Method for multi-parameter intelligent testing of transformer based on digital twinning

Abstract

This invention relates to the field of programmable control technology, specifically to a digital twin-based intelligent multi-parameter testing method for transformers. The method establishes an electromagnetic-thermal-fluid multi-physics coupling mechanism model of the transformer. Historical test data is input into a graph neural network to extract residual weight matrices, which are then superimposed onto the mechanism model coefficients to generate a digital twin. Using real-time acquired oil temperature, winding current, and top oil pressure as initial boundary conditions, an alternating iterative algorithm combining a long short-term memory network and a temporal finite difference method is employed to solve the multi-parameter evolution trajectory. The extreme points and inflection points of the evolution trajectory are extracted, and the amplitude, frequency, and application timing sequence of the excitation signal for the next test cycle are calculated using a model predictive control algorithm. This sequence is then sent to the power amplifier and signal generator for closed-loop adaptive adjustment of the test excitation. This invention eliminates the mismatch between the test excitation and the dynamic physical conditions of the equipment, reducing the cumulative deviation of multi-parameter synchronous test results from the actual physical state.

Description

Technical Field

[0001] This invention relates to the field of program control technology, and specifically to a digital twin-based intelligent multi-parameter testing method for transformers. Background Technology

[0002] Multi-parameter testing of transformers commonly employs an open-loop control architecture based on a pre-set test procedure. In conventional testing schemes, the test system outputs excitation signals such as voltage and current to the transformer under test according to a set fixed step size and timing sequence. Simultaneously, it collects response parameters such as transformer oil temperature, winding current, and top oil pressure through a sensor array. The generation process of the test excitation is isolated from the real-time physical state inside the device under test. The test system only switches the test flow of different parameters based on preset logic or time trigger conditions. The data collected by the sensors is only used for later result recording and offline analysis and is not fed back to the excitation signal generation stage to change the currently executing test sequence.

[0003] Based on the aforementioned open-loop fixed-timing implementation, the transformer exhibits a strong electromagnetic-thermal coupling time-varying effect when subjected to continuous test excitation. The energy injected in the preceding test alters the internal temperature field distribution and the physical state of the medium, causing the initial boundary conditions of subsequent parameter testing phases to drift. The fixed test sequence cannot detect the dynamic transition of the device under test's state, resulting in a mismatch between the actual excitation applied by the test system and the current true physical conditions of the device under test during subsequent parameter testing phases. This leads to a cumulative deviation in the results of multi-parameter synchronous tests from the true physical state. Summary of the Invention

[0004] The purpose of this invention is to provide a digital twin-based intelligent multi-parameter testing method for transformers, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A digital twin-based intelligent testing method for transformers with multiple parameters includes: establishing an electromagnetic-thermal-fluid multi-physics coupling mechanism model of the transformer; inputting historical test datasets into a graph neural network for residual feature extraction; and superimposing the residual weight matrix output by the graph neural network into the discretization coefficients of the partial differential equations of the multi-physics coupling mechanism model to generate a digital twin that integrates prior physical laws. The transformer's oil temperature, winding current, and top oil pressure parameters are collected in real time by a multi-source sensor array and input into the digital twin. An alternating iterative algorithm using a long short-term memory network and a temporal finite difference method is employed, with the currently collected parameters as the initial boundary conditions, to forward solve the multi-parameter evolution trajectory within a future set time window. The extreme points and inflection points in the multi-parameter evolution trajectory are extracted, and the excitation signal amplitude, frequency and application timing sequence for the next test cycle are generated through a model predictive control algorithm. The amplitude, frequency, and application timing sequence of the excitation signal are sent to the power amplifier and signal generator of the physical test terminal for closed-loop adaptive adjustment of the test excitation.

[0006] Preferably, the establishment of the electromagnetic-thermal-fluid multiphysics coupling mechanism model of the transformer includes: obtaining the geometric structure and material property parameters of the transformer, establishing a basic electromagnetic field distribution model based on Maxwell's equations, establishing a basic heat conduction model based on Fourier's law of heat conduction, establishing a basic fluid field model based on Navier-Stokes equations, and bidirectionally coupling the basic electromagnetic field distribution model, the basic heat conduction model, and the basic fluid field model through boundary coupling conditions; The step of inputting the historical test dataset into a graph neural network for residual feature extraction includes: constructing the historical test dataset into a graph topology according to the time series, using sensor nodes as graph nodes, and using the physical distance and electrical correlation between sensor nodes as edge weights, extracting the implicit nonlinear residual features in the graph topology through graph convolutional layers, and outputting the residual weight matrix.

[0007] Preferably, the method of using a long short-term memory network and a temporal finite difference method to solve the multi-parameter evolution trajectory within a future set time window includes: dividing the future set time window into multiple discrete time steps; in each discrete time step, calling the long short-term memory network, taking the current collected parameters and the hidden state vector of the previous discrete time step as input, and outputting the macroscopic state prediction value of the current discrete time step. The macroscopic state prediction value is used as the algebraic boundary condition of the time-series finite difference method to solve the partial differential equation by spatial discretization, thereby obtaining the microscopic field distribution data of the current discrete time step. The hidden state vector of the Long Short-Term Memory network is updated with the micro-field distribution data and the next discrete time step is entered. This process continues until all discrete time steps have been traversed. The micro-field distribution data of each discrete time step are then spliced ​​together to generate the multi-parameter evolution trajectory.

[0008] Preferably, the extraction of extreme points and inflection point features in the multi-parameter evolution trajectory includes: performing first-order and second-order difference operations on the oil temperature curve, winding current curve and top oil pressure curve in the multi-parameter evolution trajectory, marking the time corresponding to the zero crossing point of the first-order difference result as the extreme point, and marking the time corresponding to the zero crossing point of the second-order difference result as the inflection point. The step of generating the excitation signal amplitude, frequency, and application timing sequence for the next test cycle through the model predictive control algorithm includes: using the physical quantities at the times corresponding to the extreme points and the inflection points as the optimization target reference trajectory of the model predictive control algorithm, using the hardware output limit of the physical test terminal as the constraint condition, and calculating the excitation signal amplitude, frequency, and application timing sequence for the next test cycle that minimizes the tracking error by solving a quadratic programming problem in the finite time domain.

[0009] Preferably, the closed-loop adaptive adjustment of sending the excitation signal amplitude, frequency and application timing sequence to the power amplifier and signal generator of the physical test terminal for test excitation includes: performing timestamp-aligned encapsulation on the excitation signal amplitude, frequency and application timing sequence, parsing the timestamp-aligned encapsulated data packet through a field-programmable gate array, and extracting the digital waveform lookup table corresponding to the excitation signal amplitude and frequency and the timer trigger instruction corresponding to the application timing sequence; The field-programmable gate array controls the digital waveform lookup table to write timing information to the data bus of the digital-to-analog converter according to the timer trigger instruction. After the digital-to-analog converter converts the digital waveform into an analog signal, it synchronously outputs the signal to the power amplifier and the signal generator, driving the power amplifier and the signal generator to change their output power and output waveform according to the applied timing sequence.

[0010] Preferably, the real-time acquisition of transformer oil temperature, winding current and top oil pressure parameters through a multi-source sensor array includes: deploying an edge sampling clock at each sensor node of the multi-source sensor array, and performing frequency calibration on the edge sampling clock by receiving a second pulse signal; Each of the sensor nodes performs synchronous sampling according to the calibrated edge sampling clock, and an absolute time tag and a sensor node physical location identifier are added to each sampling point data frame; The sampling point data frame with the attached absolute time tag and the physical location identifier of the sensor node is sent to the digital twin computing platform via a high-speed bus. The digital twin computing platform aligns the oil temperature, winding current and top oil pressure parameters at the same moment according to the absolute time tag, and maps the oil temperature, winding current and top oil pressure parameters to the corresponding spatial grid nodes of the digital twin according to the physical location identifier of the sensor node.

[0011] Preferably, after generating a digital twin that incorporates prior physical laws, the method further includes online fine-tuning of the digital twin: during the closed-loop adaptive adjustment of the test stimulus, the measured oil temperature, measured winding current, and measured top oil pressure at the current moment are acquired in real time, and the measured oil temperature, measured winding current, and measured top oil pressure are subtracted from the predicted oil temperature, predicted winding current, and predicted top oil pressure output by the digital twin at the same moment to generate a multidimensional residual vector; The L2 norm of the multidimensional residual vector is calculated. When the L2 norm exceeds a preset dynamic threshold, the measured data within a preset time period before the current time is extracted to construct a local incremental graph topology. The local incremental graph topology is input into the graph neural network for backpropagation training. The updated network parameters are used to recalculate and output the residual weight matrix, which replaces the current residual weight matrix.

[0012] Preferably, after acquiring the microscopic field distribution data of the current discrete time step, the method further includes adaptive step size adjustment: calculating the relative error between the microscopic field distribution data of the current discrete time step and the microscopic field distribution data of the previous discrete time step, and simultaneously calculating the truncation error generated by the temporal finite difference method solution within the current discrete time step. The relative error and the truncation error are weighted and summed to generate a comprehensive error index; When the comprehensive error index is greater than the upper limit of the preset stable interval, the time step of the next discrete time step is reduced to 1 / 2 of the time step of the current discrete time step and the spatial discretization grid is densified. When the comprehensive error index is less than the lower limit of the preset stable interval, the time step of the next discrete time step is expanded to twice the time step of the current discrete time step and the spatial discretization grid is sparsed in order to maintain the numerical stability of the alternating iterative algorithm.

[0013] Preferably, the step of calculating the excitation signal amplitude, frequency, and application timing sequence for the next test cycle that minimizes the tracking error by solving a quadratic programming problem in the finite time domain includes: constructing a multi-objective penalty function in the quadratic programming problem, wherein the multi-objective penalty function includes a tracking error penalty term, a hardware output change rate penalty term, and a transformer insulation thermal stress penalty term, wherein the hardware output change rate penalty term suppresses step changes in the excitation signal during adjacent test cycles, and the transformer insulation thermal stress penalty term is generated by nonlinear mapping based on the oil temperature extreme point; During the solution iteration, it is determined whether there is a conflict between the hardware output limit constraint of the physical test terminal and the gradient descent direction of the multi-objective penalty function. If there is a conflict, the Lagrange multiplier corresponding to the hardware output limit constraint is amplified by a preset factor and the solution is iterated again until the excitation signal amplitude, frequency and application timing sequence that satisfy the balance of each weight are obtained.

[0014] Preferably, after the digital-to-analog converter converts the digital waveform into an analog signal and outputs it synchronously to the power amplifier and the signal generator, it also includes hardware-level latch-up protection: injecting a high-frequency current transformer in series in the output circuit to monitor the actual output current and actual output voltage in real time and inputting them into the field-programmable gate array; The field-programmable gate array compares the actual output current and the actual output voltage with the theoretical given values ​​in the excitation signal amplitude, frequency and application timing sequence in real time. When the actual output current or the actual output voltage exceeds the safety envelope of the theoretically given value, the field-programmable gate array cuts off the digital waveform lookup table from writing an enable signal to the data bus of the digital-to-analog converter, triggers a hardware relay to disconnect the AC power supply circuit, and simultaneously controls the digital twin computing platform to force the internal state variables of the digital twin to revert to the initial boundary conditions of the previous test cycle.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a digital twin that integrates a multi-physics coupling mechanism model and graph neural network residual feature extraction, transforming test stimulus planning from open-loop preset to closed-loop adaptive adjustment based on forward evolution. An alternating iterative algorithm using a long short-term memory network and a temporal finite difference method, with real-time acquired parameters as initial boundary conditions, solves for a multi-parameter evolution trajectory containing extreme points and inflection points. The model predictive control algorithm then generates the amplitude, frequency, and application timing sequence of the excitation signal matching the current dynamic physical state of the equipment. This technique eliminates the mismatch between test stimulus and equipment dynamic operating conditions, avoids the drift of subsequent test boundary conditions caused by preceding test energy injection, enables continuous tracking of changes in the internal physical state of the equipment during the testing process, and reduces the cumulative deviation of multi-parameter synchronous test results from the true physical state. 2. By synchronizing the edge sampling clock with the second pulse signal and attaching absolute time stamps and position identifiers to the data frames, the spatiotemporal alignment accuracy of multi-source sensor data input to the digital twin is ensured; the online fine-tuning mechanism dynamically updates the residual weight matrix through multi-dimensional residual vectors, enabling the digital twin to correct model drift during long-term testing; the adaptive step size adjustment dynamically changes the time step size and spatial grid density based on the comprehensive error index, maintaining the numerical stability of the alternating iterative algorithm; the multi-objective penalty function combined with Lagrange multiplier conflict handling balances the tracking error and the hardware output change rate, suppressing excitation step jumps between adjacent test cycles; hardware-level latching protection cuts off data write enable and disconnects the AC power supply circuit when the output exceeds the limit, preventing the risk of equipment damage caused by the transmission of digital spatial calculation anomalies to the physical test terminal. Attached Figure Description

[0016] Figure 1 This is the main flowchart of the intelligent multi-parameter testing method for transformers based on digital twins of the present invention; Figure 2 This is a flowchart of the digital twin construction and residual feature extraction process of the present invention; Figure 3 This is a flowchart illustrating the iterative solution process of the Long Short-Term Memory network and the Temporal Finite Difference Method of this invention. Figure 4 The flowchart for feature extraction and model prediction control generation of the excitation sequence in this invention is shown below; Figure 5 This is a flowchart illustrating the excitation signal transmission and synchronous acquisition by multiple sensors according to the present invention. Figure 6 This is a flowchart of the online fine-tuning and adaptive step size adjustment of the digital twin of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] Please refer to Figure 1This embodiment provides a digital twin-based intelligent testing method for multiple parameters of transformers. First, it establishes a multi-physics coupling mechanism model of the transformer based on electromagnetic, thermal, and fluid fields. Specifically, it acquires the geometric and material property parameters of the transformer under test. Geometric parameters include the core lamination dimensions, winding turns and wire diameter, internal tank dimensions, cooling oil channel distribution dimensions, and the spatial location and dimensions of the insulating components. Material property parameters include the core's permeability and iron loss coefficient, the winding copper's electrical and thermal conductivity, the transformer oil's dynamic viscosity and specific heat capacity, and the insulating components' dielectric constant and thermal conductivity. A fundamental electromagnetic field distribution model is established based on Maxwell's equations. This model is used to solve for the electric field strength, magnetic induction intensity, and current density distribution inside the transformer under the action of an excitation signal, as well as the corresponding Joule loss and hysteresis loss distribution. A fundamental heat conduction model is established based on Fourier's law of heat conduction. This model is used to solve for the temperature field distribution of solid components and cooling fluid inside the transformer, where the volumetric heat source term is provided by the Joule loss and hysteresis loss obtained from the fundamental electromagnetic field distribution model. A fundamental fluid field model is established based on the Navier-Stokes equations. This model is used to solve for the velocity and pressure distribution of cooling oil inside the transformer, and the temperature boundary conditions of the fluid field are provided by the solid wall temperature obtained from the fundamental heat conduction model. The fundamental electromagnetic field distribution model, the fundamental heat conduction model, and the fundamental fluid field model are bidirectionally coupled through boundary coupling conditions to form an electromagnetic-thermal-fluid multiphysics coupling mechanism model.

[0019] Table 1. Boundary Coupling Conditions for Transformer Multiphysics Coupled Model Table 1 lists the boundary coupling variables, coupling directions, mathematical expressions, and physical meanings between the physical fields during the multiphysics coupling process. This is used to clarify the specific implementation of bidirectional coupling and ensure the convergence of the numerical solution and physical consistency of the multiphysics coupling mechanism model.

[0020] In this embodiment, the differential form of Maxwell's equations is: in, The magnetic field strength, For current density, It is the electric displacement vector. For electric field strength, It represents the magnetic flux density. Free charge volume density, It is a time variable.

[0021] The expression for the Fourier heat conduction equation is: in, For material density, For isobaric specific heat capacity, For temperature, Thermal conductivity, is the volumetric heat source density.

[0022] The Navier-Stokes equations are expressed as follows: in, For fluid density, For fluid velocity vector, For fluid pressure, For fluid dynamic viscosity, This is the gravitational acceleration vector.

[0023] Historical test datasets are input into a graph neural network for residual feature extraction. The historical test dataset includes factory test data, type test data, and field operation test data for the same model of transformer under test. Data dimensions include oil temperature, winding current, top oil pressure, partial discharge, winding hot spot temperature, and core loss data under different excitation conditions. The historical test dataset is constructed as a graph topology according to time series, with the sensor nodes corresponding to the test data as graph nodes, and the physical distance and electrical correlation between sensor nodes as edge weights. Edge weights are calculated using a weighted summation method, with distance as the weight coefficient. The electrical correlation weight coefficient is set to 0.4, and the electrical correlation normalization value is set to 0.6. The electrical correlation normalization value is calculated from the electrical connection relationship and mutual inductance coefficient of the corresponding measurement position of the sensor node. The electrical correlation normalization value between current sensor nodes in the same winding is set to 1.0, and the electrical correlation normalization value between sensor nodes without electrical connection is set to 0.

[0024] The formula for calculating edge weights is: in, For nodes With nodes The edge weights between them For nodes With nodes The physical distance between them To prevent the minimum value where the denominator is zero, For nodes With nodes The normalized value of the electrical correlation between them.

[0025] The nonlinear residual features hidden in the graph topology are extracted using two graph convolutional layers. The first graph convolutional layer has 64 output channels, and the second graph convolutional layer has the same number of output channels as the coefficient matrix dimension after the finite element discretization of the multiphysics coupling mechanism model. The LeakyReLU activation function is used, with a negative slope of 0.2. The final output is the residual weight matrix. The propagation formula for the graph convolutional layer is: in, For the first The output feature matrix of layer graph convolution, It is a non-linear activation function. To add a self-loop adjacency matrix, for The degree matrix, For the first The input feature matrix of the layer, For the first The trainable weight matrix of the layer.

[0026] The residual weight matrix output from the graph neural network is superimposed onto the discretization coefficients of the partial differential equations of the multiphysics coupling mechanism model to generate a digital twin incorporating prior physical laws. Specifically, the partial differential equations of the multiphysics coupling mechanism model are spatially discretized using the finite element method to generate the original stiffness matrix, mass matrix, and load matrix. The residual weight matrix is ​​then used in an element-wise Hadamard product operation with the original stiffness matrix to obtain the corrected stiffness matrix. The corrected stiffness matrix, the original mass matrix, and the load matrix together constitute the core of the numerical solution for the digital twin. The residual weight matrix is ​​used to correct nonlinear errors in the multiphysics coupling mechanism model that cannot be fully described by first principles, including discretization errors of material properties, processing errors of geometric structures, and simplification errors of boundary conditions. The formula for calculating the corrected discretization coefficient matrix is ​​as follows: in, This is the corrected discretization coefficient matrix. The original stiffness matrix is ​​obtained by finite element discretization of the multiphysics coupling mechanism model. For Hadama accumulation, This is the residual weight matrix output by the graph neural network.

[0027] The transformer's oil temperature, winding current, and top oil pressure parameters are collected in real time using a multi-source sensor array and input into the digital twin. The multi-source sensor array includes fiber Bragg grating oil temperature sensors deployed at the bottom, middle, and top of the transformer tank; Rogowski coil current sensors deployed at the high-voltage and low-voltage winding input terminals; and a piezoresistive oil pressure sensor deployed at the top of the tank. The sampling data from each sensor node is transmitted to the digital twin computing platform via a high-speed Ethernet bus. The digital twin computing platform performs filtering preprocessing on the received sampling data to remove power frequency interference and random noise. The preprocessed data is then used as the input parameters for the digital twin.

[0028] An iterative algorithm combining Long Short-Term Memory (LSTM) networks and Temporal Finite Difference (TFD) is employed, using the currently acquired parameters as initial boundary conditions to forward solve the multi-parameter evolution trajectory within a future set time window. Specifically, the future set time window is divided into multiple discrete time steps, the length of which is determined based on the test cycle. In this embodiment, the set time window length is one test cycle, and the initial step size of the discrete time step is 1 ms. Within each discrete time step, the Long Short-Term Memory (LSTM) network is invoked, taking the currently acquired parameters and the hidden state vector of the previous discrete time step as input, and outputting the macroscopic state prediction value of the current discrete time step. The macroscopic state prediction value includes the transformer's average oil temperature, average winding current, and average top oil pressure. The macroscopic state prediction value is used as the algebraic boundary condition of the temporal finite difference method to solve the partial differential equations of the multiphysics coupling mechanism model in spatial discretization, obtaining the microscopic field distribution data of the current discrete time step. The microscopic field distribution data includes the temperature field distribution, electromagnetic field distribution, and fluid field distribution of the entire space grid inside the transformer. The hidden state vector of the LSM network is updated with the statistical characteristics of the microscopic field distribution data, and the process proceeds to the next discrete time step until all discrete time steps have been traversed. The microscopic field distribution data of each discrete time step are then concatenated in chronological order to generate a multi-parameter evolution trajectory within a set time window.

[0029] In this embodiment, the gating calculation formula for the Long Short-Term Memory network is as follows: Forget Gate Calculation Formula: in, for The output of the Forget Gate Here is the forget gate weight matrix. for The hidden state vector at time step 1. for The input vector at time t, Forget gate bias vector, It is the sigmoid activation function.

[0030] Input gate and cell state update formula: in, for The output of the input gate is always in use. The input gate weight matrix, This is the input gate bias vector. for The state of candidate cells at any given time. This is the cell state weight matrix. This is the cell state bias vector. for The constantly updated state of the cell. for The state of a cell at any given moment.

[0031] Output gate and hidden state update formula: in, for The output of the output gate is always available. This is the output gate weight matrix. This is the output gate bias vector. for The hidden state vector updated at each time step.

[0032] The discretization scheme for the heat conduction equation using the finite difference method of time is as follows: in, for Time step, spatial coordinates Temperature value at that location, for The temperature value at the corresponding spatial location for each time step. For time step, , , They are respectively , , Spatial grid step size in the direction, for The volumetric heat source density at the spatial location corresponding to the time step.

[0033] Extreme points and inflection points are extracted from the multi-parameter evolution trajectory. A model predictive control (MMC) algorithm is then used to generate the excitation signal amplitude, frequency, and application timing sequence for the next test cycle. Specifically, first-order and second-order difference operations are performed on the oil temperature curve, winding current curve, and top oil pressure curve in the multi-parameter evolution trajectory, respectively. The time corresponding to the zero-crossing point of the first-order difference result is marked as the extreme point, and the time corresponding to the zero-crossing point of the second-order difference result is marked as the inflection point. The extreme point corresponds to the moment when the parameter change trend changes from increasing to decreasing or from decreasing to increasing, and the inflection point corresponds to the extreme moment of the parameter change rate. The physical quantities corresponding to the extreme points and inflection points are used as the optimization target reference trajectory for the MMC algorithm. The hardware output limits of the physical test terminal are used as constraints. These hardware output limits include the maximum output power of the power amplifier, the maximum output voltage of the signal generator, and the frequency range. By solving a quadratic programming problem in the finite time domain, the excitation signal amplitude, frequency, and application timing sequence for the next test cycle that minimizes the tracking error are calculated.

[0034] The first-order difference and the formula for determining extreme points are: when At that time, the judgment Time is a curve The extreme points, among which For a single-parameter time series curve in a multi-parameter evolution trajectory, This represents the time interval of the sequence.

[0035] The formula for determining the second-order difference and inflection point is as follows: when At that time, the judgment Time is a curve The turning point.

[0036] The excitation signal amplitude, frequency, and application timing sequence are sent to the power amplifier and signal generator at the physical test terminal for closed-loop adaptive adjustment of the test excitation. Specifically, the excitation signal amplitude, frequency, and application timing sequence are timestamped and encapsulated. Each excitation parameter data packet is appended with a corresponding absolute trigger timestamp. The encapsulated data packets are transmitted to the field-programmable gate array (FPGA) at the physical test terminal via Gigabit Ethernet. The FPGA parses the received data packets, extracting a digital waveform lookup table corresponding to the excitation signal amplitude and frequency, and a timer trigger instruction corresponding to the application timing sequence. The digital waveform lookup table has a storage depth of 1024 bits and a data bit width of 16 bits, corresponding to the full-scale output range of the signal generator. Based on the timer trigger instruction, the FPGA controls the digital waveform lookup table to write the timing sequence to the data bus of the digital-to-analog converter (DAC). The DAC has a conversion rate of 1 MSPS. After converting the digital waveform into an analog signal, the DAC synchronously outputs it to the power amplifier and signal generator, driving the power amplifier and signal generator to change their output power and output waveform according to the application timing sequence, thus achieving closed-loop adaptive adjustment of the test excitation.

[0037] In this embodiment, a digital twin is constructed by integrating a multi-physics coupling mechanism model and graph neural network residual feature extraction, achieving high-precision mapping of the internal physical state of the transformer. The long short-term memory network and the temporal finite difference method are used to alternately iterate and solve the multi-parameter evolution trajectory. The model predictive control algorithm generates an excitation signal that matches the dynamic state of the transformer based on the features of the evolution trajectory, realizing closed-loop adaptive adjustment of the test excitation and eliminating the problem of mismatch between the test excitation and the dynamic operating conditions of the equipment.

[0038] In one alternative embodiment, refer to Figures 2 to 6When establishing the electromagnetic-thermal-fluid multiphysics coupling mechanism model of the transformer, finite element meshing was used for the geometric structural parameters. The meshing types included tetrahedral unstructured meshes and hexahedral structured meshes. The core and winding regions used hexahedral structured meshes with a mesh size of 2 mm, while the tank and insulation regions used tetrahedral unstructured meshes with a mesh size of 5 mm. After meshing, mesh independence verification was performed to ensure that the impact of mesh size changes on the numerical solution results was less than a preset threshold. When constructing the graph topology, the historical test dataset was preprocessed according to the test conditions. The preprocessing included outlier removal, data normalization, and time series alignment. Outlier removal adopted the 3σ criterion, data normalization adopted the maximum-minimum normalization method to map all test data to the [0,1] interval, and time series alignment adopted the linear interpolation method to unify test data with different sampling frequencies to the same time base. The training process of the graph neural network adopts a supervised training method, with a training set to test set ratio of 8:2. The loss function is the mean squared error loss function, the optimizer is the Adam optimizer, the initial learning rate is 0.001, and the training epochs are 200. After each training epoch, the test set is used for validation. Training stops when the validation loss no longer decreases for 10 consecutive epochs, and the network parameters after training are saved.

[0039] When acquiring transformer oil temperature, winding current, and top oil pressure parameters in real time using a multi-source sensor array, an edge sampling clock is deployed at each sensor node of the array. The edge sampling clock uses a temperature-controlled crystal oscillator with a frequency stability better than ±0.1ppm. The edge sampling clock is calibrated by receiving the second pulse signal output from the Global Navigation Satellite System (GNSS). The calibration period is 1 second. Within each calibration period, the local counter of each sensor node latches the rising edge of the second pulse signal, calculates the frequency deviation between the local clock frequency and the standard second pulse, and dynamically corrects the frequency division coefficient of the edge sampling clock based on this deviation, ensuring that the sampling clock frequency deviation of all sensor nodes is less than 10ns. Each sensor node performs synchronous sampling based on a calibrated edge sampling clock. The sampling frequency is determined according to the sensor type: 1kHz for the oil temperature sensor, 10kHz for the current sensor, and 5kHz for the oil pressure sensor. An absolute time stamp and a sensor node physical location identifier are appended to each sampling point data frame. The absolute time stamp uses Coordinated Universal Time (UTC) of the Global Navigation Satellite System (GNSS) with a time resolution of 1ns. The sensor node physical location identifier uses a 32-bit unsigned integer encoding, including the sensor type, installation spatial coordinates, and associated electrical circuit information. The sampling point data frames with the appended absolute time stamp and sensor node physical location identifier are sent to the digital twin computing platform via a high-speed serial computer extension bus. The digital twin computing platform aligns the oil temperature, winding current, and top oil pressure parameters at the same moment based on the absolute time stamp, with an alignment error of less than 50ns. It then maps the oil temperature, winding current, and top oil pressure parameters to the corresponding spatial grid nodes of the digital twin based on the sensor node physical location identifier, achieving a correspondence between the sampled data and the digital twin's spatial grid.

[0040] Table 2 Configuration Table of Synchronous Sampling Parameters for Multi-Source Sensor Array Table 2 lists the core parameter configurations for synchronous sampling of various types of sensors in the multi-source sensor array. These parameters are used to ensure the time synchronization accuracy and spatial mapping accuracy of the multi-source data, and to ensure the spatiotemporal consistency of the input data of the digital twin.

[0041] After generating a digital twin that incorporates prior physical laws, the digital twin is fine-tuned online. Specifically, during the closed-loop adaptive adjustment of the test stimulus, the measured oil temperature, measured winding current, and measured top oil pressure at the current moment are acquired in real time. The difference between the measured oil temperature, measured winding current, and measured top oil pressure and the predicted oil temperature, predicted winding current, and predicted top oil pressure output by the digital twin at the same moment is calculated to generate a multidimensional residual vector. The L2 norm of the multidimensional residual vector is calculated, and a preset dynamic threshold is dynamically adjusted according to the excitation amplitude of the test condition. The larger the excitation amplitude, the larger the preset dynamic threshold. When the L2 norm exceeds the preset dynamic threshold, the measured data within a preset time period before the current moment is extracted to construct a local incremental graph topology. The preset time period is 10 sampling periods, and the rules for constructing the node and edge weights of the local incremental graph topology are consistent with the graph topology of the historical test dataset. The local incremental graph topology is input into the graph neural network for backpropagation training. The learning rate is 0.0001, the training epochs are 10, and a mini-batch training method is used with a batch size of 32. The updated network parameters are used to recalculate and output the residual weight matrix, which replaces the current residual weight matrix, thus completing the online fine-tuning of the digital twin and correcting the model drift that occurred during long-term testing.

[0042] The formula for calculating the L2 norm of the multidimensional residual vector is: in, The L2 norm of the multidimensional residual vector. For parameter dimensions, For the first Measured values ​​of the dimension parameter. For the first The predicted value of the digital twin with dimensional parameters.

[0043] After acquiring the microscopic field distribution data for the current discrete time step using an alternating iterative algorithm employing a Long Short-Term Memory (LSTM) network and a temporal finite-difference (FDD) method, an adaptive step size adjustment is performed. Specifically, the relative error between the microscopic field distribution data of the current discrete time step and the microscopic field distribution data of the previous discrete time step is calculated, using the relative rate of change of the maximum temperature across the entire field. Simultaneously, the truncation error generated by the FDD solution within the current discrete time step is calculated, obtained from the Taylor expansion remainder term of the FDD scheme. The relative error and the truncation error are then weighted and summed to generate a comprehensive error index, with the relative error weighting coefficient... The truncation error weighting coefficient is set to 0.4. The upper limit of the preset stability interval is set to... The lower limit value is taken When the overall error index exceeds the upper limit of the preset stability interval, the time step of the next discrete time step is reduced to half the time step of the current discrete time step, and the spatial discretization grid is refined, with the grid size reduced to half the current size. When the overall error index is less than the lower limit of the preset stability interval, the time step of the next discrete time step is expanded to twice the time step of the current discrete time step, and the spatial discretization grid is sparsed, with the grid size expanded to twice the current size, in order to maintain the numerical stability of the alternating iterative algorithm and balance the solution accuracy and computational efficiency.

[0044] The formula for calculating the comprehensive error index is: in, As a comprehensive error index, The relative error of the microscopic field distribution data at adjacent time steps. This is the truncation error obtained by the finite difference method for the current time step.

[0045] In this embodiment, the time synchronization accuracy of multi-source sensor data is ensured by calibrating the edge sampling clock and the second pulse signal synchronously, and the sampling data and the spatial grid of the digital twin are accurately mapped by the physical location identification code. The online fine-tuning mechanism realizes the dynamic correction of the digital twin by monitoring the multi-dimensional residual vector and incrementally training the graph neural network, thus maintaining the prediction accuracy of the model. The adaptive step size adjustment mechanism dynamically adjusts the time step size and spatial grid density by calculating the comprehensive error index, which improves the computational efficiency of the alternating iterative algorithm while ensuring the solution accuracy.

[0046] In another optional embodiment, when extracting extreme points and inflection point features from the multi-parameter evolution trajectory, the oil temperature curve, winding current curve, and top oil pressure curve in the multi-parameter evolution trajectory are first subjected to wavelet threshold denoising processing. The wavelet basis function is db4 wavelet, the decomposition level is 5, and the threshold function is a soft threshold function to remove high-frequency noise in the evolution trajectory and avoid misjudgment of extreme points and inflection points caused by noise. First-order and second-order difference operations are performed on the denoised curves respectively. The time interval of the first-order difference operation is consistent with the discrete time step. The time corresponding to the zero crossing of the first-order difference result is marked as the extreme point, and the parameter amplitude and direction of change corresponding to the extreme point are recorded. The time corresponding to the zero crossing of the second-order difference result is marked as the inflection point, and the parameter change rate and curvature direction corresponding to the inflection point are recorded. The extracted extreme points and inflection points are filtered for effectiveness, and duplicate feature points with adjacent time intervals less than a preset minimum interval are removed. The preset minimum interval is 10 discrete time steps. The filtered extreme points and inflection point features constitute the key nodes of the optimized target reference trajectory.

[0047] When generating the excitation signal amplitude, frequency, and application timing sequence for the next test cycle using a model predictive control algorithm, a multi-objective penalty function is constructed within a quadratic programming problem. This multi-objective penalty function includes a tracking error penalty term, a hardware output change rate penalty term, and a transformer insulation thermal stress penalty term. The tracking error penalty term minimizes the deviation between the system output and the reference trajectory. The hardware output change rate penalty term suppresses step changes in the excitation signal between adjacent test cycles, avoiding equipment shocks caused by drastic fluctuations in hardware output. The transformer insulation thermal stress penalty term is generated through a nonlinear mapping based on oil temperature extreme points; the higher the oil temperature extreme point, the greater the weight of the thermal stress penalty term, used to limit the accumulation of thermal stress in the transformer insulation system. During the solution iteration, it is determined whether there is a conflict between the hardware output limit constraint of the physical test terminal and the gradient descent direction of the multi-objective penalty function. The conflict determination criterion is that the dot product of the Jacobian matrix of the constraint and the gradient descent direction is less than zero. If there is a conflict, the Lagrange multiplier corresponding to the hardware output limit constraint is amplified by a preset factor, which is 10 times. The solution is iterated again until the amplitude, frequency and application timing sequence of the excitation signal that satisfy the balance of various weights are obtained.

[0048] The expression for the multi-objective penalty function is: in, Let the objective function be a quadratic programming problem. To predict the length of the time domain, for The system output at any given time, for Reference trajectory at any moment for Rate of change of control quantity at time for Transformer insulation thermal stress at any given time , , These are the weight diagonal matrices for the corresponding terms.

[0049] When the excitation signal amplitude, frequency, and application timing sequence are sent to the power amplifier and signal generator of the physical test terminal for closed-loop adaptive adjustment of the test excitation, the excitation signal amplitude, frequency, and application timing sequence are timestamped and encapsulated. The encapsulation format adopts the User Datagram Protocol (UDP), and each data packet has a fixed length of 128 bytes. The first 8 bytes are the absolute trigger timestamp, the middle 64 bytes are the excitation signal amplitude and frequency parameters, and the last 56 bytes are the application timing sequence parameters and cyclic redundancy check (CRC) code. The timestamped and encapsulated data packets are parsed by the Field Programmable Gate Array (FPGA). First, the data packets are subjected to CRC. Data packets that fail the check are discarded. For data packets that pass the check, the digital waveform lookup table corresponding to the excitation signal amplitude and frequency, as well as the timer trigger instruction corresponding to the application timing sequence, are extracted. The digital waveform lookup table is generated based on the direct digital frequency synthesis principle. The step size value of the phase accumulator is calculated according to the frequency parameter of the excitation signal, and the waveform sample is scaled according to the amplitude parameter of the excitation signal to generate a sine wave digital sample sequence with corresponding amplitude and frequency, which is stored in the block random access memory of the FPGA. The field-programmable gate array (FPGA) controls the digital waveform lookup table to write timing information to the data bus of the digital-to-analog converter (DAC) based on timer trigger instructions. The time base of the timer trigger instructions is the same as the time base of the sensor sampling, ensuring that the excitation output and data acquisition are synchronized. After the DAC converts the digital waveform into an analog signal, it outputs it synchronously to the power amplifier and signal generator through differential signal lines. This drives the power amplifier and signal generator to change their output power and output waveform according to the applied timing sequence, realizing closed-loop adaptive adjustment of the test excitation.

[0050] After the digital-to-analog converter converts the digital waveform into an analog signal and synchronously outputs it to the power amplifier and signal generator, hardware-level latch-up protection is implemented. Specifically, a high-frequency current transformer is injected in series in the output circuit, and a voltage sensor is connected in parallel across the output circuit to monitor the actual output current and voltage in real time. The monitoring data is input to the field-programmable gate array (FPGA) in parallel in real time, with a sampling frequency of 100kHz. The FPGA compares the actual output current and voltage with the amplitude and frequency of the excitation signal and the theoretical given values ​​in the applied timing sequence in real time, with a comparison period of 10μs. A safety envelope is generated based on theoretically given values. The upper limit of the safety envelope is 1.2 times the theoretically given value, and the lower limit is 0.8 times the theoretically given value. When the actual output current or actual output voltage exceeds the theoretically given value of the safety envelope, the field-programmable gate array immediately cuts off the digital waveform lookup table from writing an enable signal to the data bus of the digital-to-analog converter, stops the output of the digital waveform, and triggers a hardware relay to disconnect the AC power supply circuit, cuts off the input power of the power amplifier and signal generator, and synchronously controls the digital twin computing platform to force the internal state variables of the digital twin to revert to the initial boundary conditions of the previous test cycle, so as to avoid the calculation anomaly in the digital space from being transmitted to the physical test terminal and prevent damage to the transformer under test and the test equipment.

[0051] Table 3 Comparison of Hardware-Level Interlocking Protection Trigger Thresholds and Action Logic Table 3 lists the monitoring parameters, trigger thresholds, trigger conditions, and corresponding action execution logic of the hardware-level interlocking protection, which is used to clarify the implementation process of hardware protection and ensure equipment safety during the testing process.

[0052] In this embodiment, wavelet threshold denoising and feature point screening improve the accuracy of extreme point and inflection point feature extraction; the multi-objective penalty function and Lagrange multiplier conflict handling mechanism balance tracking error, hardware output impact and transformer insulation thermal stress, and achieve smooth and optimized generation of excitation signal; the hardware-level interlocking protection mechanism compares output parameters with safety envelope in real time, and quickly executes hardware cut-off and state rollback when output is abnormal, avoiding the risk of equipment damage during the test process and ensuring the safe operation of the test system.

Claims

1. A digital twin-based intelligent multi-parameter testing method for transformers, characterized in that, include: An electromagnetic-thermal-fluid multiphysics coupling mechanism model of a transformer is established. Historical test datasets are input into a graph neural network for residual feature extraction. The residual weight matrix output by the graph neural network is superimposed on the discretization coefficients of the partial differential equation of the multiphysics coupling mechanism model to generate a digital twin that integrates prior physical laws. The transformer's oil temperature, winding current, and top oil pressure parameters are collected in real time by a multi-source sensor array and input into the digital twin. An alternating iterative algorithm using a long short-term memory network and a temporal finite difference method is employed, with the currently collected parameters as the initial boundary conditions, to forward solve the multi-parameter evolution trajectory within a future set time window. The extreme points and inflection points in the multi-parameter evolution trajectory are extracted, and the excitation signal amplitude, frequency and application timing sequence for the next test cycle are generated through a model predictive control algorithm. The amplitude, frequency, and application timing sequence of the excitation signal are sent to the power amplifier and signal generator of the physical test terminal for closed-loop adaptive adjustment of the test excitation.

2. The intelligent multi-parameter testing method for transformers based on digital twins according to claim 1, characterized in that, The establishment of the electromagnetic-thermal-fluid multiphysics coupling mechanism model of the transformer includes: obtaining the geometric structure and material property parameters of the transformer; establishing a basic electromagnetic field distribution model based on Maxwell's equations; establishing a basic heat conduction model based on Fourier's heat conduction law; establishing a basic fluid field model based on Navier-Stokes equations; and bidirectionally coupling the basic electromagnetic field distribution model, the basic heat conduction model, and the basic fluid field model through boundary coupling conditions. The step of inputting the historical test dataset into a graph neural network for residual feature extraction includes: constructing the historical test dataset into a graph topology according to the time series, using sensor nodes as graph nodes, and using the physical distance and electrical correlation between sensor nodes as edge weights, extracting the implicit nonlinear residual features in the graph topology through graph convolutional layers, and outputting the residual weight matrix.

3. The intelligent multi-parameter testing method for transformers based on digital twins according to claim 1, characterized in that, The method of using a forward iterative algorithm of alternating long short-term memory network and temporal finite difference method to solve the multi-parameter evolution trajectory within a future set time window includes: dividing the future set time window into multiple discrete time steps; in each discrete time step, calling the long short-term memory network, taking the current collected parameters and the hidden state vector of the previous discrete time step as input, and outputting the macroscopic state prediction value of the current discrete time step. The macroscopic state prediction value is used as the algebraic boundary condition of the time-series finite difference method to solve the partial differential equation by spatial discretization, thereby obtaining the microscopic field distribution data of the current discrete time step. The hidden state vector of the Long Short-Term Memory network is updated with the micro-field distribution data and the next discrete time step is entered. This process continues until all discrete time steps have been traversed. The micro-field distribution data of each discrete time step are then spliced ​​together to generate the multi-parameter evolution trajectory.

4. The intelligent multi-parameter testing method for transformers based on digital twins according to claim 1, characterized in that, The extraction of extreme points and inflection point features in the multi-parameter evolution trajectory includes: performing first-order and second-order difference operations on the oil temperature curve, winding current curve and top oil pressure curve in the multi-parameter evolution trajectory respectively, marking the time corresponding to the zero crossing point of the first-order difference result as the extreme point, and marking the time corresponding to the zero crossing point of the second-order difference result as the inflection point. The step of generating the excitation signal amplitude, frequency, and application timing sequence for the next test cycle through the model predictive control algorithm includes: using the physical quantities at the times corresponding to the extreme points and the inflection points as the optimization target reference trajectory of the model predictive control algorithm, using the hardware output limit of the physical test terminal as the constraint condition, and calculating the excitation signal amplitude, frequency, and application timing sequence for the next test cycle that minimizes the tracking error by solving a quadratic programming problem in the finite time domain.

5. The intelligent multi-parameter testing method for transformers based on digital twins according to claim 1, characterized in that, The closed-loop adaptive adjustment of sending the excitation signal amplitude, frequency and application timing sequence to the power amplifier and signal generator of the physical test terminal for test excitation includes: performing timestamp-aligned encapsulation on the excitation signal amplitude, frequency and application timing sequence; parsing the timestamp-aligned encapsulated data packet through a field-programmable gate array; and extracting the digital waveform lookup table corresponding to the excitation signal amplitude and frequency and the timer trigger instruction corresponding to the application timing sequence. The field-programmable gate array controls the digital waveform lookup table to write timing information to the data bus of the digital-to-analog converter according to the timer trigger instruction. After the digital-to-analog converter converts the digital waveform into an analog signal, it synchronously outputs the signal to the power amplifier and the signal generator, driving the power amplifier and the signal generator to change their output power and output waveform according to the applied timing sequence.

6. The intelligent multi-parameter testing method for transformers based on digital twins according to claim 1, characterized in that, The real-time acquisition of transformer oil temperature, winding current and top oil pressure parameters through a multi-source sensor array includes: deploying an edge sampling clock at each sensor node of the multi-source sensor array, and calibrating the edge sampling clock by receiving a second pulse signal; Each of the sensor nodes performs synchronous sampling according to the calibrated edge sampling clock, and an absolute time tag and a sensor node physical location identifier are added to each sampling point data frame; The sampling point data frame with the attached absolute time tag and the physical location identifier of the sensor node is sent to the digital twin computing platform via a high-speed bus. The digital twin computing platform aligns the oil temperature, winding current and top oil pressure parameters at the same moment according to the absolute time tag, and maps the oil temperature, winding current and top oil pressure parameters to the corresponding spatial grid nodes of the digital twin according to the physical location identifier of the sensor node.

7. The intelligent multi-parameter testing method for transformers based on digital twins according to claim 2, characterized in that, After generating a digital twin that incorporates prior physical laws, the process also includes online fine-tuning of the digital twin: during the closed-loop adaptive adjustment of the test stimulus, the measured oil temperature, measured winding current, and measured top oil pressure at the current moment are acquired in real time. The measured oil temperature, measured winding current, and measured top oil pressure are subtracted from the predicted oil temperature, predicted winding current, and predicted top oil pressure output by the digital twin at the same moment to generate a multidimensional residual vector. The L2 norm of the multidimensional residual vector is calculated. When the L2 norm exceeds a preset dynamic threshold, the measured data within a preset time period before the current time is extracted to construct a local incremental graph topology. The local incremental graph topology is input into the graph neural network for backpropagation training. The updated network parameters are used to recalculate and output the residual weight matrix, which replaces the current residual weight matrix.

8. The intelligent multi-parameter testing method for transformers based on digital twins according to claim 3, characterized in that, After acquiring the microscopic field distribution data of the current discrete time step, the adaptive step size adjustment is also included: calculating the relative error between the microscopic field distribution data of the current discrete time step and the microscopic field distribution data of the previous discrete time step, and calculating the truncation error generated by the temporal finite difference method solution within the current discrete time step. The relative error and the truncation error are weighted and summed to generate a comprehensive error index; When the comprehensive error index is greater than the upper limit of the preset stable interval, the time step of the next discrete time step is reduced to 1 / 2 of the time step of the current discrete time step and the spatial discretization grid is densified. When the comprehensive error index is less than the lower limit of the preset stable interval, the time step of the next discrete time step is expanded to twice the time step of the current discrete time step and the spatial discretization grid is sparsed in order to maintain the numerical stability of the alternating iterative algorithm.

9. The intelligent multi-parameter testing method for transformers based on digital twins according to claim 4, characterized in that, The step of calculating the excitation signal amplitude, frequency, and application timing sequence for the next test cycle that minimizes the tracking error by solving a quadratic programming problem in the finite time domain includes: constructing a multi-objective penalty function in the quadratic programming problem, wherein the multi-objective penalty function includes a tracking error penalty term, a hardware output change rate penalty term, and a transformer insulation thermal stress penalty term, wherein the hardware output change rate penalty term suppresses step changes in the excitation signal during adjacent test cycles, and the transformer insulation thermal stress penalty term is generated by nonlinear mapping based on the oil temperature extreme point; During the solution iteration, it is determined whether there is a conflict between the hardware output limit constraint of the physical test terminal and the gradient descent direction of the multi-objective penalty function. If there is a conflict, the Lagrange multiplier corresponding to the hardware output limit constraint is amplified by a preset factor and the solution is iterated again until the excitation signal amplitude, frequency and application timing sequence that satisfy the balance of each weight are obtained.

10. The intelligent multi-parameter testing method for transformers based on digital twins according to claim 5, characterized in that, After the digital-to-analog converter converts the digital waveform into an analog signal and outputs it synchronously to the power amplifier and the signal generator, it also includes hardware-level latch-up protection: a high-frequency current transformer is injected in series in the output circuit to monitor the actual output current and actual output voltage in real time and input them into the field programmable gate array; The field-programmable gate array compares the actual output current and the actual output voltage with the theoretical given values ​​in the excitation signal amplitude, frequency and application timing sequence in real time. When the actual output current or the actual output voltage exceeds the safety envelope of the theoretically given value, the field-programmable gate array cuts off the digital waveform lookup table from writing an enable signal to the data bus of the digital-to-analog converter, triggers a hardware relay to disconnect the AC power supply circuit, and simultaneously controls the digital twin computing platform to force the internal state variables of the digital twin to revert to the initial boundary conditions of the previous test cycle.

Images