A method for predicting physical parameters of a hybrid gas arc plasma

Abstract

This invention belongs to the interdisciplinary field of plasma computation and discloses a method for predicting the physical property parameters of mixed gas arc plasma. The method includes the following steps: obtaining a training set consisting of precise values ​​of various physical property parameters of a sample mixed gas arc plasma under different state parameters; encoding the categories of the physical property parameters to form encoding vectors; and constructing a hybrid expert prediction model including a gating network and multiple expert networks, wherein the gating network takes the encoding vector as input and outputs a weight distribution for selecting the expert networks. This application achieves efficient collaborative prediction of various types of physical property parameters by constructing a hybrid expert prediction model including a gating network and multiple expert networks. Furthermore, by selectively activating expert networks based on the category of physical property parameters instead of calling all of them, the computational complexity is significantly reduced, effectively solving the problems of poor model reusability and low prediction efficiency in existing technologies.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of plasma computing, specifically relating to a method for predicting the physical property parameters of mixed gas arc plasma. Background Technology

[0002] Sulfur hexafluoride (SF6) is widely used in high-voltage circuit breakers due to its excellent insulation and arc-extinguishing properties; however, its extremely high global warming potential has prompted researchers to seek sustainable alternatives. Current research mainly focuses on two directions: mixing SF6 with different buffer gases or replacing SF6 with new environmentally friendly gases. Studying the arc plasma properties of these mixed gases is not only a key basis for numerically evaluating their arc-extinguishing capabilities, but also an important basis for optimizing the design of switchgear.

[0003] Traditional methods for calculating physical property parameters typically rely on pre-calculated data, obtained through table lookups or interpolation. However, these methods often struggle to balance accuracy and computational efficiency, limiting their applicability in practical applications. In recent years, deep neural networks, with their powerful nonlinear mapping capabilities, have been introduced into the field of physical property parameter calculation. While they perform well for individual physical property parameters, different models require separate design and training for each type, leading to model redundancy and wasted training resources. Therefore, existing technologies struggle to achieve prediction of multiple types of physical property parameters using a unified model architecture, resulting in poor model reusability and low prediction efficiency. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method for predicting the physical property parameters of mixed gas arc plasma, which solves the problem that existing technologies have difficulty in predicting multiple types of physical property parameters through a unified model architecture, resulting in poor model reusability and low prediction efficiency.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for predicting the physical properties of mixed gas arc plasma includes the following steps:

[0007] Obtain a training set consisting of precise values ​​of various physical properties of the sample mixed gas arc plasma under different state parameters;

[0008] The categories of physical property parameters are encoded to form encoding vectors;

[0009] Construct a hybrid expert prediction model comprising a gating network and multiple expert networks, wherein:

[0010] Gated networks take encoded vectors as input and output weight distributions for selecting expert networks, thereby activating one or more expert networks according to different categories of physical property parameters.

[0011] The expert network takes the state parameters of the mixed gas arc plasma as input and inputs them into the activated expert network. The activated expert network outputs the corresponding prediction sub-results.

[0012] Based on the weight distribution of the gating network output, the prediction sub-results output by each expert network are weighted and fused to obtain the prediction results of the physical property parameters.

[0013] The hybrid expert prediction model is iteratively trained using the training set, and a loss function is constructed to optimize the network parameters in the hybrid expert prediction model.

[0014] The state parameters and physical property parameters of the mixed gas arc plasma to be tested are obtained and encoded vectors are input into the trained hybrid expert prediction model, and the prediction results of the physical property parameters of the test category are output.

[0015] Physical properties include the thermodynamic parameters and transport coefficients of the mixed gas arc plasma;

[0016] Thermodynamic parameters include mass density, enthalpy, specific heat at isobaric pressure, and entropy.

[0017] Transport coefficients include electrical conductivity, thermal conductivity, and viscosity.

[0018] State parameters include the gas ratio, pressure, and temperature of the mixed gas.

[0019] The gated network consists of a feedforward neural network and a noisy network;

[0020] The feedforward neural network is used to receive the encoded vector and output an initial score vector of the same length as the number of expert networks;

[0021] The noise network is used to generate a perturbation noise vector with the same dimension as the initial score vector;

[0022] The initial fraction vector is added to the disturbance noise vector to obtain the noisy fraction vector;

[0023] Based on the noisy score vector, select the Q expert networks corresponding to the Q highest scores in the noisy score vector as the activated expert networks;

[0024] The scores of the Q expert networks are normalized to obtain the weight distribution of each activated expert network.

[0025] Each expert network consists of a deep operator network composed of a branch network and a backbone network.

[0026] The branch network takes the gas ratio of the mixed gas as input and outputs the first sub-vector;

[0027] The backbone network takes the pressure and temperature of the mixed gas as input and outputs the second sub-vector.

[0028] Perform a dot product operation on the first sub-vector and the second sub-vector, and use it as the predicted sub-result output by the expert network.

[0029] The network parameters in the hybrid expert prediction model include the number of hidden layers in the gating network and the expert network, the number of neurons in each hidden layer, and the learning rate.

[0030] The hybrid expert prediction model is iteratively trained using the training set, and a loss function is constructed to optimize the network parameters in the hybrid expert prediction model. This process includes the following steps:

[0031] Randomly initialize the network parameters in the hybrid expert prediction model;

[0032] The hybrid expert prediction model was iteratively trained using the training set to obtain the prediction results for each physical property parameter under different state parameters in the training set.

[0033] A loss function is constructed based on the mean square error between the predicted results of each physical property parameter and the corresponding accurate values ​​of each physical property parameter in the training set.

[0034] The total loss function is obtained by accumulating the loss function values ​​of each physical property parameter;

[0035] Based on the total loss function, the gradient descent algorithm is used to iteratively optimize the network parameters in the hybrid expert prediction model.

[0036] Preset the lower limit or upper limit of the total loss function;

[0037] The training of the hybrid expert prediction model is complete when the value of the total loss function converges to the lower limit threshold or reaches the upper limit of the number of iterations.

[0038] The categories of physical property parameters are encoded to form encoding vectors, specifically including:

[0039] One-hot encoding is used to digitize the features of different categories of physical property parameters, generating an N-dimensional encoding vector, where N is the total number of categories of physical property parameters;

[0040] The one-hot encoding uses an N-bit state register, with each physical property parameter corresponding to one bit in the state register, and only one bit is in a valid state at any given time.

[0041] The beneficial effects of this invention are:

[0042] This application achieves efficient collaborative prediction of multiple types of physical property parameters by constructing a hybrid expert prediction model that includes a gating network and multiple expert networks. Since the gating network can activate one or more expert networks according to different physical property parameter categories, and then obtain the prediction results of the physical property parameters by weighted fusion of the outputs of each activated expert network, a single hybrid expert prediction model can replace the function of multiple independent models in the traditional way. This significantly improves the reusability of the model and avoids the resource waste caused by building a separate prediction model for each physical property parameter. At the same time, by selectively activating expert networks according to the category of physical property parameters instead of calling all of them, the computational complexity is significantly reduced and the prediction efficiency is improved. This effectively solves the problem in existing technologies where it is difficult to predict multiple types of physical property parameters through a unified model architecture, resulting in poor model reusability and low prediction efficiency. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the process for predicting the physical property parameters of mixed gas arc plasma according to the present invention;

[0045] Figure 2 This is a schematic diagram of the hybrid expert prediction model architecture of the present invention;

[0046] Figure 3 This is a comparison chart of the predicted mass density results and the accurate values ​​in an embodiment of the present invention;

[0047] Figure 4 This is a comparison chart of the predicted conductivity results and the accurate values ​​in the embodiments of the present invention;

[0048] Figure 5 This is a comparison chart of the predicted enthalpy values ​​and the accurate values ​​in the embodiments of the present invention;

[0049] Figure 6 This is a comparison chart of the predicted viscosity coefficient and the accurate value in the embodiments of the present invention;

[0050] Figure 7 This is a comparison chart of the predicted specific heat at constant pressure and the accurate value in the embodiments of the present invention;

[0051] Figure 8 This is a comparison chart of the predicted thermal conductivity results and the accurate values ​​in the embodiments of the present invention;

[0052] Figure 9This is a comparison chart of the entropy prediction result and the accurate value in an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] like Figures 1 to 2 As shown, a method for predicting the physical properties of a mixed gas arc plasma includes the following steps:

[0055] Obtain a training set consisting of precise values ​​of various physical properties of the sample mixed gas arc plasma under different state parameters;

[0056] The categories of physical property parameters are encoded to form encoding vectors;

[0057] Construct a hybrid expert prediction model comprising a gating network and multiple expert networks, wherein:

[0058] Gated networks take encoded vectors as input and output weight distributions for selecting expert networks, thereby activating one or more expert networks according to different categories of physical property parameters.

[0059] The expert network takes the state parameters of the mixed gas arc plasma as input and inputs them into the activated expert network. The activated expert network outputs the corresponding prediction sub-results.

[0060] Based on the weight distribution of the gating network output, the prediction sub-results output by each expert network are weighted and fused to obtain the prediction results of the physical property parameters.

[0061] The hybrid expert prediction model is iteratively trained using the training set, and a loss function is constructed to optimize the network parameters in the hybrid expert prediction model.

[0062] The state parameters and physical property parameters of the mixed gas arc plasma to be tested are obtained and encoded vectors are input into the trained hybrid expert prediction model, and the prediction results of the physical property parameters of the test category are output.

[0063] This application achieves efficient collaborative prediction of various types of physical property parameters by constructing a hybrid expert prediction model that includes a gating network and multiple expert networks. Since the gating network can activate one or more expert networks according to different physical property parameter categories, and then the outputs of each activated expert network are weighted and fused to obtain the prediction results, a single hybrid expert prediction model can replace the function of multiple independent models in the traditional way. This significantly improves the model's reusability and avoids the resource waste caused by building a separate prediction model for each physical property parameter. Simultaneously, by selectively activating expert networks according to the physical property parameter category instead of calling all of them, the computational complexity is significantly reduced and the prediction efficiency is improved. This effectively solves the problem in existing technologies where it is difficult to predict various types of physical property parameters through a unified model architecture, resulting in poor model reusability and low prediction efficiency.

[0064] Preferably, one-hot encoding is used to digitize the N physical property parameters used in training. That is, an N-bit state register is used to encode the N physical property parameters. Each parameter has its own independent register bit, and only one bit is in a valid state at any time, which is used as the model input for training.

[0065] This application employs one-hot encoding to generate encoding vectors, ensuring the orthogonality and equality of different property parameter categories in the input space and avoiding incorrect associations of the encoding vectors by the gating network. At the same time, one-hot encoding has excellent scalability. When it is necessary to add new property parameters for calculation, only one dimension needs to be extended in the state register without changing the core architecture, which greatly improves the practicality of this method.

[0066] Physical properties include the thermodynamic parameters and transport coefficients of the mixed gas arc plasma;

[0067] Thermodynamic parameters include mass density, enthalpy, specific heat at isobaric pressure, and entropy.

[0068] Transport coefficients include electrical conductivity, thermal conductivity, and viscosity.

[0069] It should be noted that when constructing a hybrid expert prediction model, the theoretical calculation basis of various physical property parameters must be clearly defined to ensure the accuracy of subsequent predictions.

[0070] Thermodynamic parameters can be directly calculated based on the definitions of plasma composition and internal partition function:

[0071] The specific formula for calculating the mass density ρ is as follows:

[0072]

[0073] In the formula, n i m iThese are particle number density and particle mass, respectively.

[0074] The specific calculation formula for the enthalpy value h is as follows:

[0075]

[0076] In the formula, x i and These are the mole fraction and standard state molar enthalpy of particle i, respectively, and M is the molar mass of the mixture;

[0077] The specific calculation formula for the isobaric specific heat Cp is as follows:

[0078]

[0079] The specific formula for calculating the entropy value s is as follows:

[0080]

[0081] The transport coefficient can be calculated by solving the Boltzmann equation based on Chapman-Enskog theory:

[0082] The specific formula for calculating conductivity σ is as follows:

[0083]

[0084] In the formula, e is the elementary charge, and n e Let m be the electron number density, k be the Boltzmann constant, and m be the electron number density. e For electron mass, q ij It is a function of the fractional density of each group and the collision integral between groups;

[0085] The specific formula for calculating the viscosity coefficient η is as follows:

[0086]

[0087] In the formula, x i H represents the mole fraction of particle i. ii H represents the self-interaction of particle i and its coupling effect with other particles. ij The cross-interactions between different particles i and j are described;

[0088] The specific formula for calculating thermal conductivity κ is as follows:

[0089] κ=κ tr +κ in +κ re

[0090] In the formula, κ tr κ in and κ reThese are contact thermal conductivity, internal thermal conductivity, and reaction thermal conductivity, respectively.

[0091] State parameters include the gas ratio, pressure, and temperature of the mixed gas.

[0092] The gated network consists of a feedforward neural network and a noisy network;

[0093] The feedforward neural network is used to receive the encoded vector and output an initial score vector of the same length as the number of expert networks;

[0094] The noise network is used to generate a perturbation noise vector with the same dimension as the initial score vector;

[0095] The initial fraction vector is added to the disturbance noise vector to obtain the noisy fraction vector;

[0096] Based on the noisy score vector, select the Q expert networks corresponding to the Q highest scores in the noisy score vector as the activated expert networks;

[0097] The scores of the Q expert networks are normalized to obtain the weight distribution of each activated expert network.

[0098] Each expert network consists of a deep operator network composed of a branch network and a backbone network.

[0099] The branch network takes the gas ratio of the mixed gas as input and outputs the first sub-vector;

[0100] The backbone network takes the pressure and temperature of the mixed gas as input and outputs the second sub-vector.

[0101] Perform a dot product operation on the first sub-vector and the second sub-vector, and use it as the predicted sub-result output by the expert network.

[0102] The network parameters in the hybrid expert prediction model include the number of hidden layers in the gating network and the expert network, the number of neurons in each hidden layer, and the learning rate.

[0103] The hybrid expert prediction model is iteratively trained using the training set, and a loss function is constructed to optimize the network parameters in the hybrid expert prediction model. This process includes the following steps:

[0104] Randomly initialize the network parameters in the hybrid expert prediction model;

[0105] The hybrid expert prediction model was iteratively trained using the training set to obtain the prediction results for each physical property parameter under different state parameters in the training set.

[0106] A loss function is constructed based on the mean square error between the predicted results of each physical property parameter and the corresponding accurate values ​​of each physical property parameter in the training set.

[0107] The total loss function is obtained by accumulating the loss function values ​​of each physical property parameter;

[0108] Based on the total loss function, the gradient descent algorithm is used to iteratively optimize the network parameters in the hybrid expert prediction model.

[0109] Preset the lower limit or upper limit of the total loss function;

[0110] The training of the hybrid expert prediction model is complete when the value of the total loss function converges to the lower limit threshold or reaches the upper limit of the number of iterations.

[0111] This application uses SF6 gas and N2 binary mixed gas arc plasma as the research object to verify the prediction method of this application;

[0112] The sets of gas ratio C, pressure P, and temperature T for the state parameters in the training set are as follows:

[0113] T=[300K,400K,500K,...,3000K]

[0114] P=[1bar,3bar,5bar,7bar,9bar,11bar,13bar,15bar]

[0115]

[0116] In the set of gas ratios C, the first row represents the volume fraction of SF6, and the second row represents the corresponding...

[0117] Volume fraction of N2;

[0118] Preferably, in this embodiment, the Q value is 2;

[0119] Preferably, the number of expert networks is set to 10, each expert network has 4 hidden layers, and each layer has 100 neurons;

[0120] In the gated network, the feedforward neural network has an input dimension of 7, a hidden layer of 1 containing 50 neurons, and an output layer of 10 neurons, which is used to generate a 10-dimensional initial score vector.

[0121] During training, the learning rate was set to 10%. -4 The training rounds are set to 1,000,000.

[0122] The comparison results of the predicted values ​​of various physical property parameters under different state parameters with the accurate values ​​in the training set are as follows: Figures 3 to 9As shown, the model can obtain accurate simulation results under different gas ratios and pressure conditions. Compared with ordinary neural networks, the hybrid expert architecture can leverage the advantages of each expert on different subsets of physical property parameter data, enabling the prediction of multiple physical property parameters in a single training, thereby significantly improving the overall computational efficiency of the model.

[0123] The method for calculating the physical properties of mixed gas arc plasma based on a hybrid expert architecture not only obtains accurate calculation results, but also effectively reduces computational costs by activating only relevant experts when processing input samples, rather than calling all model parameters.

[0124] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0125] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for predicting the physical property parameters of a mixed gas arc plasma, characterized in that, Includes the following steps: Obtain a training set consisting of precise values ​​of various physical properties of the sample mixed gas arc plasma under different state parameters; The categories of physical property parameters are encoded to form encoding vectors; Construct a hybrid expert prediction model comprising a gating network and multiple expert networks, wherein: Gated networks take encoded vectors as input and output weight distributions for selecting expert networks, thereby activating one or more expert networks according to different categories of physical property parameters. The expert network takes the state parameters of the mixed gas arc plasma as input and inputs them into the activated expert network. The activated expert network outputs the corresponding prediction sub-results. Based on the weight distribution of the gating network output, the prediction sub-results output by each expert network are weighted and fused to obtain the prediction results of the physical property parameters. The hybrid expert prediction model is iteratively trained using the training set, and a loss function is constructed to optimize the network parameters in the hybrid expert prediction model. The state parameters and physical property parameters of the mixed gas arc plasma to be tested are obtained and encoded vectors are input into the trained hybrid expert prediction model, and the prediction results of the physical property parameters of the test category are output. The gated network includes a feedforward neural network and a noise network; The feedforward neural network is used to receive the encoded vector and output an initial score vector of the same length as the number of expert networks; The noise network is used to generate a perturbation noise vector with the same dimension as the initial score vector; The initial fraction vector is added to the disturbance noise vector to obtain the noisy fraction vector; Based on the noisy score vector, select the Q expert networks corresponding to the Q highest scores in the noisy score vector as the activated expert networks; The scores corresponding to the Q expert networks are normalized to obtain the weight distribution of each activated expert network. Each of the expert networks consists of a deep operator network composed of a branch network and a backbone network. The branch network takes the gas ratio of the mixed gas as input and outputs the first sub-vector; The backbone network takes the pressure and temperature of the mixed gas as input and outputs the second sub-vector. Perform a dot product operation on the first sub-vector and the second sub-vector, and use it as the predicted sub-result output by the expert network.

2. The method for predicting the physical property parameters of mixed gas arc plasma according to claim 1, characterized in that, Physical properties include the thermodynamic parameters and transport coefficients of the mixed gas arc plasma; Thermodynamic parameters include mass density, enthalpy, specific heat at isobaric pressure, and entropy. Transport coefficients include electrical conductivity, thermal conductivity, and viscosity.

3. The method for predicting the physical property parameters of mixed gas arc plasma according to claim 1, characterized in that, State parameters include the gas ratio, pressure, and temperature of the mixed gas.

4. The method for predicting the physical property parameters of mixed gas arc plasma according to claim 1, characterized in that, The network parameters in the hybrid expert prediction model include the number of hidden layers in the gating network and the expert network, the number of neurons in each hidden layer, and the learning rate.

5. The method for predicting the physical property parameters of mixed gas arc plasma according to claim 1, characterized in that, The hybrid expert prediction model is iteratively trained using the training set, and a loss function is constructed to optimize the network parameters in the hybrid expert prediction model. This process includes the following steps: Randomly initialize the network parameters in the hybrid expert prediction model; The hybrid expert prediction model was iteratively trained using the training set to obtain the prediction results for each physical property parameter under different state parameters in the training set. A loss function is constructed based on the mean square error between the predicted results of each physical property parameter and the corresponding accurate values ​​of each physical property parameter in the training set. The total loss function is obtained by accumulating the loss function values ​​of each physical property parameter; Based on the total loss function, the gradient descent algorithm is used to iteratively optimize the network parameters in the hybrid expert prediction model. Preset the lower limit or upper limit of the total loss function; The training of the hybrid expert prediction model is complete when the value of the total loss function converges to the lower limit threshold or reaches the upper limit of the number of iterations.

6. The method for predicting the physical property parameters of mixed gas arc plasma according to claim 1, characterized in that, The categories of physical property parameters are encoded to form encoding vectors, specifically including: One-hot encoding is used to digitize the features of different categories of physical property parameters, generating an N-dimensional encoding vector, where N is the total number of categories of physical property parameters; The one-hot encoding uses an N-bit state register, with each physical property parameter corresponding to one bit in the state register, and only one bit is in a valid state at any given time.

Images