Optimization method of nuclear power import and export structure based on neural network and dynamic particle swarm

Abstract

The present invention belongs to the technical field of nuclear power inlet and outlet design and discloses a method for optimizing nuclear power inlet and outlet structures based on a neural network and a dynamic particle swarm optimization algorithm. The method comprises the following steps: 1. establishing a finite element model of a nuclear power integrated inlet and outlet based on fluid mechanics; wherein the geometric parameters of the finite element model include the outer tube turning radius, the angle between the inner tube and the horizontal direction, and the inner tube turning radius; 2. constructing a neural network model based on the characteristics of nuclear power pipe parameters to map the relationship between the geometric parameters and the pressure drop; and 3. searching for the optimal geometric parameter combination based on the dynamic particle swarm optimization algorithm and the neural network model, and optimizing the nuclear power integrated inlet and outlet structure based on the optimal geometric parameter combination. The present invention integrates the neural network and dynamic particle swarm optimization algorithms to optimize the nuclear power integrated inlet and outlet structure, reducing flow resistance, meeting the requirements of more efficient and reliable nuclear power reactions, while also improving optimization efficiency and reducing computational costs.

Description

Technical Field

[0001] The present invention belongs to the technical field of nuclear power import and export design, and specifically relates to a nuclear power import and export structure optimization method based on neural network and dynamic particle swarm. Background Art

[0002] As a clean, efficient, stable, and low-carbon energy source, nuclear power plays an irreplaceable role in ensuring energy security. The reactor is the core equipment for nuclear energy utilization, and the integrated inlet and outlet pipes, as key components of the reactor, have a flow path design that directly impacts the reactor's safety and operational efficiency. Optimizing the flow path design within the reactor can reduce flow resistance, ensure stable and uniform coolant flow, prevent local overheating, and thus reduce the risk of core damage. Optimizing flow resistance reduces energy consumption in the main circulation pump, lowering operating costs. Proper flow path design also extends equipment life and improves the overall economic efficiency of the reactor. Therefore, the design of the integrated inlet and outlet pipes for nuclear power plants is crucial and a key step in ensuring the safe and efficient operation of the reactor.

[0003] In the flow channel optimization study, Shi Fang and others used the finite element method to systematically analyze the influence of the structural parameters of the transition zone of the pressure vessel nozzle on the stress distribution; the study showed that increasing the nozzle wall thickness, optimizing the internal and external transition fillet diameters, reducing the nozzle opening size, and using welding reinforcement rings or fins can effectively reduce the stress intensity factor and maximum equivalent stress at the crack tip, thereby improving structural reliability. In order to adapt to the development trend of miniaturization of nuclear reactors, Chen Minrui and others proposed four integrated inlet and outlet nozzle design schemes, and compared and analyzed the flow field characteristics of each scheme based on numerical simulation methods. The results show that scheme three has the smallest total pressure drop and the best flow characteristics, which is not only conducive to reducing the system's operating energy consumption, but also improving the reactor's energy conversion efficiency; this study provides an important reference for the design of reactor miniaturization.

[0004] However, current research still faces significant challenges. Integrated nozzle design involves the complex coupling of multiple physical fields, including temperature, flow, and stress fields, requiring multi-parameter collaborative optimization to find the optimal design solution. However, traditional calculation methods suffer from high computational costs, long optimization cycles, and a tendency to fall into local optimality, making them difficult to meet the demands of rapid iteration and lean design in engineering practice. Summary of the Invention

[0005] The purpose of the present invention is to propose a nuclear power inlet and outlet structure optimization method based on neural network and dynamic particle swarm. This method integrates the neural network and dynamic particle swarm optimization algorithms to optimize the integrated nuclear power inlet and outlet structure, reduces the flow resistance, meets the needs of more efficient and reliable nuclear power reactions, and at the same time improves the optimization efficiency and reduces the computational cost.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for optimizing the import and export structure of nuclear power plants based on neural networks and dynamic particle swarms includes the following steps:

[0008] Step 1: Establish a finite element model of the nuclear power integrated inlet and outlet based on fluid mechanics; wherein the geometric parameters of the finite element model of the nuclear power integrated inlet and outlet include the radius of the outer tube turning section, the angle between the inner tube and the horizontal direction, and the inner tube corner radius;

[0009] Step 2: Construct a neural network model based on the characteristics of nuclear power takeover parameters to map the relationship between geometric parameters and voltage drop; the neural network model includes an input layer, three hidden layers, and an output layer; the input layer includes three nodes, corresponding to three geometric parameters respectively; the number of nodes in the three hidden layers is 64, 32, and 8, respectively; the first two hidden layers are connected to a batch normalization layer, and a dropout layer is connected to the batch normalization layer; the output layer is a single-node fully connected layer;

[0010] Step 3: Search for the optimal geometric parameter combination based on the dynamic particle swarm optimization algorithm and neural network model, and optimize the nuclear power integrated import and export structure according to the optimal geometric parameter combination;

[0011] The optimal geometric parameter combination is searched through the following steps:

[0012] Step 3-1: Initialize the particle swarm, where each particle represents a set of geometric parameters; set the particle swarm size N, the maximum number of iterations Tmax, and the baseline value of the dimensional decoupling parameter, and randomly initialize the particle position and velocity to ensure that the initial solution meets the geometric constraints;

[0013] Step 3-2: Initialize the inertia weight ω, learning factors c1 and c2 of the neural network model established in step 2; train the neural network model based on the particle swarm to predict the pressure drop; during the iteration process, the optimal ten particles are evaluated by particle fitness, and the true value is calculated and updated to the neural network model;

[0014] The particle fitness ; is the predicted pressure drop; is the geometric parameter penalty; Pressure drop penalty;

[0015] Step 3-3: Dynamically adjust the inertia weight ω, learning factors c1 and c2 of the neural network model; update the particle velocity according to the dimensional decoupling formula, and correct the particle position according to the updated particle velocity;

[0016] Dimension decoupling formula:

[0017] ;

[0018] Where, is the particle index, is the dimension index, is the current iteration number, For particles In dimension On the The speed of iterations, For particles In dimension On the The position of the iteration; For particles In dimension The best historical position on The global optimal particle In dimension Position on For the The inertia weight of the iteration, 、 is the individual learning factor and social learning factor of dimension k; 、 is a random number on dimension k;

[0019] Step 3-4: Update particle velocity and position to ensure that geometric constraints are met;

[0020] If the iteration t reaches the maximum number of iterations Tmax or the global optimal solution has not been improved for 20 consecutive generations, the iteration is terminated and the optimal geometric parameter combination and its corresponding pressure drop are output.

[0021] Furthermore, the height of the outer tube expansion section and the inner tube turning radius of the nuclear power integrated inlet and outlet finite model are determined by the outer tube turning section radius, the angle between the inner tube and the horizontal direction, and the inner tube turning corner radius.

[0022] Furthermore, the optimized range of the outer tube turning radius is 295~465mm, the optimized range of the inner tube corner radius is 219~405mm, and the optimized range of the angle between the inner tube and the horizontal direction is 0°~70°.

[0023] Furthermore, the loss function of the neural network model is , 、 、 are the true values ​​of the outer tube turning radius, the inner tube angle with the horizontal direction, and the inner tube turning radius. 、 、 They are respectively the predicted values ​​of the outer tube turning radius, the angle between the inner tube and the horizontal direction, and the inner tube turning radius.

[0024] Furthermore, the geometric parameter penalty The closed-formula method directly calculates whether the geometric parameters satisfy the feasible domain, and imposes a maximum fixed penalty term on particles that violate the constraints to force the exclusion of solutions that cannot be physically assembled.

[0025] ; y is the geometric parameter that the particle reacts to.

[0026] Furthermore, the voltage drop penalty To provide progressive penalty for fluid performance, a quadratic penalty function is used to gradually increase the constraint strength for particles that meet geometric constraints but exceed the pressure drop limit.

[0027] , is the pressure drop predicted by the neural network model, is the adaptive coefficient that changes with the number of iterations, is the initial coefficient, Tmax is the maximum number of iterations, is the current iteration number.

[0028] Furthermore, the inertia weight ω adopts a time linear decreasing strategy, , is the adjusted inertia weight, and are the maximum and minimum inertia weights, Tmax is the maximum number of iterations, is the current iteration number.

[0029] Furthermore, the learning factors c1 and c2 are dynamically adjusted according to the dimensional characteristics. , To learn the dimensions of factors c1 and c2 Upper adjustment value; is the current particle swarm dimension The standard deviation on is the mean of the standard deviation of each dimension, is the baseline learning factor.

[0030] Furthermore, the geometric constraints are:

[0031] ,

[0032] Where H01 is the height of the outer tube expansion section, R01 is the radius of the outer tube turning section, TH0 is the angle between the inner tube and the horizontal direction, R02 is the inner tube corner radius, and R03 is the inner tube turning radius; HH is the projection distance of the inner tube outlet center in the Y direction in the XY plane, RR is the motion trajectory radius of the inner tube outlet center, and LL is the distance between the outer tube inlet and the inner tube outlet in the XY plane.

[0033] The present invention has the following beneficial effects:

[0034] (1) A finite element numerical fluid simulation model was established based on COMSOL simulation software, and the parameters of the integrated nozzle structure were optimized by combining a neural network agent model and a particle swarm optimization algorithm. This reduces the nozzle flow resistance, achieving a more efficient and reliable nuclear power reaction, while improving optimization efficiency and reducing computational costs.

[0035] (2) A neural network model replaces high-cost simulation. Traditional finite element simulation (such as COMSOL) takes several hours to calculate a single flow resistance, while the neural network model (FCN) only takes milliseconds to predict, greatly reducing the computational cost. Through a fully connected network with a 64-32-8 node structure (including BatchNorm and Dropout), high accuracy (MSE loss function) is maintained even with small samples. Combined with dynamic PSO to accelerate convergence, dynamic inertia weights and dimensional decoupling learning factors, the number of iterations is reduced by about 30% compared to traditional PSO (convergence within 200 generations).

[0036] (3) Excellent optimization performance: Dynamic PSO uses a high inertia weight (ω≈1.1) in the early stages to promote global exploration, while a low weight (ω≤0.2) in the later stages focuses on local development, avoiding falling into local optimality. The final optimized parameter combination (R01=465mm; THO=65.6°; R02=396.864mm) reduces the pressure drop to 4800Pa, a reduction of 3260Pa (40.4%) compared to the initial design. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the process of the present invention.

[0038] Figure 2 This is a schematic diagram of the geometric parameters of the nuclear power integrated inlet and outlet of the present invention.

[0039] Figure 3 This is the pressure and flow rate diagram of the simulation results of the present invention.

[0040] Figure 4 Schematic diagrams of the nuclear power integrated import and export structures before and after optimization of the present invention, where (a) is the structural schematic diagram before optimization and (b) is the structural schematic diagram after optimization. DETAILED DESCRIPTION

[0041] like Figure 1 As shown, this embodiment provides a method for optimizing the nuclear power import and export structure based on a neural network and a dynamic particle swarm, comprising the following steps:

[0042] Step 1: Based on fluid mechanics, a finite element model of the nuclear power integrated inlet and outlet is established using COMSOL software. The geometric parameters of the finite element model of the nuclear power integrated inlet and outlet include the height H01 of the outer tube expansion section, the radius R01 of the outer tube turning section, the angle TH0 between the inner tube and the horizontal direction, the inner tube corner radius R02, and the inner tube turning radius R03. The height H01 of the outer tube expansion section and the inner tube corner radius R02 are determined by the radius R01 of the outer tube turning section, the angle TH0 between the inner tube and the horizontal direction, and the inner tube turning radius R03, and the relationship is: , .

[0043] Therefore, the actual geometric parameters that need to be optimized are the outer tube turning radius R01, the angle between the inner tube and the horizontal direction TH0, and the inner tube corner radius R02; Figure 2 As shown in the figure, the optimization range of the outer tube turning radius R01 is 295~465mm, the inner tube corner radius R02 is 219~405mm, and the angle between the inner tube and the horizontal direction (TH0) is 0°~70°.

[0044] Step 2: Construct a neural network model based on the characteristics of nuclear power takeover parameters and engineering experience to map the relationship between geometric parameters and voltage drop; the neural network model includes an input layer, three hidden layers and an output layer.

[0045] The input layer includes three nodes, corresponding to three geometric parameters (the radius of the outer tube turning section R01, the angle between the inner tube and the horizontal direction TH0, and the inner tube corner radius R02).

[0046] The number of nodes in the three hidden layers is 64, 32, and 8, respectively. This decreasing structure is suitable for regression tasks, gradually compressing feature dimensions to avoid the curse of dimensionality. A batch normalization layer is connected after the first two hidden layers to alleviate the gradient explosion problem, stabilize the network activation value distribution, and improve training stability. A dropout layer is then connected to the batch normalization layer with a retention rate of 0.3, effectively reducing the risk of overfitting.

[0047] The output layer is a single-node fully connected layer. Since the pressure drop is a dimensionless scalar parameter, a single output node can directly map the physical quantity. The output layer is used to output the predicted pressure drop. This embodiment uses the mean square error as the optimization target, and the loss function is , 、 、 are the true values ​​of the outer tube turning radius, the inner tube angle with the horizontal direction, and the inner tube turning radius. 、 、 They are respectively the predicted values ​​of the outer tube turning radius, the angle between the inner tube and the horizontal direction, and the inner tube turning radius.

[0048] Step 3: Based on the dynamic particle swarm optimization algorithm and neural network model, efficiently search for the optimal geometric parameter combination, and optimize the nuclear power integrated import and export structure according to the optimal geometric parameter combination.

[0049] The optimal geometric parameter combination is searched through the following steps:

[0050] Step 3-1: Initialize the particle swarm, where each particle represents a set of geometric parameters (the outer tube turning radius R01, the inner tube angle TH0 with the horizontal direction, and the inner tube turning radius R03). Set the particle swarm size N, the maximum number of iterations Tmax, and the baseline value of the dimensional decoupling parameter, and randomly initialize the particle position and velocity to ensure that the initial solution meets the geometric constraints.

[0051] The geometric constraints are ,

[0052] Where H01 is the height of the outer tube expansion section, R01 is the radius of the outer tube turning section, TH0 is the angle between the inner tube and the horizontal direction, R02 is the inner tube corner radius, and R03 is the inner tube turning radius; HH is the projection distance of the inner tube outlet center in the Y direction in the XY plane, RR is the motion trajectory radius of the inner tube outlet center, and LL is the distance between the outer tube inlet and the inner tube outlet in the XY plane.

[0053] Step 3-2: Initialize the parameters of the neural network model established in step 2: inertia weight ω, learning factors c1 and c2; train the neural network model based on particle swarm to predict pressure drop;

[0054] During the iteration process, the evaluated particle fitness is updated to the neural network model to improve the prediction accuracy; the particle fitness ; is the predicted pressure drop; is the geometric parameter penalty; Pressure drop penalty.

[0055] The geometric parameter penalty Whether the geometric parameters satisfy the feasible domain is directly calculated through closed-formula, and a maximum fixed penalty term is imposed on particles that violate the constraints to force the exclusion of solutions that cannot be physically assembled.

[0056] Right now ; is the geometric parameter penalty, which is the main penalty for particles in the hierarchical dynamic penalty function; y is the geometric parameter reflected by the particle.

[0057] The voltage drop penalty In order to gradually penalize the fluid performance, a quadratic penalty function is used to gradually increase the constraint strength for particles that meet the geometric constraints but exceed the pressure drop limit. , is the pressure drop predicted by the neural network model, is the adaptive coefficient that changes with the number of iterations, is the initial coefficient, which is 250 in this embodiment, and Tmax is the maximum number of iterations. is the current iteration number.

[0058] Update the individual and global optimal solutions, compare the current particle fitness with its historical optimal value, and update if it is better. At the same time, select the particle with the best fitness in the population as the global optimal solution.

[0059] Step 3-3: Dynamically adjust the parameters of the neural network model;

[0060] The dynamic inertia weight mechanism is used to regulate the time dimension. The inertia weight ω adopts a time linear decrease strategy. , is the adjusted inertia weight, and are the maximum and minimum inertia weights, Tmax is the maximum number of iterations, is the current iteration number.

[0061] In the early stage (t<50), a high inertia weight ω≈1.0-1.1 is maintained to keep the particles moving at high speed and promote global exploration. In the mid-term transition stage (50≤t<150), ω is linearly reduced from 0.8 to 0.4 to balance exploration and development, and a directed focused search is performed on the assembly size constraint area through orthogonal projection decomposition of the velocity vector. In the late convergence stage (t≥150), a low inertia weight ω≤0.2 is used to enhance the local fine-tuning capability so that the particle cluster is located within the optimal narrow band.

[0062] The learning factors (c1, c2) are dynamically adjusted according to the dimensional characteristics. ,in, To learn the dimensions of factors c1 and c2 The adjusted value, is the k-th dimension standard deviation of the current particle swarm, is the mean of the standard deviation of each dimension, is the baseline learning factor; when hour, , reduce the intensity of following the dimension with sharp fluctuations and strengthen learning in the sensitive dimension; when hour, , promoting its rapid optimization in the convergence direction.

[0063] Update particle velocity according to the dimension decoupling formula, correct particle position according to the updated velocity, and optimize for different parameter characteristics.

[0064] Dimension decoupling formula ,in, is the particle index, is the dimension index, is the current iteration number, For particles In dimension On the The speed of iterations, For particles In dimension On the The position of the iteration; For particles In dimension The best historical position on The global optimal particle In dimension Position on For the The inertia weight of the iteration, 、 is the individual learning factor and social learning factor of dimension k; 、 is a random number in dimension k (between [0,1]).

[0065] Step 3-4: Particle update and convergence judgment;

[0066] Update the particle velocity and position to ensure that the geometric constraints are met; if the maximum number of iterations Tmax is reached or the global optimal solution has not been improved for 20 consecutive generations, the iteration is terminated and the heat map is output, such as Figure 3 As shown, the optimal geometric parameters and their pressure drop values.

[0067] Analysis of the optimized pressure results reveals a lower flow resistance compared to the pre-optimization model. The particle swarm optimization model achieved an outer tube turning radius (R01) of 465 mm, an inner tube turning radius (R02) of 396.864 mm, an inner tube angle (TH0) of 65.628° with respect to the horizontal, an outer tube expansion height (H01) of 525 mm, and an inner tube turning radius (R03) of 1523.1 mm. The minimum pressure drop was 6200 Pa, a reduction of 1860 Pa compared to the pre-optimization model.

[0068] The models before and after optimization are as follows Figure 4 As shown in the figure, compared with before optimization, the flow channel cross-sectional size transition of the expansion section is better, the internal pressure distribution is more uniform, especially the pressure at the elbow is significantly reduced; the velocity distribution is obtained through the velocity distribution, which is faster and more uniform. In general, the optimization goal is achieved well.

[0069] The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any modification and replacement based on the technical solution and inventive concept provided by the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for optimizing nuclear power import and export structure based on neural network and dynamic particle swarm, characterized in that: The steps include: Step 1: Establish a finite element model of the nuclear power integrated inlet and outlet based on fluid mechanics; wherein the geometric parameters of the finite element model of the nuclear power integrated inlet and outlet include the radius of the outer tube turning section, the angle between the inner tube and the horizontal direction, and the inner tube corner radius; Step 2: Construct a neural network model based on the characteristics of nuclear power takeover parameters to map the relationship between geometric parameters and voltage drop; the neural network model includes an input layer, three hidden layers, and an output layer; the input layer includes three nodes, corresponding to three geometric parameters respectively; the number of nodes in the three hidden layers is 64, 32, and 8, respectively; the first two hidden layers are connected to a batch normalization layer, and a dropout layer is connected to the batch normalization layer; the output layer is a single-node fully connected layer; Step 3: Search for the optimal geometric parameter combination based on the dynamic particle swarm optimization algorithm and neural network model, and optimize the nuclear power integrated import and export structure according to the optimal geometric parameter combination; The optimal geometric parameter combination is searched through the following steps: Step 3-1: Initialize the particle swarm, where each particle represents a set of geometric parameters; set the particle swarm size N, the maximum number of iterations Tmax, and the baseline value of the dimensional decoupling parameter, and randomly initialize the particle position and velocity to ensure that the initial solution meets the geometric constraints; Step 3-2: Initialize the inertia weight ω, learning factors c1 and c2 of the neural network model established in step 2; train the neural network model based on the particle swarm to predict the pressure drop; during the iteration process, the optimal ten particles are evaluated by particle fitness, and the true value is calculated and updated to the neural network model; The particle fitness ; is the predicted pressure drop; is the geometric parameter penalty; Pressure drop penalty; Step 3-3: Dynamically adjust the inertia weight ω, learning factors c1 and c2 of the neural network model; update the particle velocity according to the dimensional decoupling formula, and correct the particle position according to the updated particle velocity; Dimension decoupling formula: ； Where, is the particle index, is the dimension index, is the current iteration number, For particles In dimension On the The speed of iterations, For particles In dimension On the The position of the iteration; For particles In dimension The best historical position on The global optimal particle In dimension Position on For the The inertia weight of the iteration, 、 is the individual learning factor and social learning factor of dimension k; 、 is a random number on dimension k; Step 3-4: Update particle velocity and position to ensure that geometric constraints are met; If the iteration t reaches the maximum number of iterations Tmax or the global optimal solution has not been improved for 20 consecutive generations, the iteration is terminated and the optimal geometric parameter combination and its corresponding pressure drop are output.

2. The method for optimizing nuclear power import and export structure based on neural network and dynamic particle swarm according to claim 1 is characterized in that: The height of the outer tube expansion section and the inner tube turning radius of the nuclear power integrated inlet and outlet finite model are determined by the outer tube turning section radius, the angle between the inner tube and the horizontal direction, and the inner tube turning angle radius.

3. The method for optimizing nuclear power import and export structure based on neural network and dynamic particle swarm according to claim 2 is characterized in that: The optimized range of the outer tube turning radius is 295~465mm, the optimized range of the inner tube corner radius is 219~405mm, and the optimized range of the angle between the inner tube and the horizontal direction is 0°~70°.

4. The method for optimizing nuclear power import and export structure based on neural network and dynamic particle swarm according to claim 1 is characterized in that: The loss function of the neural network model is , 、 、 are the true values ​​of the outer tube turning radius, the inner tube angle with the horizontal direction, and the inner tube turning radius, respectively. 、 、 They are respectively the predicted values ​​of the outer tube turning radius, the angle between the inner tube and the horizontal direction, and the inner tube turning radius.

5. The method for optimizing nuclear power import and export structure based on neural network and dynamic particle swarm according to claim 1 is characterized in that: The geometric parameter penalty The closed-formula method directly calculates whether the geometric parameters satisfy the feasible domain, and imposes a maximum fixed penalty term on particles that violate the constraints to force the exclusion of solutions that cannot be physically assembled. ; y is the geometric parameter that the particle reacts to.

6. The method for optimizing nuclear power import and export structure based on neural network and dynamic particle swarm according to claim 1 is characterized in that: The voltage drop penalty To provide progressive penalty for fluid performance, a quadratic penalty function is used to gradually increase the constraint strength for particles that meet geometric constraints but exceed the pressure drop limit. , is the pressure drop predicted by the neural network model, is the adaptive coefficient that changes with the number of iterations, is the initial coefficient, Tmax is the maximum number of iterations, is the current iteration number.

7. The method for optimizing nuclear power import and export structure based on neural network and dynamic particle swarm according to claim 1 is characterized in that: The inertia weight ω adopts a time linear decreasing strategy, , is the adjusted inertia weight, and are the maximum and minimum inertia weights, Tmax is the maximum number of iterations, is the current iteration number.

8. The method for optimizing nuclear power import and export structure based on neural network and dynamic particle swarm according to claim 1 is characterized in that: The learning factors c1 and c2 are dynamically adjusted according to the dimensional characteristics. , To learn the dimensions of factors c1 and c2 Upper adjustment value; is the current particle swarm dimension The standard deviation on is the mean of the standard deviation of each dimension, is the baseline learning factor.

9. The method for optimizing nuclear power import and export structure based on neural network and dynamic particle swarm according to claim 1, characterized in that: The geometric constraints are: ， Where H01 is the height of the outer tube expansion section, R01 is the radius of the outer tube turning section, TH0 is the angle between the inner tube and the horizontal direction, R02 is the inner tube corner radius, and R03 is the inner tube turning radius; HH is the projection distance of the inner tube outlet center in the Y direction in the XY plane, RR is the motion trajectory radius of the inner tube outlet center, and LL is the distance between the outer tube inlet and the inner tube outlet in the XY plane.

Images