Accident sequence screening method based on FCNN and PSO

Through the combination of FCNN and PSO, deep learning databases and alternative models are quickly built, and accident sequences are screened using PSO algorithm, which solves the problem of large amount of calculations in accident sequence analysis of nuclear power plants and improves analysis efficiency and stability.

CN114970324BActive Publication Date: 2024-11-22SICHUAN KERUI EDDIE TECH CONSULTING CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210473115.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-29
Publication Date
2024-11-22
Estimated Expiration
2042-04-29

AI Technical Summary

Technical Problem

The existing safety evaluation methods for nuclear power plants are large in the analysis of accident sequences, resulting in inefficiency, especially when the operating status of nuclear power plants is slightly changed, it is difficult to respond quickly.

Method used

Using a combination of FCNN and PSO, a deep learning database is quickly constructed through concurrent computing, a deep learning alternative model is constructed, and accident analysis is performed instead of RELAP5, and the PSO algorithm is used to quickly capture the optimal solution of the accident sequence, and filter out the BEPU analysis that needs to be performed. sequence.

Benefits of technology

The efficiency of single accident analysis is significantly improved, the amount of calculation is reduced, the number of BEPU analysis sequences is reduced, and the analysis efficiency and program stability are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114970324B_ABST
    Figure CN114970324B_ABST
Patent Text Reader

Abstract

The present invention discloses an accident sequence screening method based on the combination of FCNN and PSO, which relates to the technical field of accident sequence screening, and includes the following steps: S101. Clarify the research object and target parameters, and complete deterministic and probabilistic modeling; S201. Use concurrent calculation methods, concurrently calculate RELAP5 programs, and quickly build a deep learning database; S301. Use FNCC analysis methods to build a deep learning alternative model to replace RELAP5 for accident analysis; S401. Use PSO methods to call alternative models for accident analysis, quickly capture the optimal solution of each accident sequence, and screen out sequences that need to be analyzed by BEPU. The present invention uses a fully connected neural network to construct an alternative model to replace the traditional system simulation program, improve the efficiency of single accident analysis, use a particle swarm optimization algorithm to optimize the constructed alternative model, and reduce the number of analysis and calculation times to achieve an efficient accident sequence screening method.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of accident sequence screening, and in particular to an accident sequence screening method based on the combination of FCNN and PSO. Background Art

[0002] Traditional safety assessment methods for nuclear power plants are deterministic analysis (DSA) and probabilistic analysis (PSA). However, both methods have certain limitations: the deterministic safety analysis method only performs safety analysis on specific sequences of representative design basis accidents with extremely low probability of occurrence, and the decision criteria are too conservative. The probabilistic method can well consider the random uncertainty in the accident process, but the calculation assumptions of its success criteria are very simplified and conservative, and its ability to handle small changes in the operating status of nuclear power plants is poor. Therefore, the risk-guided safety margin characteristic analysis method (RISMC) of DSA and PSA has become a research hotspot.

[0003] The existing RISMC method is to model all accident sequences using the PSA analysis method, and then use the best estimate plus uncertainty analysis method (BEPU) to analyze each accident sequence to obtain the final target parameter distribution, thereby judging the safety performance of the nuclear power plant. Figure 1 The figure shows the analysis process of the RISMC method for a small break loss of coolant accident in a nuclear power plant. In order to measure the integrity of the nuclear reactor core, the cladding peak temperature (PCT) is used as the target parameter.

[0004] Taking the best estimation program RELAP5 as an example, it takes about 6 minutes to perform a 5000s simulation calculation using an i7-9700K 3.6GHz CPU computer. In order to ensure the reliability of the uncertainty analysis results, nearly a thousand groups of calculations are required for a sequence using random sampling. The serial calculation analysis of a single sequence takes about 100 hours. When the number of accident sequences is large, the amount of calculation is too large.

[0005] Therefore, it is an urgent problem for technical personnel in this field to propose an accident sequence screening method based on the combination of FCNN and PSO to improve the efficiency of single accident analysis and reduce the amount of calculation. Summary of the invention

[0006] In view of this, the present invention provides an accident sequence screening method based on the combination of FCNN and PSO, which effectively improves the efficiency of single accident analysis and reduces the amount of calculation.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] The accident sequence screening method based on the combination of FCNN and PSO includes the following steps:

[0009] S101. Clarify the research object and target parameters, and complete deterministic and probabilistic modeling;

[0010] S201. Use concurrent computing methods and concurrent computing RELAP5 programs to quickly build a deep learning database;

[0011] S301. Use FNCC analysis method to build a deep learning alternative model to replace RELAP5 for accident analysis;

[0012] S401. Use the PSO method to call the alternative model for accident analysis, quickly capture the optimal solution for each accident sequence, and screen out the sequences that need BEPU analysis.

[0013] Optionally, S101 specifically includes the following steps:

[0014] S1011. Model the research object based on determinism;

[0015] S1012. Determine all accident sequences based on the probability model;

[0016] S1013. Determine uncertainty parameters, distribution and target parameters;

[0017] S1014. Conduct parameter sensitivity analysis and screen key parameters.

[0018] Optionally, S1011. Modeling the research object according to determinism specifically includes the following steps:

[0019] S10111. Clarify the analysis object and the accident;

[0020] S10112. Obtain all parameter information required in the modeling process;

[0021] S10113. After completing the object node diagram according to the key parameters, write the input card.

[0022] S10114. After modeling is completed, the simulation results need to be compared and analyzed with the design parameters to ensure that the modeling accuracy meets the analysis requirements.

[0023] Optionally, S201 specifically includes the following steps:

[0024] S2011. Concurrent initialization: Before concurrent computing, you need to initialize and set input parameters, input parameter range, distribution, and target parameters. Clarify the input and output of database construction;

[0025] S2012. Computer equipment performance test, complete optimization program initialization settings;

[0026] S2013.Multithreading to implement batch RELAP5 input file update;

[0027] S2014. Multithreaded RELAP5 calculations;

[0028] S2015. Build input and output databases.

[0029] Optionally, S301 specifically includes the following steps:

[0030] S3011. Construct FCNN deep learning alternative model;

[0031] S3012. Determine whether a FCNN deep learning alternative model that meets the accuracy requirements can be generated while the current database sample size remains unchanged; if so, the model is built and packaged, and then enters S401 to participate in concurrent computing; otherwise, returns to S201 to increase the number of learning database samples.

[0032] Optionally, S3011. constructing the FCNN deep learning alternative model specifically includes the following steps:

[0033] S30111. Calling the learning database: importing the generated database samples into this step for model building;

[0034] S30112. Initialization of input layer learning sample data and test sample data: Update the nominal values ​​of the input layer and output layer of FCNN according to the input parameters and output parameters in the database. And initialize the activation function and key information of FCNN;

[0035] S30113. Hidden layer fitting: The data of the input layer enters the hidden layer and undergoes nonlinear fitting through the activation function; when there are multiple hidden layers, the output of the previous hidden layer is used as the input of the next hidden layer for data transmission;

[0036] S30114. Output layer data output: After an input layer initial parameter is fitted by the hidden layer, a fitting target parameter will be output;

[0037] S30115. Determine whether the error between the fitting output data and the nominal value meets the requirements: compare the fitting target parameters of the output layer with the standard target parameters calculated by the RELAP5 program in the database, and enter S30116 when the error meets the specified value; if the error does not meet the requirements, the program automatically adjusts the parameters through the Adam algorithm and continues to enter S30113 until it meets the specified value and enters S30116;

[0038] S30116. Determine whether the accuracy of the test sample meets the requirements: import the input data of the test sample into the model, compare the obtained output value with the standard output value in the test sample, and proceed to the next step if it meets the requirements; if it does not meet the requirements, continue to adjust the parameters of the Adam algorithm until the accuracy requirements are met.

[0039] Optionally, S401 specifically includes the following steps:

[0040] S4011. Calculate the nominal input values ​​of all accident sequences in S101 using the RELAP5 program to obtain the nominal target parameters of each sequence;

[0041] S4012. Classify the accident sequence and determine the target parameter solution requirements of the accident sequence;

[0042] S4013. Use the POS algorithm to calculate the optimal solution for the accident sequence.

[0043] Optionally, S4013. Calculating the optimal solution of the accident sequence using the POS algorithm specifically includes the following steps:

[0044] S40131. Initialize particle parameters: according to the calculation requirements, set the number of particles per generation, particle dimensions, total number of iterations, inertia weight factor, learning factor, key parameters, and control the scale and efficiency of PSO optimization;

[0045] S40132. Determine input parameters and target parameters, and complete program initialization settings: According to the results of the sensitivity analysis obtained in S101, update the input parameter information and target parameter information, and initialize;

[0046] S40133. Calling the deep learning alternative model and concurrent calculation: calling the deep learning alternative model encapsulated in S301 to replace RELAP5 for accident analysis calculation;

[0047] S40134. Obtain target parameters and update the best fitness value: Output the target parameters after calculation through the deep learning alternative model, and compare and update the optimal solutions of each generation of target parameters;

[0048] S40135. Determine whether the convergence conditions are met: When any convergence condition is met, the program completes the calculation and enters S40136; if not, the initial parameters of each accident analysis sequence are updated according to the PSO algorithm and the iterative calculation continues to enter S40133;

[0049] S40136. Output optimal solution data;

[0050] S40137. Output sequence screening results: Based on the comparison between the optimal solution data and the safety limit and the judgment and analysis of S40135, the accident sequence numbers that require BEPU analysis will be output.

[0051] Optionally, the PSO calculation convergence conditions include: (1) reaching the maximum number of iterations; (2) the difference between the optimal solutions of multiple consecutive generations is within the error range; (3) a special value appears in the calculated target parameter, where the special value refers to the relationship between the calculated target parameter and the safety limit that is different from the relationship between the nominal target parameter and the safety limit of the sequence.

[0052] It can be seen from the above technical solutions that, compared with the prior art, the present invention provides an accident sequence screening method based on the combination of FCNN and PSO: concurrent calculation of RELAP5 is realized to quickly build a deep learning database; FCNN is used to build a deep learning alternative model with sufficient accuracy to replace RELAP5 for accident analysis, thereby improving the calculation efficiency of a single accident analysis case and the program stability when the PSO analysis is called; the PSO method quickly captures the optimal solution of the target parameters of the accident sequence, completes the accident sequence screening, and reduces the number of BEPU analysis sequences. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] 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.

[0054] Figure 1 RISMC analysis process for small break loss of coolant accidents in nuclear power plants;

[0055] Figure 2 This is a flow chart of the accident sequence screening method based on the combination of FCNN and PSO of the present invention;

[0056] Figure 3 This is a specific flow chart of the accident sequence screening method based on the combination of FCNN and PSO of the present invention;

[0057] Figure 4 Rapidly construct a deep learning database flow chart for concurrent computing of the present invention;

[0058] Figure 5 This is the schematic diagram of FCNN;

[0059] Figure 6 Construct a flow chart for the FCNN deep learning alternative model of the present invention;

[0060] Figure 7 The method flow of the PSO method of the present invention for finding the optimal solution of the accident sequence;

[0061] Figure 8This is a schematic diagram of the data iteration principle of the particle swarm algorithm of the present invention. DETAILED DESCRIPTION

[0062] 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.

[0063] Reference Figure 1 The figure shows the analysis process of the RISMC method for small break loss of coolant accidents in nuclear power plants. In order to measure the integrity of the nuclear reactor core, the peak cladding temperature (PCT) is used as the target parameter. Taking the best estimation program RELAP5 as an example, it takes about 6 minutes to perform a 5000s simulation calculation using an i7-9700K 3.6GHz CPU computer. In order to ensure the reliability of the uncertainty analysis results, nearly a thousand groups of calculations are required for a sequence using random sampling. The serial calculation and analysis of a single sequence requires about 100 hours. When the number of accident sequences is large, the amount of calculation is too large.

[0064] Based on this, this paper proposes a nuclear reactor accident sequence screening method based on a fully connected neural network combined with a particle swarm optimization algorithm (FCNN-PSO). Figure 2 shown.

[0065] In order to solve the problem of low efficiency in single accident analysis, this method uses a substitute model to replace RELAP5 for accident calculation. The FCNN method is an effective method for constructing a substitute model because it is easy to implement and has strong nonlinear fitting ability. The deep learning database is quickly trained by concurrently calculating the RELAP5 calculation program. The FCNN method is used to construct a substitute model based on the database to replace the RELAP5 program for accident analysis, reducing the single-time time of several minutes to a few seconds, effectively improving the efficiency of single accident analysis. When performing BEPU analysis on all accident sequences, the number of case calculations is large. This method uses an optimization algorithm to quickly capture the optimal solution of each accident sequence, classify the accident sequences, and screen out the sequences that need to be analyzed by BEPU. Since the PSO method is easy to implement, has high accuracy, fast convergence and is easy to parallelize, the PSO method is selected as the optimization algorithm used for screening.

[0066] Reference Figure 2 As shown, the present invention discloses an accident sequence screening method based on the combination of FCNN and PSO, comprising the following steps:

[0067] S101. Clarify the research object and target parameters, and complete deterministic and probabilistic modeling;

[0068] S201. Use concurrent computing methods and concurrent computing RELAP5 programs to quickly build a deep learning database;

[0069] S301. Use FNCC analysis method to build a deep learning alternative model to replace RELAP5 for accident analysis;

[0070] S401. Use the PSO method to call the alternative model for accident analysis, quickly capture the optimal solution for each accident sequence, and screen out the sequences that need BEPU analysis.

[0071] Reference Figure 3 As shown in FIG. 1 , a specific flow chart of the accident sequence screening method based on the combination of FCNN and PSO is shown.

[0072] Furthermore, S101 specifically includes the following steps:

[0073] S1011. Model the research object based on determinism;

[0074] S1012. Determine all accident sequences based on the probability model;

[0075] Among them, an event tree model of corresponding accuracy is established based on the power plant and accident information that has been clarified in the process of deterministic modeling to clarify each accident sequence. The more sophisticated the deterministic modeling is, the more systems and equipment are involved, and the more complex the event tree is. The event tree model must match the deterministic model;

[0076] S1013. Determine uncertainty parameters, distribution and target parameters;

[0077] Among them, it is necessary to introduce uncertainty in the input parameters during the BEPU analysis process, so it is necessary to clarify the specific uncertainty parameters, distribution and range before the analysis. The determination of uncertainty parameters and their distribution is mainly based on the design data of the nuclear power plant, the phenomenon identification and classification table (PIRT) and experience. Since each nuclear power plant has a certain specificity, there will be a certain deviation in the selection of uncertainty parameters, so a larger number of uncertainty parameters should be selected for calculation initially. The target parameter is the key parameter for judging whether the core is damaged under the accident condition. Taking the small break loss of coolant accident as an example, PCT is the target parameter.

[0078] S1014. Perform parameter sensitivity analysis and screen key parameters;

[0079] Among them, S1013 determines a large number of uncertainty parameters, which will lead to too high problem dimensions, low calculation accuracy, and slow convergence speed in the deep learning and optimization algorithm calculation process. Therefore, it is necessary to conduct sensitivity analysis before analysis to screen out important uncertainty parameters, reduce the number of input uncertainty parameters, and reduce the difficulty of subsequent analysis.

[0080] Furthermore, S1011. Modeling the research object according to determinism specifically includes the following steps:

[0081] S10111. Clarify the analysis object and the accident;

[0082] Different nuclear power plants have significantly different structures and characteristics, so specific modeling is required for the power plants. After determining the modeling object, the accident to be analyzed needs to be determined. Different accidents involve different key phenomena and processes, and the accuracy requirements for modeling in various systems and equipment are significantly different;

[0083] S10112. Obtain all parameter information required in the modeling process;

[0084] Among them, it is determined that sufficient nuclear power plant parameter information is needed to study power plants and accidents. By investigating nuclear power plant design documents, safety analysis reports and other materials, all parameter information required in the nuclear power plant modeling process is determined. For example, the power of the core, the structural parameters of the pipeline, the core inlet and outlet coolant temperature and flow, etc.;

[0085] S10113. After completing the object node diagram according to key parameters, write the input card;

[0086] Taking the RELAP5 program as an example, the node diagram of the nuclear power plant is first completed according to the key parameters. On this basis, the input card is written;

[0087] S10114. After modeling is completed, the simulation results need to be compared and analyzed with the design parameters to ensure that the modeling accuracy meets the analysis requirements.

[0088] Among them, the uncertainty parameters, range, distribution and target parameters are determined according to the deterministic model. On this basis, the multi-threading module and sampling module of the Python program are used to batch generate RELAP5 input cards.

[0089] Furthermore, S201 specifically includes the following steps:

[0090] S2011. Concurrency initialization: before concurrent computing, you need to initialize and set input parameters, input parameter range, distribution, and target parameters; clarify the input and output of database construction;

[0091] S2012. Computer equipment performance test, complete optimization program initialization settings;

[0092] In order to ensure that multi-threaded data do not affect each other during concurrent calculations, thread locks need to be set, and a certain time interval needs to be set to prevent simultaneous reading of physical property parameter tables when multiple RELAP5s are started. Therefore, it is necessary to test the computer's response performance and set a suitable delay time;

[0093] S2013.Multithreading to implement batch RELAP5 input file update;

[0094] Among them, the traditional RELAP5 is a serial calculation on the computer, which does not make full use of computing resources and is inefficient. In order to improve the analysis efficiency, the present invention adopts Python program to develop concurrent computing functions to replace serial calculations. Before performing RELAP5 calculations, it is necessary to use multi-threading to batch generate new input cards. Call n threads in a computer process, and each thread performs parameter sampling, input card modification, and update tasks for an input card. When calling the IO_function function in Python for multi-threaded calculations, the modification of global variables will generate multiple duplicate input cards due to multi-threaded data sharing. Therefore, it is necessary to set thread locks between each thread to ensure that only the current input card is updated during the data card update process. After the input card update is completed, the new input card is named in the form of a timestamp plus a thread number to ensure that the input card naming will not be repeated;

[0095] S2014. Multithreaded RELAP5 calculation:

[0096] Only one RELAP5 program can be started in a thread at a time for calculation, so the multi-process module in Python is called together with the multi-threading module to start a RELAP5 program in each thread of each process for calculation. When RELAP5 starts calculation, it will read the water property table, and when multiple processes read a water property table at the same time, an error will occur, so a small time interval time.sleep() is set between each two RELAP5 startups to ensure that multiple processes do not read the water property table at the same time;

[0097] S2015. Build input and output databases;

[0098] After RELAP5 calculation is completed, complete o and r files will be generated. Since each file name is unique, multiple threads in multiple processes work simultaneously to read the target parameters in the o file according to each file name. The o and r files generated after RELAP5 calculation are large. In order to effectively save memory, the o and r files are deleted immediately after each data reading. After all accident sequence calculations are completed, a database with one-to-one correspondence between input parameters and output parameters is generated.

[0099] The traditional RELAP5 program can only perform serial calculations in the computer. In order to improve the computing efficiency and more efficiently generate the database required for deep learning, this method uses concurrent computing. Taking the i7-9700K 3.6GHz CPU 8-core computer as an example, when performing single-threaded RELAP5 serial calculations, the CPU utilization rate is about 13%, and it takes about 6 minutes to calculate a single 5000s case. Using 8 processes to calculate 8 5000s cases at the same time, the CPU utilization rate is about 100%, and it takes about 9 minutes in total. Multi-process concurrent computing can make full use of CPU computing resources and improve computing efficiency.

[0100] First, the uncertainty parameters, range, distribution and target parameters are determined according to the deterministic model. On this basis, the multi-threading module and sampling module of the Python program are used to batch generate RELAP5 input cards. The multi-threading and multi-process modules are used to control multiple RELAP5 concurrent calculations, complete the key parameter extraction, and generate a deep learning database. The specific implementation process is as follows Figure 4 shown.

[0101] Furthermore, S301 specifically includes the following steps:

[0102] S3011. Construct FCNN deep learning alternative model;

[0103] S3012. Determine whether a FCNN deep learning alternative model that meets the accuracy requirements can be generated while the current database sample size remains unchanged; if so, the model is built and packaged, and then enters S401 to participate in concurrent computing; otherwise, returns to S201 to increase the number of learning database samples.

[0104] Among them, the existing RISMC analysis uses system programs for accident case analysis, and the single calculation time is long. Taking the i7-9700K 3.6GHz CPU 8-core computer as an example, it takes about 6 minutes to calculate a single 5000s case with the RELAP5 program. In the process of finding the optimal solution of the PSO algorithm, multiple particles and multiple generations of calculations are required to obtain convergence results. When the number of calculation cases is large, the overall calculation time is too long. And multiple PSO algorithm iterations of RELAP5 calculations have a great impact on the stability of the program. Therefore, in order to improve the calculation efficiency of a single accident analysis case and optimize the stability of the program, it is necessary to construct an alternative model to replace the RELAP5 program for accident analysis. However, due to the drastic data fluctuations during the accident analysis process, it is difficult to accurately describe the accident process using a simple fitting formula. Therefore, the FCNN method, which is easy to implement and has strong nonlinear fitting ability, is used to construct an alternative model.

[0105] The schematic diagram of FCNN is as follows Figure 5 As shown. FCNN is mainly composed of three parts: input layer, hidden layer and output layer. After the deep learning database of the present invention is imported, the parameter information input in the database updates the FCNN input layer information, and the corresponding output parameters in the database are used as the nominal values ​​in the output layer to compare the fitting accuracy. After the input layer data is imported, the data is passed to the hidden layer. The activation function in the hidden layer ensures that the hidden layer can fit the nonlinear data. The output data of hidden layer 1 after fitting is used as the input of hidden layer 2, and the data is passed in sequence to the output layer. The fitting target parameters output by the output layer are compared with the nominal values ​​imported from the database. If the error accuracy requirements are not met, the learning rate is automatically adjusted, and the data is put into the hidden layer again for fitting, and it is iterated until the accuracy requirements are met.

[0106] Further, see Figure 6 As shown, S3011. Constructing the FCNN deep learning alternative model specifically includes the following steps:

[0107] S30111. Calling the learning database: importing the generated database samples into this step for model building;

[0108] S30112. Initialization of input layer learning sample data and test sample data: Update the nominal values ​​of the input layer and output layer of FCNN according to the input parameters and output parameters in the database. And initialize the activation function and key information of FCNN;

[0109] S30113. Hidden layer fitting: The data of the input layer enters the hidden layer and undergoes nonlinear fitting through the activation function; when there are multiple hidden layers, the output of the previous hidden layer is used as the input of the next hidden layer for data transmission;

[0110] S30114. Output layer data output: After an input layer initial parameter is fitted by the hidden layer, a fitting target parameter will be output;

[0111] S30115. Determine whether the error between the fitting output data and the nominal value meets the requirements: compare the fitting target parameters of the output layer with the standard target parameters calculated by the RELAP5 program in the database, and enter S30116 when the error meets the specified value; if the error does not meet the requirements, the program automatically adjusts the parameters through the Adam algorithm and continues to enter S30113 until it meets the specified value and enters S30116;

[0112] S30116. Determine whether the accuracy of the test sample meets the requirements: import the input data of the test sample into the model, compare the obtained output value with the standard output value in the test sample, and proceed to the next step if it meets the requirements; if it does not meet the requirements, continue to adjust the parameters of the Adam algorithm until the accuracy requirements are met.

[0113] Furthermore, S401 specifically includes the following steps:

[0114] S4011. Calculate the nominal input values ​​of all accident sequences in S101 using the RELAP5 program to obtain the nominal target parameters of each sequence;

[0115] S4012. Classify the accident sequence and determine the target parameter solution requirements of the accident sequence;

[0116] S4013. Use the POS algorithm to calculate the optimal solution for the accident sequence.

[0117] Further, see Figure 7 As shown, S4013. Calculating the optimal solution of the accident sequence using the POS algorithm specifically includes the following steps:

[0118] S40131. Initialize particle parameters: according to the calculation requirements, set the number of particles per generation, particle dimensions, total number of iterations, inertia weight factor, learning factor, key parameters, and control the scale and efficiency of PSO optimization;

[0119] S40132. Determine input parameters and target parameters, and complete program initialization settings: According to the results of the sensitivity analysis obtained in S101, update the input parameter information and target parameter information, and initialize;

[0120] S40133. Calling the deep learning alternative model and concurrent calculation: calling the deep learning alternative model encapsulated in S301 to replace RELAP5 for accident analysis calculation;

[0121] S40134. Obtain target parameters and update the best fitness value: Output the target parameters after calculation through the deep learning alternative model, and compare and update the optimal solutions of each generation of target parameters;

[0122] S40135. Determine whether the convergence conditions are met: When any convergence condition is met, the program completes the calculation and enters S40136; if not, the initial parameters of each accident analysis sequence are updated according to the PSO algorithm and the iterative calculation continues to enter S40133;

[0123] S40136. Output optimal solution data;

[0124] S40137. Output sequence screening results: Based on the comparison between the optimal solution data and the safety limit and the judgment and analysis of S40135, the accident sequence numbers that require BEPU analysis will be output.

[0125] Furthermore, the PSO calculation convergence conditions include: (1) reaching the maximum number of iterations; (2) the difference between the optimal solutions of multiple consecutive generations is within the error range; (3) outliers appear in the calculated target parameters, where the outliers refer to the relationship between the calculated target parameters and the safety limit that is different from the relationship between the nominal target parameters and the safety limit of the sequence.

[0126] Among them, reference Figure 8 As shown, it is a principle diagram of data iteration of the particle swarm algorithm, where A1,1 represents the first particle of the first generation, and An,m represents the mth particle of the nth generation. Take solving the maximum value of the target parameter of the accident sequence as an example. In this invention, each particle represents a target parameter calculated by an alternative model. First, all the results of the first generation are calculated according to the initialized input parameters, and the maximum value W1 is screened out from all the results. Then, a new set of input parameters is updated according to the PSO algorithm, the second generation target parameters are calculated and the maximum value W2 is screened out. If W2 is greater than W1, the optimal solution on the far right is updated to W2. If W2 is less than W1, the optimal solution on the far right is maintained as W1. And so on, iterate and update the optimal solution generation by generation. When the difference between the optimal solutions of 10 consecutive generations meets the accuracy requirements, the calculation is terminated and the optimal solution is output.

[0127] The above description of the disclosed embodiments is provided in a progressive manner to enable 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. The accident sequence screening method based on the combination of FCNN and PSO is characterized by: The following steps are involved: S101. Clarify the research object and target parameters, and complete deterministic and probabilistic modeling; S201. Use concurrent computing methods and concurrent computing RELAP5 programs to quickly build a deep learning database; S301. Use FNCC analysis method to build a deep learning alternative model to replace RELAP5 for accident analysis; S401. Use the PSO method to call the alternative model for accident analysis, quickly capture the optimal solution of each accident sequence, and screen out the sequences that need BEPU analysis; S301 specifically includes the following steps: S3011. Construct FCNN deep learning alternative model; S3012. Determine whether a FCNN deep learning alternative model that meets the accuracy requirements can be generated when the current database sample size remains unchanged; if so, the model is built and packaged, and then enters S401 to participate in concurrent computing; otherwise, return to S201 to increase the number of learning database samples; S3011. Constructing the FCNN deep learning alternative model specifically includes the following steps: S30111. Calling the learning database: importing the generated database samples into this step for model building; S30112. Initialization of input layer learning sample data and test sample data: Update the nominal values ​​of the input layer and output layer of FCNN according to the input parameters and output parameters in the database; and initialize the activation function and key information of FCNN; S30113. Hidden layer fitting: The data of the input layer enters the hidden layer and undergoes nonlinear fitting through the activation function; when there are multiple hidden layers, the output of the previous hidden layer is used as the input of the next hidden layer for data transmission; S30114. Output layer data output: After an input layer initial parameter is fitted by the hidden layer, a fitting target parameter will be output; S30115. Determine whether the error between the fitting output data and the nominal value meets the requirements: compare the fitting target parameters of the output layer with the standard target parameters calculated by the RELAP5 program in the database, and enter S30116 when the error meets the specified value; if the error does not meet the requirements, the program automatically adjusts the parameters through the Adam algorithm and continues to enter S30113 until it meets the specified value and enters S30116; S30116. Determine whether the accuracy of the test sample meets the requirements: import the input data of the test sample into the model, compare the obtained output value with the standard output value in the test sample, and proceed to the next step if it meets the requirements; if it does not meet the requirements, continue to adjust the parameters of the Adam algorithm until the accuracy requirements are met.

2. The accident sequence screening method based on the combination of FCNN and PSO according to claim 1 is characterized in that: S101 specifically includes the following steps: S1011. Model the research object based on determinism; S1012. Determine all accident sequences based on the probability model; S1013. Determine uncertainty parameters, distribution and target parameters; S1014. Conduct parameter sensitivity analysis and screen key parameters.

3. The accident sequence screening method based on the combination of FCNN and PSO according to claim 2 is characterized in that: S1011. Modeling the research object based on determinism specifically includes the following steps: S10111. Clarify the analysis object and the accident; S10112. Obtain all parameter information required in the modeling process; S10113. After completing the object node diagram according to the key parameters, write the input card. S10114. After modeling is completed, the simulation results need to be compared and analyzed with the design parameters to ensure that the modeling accuracy meets the analysis requirements.

4. The accident sequence screening method based on the combination of FCNN and PSO according to claim 1 is characterized in that: S201 specifically includes the following steps: S2011. Concurrency initialization: before concurrent computing, you need to initialize and set input parameters, input parameter range, distribution, and target parameters; clarify the input and output of database construction; S2012. Computer equipment performance test, complete optimization program initialization settings; S2013.Multithreading to implement batch RELAP5 input file update; S2014. Multithreaded RELAP5 calculations; S2015. Build input and output databases.

5. The accident sequence screening method based on the combination of FCNN and PSO according to claim 1 is characterized in that: S401 specifically includes the following steps: S4011. Calculate the nominal input values ​​of all accident sequences in S101 using the RELAP5 program to obtain the nominal target parameters of each sequence; S4012. Classify the accident sequence and determine the target parameter solution requirements of the accident sequence; S4013. Use the POS algorithm to calculate the optimal solution for the accident sequence.

6. The accident sequence screening method based on the combination of FCNN and PSO according to claim 5 is characterized in that: S4013. Calculating the optimal solution of the accident sequence using the POS algorithm specifically includes the following steps: S40131. Initialize particle parameters: according to the calculation requirements, set the number of particles per generation, particle dimensions, total number of iterations, inertia weight factor, learning factor, key parameters, and control the scale and efficiency of PSO optimization; S40132. Determine input parameters and target parameters, and complete program initialization settings: According to the results of the sensitivity analysis obtained in S101, update the input parameter information and target parameter information, and initialize; S40133. Calling the deep learning alternative model and concurrent calculation: calling the deep learning alternative model encapsulated in S301 to replace RELAP5 for accident analysis calculation; S40134. Obtain target parameters and update the best fitness value: Output the target parameters after calculation through the deep learning alternative model, and compare and update the optimal solutions of each generation of target parameters; S40135. Determine whether the convergence conditions are met: When any convergence condition is met, the program completes the calculation and enters S40136; if not, the initial parameters of each accident analysis sequence are updated according to the PSO algorithm and the iterative calculation continues to enter S40133; S40136. Output optimal solution data; S40137. Output sequence screening results: Based on the comparison between the optimal solution data and the safety limit and the judgment and analysis of S40135, the accident sequence numbers that require BEPU analysis will be output.

7. The accident sequence screening method based on the combination of FCNN and PSO according to claim 6 is characterized in that: The PSO calculation convergence conditions include: (1) reaching the maximum number of iterations; (2) the difference between the optimal solutions of multiple consecutive generations is within the error range; (3) special values ​​appear in the calculated target parameters, where the special value refers to the relationship between the calculated target parameters and the safety limit that is different from the relationship between the nominal target parameters and the safety limit of the sequence.