Nuclear power control system safety risk prediction method based on Transform model

Abstract

The invention discloses a nuclear power control system safety risk prediction method based on a Transform model, which is characterized in that a nuclear power control system safety risk prediction model is constructed by adopting a nuclear power control system risk feature selection and small probability feature value generation algorithm, and intelligent safety risk prediction of a nuclear power control system is realized. The method specifically comprises the steps of performing preliminary selection on nuclear power system characteristics, performing a small probability characteristic value generation algorithm, constructing a nuclear power control system intelligent safety risk prediction model based on Transform and the like. Compared with the prior art, the method has the advantages that safe operation of the nuclear power reactor is ensured, various parameters of the reactor are monitored to be in a safe range, actions are automatically taken when abnormal conditions are detected to prevent accidents, and the method is convenient, simple, good in monitoring effect and high in safety. The method has important significance and a good application prospect for safety management of nuclear power.

Description

Technical Field

[0001] The invention relates to the technical field of risk prediction, and in particular to a method for predicting safety risks of a nuclear power control system based on a Transformer model. Background Art

[0002] The nuclear power reactor control system is a vital component of a nuclear power plant. Its main function is to monitor the operating status of the reactor and ensure that safety measures can be automatically triggered when abnormal conditions occur to protect personnel, equipment and the environment from the threat of radioactive substances. Real-time monitoring of key parameters of the reactor, such as neutron flux, temperature, pressure, etc. Take measures to prevent potential dangers before the parameters exceed the safe range. When an emergency is detected, the reactor is automatically triggered to shut down to avoid accidents.

[0003] Statistical model checking is a simulation-based model verification technology that combines fast simulation and statistical analysis techniques of the system to effectively improve the efficiency of model checking. Small probability eigenvalues ​​play an important role in the safety risk prediction of nuclear power control systems. Small probability eigenvalues ​​are usually associated with rare events that may seriously affect system safety. For example, the probability of certain equipment failures or abnormal conditions occurring is very small, but once they occur, they may lead to serious safety accidents. When conducting safety risk analysis, considering small probability events helps ensure the integrity of the system analysis and ensures that all possible risk factors are taken into account, thereby improving the overall safety of the system.

[0004] The Transformer model is a deep learning architecture proposed by the Google team in 2017 and has performed well in natural language processing and other sequence-to-sequence tasks. The core of the Transformer model is the self-attention mechanism, which allows the model to consider all positions in the sequence simultaneously when processing sequence data, unlike traditional recurrent neural networks or convolutional neural networks, which process sequence data step by step.

[0005] The models used in current nuclear power reactor control systems are mostly deterministic models or recurrent neural network models. In the rapidly changing nuclear power plant environment, existing models may not be able to update and respond to new safety risks in real time. Although rare events have a low probability of occurrence, they may lead to serious consequences once they occur. Existing models fail to fully consider these low-probability but high-risk events and have limitations in prediction accuracy, especially when predicting rare events or complex system behaviors. Summary of the invention

[0006] The purpose of the present invention is to propose a method for safety risk prediction of a nuclear power control system based on a Transformer model in response to the deficiencies of the prior art. The method adopts a risk feature selection method for a nuclear power control system, uses a small probability eigenvalue generation algorithm based on statistical model testing, constructs a safety risk prediction model for a nuclear power control system, realizes intelligent safety risk prediction of the nuclear power control system, ensures the safe operation of the reactor, monitors various parameters of the reactor to ensure that they are within a safe range, and automatically takes action when an abnormal situation is detected to prevent accidents. The method is convenient and simple, has a good monitoring effect, and is of great significance and good application prospects for the safety management of nuclear power.

[0007] The specific technical scheme for achieving the purpose of the present invention is: a method for safety risk prediction of nuclear power control system based on Transformer model, characterized by adopting risk feature selection of nuclear power control system and small probability feature value generation algorithm based on statistical model verification, constructing safety risk prediction model of nuclear power control system, and realizing intelligent safety risk prediction of nuclear power control system, which specifically includes the following steps:

[0008] Step 1: Construct a risk feature selection method for nuclear power reactor safety control system and make a preliminary selection of features;

[0009] Step 2: Then, for the small probability feature items, a small probability feature value selection method based on statistical model detection is proposed;

[0010] Step 3: Construct an intelligent safety risk prediction model for nuclear power control systems based on the Transformer model.

[0011] The step 1 constructs a risk feature selection method for a nuclear power control system, and the preliminary selection of features further includes: the nuclear power control system ensures the safe operation of the reactor by monitoring a variety of key parameters, and combines the existing reactor protection parameter table to preliminarily select features for safety risk prediction analysis.

[0012] The second step proposes a method for selecting small probability feature values ​​based on statistical model detection for small probability feature items, which specifically includes the following steps:

[0013] Step B1: Model the nuclear power control system and describe the attributes of safety requirements.

[0014] Step B2: Use statistical model detection to calculate the required small probability feature values.

[0015] Step B3: Use importance sampling to optimize the original statistical model detection method.

[0016] The step 3 proposes a nuclear power control system safety risk prediction model based on the Transformer model, which specifically includes the following steps:

[0017] Step C1: Vector representation of feature set

[0018] The initially selected risk features and small probability features are constructed into a feature set and converted into a series of vector representations, which are positionally encoded and input into the encoder, and the nuclear power control system status is input into the decoder.

[0019] Step C2: Design training sample dataset

[0020] The sample data set is randomly divided into three parts, 70% of which are training samples; 15% are validation samples to verify the generalization ability of the network; and 15% are test samples to detect the model's judgment performance for unknown classification sample sets.

[0021] Step C3: Validation test sample set

[0022] The test sample set is input into the trained security risk prediction model for prediction and classification, the classification results are obtained, and the prediction performance of the model is evaluated.

[0023] Compared with the prior art, the present invention ensures the safe operation of nuclear power reactors by monitoring various parameters of the reactor to ensure that they are within a safe range and automatically takes action when an abnormal situation is detected to prevent accidents. The present invention is convenient and simple, has good monitoring effect, and is of great significance and good application prospect for the safety management of nuclear power. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of the present invention;

[0025] Figure 2 This is the Transformer model structure diagram. DETAILED DESCRIPTION

[0026] The present invention is further described in detail with reference to the following specific examples and drawings. The process, conditions, experimental methods, etc. for implementing the present invention, except for the contents specifically mentioned below, are all common knowledge and common common sense in the art and are not particularly limited by the present invention.

[0027] Example 1

[0028] See also Figure 1 , the present invention specifically comprises the following steps:

[0029] Step S1: construct a nuclear power control system risk feature selection method and perform preliminary feature selection;

[0030] Step S2: For the small probability feature items, a small probability feature value generation algorithm based on statistical model testing is proposed;

[0031] Step S3: A nuclear power control system safety risk prediction model based on the Transformer model is proposed.

[0032] The step S1 constructs a risk feature selection method for a nuclear power control system, and the preliminary selection of features further includes: the nuclear power control system ensures the safe operation of the reactor by monitoring a variety of key parameters, and combines the existing reactor protection parameter table to make a preliminary selection of features for safety risk prediction analysis. The nuclear power reactor safety control system is a key component of a nuclear power plant, and its main function is to ensure the safe operation of the reactor. The system monitors various parameters of the reactor to ensure that they are within a safe range, and automatically takes action when an abnormal situation is detected to prevent accidents. The state of the nuclear power reactor system can be defined by the following formula (II):

[0033]

[0034] Assume that the sample data set is in the form of (F, y), where F is the feature vector of the safety risk prediction task, and y is the state of the nuclear power reactor system under the current state parameters, i.e., C1, C2. Therefore, the problem is transformed into finding a function f that can establish a mapping relationship between the risk prediction feature F and the nuclear power reactor system state y, and is expressed by the following formula (III):

[0035] f: F→y(III).

[0036] The above steps S1 and S2 are for constructing a risk prediction feature set, and step S3 is for constructing a risk prediction method between the risk prediction features and the nuclear reactor system status.

[0037] The nuclear power reactor safety control system ensures the safe operation of the reactor by monitoring a variety of key parameters. The key prediction characteristics and characteristic variable descriptions of nuclear power reactors are detailed in Table 1 below:

[0038] Table 1 Key prediction characteristics of nuclear power reactors and description of characteristic variables

[0039]

[0040]

[0041] In nuclear power control systems, the probability of failure of certain key functions is particularly low, but once these low-probability events occur, they may bring serious consequences, such as the probability of failure of the fuel assembly state and the probability of failure of the auxiliary system state in Table 1 above. Traditional methods are difficult to accurately estimate the probability of occurrence of these low-probability events, while the algorithm of the present invention can generate these low-probability characteristic values, providing key data for risk prediction. The algorithm uses statistical model verification technology, which is a method suitable for verifying relevant safety attributes of random systems. By simulating system behavior and calculating the probability of attribute satisfaction, the possibility of low-probability events can be effectively evaluated. In order to reduce the number of samples required for simulation and increase the number of occurrences of low-probability events, the algorithm uses importance sampling technology. By selecting a suitable sampling distribution, the frequency of occurrence of low-probability events in simulation can be increased, thereby improving the efficiency of the algorithm.

[0042] The step S2 proposes a small probability feature value generation algorithm based on statistical model testing for small probability feature items, which specifically includes the following steps:

[0043] Step B1: Model the nuclear power control system and describe the safety requirements in a standardized manner

[0044] The process of the statistical model verification method is as follows: first, simulate the running system model M and sample the model execution path (Trace) σ; then verify whether each path σ satisfies the attribute formula φ; finally, calculate the probability p that the model M satisfies the attribute formula φ based on the sampling result that the execution path σ satisfies the attribute formula φ.

[0045] Step B2: Use statistical model detection to calculate the required small probability eigenvalues

[0046] The statistical analysis process will generate more model execution paths σ so that the value of probability p is within an acceptable error θ. In statistical model detection, the principle of verification is the Monte Carlo method. The process in the present invention is as follows: assuming that the verified attribute is the probability of failure of the nuclear power fuel assembly state, a prediction model is established to determine the independent variables driving the prediction; specify the probability distribution of the independent variables, use historical data or the analyst's subjective judgment, define a series of possible values, and distribute probability weights for each value; repeatedly run the simulation to generate random values ​​for the independent variables. Repeat until a sufficient number of results are collected to form a representative sample of almost infinite possible combinations, which can be expressed by the following formula (IV):

[0047]

[0048] Where N1 is the number of simulation samples. If the sample meets the fuel assembly failure condition, then y(x i )=1, otherwise y(x i )=0.

[0049] Step B3: Optimize the original statistical model detection method using importance sampling

[0050] For nuclear power control systems, the safety integrity level is relatively high, that is, the probability of failure of the fuel assembly state and the auxiliary system state failure is less than 10. -8 If statistical model detection is used, in order to keep a small relative error, the required sample size is very large and requires a lot of simulation time. Therefore, it is proposed to use importance sampling to optimize the original statistical model detection algorithm.

[0051] The process of importance sampling is as follows: In importance sampling, you first need to select or design a sampling distribution, which should make the probability of the target event higher, thereby increasing the chance of sampling the event of interest; generate samples under the selected sampling distribution, rather than the original distribution, which means that the generation of samples is biased and tends to be more inclined to those areas that have a greater impact on the estimation results; for each sample, calculate its weight, which is a function of the ratio of the original distribution to the sampling distribution, and is used to adjust the sample to reflect the characteristics of the original distribution. The weight formula is expressed by the following (V) formula:

[0052]

[0053] Where f(x) is the probability density function of the original distribution, and g(x) is the probability density function of the sampled distribution. The weighted samples are then used to estimate the expected value or other statistics of the target function. The estimation of the expected value can be expressed as follows (VI):

[0054]

[0055] Where N2 is the number of samples, f(x i ) is the function value, W(x i ) is the weight of the corresponding sample. Therefore, the probability estimate of importance sampling is calculated using the following formula (VII):

[0056]

[0057] Among them, N2< <N1。

[0058] The step S3 proposes a nuclear power control system safety risk prediction model based on the Transformer model, which specifically includes the following steps:

[0059] Step C1: construct the initially selected risk features and small probability features into a feature set and convert them into a series of vector representations, and input them into the encoder by position encoding, and input the nuclear power control system state into the decoder. The abnormality of the nuclear reactor is closely related to the characteristic parameters of the nuclear reactor during this period. Specifically, the system parameter variables that have abnormalities at time T are different from those under normal operating conditions, and the system operating variables in the period before and after the abnormality will also have time correlation. Therefore, to establish an effective safety risk prediction model, it is necessary to extract and divide the feature data from a time perspective and process the system state parameter time series data, which is crucial to improving the accuracy of nuclear power reactor system state prediction classification.

[0060] See also Figure 2 , the input of the Transformer encoder is a sequence data. In the present invention, time series samples are used as input, and the input samples are in the form of a set of time series data samples X represented by the following formula (VIII):

[0061] X={(x1,y1), (x2,y2),......(x p ,y p )} (VIII).

[0062] Among them, x1, x2...x p is a time series sample; y1, y2, ...y p It is the status of nuclear power reactor system.

[0063] Each sequence sample has t observation points at different times, and is expressed by the following formula (IX):

[0064] x i ={x(1), x(2),...x(t)}1≤i≤p(IX).

[0065] Each observation point includes 13 system characteristic data, including 11 preliminary prediction characteristics and 2 small probability characteristic values, and is expressed by the following formula (X):

[0066] x(t)=F(t)=[f1(t), f2(t), f3(t),...f13(t)](X).

[0067] Among them, f1(t), f2(t), ... f 13 (t) are 13 system characteristic data respectively.

[0068] The Transformer model is different from RNN or CNN. It does not have position information, so it is necessary to introduce position encoding. The position encoding is expressed by the following formulas (XI) to (XII):

[0069]

[0070] Among them, d model represents the dimension of the feature vector, pos represents the absolute position of the observation point in the sequence sample, and i represents the dimension of the observation point. Adding the position encoding to the feature vector of the observation point is the input embedding of the decoder of the Transformer model.

[0071] The input of the Transformer decoder is processed in the same way as the input of the encoder. In the nuclear power control system, the encoder receives the time series sample x i , the decoder receives the nuclear power reactor system state y i When making predictions, the decoder input does not receive any information.

[0072] The Transformer encoder is composed of N=6 identical layers stacked together, each of which contains two sublayers. The first sublayer is a multi-layer self-attention mechanism (Multi-HeadAttention), and the second sublayer is a fully connected neural network (FeedForwardNetwork).

[0073] In the Transformer model, the attention function provides three weight matrices W Q , W K , W V , the input vector is multiplied by the weight matrix respectively to obtain the Query, Key, and Value matrices required to calculate the attention value, which are abbreviated as Q, K, and V matrices; the attention is calculated using dot multiplication. The specific steps are as follows: calculate the dot product of each vector in Q and each vector in K; then divide the result after the dot product by where d k is the dimension of the K matrix; then the result is converted into a probability distribution between [0, 1] through the softmax function; finally, it is multiplied by the V matrix and expressed by the following formula (XIII):

[0074]

[0075] The multi-head self-attention mechanism is to use multiple sets of weight matrices W Q , W K , W V To obtain multiple groups of Query, Key, and Value matrices, each group is calculated to obtain a Z matrix, and finally the multiple Z matrices are spliced.

[0076] The results obtained by the multi-head self-attention mechanism are residually linked and layer normalized and then input into the fully connected neural network, whose fully connected layer is expressed by the following formula (XIV):

[0077] FFN(x)=max(0,xW1+b1)W2+b2(XIV).

[0078] The fully connected layer here is a two-layer neural network, which first performs a linear transformation, then passes through the ReLU nonlinearity, and finally performs a linear transformation. The input x is the output result of the multi-head self-attention layer.

[0079] See also Figure 2 The decoder is also composed of N = 6 identical layers stacked together. Figure 2 It can be seen that a decoder is composed of a masked multi-head self-attention sublayer, a multi-head self-attention sublayer and a fully connected neural network. There is one more multi-head self-attention sublayer than the encoder, and the rest of the structure is the same as the encoder.

[0080] The output layer first undergoes a linear transformation on the result of the last layer of the decoder, and then uses the Softmax function to obtain the output probability distribution C1 or C2 shown in the following formula (XV):

[0081] y=softmax(W s h f +b)(XV).

[0082] Among them, y represents the state of the nuclear power reactor system, W s represents the output layer weight matrix, b represents the output layer bias value, h f It is the result of the linear change of the last layer of the decoder.

[0083] Step C2: Design training sample dataset

[0084] The sample data set is randomly divided into three parts, of which 70% are training samples; 15% are validation samples to verify the generalization ability of the network; and 15% are test samples to detect the model's judgment performance for unknown classification sample sets. The loss function uses the cross entropy loss function Loss shown in the following formula (I):

[0085] Loss=-(ylog(p)+(1-y)log(1-p))(I);

[0086] Where y is the actual state of the nuclear power reactor system: y(C1)=1, y(C2)=0; p is the probability that the model predicts that the sample is of category C1.

[0087] Step C3: Validation test sample set

[0088] The test sample set is input into the trained security risk prediction model for prediction and classification, the classification results are obtained, and the prediction performance of the model is evaluated. The experimental results are shown in Table 2 below:

[0089] Table 2 Accuracy of the model before and after use

[0090] category Prediction accuracy Recurrent Neural Network Prediction Model 82％ Prediction model of the present invention 94％

[0091] It can be seen from Table 2 above that the accuracy of the recurrent neural network prediction model is 82%, and the accuracy can reach 94% after using the safety risk prediction model of the present invention, which fully demonstrates the beneficial technical effects and significant technical progress of the present invention.

[0092] The above specific implementations are only for further explanation of the present invention, and are not intended to limit the patent of the present invention. All equivalent implementations of the present invention should be included in the scope of the claims of the patent of the present invention.

Claims

1. A safety risk prediction method for nuclear power control system based on Transformer model, characterized in that: The nuclear power control system risk feature selection and small probability feature value generation algorithm are used to build a nuclear power control system safety risk prediction model to realize the intelligent safety risk prediction of the nuclear power control system, which specifically includes the following steps: Step 1: Preliminary selection of nuclear power system characteristics Conduct forecast analysis of nuclear power control system safety risks based on monitored operating parameters and reactor protection parameter tables, and make preliminary selections of nuclear power system risk characteristics; Step 2: Small probability eigenvalue generation algorithm B1: A small probability eigenvalue generation algorithm based on statistical model testing is used to model the nuclear power control system and to describe the safety requirements in a standardized manner; B2: Use statistical models to detect and calculate small probability eigenvalues; B3: Use importance sampling to optimize the original statistical model detection method; Step 3: Build a Transformer-based intelligent safety risk prediction model for nuclear power control systems C1: construct the initially selected risk features and small probability feature values ​​into a feature set, convert them into a series of vector representations, input them into the encoder after position encoding, and input the nuclear power control system state into the decoder; C2: The sample data set is randomly divided into training samples, validation samples and test samples, of which 70% of the sample data set is the training sample data set; 15% of the sample data set is the validation sample data set to verify the generalization ability of the network; 15% of the sample data set is a test sample data set to test the model's judgment performance on unknown classification sample sets; C3: Input the test sample set into the trained security risk prediction model for prediction and classification, obtain the classification results, and evaluate the prediction performance of the model.

2. According to claim 1, the method for predicting safety risks of nuclear power control systems based on the Transformer model is characterized in that: The training sample data set is used for training using the following ( ) is the cross entropy loss function shown in : （ ）； in, The status of the real nuclear power reactor system: The probability that the model predicts for a sample.

Images