Rapid prediction system and model construction method for key parameters of nuclear reactor containment

Abstract

A method for constructing a rapid prediction system and model for key parameters of a nuclear reactor containment, which is used for accident prediction of a passive containment cooling system, comprises the following steps: initializing containment parameters and constructing a data set; constructing an LSTM rapid prediction model, which includes a single parameter model and a multi-parameter collaborative model, and using the single parameter model and the multi-parameter collaborative model to predict the transient state of safety parameters; training the model using the data set to obtain the optimal LSTM rapid prediction model; inputting containment parameters into the optimal LSTM rapid prediction model to predict the safety performance of the containment. The system comprises an input parameter acquisition module, a data processing module, a prediction module, and a prediction result verification module. The results of the present invention can provide a rapid prediction intelligent analysis model for the sensitivity, uncertainty analysis, and system design optimization of key safety parameters of a nuclear reactor containment.

Description

Technical Field

[0001] The present invention relates to the field of accident prediction of a passive containment cooling system of a nuclear reactor, and in particular to a system for rapidly predicting key parameters of a nuclear reactor containment and a method for constructing a model. Background Art

[0002] The main steam pipe rupture accident MSLB is a typical design basis accident that threatens the integrity of the nuclear reactor containment. In the event of an MSLB accident, the passive containment cooling system PCCS transfers the heat in the containment to the final heat sink through passive natural circulation, which is used to alleviate the sudden rise in pressure and temperature caused by the mass and energy release of the nuclear reactor primary loop rupture under accident conditions, ensure the integrity of the nuclear reactor containment and prevent the leakage of radioactive materials. The existing technology widely uses two-phase flow thermal hydraulic analysis programs to model the nuclear reactor containment. By solving complex two-phase flow equations, namely, the conservation of mass, momentum, and energy of the vapor phase, and the conservation of mass, momentum, and energy of the liquid phase, the complex vapor-liquid mass-energy exchange between the two phases is considered, and this is used as the basic physical model to model the containment and analyze the transient characteristics under the MSLB accident. However, the above methods generally have the following problems: slow solution speed, low analysis efficiency, and the solution accuracy depends on the empirical relationship of two-phase flow heat transfer.

[0003] Therefore, how to more accurately and quickly predict the transient response of key safety parameters of the passive containment cooling system PCCS of a nuclear reactor under a MSLB accident over time is a technical problem that the present invention urgently needs to solve. Summary of the invention

[0004] The purpose of the present invention is to provide a nuclear reactor containment key parameter rapid prediction system and model construction method, which can more accurately and quickly predict the transient response of the key safety parameters of the passive containment cooling system PCCS changing with time under MSLB accidents.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A method for constructing a rapid prediction model for key parameters of a containment shell comprises the following steps:

[0007] S1: Initialize containment parameters, construct a data set, and normalize, segment, and divide the data set;

[0008] S2: constructing an LSTM fast prediction model, wherein the LSTM fast prediction model includes a single parameter model and a multi-parameter collaborative model, and using the single parameter model and the multi-parameter collaborative model to predict the transient response characteristics of the safety parameters;

[0009] Using the data set to train the LSTM fast prediction model to obtain an optimal LSTM fast prediction model;

[0010] S3: Input the containment parameters into the optimal LSTM fast prediction model to predict the containment safety performance.

[0011] Optionally, partitioning of the data set:

[0012] According to the length of time series, it is divided into small sample data set and large sample data set.

[0013] Optionally, the data segmentation is a segmentation of data set feature-label pairs, where the data set features and labels are segmented in a one-to-one correspondence.

[0014] Optionally, the single parameter model: first LSTM loop calculation layer-Dropout layer-second LSTM loop calculation layer-Dropout layer-fully connected layer;

[0015] The multi-parameter collaborative model: the third LSTM loop calculation layer-Dropout layer-fully connected layer;

[0016] The first LSTM loop calculation layer, the second LSTM loop calculation layer, and the third LSTM loop calculation layer have different numbers of memory nodes.

[0017] Optionally, in the construction of the LSTM fast prediction model, the LSTM fast prediction model is trained using a training set; the effect of the training process is observed using a validation set; and the performance of the LSTM fast prediction model is tested using a test set;

[0018] When the test set is used to test the LSTM fast prediction model, the time series of the test set does not change.

[0019] Optionally, the training set, the validation set, and the test set in the single parameter model all use small sample data sets;

[0020] The training set and the validation set in the multi-parameter collaborative model also use small sample data sets and the data in the multi-parameter collaborative model needs to be preprocessed. The distribution ratio of the training set and the validation set after the data preprocessing process is consistent with that of the single-parameter model;

[0021] In the multi-parameter collaborative model, the entire large sample data set is used as the test set.

[0022] A containment key parameter rapid prediction system, comprising:

[0023] Input parameter acquisition module, where the input parameters include initial pressure in the containment, initial liquid film coverage, cooling water flow, and wind speed;

[0024] A data processing module pre-processes the parameter data acquired by the input parameter acquisition module;

[0025] The prediction module inputs the input parameter data processed by the data processing module into the LSTM fast prediction model to obtain the prediction value;

[0026] The prediction result verification module compares the predicted value with the actual value, and uses the mean square error (MSE) to evaluate the prediction effect of the model.

[0027] Optionally, the LSTM fast prediction model is obtained through claims 1-6.

[0028] From the above content, it can be known that compared with the prior art, the beneficial effects of the present invention are:

[0029] The method of training a small sample data set to predict a large sample data set with a length of ten times the time series is feasible. The multi-parameter collaborative model is applicable to the case of different working conditions of similar accidents. Both the single parameter model and the multi-parameter collaborative model can effectively predict the transient state of safety parameters with high prediction accuracy. The single parameter model prediction error MSE is 10 -2 The prediction error MSE of the multi-parameter collaborative model is as low as 10 -8 The prediction accuracy of the single parameter and multi-parameter collaborative model using LSTM is generally higher than that of the model using RNN network. The prediction is closer to the analysis results of the complex containment two-phase flow thermal analysis model, and can more accurately reflect the change trend of transient safety parameters under MSLB accidents. The results of the present invention can provide a fast prediction intelligent analysis model for the sensitivity, uncertainty analysis and system design optimization of key safety parameters of nuclear reactor containment. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0031] Figure 1 Detailed step flow chart of Example 1 of the present invention;

[0032] Figure 2 Schematic diagram of segmentation of data set-label pairs of the present invention;

[0033] Figure 3 Schematic diagram of the LSTM model established for the present invention;

[0034] Figure 4 The pressure prediction result of the containment of the multi-parameter system model in Example 1 of the present invention;

[0035] Figure 5 The comparison of the prediction of the evaporation of the liquid film on the outer wall of the containment vessel by the multi-parameter collaborative model in Example 1 of the present invention;

[0036] Figure 6 This is a comparison chart of the predicted results and actual values ​​of the RNN model and LSTM model established using the method in Example 1 of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Embodiment 1:

[0039] The present invention is based on the LSTM time series deep learning model to perform high-precision and rapid prediction and analysis of the transient parameters of the containment in the event of a main steam pipe rupture MSLB accident in the passive containment cooling system. The specific operation steps are as follows: Figure 1 shown.

[0040] (1) Setting initial input parameters

[0041] Input parameters: initial pressure in the containment, initial liquid film coverage, cooling water flow rate, wind speed; as shown in Table 1, these are the parameters selected during the demonstration and implementation process in this embodiment.

[0042] Table 1 Initial input parameters

[0043]

[0044] (2) Output the MSLB transient safety parameter as the actual value

[0045] (3) Setting the working condition data set in this embodiment

[0046] As shown in Table 2, this is the operating condition data table set in this embodiment, wherein the analysis program uses 1 s as the running step length to obtain a design operating condition data set with a time series length of 1000 s when the atmospheric temperature is 16°C, i.e., a small sample data set; at the same time, the analysis program uses 1 s as the running step length to obtain a test operating condition data set with a time series length of 10000 s when the atmospheric temperature is 10°C, i.e., a large sample data set.

[0047] Table 2 Working condition data set

[0048]

[0049] To facilitate identification and operation training, the data are placed in a set one by one to set different data sets. The data sets for design conditions and test conditions are set to verify the applicability of the model under different similar accident conditions. The size of the data set is artificially selected and should not be too small or too large. Too small will result in fewer training samples and incomplete training, while too large will increase the amount of calculation.

[0050] The data within the selected range are more representative, and the ten-fold difference in size between the small sample data set and the large sample data set can verify the method of training and predicting large sample data by small sample data sets, so the number of data sets selected in the experiment process of the present invention is 1000 and 10000.

[0051] (4) Perform linear normalization on the data

[0052] In order to improve the convergence speed of the time series deep learning model, the data is normalized uniformly. The formula is as follows:

[0053]

[0054] Among them, X norm is the normalized data, X is the original data, and X max , X min are the maximum and minimum values ​​in the data set.

[0055] (5) Segmentation of dataset feature-label pairs

[0056] like Figure 2 As shown, the dataset is segmented into feature-label pairs, with feature labels corresponding one to one. That is, each set of features is {x i ,x i+1 ,x i+2 ,……,x i+historysteps-1}, the corresponding label is {y i+historysteps},i∈[n0,n s-historysteps+1 ], where the subscript historysteps is the length of the historical time step.

[0057] The starting time point of the segmented data set is n0 and the ending time point is n s The label starting point n0′ has a time lag of the length of the historical time step, and the last feature segmentation window corresponds to n s ~n f To facilitate comparison with known data to understand the prediction effect, the features are selected until the second-to-last feature segmentation window.

[0058] (6) Divide the data set

[0059] The data sets are divided into two types according to their usage. The purpose is to distinguish the data used in the training and testing processes, and to allow the two data sets to be operated independently. In this way, the data used for training and prediction will not affect each other and cause inaccurate prediction accuracy.

[0060] Principles for dividing data sets:

[0061] 1) Datasets used for the training process: training set and validation set.

[0062] 2) Dataset used for testing: test set.

[0063] In order to reduce the possibility of overfitting, the data in the training set are shuffled, that is, there is no requirement for the time series of the data; the test set shows the prediction effect according to the time development progress, and the time series needs to remain unchanged, so the original order is maintained.

[0064] As shown in Table 3, there are 1000 sets of data on the temperature parameters in the containment in the single parameter model design condition data set. The data are linearly normalized. The training process of the model requires more data, so most of the data are allocated to the training process. Therefore, in the verification process of the present invention, the first 800 sets of data are used for the training process, and the remaining data are used for the test set.

[0065] For example, the first 800 groups of data are normalized and used for model training. The data is rolled and split with a historical time step window size of 40s, that is, the first group uses the first 40s of data as features, and the 41st second of data as the corresponding label, and so on. Finally, 760 groups of feature-label pairs are formed. The first 700 groups are used as training sets, and the last 60 groups are used as validation sets for model training. The training sets and validation sets are randomly sorted to randomize the data to avoid overfitting. The test set keeps the original data order without randomization, and its features are rolled and split into 160 groups in total.

[0066] As shown in Table 4, the design condition data set used for training the multi-parameter collaborative model includes 800 sets of data on the response of the pressure inside the containment, the liquid film coverage rate on the outer wall of the containment, and the liquid film evaporation amount on the outer wall of the containment over time. Currently, this model can only be used to predict the above three parameters. After the data preprocessing process, the distribution ratio of its training set and validation set is consistent with that of the single-parameter model. Finally, each feature-label pair contains 40 sets of features and 1 set of labels. Each set of data corresponds to the transient values ​​of the three safety parameters P, Ra, and Me at that time point. The multi-parameter collaborative model uses the entire test condition data set as the test set, which contains a total of 9960 sets of transient response values ​​of the above three safety parameters.

[0067] Table 3 Dataset division (single parameter model)

[0068]

[0069]

[0070] Table 4 Dataset division (multi-parameter collaborative model)

[0071]

[0072] (7) Setting the structure of the time-depth model to select the best hyperparameters

[0073] like Figure 3 The figure shows a schematic diagram of the LSTM time series deep learning model used in Example 1. The historical time step length is set to 40, that is, the memory unit is expanded by 40 steps along the time axis, x represents the input data, and y represents the output result. The cylinder represents the memory unit, and the number is the number of memory nodes contained in the memory unit. The single parameter model is above the time axis, and the multi-parameter collaborative model is below.

[0074] The input dimension of the single-parameter model training set is [760, 40, 1], and the input dimension of the multi-parameter collaborative model training set is [760, 40, 3].

[0075] The single-parameter model consists of two layers of LSTM. The first LSTM loop calculation layer memory unit is set to contain 80 memory nodes, and each time step pushes h t To the next layer; the memory unit of the second LSTM recurrent computation layer contains 100 memory nodes, and only the last time step pushes h t ; A Dropout layer is added after each LSTM layer to improve the generalization ability of the model and avoid overfitting due to less training data; finally, y is obtained through the fully connected layer t The parameter settings are shown in Table 5.

[0076] The multi-parameter model contains a layer of LSTM. The memory unit of the LSTM cyclic calculation layer is set to contain 128 memory nodes, and each time step pushes h t For the next layer, add a Dropout layer after LSTM to improve the generalization ability of the model and avoid overfitting due to less training data; finally, pass the fully connected layer to obtain y t The parameter settings are shown in Table 5.

[0077] Table 5 LSTM neural network parameters

[0078]

[0079] Hyperparameters include the length of the historical time step and the parameters in Table 5. Generally, the larger the length of the historical time step, the better the model learning effect, but it increases the training operation burden, time and the probability of overfitting, so it should be set reasonably. The experimental results of taking multiple values ​​every 5 or 10 as the span are comprehensive: 40 historical time steps are more appropriate.

[0080] In Table 5, the number of Dense nodes is equal to the number of parameters to be predicted by the model. In the research and application of neural network, the model with Dropout set to 0.2, Adam as the optimizer, and MSE as the loss function generally has good results. The number of LSTM nodes is generally rounded to multiples of 10 or 16; batch_size is a multiple of 16, which should not be too small or too large; epoch is rounded to 10, and the larger the value, the more sufficient the training, but too large a value may cause overfitting. The hyperparameters are set based on the above-mentioned empirical principles and the performance of multiple experiments.

[0081] (8) Call the training set to train the selected model

[0082] Put the training set divided in step 6 into the LSTM neural network in step 7 for training, and then use the validation set divided in step 6 for verification. If the loss error of the validation set gradually decreases and approaches 0, it preliminarily meets the accuracy requirements.

[0083] (9) Predict future time transient responses using the trained time series model

[0084] Send the test set divided in step 6 into the trained model to calculate and obtain the model prediction results.

[0085] (10) Comparative analysis of result accuracy

[0086] The trained LSTM operation obtains the predicted value and compares it with the actual value.

[0087] At the same time, the mean square error (MSE) is used to evaluate the prediction effect of the model. The smaller the value, the closer the predicted value is to the actual value.

[0088] (11) Determine whether the accuracy meets the requirements

[0089] In the prediction of transient parameters of the passive containment cooling system in the event of a main steam pipe rupture MSLB accident, the mean square error (MSE) is within 10 -2 If the predicted curve is consistent with the actual curve without too much deviation, the accuracy requirement is met. If it does not meet the requirements, proceed to step 12. If it meets the requirements, proceed to step 13.

[0090] (12) Re-divide the dataset or adjust the model structure and hyperparameters, and return to step 8.

[0091] (13) Save the model

[0092] Save the trained model for training and prediction based on actual needs.

[0093] The present invention has completed the prediction results display:

[0094] Table 6 Comparison of MSE prediction errors of multi-parameter collaborative model LSTM and RNN network prediction

[0095]

[0096] (1) About the multi-parameter collaborative model

[0097] As shown in Table 6, the MSE error between the predicted value of the containment pressure (p) using the LSTM model and the actual value is only 4.2×10 -5 For the prediction of liquid film coverage, whether the LSTM model or the RNN model is used, the prediction error of liquid film coverage (Ra) is the largest compared with the other two prediction parameters, but the maximum error between the predicted value and the actual value is less than 0.1. For the prediction of Ra time series model, the maximum MSE error is only 2.5×10- 2 According to the MSE results, the prediction effect of the RNN model is better than that of the LSTM model. At the same time, compared with the other two prediction parameters, the prediction error of the outer wall liquid film evaporation (Me) is the smallest, and the minimum MSE value is 4.6×10 -9 .

[0098] (2) About the single parameter model

[0099] In the design condition test set containing 160 sets of data, the transient response change of temperature in the containment obtained by the time series deep learning LSTM and RNN model (predicted value) is consistent with the transient response of temperature in the containment obtained by the complex containment two-phase flow thermal analysis model (actual value). The MSE error of the LSTM model prediction is 1.88×10 -2 , the MSE error predicted by the RNN model is 1.21×10 -1 , indicating that the prediction accuracy of the LSTM model based on time series deep learning is higher. It also shows that the single-parameter time series deep learning model established in this paper can accurately predict the transient response of the temperature inside the containment under MSLB accidents.

[0100] (3) Comparison between the LSTM model established by the construction method of Example 1 and the containment two-phase flow thermal analysis model Compare

[0101] like Figure 4 The predicted value of the pressure inside the containment by the multi-parameter collaborative model under the test conditions shown is highly consistent with the transient analysis result of the complex containment two-phase flow thermal analysis model, i.e. the actual value. The LSTM model predicts more accurately than the RNN model.

[0102] At the beginning of the accident, the temperature of the outer wall rose slowly and was close to the atmospheric temperature. Therefore, the evaporation of the liquid film on the outer wall of the containment was not large and fluctuated with the temperature of the outer wall. Then, the evaporation increased rapidly due to the rapid rise in the temperature of the outer wall. After that, the evaporation decreased as the coverage of the liquid film decreased. Figure 5 The figure shows the prediction results of the evaporation of the liquid film on the outer wall of the containment. The predicted values ​​of the time series deep learning model are in good agreement with the transient analysis results of the thermal analysis program, that is, the actual values. The fluctuation of the evaporation in the early stage of the accident and the turning point in the middle and late stages of the accident can be accurately predicted.

[0103] For MSLB accidents, the discharge of a large amount of high-energy steam causes the temperature of the containment to rise sharply after the accident. At this time, PCCS is put into operation. The PCCS system forms a water film on the surface of the steel containment to enhance the condensation heat exchange of water vapor in the containment. As the heat is discharged, the temperature and pressure in the containment gradually decrease and tend to stabilize. Figure 6 The figure shows the comparison between the prediction results and the actual values ​​of the RNN model and LSTM model established by the steps of the present invention. In the design condition test set containing 160 sets of data, the transient response change of the temperature in the containment obtained by the time series deep learning LSTM and RNN model, i.e. the predicted value, is in good agreement with the transient response of the temperature in the containment obtained by the complex containment two-phase flow thermal analysis model, i.e. the actual value. Figure 6 As shown in the figure, the predicted value of the single parameter time series deep learning model is compared with the actual value of the complex containment two-phase flow thermal analysis model. The MSE error predicted by the LSTM model is 1.88×10 -2 , the MSE error predicted by the RNN model is 1.21×10 -1 , indicating that the prediction accuracy of the LSTM model based on time series deep learning is higher. It also shows that the single-parameter time series deep learning model established in this paper can accurately predict the transient response of the temperature inside the containment under MSLB accidents.

[0104] Embodiment 2:

[0105] A containment key parameter rapid prediction system, comprising:

[0106] Input parameter acquisition module, where the input parameters include initial pressure in the containment, initial liquid film coverage, cooling water flow, wind speed, etc.;

[0107] A data processing module pre-processes the parameter data acquired by the input parameter acquisition module;

[0108] The preprocessing method is to normalize the data;

[0109] A prediction module, inputting the input parameter data processed by the data processing module into the LSTM fast prediction model to obtain a prediction value; the LSTM fast prediction model is the fast prediction model for key parameters of the containment constructed in Example 1;

[0110] The prediction result verification module compares the predicted value with the actual value, and uses the mean square error MSE to evaluate the prediction effect of the model. The result verification module compares the predicted value obtained by the prediction module in the system with the actual value obtained in advance by the traditional method to determine whether the prediction result of the system is accurate.

[0111] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a rapid prediction model for key parameters of a nuclear reactor containment, characterized in that: The following steps are involved: S1: Initialize containment parameters, construct a data set, and normalize, segment, and divide the data set; S2: constructing an LSTM fast prediction model, wherein the LSTM fast prediction model includes a single parameter model and a multi-parameter collaborative model, and using the single parameter model and the multi-parameter collaborative model to predict the transient state of the safety parameter; The single parameter model: first LSTM loop calculation layer-Dropout layer-second LSTM loop calculation layer-Dropout layer-fully connected layer; The multi-parameter collaborative model: the third LSTM loop calculation layer-Dropout layer-fully connected layer; The first LSTM loop calculation layer is set to contain 80 memory nodes, the second LSTM loop calculation layer is set to contain 100 memory nodes, and the third LSTM loop calculation layer is set to contain 128 memory nodes; Using the data set to train the LSTM fast prediction model to obtain an optimal LSTM fast prediction model; S3: Input the containment parameters into the optimal LSTM fast prediction model to predict the containment safety performance.

2. The method for constructing a rapid prediction model for key parameters of a nuclear reactor containment vessel according to claim 1, characterized in that: Partitioning of the dataset: According to the length of time series, it is divided into small sample data set and large sample data set.

3. The method for constructing a rapid prediction model for key parameters of a nuclear reactor containment according to claim 1, characterized in that: The data segmentation is the segmentation of data set feature-label pairs, and the data set features and labels are segmented in a one-to-one correspondence.

4. The method for constructing a rapid prediction model for key parameters of a nuclear reactor containment vessel according to claim 1, characterized in that In the construction of the LSTM fast prediction model, the LSTM fast prediction model is trained using a training set; the effect of the training process is observed using a validation set; and the performance of the LSTM fast prediction model is tested using a test set; When the test set is used to test the LSTM fast prediction model, the time series of the test set does not change.

5. The method for constructing a rapid prediction model for key parameters of a nuclear reactor containment vessel according to claim 4, characterized in that The training set, the validation set, and the test set in the single parameter model all use small sample data sets; The training set and the validation set in the multi-parameter collaborative model also use small sample data sets and the data in the multi-parameter collaborative model needs to be preprocessed. The distribution ratio of the training set and the validation set after the data preprocessing process is consistent with that of the single-parameter model; In the multi-parameter collaborative model, the entire large sample data set is used as the test set.

6. A nuclear reactor containment key parameter rapid prediction system, characterized in that include: Input parameter acquisition module, where the input parameters include initial pressure in the containment, initial liquid film coverage, cooling water flow, and wind speed; A data processing module pre-processes the parameter data acquired by the input parameter acquisition module; A prediction module, inputting the input parameter data processed by the data processing module into an LSTM fast prediction model to obtain a prediction value; the LSTM fast prediction model is obtained by the method for constructing a nuclear reactor containment key parameter fast prediction model according to any one of claims 1 to 5; The prediction result verification module compares the predicted value with the actual value, and uses the mean square error (MSE) to evaluate the prediction effect of the model.

Images