High-accuracy nuclear power key index predictor based on space-time correlation

Abstract

The invention discloses a high-accuracy nuclear power key index predictor based on spatial-temporal correlation. The high-accuracy nuclear power key index predictor is formed by connecting a nuclear power radiation station database, an upper computer and a sensor array. And the nuclear power radiation station database stores historical records of monitoring data of the nuclear power radiation station. The upper computer comprises a generative adversarial module and a nuclear power key index prediction module, data in a database is used for training a generator and a discriminator in the generative adversarial module, and then the trained generator serves as a nuclear power key index prediction model to be stored in the nuclear power key index prediction module. And the sensor array receives new radiation related data and transmits the new radiation related data to the nuclear power key index prediction module of the upper computer for nuclear power key index prediction. According to the method, the generative attention model is innovatively introduced to fully mine the temporal correlation and spatial correlation of the data of the nuclear power radiation station, the anti-interference capability of the predictor is enhanced, and the prediction accuracy of the nuclear power key indexes is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of nuclear power plant index monitoring and data-driven, and particularly relates to a high-accuracy nuclear power key index predictor based on spatio-temporal correlation. Background Technique

[0002] The basic principle of nuclear power key index prediction technology is to simulate and predict the diffusion and propagation process of radioactive substances in the atmosphere, water bodies and soil. These models are usually based on principles in fields such as physics, meteorology and earth science, and combine actual monitoring data and characteristic parameters of radiation sources for complex calculations and simulations. Through these models, nuclear power key indicators such as the propagation path, concentration distribution and arrival time of radioactive substances can be predicted, providing important information for decision-makers to formulate emergency response plans and protection measures. However, there are still many deficiencies in existing nuclear power key index prediction technologies. One challenge is that the accuracy of the model is affected by many factors, such as meteorological conditions, terrain and landforms, radiation source characteristics, etc. The uncertainty and complexity of these factors make the prediction results may have a certain degree of error, and existing methods cannot comprehensively utilize the correlation information of multiple variables, resulting in low accuracy. Another challenge is the real-time nuclear power key index prediction in emergency situations. Nuclear power radiation accidents often require rapid responses, but real-time prediction requires a large amount of data and calculations, and accurate results need to be provided within a short time. This poses high requirements for computing power and data transmission, and also challenges the real-time performance and operability of the prediction model.

[0003] With the increasing maturity of artificial intelligence technology, data-driven related technologies have also been fully developed. Applying data-driven methods to nuclear power key index prediction can bring multiple benefits. First of all, data-driven technologies can utilize a large amount of measured data and historical data to predict the propagation and diffusion of nuclear power key indicators by establishing complex models and algorithms. These models can consider various factors, such as meteorological conditions, terrain and landforms, radiation source characteristics, etc., thereby improving the accuracy of prediction. Secondly, data-driven technologies can quickly generate prediction results through the input of real-time monitoring data. In this way, decision-makers can timely understand the radiation situation and take appropriate emergency measures and protection measures. Real-time index prediction can help reduce the risk of personnel exposure and minimize the impact of accidents. Finally, data-driven technologies can provide scientific basis and support for decision-makers to help them formulate more effective emergency response plans and protection strategies. The prediction results can reveal the propagation path, concentration distribution and arrival time of nuclear power key indicators, thus guiding decisions such as personnel evacuation, area isolation and resource allocation.

[0004] Current nuclear power key index prediction methods and devices, including instruments applying data-driven related technologies, do not fully utilize the temporal and spatial correlations in the historical monitoring data of nuclear power radiation stations. It is also difficult for model design to notice the most critical factors for future changes in nuclear power key indices. Therefore, there are deficiencies such as low anti-interference ability, poor feature extraction ability, and low prediction accuracy. This technology is a difficult point and a hot topic in the research of related fields at home and abroad, and has important application value for national and social security. There is an urgent need to invent a more stable and accurate predictor to complete real-time high-precision prediction of nuclear power key indices. Summary of the Invention

[0005] The purpose of the present invention is to provide a high-accuracy nuclear power key index predictor based on spatio-temporal correlation in view of the deficiencies of the prior art.

[0006] The purpose of the present invention is achieved through the following technical solutions: A high-accuracy nuclear power key index predictor based on spatio-temporal correlation, the predictor includes a nuclear power radiation station database, a host computer, and a sensor array; the nuclear power radiation station database and the sensor array are respectively connected to the host computer; the host computer includes a generative adversarial module and a nuclear power key index prediction module connected in sequence;

[0007] The nuclear power radiation station database is used to store the historical monitoring data D of the nuclear power radiation station and upload it to the generative adversarial module in the host computer;

[0008] The generative adversarial module is used to train the generative adversarial attention model using the historical monitoring data, and then upload the generator in the trained generative adversarial attention model as a prediction model to the nuclear power key index prediction module;

[0009] The sensor array is used to monitor and obtain real-time radiation-related data and upload it to the nuclear power key index prediction module in the host computer;

[0010] The nuclear power key index prediction module is used to use the prediction model to make a prediction according to the real-time radiation-related data, and obtain the prediction result of the future nuclear power key index.

[0011] Further, the historical monitoring data D is D = {X 1 , X 2 , …, X a , …, X n}, where X a represents the monitoring data at the a-th moment, n represents the monitoring data at a total of n moments in the historical monitoring data D, and a = 1, 2, …, a, …, n; m represents that each moment of monitoring data X a is respectively composed of m monitoring variable values, Denote the monitoring data as X a which is the value of the b-th monitoring variable in a , where b = 1, 2, …, b, …, m.

[0012] Furthermore, the generative adversarial module uses the historical monitoring data to train the generative adversarial attention model, and then uploads the generator in the trained generative adversarial attention model to the nuclear power key index prediction module as the prediction model. Specifically:

[0013] (a.1) The generative adversarial module contains a generative adversarial attention model, which includes a generator and a discriminator;

[0014] The generator consists of a first temporal correlation extraction network, a first spatial correlation extraction network, a second temporal correlation extraction network, a second spatial correlation extraction network, a gated recurrent unit, and a fourth fully connected neural network; the discriminator consists of a fifth fully connected neural network;

[0015] (a.2) The generative adversarial module extracts a data segment S of length T from the historical monitoring data D: S = {X t-T+1 , X t-T+2 , …, X c , …, X t} and inputs it into the first temporal correlation extraction network, where t represents any moment, and X c represents the monitoring data at the c-th moment, where c = t - T + 1, t - T + 2, …, c, …, t;

[0016] (a.3) The generator predicts the data segment S to obtain a prediction result and inputs it into the discriminator;

[0017] (a.4) Meanwhile, the discriminator obtains the data y d = {X t+1 , X t+2 , …, X t-H} of the data segment S at H moments after the data segment S is obtained from the historical monitoring data D; and constructs a total loss function l based on the prediction result and the data y d ;

[0018] (a.5) According to the total loss function l, backpropagation is performed on the generator and the discriminator, and all neural network parameters in the generative adversarial attention model are updated using the stochastic gradient descent method;

[0019] (a.6) Repeat steps (a.3) - (a.5) until the total loss function l no longer decreases, complete the training of the generative adversarial attention model, and obtain the trained generative adversarial attention model; and use the generator in the trained generative adversarial attention model as the prediction model and upload it to the nuclear power key index prediction module.

[0020] Further, the specific steps of step (a.3) include the following sub-steps:

[0021] (a.3.1) The first temporal correlation extraction network first processes the data segment S through the first multi-layer perceptron (MLP) 1 , the second multi-layer perceptron (MLP) 2 and the third multi-layer perceptron (MLP) 3 to obtain the output vectors output vector and output vector The first multi-layer perceptron (MLP) 1 , the second multi-layer perceptron (MLP) 2 and the third multi-layer perceptron (MLP) 3 have the same structure, all consisting of an input layer, two hidden layers and an output layer. The dimension of the input layer is m, the dimension of the hidden layer is p, and the dimension of the output layer is q. The calculation formula is as follows:

[0022]

[0023] Among them, the output vectors output vector and output vector V t 1 are respectively of dimension T×q;

[0024] Subsequently, the output vectors output vector and output vector V t 1 are subjected to attention calculation to obtain the output vector where SoftMax(·) is the SoftMax function and d is a constant;

[0025] Finally, the output vector is multiplied by the parameter matrix W 0 to obtain the output vector of the first temporal correlation extraction network and input it into the first spatial correlation extraction network, where W 0 has a size of q×q and consists of learnable numerical parameters;

[0026] (a.3.2) First, transpose the output vector to obtain the transposed and input it into the first spatial correlation extraction network. The first spatial correlation extraction network processes the transposed separately through the first fully connected neural network MLP 1 , the second fully connected neural network MLP 2 and the third fully connected neural network MLP 3 to obtain the output vector The output vector and the output vector V t 2 . The calculation formula is as follows:

[0027]

[0028] Subsequently, perform attention calculation on the output vector the output vector and the output vector V t 2 to obtain the output vector

[0029] Finally, multiply the output vector by the parameter matrix W 0 to obtain the output vector S of the first spatial correlation extraction network t 2 : and input it into the second temporal correlation extraction network;

[0030] (a.3.3) First, transpose the output vector to obtain the transposed and input it into the second temporal correlation extraction network. The second temporal correlation extraction network processes the transposed separately through the first fully connected neural network MLP 1 , the second fully connected neural network MLP 2 and the third fully connected neural network MLP 3 to obtain the output vector The output vector and the output vector V t 3 . The calculation formula is as follows:

[0031]

[0032] Subsequently, perform attention calculation on the output vector the output vector and the output vector V t 3 to obtain the output vector

[0033] Finally, multiply the output vector by the parameter matrix W 0 to obtain the output vector of the second temporal correlation extraction network and input it into the second spatial correlation extraction network;

[0034] (a.3.4) First, transpose the output vector to obtain the transpose and input it into the second spatial correlation extraction network. The second spatial correlation extraction network processes the transpose through the first fully connected neural network MLP 1 , the second fully connected neural network MLP 2 and the third fully connected neural network MLP 3 to obtain the output vector The output vector and the output vector V t 4 . The calculation formula is as follows:

[0035]

[0036] Subsequently, perform attention calculation on the output vector the output vector and the output vector V t 4 to obtain the output vector

[0037] Finally, multiply the output vector by the parameter matrix W 0 to obtain the output vector of the second spatial correlation extraction network and input it into the gated recurrent unit;

[0038] (a.3.5) First, transpose the output vector to obtain the transpose and input it into the gated recurrent unit. The gated recurrent unit processes the transpose to obtain the output vector h g and input it into the fourth fully connected neural network;

[0039] (a.3.6) Subsequently, the fourth fully connected neural network makes a prediction on the output vector h g to obtain the prediction result and input it into the discriminator.

[0040] Further, (a.3.5) is specifically as follows:

[0041] First, transpose the output vector to obtain the transpose and input it into the gated recurrent unit. The transpose is where x c is the data at any moment in the transpose ;

[0042] When the transpose is input into the gated recurrent unit, the calculation process at the first moment is as follows:

[0043] z 1 = σ(W z [h 0 , x t-T+1 );

[0044] r 1 = σ(W r [h 0 , x t-T+1 );

[0045]

[0046] where W z , W r and W h are respectively learnable parameter matrices; σ(·) is the Sigmoid function; tanh(·) is the tanh function; h 0 is the initial hidden vector of the gated recurrent unit, randomly sampled from a normal distribution, with a dimension of q;

[0047] The calculation process at the e-th moment of the gated recurrent unit is as follows:

[0048] z e = σ(W z [h e-1 , x t-T+e );

[0049] r e = σ(W r [h e-1 , x t-T+e );

[0050]

[0051] where h e-1 is the output vector of the gated recurrent unit at the previous moment; z e , r e and is the intermediate output generated within the gated recurrent unit; h e is the output vector at this moment;

[0052] Repeat the above calculation to obtain the output vector of the gated recurrent unit at the t-th moment as h g and input it into the fourth fully connected neural network.

[0053] Furthermore, the step (a.4) specifically includes the following sub-steps:

[0054] (a.4.1) The discriminator obtains the data y at H moments after the data segment S from the historical monitoring data D d ={X t+1 , X t+2 , …, X t-H}; Subsequently, the discriminator constructs the adversarial loss function l and the data y d : adv :

[0055]

[0056] where E represents the mathematical expectation, y d ~p data means that the data y d follows the data distribution p in the database data ; represents the prediction result follows the prior model distribution p of the generator g ; D(y d ) represents the probability that the discriminator gives the data y d as a real sample; represents the probability that the discriminator gives the data as a real sample; ΔD(y d ) represents the gradient of the discriminator for the input y d ; ||·|| 2 represents the L2 norm; λ is the penalty term weight;

[0057] (a.4.2) And construct the mean squared error loss function l and the data y d : reg :

[0058]

[0059] (a.4.3) Construct the total loss function l through the adversarial loss function l adv and the mean squared error loss function l reg : l = αl reg + βladv , where α and β are constants greater than 0 and satisfy α + β.

[0060] Furthermore, the sensor array is a sensor array composed of m sensors, and each sensor is responsible for monitoring a variable.

[0061] The beneficial effects of the present invention are as follows:

[0062] 1) Innovatively introduce the time and space attention network, enabling the nuclear power key index prediction model to fully utilize the time correlation in the monitoring data and the space correlation between variables, and improving the prediction accuracy of the predictor;

[0063] 2) Innovatively introduce the generative adversarial architecture, which can improve the stability and anti-interference ability of the predictor, making the predictor have strong prediction ability and high prediction accuracy;

[0064] 3) The predictor can predict the future value in real time according to the monitoring value, and provide timely and effective guidance for behavior decision-making, with strong real-time performance and intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a structural diagram of a high-accuracy nuclear power key index predictor based on spatio-temporal correlation;

[0066] Figure 2 It is a structural diagram of the generative adversarial module;

[0067] In the figure, 1: Nuclear power radiation station database; 2: Host computer; 3: Generative adversarial module; 4: Nuclear power key index prediction module; 5: Sensor array; 3-1: First time correlation extraction network; 3-2: First space correlation extraction network; 3-3: Second time correlation extraction network; 3-4: Second space correlation extraction network; 3-5: Gated recurrent unit; 3-6: Fourth fully connected neural network; 3-7: Fifth fully connected neural network. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0069] The technical concept of the present invention is as follows: high-precision intelligent prediction of key nuclear power indicators, monitoring of variable data by nuclear power radiation stations, constructing and storing a data set, using the data set to train a generative adversarial attention model, obtaining a trained radiation source prediction model, and using it to predict future key nuclear power indicators based on new variable values monitored by sensors.

[0070] Embodiment 1

[0071] As Figure 1 shown, the present invention provides a high-accuracy nuclear power key indicator predictor based on spatio-temporal correlation. The predictor includes a nuclear power radiation station database 1, a host computer 2, and a sensor array 5; the nuclear power radiation station database 1 and the sensor array 5 are respectively connected to the host computer 2. The host computer 2 includes a generative adversarial module 3 and a nuclear power key indicator prediction module 4 connected in sequence.

[0072] The nuclear power radiation station database 1 is used to store the historical monitoring data D of the nuclear power radiation station and upload it to the generative adversarial module 3 in the host computer 2.

[0073] The historical monitoring data D is D = {X 1 , X 2 , …, X a , …, X n}, where X a represents the monitoring data at the a-th moment, n represents the monitoring data at n moments in the historical monitoring data D, and a = 1, 2, …, a, …, n; m represents that each moment of the monitoring data X a is respectively composed of m monitoring variable values, represents the b-th monitoring variable value in the monitoring data X a , and b = 1, 2, …, b, …, m.

[0074] The generative adversarial module 3 is used to train the generative adversarial attention model using the historical monitoring data, and then upload the generator in the trained generative adversarial attention model as a prediction model to the nuclear power key indicator prediction module 4. By innovatively introducing a generative adversarial architecture, the stability and anti-interference ability of the predictor are improved, so that the predictor has strong prediction ability and high prediction accuracy, and can predict future values in real time according to the monitored values, providing guidance for behavior decision-making in a timely and effective manner, with strong real-time performance and intelligence. Specifically:

[0075] (a.1) The generative adversarial module 3 contains a generative adversarial attention model, and the generative adversarial attention model includes a generator and a discriminator;

[0076] As Figure 2As shown, the generator consists of a first temporal correlation extraction network 3-1, a first spatial correlation extraction network 3-2, a second temporal correlation extraction network 3-3, a second spatial correlation extraction network 3-4, a gated recurrent unit 3-5, and a fourth fully-connected neural network 3-6. The discriminator consists of a fifth fully-connected neural network 3-7.

[0077] (a.2) The generative adversarial module 3 extracts a data segment S of length T from the historical monitoring data D: S = {X t-T+1 , X t-T+2 , …, X c , …, X t} and inputs it into the first temporal correlation extraction network 3-1, where t represents any moment, and X c represents the monitoring data at the c-th moment, and c = t - T + 1, t - T + 2, …, c, …, t.

[0078] (a.3) The generator makes a prediction on the data segment S to obtain a prediction result and inputs it into the discriminator.

[0079] The step (a.3) specifically includes the following sub-steps:

[0080] (a.3.1) The data segment S is first fed into the first temporal correlation extraction network 3-1 to extract the temporal correlation between the monitoring data at each moment in the data segment S.

[0081] The first temporal correlation extraction network 3-1 first processes the data segment S through a first multi-layer perceptron MLP 1 , a second multi-layer perceptron MLP 2 and a third multi-layer perceptron MLP 3 to obtain output vectors output vector and output vector V t 1 , and the first multi-layer perceptron MLP 1 , the second multi-layer perceptron MLP 2 and the third multi-layer perceptron MLP 3 have the same structure, all consisting of an input layer, two hidden layers, and an output layer. The dimension of the input layer is m, the dimension of the hidden layer is p, and the dimension of the output layer is q. The calculation formula is as follows:

[0082]

[0083] Among them, the output vector output vector and output vector V t 1The dimensions are T×q respectively.

[0084] Subsequently, the output vector Output vector and the output vector V t 1 are subjected to attention calculation to obtain the output vector where SoftMax(·) is the SoftMax function and d is a constant;

[0085] Finally, the output vector is multiplied by the parameter matrix W 0 to obtain the output vector of the first temporal correlation extraction network 3-1 and is input into the first spatial correlation extraction network 3-2, where W 0 has a size of q×q and is composed of learnable numerical parameters.

[0086] (a.3.2) The transpose of the output vector is fed into the first spatial correlation extraction network 3-2 to extract the correlation between the monitored variable values in the monitored data at each moment; the structure of the first spatial correlation extraction network 3-2 is exactly the same as that of the first temporal correlation extraction network 3-1, and the calculation steps in the first spatial correlation extraction network 3-2 are also exactly the same as the calculation steps of the data segment S in the first temporal correlation extraction network 3-1. The output vector of the first spatial correlation extraction network 3-2 obtained is Thereby overcoming the drawback that the existing method cannot comprehensively utilize the correlation information of multiple variables, resulting in a low prediction accuracy.

[0087] First, the output vector is transposed to obtain the transpose and is input into the first spatial correlation extraction network 3-2. The first spatial correlation extraction network 3-2 processes the transpose through the first fully connected neural network MLP 1 respectively, the second fully connected neural network MLP 2 and the third fully connected neural network MLP 3 to obtain the output vector Output vector and the output vector V t 2 , and the calculation formula is as follows:

[0088]

[0089] Subsequently, the output vector Output vector and the output vector V t 2 perform attention calculation to obtain the output vector

[0090] Finally, the output vector is multiplied by the parameter matrix W 0 to obtain the output vector of the first spatial correlation extraction network 3-2 and input it into the second temporal correlation extraction network 3-3

[0091] (a.3.3) Transpose of the output vector is fed into the second temporal correlation extraction network 3-3 for in-depth extraction of the temporal correlation between the monitoring data at each moment; the structure of the second temporal correlation extraction network 3-3 is also exactly the same as that of the first temporal correlation extraction network 3-1 and the calculation steps in the second temporal correlation extraction network 3-3 are also exactly the same as those of the data segment S in the first temporal correlation extraction network 3-1. The output vector of the second temporal correlation extraction network 3-3 obtained is

[0092] First, the output vector is transposed to obtain the transpose and input it into the second temporal correlation extraction network 3-3, and the second temporal correlation extraction network 3-3 processes the transpose through the first fully connected neural network MLP 1 , the second fully connected neural network MLP 2 and the third fully connected neural network MLP 3 to obtain the output vector Output vector and the output vector V t 3 , and the calculation formula is as follows

[0093]

[0094] Subsequently, the output vector Output vector and the output vector V t 3 perform attention calculation to obtain the output vector

[0095] Finally, the output vector is multiplied by the parameter matrix W 0Multiply to obtain the output vector of the second temporal correlation extraction network 3-3 and input it into the second spatial correlation extraction network 3-4.

[0096] (a.3.4) Output vector The transpose of is fed into the second spatial correlation extraction network 3-4 to extract the correlation between the monitored variable values in the monitored data at each moment; the structure of the second spatial correlation extraction network 3-4 is also exactly the same as that of the first temporal correlation extraction network 3-1, and the calculation steps in the second spatial correlation extraction network 3-4 are also exactly the same as the calculation steps of the data segment S in the first temporal correlation extraction network 3-1. The output vector of the obtained second spatial correlation extraction network 3-4 is Thereby further overcoming the drawback that the existing method cannot comprehensively utilize the correlation information of multiple variables, resulting in a relatively low prediction accuracy.

[0097] First, transpose the output vector to obtain the transpose and input it into the second spatial correlation extraction network 3-4. The second spatial correlation extraction network 3-4 transposes and processes it through the first fully connected neural network MLP 1 , the second fully connected neural network MLP 2 and the third fully connected neural network MLP 3 to obtain the output vector Output vector and output vector V t 4 . The calculation formula is as follows:

[0098]

[0099] Subsequently, perform attention calculation on the output vector Output vector and output vector V t 4 to obtain the output vector

[0100] Finally, multiply the output vector by the parameter matrix W 0 to obtain the output vector of the second spatial correlation extraction network 3-4 and input it into the gated recurrent unit 3-5.

[0101] (a.3.5) First, transpose the output vector to obtain the transpose and input it into the gated recurrent unit 3-5, and the gated recurrent unit 3-5 transposes for processing to obtain the output vector h g and input it into the fourth fully connected neural network 3-6, and the calculation formula is as follows:

[0102]

[0103] where GRU(·) represents the processing of the gated recurrent unit.

[0104] The specific step (a.3.5) is as follows:

[0105] First, transpose the output vector to obtain the transposed and input it into the gated recurrent unit 3-5, and the transposed is where x c is the transposed data at any moment in

[0106] When the transposed is input into the gated recurrent unit 3-5, the calculation process at the first moment is as follows:

[0107] z 1 =σ(W z [h 0 , x t-T+1 );

[0108] r 1 =σ(W r [h 0 , x t-T+1 );

[0109]

[0110] where W z , W r and W h are respectively learnable parameter matrices; σ(·) is the Sigmoid function; tanh(·) is the tanh function; h 0 is the initial hidden vector of the gated recurrent unit, randomly sampled from a normal distribution, with a dimension of q.

[0111] The calculation process at the e-th moment of the gated recurrent unit is as follows:

[0112] z e =σ(W z [h e-1 , x t-T+e );

[0113] r e = σ(W r [h e-1 ,x t-T+e );

[0114]

[0115] where h e-1 is the output vector of the gated recurrent unit at the previous moment; z e , r e and are the intermediate outputs generated within the gated recurrent unit; h e is the output vector at this moment.

[0116] Repeat the above calculation to obtain the output vector of the gated recurrent unit at the t-th moment as h g and input it into the fourth fully connected neural network 3-6.

[0117] (a.3.6) Subsequently, the fourth fully connected neural network 3-6 makes a prediction on the output vector h g to obtain a prediction result and input it into the discriminator.

[0118] (a.4) At the same time, the discriminator obtains the data y at H moments after the data segment S from the historical monitoring data D d = {X t+1 , X t+2 , …, X t-H}; and constructs the total loss function l based on the prediction result and the data y d At the same time, the discriminator obtains the data y at H moments after the data segment S from the historical monitoring data D d = {X t+1 , X t+2 , …, X t-H}; and constructs the total loss function l based on the prediction result and the data y d .

[0119] Step (a.4) specifically includes the following sub-steps:

[0120] (a.4.1) At the same time, the discriminator obtains the data y at H moments after the data segment S from the historical monitoring data D d = {X t+1 , X t+2 , …, X t-H}; Subsequently, the discriminator constructs the adversarial loss function l through the prediction result d and the data y adv :

[0121]

[0122] Among them, E represents the mathematical expectation of y d ~p data represents the data y d obeys the data distribution p in the database data ; represents the prediction result obeys the generator prior model distribution p g ; D(y d ) represents the probability that the discriminator gives the data y d as a real sample; represents the probability that the discriminator gives the data as a real sample; ΔD(y d ) represents the gradient of the discriminator with respect to the input y d ; ||·|| 2 represents the L2 norm; λ is the penalty term weight, used to control the intensity of the gradient penalty, generally taken as 10.

[0123] (a.4.2) And according to the prediction result and the data y d construct the mean square error loss function l reg :

[0124]

[0125] (a.4.3) Through the adversarial loss function l adv and the mean square error loss function l reg construct the total loss function l: l = αl reg +βl adv , where α and β are constants greater than 0 and satisfy α + β; control the proportions of the adversarial loss function l adv and the mean square error loss function l reg in the total loss function l.

[0126] (a.5) According to the total loss function l, perform backpropagation on the generator and the discriminator, and update all neural network parameters in the generative adversarial attention model using the stochastic gradient descent method.

[0127] (a.6) Repeat steps (a.3) - step (a.5) until the total loss function l no longer decreases, complete the training of the generative adversarial attention model, and obtain the trained generative adversarial attention model; and upload the generator in the trained generative adversarial attention model as the prediction model to the nuclear power key index prediction module.

[0128] The sensor array 5 is used to monitor real-time radiation-related data and upload it to the nuclear power key index prediction module 4 in the host computer 2.

[0129] The sensor array 5 is a sensor array composed of m sensors, and each sensor is responsible for monitoring a variable.

[0130] The nuclear power key index prediction module 4 is used to make a prediction based on the real-time radiation-related data using the prediction model to obtain the prediction result of future nuclear power key indexes.

[0131] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A high-accuracy nuclear power key indicator prediction instrument based on time-space correlation, characterized in that: The predictor comprises a nuclear power radiation station database, a host computer and a sensor array; the nuclear power radiation station database and the sensor array are respectively connected to the host computer; the host computer comprises a generative adversarial module and a nuclear power key indicator prediction module which are connected in sequence; The nuclear power radiation station database is used to store the historical monitoring data D of the nuclear power radiation station and upload it to the generative adversarial module in the host computer; The generative adversarial module is used to train the generative adversarial attention model using the historical monitoring data, and then upload the generator in the trained generative adversarial attention model as a prediction model to the nuclear power key indicator prediction module; The sensor array is used to monitor and obtain real-time radiation-related data and upload it to the nuclear power key indicator prediction module in the host computer; The nuclear power key indicator prediction module is used to use the prediction model to make predictions based on the real-time radiation-related data to obtain prediction results of future nuclear power key indicators.

2. The high-accuracy nuclear power key indicator prediction instrument based on time-space correlation according to claim 1 is characterized in that: The historical monitoring data D is D={X1, X2, ..., X a ,…,X n }, where X a represents the monitoring data at the ath moment, n represents the monitoring data at the nth moment in the historical monitoring data D, a=1,2,…,a,…,n; m represents the monitoring data X at each moment a They are composed of m monitoring variable values ​​respectively. Indicates monitoring data X a The bth monitored variable value in , b=1,2,…,b,…,m.

3. The high-accuracy nuclear power key indicator prediction instrument based on time-space correlation according to claim 2 is characterized in that: The generative adversarial module uses the historical monitoring data to train the generative adversarial attention model, and then uploads the generator in the trained generative adversarial attention model as a prediction model to the nuclear power key indicator prediction module, specifically: (a.1) The generative adversarial module includes a generative adversarial attention model, and the generative adversarial attention model includes a generator and a discriminator; The generator is composed of a first time correlation extraction network, a first space correlation extraction network, a second time correlation extraction network, a second space correlation extraction network, a gated recurrent unit and a fourth fully connected neural network; the discriminator is composed of a fifth fully connected neural network; (a.2) The generative adversarial module takes a data segment S of length T from the historical monitoring data D: S = {X t-T+1 ,X t-T+2 ,…,X c ,…,X t } and input to the first time correlation extraction network, where t represents any moment, X c represents the monitoring data at the cth moment, c = t-T+1, t-T+2, …, c, …, t; (a.3) The generator predicts the data segment S and obtains the prediction result And input to the discriminator; (a.4) At the same time, the discriminator obtains the data y at H moments after the data segment S from the historical monitoring data D d ={X t+1 ,X t+2 ,…,X t-H }; and according to the prediction results and data y d Construct the total loss function l; (a.5) Back-propagate the generator and discriminator according to the total loss function l, and use the stochastic gradient descent method to update all neural network parameters in the generative adversarial attention model; (a.6) Repeat steps (a.3) to (a.5) until the total loss function l no longer decreases, completing the training of the generative adversarial attention model and obtaining the trained generative adversarial attention model; and upload the generator in the trained generative adversarial attention model as a prediction model to the nuclear power key indicator prediction module.

4. The high-accuracy nuclear power key indicator prediction instrument based on time-space correlation according to claim 2 is characterized in that: The step (a.3) specifically includes the following sub-steps: (a.3.1) The first time correlation extraction network first processes the data segment S through the first fully connected neural network MLP1, the second fully connected neural network MLP2 and the third fully connected neural network MLP3 to obtain an output vector Output vector and the output vector The first fully connected neural network MLP1, the second fully connected neural network MLP2 and the third fully connected neural network MLP3 have the same structure, and are composed of an input layer, two hidden layers and an output layer. The dimension of the input layer is m, the dimension of the hidden layer is p, and the dimension of the output layer is q. The calculation formula is as follows: Among them, the output vector Output vector And the output vector V t 1 The dimensions are T×q; The output vector Output vector And the output vector V t 1 Perform attention calculation and get the output vector Among them, SoftMax(·) is the SoftMax function, and d is a constant; Finally, the output vector Multiply it with the parameter matrix W0 to get the output vector of the first time correlation extraction network And input into the first spatial correlation extraction network, where the size of W0 is q×q, which is composed of learnable numerical parameters; (a.3.2) First, the output vector Transpose to get the transpose And input to the first spatial correlation extraction network, the first spatial correlation extraction network will transpose The output vector is obtained by processing the first fully connected neural network MLP1, the second fully connected neural network MLP2 and the third fully connected neural network MLP3 respectively. Output vector And the output vector V t 2 , the calculation formula is as follows: The output vector Output vector And the output vector V t 2 Perform attention calculation and get the output vector Finally, the output vector Multiply it with the parameter matrix W0 to get the output vector of the first spatial correlation extraction network and input into the second time correlation extraction network; (a.3.3) First, the output vector Transpose to get the transpose And input to the second time correlation extraction network, the second time correlation extraction network will transpose The output vector is obtained by processing the first fully connected neural network MLP1, the second fully connected neural network MLP2 and the third fully connected neural network MLP3 respectively. Output vector And the output vector V t 3 , the calculation formula is as follows: The output vector Output vector And the output vector V t 3 Perform attention calculation and get the output vector Finally, the output vector Multiply it with the parameter matrix W0 to get the output vector of the second time correlation extraction network And input into the second spatial correlation extraction network; (a.3.4) First, the output vector Transpose to get the transpose And input to the second spatial correlation extraction network, the second spatial correlation extraction network will transpose The output vector is obtained by processing the first fully connected neural network MLP1, the second fully connected neural network MLP2 and the third fully connected neural network MLP3 respectively. Output vector And the output vector V t 4 , the calculation formula is as follows: The output vector Output vector And the output vector V t 4 Perform attention calculation and get the output vector Finally, the output vector Multiply it with the parameter matrix W0 to get the output vector of the second spatial correlation extraction network And input into the gated recurrent unit; (a.3.5) First, the output vector Transpose to get the transpose and input to the gated recurrent unit, which transposes Processing is performed to obtain the output vector h g And input into the fourth fully connected neural network; (a.3.6) Then the fourth fully connected neural network outputs the vector h g Make predictions and get prediction results And input to the discriminator.

5. The high-accuracy nuclear power key indicator prediction instrument based on time-space correlation according to claim 4 is characterized in that: The (a.3.5) is specifically: First, the output vector Transpose to get the transpose And input to the gated recurrent unit, the transposed for Among them, x c Transpose Data at any time in When transposed Input into the gated recurrent unit, the calculation process at the first moment is as follows: z1=σ(W z [h0,x t-T+1 ]); r1=σ(W r [h0,x t-T+1 ]); Among them, W z , W r and W h are learnable parameter matrices; σ(·) is the Sigmoid function; tanh(·) is the tanh function; h0 is the initial hidden vector of the gated recurrent unit, randomly sampled from a normal distribution, with a dimension of q; The calculation process at the e-th moment of the gated recurrent unit is as follows: z e =σ(W z [h e-1 ,x t-T+e ]); r e =σ(W r [h e-1 ,x t-T+e ]); Among them, h e-1 is the output vector of the gated recurrent unit at the previous moment; z e 、r e and is the intermediate output generated in the gated recurrent unit; h e is the output vector at this moment; Repeat the above calculation to obtain the output vector of the gated recurrent unit at the tth moment as h g And input into the fourth fully connected neural network.

6. The high-accuracy nuclear power key indicator prediction instrument based on time-space correlation according to claim 4 is characterized in that: The step (a.4) specifically includes the following sub-steps: (a.4.1) At the same time, the discriminator obtains the data y H moments after the data segment S from the historical monitoring data D d ={X t+1 ,X t+2 ,…,X t-H }; The discriminator then predicts the result and data y d Construct the adversarial loss function l adv : Among them, E represents the mathematical expectation y d ~p data Represents data y d Obey the data distribution p in the database data ; Represents the prediction results Obey the generator prior model distribution p g ; D(y d ) indicates that the discriminator gives data y d The probability of being a true sample; Indicates that the discriminator gives data is the probability of the true sample; ΔD(y d ) represents the discriminator's response to the input y d The gradient of ; ||·||2 represents the L2 norm; λ is the penalty weight; (a.4.2) and based on the prediction results and data y d Construct the mean square error loss function l reg : (a.4.3) By adversarial loss function l adv And the mean square error loss function l reg Construct the total loss function l: l = αl reg +βl adv , where α and β are constants greater than 0 and satisfy α+β.

7. The high-accuracy nuclear power key indicator prediction instrument based on time-space correlation according to claim 1 is characterized in that: The sensor array is composed of m sensors, and each sensor is responsible for monitoring one variable.

Images