Transformer model construction method and system based on detection impedance matrix

Abstract

The invention relates to the technical field of transformer detection, in particular to a transformer model construction method and system based on a detection impedance matrix, and the method comprises the steps: determining a to-be-optimized parameter vector theta based on a transformer equivalent circuit model, and defining a constraint domain omega of the to-be-optimized parameter vector theta; forming a training data set by using all the detection impedance matrixes under the same frequency point; building a neural network model by using the constraint domain omega, the to-be-optimized parameter vector theta and the training data set, and training the neural network model by adopting an Adam optimizer until the loss function is converged; based on a particle swarm optimization algorithm, taking a difference between a minimum detection impedance matrix and a prediction impedance matrix generated by the neural network model as a target, and searching an optimal parameter vector; reconstructing an evaluation impedance matrix by using the optimal parameter vector and the corresponding frequency point f, and representing the transformer model by the optimal parameter vector and the evaluation impedance matrix; the problem that the evaluation of various aspects of conditions of the transformer is wrong due to bad data in detection impedance matrix data is solved.

Description

Technical Field

[0001] This invention relates to the field of transformer testing technology, and in particular to a method and system for constructing a transformer model based on a detection impedance matrix. Background Technology

[0002] Transformers are critical equipment in power systems, and their operating status directly affects the reliability and security of the power grid. Currently, transformer condition monitoring and fault diagnosis mainly rely on port impedance measurement methods, such as detecting the current and voltage at each port, calculating the impedance of each port, and then arranging the impedances of each port according to the port number to form an impedance matrix. The impedance matrix is ​​then used to further analyze and evaluate the condition of the transformer (such as whether the internal core, windings, etc. are damaged or deformed).

[0003] After obtaining the detection impedance matrix through transformer port detection, a large amount of detection impedance matrix data is obtained based on different frequencies during transformer operation (such as the power frequency fundamental wave 50Hz / 60Hz and specific harmonics with significant energy: 2nd harmonic 100Hz, 3rd harmonic 150Hz, 5th harmonic 250Hz, etc.) and the number of port detections. The usual practice is to traverse all detection impedance matrices to evaluate various aspects of the transformer's condition.

[0004] However, when evaluating the condition of a transformer in this way, since the impedance matrix data is not subject to any constraints, some impedance matrix data (which is composed of the self-impedance and mutual impedance of the measured voltage and current calculation ports and is a calculation result) has deviated from the range of the transformer model (for example, the transformer winding resistance value represented by the impedance matrix has exceeded the design range of resistance value), and cannot be used to represent the transformer model, thus leading to errors in the evaluation of various aspects of the transformer's condition. Summary of the Invention

[0005] To address the aforementioned shortcomings, the present invention aims to propose a method and system for constructing a transformer model based on the impedance matrix, thereby solving the problem of incorrect evaluation of various aspects of the transformer due to poor data in the impedance matrix.

[0006] To achieve this objective, the present invention adopts the following technical solution: A method for constructing a transformer model based on the sensing impedance matrix includes the following steps: A1: Determine the parameter vector θ to be optimized based on the transformer equivalent circuit model, and define the constraint domain Ω of the parameter vector θ to be optimized; A2: Compile all detection impedance matrices at the same frequency point into a training dataset; A3: Build a neural network model with the constraint domain Ω, the parameter vector to be optimized θ and the training dataset, and train the neural network model using the Adam optimizer until the loss function converges; A4: Based on the particle swarm optimization algorithm, the optimal parameter vector is searched to minimize the difference between the detection impedance matrix and the predicted impedance matrix generated by the neural network model. ; A5: Using the optimal parameter vector And the corresponding frequency point f is used to reconstruct the evaluation impedance matrix, which is derived from the optimal parameter vector. The impedance matrix represents the transformer model.

[0007] Furthermore, in step A1, the parameter vector θ to be optimized is: θ=[ ] ; in, For high voltage winding resistance, For leakage inductance of the high-voltage winding, For low-voltage winding resistance, For low-voltage winding leakage inductance, For iron loss resistance, It is the magnetizing inductor.

[0008] Furthermore, in step A1, the constraint domain Ω includes: (1) Non-negative constraints on the resistance and inductance of the parameter vector θ to be optimized; (2) Constraints on the range of values ​​for each parameter in the parameter vector θ to be optimized; (3) Short-circuit impedance calibration constraint of the parameter vector θ to be optimized.

[0009] Furthermore, step A4 includes the following sub-steps: A41: Initialize the particle swarm, with particle dimensions consistent with the parameter vector to be optimized θ, and particle positions strictly constrained within the constraint domain Ω; A42: For each particle, the candidate solution for the parameter vector θ to be optimized is input into the trained neural network model to obtain the predicted impedance matrix; A43: Calculate the particle fitness, which is the sum of squares of the Frobenius norms of the detection impedance matrix and the predicted impedance matrix at each frequency point. Add a penalty term for particles that violate the constraint domain Ω. A44: Update the individual optimal pbest, global optimal gbest, and particle position / velocity. After updating, check whether the particle position satisfies the constraint domain Ω. Iterate until fitness converges. A45: Use the globally optimal gbest as the optimal parameter vector. .

[0010] Furthermore, in step A3, the neural network model includes an input layer, two hidden layers, and an output layer; the activation function for the two hidden layers is ReLU, the output layer uses linear activation, and the loss function is mean squared error. Training is performed until the mean squared error is less than [a certain value]. .

[0011] Furthermore, in step A5, based on the optimal parameter vector At frequency point f, each element in the evaluation impedance matrix is ​​calculated using a T-type or π-type equivalent circuit model, thereby reconstructing the evaluation impedance matrix.

[0012] Furthermore, in step A5, the matrix elements are set to... Where i and j represent the row and column respectively, the calculation equation is: Angular frequency ω = 2πf; Magnetizing impedance ; when When =j, ; When i≠j .

[0013] A transformer model construction system based on a sensed impedance matrix, used to execute the aforementioned transformer model construction method based on a sensed impedance matrix, includes: Definition module: Used to input and store custom parameter vectors θ and constraint domains Ω; Training data module: used for inputting and storing detection impedance matrices, and assigning detection impedance matrices at the same frequency point to the same training dataset; Neural network module: used to connect the definition module and the training data module to build a neural network model; The optimization module is used to coordinate with the neural network training module to execute a particle swarm optimization algorithm, aiming to minimize the difference between the detection impedance matrix and the predicted impedance matrix generated by the neural network model, and to search for the optimal parameter vector. ; Matrix reconstruction module: used to obtain the optimal parameter vector from the optimization module. To set an algorithm and generate an evaluation impedance matrix; Transformer model module: Used to store the optimal parameter vector generated by the optimization module. The evaluation impedance matrix generated by the matrix reconstruction module.

[0014] The technical solution provided by this invention can include the following beneficial effects: Determine the parameter vector θ to be optimized and the constraint domain Ω using an equivalent circuit model (e.g., T-type or π-type) that conforms to transformer design specifications or physical requirements; establish a mapping between the parameter vector θ to be optimized and the detection impedance matrix through a neural network (NN) to obtain the predicted impedance matrix; combine the impedance fitting error (i.e., minimization) and the penalty term (i.e., the constraint domain Ω, substituted into the loss function so that detection impedance matrices that violate the constraint domain Ω are eliminated by the algorithm) as the optimization objective of the particle swarm optimization (PSO) algorithm; perform a global search on the detection impedance matrices (i.e., the training dataset) at the same frequency point; and through iterative optimization using PSO, optimize the predicted impedance matrix (i.e., the optimal particle) from numerous detection impedance matrix data to best meet the transformer model requirements (or transformer design specifications) at that frequency point f. The parameter vector θ corresponding to the globally optimal particle of PSO is the optimal parameter vector. This represents the state of each physical parameter of the transformer at that frequency point f, thus avoiding misleading evaluation results from poor data in the evaluation data.

[0015] Considering that the impedance matrix may also be needed when evaluating various aspects of the transformer's condition, the optimal parameter vector is used. And the evaluation impedance matrix at the corresponding frequency point f is reconstructed (equivalent to the detection impedance matrix in the training dataset, and the most representative matrix fitted / selected after eliminating matrices that do not meet the requirements). Attached Figure Description

[0016] Figure 1 This is a flowchart of a transformer model construction method based on a detection impedance matrix, which is one embodiment of the present invention.

[0017] Figure 2 This is a structural block diagram of a transformer model construction system based on a detection impedance matrix, which is one embodiment of the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] In the description of embodiments of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of the present invention according to the specific circumstances.

[0021] The following is combined with Figures 1 to 2 This invention describes a method and system for constructing a transformer model based on a detection impedance matrix, according to an embodiment of the present invention.

[0022] Example 1 A method for constructing a transformer model based on the sensing impedance matrix includes the following steps: A1: Determine the parameter vector θ to be optimized based on the transformer equivalent circuit model, and define the constraint domain Ω of the parameter vector θ to be optimized; A2: Compile all detection impedance matrices at the same frequency point into a training dataset; A3: Build a neural network model with the constraint domain Ω, the parameter vector to be optimized θ and the training dataset, and train the neural network model using the Adam optimizer until the loss function converges; A4: Based on the particle swarm optimization algorithm, the optimal parameter vector is searched to minimize the difference between the detection impedance matrix and the predicted impedance matrix generated by the neural network model. ; A5: Using the optimal parameter vector And the corresponding frequency point f is used to reconstruct the evaluation impedance matrix, which is derived from the optimal parameter vector. The impedance matrix represents the transformer model.

[0023] In a preferred embodiment of the transformer model construction method based on the detection impedance matrix proposed in this invention, such as... Figure 1As shown, an equivalent circuit model (e.g., T-type or π-type) conforming to transformer design specifications or physical requirements is used to determine the parameter vector θ to be optimized and the constraint domain Ω. A neural network (NN) is used to establish a mapping between the parameter vector θ to be optimized and the detection impedance matrix, resulting in a predicted impedance matrix. The impedance fitting error (i.e., minimizing it) and the penalty term (i.e., the constraint domain Ω, substituted into the loss function so that detection impedance matrices violating the constraint domain Ω are eliminated by the algorithm) are combined as the optimization objective of the particle swarm optimization (PSO) algorithm. A global search is performed on the detection impedance matrices at the same frequency point (i.e., the training dataset). Through iterative optimization by PSO, the predicted impedance matrix (i.e., the optimal particle) that best conforms to the transformer model requirements (or transformer design specifications) at that frequency point f is optimized from numerous detection impedance matrix data. The parameter vector θ corresponding to the globally optimal particle of PSO is the optimal parameter vector. This represents the state of each physical parameter of the transformer at that frequency point f, thus avoiding misleading evaluation results from poor data in the evaluation data.

[0024] Considering that the impedance matrix may also be needed when evaluating various aspects of the transformer's condition, the optimal parameter vector is used. And the evaluation impedance matrix at the corresponding frequency point f is reconstructed (equivalent to the detection impedance matrix in the training dataset, and the most representative matrix fitted / selected after eliminating matrices that do not meet the requirements).

[0025] It should be noted that, generally speaking, the results of evaluating the transformer state are similar for the detection impedance matrix at the same frequency point, except for those detection impedance matrices that do not meet the requirements. Therefore, it is sufficient to reconstruct an optimal evaluation impedance matrix for each frequency point for evaluation.

[0026] Furthermore, in step A1, the parameter vector to be optimized is θ: θ=[ ] ; in, For high voltage winding resistance, For leakage inductance of the high-voltage winding, For low-voltage winding resistance, For low-voltage winding leakage inductance, For iron loss resistance, It is the magnetizing inductor.

[0027] In this embodiment, the parameter vector θ to be optimized serves two purposes: first, to establish a mapping relationship with the detection impedance matrix; and second, to derive the optimal parameter vector. For subsequent evaluation of transformer condition, its constituent elements must have the corresponding physical definition of a transformer. Therefore, it consists of high-voltage winding resistance, high-voltage winding leakage inductance, etc., and can be obtained through the optimal parameter vector. One or more elements can be used to determine the condition of a transformer.

[0028] Furthermore, in step A1, the constraint domain Ω includes: (1) Non-negative constraints on the resistance and inductance of the parameter vector θ to be optimized; (2) Constraints on the range of values ​​for each parameter in the parameter vector θ to be optimized; (3) Short-circuit impedance calibration constraint of the parameter vector θ to be optimized.

[0029] In this embodiment, the significance of setting the constraint domain Ω is as follows: (1) Non-negativity constraints on the resistance and inductance of the parameter vector θ to be optimized. For example, non-negativity constraints on resistance, because it is impossible for a real transformer to have negative values ​​or even zero values. Therefore, the optimization is limited. >0, >0, >0; (2) Constraints on the range of values ​​for each parameter in the parameter vector θ to be optimized, such as The range of parameter values ​​is limited by the maximum and minimum values. If the value is not within this range, it proves that the transformer design specifications are violated. (3) The short-circuit impedance calibration constraint of the parameter vector θ to be optimized is to require that the short-circuit impedance in the impedance matrix must be consistent with the short-circuit impedance on the transformer's nameplate. A high degree of agreement is required; otherwise, it cannot represent the transformer. For example: ; in, It is the short-circuit impedance calculated from the short-circuit model parameters (for example, the impedance when the low-voltage side is short-circuited, viewed from the high-voltage side). E is the reference impedance, and E is the allowable error.

[0030] Furthermore, step A4 includes the following sub-steps: A41: Initialize the particle swarm, with particle dimensions consistent with the parameter vector to be optimized θ, and particle positions strictly constrained within the constraint domain Ω; A42: For each particle, the candidate solution for the parameter vector θ to be optimized is input into the trained neural network model to obtain the predicted impedance matrix; A43: Calculate the particle fitness, which is the sum of squares of the Frobenius norms of the detection impedance matrix and the predicted impedance matrix at each frequency point. Add a penalty term for particles that violate the constraint domain Ω. A44: Update the individual optimal pbest, global optimal gbest, and particle position / velocity. After updating, check whether the particle position satisfies the constraint domain Ω. Iterate until fitness converges. A45: Use the globally optimal gbest as the optimal parameter vector. .

[0031] In this embodiment, the objective function of the particle swarm optimization (PSO) algorithm in step A4 is: ; in, Indicates in The detection impedance matrix detected at frequency points, k=1 to N , ... This represents the sum of the detection impedance matrices at the same frequency point, which is equivalent to representing the training dataset. This represents the predicted impedance matrix obtained in sub-step A42; This indicates the penalty term in sub-step A43, which is included in the loss function of step A3. It is a regularization term based on the parameter vector θ to be optimized, commonly using Tikhonov regularization (ridge regression): ; This is a weight matrix, usually an identity matrix. It is a regularization parameter that controls the severity of the penalty.

[0032] The PSO iterative model is: Treat the parameter vector θ to be optimized as the particle position (i.e., consistent dimension) and initialize the particle swarm.

[0033] Iteratively update particle position and velocity: ; ; in, Where 'w' is the particle velocity and 'w' is the inertial weight, It is a learning factor. It is a random number. It is the best in particle history. It is the global optimum.

[0034] Furthermore, in step A3, the neural network model includes an input layer, two hidden layers, and an output layer; the activation function for the two hidden layers is ReLU, the output layer uses linear activation, and the loss function is mean squared error. Training is performed until the mean squared error is less than [a certain value]. .

[0035] In this embodiment, the neural network model is trained using the backpropagation algorithm. The structure is: Input layer → Hidden layer 1 (ReLU) → Hidden layer 2 (ReLU) → Output layer (linear). The core steps of the backpropagation algorithm are an iterative loop of "forward propagation to calculate loss → backpropagation to calculate gradient → parameter update". (1) Forward propagation calculates the weighted sum and activation value of neurons layer by layer from the input layer to the output layer, and finally obtains the predicted value and calculates the loss; (2) The core of backpropagation is to use the chain rule to calculate the gradient of the loss with respect to the weights and biases of each layer, starting from the loss function. (3) Parameter update: Using the obtained gradient, the Adam optimizer updates the weights and biases of each layer according to the gradient, so that the gradient decreases.

[0036] Furthermore, in step A5, based on the optimal parameter vector At frequency point f, each element in the evaluation impedance matrix is ​​calculated using a T-type or π-type equivalent circuit model, thereby reconstructing the evaluation impedance matrix.

[0037] In this embodiment, since the elements of the parameter vector θ to be optimized are designed based on the physical definition of the transformer by the transformer equivalent circuit model, each matrix element in the evaluation impedance matrix can be deduced from the transformer equivalent circuit model, thereby reconstructing the evaluation impedance matrix. Commonly used transformer equivalent circuit models include T-type and π-type equivalent circuit models.

[0038] Furthermore, in step A5, the matrix elements are set to... Where i and j represent the row and column respectively, the calculation equation is: Angular frequency ω = 2πf; Magnetizing impedance ; when When =j, ; When i≠j .

[0039] The above calculation equations are based on the T-type equivalent circuit model. For example, the evaluation impedance matrix to be reconstructed is:

[0040] when for hour: = ; when for hour: = ; And so on, the impedance matrix will be evaluated. Each matrix element in the matrix is ​​calculated to achieve reconstruction.

[0041] For example: optimal parameter vector =[ Given f=50Hz, the calculation yields: .

[0042] Example 2 A transformer model construction system based on a sensed impedance matrix, used to execute the aforementioned transformer model construction method based on a sensed impedance matrix, includes: Define module 1: Used to input and store the custom parameter vector θ to be optimized and the constraint domain Ω; Training data module 2: used to input and store the detection impedance matrix, and to assign the detection impedance matrix at the same frequency point to the same training dataset; Neural Network Module 3: Used to connect Definition Module 1 and Training Data Module 2 to build a neural network model; Optimization module 4 is used to connect the neural network training module 3 to execute the particle swarm optimization algorithm, aiming to minimize the difference between the detection impedance matrix and the predicted impedance matrix generated by the neural network model, and search for the optimal parameter vector. ; Matrix reconstruction module 5: Used to obtain the optimal parameter vector from optimization module 4. To set an algorithm and generate an evaluation impedance matrix; Transformer Model Module 6: Used to store the optimal parameter vector generated by Optimization Module 4. The evaluation impedance matrix generated by the matrix reconstruction module 5.

[0043] In this embodiment, a preferred embodiment of a transformer model construction system based on the detection impedance matrix is ​​also proposed, such as... Figure 2 As shown, the system is built in an electronic device or computer, and the processor of the electronic device or computer can execute each step of the transformer model construction method based on the detection impedance matrix according to the program or instructions.

[0044] Other components and operations of the transformer model construction method and system based on the detection impedance matrix according to the present invention are known to those skilled in the art and will not be described in detail here.

[0045] In the description of this specification, references to terms such as "embodiment," "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.

[0046] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for constructing a transformer model based on a detection impedance matrix, characterized in that: Includes the following steps: A1: Determine the parameter vector θ to be optimized based on the transformer equivalent circuit model, and define the constraint domain Ω of the parameter vector θ to be optimized; A2: Compile all detection impedance matrices at the same frequency point into a training dataset; A3: Build a neural network model with the constraint domain Ω, the parameter vector to be optimized θ and the training dataset, and train the neural network model using the Adam optimizer until the loss function converges; A4: Based on the particle swarm optimization algorithm, the optimal parameter vector is searched to minimize the difference between the detection impedance matrix and the predicted impedance matrix generated by the neural network model. ; A5: Using the optimal parameter vector And the corresponding frequency point f is used to reconstruct the evaluation impedance matrix, which is derived from the optimal parameter vector. The impedance matrix represents the transformer model.

2. The method for constructing a transformer model based on a detection impedance matrix according to claim 1, characterized in that: In step A1, the parameter vector to be optimized is θ: θ=[ ]; in, For high voltage winding resistance, For leakage inductance of the high-voltage winding, For low-voltage winding resistance, For low-voltage winding leakage inductance, For iron loss resistance, It is the magnetizing inductor.

3. The method for constructing a transformer model based on a detection impedance matrix according to claim 1, characterized in that: In step A1, the constraint domain Ω includes: (1) Non-negative constraints on the resistance and inductance of the parameter vector θ to be optimized; (2) Constraints on the range of values ​​for each parameter in the parameter vector θ to be optimized; (3) Short-circuit impedance calibration constraint of the parameter vector θ to be optimized.

4. The method for constructing a transformer model based on a detection impedance matrix according to claim 1, characterized in that: Step A4 includes the following sub-steps: A41: Initialize the particle swarm, with particle dimensions consistent with the parameter vector to be optimized θ, and particle positions strictly constrained within the constraint domain Ω; A42: For each particle, the candidate solution for the parameter vector θ to be optimized is input into the trained neural network model to obtain the predicted impedance matrix; A43: Calculate the particle fitness, which is the sum of squares of the Frobenius norms of the detection impedance matrix and the predicted impedance matrix at each frequency point. Add a penalty term for particles that violate the constraint domain Ω. A44: Update the individual optimal pbest, global optimal gbest, and particle position / velocity. After updating, check whether the particle position satisfies the constraint domain Ω. Iterate until fitness converges. A45: Use the globally optimal gbest as the optimal parameter vector. .

5. The method for constructing a transformer model based on a detection impedance matrix according to claim 1, characterized in that: In step A3, the neural network model includes an input layer, two hidden layers, and an output layer; the activation function of the two hidden layers is ReLU, the output layer is linear activation, and the loss function is mean squared error. Training is performed until the mean squared error is less than [a certain value]. .

6. The method for constructing a transformer model based on a detection impedance matrix according to claim 1, characterized in that: In step A5, based on the optimal parameter vector At frequency point f, each element in the evaluation impedance matrix is ​​calculated using a T-type or π-type equivalent circuit model, thereby reconstructing the evaluation impedance matrix.

7. The method for constructing a transformer model based on a detection impedance matrix according to claim 6, characterized in that: In step A5, the matrix elements are set to... Where i and j represent the row and column respectively, the calculation equation is: Angular frequency ω = 2πf; Magnetizing impedance ; when When =j, ; When i≠j .

8. A transformer model construction system based on the detection impedance matrix, characterized in that: A method for constructing a transformer model based on a detection impedance matrix according to any one of claims 1 to 7 includes: Definition module: Used to input and store custom parameter vectors θ and constraint domains Ω; Training data module: used for inputting and storing detection impedance matrices, and assigning detection impedance matrices at the same frequency point to the same training dataset; Neural network module: used to connect the definition module and the training data module to build a neural network model; The optimization module is used to coordinate with the neural network training module to execute a particle swarm optimization algorithm, aiming to minimize the difference between the detection impedance matrix and the predicted impedance matrix generated by the neural network model, and to search for the optimal parameter vector. ; Matrix reconstruction module: used to obtain the optimal parameter vector from the optimization module. To set an algorithm and generate an evaluation impedance matrix; Transformer model module: Used to store the optimal parameter vector generated by the optimization module. The evaluation impedance matrix generated by the matrix reconstruction module.

Images