit tests the trained digital twin model of the nuclear power plant to obtain the prediction result of the model, compares and evaluates the prediction result with the actual data, and calculates the result evaluation index. The prediction accuracy and generalization ability of the digital twin model are evaluated based on the evaluation index, the result comparison chart and the frequency distribution histogram of the absolute error.

[0008] Technical Effects

[0009] The present invention uses a physical-data hybrid driven ResDPNN neural network with residual connections to construct a digital twin model of a nuclear power plant, and integrates the law of conservation of energy and residual connection technology into a feedforward-backpropagation fully connected neural network to construct a ResDPNN digital twin model, which enhances the interpretability of the model and significantly reduces the difficulty and amount of calculation. At the same time, the present invention also has generalization capabilities and is suitable for more scenarios of the same type, and can achieve high-fidelity real-time mapping of real nuclear power plants in virtual space. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flow chart of the present invention;

[0011] Figure 2 It is a schematic diagram of the data preprocessing module;

[0012] Figure 3 It is a structural diagram of residual connection;

[0013] Figure 4 This is a schematic diagram of the specific structure of the ResDPNN neural network;

[0014] Figure 5 It is a schematic diagram of the effect of the embodiment. DETAILED DESCRIPTION

[0015] like Figure 1As shown, this embodiment involves a method for implementing a digital twin model of the average temperature of the primary coolant of a nuclear reactor in the event of a nuclear power plant losing its main power supply. A feedforward-backward propagation fully connected neural network is constructed. After the network loss function is embedded in physical knowledge to construct a DPNN neural network, a residual connection is introduced into the DPNN neural network, and then the following is constructed: Figure 4 The ResDPNN neural network shown in Figure 2 The preprocessing method shown obtains the target data set to train the ResDPNN neural network after hyperparameter optimization; in the online stage, the trained ResDPNN neural network is used as the digital twin model of the nuclear power plant. In the application, the operating conditions of the nuclear power plant can be predicted based on the digital twin model, the digital twin can be assimilated with the newly added nuclear power plant data, and the online monitoring of the nuclear power plant, the life prediction of the nuclear power plant equipment, fault diagnosis and the construction optimization of the nuclear power plant can be realized.

[0016] like Figure 2 As shown, the preprocessing method is to obtain the actual data of the detector from the real nuclear power plant. Since most working conditions in the nuclear power plant have multiple detection points, it is necessary to select the required detection point data according to the selected specific digital twin model mapping object, and combine the domain knowledge provided by the physical information to obtain the measured model training original data; since the measured original data usually contains abnormal values ​​and missing values, the obvious abnormal values ​​are judged by the comprehensive analysis of the physical information, and the missing values ​​are found by data statistics and manual screening. Therefore, it is necessary to remove and fill them, and the average value of the detection data is used to fill them; since there are 46 working conditions contained in the original data, it is necessary to select the required working conditions as feature parameters according to actual needs, specifically: select nuclear power, main pump speed, regulator pressure, regulator water level and steam generator secondary side feed water flow as feature parameters, and use the average temperature of the primary circuit coolant as the mapping object of the digital twin; in addition, the embedded physical knowledge requires the use of certain physical parameters, which are not included in the original data and need to be supplemented to form a complete target data set.

[0017] The physical property parameters include: coolant is the density of the coolant in the hot pipe section, C p_coolant is the constant pressure specific heat capacity of the coolant in the hot pipe section, G coolant is the coolant flow rate, is the output of the digital twin model, i.e., the average temperature of the primary coolant, T cold_leg is the temperature of the coolant in the cold pipe section, ρ SG_in is the density of the feed water flow rate on the secondary side of the steam generator, C p_SG_in is the constant pressure specific heat capacity of the secondary side feed water flow, G SG_in is the secondary side water flow rate, T SG_in is the temperature of the secondary side feed water flow, ρSG_out is the density of steam flow on the secondary side of the steam generator, C p_SG_out is the constant pressure specific heat capacity of the secondary side steam flow, G SG_out is the secondary side steam flow rate, T SG_out is the temperature of the secondary side steam flow. Therefore, the physical parameters that need to be supplemented are the density and constant pressure specific heat capacity mentioned above.

[0018] The target data set is preferably divided into a training set, a validation set, and a test set. The measurement data of the first 7400 seconds of the event process is selected as the original data set, and the detector detects data every 10 seconds. The characteristic parameters of these time nodes and the corresponding physical property parameters are selected, and the data set is randomly divided according to the ratio of 50%, 20%, and 30%, and the data is reduced by Z value standardization to obtain the target data set that can be used for the digital twin model.

[0019] The hyperparameter optimization refers to that machine learning or deep learning models usually have certain parameters that need to be set manually, which often have a great impact on model performance. By adjusting the hyperparameters, the structure and configuration of the model can be optimized, thereby improving its performance on specific tasks. At the same time, reasonable hyperparameter settings can help the model obtain good generalization ability on training data and prevent overfitting. The hyperparameters of the digital twin model are adjusted specifically using the Bayesian optimization algorithm, which is a search function based on the Bayesian method, especially the method of global extreme values ​​of high-dimensional nonlinear non-convex functions. According to the currently tested hyperparameter combination, the next combination that may bring the greatest benefit is predicted. Specifically, the absolute mean error of the validation set is used as the optimization indicator to search the number of hidden layers, the number of hidden layer nodes and the batch size of the digital twin model, and the optimal hyperparameters of the digital twin model are obtained at the end of the optimization. The number of Bayesian optimizations is set to 250 times, and the number of rounds of each model optimization is set to 100. The optimal number of hidden layers of the model obtained is 49 layers, the optimal number of hidden layer nodes is 82 layers, and the optimal batch size is 32.

[0020] The feedforward-backward propagation fully connected neural network includes: a characteristic parameter calculation unit, a model hyperparameter optimization unit, an error back propagation unit, an optimal model parameter retention unit and a model performance evaluation unit, wherein: the characteristic parameter calculation unit inputs the nuclear power, the main pump speed, the regulator pressure, the regulator water level and the steam generator secondary side feed water flow as characteristic parameters, and the expected output obtained by calculating the parameters of the fully connected neural network is the average temperature of the primary circuit coolant; the model hyperparameter optimization unit uses the absolute average error of the validation set as the optimization index, uses the Bayesian optimization method to search the number of hidden layers, the number of hidden layer nodes and the batch size of the digital twin model, and obtains the optimal hyperparameters of the digital twin model at the end of the optimization; the error back propagation unit takes the absolute average error of the validation set as the optimization index, and .... The error is obtained by comparing the output value of the fully connected neural network with the actual value of the target parameter, and the weight and bias of the fully connected neural network are corrected by the back propagation algorithm to gradually reduce the prediction error, so that the expected output obtained by calculating the parameters of the fully connected neural network is as close to the average temperature of the primary coolant as possible. Specifically, RMSE is selected as the loss function, and the Adam optimizer is used for optimization. The default learning rate of the optimizer is selected as the learning rate. The best model parameter retention unit records the number of training rounds when the validation set is optimal according to the results of the training process, and saves the model parameters of the number of rounds. After the model training converges, the model performance evaluation unit applies the obtained neural network to the test set, and evaluates its prediction accuracy and generalization ability according to the error between the predicted value and the actual value.

[0021] The embedded physical knowledge is specifically: the energy conservation of the first and second loops of the nuclear power plant, which is embedded in the digital twin model. The specific form in the loss function of the feedforward-backward propagation fully connected neural network is y is the actual value of the average temperature of the coolant in the first circuit, is the output value of the digital twin model, Δenergy1 is the energy change value of the reactor primary circuit, and Δenergy2 is the energy change value of the reactor secondary circuit. Before calculation, the variables contained in Δenergy1 and Δenergy2 need to be denormalized to ensure that the original physical meaning remains unchanged.

[0022] The residual connection mentioned here refers to the residual connection structure of the residual network (ResNet). ResNet improves the efficiency of information transmission and solves the problem of gradient vanishing or gradient exploding in deep networks by residual connection, and extends the depth of the network to more than dozens of layers, thereby improving the performance of the model. The basic idea of ​​residual connection is to directly connect the input signal to the output in certain layers of the network, thereby introducing "cross-layer connection" in the network. Figure 3As shown, the output of each hidden layer of the neural network is added with the input signal of the hidden layer. Its mathematical description is: the expectation of the deep learning model is to let the model f(x; θ) approximate the target function h(x), so the target function is split into the identity function x and the residual function h(x)-x, let the model f(x; θ) approximate the residual function h(x)-x, and use f(x; θ)+x to approximate h(x).

[0023] Preferably, in order to evaluate the error between the output value of the nuclear power plant digital twin model and the actual value of the target parameter and to measure the prediction accuracy and generalization ability of the digital twin model, it is necessary to formulate evaluation indicators to evaluate the results, specifically the mean absolute error MAE, the mean relative error MAPE, the maximum absolute prediction error MaxAE and the determination coefficient R 2 ,in: Where: y i is a certain actual value of the average temperature of the coolant in a circuit, A corresponding value output by the digital twin model.

[0024] like Figure 4 As shown, the ResDPNN neural network includes: a feedforward-backward propagation fully connected neural network unit with residual connection, a digital twin model hyperparameter optimization unit, a loss function calculation unit, and a digital twin model parameter optimization unit, wherein: the feedforward-backward propagation fully connected neural network unit with residual connection introduces residual connection according to the aforementioned feedforward-backward propagation fully connected neural network, and adds the output of each hidden layer of the neural network to the input signal of the hidden layer to obtain a feedforward-backward propagation fully connected neural network with residual connection; the digital twin model hyperparameter optimization unit uses the absolute average error of the model on the validation set as the optimization index, searches the number of hidden layers, the number of hidden layer nodes and the batch size of the digital twin model through the hyperparameter optimization method, and obtains the optimal hyperparameters of the digital twin model at the end of the optimization; the loss function calculation unit The calculation unit calculates the expected output calculated by the feedforward-backward propagation fully connected neural network unit with residual connection after hyperparameter optimization, and calculates the loss with the actual value to obtain the data loss term. At the same time, according to the expected output, it is denormalized to ensure that the original physical meaning remains unchanged, and the loss is calculated by the above-mentioned physical knowledge to obtain the physical knowledge loss term. Specifically, RMSE is selected to calculate the loss, and the weighted average of the data loss term and the physical knowledge loss term is used as the loss function of the digital twin model; the digital twin model parameter optimization unit corrects the weight and bias of the model through the back propagation algorithm according to the result obtained by the loss function calculation unit to gradually reduce the prediction error, so that the expected output calculated by the model is as consistent as possible with the actual value of the average temperature of the primary coolant, and the Adam optimizer is used for optimization, and the learning rate selects the default learning rate of the optimizer.

[0025] After specific experiments, using real data from 740 nuclear power plants, a ResDPNN neural network was established for the primary coolant flow rate of nuclear reactors. The digital twin model of the average temperature of the primary coolant of the nuclear reactor obtained showed a schematic diagram of the predicted value and actual value in the test set and a frequency distribution histogram of the absolute error. Figure 5 As shown in the figure, most of the data points in the diagram of predicted values ​​and actual values ​​are concentrated on the diagonal line, and the frequency distribution histogram of absolute error shows that the absolute errors of most data points are small, concentrated in the range of 0 to 0.2, with the maximum absolute error being 0.9625. In addition, the coefficient of determination R 2 The average absolute error MAE is 0.99946, the average absolute error MAE is 0.09966, and the average relative error MAPE is 0.03558%. The above results show that this method can effectively map the coolant temperature of the primary circuit of the nuclear reactor. This method can be used in the construction of the basic model of the digital twin system of the nuclear power plant, and is suitable for building any working conditions, so that the mirror synchronization between the digital twin of the nuclear power plant and the physical entity of the nuclear power plant can be faster and more accurate, and can also achieve the purpose of reducing the calculation dimension, reducing the amount of calculation, and saving calculation time. At the same time, the embedding of physical information enhances the interpretability and generalization ability of the digital twin model, and the introduction of residual connection improves the prediction accuracy and expression ability of the model, making the model more reliable. Therefore, this method can provide a reference for the construction of the basic model of the digital twin of the nuclear power plant in the future.

[0026] Compared with the existing nuclear power plant digital twin model construction method, the determination coefficient R obtained by using K-nearest neighbor regression (KNN Regression) to compare the same data set is 2 The determination coefficient R is 0.99706 and the mean absolute error MAE is 0.1479. The same data set is compared and tested using a feedforward-backward propagation fully connected neural network. 2 The determination coefficient R is 0.99769 and the mean absolute error MAE is 0.1403. The DPNN neural network is used to compare the same data set and get the determination coefficient R 2 The average absolute error (MAE) is 0.99899 and the average absolute error (MAE) is 0.107925.

[0027] The present invention constructs a ResDPNN digital twin model by integrating the law of conservation of energy and residual connection technology into a feedforward-backward propagation fully connected neural network. This method enhances the interpretability of the model. The experimental results show that the generalization performance of the model is improved, and the experimental results show that the model has a deeper number of layers, which means that the model has a stronger expression ability. At the same time, the evaluation index shows that the prediction accuracy of the model is improved. This method also combines the real data of the nuclear power plant losing the main power supply outside the plant, and uses the ResDPNN digital twin model proposed by this method to achieve high-precision mapping of the average temperature of the primary coolant of the nuclear reactor. The evaluation index shows that its results are better than several more representative data-driven digital twin model construction methods.

[0028] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A method for implementing a nuclear power plant digital twin model based on physical-data hybrid drive, characterized in that: In the offline stage, a feedforward-backward propagation fully connected neural network is constructed. After the loss function of the network is embedded in physical knowledge to construct a DPNN neural network, the residual connection is introduced into the DPNN neural network to construct a ResDPNN neural network. The target data set is obtained by preprocessing methods to train the ResDPNN neural network after hyperparameter optimization. In the online stage, the trained ResDPNN neural network is used to construct the digital twin of the nuclear power plant. The target data set is obtained by using the real operating conditions of a nuclear power plant losing its main external power supply at level 0, performing data preprocessing and selecting characteristic parameters; The preprocessing method is as follows: after obtaining the actual measured data of the detector from the real nuclear power plant, appropriate detection point data is selected, and the original data for model training is obtained by combination, and the nuclear power, main pump speed, pressurizer pressure, pressurizer water level and steam generator secondary side feed water flow are selected as characteristic parameters after elimination and filling, and the average temperature of the primary circuit coolant is used as the mapping object of the digital twin, and the physical property parameters required by the embedded physical knowledge are additionally supplemented, and finally a complete target data set is formed; The embedded physical knowledge is specifically: the energy conservation of the first and second loops of the nuclear power plant, which is embedded in the digital twin model. The specific form in the loss function of the feedforward-backward propagation fully connected neural network is in: Δenergy2=|ρ SG_in *C p_SG_in *G SG_in *T SG_in -ρ SG_out *C p_SG_out *G SG_out *T SG_out |, y is the actual value of the average temperature of the coolant in the first circuit, is the output value of the digital twin model, Δenergy1 is the energy change value of the reactor primary circuit, and Δenergy2 is the energy change value of the reactor secondary circuit. Before calculation, the variables contained in Δenergy1 and Δenergy2 need to be denormalized to ensure that the original physical meaning remains unchanged; The physical property parameters include: coolant is the density of the coolant in the hot pipe section, C p_coolant is the constant pressure specific heat capacity of the coolant in the hot pipe section, G coolant is the coolant flow rate, is the output of the digital twin model, i.e., the average temperature of the primary coolant, T cold_leg is the temperature of the coolant in the cold pipe section, ρ SG_in is the density of feed water flow rate on the secondary side of the steam generator, C p_SG_in is the constant pressure specific heat capacity of the secondary side feed water flow, G SG_in is the secondary side water flow rate, T SG_in is the temperature of the secondary side feed water flow, ρ SG_out is the density of steam flow on the secondary side of the steam generator, C p_SG_out is the constant pressure specific heat capacity of the secondary side steam flow, G SG_out is the secondary side steam flow rate, T SG_out is the temperature of the steam flow on the secondary side, so the physical parameters that need to be supplemented are the density and constant-pressure specific heat capacity mentioned above.

2. The method for implementing a nuclear power plant digital twin model based on physical-data hybrid drive according to claim 1 is characterized in that: The hyperparameter optimization refers to: using the Bayesian Optimization algorithm to predict the next combination that may bring the greatest benefit based on the currently tested hyperparameter combination, specifically: using the absolute mean error of the validation set as the optimization indicator, searching the number of hidden layers, the number of hidden layer nodes and the batch size of the digital twin model, and obtaining the optimal hyperparameters of the digital twin model at the end of the optimization. The number of Bayesian optimizations is set to 250 times, and the number of rounds of each model optimization is set to 100. The optimal number of hidden layers of the model is 49, the optimal number of hidden layer nodes is 82, and the optimal batch size is 32.

3. The method for implementing a nuclear power plant digital twin model based on physical-data hybrid drive according to claim 1 is characterized in that: The feedforward-backward propagation fully connected neural network includes: a characteristic parameter calculation unit, a model hyperparameter optimization unit, an error back propagation unit, an optimal model parameter retention unit and a model performance evaluation unit, wherein: the characteristic parameter calculation unit inputs the nuclear power, the main pump speed, the regulator pressure, the regulator water level and the steam generator secondary side feed water flow as characteristic parameters, and the expected output obtained by calculating the parameters of the fully connected neural network is the average temperature of the primary circuit coolant; the model hyperparameter optimization unit uses the absolute average error of the validation set as the optimization index, uses the Bayesian optimization method to search the number of hidden layers, the number of hidden layer nodes and the batch size of the digital twin model, and obtains the optimal hyperparameters of the digital twin model at the end of the optimization; the error back propagation unit takes the absolute average error of the validation set as the optimization index, and .... The error is obtained by comparing the output value of the fully connected neural network with the actual value of the target parameter, and the weight and bias of the fully connected neural network are corrected by the back propagation algorithm to gradually reduce the prediction error, so that the expected output obtained by calculating the parameters of the fully connected neural network is as close to the average temperature of the primary coolant as possible. Specifically, RMSE is selected as the loss function, and the Adam optimizer is used for optimization. The default learning rate of the optimizer is selected as the learning rate. The best model parameter retention unit records the number of training rounds when the validation set is optimal according to the results of the training process, and saves the model parameters of the number of rounds. After the model training converges, the model performance evaluation unit applies the obtained neural network to the test set, and evaluates its prediction accuracy and generalization ability according to the error between the predicted value and the actual value.

4. The method for implementing a nuclear power plant digital twin model based on physical-data hybrid drive according to claim 1 is characterized in that: The residual connection means that the input signal of each hidden layer is added to the output of each hidden layer of the neural network, that is, the objective function is split into the identity function x and the residual function h(x)-x, the model f(x; θ) is used to approximate the residual function h(x)-x, and f(x; θ)+x is used to approximate h(x).

5. The method for implementing a nuclear power plant digital twin model based on physical-data hybrid drive according to claim 1 is characterized in that: The ResDPNN neural network includes: a feedforward-backward propagation fully connected neural network unit with residual connection, a digital twin model hyperparameter optimization unit, a loss function calculation unit, and a digital twin model parameter optimization unit, wherein: the feedforward-backward propagation fully connected neural network unit with residual connection introduces residual connection according to the aforementioned feedforward-backward propagation fully connected neural network, and adds the output of each hidden layer of the neural network to the input signal of the hidden layer to obtain a feedforward-backward propagation fully connected neural network with residual connection; the digital twin model hyperparameter optimization unit uses the absolute average error of the model on the validation set as an optimization index, searches the number of hidden layers, the number of hidden layer nodes and the batch size of the digital twin model through hyperparameter optimization, and obtains the optimal hyperparameters of the digital twin model at the end of the optimization; the loss function calculation unit The expected output calculated by the feedforward-backward propagation fully connected neural network unit with residual connection after hyperparameter optimization is compared with the actual value to obtain the data loss term. At the same time, according to the expected output, it is denormalized to ensure that the original physical meaning remains unchanged. The loss is calculated by the above-mentioned physical knowledge to obtain the physical knowledge loss term. Specifically, RMSE is selected to calculate the loss, and the weighted average of the data loss term and the physical knowledge loss term is used as the loss function of the digital twin model. The digital twin model parameter optimization unit corrects the weights and biases of the model through the back propagation algorithm according to the results obtained by the loss function calculation unit to gradually reduce the prediction error, so that the expected output calculated by the model is as similar as possible to the actual value of the average temperature of the primary coolant. The Adam optimizer is used for optimization, and the learning rate selects the default learning rate of the optimizer.

6. The method for implementing a nuclear power plant digital twin model based on physical-data hybrid drive according to claim 1 is characterized in that: In order to evaluate the error between the output value of the nuclear power plant digital twin model and the actual value of the target parameter and measure the prediction accuracy and generalization ability of the digital twin model, it is necessary to formulate evaluation indicators to evaluate the results, specifically the mean absolute error MAE, mean relative error MAPE, maximum absolute prediction error MaxAE and determination coefficient R 2 ,in: Where: y i is a certain actual value of the average temperature of the coolant in a circuit, A corresponding value output by the digital twin model.

7. A system for constructing a digital twin model of a nuclear power plant implementing the method described in any one of claims 1 to 6, characterized in that: include: Data preprocessing unit, physical-data hybrid driven digital twin model construction unit and result evaluation and comparison unit, where: the data preprocessing unit cleans, extracts and integrates the real data of the nuclear power plant, selects characteristic parameters according to specific goals and searches for the physical property parameters required in the physical information, obtains the data set that can be input into the digital twin model and further divides it into training set, verification set and test set; The digital twin model construction unit is based on the back-propagation fully connected neural network, embeds physical knowledge into the loss function of the back-propagation fully connected neural network, constructs the DPNN neural network, and further introduces residual connection to the DPNN neural network to construct the ResDPNN neural network. It then optimizes and trains its hyperparameters to obtain the digital twin model of the nuclear power plant. The result evaluation and comparison unit tests the trained digital twin model of the nuclear power plant to obtain the prediction results of the model, compares and evaluates the prediction results with the actual data, and calculates the result evaluation indicators.

Images