A pressurized water reactor activation corrosion product source term prediction method based on recurrent neural network

Abstract

The present application belongs to the technical field of radiation protection and pressurized water reactor source term analysis, and particularly relates to a pressurized water reactor activated corrosion product source term prediction method. The method comprises the following steps: step 1, reading reactor historical operation data and performing data cleaning; step 2, reading reactor primary loop activated corrosion product nuclide specific activity data and performing data cleaning; step 3, selecting operation parameters as input and target nuclide specific activity as output in a recurrent neural network for normalization processing; step 4, dividing a historical data set into a training set and a test set; step 5, determining a recurrent neural network structure for predicting activated corrosion product source terms; and step 6, training the neural network using the training set and testing the neural network using the test set. The present application uses a recurrent neural network to predict the change trend of activated corrosion product source terms, and only the network training process requires certain computing resources, the computing resources required for subsequent source term prediction can be ignored, compared with traditional methods, the present application does not need to solve equation sets, has a faster prediction speed, and requires less computing resources; traditional activated corrosion product source term prediction models lack the use of historical data of reactor operation, and cannot combine the characteristics of target reactors for source term prediction.

Description

Technical Field

[0001] This invention belongs to the field of radiation protection and pressurized water reactor source term analysis technology, specifically relating to a method for predicting source terms of activated corrosion products in pressurized water reactors. Background Technology

[0002] During normal reactor operation, corrosion of the surfaces of coolant system pipes and structural materials generates corrosion products, which are transported to the active zone of the reactor core via the coolant. These products react with neutrons to form activated corrosion product source terms. Alternatively, materials located in the active zone of the reactor core may first react with neutrons and then be activated, subsequently corroding and entering the coolant to form activated corrosion product source terms. In pressurized water reactors, the main activated corrosion products include... 51 Cr 54 Mn, 59 Fe、 58 Co、 60 Co、 110m Ag、 122 Sb, 124 Sb, etc. Among them, the first four nuclides are activated corrosion products generated by the base metal material, while the latter four are activated corrosion products generated by a small number of special components.

[0003] Activated corrosion product nuclides typically decay continuously, releasing gamma rays, which are a major source of dose rate contribution during normal operation and after reactor shutdown in nuclear power plants. In a reactor environment under normal steady-state operation, the source term of activated corrosion products in the coolant system contributes over 90% of the radiation field outside the reactor core. Particularly for long-half-life nuclides, their contribution to the radiation field dose rate increases continuously as the reactor operates. Furthermore, during the migration of activated corrosion products within various reactor systems, they easily deposit at locations such as bends and blind ends of pipes, forming hotspots. The dose rate in these areas increases significantly, thus increasing the collective dose during major nuclear power plant outages.

[0004] Controlling the source term of activated corrosion products (ACTS) is the most important means of reducing the collective dose in nuclear power plants. The foundation of ACCS source term control lies in establishing a source term model and predicting future source term trends based on current reactor operating parameters. Currently, only mechanistic models exist for ACCS source terms, which theoretically calculate the generation and deposition rates of corrosion products and model the source term based on first-order kinetic equations. Mechanistic ACCS source term models cannot utilize historical monitoring data from nuclear power plants. However, the ACCS source term prediction method based on recurrent neural networks proposed in this invention can predict future source terms using historical monitoring data, providing technical support for ACCS source term control. This invention helps protect the safety of nuclear power plant maintenance personnel, improves maintenance efficiency, and thus enhances the safety and economy of nuclear power plant operation. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the source terms of activated corrosion products in pressurized water reactors based on recurrent neural networks. This method predicts the future trend of source term changes by inputting historical monitoring data of pressurized water reactor operation, thus providing technical support for the control of activated corrosion product source terms.

[0006] The technical solution of the present invention is as follows: A method for predicting source terms of activated corrosion products in pressurized water reactors based on recurrent neural networks, comprising the following steps:

[0007] Step 1: Read historical reactor operating data and perform data cleaning;

[0008] Step 2: Read the specific activity data of the nuclides in the activation corrosion products of the reactor primary loop and perform data cleaning;

[0009] Step 3: Normalize the operating parameters used as input and the specific activity of the target nuclide used as output in the recurrent neural network;

[0010] Step 4: Divide the historical dataset into training and test sets;

[0011] Step 5: Determine the recurrent neural network structure for predicting the source terms of activated corrosion products;

[0012] Step 6: Train the neural network using the training set and test the neural network using the test set.

[0013] Preferably, the reactor historical operation data cleaning in step 1 includes removing duplicate data, interpolating missing data, and removing abnormal data at 1-day intervals.

[0014] Preferably, the data cleaning of the specific activity of the reactor primary loop activation corrosion product nuclides in step 2 includes: removing duplicate data at 1-day intervals, interpolating missing data, and removing abnormal data.

[0015] Preferably, in step 2, for cases where the interval between recording the specific activity data of nuclides is greater than one day, a linear interpolation method is used to fit the missing data.

[0016] Preferably, the normalization formula for the reactor operating parameters in step 3 is as follows:

[0017]

[0018] In the formula: i is the reactor operating parameter number; x i x is the value of reactor operating parameter i; i,min x is the historical minimum value of reactor operating parameter i; i,max x represents the historical maximum value of reactor operating parameter i; i,norm This represents the normalized value of reactor operating parameter i;

[0019] The formula for normalizing the specific activity of the target nuclide is as follows:

[0020]

[0021] In the formula: j is the nuclide number; y j y represents the specific activity value of the target nuclide. j,min The minimum specific activity of the target nuclide; y j,max The maximum specific activity of the target nuclide; y j,norm This is the normalized value of the specific activity of the target nuclide.

[0022] Preferably, in step 4, the historical dataset is divided such that the first 60% to 80% is used as the training set and the remainder as the test set.

[0023] Preferably, in step 5, the recurrent neural network structure for predicting the source terms of activated corrosion products is determined using an open-loop network structure. The network structure parameters to be determined include the types of historical operating parameters, the types of target nuclide specific activities, the number of days of input historical parameters, the number of hidden layer nodes, and the neural network activation function.

[0024] Preferably, in step 5, when predicting the source term of activated corrosion products several days later, it is assumed that the reactor operating parameters remain unchanged, and the predicted value output by the neural network is used as the input.

[0025] Preferably, when predicting the source term of activated corrosion products several days in the future, an offset of the prediction days can be preset between the input and output of the neural network, and the source term several days in the future can be directly predicted using the current data.

[0026] Preferably, in step 6, the mean squared error decreases by more than two orders of magnitude when training the neural network.

[0027] The advantages of this invention are as follows: Traditional activation corrosion product source term prediction models require modeling based on the generation, transport, and deposition mechanisms of corrosion products in the reactor. After modeling, numerical methods are needed to solve a system of differential equations to predict the changing trends of corrosion product source terms. This process is computationally complex and requires significant computational resources. This invention uses a recurrent neural network to predict the changing trends of activation corrosion product source terms. Only the network training process requires certain computational resources; the computational resources required for subsequent source term prediction are negligible. Compared to traditional methods, it does not require solving a system of equations, resulting in faster prediction speed and fewer computational resources. Furthermore, traditional activation corrosion product source term prediction models lack the use of historical reactor operation data and cannot combine the specific characteristics of the target reactor for source term prediction. This invention uses a recurrent neural network to predict activation corrosion product source terms, using the historical operating data of the target reactor as neural network input. Compared to traditional models, this better fits the specific characteristics of the reactor and has higher prediction accuracy. Traditional activation corrosion product source term prediction methods suffer from uncertain model parameters, requiring the setting of hypothetical values ​​based on experience, leading to significant uncertainty in the prediction results. The source term prediction model in this invention does not require pre-input parameters. After the target reactor is determined, the model can be adaptively generated for predicting the source terms of activation corrosion products. The uncertainty of the prediction results is much smaller than that of traditional methods. Attached Figure Description

[0028] Figure 1 A schematic diagram of the recurrent neural network structure in the pressurized water reactor activation corrosion product source term prediction method based on recurrent neural network provided by the present invention;

[0029] Figure 2 This is a schematic diagram of the recurrent neural network structure in Example 1;

[0030] Figure 3 This is a schematic diagram of the recurrent neural network structure in Example 2. Detailed Implementation

[0031] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0032] A method for predicting source terms of activation corrosion products in pressurized water reactors based on recurrent neural networks includes the following steps:

[0033] Step 1: Read historical data of reactor operating parameters (including core power, discharge flow rate, coolant pH, coolant dissolved oxygen concentration, coolant dissolved hydrogen concentration, coolant boron concentration, etc.) and perform data cleaning at certain time intervals, such as one day. Data cleaning includes removing duplicate data, interpolating missing data, and removing outlier data.

[0034] Step 2: Read the historical monitoring data of the specific activity of the activated corrosion products nuclides in the reactor primary coolant, and clean the data at certain time intervals, such as one day. For cases where the data recording interval of the specific activity of nuclides is greater than one day, use the linear interpolation method to fit the missing data.

[0035] Step 3: Select as follows Figure 1 The recurrent neural network shown is normalized using its input operating parameters and its output target nuclide specific activity.

[0036] The input runtime parameters are normalized using the following formula:

[0037]

[0038] In the formula, k is the reactor operating parameter number, x k x is the value of reactor operating parameter k. k,min x is the historical minimum value of reactor operating parameter k. k,max x is the historical maximum value of reactor operating parameter k. k,norm This represents the normalized value of the reactor operating parameter k.

[0039] When normalizing reactor operating parameters, the corresponding reactor operating parameter limits can be referenced, and x can be used as a reference. k,min Set x to the lower limit of the reactor operating parameters. k,max Set to the upper limit value of the reactor operating parameters;

[0040] The specific activity of the target nuclide was normalized using the following formula:

[0041]

[0042] In the formula, l is the nuclide number; y l y represents the specific activity value of the target nuclide. l,min The minimum specific activity of the target nuclide; y l,max The maximum specific activity of the target nuclide; y l,norm This is the normalized value of the specific activity of the target nuclide.

[0043] Step 4: Divide the historical dataset into training and test sets.

[0044] When dividing a dataset into training and testing sets, the data should be sorted according to the measurement time order, and the ratio of training to testing sets should be determined based on the total amount of data in the dataset.

[0045] The first 60% to 80% is used as the training set, and the last 40% to 20% is used as the test set.

[0046] Step 5: Determine the recurrent neural network structure for predicting the source terms of activated corrosion products.

[0047] Among them, the recurrent neural network that predicts the source terms of activated corrosion products can use ordinary RNN or LSTM, etc. The network structure uses an open-loop structure, that is, the input of the recurrent neural network is the historical operating parameters of the reactor and the historical monitoring values ​​of the specific activity of nuclides, rather than the historical operating parameters of the reactor and the output values ​​of the recurrent neural network.

[0048] The main structural parameters of the recurrent neural network that need to be determined include: the types of historical operating parameters, the types of target nuclide specific activities, the number of days for inputting historical parameters, the number of hidden layer nodes, and the neural network activation function. Preferably, the number of days for inputting historical parameters is 3 days, and the tanh function is used as the neural network activation function.

[0049] When predicting the source terms of activated corrosion products several days in advance, the reactor operating parameters are assumed to remain unchanged, and the predicted values ​​output by the neural network are used as input.

[0050] When predicting the source terms of activated corrosion products several days in advance, an offset of the prediction days can be pre-set between the input and output of the neural network, and the source terms several days in advance can be directly predicted using the current data.

[0051] Step 6: Train the neural network using the training set.

[0052] The training set is input into the neural network, and the output of the neural network is obtained after forward propagation.

[0053] The loss function value of the neural network is calculated based on its output and the training set. Preferably, the loss function is the mean squared error (MSE).

[0054] The gradient of the loss function with respect to each input parameter is calculated through backpropagation. After gradient calculation, an optimization algorithm is used to update the neural network weights. Preferably, the mean squared error decreases by more than two orders of magnitude during neural network training.

[0055] Step 7: Test the neural network performance using the test set and optimize the training parameters. Repeat Step 6 until the neural network performance meets the requirements.

[0056] Example 1

[0057] A method for predicting source terms of activation corrosion products in pressurized water reactors based on recurrent neural networks includes the following steps:

[0058] Step 1: Backtrack the historical parameters of reactor operation within 400 days from the current time, remove abnormal monitoring data, and interpolate to make the data time interval 1 day.

[0059] Step 2: Backtrack the primary loop nuclide specific activity monitoring values ​​within 400 days from the current time, remove abnormal monitoring data, and interpolate to make the data time interval 1 day.

[0060] Step 3: Select coolant pH, discharge flow rate, reactor power, and dissolved oxygen concentration as inputs. 58 Co and 60 Co is used as the target nuclide, and its specific activity is used as the output. The input parameters are normalized according to their operating limits.

[0061]

[0062] In the formula, m is the reactor operating parameter number; x m These are reactor operating parameter values; x m,low This represents the lower limit of the reactor operating parameters; x m,high This represents the upper limit of the reactor operating parameters; x m,norm These are the normalized values ​​of the reactor operating parameters. For output parameters, normalization is performed based on their monitored maximum and minimum values:

[0063]

[0064] In the formula, n is the nuclide number; y n nuclide 58 Co and 60 Co specific activity value; y n,min for 58 Co and 60 Minimum specific activity of Co; y n,max for 58 Co and 60 Maximum specific activity of Co; y n,norm for 58 Co and 60 The normalized value of Co specific activity.

[0065] Step 4: Sort the 400 days of data in chronological order. The first 300 days of data will be used for neural network training, and the last 100 days of data will be used for neural network testing and verification.

[0066] Step 5: Using reactor operating parameters and nuclide specific activity monitoring data from the past 3 days, predict the nuclide specific activity values ​​for the next 3 days, and construct a recurrent neural network for predicting the source terms of activation corrosion products, as shown in the following structure. Figure 2 As shown. When predicting future source terms, it is assumed that the reactor operating parameters remain constant, and the predicted value output by the neural network is used as input. For example, based on historical data I... t-3 O t-3 I t-2 O t-2 I t-1 O t-1 The predicted specific activity of the nuclide at time t is O. t Theoretical prediction of nuclide specific activity at time t+1 requires measured operating parameters and specific activity at time t. Since measured values ​​are not available at time t, It is used. t-1 As the measured value of the operating parameter, the predicted value O is used. t Based on the measured value of the specific activity of the nuclide, the specific activity value of the nuclide at time t+1 is predicted. The number of nodes in the hidden layer of the recurrent neural network is set to 30, and the tanh function is used as the activation function of the neural network.

[0067] Step 6: Train the neural network using the ADAM optimizer and test its performance on the test set.

[0068] Example 2

[0069] A method for predicting source terms of activation corrosion products in pressurized water reactors based on recurrent neural networks includes the following steps:

[0070] Step 1: Backtrack the historical parameters of reactor operation within 600 days from the current time, remove abnormal monitoring data, and interpolate to make the data time interval 1 day.

[0071] Step 2: Backtrack the primary loop nuclide specific activity monitoring values ​​within 600 days from the current time, remove abnormal monitoring data, and interpolate to make the data time interval 1 day.

[0072] Step 3: Select coolant pH value, discharge flow rate change rate, reactor power, dissolved hydrogen concentration, and zinc concentration as inputs, and select... 58 Co and 60 Co is used as the target nuclide, and its specific activity is used as the output. The input parameters are normalized according to their operating limits.

[0073]

[0074] In the formula: p is the reactor operating parameter number; x p These are reactor operating parameter values; x p,low This represents the lower limit of the reactor operating parameters; x p,highThis represents the upper limit of the reactor operating parameters; x p,norm These are the normalized values ​​of the reactor operating parameters. For output parameters, normalization is performed based on their monitored maximum and minimum values:

[0075]

[0076] In the formula: q is the nuclide number; y q nuclide 58 Co and 60 Co specific activity value; y q,min for 58 Co and 60 Minimum specific activity of Co; y q,max for 58 Co and 60 Maximum specific activity of Co; y q,norm for 58 Co and 60 The normalized value of Co specific activity.

[0077] Step 4: Sort the 600 days of data in chronological order. The first 500 days of data will be used for neural network training, and the last 100 days of data will be used for neural network testing and verification.

[0078] Step 5: Construct the structure as follows Figure 3 The recurrent neural network shown is used for source term prediction. There is a 3-day offset between the reactor's historical operating parameters, nuclide specific activity monitoring values, and the nuclide specific activity values; that is, the values ​​at times t-3, t-4, and t-5 are used to predict the value at time t. Under this network structure, historical data can be used to directly predict the source term of activation corrosion products 3 days later: neural network input I. t-3 O t-3 I t-2 O t-2 I t-1 O t-1 Directly output O t+2 The number of hidden layer nodes in the recurrent neural network is set to 30, and the tanh function is used as the activation function.

[0079] Step 6: Train the neural network using the SGD optimizer and test its performance on the test set.

[0080] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for predicting source terms of activated corrosion products in pressurized water reactors based on recurrent neural networks, characterized in that, Includes the following steps: Step 1: Read historical reactor operating data and perform data cleaning; Step 2: Read the specific activity data of the nuclides in the activation corrosion products of the reactor primary loop and perform data cleaning; Step 3: Normalize the operating parameters used as input and the specific activity of the target nuclide used as output in the recurrent neural network; The normalization formula for the reactor operating parameters in step 3 is as follows: ， In the formula: i Number the reactor operating parameters; x i Reactor operating parameters i The value; x i,min Reactor operating parameters i The historical minimum value; x i,max Reactor operating parameters i The historical maximum value; x i,norm Reactor operating parameters i The normalized value; The formula for normalizing the specific activity of the target nuclide is as follows: ， In the formula: j Number the nuclide; y j The specific activity value of the target nuclide; y j,min The minimum specific activity of the target nuclide; y j,max The maximum specific activity of the target nuclide; y j,norm This is the normalized value of the specific activity of the target nuclide; Step 4: Divide the historical dataset into training and test sets; Step 5: Determine the recurrent neural network structure for predicting the source terms of activated corrosion products; In step 5, the recurrent neural network structure for predicting the source terms of activated corrosion products is determined. An open-loop network structure is used, that is, the input of the recurrent neural network is the historical operating parameters of the reactor and the historical monitoring values ​​of the specific activity of the nuclide, rather than the output value of the recurrent neural network. The network structure parameters to be determined include the types of historical operating parameters, the types of specific activities of the target nuclide, the number of days of input historical parameters, the number of hidden layer nodes, and the activation function of the neural network. Step 6: Train the neural network using the training set and test the neural network using the test set.

2. The method for predicting source terms of activated corrosion products in pressurized water reactors based on recurrent neural networks as described in claim 1, characterized in that: The reactor historical operation data cleaning in step 1 includes removing duplicate data, interpolating missing data, and removing abnormal data at 1-day intervals.

3. The method for predicting source terms of activated corrosion products in pressurized water reactors based on recurrent neural networks as described in claim 1, characterized in that: The reactor primary loop activation corrosion product nuclide specific activity data cleaning in step 2 includes: removing duplicate data at 1-day intervals, interpolating missing data, and removing abnormal data.

4. The method for predicting source terms of activated corrosion products in pressurized water reactors based on recurrent neural networks as described in claim 3, characterized in that: In step 2, for cases where the interval between recording the specific activity data of nuclides is greater than one day, a linear interpolation method is used to fit the missing data.

5. The method for predicting source terms of activated corrosion products in pressurized water reactors based on recurrent neural networks as described in claim 1, characterized in that: In step 4, the historical dataset is divided into two parts: the first 60% to 80% is used as the training set, and the remainder is used as the test set.

6. The method for predicting source terms of activated corrosion products in pressurized water reactors based on recurrent neural networks as described in claim 1, characterized in that: In step 5, when predicting the source term of activated corrosion products several days later, it is assumed that the reactor operating parameters remain unchanged, and the predicted value output by the neural network is used as the input.

7. The method for predicting source terms of activated corrosion products in pressurized water reactors based on recurrent neural networks as described in claim 1, characterized in that: In step 5, when predicting the source term of activated corrosion products several days later, an offset of the prediction days is preset between the input and output of the neural network, and the source term several days later is directly predicted using the current data.

8. The method for predicting source terms of activated corrosion products in pressurized water reactors based on recurrent neural networks as described in claim 1, characterized in that: In step 6, the mean squared error decreases by more than two orders of magnitude when training the neural network.

Images