A design method for composite insulator grading ring structure based on deep learning

Abstract

A deep learning-based method for designing the grading ring structure of composite insulators can quickly derive the structural parameters of the grading ring suitable for a new type of composite insulator based on experience. This method provides an initial reference for the design of the grading ring, speeds up the design process, and significantly reduces the time and economic costs required for the design process compared to traditional design methods.

Description

Technical Field

[0001] The present invention relates to the technical field of insulation structure design of power equipment, and in particular to a composite insulator grading ring structure design method based on deep learning. Background Art

[0002] With the construction of a robust smart grid encompassing three horizontal, three vertical, and one ring network, my country's high-voltage, high-capacity transmission technology continues to advance. Composite insulators are widely used in overhead transmission lines. Compared to traditional suspension insulators, composite insulators offer advantages such as lightweight, superior pollution flashover resistance, high mechanical strength, and ease of maintenance. After more than a decade of grid-connected operation, they have earned recognition from the power sector and are now a priority for use in my country's heavily polluted areas.

[0003] The electric field distribution on composite insulators is extremely uneven, with the axial electric field exhibiting a U-shaped curve. This means that the high and low voltage sides experience strong electric fields, potentially causing corona discharge, which accelerates insulation aging and, in severe cases, can lead to core rod brittleness and fracture. Therefore, grading rings can be installed on the high and low voltage side fittings of composite insulators to reduce the maximum field strength near the ends, making the electric field distribution more uniform, reducing the occurrence of corona and flashover, and slowing insulation aging. However, not all grading ring sizes can achieve uniform electric field distribution, and improper structural design can degrade the insulation performance of composite insulators.

[0004] In the early days, the exhaustive method was generally used to design the parameters of the grading ring structure suitable for different composite insulators. However, it required continuous adjustment of the structural size parameters of the grading ring, which was computationally intensive and time-consuming. Summary of the Invention

[0005] In order to overcome the deficiencies of the above technologies, the present invention provides a method for structural design of a grading ring adapted to a composite insulator through an artificial intelligence algorithm.

[0006] The technical solution adopted by the present invention to overcome the technical problems is:

[0007] A method for designing a composite insulator grading ring structure based on deep learning, comprising:

[0008] a) Obtaining structural parameters from the design drawings of several composite insulators and their grading rings, and establishing a database consisting of structural parameters and electric field strength values;

[0009] b) Using ANSYS MAXWELL 3D software, a three-dimensional finite element model of the composite insulator and its grading ring was established in actual dimensions. The electric field distribution of the composite insulator before and after the grading ring was installed was calculated, and the calculation results were imported into a database.

[0010] c) Establish an LSTM deep learning model, normalize the database, divide the normalized database into data sets proportionally, and then perform repeated debugging and training on the LSTM deep learning model;

[0011] d) Inputting the structural parameters of the newly produced composite insulator to be calculated into the LSTM deep learning model trained in step c), calculating the adapted grading ring structural parameters and the corresponding electric field distribution results, and importing the calculation results into the database.

[0012] Furthermore, the structural parameters of the composite insulator in step a) include structural height, insulation distance, minimum dry arc distance, minimum creepage distance, umbrella diameter and length of large and small sheds, tube diameter and length of the sheath, radius and length of the core rod, and rated voltage and rated mechanical load; the structural parameters of the grading ring include the tube diameter, ring diameter and shielding depth of the grading ring.

[0013] Furthermore, in step b), a three-dimensional finite element model of the actual size of the composite insulator and its grading ring is established using ANSYS MAXWELL 3D software, and then the dielectric constant parameter values ​​of the core rod, sheath, and hardware materials are set. Then, according to the actual situation simulated by the three-dimensional finite element model, a DC voltage excitation is applied to the composite insulator model without a grading ring and the composite insulator model with a grading ring. The simulation calculation function of ANSYS MAXWELL software is used to solve the electric field distribution of the composite insulator model without a grading ring and the composite insulator model with a grading ring.

[0014] Furthermore, the electric field strength values ​​at the junction of the hardware, sheath and air of the four composite insulator models are taken as reference values ​​and stored in the database.

[0015] Furthermore, step c) includes:

[0016] c-1) Establish an LSTM deep learning model, use the various parameters in the database as the input units of the LSTM network input layer, and combine the various input units to form the input layer of the LSTM deep learning model. The input activation function of the LSTM deep learning model is the sigmoid function, and its output function is the tanh function;

[0017] c-2) By formula Calculate the forget gate weight f at time t i (t) , where σ is the sigmoid function, is the input vector x at time t (t) The data of the jth column, h (t-1) The data of the jth column, h (t-1) is the hidden layer vector at time t-1, for b f The data of column i, b f is the bias vector of the forget gate, For U f The data of row i and column j, U f is the input weight matrix of the forget gate, W f The data of row i and column j, W f is the cyclic weight matrix of the forget gate;

[0018] c-3) Through the formula Calculate the input information at time t Where, for b g The data of row i, b g is the bias vector of the input information, For U g The data of row i and column j, U g is the input weight matrix of the input information, W g The data of row i and column j, W g is the cyclic weight matrix of the input information, through the formula Calculate the output information at time t In the formula for b q The data of row i, b q is the bias vector of the output information, For U q The data of row i and column j, U q is the input weight matrix of the output information, W q The data of row i and column j, W q is the cyclic weight matrix of the output information;

[0019] c-4) Through the formula Calculate the state update expression inside the LSTM cell at time t In the formula is the cell internal state update expression at time t-1, b i is the data of the i-th row of b, b is the bias vector, U i,j is the data of the i-th row and j-th column of U, where U is the input weight matrix, W i,j is the data of the i-th row and j-th column of W, W is the cyclic weight matrix, through the formula Calculate the output result c-5) The number of input neurons in the input layer of the LSTM deep learning model is 14, the number of neurons in the hidden layer of the LSTM deep learning model is 128, the ReLU function is used as the activation function, and the output layer of the LSTM deep learning model uses Softmax, with 7 output neurons;

[0020] c-6) Through the formula Calculate the data x in row i and column j of the input vector i,j The normalized result Where min x i is the minimum value of the i-th row of the input vector, max x i is the maximum value of the data in the i-th row of the input vector;

[0021] c-7) The processed database is divided into training, test, and validation sets in a ratio of 7:2:1. Hyperparameters are adjusted during the repetitive training of the LSTM deep learning model. After the LSTM deep learning model is established, the validation set is input for testing, and the root mean square error is used as the evaluation criterion to form a prediction model.

[0022] The beneficial effect of the present invention is that when designing and producing a new type of composite insulator, the structural parameters of the grading ring suitable for this insulator can be quickly obtained based on experience, providing an initial reference basis for the design of its grading ring, accelerating the design speed, and greatly reducing the time cost and economic cost required for the design link compared with traditional design methods. DETAILED DESCRIPTION

[0023] The present invention will be further described below.

[0024] A method for designing a composite insulator grading ring structure based on deep learning, comprising:

[0025] a) Organize the design drawings of various existing composite insulators and their grading rings. Obtain the structural parameters from the design drawings of several composite insulators and their grading rings, and establish a database consisting of structural parameters and electric field strength values. b) Use ANSYS MAXWELL 3D software to create a three-dimensional finite element model of the composite insulator and its grading ring in actual dimensions. Calculate the electric field distribution of the composite insulator before and after the grading ring is installed, and import the calculation results into the database.

[0026] c) Establish an LSTM deep learning model, normalize the database, divide the normalized database into data sets proportionally, and then perform repeated debugging and training on the LSTM deep learning model;

[0027] d) Inputting the structural parameters of the newly produced composite insulator to be calculated into the LSTM deep learning model trained in step c), calculating the adapted grading ring structural parameters and the corresponding electric field distribution results, and importing the calculation results into the database.

[0028] When designing and producing a new type of composite insulator, the structural parameters of the grading ring suitable for this insulator can be quickly derived based on experience, providing an initial reference basis for the design of its grading ring, speeding up the design process, and greatly reducing the time and economic costs required for the design process compared to traditional design methods.

[0029] Example 1:

[0030] Furthermore, the structural parameters of the composite insulator in step a) include 14 parameters: structural height, insulation distance, minimum dry arc distance, minimum creepage distance, shed diameter and length, sheath diameter and length, core rod radius and length, as well as rated voltage and rated mechanical load. The structural parameters of the grading ring include three parameters: diameter, ring diameter, and shielding depth. A structural parameter database is established, and the above 17 parameters are stored as a set of data in a structural parameter and electric field strength value database. Each data set reserves four parameter positions for the next step, when four electric field strength values ​​are added.

[0031] Example 2:

[0032] In step b), after establishing the actual size three-dimensional finite element model of the composite insulator and its grading ring by ANSYS MAXWELL 3D software, the dielectric constant parameter value of the core rod, sheath and hardware material is set, and then the composite insulator model without grading ring and the composite insulator model with grading ring are applied with DC voltage excitation according to the actual situation simulated by the three-dimensional finite element model, and the electric field distribution of the composite insulator model without grading ring and the composite insulator model with grading ring is solved by the simulation calculation function of ANSYS MAXWELL software. The two cases of composite insulator assembly and non-assembly of grading ring are simulated, and the electric field strength value at the junction of high and low voltage side hardware, sheath and air is taken, a total of 4 data, and the 4 electric field strength values ​​and the 17 parameters in step 1 are stored in the database as a set of data.

[0033] Example 3:

[0034] The electric field strength values ​​at the junction of the hardware, sheath and air of the four composite insulator models are taken as reference values ​​and stored in the database.

[0035] Example 4:

[0036] The steps of step c) include:

[0037] c-1) Establish an LSTM deep learning model, use the various parameters in the database as the input units of the LSTM network input layer, and combine the various input units to form the input layer of the LSTM deep learning model. The input activation function of the LSTM deep learning model is the sigmoid function, and its output function is the tanh function.

[0038] c-2) By formula Calculate the forget gate weight f at time t i (t) , where σ is the sigmoid function, is the input vector x at time t (t) The data of the jth column, h (t-1) The data of the jth column, h (t-1) is the hidden layer vector at time t-1. t is the hidden layer vector at the current time t, which contains the output of all LSTM cells. for b f The data of column i, b f is the bias vector of the forget gate, For U f The data of row i and column j, U f is the input weight matrix of the forget gate, W f The data of row i and column j, W f is the cyclic weight matrix of the forget gate. c-3) Through the formula Calculate the input information at time t Where, for b g The data of row i, b g is the bias vector of the input information, For U g The data of row i and column j, U g is the input weight matrix of the input information, W g The data of row i and column j, W g is the cyclic weight matrix of the input information, through the formula Calculate the output information at time t In the formula for b q The data of row i, b q is the bias vector of the output information, For U q The data of row i and column j, U q is the input weight matrix of the output information, W qThe data of row i and column j, W q is the cyclic weight matrix of the output information.

[0039] c-4) Through the formula Calculate the state update expression inside the LSTM cell at time t In the formula is the cell internal state update expression at time t-1, b i is the data of the i-th row of b, b is the bias vector, U i,j is the data of the i-th row and j-th column of U, where U is the input weight matrix, W i,j is the data of the i-th row and j-th column of W, W is the cyclic weight matrix, through the formula Calculate the output result c-5) The overall structure of the LSTM deep learning model consists of an input layer, a hidden layer, and an output layer. The input layer has 14 input neurons, which conforms to the structural parameters of the insulator. The hidden layer of the LSTM deep learning model has 128 neurons and uses the ReLU function as the activation function. Its mathematical expression is f(x) = max(x, λx), where λ = 0.01. The output layer of the LSTM deep learning model uses Softmax and has 7 output neurons. This adapts to the three structural parameters of the grading ring and the electric field strength values ​​on the high and low voltage sides before and after the grading ring is installed.

[0040] c-6) Through the formula Calculate the data x in row i and column j of the input vector i,j The normalized result Where min x i is the minimum value of the i-th row of the input vector, max x i is the maximum value of the data in the i-th row of the input vector.

[0041] c-7) The processed database is divided into training, test, and validation sets in a ratio of 7:2:1. Hyperparameters are adjusted during the repetitive training of the LSTM deep learning model. After the LSTM deep learning model is established, the validation set is input for testing, and the root mean square error is used as the evaluation criterion to form a prediction model.

[0042] Finally, the 14 parameters of the newly produced composite insulator to be calculated—the structural height, insulation distance, minimum dry arc distance, minimum creepage distance, shed diameter and length, sheath diameter and length, core rod radius and length, rated voltage, and rated mechanical load—are grouped together and fed into the established LSTM network deep learning model as a column vector. The model then calculates the structural parameters of the corresponding grading ring and obtains the electric field strength values ​​given by the model, namely the output grading ring diameter, ring diameter, and shielding depth, as well as the high and low voltage side electric field strength values ​​before and after the grading ring is installed. Finally, electric field simulation is performed using ANSYS software, and the results are compared to verify the rationality of the results. Once confirmed, the above results are imported into a database.

[0043] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A composite insulator grading ring structure design method based on deep learning, characterized in that: include: a) Obtaining structural parameters from the design drawings of several composite insulators and their grading rings, and establishing a database consisting of structural parameters and electric field strength values; b) Using ANSYS MAXWELL 3D software, a three-dimensional finite element model of the composite insulator and its grading ring was established in actual dimensions. The electric field distribution of the composite insulator before and after the grading ring was installed was calculated, and the calculation results were imported into a database. c) Establish an LSTM deep learning model, normalize the database, divide the normalized database into data sets proportionally, and then perform repeated debugging and training on the LSTM deep learning model; d) Inputting the structural parameters of the newly produced composite insulator to be calculated into the LSTM deep learning model trained in step c) to calculate the structural parameters of the grading ring adapted to it and the corresponding electric field distribution results, wherein the electric field distribution results are the high and low voltage side electric field strength values ​​before and after the grading ring is assembled, and importing the calculation results into the database.

2. The method for designing composite insulator grading ring structures based on deep learning according to claim 1, characterized in that: The structural parameters of the composite insulator in step a) include structural height, insulation distance, minimum dry arc distance, minimum creepage distance, shed diameter and length of large and small sheds, tube diameter and length of the sheath, radius and length of the core rod, and rated voltage and rated mechanical load; the structural parameters of the grading ring include tube diameter, ring diameter, and shielding depth of the grading ring.

3. The method for designing composite insulator grading ring structures based on deep learning according to claim 1, characterized in that: In step b), a three-dimensional finite element model of the actual size of the composite insulator and its grading ring is established by ANSYS MAXWELL 3D software, and then the dielectric constant parameter values ​​of the core rod, sheath and hardware materials are set. Then, according to the actual situation simulated by the three-dimensional finite element model, a DC voltage excitation is applied to the composite insulator model without a grading ring and the composite insulator model with a grading ring assembled, and the electric field distribution of the composite insulator model without a grading ring and the composite insulator model with a grading ring assembled is obtained by using the simulation calculation function of ANSYS MAXWELL software.

4. The method for designing composite insulator grading ring structures based on deep learning according to claim 3, characterized in that: The electric field strength values ​​at the junction of the hardware, sheath and air of the four composite insulator models are taken as reference values ​​and stored in the database.

5. The composite insulator grading ring structure design method based on deep learning according to claim 1, characterized in that: The steps of step c) include: c-1) Establish an LSTM deep learning model, use the various parameters in the database as the input units of the LSTM network input layer, and combine the various input units to form the input layer of the LSTM deep learning model. The input activation function of the LSTM deep learning model is the sigmoid function, and its output function is the tanh function; c-2) By formula Calculate the forget gate weight f at time t i (t) , where σ is the sigmoid function, is the input vector x at time t (t) The data of the jth column, h (t-1) The data of the jth column, h (t-1) is the hidden layer vector at time t-1, for b f The data of row i, b f is the bias vector of the forget gate, For U f The data of row i and column j, U f is the input weight matrix of the forget gate, W f The data of row i and column j, W f is the cyclic weight matrix of the forget gate; c-3) Through the formula Calculate the input information at time t Where, for b g The data of row i, b g is the bias vector of the input information, For U g The data of row i and column j, U g is the input weight matrix of the input information, W g The data of row i and column j, W g is the cyclic weight matrix of the input information, through the formula Calculate the output information at time t In the formula for b q The data of row i, b q is the bias vector of the output information, For U q The data of row i and column j, U q is the input weight matrix of the output information, W q The data of row i and column j, W q is the cyclic weight matrix of the output information; c-4) Through the formula Calculate the state update expression inside the LSTM cell at time t In the formula is the cell internal state update expression at time t-1, b i is the data of the i-th row of b, b is the bias vector, U i,j is the data of the i-th row and j-th column of U, where U is the input weight matrix, W i,j is the data of the i-th row and j-th column of W, W is the cyclic weight matrix, through the formula Calculate the output result c-5) The number of input neurons in the input layer of the LSTM deep learning model is 14, the number of neurons in the hidden layer of the LSTM deep learning model is 128, the ReLU function is used as the activation function, and the output layer of the LSTM deep learning model uses Softmax, with 7 output neurons; c-6) Through the formula Calculate the data x in row i and column j of the input vector i,j The normalized result Where min x i is the minimum value of the i-th row of the input vector, max x i is the maximum value of the data in the i-th row of the input vector; c-7) The processed database is divided into training, test, and validation sets in a ratio of 7:2:

1. Hyperparameters are adjusted during the repetitive training of the LSTM deep learning model. After the LSTM deep learning model is established, the validation set is input for testing, and the root mean square error is used as the evaluation criterion to form a prediction model.

Images