Searching method and device for boron-free operation of pressurized water reactor

Abstract

The application discloses a search method for a boron-unregulated working condition of a pressurized water reactor, and comprises the following steps: constructing a geometric model of a target pressurized water reactor; obtaining a reactor core parameter data set through an iterative search of a neutron solver and automatic search based on a burnup time corresponding to a position condition of a control rod of the pressurized water reactor; and training a prediction model of unknown reactor core parameters through a back propagation neural network based on the reactor core parameter data set.

Description

Technical Field

[0001] This invention belongs to the field of neutron transport numerical calculation technology, and particularly relates to a search method and apparatus for the unadjusted boron condition of a pressurized water reactor. Background Technology

[0002] To address the aforementioned issues, advanced pressurized water reactors (PWRs) have introduced boron-free load tracking technology. This technology achieves reactivity compensation without frequent boron adjustments, significantly reducing the generation of highly radioactive boron waste liquid. Furthermore, the boron-free load tracking process can be automated through the rod control system, effectively reducing the operator's workload. Under boron-free operating conditions, core operation relies on the real-time movement of control rods to maintain the core's critical state with a constant coolant boron concentration. Therefore, the precise determination of the control rod positions significantly impacts the core's reactivity distribution and boron concentration control.

[0003] Currently, achieving core operation without boron adjustment mainly relies on traditional neutronics calculations. However, since the control rods need to move constantly under boron-free conditions, involving complex core physics processes, traditional neutronics calculations face technical bottlenecks such as long computation time and slow iteration speed when searching for the critical burn-up time corresponding to the control rod positions, making it difficult to meet the needs of efficient calculation and real-time analysis.

[0004] Therefore, to address the problem of low computational efficiency in pressurized water reactors operating without boron adjustment, it is urgent to develop a computational method and device that couples control rod position and burnup time search with neural network prediction to accelerate the computational process of core operation without boron adjustment, significantly reduce computation time, and improve overall computational efficiency. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a search method and apparatus for the uncontrolled boron condition in pressurized water reactors (PWRs). By establishing a geometric model of the target reactor assembly, setting the initial burnup time through an automated search program, and determining the control rod positions, the initial burnup time is then input into a neutronics solver for burnup calculation. This effectively reduces the workload of experimental operators and accelerates the critical search calculation rate for the uncontrolled boron control process in PWRs. Furthermore, the dataset obtained by the automated search program is introduced into a neural network. Leveraging its powerful nonlinear mapping capabilities, the iterative search involving numerous burnup steps, originally relying on automated search methods, is transformed into a data-driven, rapid prediction, further enhancing the critical search calculation rate for the uncontrolled boron control process in PWRs.

[0006] The first aspect of the present invention provides a method for searching for boron-free operating conditions in a pressurized water reactor, comprising: A geometric model of the target pressurized water reactor is constructed. Based on the burnup time corresponding to the position conditions of the pressurized water reactor control rods, the core parameter dataset is obtained through iterative search using a neutronics solver and an automated search. Based on the core parameter dataset, a prediction model for unknown core parameters is trained using a backpropagation neural network.

[0007] Preferably, the step of obtaining the core parameter dataset further includes: S110: Set the initial state of the core without boron adjustment based on material parameters, and determine the position conditions of the pressurized water reactor control rods according to the non-boron load tracking process; S120: Based on the control rod position conditions, the initial burnup time is set through an automated search program, and the neutronics solver is called to perform calculations to extract the critical boron concentration of the core coolant at the end of the burnup step. S130: Based on the burnup time of the current iteration, call the neutronics solver to calculate and extract the critical boron concentration for the corresponding iteration; S140: Based on the critical boron concentration of the current iteration and the critical boron concentration of the previous iteration, update the fuel consumption time step of the next iteration through convergence judgment; S150: Obtain the core parameter dataset by real-time monitoring of the critical boron concentration. The specific rule is: if the control rod position conditions have been traversed and the critical boron concentration is less than 0, output and obtain the core parameter dataset; if the traversal has not been completed, update the next control rod position conditions and jump to S120 for execution.

[0008] Preferably, the step of determining the burnup time step for the next iteration based on convergence further includes: Calculate the absolute difference between the critical boron concentration in the current iteration and the critical boron concentration in the previous iteration; The fuel consumption time step for the next iteration is updated based on a comparison between the absolute difference and a preset threshold. Specifically, if the absolute difference is less than the preset threshold, the convergence condition is met, the fuel consumption time of the current iteration is confirmed as the target fuel consumption time, and the current iteration is terminated. If the absolute difference is greater than or equal to the preset threshold, the fuel consumption time step for the next iteration is updated based on the critical boron concentration of the current iteration and the critical boron concentration of the previous iteration. Specifically, if the critical boron concentration of the current iteration is greater than or equal to the critical boron concentration of the previous iteration, the fuel consumption time step for the next iteration is the current fuel consumption time step plus a preset fixed step. If the critical boron concentration of the current iteration is less than the critical boron concentration of the previous iteration, the fuel consumption time step for the next iteration is the average of the current fuel consumption time step and the fuel consumption time step of the previous iteration.

[0009] Preferably, the step of training a predictive model for unknown core parameters further includes: Construct a backpropagation neural network; Based on the core parameter dataset, a training set and a validation set are obtained by partitioning the dataset. The prediction model is trained using a propagation neural network based on the training set and the validation set.

[0010] Preferably, the step of constructing the backpropagation neural network further includes: A multi-layered backpropagation neural network is constructed, comprising an input layer, hidden layers, and an output layer. The number of nodes in the input layer is adapted to the dimensionality of the core parameter dataset. The hidden layers have several nodes and are configured with activation functions. The hidden layers are five fully connected layers. The output layer has nodes to output the predicted value of the critical boron concentration. The weight parameters of the backpropagation neural network are initialized and updated using Kaiming. The weight parameters follow a normal distribution, and their calculation expression is: In the formula, The number of neurons in the i-th layer is the input.

[0011] Preferably, the activation function is the LeakyReLU function, and the calculation expression is: In the formula, x is the neural network input vector; α is the slope.

[0012] Preferably, during the backpropagation neural network update of weight parameters, the AdamW optimizer with decoupled weight decay is used, and the network weights are updated based on a preset initial learning rate, a preset weight decay coefficient, and preset first-order and second-order momentum exponential decay rates.

[0013] Preferably, the loss function of the backpropagation neural network is calculated based on the mean squared error function and the mean absolute error function, and the expression is as follows: In the formula, Let be the mean square error function. Let y be the mean absolute error function. i Let i be the i-th true critical boron concentration value in the training set. This represents the predicted critical boron concentration output by the neural network. The probability is preset.

[0014] Preferably, the step of training the prediction model using a backpropagation neural network further includes: The Dropout mechanism is used in each forward propagation iteration. The specific rules are as follows: the nodes in the hidden layer of the backpropagation neural network are set to an inactive state by a preset drop probability; the forward propagation calculation and error gradient backpropagation update are completed based on the nodes in the active state; the above training process including node dropping is executed repeatedly until the total number of iterations reaches the preset iteration threshold.

[0015] A second aspect of the present invention provides a search device for a pressurized water reactor operating under non-boron adjustment conditions, comprising: The dataset construction module is used to construct the geometric model of the target pressurized water reactor. Based on the burnup time corresponding to the position conditions of the pressurized water reactor control rods, the core parameter dataset is obtained through iterative search using a neutronics solver and an automated search. The model training module is used to train a prediction model for unknown core parameters using a backpropagation neural network based on the core parameter dataset.

[0016] Because the present invention adopts the above technical solution, it has the following advantages and positive effects compared with the prior art: This invention proposes a search method and apparatus for the boron-free operation of a pressurized water reactor. By establishing a geometric model of the target reactor assembly, setting the initial burnup time through an automated search program, determining the control rod positions, and then inputting the initial burnup time into a neutronics solver for burnup calculation, the invention effectively reduces the burden on experimental operators and accelerates the critical search calculation rate of the boron-free control process of a pressurized water reactor.

[0017] By introducing neural networks into the dataset obtained by the automated search program and leveraging their powerful nonlinear mapping capabilities, the original iterative search involving numerous burnup steps, which relied on automated search methods, was transformed into a rapid prediction based on data-driven approaches. This further improved the critical search calculation rate for the boron-free control process of pressurized water reactors. Attached Figure Description

[0018] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is the main flowchart of the search method for the non-boron adjustment condition of the pressurized water reactor in this invention; Figure 2 This is an embodiment of the critical boron concentration of the core coolant at the end of the extraction burnup step in this invention; Figure 3 This is one embodiment of the learning residual in the present invention; Figure 4 This is one embodiment of the Dropout mechanism in this invention; Figure 5 This is the first embodiment of the search method for the non-boron adjustment condition of pressurized water reactor in this invention; Figure 6 This is a second embodiment of the search method for the non-boron adjustment condition of pressurized water reactor in this invention; Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer from the following description and claims. It should be noted that the drawings are all in a very simplified form and use non-precise ratios, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0020] It should be noted that all directional indicators (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicator will also change accordingly.

[0021] First Embodiment See Figure 1 , Figure 5 and Figure 6 The first aspect of the present invention provides a method for searching for boron-free operating conditions in a pressurized water reactor, comprising: S100: Construct the geometric model of the target pressurized water reactor, and obtain the core parameter dataset through iterative search using a neutronics solver and an automated search based on the burnup time corresponding to the position conditions of the pressurized water reactor control rods. S200: A predictive model for unknown core parameters trained using a backpropagation neural network based on a core parameter dataset.

[0022] The specific process of this method is as follows: S100: Establish the geometric model of the target pressurized water reactor, determine its material parameters, state parameters, and calculation control parameters, and use a suitable neutronics solver to perform neutron transport calculations to obtain various data under the unadjusted boron condition of the reactor core; S200: The automated search program will perform the initial burnup step settings in the neutronics solver, then add the next burnup step, and set the control rod position and burnup time for the burnup step, and perform burnup calculation; After the calculation is completed, extract the boron concentration of the corresponding burnup step, and determine whether the current boron concentration is greater than 0. If so, calculate the difference between the current boron concentration Bi+1 and the absolute value of the boron concentration Bi in the previous burnup step. If the absolute value of the current boron concentration difference is less than 1 ppm, the control rod position search process is completed, and the search calculation for the next burnup step is performed; if it is greater than 1 ppm, the control rod position search process is completed, and the search calculation for the next burnup step is performed; if it is greater than 1 ppm, the control rod position search process is completed, and the search calculation for the next burnup step is performed. A ppm value indicates that the burnup time step is either too small or too large. Therefore, it is necessary to search for the burnup time step at this control rod position again. Then, it is determined whether the current boron concentration Bi+1 is greater than the boron concentration Bi of the previous burnup step. If it is greater, it means the burnup time value is too small, so 5 is added to the current burnup time, and the burnup step calculation is performed again. If the current boron concentration is less than the boron concentration of the previous burnup step, it means the burnup time value is too large. The average burnup time of the two most recent burnup steps is taken as the time value for the next burnup step, and the burnup calculation is performed again. After the calculation is completed, the current boron concentration value is checked, and the calculation is iterated repeatedly until the absolute difference of the boron concentration of the two most recent burnup steps is less than 1 ppm. This completes the burnup time search for the control rod position at this burnup step. Then, the time search for the next burnup step is performed. When the core boron concentration is less than 0, meaning the core coolant no longer needs boron addition, the entire calculation is completed. After obtaining a sufficient dataset through an automated search program, the existing dataset is read by a neural network and divided into training and validation sets. Before training begins, the neural network parameters are initialized, and appropriate hidden layer node numbers, activation functions, loss functions, and optimizers are set. Upon inputting a command to start training, the neural network reads the data from the training set, performs forward propagation based on the initialized parameters to obtain the loss function, and updates the parameters accordingly. Next, the loss function and the number of training iterations are evaluated; if neither meets the criteria, the next iteration is performed. Training ends when either the loss function or the number of training iterations meets the criteria. Validation set data is input into the model for prediction. If the obtained validation set loss function does not meet the criteria, the neural network functions and optimizer are optimized, and the neural network is retrained. If the obtained validation set loss function meets the criteria, the required pressurized water reactor parameter prediction model is obtained. Finding the matching relationship between control rod position and burnup time under non-boron-adjusted operating conditions relies on engineers' manual experience for continuous trial calculations, and each calculation requires calling a neutronics solver for extremely time-consuming low-level physics calculations.This invention avoids blind trial and error through an automated search iterative mechanism, achieving a closed loop of physical optimization. This efficiently and automatically generates high-quality datasets with rigorous physical meaning, reducing time and labor costs. Repeatedly calling the neutronics solver each time is simply unacceptable for real-time requirements. This invention utilizes a backpropagation neural network to instantly output core parameters from the core parameter dataset obtained from S100, given the input conditions. By constructing a high-fidelity physical dataset through automated iterative search and combining it with a neural network model for learning, the core physics solution process is transformed into prediction, balancing the physical accuracy of the data with the real-time requirements of engineering applications.

[0023] See Figure 2 Preferably, the step of obtaining the core parameter dataset further includes: S110: Set the initial state of the core without boron adjustment based on material parameters, and determine the position conditions of the pressurized water reactor control rods according to the non-boron load tracking process; S120: Based on the control rod position conditions, the initial burnup time is set through an automated search program, and the neutronics solver is called to calculate and extract the critical boron concentration of the core coolant at the end of the burnup step. S130: Based on the burnup time of the current iteration, call the neutronics solver to calculate and extract the critical boron concentration for the corresponding iteration; S140: Based on the critical boron concentration of the current iteration and the critical boron concentration of the previous iteration, update the fuel consumption time step of the next iteration through convergence judgment; S150: Obtain the core parameter dataset by real-time monitoring of the critical boron concentration. The specific rule is: if the control rod position conditions have been traversed and the critical boron concentration is less than 0, output and obtain the core parameter dataset; if the traversal has not been completed, update the next control rod position conditions and jump to S120 for execution.

[0024] This invention establishes an automated closed loop through steps S110 to S150, encompassing setting conditions, automatic calculation, automatic extraction, automatic convergence judgment, automatic step size update, and automatic traversal, eliminating the cost of manual trial and error. In the process of approaching the critical boron concentration, a fixed step size can easily lead to crossing the critical point or ineffective calculations in the non-critical region. Step S140 introduces the critical boron concentration from two adjacent iterations for convergence comparison, dynamically and adaptively updating the burnup time step for the next iteration. This enables the search algorithm to capture the extreme critical inflection point before the boron concentration drops to 0. Step S150 sets a dual-termination condition of completing the traversal and the critical boron concentration being less than 0, corresponding to the physical endpoint of the non-boron-adjusted operating condition. The positional conditions include at least the number of pilot rod steps, the overlap ratio between pressurized water reactor control rod groups, and the absolute number of R-rod insertion steps.

[0025] Preferably, the step of determining the burnup time step for the next iteration based on convergence further includes: Calculate the absolute difference between the critical boron concentration in the current iteration and the critical boron concentration in the previous iteration; The fuel consumption time step for the next iteration is updated based on a comparison between the absolute difference and a preset threshold. Specifically, if the absolute difference is less than the preset threshold, the convergence condition is met, the fuel consumption time of the current iteration is confirmed as the target fuel consumption time, and the current iteration is terminated. If the absolute difference is greater than or equal to the preset threshold, the fuel consumption time step for the next iteration is updated based on the critical boron concentration of the current iteration and the critical boron concentration of the previous iteration. Specifically, if the critical boron concentration of the current iteration is greater than or equal to the critical boron concentration of the previous iteration, the fuel consumption time step for the next iteration is the current fuel consumption time step plus a preset fixed step. If the critical boron concentration of the current iteration is less than the critical boron concentration of the previous iteration, the fuel consumption time step for the next iteration is the average of the current fuel consumption time step and the fuel consumption time step of the previous iteration.

[0026] An automated search program sets the initial burnup time, determines the pressurized water reactor (PWR) control rod positions, and then inputs the initial burnup time into the neutronics solver for burnup calculation. The boron concentration is extracted from the current burnup step calculation results, and the absolute difference in boron concentration is evaluated. If it meets a preset standard, the burnup time search for the current PWR control rod position is completed, and the next burnup step begins. If the absolute difference does not meet the preset standard, the burnup time is adjusted based on the comparison between the current and previous boron concentrations, and the burnup calculation is repeated. This process is repeated until the core boron concentration is less than 0, at which point the automated search program ends, generating a core parameter dataset containing burnup time, boron adjustment counts, and boron concentration. By comparing the magnitudes of adjacent boron concentrations, the step size update mode is switched, enabling rapid large-step updates in non-sensitive areas and fine-tuning near the critical zone, thus improving search efficiency. If the current boron concentration is greater than or equal to the previous boron concentration, the burnup time value is too small, and the core has not yet reached the sensitive range for reactivity changes, resulting in a slow decrease in boron concentration. By employing an aggressive scaling strategy with a current step size and a fixed step size, these insignificant flat regions can be quickly skipped, avoiding wasted computing power and significantly accelerating the approach to the critical state. When the current boron concentration is less than the previous boron concentration, it indicates that the core has entered the critical range, and the boron concentration drops rapidly. Continuing to increase the step size at this point is extremely dangerous. This invention uses the average of the current and previous step sizes to reduce the search step size, creating a high-density interception of critical inflection points and ensuring the accuracy of the underlying dataset.

[0027] Preferably, the step of training a predictive model for unknown core parameters further includes: Construct a backpropagation neural network; Training and validation sets are obtained by partitioning the core parameter dataset. A prediction model is trained by propagating a neural network based on the training and validation sets.

[0028] This invention strictly divides the dataset into training and validation sets, ensuring that the model is validated while learning. The core parameters of a pressurized water reactor under non-boron-adjusted operating conditions exhibit extremely strong nonlinear characteristics. By clearly defining the standardized process of first constructing a backpropagation neural network and then partitioning and inputting the data, this ensures that the network structure with its powerful nonlinear mapping capabilities is compatible with the core parameter dataset obtained in step S100.

[0029] Preferably, the step of constructing the backpropagation neural network further includes: A multi-layered backpropagation neural network is constructed, consisting of an input layer, hidden layers, and an output layer. The number of nodes in the input layer is adapted to the dimensionality of the core parameter dataset. The hidden layers have several nodes and are configured with activation functions. The hidden layers are five fully connected layers. The output layer has nodes to output the predicted critical boron concentration. The weight parameters of the backpropagation neural network are initialized and updated using Kaiming. The weight parameters follow a normal distribution, and their calculation expression is: In the formula, The number of neurons in the i-th layer is the input.

[0030] This invention employs a five-layer fully connected deep hidden layer structure, significantly expanding the network's feature representation space and enabling it to extract abstract physical laws layer by layer, ensuring a high-precision approximation of complex critical boron concentration variation trends. The number of input layer nodes is strictly adapted to the dataset dimension, and the output layer is dedicated to outputting the critical boron concentration prediction value. This end-to-end customized design ensures that various core operating conditions are handled without information loss or dimensional mismatch when entering the network, and without unnecessary computational overhead at output, achieving optimal physical matching for the entire network's signal transmission channels. This invention uses Kaiming initialization, ensuring it follows a normal distribution, maintaining the variance of each neural network layer at the same level from the start of training. This achieves smooth training without cold-start obstacles, greatly accelerating the convergence speed of the loss function. This neural network has six input vectors and one output vector, with five fully connected hidden layers.

[0031] Preferably, the activation function is the LeakyReLU function, and the calculation expression is: In the formula, x is the neural network input vector; α is the slope.

[0032] In deep learning, the traditional ReLU activation function always outputs zero when the input is negative, causing the gradient of that neuron to remain perpetually zero and never update. The LeakyReLU function introduces a slope α when the input is negative, ensuring a small non-zero gradient remains in the negative interval. This guarantees that all neurons remain active and continue learning under all conditions, fully utilizing the computational potential of a five-layer deep network. LeakyReLU maintains the same linear output characteristics as traditional ReLU. This not only effectively alleviates the gradient vanishing problem easily caused by traditional activation functions such as Sigmoid, accelerating backpropagation computation, but also preserves the sparse activation of the neural network, allowing the model to maintain forward inference efficiency even when faced with massive amounts of core parameter data. In this embodiment, α can be set to 0.1 or other values; this embodiment is not limited to any particular value.

[0033] Preferably, during the backpropagation neural network update of weight parameters, the AdamW optimizer with decoupled weight decay is used, and the network weights are updated based on a preset initial learning rate, a preset weight decay coefficient, and preset first-order and second-order momentum exponential decay rates.

[0034] This invention employs decoupled weight decay, performing the decay operation independently after each weight update, achieving a pure L2 regularization effect. This ensures that the trained model does not overly rely on local data noise, exhibiting stronger and more stable predictive generalization capabilities when facing unknown pressurized water reactor operating conditions. The AdamW optimizer, based on preset first- and second-order momentum exponential decay rates, adaptively adjusts the update step size according to the historical gradient of each weight parameter. This mechanism allows the network to automatically decrease the step size in steep feature spaces and automatically increase the step size in flat spaces, fundamentally avoiding gradient explosion or getting trapped in local optima, achieving smooth and fast convergence under complex loss function landscapes.

[0035] The optimizer uses AdamW: learning rate lr=0.001, preset weight decay weight_decay=1e-4, betas=(0.9,0.999), but other values ​​are also possible. This embodiment does not impose any restrictions.

[0036] Preferably, the loss function of the backpropagation neural network is calculated based on the mean squared error function and the mean absolute error function, and the expression is as follows: In the formula, Let be the mean square error function. Let y be the mean absolute error function. iLet i be the i-th true critical boron concentration value in the training set. This represents the predicted critical boron concentration output by the neural network. The probability is preset.

[0037] This invention employs a combined loss function of Mean Squared Error (MSE) and Mean Absolute Error (MAE). MSE amplifies the error by square, meaning that when the model predicts a large deviation, MSE generates a significant penalty gradient, forcibly pulling back the network weights and ensuring the model avoids wildly inaccurate predictions. MAE, on the other hand, takes the absolute value of the error; when the predicted value is very close to the true value, MAE maintains a stable linear gradient, guiding the model for fine-tuning. When obtaining massive core parameter datasets using neutronics solvers, a few abnormal fluctuations due to computational convergence difficulties are inevitable. This invention introduces the MAE function to neutralize these fluctuations. Because MAE penalizes outliers linearly rather than exponentially, it effectively buffers the impact of outlier data on network weight updates, significantly improving the robustness of the entire prediction model to noisy data. The value used in this embodiment is 0.7, but other values ​​are not limited.

[0038] See Figure 4 Preferably, the step of training the prediction model using a backpropagation neural network further includes: The Dropout mechanism is used in each forward propagation iteration. The specific rules are as follows: the nodes in the hidden layer of the backpropagation neural network are set to an inactive state by a preset dropout probability; the forward propagation calculation and backpropagation update of the error gradient are completed based on the nodes in the active state; the above training process including node dropping is executed repeatedly until the total number of iterations reaches the preset iteration threshold.

[0039] See Figure 4 Preferably, the backpropagation neural network uses the residual connection neural network training method. The essence of residual connections is to transform the neural network learning task from directly fitting the objective function H(x) to learning the residual F(x) = H(x) - x, where x is the neural network input vector. This can solve the gradient vanishing problem that occurs in the training of deep neural networks. For details, see [link to documentation]. Figure 4 .

[0040] By randomly disabling hidden layer nodes, this mechanism effectively mitigates the severe overfitting risk caused by the large parameters of a five-layer fully connected network, forcing the network to learn redundant physical features with greater generalization ability. This mechanism dynamically generates different sub-network structures in each iteration, implicitly achieving multi-model ensemble effects and reducing the variance of prediction results. Ultimately, it ensures that the prediction model can output smooth and stable critical boron concentration values ​​when facing unknown pressurized water reactor conditions. This mechanism prevents overfitting by randomly discarding nodes during neural network training, forcing the entire network to be independent of any single node. A schematic diagram of the forward propagation of the neural network with the Dropout mechanism is shown below; in this neural network, the Dropout mechanism randomly discards nodes in the hidden layer with a probability of p=0.3 in each iteration, and performs at least 80 iterations.

[0041] Second Embodiment A second aspect of the present invention provides a search device for a pressurized water reactor operating under non-boron adjustment conditions, comprising: The dataset construction module is used to construct the geometric model of the target pressurized water reactor. Based on the burnup time corresponding to the position conditions of the pressurized water reactor control rods, the core parameter dataset is obtained through iterative search using a neutronics solver and an automated search. The model training module is used to train a prediction model for unknown core parameters using a backpropagation neural network based on a core parameter dataset.

[0042] The search device of this invention, through a collaborative architecture of a dataset construction module and a model training module, achieves a high degree of decoupling and modular encapsulation of the complex search process that originally relied on manual calculation and intensive physics solutions. It establishes a data link between offline automated physics optimization and online neural network prediction, providing an integrated, efficient, and plug-and-play execution platform for rapid search and accurate prediction of pressurized water reactors under non-boron-adjustment conditions.

[0043] In the description of this application, it should be noted that the terms "inner" and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing this application and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0044] It should also be noted that, unless otherwise explicitly specified and limited, the terms "setup" and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0045] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific identification content executed by the system and device described above can be referred to the corresponding process in the foregoing method embodiments.

[0046] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments. Even if various changes are made to the present invention, if these changes fall within the scope of the claims of the present invention and their equivalents, they shall still fall within the protection scope of the present invention.

Claims

1. A method for searching the non-boron-adjustment operating condition of a pressurized water reactor, characterized in that, include: A geometric model of the target pressurized water reactor is constructed. Based on the burnup time corresponding to the position conditions of the pressurized water reactor control rods, the core parameter dataset is obtained through iterative search using a neutronics solver and an automated search. Based on the core parameter dataset, a prediction model for unknown core parameters is trained using a backpropagation neural network.

2. The search method for the non-boron-adjustment condition of a pressurized water reactor according to claim 1, characterized in that, The steps for obtaining the core parameter dataset further include: S110: Set the initial state of the core without boron adjustment based on material parameters, and determine the position conditions of the pressurized water reactor control rods according to the non-boron load tracking process; S120: Based on the control rod position conditions, the initial burnup time is set through an automated search program, and the neutronics solver is called to perform calculations to extract the critical boron concentration of the core coolant at the end of the burnup step. S130: Based on the burnup time of the current iteration, call the neutronics solver to calculate and extract the critical boron concentration for the corresponding iteration; S140: Based on the critical boron concentration of the current iteration and the critical boron concentration of the previous iteration, update the fuel consumption time step of the next iteration through convergence judgment; S150: Obtain the core parameter dataset by real-time monitoring of the critical boron concentration. The specific rule is: if the control rod position conditions have been traversed and the critical boron concentration is less than 0, output and obtain the core parameter dataset; if the traversal has not been completed, update the next control rod position conditions and jump to S120 for execution.

3. The search method for the non-boron-adjustment condition of a pressurized water reactor according to claim 2, characterized in that, The steps for determining the burn time step for the next iteration based on convergence further include: Calculate the absolute difference between the critical boron concentration in the current iteration and the critical boron concentration in the previous iteration; The fuel consumption time step for the next iteration is updated based on a comparison between the absolute difference and a preset threshold. Specifically, if the absolute difference is less than the preset threshold, the convergence condition is met, the fuel consumption time of the current iteration is confirmed as the target fuel consumption time, and the current iteration is terminated. If the absolute difference is greater than or equal to the preset threshold, the fuel consumption time step for the next iteration is updated based on the critical boron concentration of the current iteration and the critical boron concentration of the previous iteration. Specifically, if the critical boron concentration of the current iteration is greater than or equal to the critical boron concentration of the previous iteration, the fuel consumption time step for the next iteration is the current fuel consumption time step plus a preset fixed step. If the critical boron concentration of the current iteration is less than the critical boron concentration of the previous iteration, the fuel consumption time step for the next iteration is the average of the current fuel consumption time step and the fuel consumption time step of the previous iteration.

4. The search method for the non-boron adjustment condition of a pressurized water reactor according to claim 1, characterized in that, The steps for training a predictive model for unknown core parameters further include: Construct a backpropagation neural network; Based on the core parameter dataset, a training set and a validation set are obtained by partitioning the dataset. The prediction model is trained using a propagation neural network based on the training set and the validation set.

5. The search method for the non-boron adjustment condition of a pressurized water reactor according to claim 4, characterized in that, The steps for constructing a backpropagation neural network further include: A multi-layered backpropagation neural network is constructed, comprising an input layer, hidden layers, and an output layer. The number of nodes in the input layer is adapted to the dimensionality of the core parameter dataset. The hidden layers have several nodes and are configured with activation functions. The hidden layers are five fully connected layers. The output layer has nodes to output the predicted value of the critical boron concentration. The weight parameters of the backpropagation neural network are initialized and updated using Kaiming. The weight parameters follow a normal distribution, and their calculation expression is: In the formula, The number of neurons in the i-th layer is the input.

6. The search method for the non-boron adjustment condition of a pressurized water reactor according to claim 5, characterized in that, The activation function is the LeakyReLU function, and the calculation expression is: In the formula, x is the neural network input vector; α is the slope.

7. The method for searching the non-boron-adjustment condition of a pressurized water reactor according to claim 5, characterized in that, During the backpropagation neural network's weight parameter update process, the AdamW optimizer with decoupled weight decay is used, and the network weights are updated based on the preset initial learning rate, preset weight decay coefficient, and preset first-order and second-order momentum exponential decay rates.

8. The method for searching the non-boron-adjustment condition of a pressurized water reactor according to claim 7, characterized in that, The loss function of the backpropagation neural network is calculated based on the mean squared error function and the mean absolute error function, and is expressed as follows: In the formula, Let be the mean square error function. Let y be the mean absolute error function. i Let i be the i-th true critical boron concentration value in the training set. This represents the predicted critical boron concentration output by the neural network. The probability is preset.

9. The search method for the non-boron adjustment condition of a pressurized water reactor according to claim 4, characterized in that, The step of training the prediction model using a backpropagation neural network further includes: The Dropout mechanism is used in each forward propagation iteration. The specific rules are as follows: the nodes in the hidden layer of the backpropagation neural network are set to an inactive state by a preset drop probability; the forward propagation calculation and error gradient backpropagation update are completed based on the nodes in the active state; the above training process including node dropping is executed repeatedly until the total number of iterations reaches the preset iteration threshold.

10. A search device for a pressurized water reactor operating under non-boron adjustment conditions, characterized in that, include: The dataset construction module is used to construct the geometric model of the target pressurized water reactor. Based on the burnup time corresponding to the position conditions of the pressurized water reactor control rods, the core parameter dataset is obtained through iterative search using a neutronics solver and an automated search. The model training module is used to train a prediction model for unknown core parameters using a backpropagation neural network based on the core parameter dataset.

Images