Method for predicting plate-like element rectangular narrow-gap re-entrainment liquid droplet entrainment behavior

Abstract

The application discloses a method for predicting rectangular narrow-gap re-submerged liquid droplet entrainment behavior of a plate-shaped element, which comprises the following steps: in an offline stage, constructing and initializing a BP neural network, training and optimizing weight values and threshold values by taking experimental boundary conditions as input variables and liquid entrainment rate as an output variable, and predicting in an online stage. The application can rapidly and accurately calculate the liquid entrainment rate value suitable for the re-submerged working condition of the bottom of the plate-shaped element by training the input-output relationship through a large amount of data by using the artificial neural network prediction method.

Description

Method for predicting the entrainment behavior of droplets in rectangular narrow slits of plate-shaped elements Technical Field

[0001] This invention relates to a technology in the field of reactor safety control, specifically a method for predicting the entrainment behavior of reflooded droplets in a rectangular narrow slit of a plate-shaped element based on an artificial neural network. Background Technology

[0002] Large Breach Loss-of-Coolant Accident (LBLOCA) is a potential extreme accident in plate element reactors. Following an accident, coolant loss can expose the reactor core, causing a rapid increase in plate element temperature due to decay heat, potentially leading to plate element meltdown. After the emergency core cooling system is activated, coolant enters the core from the bottom, initiating the bottom reflooding phase. When the coolant contacts the high-temperature cladding surface, boiling occurs, generating steam. Due to the narrow rectangular slots of the plate elements, high-speed steam can entrain coolant downstream, including droplets, liquid fragments, and liquid films. This droplet entrainment behavior enhances heat transfer at the downstream high-temperature wall during the film boiling phase of the quenching front: for example, entrained droplets can enhance heat transfer during collisions with the high-temperature cladding and can also act as radiative heat sinks for the high-temperature wall. Therefore, droplet entrainment behavior during the bottom reflooding phase of the plate elements is crucial for enhancing the cooling performance of the high-temperature cladding and improving the safety characteristics of the plate elements. However, droplet entrainment behavior is difficult to measure locally in experiments, and the mechanism of flow and heat transfer characteristics of liquid entrainment downstream of the quenching front is unclear. Therefore, existing mechanistic models, system programs, and other research methods cannot accurately simulate the liquid entrainment behavior during the bottom resubmersion stage of plate-shaped elements. There is an urgent need to provide a novel calculation method that can quickly and accurately calculate the liquid entrainment rate applicable to the bottom resubmersion conditions of most plate-shaped elements. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies in accurately predicting droplet entrainment behavior during the resubmergence stage of rectangular narrow slits in plate-shaped elements. It proposes a method for predicting droplet entrainment behavior during resubmergence in rectangular narrow slits of plate-shaped elements. This method utilizes an artificial neural network prediction method to train and predict the input-output relationship using a large amount of data, thereby enabling the rapid and accurate calculation of the liquid entrainment rate applicable to most bottom resubmergence conditions of plate-shaped elements.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a method for predicting the entrainment behavior of droplets resubmerged by a rectangular narrow slit in a plate-shaped element. In the offline stage, a BP neural network is constructed and initialized, and the experimental boundary conditions are used as input variables and the liquid entrainment rate is used as output variables. After training and optimizing the weights and thresholds, the network is used for prediction in the online stage.

[0006] The experimental boundary conditions include: inlet cooling water temperature, cooling water flow rate, system pressure, and heating power.

[0007] The prediction method specifically includes:

[0008] 1) Conduct experimental research: Conduct re-submergence experiments on rectangular narrow slit channels of plate-shaped elements to explore the influence of different boundary conditions on droplet entrainment behavior. Only in this way can we determine accurate and appropriate neural network input parameters.

[0009] 2) Expanding the database: Due to the inherent difficulty and destructive nature of the experiments, it is challenging to conduct large-scale experiments. Therefore, this invention utilizes the system program RELAP5 or a self-developed model and program to verify the accuracy of the simulation through experimental data and calculate droplet entrainment data for other operating conditions, thereby expanding the database.

[0010] 3) Training the neural network: The expanded database is grouped, and the network structure and learning rate and other hyperparameters are determined through the training set and validation set, so as to obtain the predicted output results of droplet entrainment based on the input parameters.

[0011] Technical effect

[0012] This invention utilizes experimental research to determine accurate input parameters; verifies the accuracy of numerical simulation methods based on experimental data, and then uses system programs or developed models to calculate and expand the droplet entrainment database; it uses a BP neural network method to train on existing plate-shaped elements with bottom-submerged droplet entrainment, ultimately enabling rapid and accurate calculation of the corresponding droplet entrainment behavior. This invention utilizes artificial neural network algorithms to train and learn on the complex nonlinear physical phenomenon of bottom-submerged droplet entrainment in plate-shaped elements, thereby identifying the relationship between droplet entrainment and input parameters and making accurate predictions: compared to experiments, it significantly reduces working time and energy consumption; compared to program simulation methods, it also improves accuracy and reduces prediction time. Attached Figure Description

[0013] Figure 1 is a schematic diagram of the BP neural network structure of the present invention;

[0014] Figure 2 is a schematic diagram of the effect of the embodiment. Detailed Implementation

[0015] This embodiment relates to a method for predicting the entrainment behavior of droplets in a rectangular narrow slit of a plate-shaped element. By constructing and initializing a BP neural network, four experimental boundary conditions—inlet cooling water temperature, cooling water flow rate, system pressure, and heating power—are used as input variables, and the liquid entrainment rate is used as the output variable. After training and optimizing the weights and thresholds, the method is used for prediction in the online stage.

[0016] The BP neural network includes an input layer, a hidden layer, and an output layer.

[0017] The initialization refers to setting the weights and thresholds of all neurons in the network to random, and normalizing the values ​​of all input data (training group) (normalizing the data to 0.1 to 0.9) to avoid overfitting of the activation function when the values ​​are large.

[0018] As shown in Figure 1, the training specifically includes:

[0019] 1) After inputting a set of experimental boundary condition data, perform calculations on each neuron from front to back for the data input to obtain the net input of the hidden layer neurons;

[0020] The experimental boundary conditions in this invention are based on the re-flooding experimental conditions of plate-shaped elements, and specifically include:

[0021] 1) Experimental studies were conducted under different boundary conditions, and it was found that parameters such as inlet flow rate, system pressure, liquid subcooling, and heating power have a significant impact on droplet entrainment, so they can be used as input parameters for neural networks;

[0022] 2) The specific values ​​are measured in experiments and can be used to train neural network structures, etc.

[0023] 2) Use the Sigmoid function (S-function) as the activation function to calculate the output of the hidden layer neurons until the data of the output layer is obtained;

[0024] Preferably, this invention further adds a term with specific physical significance for the flooding experiment to the loss function, namely, the maximum value of LossP between the difference between "" and "0" (calculated by the neural network output) and "" (the difference between the outlet entrainment amount and the inlet flow rate). This is because a portion of the inlet fluid becomes steam, and the gas contains droplets; the amount of droplets is the entrainment amount. Therefore, in reality, the outlet entrainment amount is always less than the inlet flow rate. However, during training, the neural network output may produce results that violate physical laws. When the neural network produces a result where the outlet entrainment amount is greater than the inlet flow rate, LossP will be greater than 0 (if the neural network produces a result where the outlet entrainment amount is less than the inlet flow rate, this is correct and has no impact). In other words, the loss function of the neural network increases, giving the current neural network training a penalty, guiding the training of the neural network in a direction that conforms to physical laws, so that the final training result of the neural network conforms to physical laws, and the results produced by it will be more meaningful.

[0025] 3) First, calculate the output layer error, and then propagate the error backward from the output layer to calculate the error of all nodes;

[0026] 4) After calculating the total error, the weight thresholds of each layer are corrected using methods such as gradient descent;

[0027] 5) Input another set of experimental boundary condition data and calculate according to the corrected weights and thresholds until the output of all layers in each set of data is less than the specified minimum error, or the number of training iterations reaches the maximum value.

[0028] Twenty calculations were performed for different network structures, and the final average error was calculated. Based on the error calculation results, the learning rate was finally determined to be 0.5, and the network structure was 4-9-5-1. The learning rate parameter was also obtained through calculation and comparison. Choosing 0.5 avoids the problems of slow gradient descent, excessive network training time, getting stuck in local minima and missing the global minimum due to an excessively small learning rate; similarly, it also avoids the problem of the calculation failing to converge due to overshooting the minimum value due to an excessively large learning rate.

[0029] In this embodiment, the Relap5 program, a thermo-hydraulic optimal estimation program, is used to expand the data based on experimental data: first, the accuracy of the Relap5 program simulation is verified by experimental results, and then the program is used to expand the database of droplet entrainment at the bottom of the plate-shaped element.

[0030] Through specific practical experiments, the data was divided into three groups during the calculation process. The experimental data served as the validation group, while the other data were divided into training and test groups according to the commonly used 3:1 ratio in neural networks. The training set data was mainly used to train the network to meet the prediction requirements; the test set used untrained data to input into the network to verify the accuracy and scalability of the network structure; the validation set provided a direct comparison between the neural network prediction results and the experimental results, as shown in Table 1, for convenient data comparison and analysis.

[0031] Table 1. Prediction results for different network structures with a single hidden layer.

[0032]

[0033] Table 2. Partial prediction results for different network structures with two hidden layers.

[0034]

[0035] It can be seen that the two-hidden-layer neural network is significantly better than the single-hidden-layer neural network in terms of training quality and test error. Based on the error magnitude on the test set, a 4-9-5-1 network structure was chosen after comprehensive evaluation.

[0036] The learning rate parameter is related to the attributes of the dataset itself, the learning method, the choice of activation function, and other factors. A comprehensive analysis of the data characteristics in the database, the neural network training time, and the test set error was conducted. A trial-and-error method was used, with learning rates ranging from 0.01 to 1 applied to the neural network. Each learning rate was calculated 10 times and the average was taken. Ultimately, a learning rate of 0.5 was found to provide a moderate training time and minimize the test set error; therefore, a learning rate of 0.5 was chosen.

[0037] As shown in Figure 2, the results show that: 1) the average relative error between the predicted data and the training set input data is 0.86%, and the average root mean square error is 1.21%; 2) the average relative error of the test set data is 0.69%, and the average root mean square error is 0.94%; 3) the error between the neural network prediction results and the experimental data corresponding to Figure 2, where the maximum relative error is 10.82%, shows a significant improvement in accuracy compared to the calculation results of other models.

[0038] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for predicting the entrainment behavior of droplets re-submerged in a rectangular narrow slit of a plate-shaped element, characterized in that, In the offline phase, a BP neural network was constructed and initialized. The experimental boundary conditions were used as input variables and the liquid entrainment rate was used as output variables. After training and optimizing the weights and thresholds, the network was used for prediction in the online phase. The experimental boundary conditions include: inlet cooling water temperature, cooling water flow rate, system pressure, and heating power. The training utilizes the Relap5 thermal hydraulic optimal estimation program, supplemented by experimental data: first, the accuracy of the Relap5 program simulation is verified using experimental results; then, the program is used to expand the database of droplet entrainment at the bottom of the plate-shaped element. The loss function includes the larger of the difference between the outlet entrainment amount calculated from the neural network output and the inlet flow rate, and 0. Twenty calculations are performed for different network structures, and the final average error is calculated. Based on the error calculation results, the learning rate is ultimately determined to be 0.5, and the network structure is 4-9-5-1. The learning rate parameter is also obtained through calculation and comparison; a value of 0.5 avoids the problems of slow gradient descent, excessively long network training time, getting stuck in local minima and missing the global minimum due to an excessively small learning rate, and also avoids the problem of exceeding the minimum value and causing non-convergence due to an excessively large learning rate.

2. The method for predicting the entrainment behavior of re-submerged droplets in a rectangular narrow slit of a plate-like element according to claim 1, characterized in that, The initialization refers to setting the weights and thresholds of all neurons in the network to random, and normalizing all input data, i.e., the values ​​of the training group, to avoid overfitting of the activation function when the values ​​are large.

3. The method for predicting the entrainment behavior of re-submerged droplets in a rectangular narrow slit of a plate-shaped element according to claim 1, characterized in that, The training specifically includes: 1) After inputting a set of experimental boundary condition data, calculating for each neuron from front to back for the data input to obtain the net input of the hidden layer neurons; 2) Using the Sigmoid function as the activation function to calculate the output of the hidden layer neurons until the data of the output layer is obtained; 3) First calculating the output layer error, and then propagating the error backward from the output layer to calculate the error of all nodes; 4) After calculating the total error, using the gradient descent method to correct the weight thresholds of each layer; 5) Inputting another set of experimental boundary condition data and calculating according to the corrected weights and thresholds until the output of all layers in each set of data is less than the specified minimum error or the number of training iterations reaches the maximum value.

Images