Neural network-based axial power offset prediction method for peak shaving process of pressurized water reactor

Abstract

The invention discloses a neural network-based axial power offset prediction method for a peak regulation process of a pressurized water reactor, and the method comprises the steps: 1, arranging power, fuel enrichment degree, the number of burnable poison rods in a fuel assembly, control rod positions, burnup level, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, radial samarium poison distribution and the like according to a specific format; and making a standard input file. Secondly, arranging the AO (axial power offset) value of the reactor core under each target point in a specific format, and making a standard output file; thirdly, building a convolutional neural network model; and 4, putting the standard input file and the standard output file into the model for training until model parameters meet requirements, and exporting the trained model parameters. And 5, making a standard input file of a target point needing to be predicted according to the method in the first step. Sixthly, the standard input file of the target points needing to be predicted is calculated by loading model parameters, and the AO predicted value of each target point is obtained.

Description

Technical Field

[0001] The invention relates to the technical field of commercial pressurized water reactor operation optimization, and in particular to a pressurized water reactor axial power offset (AO) prediction method. Background Art

[0002] With the development of my country's nuclear power industry, not only are there more and more large-scale pressurized water reactor nuclear power plants in service, but there are also more and more pressurized water reactor nuclear power plants under construction, and the proportion of nuclear power in the entire power system is growing rapidly. At the same time, in the context of some "large machines and small grids", in order to ensure the stable operation of the power grid, there is an urgent need for nuclear power to participate in peak regulation. At present, most nuclear power units in China need to frequently participate in power regulation according to grid dispatch. During the peak regulation process, any action of the control rod group may cause xenon oscillation, resulting in a large local power peak and bringing safety risks. Therefore, it is very important to control the axial power offset (AO) during the peak regulation process. At present, nuclear power plant operators generally follow the normal AO control during the peak regulation process, and try to keep the AO within a certain range. However, due to rapid power changes, poison oscillations and other reasons, for the control of the control rod position, in addition to the need for the technical support department to give a plan in advance, it is also extremely dependent on personal experience. This is not only difficult to meet the high-frequency and deep-amplitude peak regulation needs of pressurized water reactor nuclear power units, but also easy to cause human accidents, which is not conducive to the safety of the reactor. If the AO can be predicted intuitively and the AO value at a certain time and power in the future can be obtained, this will be of great benefit to the operators in the main control room. It can assist the operators in controlling the rod position and adjusting the AO, which can further ensure the safety of the core. Summary of the invention

[0003] In order to meet the high-frequency and deep-amplitude peak-shaving needs of my country's commercial pressurized water reactor nuclear power plants, assist operators in rod position control, and further ensure core safety, the present invention proposes a method for predicting axial power offset in the peak-shaving process of a pressurized water reactor based on a neural network. According to the fuel loading information of the nuclear power plant, the historical operation information of the core, and the information such as burnup and poisons calculated by the core physics calculation software, the present invention can accurately and quickly predict the axial power offset (AO) information at a certain moment or power in the future, avoiding the disadvantages of traditional methods such as over-reliance on core physics analysis software, long calculation time, large calculation errors, deviation in the calibration coefficient of the external detector, and extreme reliance on the operator's personal experience.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] A method for predicting axial power deviation in a pressurized water reactor peak regulation process based on a neural network comprises the following steps:

[0006] Step 1: Process input data and make standard input files: First, extract the fuel enrichment and burnable poison rod quantity information of the fuel assembly according to the fuel assembly layout diagram of the unit; then, extract the core power and control rod position information of each target point based on the measured data of the peak-shaving process of the nuclear power plant, where the time step between different target points does not exceed 60 minutes; then, perform power history simulation through the core physics calculation software to extract the burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, and radial samarium poison distribution information at the corresponding time of the peak-shaving process; finally, arrange the core power, fuel enrichment, burnable poison rod quantity, control rod position, burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, and radial samarium poison distribution of all target points in the format of N×X×Y×M to make a standard input file; where N is the number of target points, that is, the number of samples; X represents the number of fuel assemblies in the core x direction; Y represents the number of fuel assemblies in the core y direction; and M represents the number of categories of extracted parameter information;

[0007] Step 2: Create a standard output file: Arrange the AO values ​​of each target point into a column in the format of N×1 and create a standard output file; where N is exactly the same as N in step 1;

[0008] Step 3: Build a convolutional neural network model: The convolutional neural network model is built based on the Inception-Resnet structure, including multiple small convolution kernels, multiple Inception-Resnet modules, a global average pooling layer and multiple fully connected layers;

[0009] Step 4: Training the convolutional neural network model: The standard input file obtained in step 1 and the standard output file obtained in step 2 corresponding to each target point together form a set of training data. All sets of training data are divided into training set, validation set and test set for training the convolutional neural network model. The mean square error function is used as the loss function for training the convolutional neural network model, and the Adam algorithm is used to update the neural network parameters. After the training is completed, the convolutional neural network model parameters are exported and used in the subsequent step 6.

[0010] Step 5: Prepare a standard input file for the target points to be predicted: Prepare a standard input file for the information of the stage to be predicted according to the process in step 1 according to the actual situation; first, extract the fuel enrichment and the number of burnable poison rods of the fuel assembly; then, extract the core power and control rod position information of each target point; then, perform power history simulation through the core physics calculation software, and extract the burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, and radial samarium poison distribution information at the corresponding time of the peak regulation process; finally, arrange the core power, fuel enrichment of the fuel assembly, the number of burnable poison rods, control rod position, burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, and radial samarium poison distribution in the format of N×X×Y×M to prepare a standard input file;

[0011] Step 6: Perform AO prediction: The standard input file of the target point to be predicted in step 5 is calculated through the convolutional neural network model parameters obtained in step 4 to obtain the AO prediction value of each target point.

[0012] Compared with the prior art, the present invention has the following advantages:

[0013] The present invention uses a convolutional neural network to achieve accurate and rapid AO prediction in the process of rapid power changes such as peak regulation in commercial pressurized water reactor nuclear power plants. A single AO prediction only takes tens of milliseconds, avoiding the drawbacks of traditional methods such as large amount of time spent on core physics analysis software calculations, limited calculation accuracy, external detector calibration coefficient deviation, and extreme reliance on the operator's personal experience in the peak regulation process. It can assist the operator in making judgments and controlling rod position adjustments, meeting major requirements for high-frequency and deep-amplitude peak regulation of pressurized water reactor nuclear power units, and is beneficial to ensuring core safety and improving the economy of nuclear power plants. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The figure is a flow chart of the method of the present invention.

[0015] Figure 2 The convolutional neural network structure diagram used in the present invention, wherein (a) is the convolutional neural network flow chart and (b) is its schematic diagram.

[0016] Figure 3 Schematic diagram of the Inception-Resnet model structure used in the present invention, where (a) is a schematic diagram of network structure A, (b) is a schematic diagram of network structure B, (c) is a schematic diagram of network structure C, and (d) is a schematic diagram of the combination method.

[0017] Figure 4 This is a schematic diagram comparing the predicted AO and actual AO of a certain M310 unit during the first phase of peak regulation in January 2023.

[0018] Figure 5 This is a schematic diagram of the changes in power and R rod position during the first stage of peak regulation of a certain M310 unit in January 2023.

[0019] Figure 6 This is a schematic diagram comparing the predicted AO and actual AO of a certain M310 unit during the fourth phase of peak regulation in February 2023.

[0020] Figure 7 This is a schematic diagram of the changes in power and R rod position during the fourth stage of peak regulation of a certain M310 unit in February 2023.

[0021] Figure 8 This is a schematic diagram comparing the predicted AO and actual AO during the only peak-shaving process of a certain M310 unit in May 2023.

[0022] Fig. 9 This is a schematic diagram of the changes in power and R-bar position during the only peak-shaving process of a certain M310 unit in May 2023. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0024] like Figure 1 As shown, the present invention is a method for predicting axial power offset in the peak-shaving process of a pressurized water reactor based on a neural network, which can accurately and quickly predict the core axial power offset (AO) at a certain time point in the future, especially in the dynamic process of rapid power change such as peak-shaving and when neutron poisons such as xenon poisoning and samarium poisoning have not yet reached equilibrium. Predicting AO changes is of great significance for mastering the core axial power distribution and assisting the operator in core control. Specifically, it includes the following steps:

[0025] Step 1: Process input data and create a standard input file. First, extract the fuel enrichment and burnable poison rod quantity information of the fuel assembly according to the fuel assembly layout diagram of the unit; then, find the peak-shaving period based on the measured data of the peak-shaving process of the nuclear power plant. In this embodiment, the peak-shaving period is 4 minutes in step, and a target point is selected every 4 minutes to extract the power and N2, N1, G2, G1 and R rod position information at the corresponding target point; then, perform power history simulation through the core physics calculation software to extract the burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, and radial samarium poison distribution information at the corresponding moment of the peak-shaving process; finally, arrange the core power, fuel enrichment of the fuel assembly, the number of burnable poison rods, control rod position, burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, and radial samarium poison distribution in the format of N×X×Y×M to create a standard input file. Where N is the number of target points, that is, the number of samples; X represents the number of fuel assemblies in the core x direction; Y represents the number of fuel assemblies in the core y direction; M represents the number of categories of parameter information to be extracted. Taking a certain M310 unit as an example, each type of parameter information is input in a 15×15 format. For example, the fuel enrichment information of the fuel assembly is completed into a 15*15 "square" format, and the rows with less than 15 are filled with 0 in the missing positions. The axial xenon poison distribution starts from the first layer and is written backward in the first position of the 15×15 "square", and the vacant positions are filled with 0. The number of parameter information categories is 13, namely, core power, fuel enrichment of fuel assemblies, number of burnable poison rods, N2 rod position, N1 rod position, G2 rod position, G1 rod position, R rod position, burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, radial samarium poison distribution, a total of 13 information, that is, M is 13. Make a standard input file according to the above process.

[0026] Step 2: Create a standard output file. Extract the AO value of each target point selected in step 1, arrange them in a column in the format of N×1, and create a standard output file. N is exactly the same as N in step 1.

[0027] Step 3: Build a convolutional neural network model. The convolutional neural network model is built based on the Inception-Resnet structure. In this embodiment, it mainly includes a 32-channel 1×1 convolution kernel, a 64-channel 1×1 convolution kernel, 3 Inception-Resnet modules, a global average pooling layer, a 64-layer fully connected layer, a 32-layer fully connected layer and a 16-layer fully connected layer. The main structure of the network model is as follows Figure 2 As shown in a and b, the Inception-Resnet module structure is as follows Figure 3As shown in a, b, c, and d. The Inception-Resnet structure is composed of three network structures in series: network structure A, network structure B, and network structure C; network structure A has five series layers, the first layer is a 15*15*C data input layer, where C is the number of channels of the previous network structure; the second layer is four functional blocks in parallel, the first functional block is 32 1×1 convolution kernels, the second is a functional block is 32 2×2 convolution kernels, the third functional block is a layer with 32 1×1 convolution kernels and a 2×2 average pooling layer in series, and the fourth functional block is a layer with 32 1×1 convolution kernels. The first layer is a 15×15×128 data input layer, and the second layer is a 2×2 maximum pooling layer in series; the third layer is a 15×15×128 feature map connection layer; the fourth layer is a ReLU activation function layer; the fifth layer is a 15×15×128 data output layer; the network structure B has ten layers in series, the first layer is a 15*15*128 data input layer; the second layer is divided into three functional blocks in parallel, the first functional block is 64 1×1 convolution kernels, the second functional block is a layer with 64 1×1 convolution kernels and a layer with 32 2×2 convolution kernels in series, the third functional block is a layer with 64 1×1 convolution kernels and a The first layer is a layer with 32 3×3 convolution kernels connected in series; the third layer is a 15×15×128 feature map connection layer; the fourth layer is a ReLU activation function layer; the fifth layer is a layer with 32 1×1 convolution kernels; the sixth layer is a layer with 64 3×3 convolution kernels; the seventh layer is a layer with 128 1×1 convolution kernels; the eighth layer is a 15×15×128 feature map connection layer; the ninth layer is a ReLU activation function layer; the tenth layer is a 15×15×128 data output layer; the first layer is connected to the eighth layer in addition to the second layer; network structure C There are five serial layers in total. The first layer is a 15*15*128 data input layer. The second layer is two functional blocks in parallel. The first functional block is a layer with 32 1×1 convolution kernels and a layer with 64 3×3 convolution kernels connected in series. The second functional block is a layer with 32 1×1 convolution kernels and two layers with 64 3×3 convolution kernels connected in series. The third layer is a 15×15×128 feature map connection layer. The fourth layer is a ReLU activation function layer. The fifth layer is a 15×15×128 data output layer. In addition to being connected to the second layer, the first layer is also connected to the third layer.

[0028] Step 4: Train the convolutional neural network model. The convolutional neural network model uses the mean square error function as the loss function. Taking the M310 unit as an example, a total of 26 peak-shaving data are used as training data, where the training set, validation set, and test set are divided in a ratio of 8:1:1. The Adam algorithm is used to update the convolutional neural network parameters. After the training is completed, the convolutional neural network model parameters are exported and used in the subsequent step 6.

[0029] Step 5: Make a standard input file for the target points that need to be predicted. Make a standard input file for the information of the stage that needs to be predicted according to the process in step 1. First, extract the fuel enrichment and the number of burnable poison rods of the fuel assembly; then, extract the core power and control rod position information of each target point; then, perform power history simulation through the core physics calculation software, and extract the burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, and radial samarium poison distribution information at the corresponding time of the peak regulation process; finally, arrange the core power, fuel enrichment of the fuel assembly, the number of burnable poison rods, control rod position, burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, and radial samarium poison distribution in the format of N×X×Y×M to make a standard input file.

[0030] Step 6: Perform AO prediction. The standard input file of the target point to be predicted in step 5 is calculated through the convolutional neural network model parameters obtained in step 4 to obtain the AO prediction value of each target point.

[0031] in Figure 4 This is a schematic diagram showing the comparison between the predicted AO and the actual AO during the first phase of peak regulation of a certain M310 unit in January 2023. Figure 5 It is a schematic diagram of the relative power and R stick position changing with time during this period; Figure 6 This is a schematic diagram showing the comparison between the predicted AO and the actual AO of a certain M310 unit during the fourth phase of peak regulation in February 2023. Figure 7 It is a schematic diagram of the relative power and R stick position changing with time during this period; Figure 8 This is a schematic diagram showing the comparison between the predicted AO and the actual AO during the only peak regulation period of a certain M310 unit in May 2023. Fig. 9 This is a schematic diagram of the relative power and R position changing with time during this period. Figures 4 to 9 It can be seen that in the process of power and rod position changes, the AO predicted value is consistent with the actual value with a small deviation, which proves that the method of the present invention has a good prediction result for the axial power offset of the PWR.

Claims

1. A method for predicting axial power deviation in a pressurized water reactor peak regulation process based on a neural network, characterized in that: It can accurately and quickly predict the core axial power offset at a certain time point in the future, hereinafter referred to as the core axial power offset is AO; The steps include: Step 1: Process input data and prepare standard input files: First, extract the fuel enrichment and burnable poison rod quantity information of the fuel assembly according to the fuel assembly layout diagram of the unit; then, extract the core power and control rod position information of each target point based on the measured data of the peak-shaving process of the nuclear power plant, where the time step between different target points does not exceed 60 minutes; then, perform power history simulation through the core physics calculation software, and calculate the burnup distribution, axial xenon poison distribution, axial samarium poison distribution, and radial The xenon poisoning distribution and the samarium poisoning distribution information are extracted; finally, the core power, fuel enrichment of the fuel assembly, the number of burnable poison rods, the position of the control rods, the burnup distribution, the axial xenon poisoning distribution, the axial samarium poisoning distribution, the radial xenon poisoning distribution, and the radial samarium poisoning distribution of all target points are arranged in the format of N×X×Y×M to make a standard input file; where N is the number of target points, that is, the number of samples; X represents the number of fuel assemblies in the core x direction; Y represents the number of fuel assemblies in the core y direction; M represents the number of categories of extracted parameter information; Step 2: Create a standard output file: Arrange the AO values ​​of each target point into a column in the format of N×1 and create a standard output file; where N is exactly the same as N in step 1; Step 3: Build a convolutional neural network model: The convolutional neural network model is built based on the Inception-Resnet structure, including multiple small convolution kernels, multiple Inception-Resnet modules, a global average pooling layer and multiple fully connected layers; Step 4: Training the convolutional neural network model: The standard input file obtained in step 1 and the standard output file obtained in step 2 corresponding to each target point together constitute a set of training data; all sets of training data are divided into training set, validation set and test set for training the convolutional neural network model. The mean square error function is used as the loss function for training the convolutional neural network model, and the Adam algorithm is used to update the neural network parameters; After training is completed, the convolutional neural network model parameters are exported and used in the subsequent step 6; Step 5: Prepare a standard input file for the target points to be predicted: Prepare a standard input file for the target points to be predicted according to the process in step 1 based on the actual situation; first, extract the fuel enrichment and burnable poison rod quantity information of the fuel assembly; then, extract the core power and control rod position information of each target point; then, perform power history simulation through the core physics calculation software, and extract the burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, and radial samarium poison distribution information at the corresponding time of the peak regulation process; finally, arrange the core power, fuel enrichment of the fuel assembly, the number of burnable poison rods, control rod position, burnup distribution, axial xenon poison distribution, axial samarium poison distribution, radial xenon poison distribution, and radial samarium poison distribution in the format of N×X×Y×M to prepare a standard input file; Step 6: Perform AO prediction: The standard input file of the target point to be predicted in step 5 is calculated through the convolutional neural network model parameters obtained in step 4, that is, the AO prediction value of each target point is obtained.

2. The method for predicting axial power deviation in a pressurized water reactor peak regulation process based on a neural network according to claim 1, characterized in that: In step 3, the Inception-Resnet structure is constructed by connecting three network structures in series: network structure A, network structure B, and network structure C. The network structure A has five serial layers. The first layer is an X*Y*C data input layer, where X represents the number of fuel assemblies in the core x direction, Y represents the number of fuel assemblies in the core y direction, and C is the number of channels of the previous network structure. The second layer is four functional blocks in parallel. The first functional block is 32 1×1 convolution kernels, the second functional block is 32 2×2 convolution kernels, the third functional block is a layer with 32 1×1 convolution kernels and a 2×2 pooling layer in series, and the fourth functional block is a layer with 32 1×1 convolution kernels and a 2×2 maximum pooling layer in series. The third layer is an X×Y×128 feature map connection layer. The fourth layer is a ReLU activation function layer; the fifth layer is an X×Y×128 data output layer; The network structure B has a total of ten serial layers. The first layer is an X*Y*128 data input layer. The second layer is divided into three functional blocks in parallel. The first functional block is 64 1×1 convolution kernels. The second functional block is a layer with 64 1×1 convolution kernels and a layer with 32 2×2 convolution kernels in series. The third functional block is a layer with 64 1×1 convolution kernels and a layer with 32 3×3 convolution kernels in series. The third layer is an X×Y×128 feature map connection layer. The fourth layer is a ReLU activation function layer; the fifth layer is a layer with 32 1×1 convolution kernels; the sixth layer is a layer with 64 3×3 convolution kernels; the seventh layer is a layer with 128 1×1 convolution kernels; The eighth layer is an X×Y×128 feature map connection layer; The ninth layer is a ReLU activation function layer; the tenth layer is an X×Y×128 data output layer; the first layer is connected to the eighth layer in addition to the second layer; The network structure C has five serial layers, the first layer is the X*Y*128 data input layer; The second layer is two functional blocks in parallel. The first functional block is a layer with 32 1×1 convolution kernels and a layer with 64 3×3 convolution kernels connected in series. The second functional block is a layer with 32 1×1 convolution kernels and two layers with 64 3×3 convolution kernels connected in series. The third layer is an X×Y×128 feature map connection layer. The fourth layer is a ReLU activation function layer; the fifth layer is an X×Y×128 data output layer; The first layer is connected to the third layer in addition to being connected to the second layer.

Images