Power distribution device self-tuning method based on neural network digital twinning algorithm

Abstract

The invention discloses a power distribution unit self-tuning method based on a neural network digital twinning algorithm, and the method comprises the steps: firstly building an equivalent test loop, carrying out the mathematical modeling, carrying out the per-unit processing of test loop elements (an inductor, a capacitor, a resistor and the like), and carrying out the re-correction of a model, so as to facilitate the subsequent optimization; and establishing a conversion relation formula between the dimensionless parameters and the dimensional parameters. And taking the dimensionless parameter as a decision variable, taking a differential equation corresponding to the model as a constraint condition, taking the sum of absolute values of differences between obtained values and standard values as a value function, continuously training the neural network model, and calculating a minimum value of the value function by using a neural network digital twinning algorithm. And finally, performing simulation verification on an optimization result. The method is suitable for solving the problem that complex nonlinear correlation exists between input and output; compared with a linear model (such as linear regression), finer nonlinear features in the data can be captured.

Description

Technical Field

[0001] This invention belongs to the field of testing technology for high-voltage electrical equipment in power systems, and particularly relates to a self-tuning method for power distribution devices based on a neural network digital twin algorithm. Background Technology

[0002] With rapid societal development, electricity demand is constantly increasing, power grid capacity is rising year by year, and short-circuit current values ​​in power systems are becoming increasingly larger. The superior performance of power transmission and transformation equipment such as switches and transformers is a key factor in ensuring the safe operation of the power system. Therefore, conducting high-voltage and high-current tests on power transmission and transformation equipment such as switches and transformers is of great significance.

[0003] Among numerous power transmission and transformation equipment, switching devices play a crucial role in ensuring the safe and stable operation of the entire power system and providing high-quality electrical energy. Representative switching devices in high-voltage, high-current transmission and transformation circuits include disconnect switches, load switches, and circuit breakers. Among these, circuit breakers are the most important and complex type of high-voltage switch. Circuit breakers used in 3kV and above power systems are generally called high-voltage circuit breakers. They serve as the last line of defense for the power system and are one of the key components ensuring its safe operation. Regardless of the power system's state—no-load, loaded, or under short-circuit fault—when required to operate, the circuit breaker must reliably operate, either closing or opening the circuit. Therefore, to ensure that circuit breakers can operate normally, reliably, and quickly during short-circuit faults in the power system, short-circuit breaking tests on high-voltage circuit breakers are essential.

[0004] To verify and assess the breaking performance of high-voltage AC circuit breakers, international standards (IEC 62271-100) and Chinese standards (GB_T 1984-2014) specify stringent short-circuit breaking test requirements. Direct testing, as the test method closest to actual power grid operating conditions, is widely used due to its realistic physical process and lack of complex equivalence theory. However, direct testing involves numerous test items. For example, switch short-circuit tests require multiple test items under different operating conditions, such as T10, T30, T60, T100a, and T100s. Each item requires parameter adjustment of the test circuit. Currently, parameter adjustment methods rely heavily on experimental experience, which cannot meet the needs of rapid parameter setting and accurate short-circuit current acquisition for short-circuit breaking capacity testing. Therefore, a new method is needed to efficiently, accurately, and automatically complete the multi-parameter coordinated tuning of the direct test circuit for high-voltage AC circuit breakers' short-circuit breaking, ensuring that the test waveform strictly conforms to standard requirements and accurately reflects the circuit breaker's breaking capacity. Summary of the Invention

[0005] Purpose of the Invention: The purpose of this invention is to provide a self-tuning method for power distribution devices based on a neural network digital twin algorithm. This method utilizes neural network digital twin algorithm processing technology to solve the problem of multi-parameter coordinated tuning in direct test circuits.

[0006] Technical solution: The present invention provides a self-tuning method for power distribution equipment based on a neural network digital twin algorithm, comprising the following steps:

[0007] Step 1: Convert the three-phase direct test circuit into a single-phase equivalent test circuit based on symmetry. Take the phase where the circuit breaker that is first broken is the equivalent single phase. At the same time, convert the transformer and reactor in the circuit into inductors, and the power grid or impulse generator into a voltage source. Establish the equivalent model.

[0008] Step 2: Based on Kirchhoff's Current Law (KCL), Kirchhoff's Voltage Law (KVL), and the characteristics of inductor and capacitor components, write the equations and initial conditions of the equivalent model.

[0009] Step 3: Select standard values ​​for voltage and current, and convert the time domain to the angle domain. Rewrite the equations and initial conditions to obtain the equivalent model under dimensionless parameters.

[0010] Step 4: Based on the equivalent model under dimensionless parameters, organize the differential equations relating the per-unit voltage across the tested circuit breaker to other dimensionless parameters.

[0011] Step 5: Use dimensionless parameters as decision variables for the neural network digital twin algorithm;

[0012] Step 6: Determine the value function of the neural network digital twin algorithm based on the differential equation relationship;

[0013] Step 7: Train the neural network to obtain the dimensionless parameters corresponding to the minimum value function. Convert these parameters to inductance, capacitance, and resistance values ​​using formulas. Then perform Simulink simulation. Obtain the waveform parameters from the obtained TRV data using the differential equation relationship from Step 6. Compare the waveform parameters with standard values ​​to verify the effectiveness of the method.

[0014] Furthermore, step 2 specifically involves: based on KCL, KVL, and the characteristics of inductor and capacitor components, the equations and initial conditions of the equivalent model are as follows:

[0015]

[0016] In the formula, and respectively flowing through the inductor and The current, u s As a voltage source, C0 and C1 are three-phase capacitors, Rs For the load impedance, u tr The transient recovery voltage across the circuit breaker is given by t, where t is time, and the asterisk (*) indicates a dimensional parameter.

[0017] Furthermore, step 3 specifically involves: selecting standard values ​​for voltage and current, transforming the time domain to the angle domain, rewriting the equations and initial conditions, and obtaining the equivalent model under dimensionless parameters:

[0018]

[0019] in, The maximum value in the y-axis data. Let t be the natural angular frequency and t be time. For the power frequency recovery voltage peak value, u c The peak value of the transient recovery voltage across the circuit breaker;

[0020] Will , These are used as standard values ​​for voltage and current, respectively, and simultaneously... replace , Indicates angle, which is time. With no damping in the test circuit natural angular frequency The product of these equations, the dimensionless form of the system of equations (1) is:

[0021]

[0022] Choose the following correspondence:

[0023]

[0024] In the formula, , These represent dimensionless parameters for capacitance and resistance, respectively.

[0025] According to equation (4), equation (3) can be rearranged as follows:

[0026]

[0027] Equation (5) yields the equivalent model under dimensionless parameters.

[0028] Furthermore, step 4 specifically involves: based on the equivalent model under dimensionless parameters, organizing the differential equation relationships between the per-unit voltage values ​​across the tested circuit breaker and other dimensionless parameters:

[0029]

[0030] The initial conditions of equation (6) are:

[0031] .

[0032] Transient recovery voltage across the circuit breaker The first derivative, Transient recovery voltage across the circuit breaker The second derivative of .

[0033] Further, step 5 specifically involves: using dimensionless parameters as decision variables for the neural network digital twin algorithm, namely the number of input layer nodes, the number of hidden layers, the number of hidden layer nodes, the number of output layer nodes, and initializing weights and bias terms; and determining the number of hidden layer neurons based on empirical formula (8).

[0034] (8)

[0035] In the formula, This indicates the number of neurons in the hidden layer. This indicates the number of neurons in the input layer. Indicates the number of neurons in the output layer. It is an empirical constant with a value range of 1 to 10;

[0036] Construct Sigmoid, Tanh, and ReLU activation functions;

[0037] The Sigmoid function is shown in formula (9):

[0038] (9)

[0039] In the formula, e represents the natural constant, and x represents the independent variable;

[0040] The Tanh function is shown in formula (10):

[0041] (10)

[0042] The ReLU function is shown in formula (11):

[0043] (11).

[0044] Furthermore, step 6 specifically includes the following steps:

[0045] Step 6.1: Solve the differential equation using the ODE solver in MATLAB to obtain the per-unit voltage and angle across the circuit breaker. The waveform and its vertical and horizontal coordinate data;

[0046] Step 6.2: Find the maximum value in the y-axis data using the max function. :

[0047] Step 6.3: Divide the ordinate of each data point by its abscissa, find the maximum value, and use it as the slope. It passes through the origin and has a slope of The straight line with a vertical coordinate value of Time corresponding to x-coordinate ,Right now

[0048] (12)

[0049] Calculate all data points before the data point corresponding to the maximum value of the y-axis. Distance, distance Represented as

[0050] (13)

[0051] Find the maximum distance value The corresponding data point, passing through this data point and with a slope of The intersection of the straight line and the x-axis is denoted as . ,Right now

[0052] (14)

[0053] make , and These are the parameters that represent the waveform of the differential equation;

[0054] Step 6.4: Convert the dimensional standard value into a dimensionless standard value, i.e., the amplitude coefficient. and The result obtained in step 6.3 and The sum of the absolute values ​​of the differences between the individual values ​​is used as the value function of the neural network digital twin algorithm.

[0055] The value function is expressed as

[0056] (15)

[0057] Step 6.5: The standard stipulates that the envelope of the TRV waveform obtained from the experiment must be above the standard specified value. A weighting coefficient and a penalty term (Penalty) are then added to the value function. Representing the weighting coefficients, the final value function is expressed as:

[0058] (16).

[0059] Furthermore, step 7 specifically involves: performing multiple training iterations of the neural network to obtain the dimensionless parameters corresponding to the minimum value function, converting them into inductance, capacitance, and resistance values ​​using formulas, and performing Simulink simulation. The obtained TRV data is then used to obtain waveform parameters in step 6, which are compared with standard values ​​to verify the effectiveness of the method.

[0060] The conversion relationship between dimensionless and dimensional parameters is as follows:

[0061] (17)

[0062] In the formula, The first polarity coefficient, I is the rated voltage of the circuit breaker, and I is the breaking current of the circuit breaker. , , This is the angle corresponding to a ordinate of 1 in the dimensionless envelope.

[0063] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0064] The present invention also discloses a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0065] The present invention also discloses a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method of the present invention.

[0066] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0067] 1. Strong nonlinear fitting ability: Neural networks, through multiple hidden layers and nonlinear activation functions, can approximate arbitrarily complex nonlinear mapping relationships; they are suitable for solving problems where there are complex nonlinear relationships between input and output; compared with linear models (such as linear regression), they can capture more subtle nonlinear features in the data.

[0068] 2. Strong self-learning and adaptive capabilities: Through the backpropagation algorithm, the model can automatically learn the mapping pattern between input and output from the training data, without the need for manual feature design. When the distribution of input data changes slightly, the model can adapt to the new data by fine-tuning the parameters, demonstrating a certain degree of generalization ability.

[0069] 3. Suitable for multi-input, multi-output scenarios: The number of hidden layers and the number of neurons per layer can be flexibly designed according to the complexity of the problem, balancing the model's capabilities and computational costs.

[0070] 4. Good robustness: During training, through the averaging of the loss function, the neural network is not sensitive to slight noise in the training data and can learn stable patterns in noisy real data.

[0071] 5. Excellent scalability: The algorithm framework can be extended to multi-parameter optimization problems. Attached Figure Description

[0072] Figure 1 This invention provides a three-phase direct test circuit, where AC represents a three-phase AC source and MB represents a protective circuit breaker that promptly interrupts the short-circuit current in case of test failure. T represents a reactor, and T represents a three-phase winding transformer. , , Composed of a frequency modulation device, TB represents the circuit breaker under test;

[0073] Figure 2 This is the single-phase equivalent circuit of the first-phase circuit breaker, where... Represents an equivalent voltage source. Represents equivalent inductance. , , It is a frequency modulation device; TB represents the circuit breaker under test that is switched off first. It is the transient recovery voltage across the circuit breaker;

[0074] Figure 3 This is a single-phase equivalent circuit composed of dimensionless parameters, where 1 for both inductance and capacitance represents a fixed value at the per-unit level. Represents an equivalent voltage source. This represents the per-unit value of series damping. This represents the per-unit value of the time delay capacitance;

[0075] Figure 4 A flowchart illustrating the working principle of a neural network digital twin algorithm;

[0076] Figure 5 This is a flowchart illustrating the process of obtaining the parameters for waveform generation from differential equations in this invention.

[0077] Figure 6 This is a schematic diagram of the parameters to be extracted from the TRV waveform;

[0078] Figure 7 Here are the waveform envelope parameters, where (a) is the TRV waveform envelope parameter; and (b) is the dimensionless TRV waveform envelope parameter.

[0079] Figure 8 This is the simulation model of the present invention in Simulink;

[0080] Figure 9 The present invention provides the TRV simulation waveforms for each test mode under the rated voltage of the circuit breaker of 24kV and the rated short-circuit breaking current of 63kA. Among them, (a) is the simulation waveform under test mode T100, (b) is the simulation waveform under test mode T60, (c) is the simulation waveform under test mode T30, and (d) is the simulation waveform under test mode T10.

[0081] Figure 10 The present invention provides the TRV simulation waveforms for each test mode under the rated voltage of the circuit breaker of 40.5kV and the rated short-circuit breaking current of 63kA. Among them, (a) is the simulation waveform under test mode T100, (b) is the simulation waveform under test mode T60, (c) is the simulation waveform under test mode T30, and (d) is the simulation waveform under test mode T10.

[0082] Figure 11 The present invention provides the TRV simulation waveforms for each test mode under the rated voltage of the circuit breaker of 72.5kV and the rated short-circuit breaking current of 63kA. Among them, (a) is the simulation waveform under test mode T100, (b) is the simulation waveform under test mode T60, (c) is the simulation waveform under test mode T30, and (d) is the simulation waveform under test mode T10.

[0083] Figure 12 Performance graph of the LM algorithm;

[0084] Figure 13 Performance graph of Bayes algorithm;

[0085] Figure 14 The performance graph for the quantization conjugate gradient algorithm. Detailed Implementation

[0086] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0087] Step 1: Combining Figure 1 and Figure 2 To explain, Figure 1 The three-phase direct test circuit in the middle is equivalent to the symmetry. Figure 2 In the single-phase equivalent test circuit, the phase where the circuit breaker that is first broken is taken as the equivalent single phase. At the same time, the transformer and reactor in the circuit are equivalent to inductors, and the power grid or impulse generator is equivalent to a voltage source, and the equivalent model is established.

[0088] Step 2: Based on KCL, KVL, and the characteristics of inductor and capacitor components, Figure 2 The equations and initial conditions of the equivalent model are as follows:

[0089]

[0090] In the formula, and respectively flowing through the inductor and The current, As a voltage source, It is a three-phase capacitor. For load impedance, The transient recovery voltage across the circuit breaker is represented by an asterisk (*) to distinguish it from the dimensionless equations.

[0091] Step 3: Select standard values ​​for voltage and current, transform the time domain to the angle domain, and rewrite the equations and initial conditions to obtain... Figure 3 The equivalent model under dimensionless parameters is shown.

[0092] Combination Figure 6 and Figure 7 Detailed explanation:

[0093] Figure 6 The TRV waveform of the circuit breaker under test. Figure 7 (a) in the figure represents the envelope of the TRV waveform. The vertical axis is plotted with the peak value of the power frequency recovery voltage. The per-unit value is used to represent the x-axis. express( For time The natural angular frequency of the undamped TRV in the test circuit (the product of), then Figure 7 (a) in the middle can be converted to Figure 7 The dimensionless form of (b) in the equation is transformed into the following relation:

[0094]

[0095] Will , These are used as standard values ​​for voltage and current, respectively, and simultaneously... replace , For time The natural angular frequency of the undamped TRV in the test circuit The accumulation of, For power frequency recovery voltage peak, The dimensionless form of the equation system (1) for the peak transient recovery voltage across the circuit breaker is:

[0096]

[0097] Choose the following correspondence

[0098]

[0099] In the formula, , These represent dimensionless parameters for capacitance and resistance, respectively.

[0100] According to equation (4), equation (3) can be rearranged as follows:

[0101]

[0102] From equation (5), we can obtain Figure 3 The equivalent model under dimensionless parameters is shown.

[0103] Step 4: Based on the equations obtained in Step 3, organize the differential equations relating the per-unit voltage across the tested circuit breaker to other dimensionless parameters:

[0104] The final result obtained is

[0105]

[0106] The initial conditions of equation (6) are:

[0107]

[0108] Transient recovery voltage across the circuit breaker The first derivative, Transient recovery voltage across the circuit breaker The second derivative of .

[0109] Step 5: Use dimensionless parameters as decision variables for the neural network digital twin algorithm, namely the number of input layer nodes, the number of hidden layers, the number of hidden layer nodes, the number of output layer nodes, as well as initializing weights and bias terms, etc.

[0110] Combination Figure 4 The working principle of the neural network digital twin algorithm is explained using a flowchart:

[0111] The input variables are set as target values, and each variable has a dimension of 2, representing respectively... , The maximum number of training iterations is set to 1000.

[0112] The learning rate is 0.01, the minimum training error is 1e-5, and the rand function is used to sort each variable. and Random assignment. Since the decision variables are dimensionless parameters, The range is (0, 1). The range is (0, 0.5). After initialization, the data is fed into the input layer of the network. Node values ​​are calculated through forward propagation. For each hidden layer node, its input value (the weighted sum of input data and connection weights) is calculated. The activation function is then applied to obtain the node's output value, continuing until the output layer is reached, yielding the final predicted output. Finally, backpropagation is performed using the loss function value. This involves calculating the error of the output layer neurons, applying the chain rule to calculate the gradient of each connection weight and bias term, and using gradient descent or other optimization algorithms to update the weights and bias terms, minimizing the loss function. The training process of model building requires repeating the forward and backward propagation processes until the stopping condition is met, iterating until the maximum number of training iterations is reached. The result of each variable is input into the value function. The value function value is compared with the previous local optimum, and the smaller value is selected as the new local optimum, with the corresponding variable as the local optimum result. Then, the minimum value is found among all local optima and selected as the global optimum, with the corresponding variable as the global optimum basis result.

[0113] According to the rules for determining hidden layers, the number of hidden layers in a neural network has a significant impact on its capabilities. A network without hidden layers can only handle linearly separable problems; a single hidden layer can perform a continuous mapping between any two finite spaces; two hidden layers combined with an appropriate activation function can fit decision boundaries and smooth mappings of arbitrary complexity; with more than two hidden layers, the additional layers can be used to learn higher-order features, achieving automatic feature extraction. Therefore, this invention employs a network structure with two hidden layers, resulting in lower model complexity and thus reducing the possibility of overfitting.

[0114] The number of neurons in the hidden layer is determined according to the empirical formula (8).

[0115] (8)

[0116] In the formula This indicates the number of neurons in the hidden layer. This indicates the number of neurons in the input layer. Indicates the number of neurons in the output layer. This is an empirical constant ranging from 1 to 10. The modeling process uses iterative calculations to determine the number of hidden layer neurons that minimizes the training error. Finally, the neural network model is created using the optimal number of hidden layer neurons to achieve better training results.

[0117] Construction of activation functions: Activation functions are an important component of artificial neural networks. They transform the input of a neuron into its output by applying a nonlinear function. In a neuron, if the received signal strength exceeds a threshold, it is activated and the signal is passed to the next layer; otherwise, the signal is stopped. Commonly used activation functions include Sigmoid, Tanh, and ReLU.

[0118] The Sigmoid function is shown in formula (9):

[0119] (9)

[0120] In the formula, Represents the natural constant. It represents the independent variable.

[0121] The function exhibits a smooth S-shaped curve, mapping any real number to a probability value, and operates in the non-linear region. However, due to output saturation as the input approaches positive and negative infinity, it suffers from the gradient vanishing problem. This prevents gradients from propagating to shallower layers during neural network training, slowing down network learning. Therefore, it is more suitable as the activation function of the output layer in binary classification problems, representing the probabilities of two classes.

[0122] The Tanh function is shown in formula (10):

[0123] (10)

[0124] The function's output ranges from -1 to 1, similar to the Sigmoid function, but with a larger dynamic range. When the input is large or small, the activation value saturates to near 1 or -1, causing the gradient to become very small. In some cases, it can be used as an activation function for hidden layers to handle positive and negative changes in the input.

[0125] The ReLU function is shown in formula (11):

[0126] (11)

[0127] The output of the function ranges from 0 to positive infinity. When the input is positive, the activation function is linear, simple, and computationally efficient, avoiding the vanishing gradient problem and exhibiting good convergence performance. For negative inputs, the output is always zero, hence the term "dead zone." In most cases, ReLU is the most commonly used activation function, offering better training speed and accuracy.

[0128] In summary, the ReLU function, compared to the Sigmoid and Tanh functions, exhibits non-saturation, maintains a larger gradient, and enables faster gradient propagation and weight updates, thus accelerating the convergence speed of neural networks. Furthermore, parasitic capacitance is not considered in the case of negative values, making the ReLU function a suitable activation function.

[0129] Training Algorithm: The performance of training algorithms varies depending on the problem, requiring selection and comparison based on specific circumstances. This paper determines the optimal training algorithm for the neural network by comparing it with the LM algorithm, Bayesian algorithm, and quantized conjugate gradient algorithm.

[0130] By comparing the performance of three training algorithms Figure 12 , Figure 13 and Figure 14 The results show that the LM algorithm can reach the target error in fewer rounds compared to the Bayes algorithm and the quantized conjugate gradient algorithm, and the mean square error of the LM algorithm is smaller. Therefore, the LM algorithm is selected as the training algorithm for this neural network.

[0131] Step Six: Based on the differential equation relationship obtained in Step Four, determine the value function of the neural network digital twin algorithm, including the following steps:

[0132] Combination Figure 5 Explanation:

[0133] Step 6.1: Since each dimensionless parameter has an initial value after initialization, the differential equation can be solved using the ODE solver in MATLAB to obtain the per-unit voltage and angle across the circuit breaker. The waveform and its vertical and horizontal coordinate data;

[0134] Step 6.2: Find the maximum value in the y-axis data. :

[0135] This can be achieved using the max function;

[0136] Step 6.3: Divide the ordinate of each data point by its abscissa, find the maximum value, and use it as the slope. It passes through the origin and has a slope of The straight line with a vertical coordinate value of Time corresponding to x-coordinate ,Right now

[0137] (12)

[0138] Calculate all data points before the data point corresponding to the maximum value of the y-axis. Distance, distance Represented as

[0139] (13)

[0140] Find the maximum distance value The corresponding data point, passing through this data point and with a slope of The intersection of the straight line and the x-axis is denoted as . ,Right now

[0141] (14)

[0142] make , and These are the parameters that represent the waveform of the differential equation.

[0143] Step 6.4: As shown in Table 1, convert the dimensional standard values ​​for direct testing of high-voltage AC circuit breakers in GB / T 1984-2014 into dimensionless standard values, i.e., amplitude coefficients. and The result obtained in step 6.3 and The sum of the absolute values ​​of the differences between the individual values ​​is used as the value function of the neural network digital twin algorithm.

[0144] Table 1

[0145]

[0146] The value function is expressed as

[0147] (15)

[0148] Step 6.5: The standard stipulates that the envelope of the TRV waveform obtained from the test must be above the standard specified value. Therefore, weighting coefficients and penalty terms need to be added to the value function to meet the requirements.

[0149] Standards stipulate The value cannot be less than the specified value, delay As small as possible, therefore, It can be smaller than the standard ,but Cannot exceed Too much energy will generate a very large peak value, causing the circuit breaker arc to reignite and leading to test failure.

[0150] To ensure priority A weighting coefficient needs to be added before the item; at the same time, in order to... Not less than A penalty term, Penalty, needs to be added to the value function, and the final value function is expressed as follows:

[0151] (16)

[0152] Step 7: Perform neural network training multiple times to obtain the dimensionless parameters corresponding to the minimum value function. Convert these parameters to inductance, capacitance, and resistance values ​​using formulas. Perform Simulink simulation on these values. Obtain the waveform parameters from the obtained TRV data using the method described in Step 6. Compare these parameters with standard values ​​to verify the effectiveness of the method.

[0153] Due to the randomness of training results, the output will not be the same each time. To ensure the effectiveness of the method, it is necessary to run the neural network training multiple times, such as 500 or 1000 times, and find the minimum value function and its corresponding value function among all the output results. and .

[0154] After the conversion is complete, input the inductor, capacitor, and resistor values. Figure 11 In the simulation model, the simulation is run to obtain the TRV waveform curve and They are then compared with standard values ​​to verify the effectiveness of the method.

[0155] Table 2 shows the output results of this invention for each test mode under the circuit breaker's rated voltage of 24kV and rated short-circuit breaking current of 63kA. and The decision variables representing the input, that is Figure 2 In and Minimizing the error means outputting the minimum value function after multiple training iterations, where theta10 represents... Figure 7 In (b), the x-coordinate value corresponding to the y-coordinate being 1 is... , , , The table shows the output frequency modulation device parameter values, including parameter results for test modes T100, T60, T30, and T10.

[0156] Table 2

[0157]

[0158] Table 3 shows the output results of the present invention under various test modes with a rated voltage of 40.5kV and a rated short-circuit breaking current of 63kA. The table includes the parameter results for test modes T100, T60, T30 and T10.

[0159] Table 3

[0160]

[0161] Table 4 shows the output results of the present invention under various test modes with a rated voltage of 72.5kV and a rated short-circuit breaking current of 63kA. The table includes the parameter results for test modes T100, T60, T30 and T10.

[0162] Table 4

[0163]

[0164] Table 5 shows the extraction... Figure 9 The simulation waveforms and parameter results are shown in the table, which includes parameters for test modes T100, T60, T30, and T10. , and The result; The peak value of the transient recovery voltage across the circuit breaker. For reference time, For time delay.

[0165] Table 5

[0166]

[0167] Table 6 shows the extraction... Figure 11 The simulation waveforms and parameter results are shown in the table, which includes parameters for test modes T100, T60, T30, and T10. , and The result;

[0168] Table 6

[0169]

[0170] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A power distribution device self-tuning method based on a neural network digital twin algorithm, characterized in that, Comprise the following steps: Step 1, the three-phase direct test circuit is equivalent to a single-phase equivalent test circuit according to symmetry, taking the phase where the first open circuit breaker is located as the equivalent single phase, while the transformer and the reactor in the circuit are equivalent to inductance, and the power grid or the impulse generator is equivalent to a voltage source, to establish the model after equivalence; Step 2, according to Kirchhoff's current law KCL, Kirchhoff's voltage law KVL and the characteristics of inductive and capacitive elements, write the equation set and initial conditions of the equivalent model; Step 3, select the standard value of voltage and current, and convert the time domain to the angle domain, rewrite the equation set and initial conditions to obtain the equivalent model under dimensionless parameters; Step 4, according to the equivalent model under dimensionless parameters, arrange the differential equation relationship between the voltage per unit value of the test circuit breaker and other dimensionless parameters; Step 5, take the dimensionless parameters as the decision variables of the neural network digital twin algorithm; Step 6, determine the value function of the neural network digital twin algorithm according to the differential equation relationship; Step 7, train the neural network to obtain the dimensionless parameters corresponding to the minimum value function, convert to inductance, capacitance and resistance value through formula, then perform Simulink simulation, obtain the TRV data through the differential equation relationship in step 6 to obtain the waveform parameters, compare the waveform parameters with the standard value to verify the effectiveness of the method.

2. The power distribution device self-tuning method based on the neural network digital twin algorithm according to claim 1, characterized in that, Step 2 specifically: according to KCL, KVL and the characteristics of inductive and capacitive elements, the equation set and initial conditions of the equivalent model are ； where and are the currents flowing through the inductances and respectively, u s is the voltage source, C0and C1are the three-phase capacitances, R s is the load impedance, u tr is the transient recovery voltage across the circuit breaker, t is the time, and the * symbol indicates a dimensioned parameter.

3. The power distribution device self-tuning method based on the neural network digital twin algorithm according to claim 2, characterized in that, Step 3 specifically: select the standard value of voltage and current, and convert the time domain to the angle domain, rewrite the equation set and initial conditions to obtain the equivalent model under dimensionless parameters: ； wherein, is the maximum value in the ordinate data, is the natural angular frequency, t is time, is the power frequency recovery voltage peak value, u c is the transient recovery voltage peak value across the circuit breaker; Will , be the standard values of voltage and current, respectively, while replacing , the angle, is the product of the time and the undamped natural angular frequency of the test circuit , the dimensionless form of the system of equations (1) is: ； The following correspondence is selected: ； wherein , respectively represent dimensionless parameters of capacitance and resistance; According to formula (4), formula (3) can be arranged as: ； From formula (5), the equivalent model under dimensionless parameters is obtained.

4. The power distribution device self-tuning method based on the neural network digital twin algorithm according to claim 3, characterized in that, Step 4 specifically: according to the equivalent model under dimensionless parameters, arrange the differential equation relationship between the voltage per unit value of the test circuit breaker and other dimensionless parameters: ； The initial conditions of formula (6) are: ； Transient recovery voltage across the circuit breaker The first derivative, Transient recovery voltage across the circuit breaker The second derivative of .

5. The power distribution device self-tuning method based on the neural network digital twin algorithm according to claim 4, characterized in that, Step 5 specifically: take the dimensionless parameters as the decision variables of the neural network digital twin algorithm, that is, the number of input layer nodes, the number of hidden layers, the number of hidden layer nodes, the number of output layer nodes, and the initialization of weights and bias terms; For the determination of the number of hidden layer neurons, the empirical formula (8) is used to determine it; (8)； wherein, represents the number of hidden layer neurons, represents the number of input layer neurons, represents the number of output layer neurons, is an empirical constant with a value ranging from 1 to 10; Sigmoid, Tanh and ReLU activation functions are constructed; The Sigmoid function is shown in formula (9): (9)； In the formula, e represents the natural constant, and x represents the independent variable; The Tanh function is shown in formula (10): (10)； The ReLU function is shown in formula (11): (11) 。 6. The power distribution device self-tuning method based on the neural network digital twin algorithm according to claim 5, characterized in that, Step 6 specifically includes the following steps: Step 6.1, Solving the differential equation by ODE solver in MATLAB to get the waveform of the voltage in per unit and angle at the terminals of the circuit breaker and the data of the longitudinal coordinate and the horizontal coordinate ; Step 6.2 Find the maximum value in the ordinate data by the max function : Step 6.3: Divide the ordinate of each data point by the abscissa, find the maximum value in it, and take it as the slope The straight line passing through the origin and having a slope of corresponds to the abscissa when the ordinate value is , i.e. (12)； The distance of all data points before the data point corresponding to the maximum value of the longitudinal coordinate to is denoted as ​ (13)； Finding the maximum distance value The intersection of the line with slope passing through the data point and the abscissa is denoted by i.e. (14)； Let , and represent parameters of the differential equation waveform; Step 6.

4. Convert the dimensioned standard value into a dimensionless standard value, i.e. the amplitude factor and Step 6.

3. Calculate the difference between the dimensionless standard value and the dimensionless value of the measured variable and Step 6.

4. Calculate the sum of the absolute values of the differences between the dimensionless standard value and the dimensionless values of the measured variables, as the value function of the neural network digital twin algorithm. The value function is represented as (15)； Step 6.5, The envelope of the TRV waveform resulting from the test is above the standard requirement value. A weight coefficient and a penalty term Penalty are added to the value function, representing the weight coefficient, the final value function is expressed as (16)。 7. The power distribution device self-tuning method based on the neural network digital twin algorithm according to claim 6, characterized in that, Step 7 specifically: multiple neural network training is performed to obtain the dimensionless parameters corresponding to the minimum value function, which are converted to inductance, capacitance and resistance value through formula, and Simulink simulation is performed, the TRV data obtained through the method in step 6 is used to obtain the waveform parameters, which are compared with the standard value to verify the effectiveness of the method: The conversion relationship between dimensionless parameters and dimensional parameters is (17)； wherein is the first opening coefficient, is the rated voltage of the circuit breaker, I is the breaking current of the circuit breaker, , , is the angle corresponding to a dimensionless ordinate of the envelope of 1.

8. A computer apparatus comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program, when executed by the processor, causes the processor to perform the method of any one of claims 1 to 7. The processor executes the computer program to implement the steps of the method of claim 1.

9. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions, when executed by the processor, implement the steps of the method of claim 1.

10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions, when executed by the processor, implement the steps of the method of claim 1.

Images