A transformer loss detection method based on CNN-BiLSTM and SFG

Abstract

This invention provides a transformer loss detection method based on CNN-BiLSTM and SFG, belonging to the field of transformer technology. This invention establishes a finite element simulation model of a converter transformer, including harmonic excitation and DC bias conditions, using the SFG adaptive sparse mesh partitioning algorithm. It generates a training dataset of three-phase voltage and current waveforms and magnetic field loss temperature distribution. A physical constraint model is constructed, integrating a CNN feature extraction layer, a BiLSTM time-dependent layer, and Joule's law constraint terms embedded in Maxwell's equations for heat conduction. The model outputs predicted loss results using real-time acquired measured three-phase voltage and current waveforms as input, and sends adjustment commands to the monitoring system or cooling control system based on the loss difference. This solves the technical problem of difficulty in real-time and accurate detection of the loss state during converter transformer operation.

Description

Technical Field

[0001] This invention belongs to the field of transformer technology, and more specifically, relates to a transformer loss detection method based on CNN-BiLSTM and SFG. Background Technology

[0002] Converter transformers play a crucial role in power systems, performing voltage level conversion and energy transmission. Traditional loss detection methods primarily employ offline testing or estimation based on empirical formulas. These methods obtain iron and copper loss parameters through no-load and short-circuit tests while the transformer is out of service, or estimate losses using empirical formulas based on load current and temperature rise data. However, current technologies struggle to capture the transient loss characteristics of converter transformers during dynamic operation due to the complex and variable operating conditions, including harmonic excitation and DC bias. Furthermore, traditional methods rely on manual measurement and offline analysis, making it difficult to achieve continuous online monitoring of loss status. In other words, existing technologies suffer from the technical challenge of accurately and in real-time detecting the loss status of converter transformers during operation. Summary of the Invention

[0003] In view of this, the present invention provides a transformer loss detection method based on CNN-BiLSTM and SFG, which can solve the technical problem in the prior art that it is difficult to detect the loss status of converter transformers in real time and accurately.

[0004] This invention is implemented as follows: It provides a transformer loss detection method based on CNN-BiLSTM and SFG to obtain converter transformer parameters, establish a finite element simulation model of the converter transformer, and simulate various operating conditions by applying an input voltage source to the external circuit. The SFG mesh generation algorithm is used to mesh the finite element simulation model of the converter transformer and solve for three-phase voltage waveform data, three-phase current waveform data, magnetic field distribution data, iron loss data, copper loss data, and temperature distribution data. After time synchronization and standardization of the three-phase voltage and current waveform data, they are divided into fixed-length time-series sample sequences using a sliding window method. A physical constraint model is then constructed to divide the fixed-length time-series sample sequences. Using time-series sample sequences as input data, magnetic field distribution data, iron loss data, copper loss data, and temperature distribution data as output target data, a training dataset is established for training to obtain a trained physical constraint model. The measured waveforms of three-phase voltage and three-phase current during transformer operation are collected in real time and input into the trained physical constraint model. The output of the physical constraint model is the predicted magnetic field distribution result, predicted iron loss result, predicted copper loss result, and predicted temperature distribution result. The iron loss difference between the predicted iron loss result and the preset iron loss threshold and the copper loss difference between the predicted copper loss result and the preset copper loss threshold are calculated, and the larger value is selected as the loss difference. When the loss difference is greater than the preset difference threshold, a load adjustment command or a cooling enhancement command is sent.

[0005] This invention constructs a physical constraint model that integrates a convolutional neural network and a bidirectional long short-term memory network. It trains the model on finite element simulation data established using an adaptive sparse mesh partitioning algorithm, mapping real-time acquired three-phase voltage and current waveforms into predicted results for magnetic field distribution, iron and copper losses, and temperature distribution. The method utilizes a CNN feature extraction layer to capture local pattern features of the voltage and current waveforms, and a BiLSTM time-dependent layer to learn the long-term temporal laws of loss evolution. Maxwell's equations, the heat conduction equation, and Joule's law constraint terms are embedded in the physical constraint fusion layer to ensure that the predicted results conform to the fundamental laws of electromagnetic fields and thermodynamics. This achieves real-time and accurate detection of loss status based on online electrical quantity measurement. In summary, this invention solves the technical problem mentioned in the background art of the difficulty in real-time and accurate detection of loss status during converter transformer operation. Attached Figure Description

[0006] Figure 1 This is a complete flowchart of the transformer loss detection method in the embodiment.

[0007] Figure 2 A spatial distribution map of the selected monitoring points for the transformer model in the embodiment is provided.

[0008] Figure 3 The image shows a three-dimensional cloud map of the magnetic field distribution under different operating conditions in the embodiment.

[0009] Figure 4 This is a graph showing the convergence curve of the validation set loss function during the training process in the example. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0011] This invention provides a transformer loss detection method based on CNN-BiLSTM and SFG, comprising the following steps: S01. Obtain the core geometry, winding geometry, electromagnetic shielding structure dimensions, core magnetization curve, core loss curve, winding resistance, winding inductance, rated voltage, rated current, operating frequency, and short-circuit impedance of the converter transformer. Establish a finite element simulation model of the converter transformer. Set up an external circuit and apply an input voltage source in the finite element simulation model of the converter transformer to simulate no-load, short-circuit, and rated load conditions. Set harmonic excitation conditions and DC bias conditions. S02. The SFG mesh generation algorithm is used to mesh the finite element simulation model of the converter transformer. The transient solver is used to solve the finite element simulation model of the converter transformer to obtain three-phase voltage waveform data, three-phase current waveform data, magnetic field distribution data, iron loss data, copper loss data, and temperature distribution data under different operating conditions. The three-phase voltage waveform data and three-phase current waveform data are time-synchronized and standardized. The processed data are divided into fixed-length time-series sample sequences according to the sliding window method. S03. Construct a physical constraint model. Use the fixed-length time-series sample sequence as input data and the magnetic field distribution data, iron loss data, copper loss data, and temperature distribution data as output target data to establish a training dataset. Use the physical constraint model to train the training dataset to obtain the trained physical constraint model. S04. Install voltage and current sensors at the high-voltage side bushing outlet, low-voltage side bushing outlet, and neutral grounding terminal of the transformer to collect the measured waveforms of the three-phase voltage and three-phase current during transformer operation in real time, and input the measured waveforms of the three-phase voltage and three-phase current into the trained physical constraint model. S05. The trained physical constraint model outputs the predicted magnetic field distribution result, the predicted iron loss result, the predicted copper loss result, and the predicted temperature distribution result. The iron loss difference between the predicted iron loss result and the preset iron loss threshold and the copper loss difference between the predicted copper loss result and the preset copper loss threshold are calculated. The larger value between the iron loss difference and the copper loss difference is selected as the loss difference. S06. When the loss difference is greater than the preset difference threshold, a load adjustment command is sent to the transformer monitoring system or a cooling enhancement command is sent to the transformer cooling control system. The load adjustment command includes a suggested reduction in load power percentage, and the cooling enhancement command is used to start the standby cooling fan or increase the oil pump speed.

[0012] The SFG mesh generation algorithm is an adaptive sparse mesh generation algorithm. The steps of the SFG mesh generation algorithm include: performing initial mesh generation on the finite element simulation model of the converter transformer to obtain an initial mesh element set; calculating the magnetic field gradient value of each initial mesh element in the initial mesh element set; selecting initial mesh elements with magnetic field gradient values ​​greater than a magnetic field gradient threshold as high-gradient mesh elements; performing a first mesh refinement process on the high-gradient mesh elements to obtain a first-refinement mesh element set; calculating the eddy current density value of each first-refinement mesh element in the first-refinement mesh element set; selecting first-refinement mesh elements with eddy current density values ​​greater than an eddy current density threshold as high-eddy current mesh elements; performing a second mesh refinement process on the high-eddy current mesh elements to obtain a final mesh element set; and applying surface boundary impedance conditions to the surfaces of the tank wall structure, clamping structure, and tension plate structure in the final mesh element set.

[0013] The method for obtaining the magnetic field gradient threshold is as follows: The magnetic field gradient values ​​of all initial grid cells in the initial grid cell set are sorted in descending order, and the magnetic field gradient values ​​corresponding to the top 30% of the sorted values ​​are selected as the magnetic field gradient threshold. The numerical range of the magnetic field gradient threshold is 0.5. Up to 5.0 The preferred value is 2.0. .

[0014] The method for obtaining the eddy current density threshold is as follows: The eddy current density values ​​of all first-stage encrypted grid cells in the first-stage encrypted grid cell set are sorted in descending order, and the eddy current density value corresponding to the top 25% position after sorting is selected as the eddy current density threshold. The numerical range of the eddy current density threshold is... to The preferred value is .

[0015] The surface boundary impedance condition is applied as follows: The conductivity values ​​of the box wall structural components, clamping structural components, and pull plate structural components are obtained; the magnetic permeability values ​​of the box wall structural components, clamping structural components, and pull plate structural components are also obtained; the skin depth values ​​of the box wall structural components, clamping structural components, and pull plate structural components are calculated based on the operating frequency value; the skin depth values ​​of the box wall structural components, clamping structural components, and pull plate structural components are then combined with the conductivity and magnetic permeability values ​​of the box wall structural components to obtain the surface boundary impedance condition of the box wall. Surface impedance value; the surface impedance value of the clamping component is calculated by combining the skin depth value of the clamping component material with the electrical conductivity and magnetic permeability values ​​of the clamping component material; the surface impedance value of the pull plate is calculated by combining the skin depth value of the pull plate component material with the electrical conductivity and magnetic permeability values ​​of the pull plate component material; the surface impedance value of the tank wall is applied to the surface of the box wall component, the surface impedance value of the clamping component is applied to the surface of the clamping component, and the surface impedance value of the pull plate is applied to the surface of the pull plate component in the finite element simulation model of the converter transformer.

[0016] The harmonic excitation conditions are set as follows: 2nd, 3rd, 5th, 7th, 11th, and 13th harmonic components are superimposed on the input voltage source. The amplitude of the 2nd harmonic component is 3% to 8% of the fundamental voltage amplitude, the amplitude of the 3rd harmonic component is 3% to 8% of the fundamental voltage amplitude, the amplitude of the 5th harmonic component is 3% to 8% of the fundamental voltage amplitude, the amplitude of the 7th harmonic component is 3% to 8% of the fundamental voltage amplitude, the amplitude of the 11th harmonic component is 3% to 8% of the fundamental voltage amplitude, and the amplitude of the 13th harmonic component is 3% to 8% of the fundamental voltage amplitude. The phase angle of each harmonic component is calculated based on the converter firing angle.

[0017] The DC bias condition is set as follows: a DC current component is injected into the neutral point of the finite element simulation model of the converter transformer, and the amplitude of the DC current component is 5% to 15% of the rated current value, preferably 10% of the rated current value.

[0018] The time synchronization process is as follows: extract the timestamps of the three-phase voltage waveform data and the three-phase current waveform data, and use linear interpolation to fill in the misaligned data points so that the sampling times of the three-phase voltage waveform data and the three-phase current waveform data are completely corresponding.

[0019] The standardization process employs a channel-by-channel standardization method. The mean and standard deviation of the A-phase, B-phase, and C-phase voltage waveforms are calculated separately. Similarly, the mean and standard deviation of the A-phase, B-phase, and C-phase current waveforms are calculated separately. The standardized A-phase voltage waveform data is obtained by subtracting the mean from the A-phase voltage waveform data and then dividing by the standard deviation. The standardized B-phase voltage waveform data is obtained by subtracting the mean from the B-phase voltage waveform data and then dividing by the standard deviation. The standardized phase B voltage waveform data is obtained by subtracting the mean of phase C voltage from the phase C voltage waveform data and then dividing by the standard deviation of phase C voltage. The standardized phase C voltage waveform data is obtained by subtracting the mean of phase A current from the phase A current waveform data and then dividing by the standard deviation of phase A current. The standardized phase B current waveform data is obtained by subtracting the mean of phase B current from the phase B current waveform data and then dividing by the standard deviation of phase B current. The standardized phase C current waveform data is obtained by subtracting the mean of phase C current from the phase C current waveform data and then dividing by the standard deviation of phase C current.

[0020] The sliding window method has a window length of 1000 sampling points and a sliding step size of 500 sampling points. The fixed-length time-series sample sequence is obtained by segmenting the standardized A-phase voltage waveform data, standardized B-phase voltage waveform data, standardized C-phase voltage waveform data, standardized A-phase current waveform data, standardized B-phase current waveform data, and standardized C-phase current waveform data according to the window length and sliding step size.

[0021] The structure of the physical constraint model includes an input layer, a CNN feature extraction layer, a BiLSTM temporal dependency layer, a physical constraint fusion layer, and a multi-task output layer.

[0022] The input layer receives the fixed-length time-series sample sequence, which contains six-dimensional time-series data consisting of standardized A-phase voltage waveform data, standardized B-phase voltage waveform data, standardized C-phase voltage waveform data, standardized A-phase current waveform data, standardized B-phase current waveform data, and standardized C-phase current waveform data.

[0023] The CNN feature extraction layer includes a first convolutional sub-layer, a second convolutional sub-layer, and a third convolutional sub-layer. The first convolutional sub-layer contains 32 first convolutional kernels with a kernel size of 3. The second convolutional sub-layer contains 32 second convolutional kernels with a kernel size of 5. The third convolutional sub-layer contains 32 third convolutional kernels with a kernel size of 7. The first, second, and third convolutional sub-layers are respectively connected to a first batch of normalization layers, a second batch of normalization layers, and a third batch of normalization layers. The first, second, and third batch of normalization layers are respectively connected to a first max pooling layer, a second max pooling layer, and a third max pooling layer.

[0024] The BiLSTM timing-dependent layer includes a first bidirectional long short-term memory (BSSM) layer and a second bidirectional BSSM layer. The first BSSM layer contains 128 first hidden units, and the second BSSM layer contains 128 second hidden units. The first BSSM layer is followed by a first discard layer, and the second BSSM layer is followed by a second discard layer.

[0025] The physical constraint fusion layer embeds constraints from Maxwell's equations, the heat conduction equation, and Joule's law.

[0026] The multi-task output layer includes a magnetic field prediction branch, a loss prediction branch, and a temperature prediction branch. The magnetic field prediction branch outputs the X-direction component value, the Y-direction component value, and the Z-direction component value of the magnetic field. The loss prediction branch outputs the predicted iron loss result and the predicted copper loss result. The temperature prediction branch outputs the temperature values ​​of the winding hot spot location, the core column center location, the core yoke location, the inner surface of the box wall location, the upper oil temperature location, and the lower oil temperature location.

[0027] The steps for establishing the training dataset for the physical constraint model include: extracting the three-phase voltage waveform data and the three-phase current waveform data from the solution results of the converter transformer finite element simulation model as input samples, and extracting the magnetic field distribution data, iron loss data, copper loss data, and temperature distribution data as output samples; performing timestamp alignment on the input samples and the output samples; performing standardization processing on the input samples and the output samples respectively, with a mean of 0 and a standard deviation of 1; dividing the standardized input samples into sequences of 1000 sampling points, and simultaneously dividing the corresponding output samples; dividing the segmented input samples and output samples into training sets and validation sets in an 8:2 ratio; and adding normally distributed noise to the input samples in the training set for data augmentation, with the noise amplitude being 5% of the standard deviation of the input samples.

[0028] The steps for training the physical constraint model include: inputting the training set into the physical constraint model for forward propagation to obtain predicted output values; calculating the mean squared error loss between the predicted output values ​​and the true output values; calculating the physical consistency loss of the predicted output values ​​violating Maxwell's equations; calculating the coupling consistency loss of the predicted output values ​​violating the heat conduction equation and Joule's law; weighting and summing the mean squared error loss, physical consistency loss, and coupling consistency loss with weight coefficients of 0.5, 0.3, and 0.2 to obtain the total loss; using an adaptive moment estimation optimization algorithm to backpropagate the total loss to update the model parameters; setting the initial learning rate to 0.001, and reducing the learning rate to 0.5 times the original learning rate when the validation set loss does not decrease for 10 consecutive rounds; stopping training when the validation set loss does not decrease for 20 consecutive rounds to obtain the trained physical constraint model.

[0029] The calculation method for the constraint terms of the Maxwell's equations is as follows: automatically differentiate the hidden state features output by the second bidirectional long short-term memory unit layer to obtain the spatial gradient of the magnetic field prediction value; calculate the curl current residual between the curl of the magnetic field prediction value and the current density prediction value; and use the square of the curl current residual as the constraint term of the Maxwell's equations.

[0030] The calculation method for the constraint terms of the heat conduction equation is as follows: Calculate the heat source density distribution based on the predicted iron loss and predicted copper loss results output by the loss prediction branch; perform spatial gradient calculation on the temperature values ​​at the winding hot spot location, the core column center location, the core yoke location, the inner surface of the box wall, the upper oil temperature location, and the lower oil temperature location output by the temperature prediction branch; calculate the heat conduction residuals that satisfy the heat conduction equation for the temperature values ​​at the winding hot spot location, the core column center location, the core yoke location, the inner surface of the box wall, the upper oil temperature location, and the lower oil temperature location; and use the square of the heat conduction residuals as the constraint terms of the heat conduction equation.

[0031] The Joule's law constraint term is calculated as follows: the equivalent current density is calculated based on the X-direction component value, Y-direction component value, and Z-direction component value of the magnetic field output by the magnetic field prediction branch; the theoretical loss value corresponding to the equivalent current density is calculated based on Joule's law; the Joule's law residual between the theoretical loss value and the predicted copper loss result output by the loss prediction branch is calculated; and the square of the Joule's law residual is used as the Joule's law constraint term.

[0032] The training of the physical constraint model adopts a phased training strategy, which includes: a first training phase that trains only the magnetic field prediction branch, fixing the parameters of the loss prediction branch and the temperature prediction branch, with 50 training rounds in the first training phase; a second training phase that trains the magnetic field prediction branch and the loss prediction branch together, fixing the parameters of the temperature prediction branch, with 50 training rounds in the second training phase; and a third training phase that trains the magnetic field prediction branch, the loss prediction branch, and the temperature prediction branch together, with 100 training rounds in the third training phase.

[0033] The training of the physical constraint model adopts a course learning strategy, which includes: in the early stage of training, only steady-state operating condition data is used for training, which is the rated load condition without harmonic excitation and DC bias; in the middle stage of training, load variation condition data is added for training, which is the load variation between 50% and 100% of the rated load; and in the later stage of training, harmonic excitation condition data and DC bias condition data are added for training.

[0034] The iron loss difference is calculated as follows: the absolute value of the difference between the predicted iron loss result and the preset iron loss threshold is calculated, and the iron loss difference is obtained by dividing the absolute value of the difference between the predicted iron loss result and the preset iron loss threshold by the preset iron loss threshold.

[0035] The copper loss difference is calculated as follows: the absolute value of the difference between the predicted copper loss result and the preset copper loss threshold is calculated, and the absolute value of the difference between the predicted copper loss result and the preset copper loss threshold is divided by the preset copper loss threshold to obtain the copper loss difference.

[0036] The preset iron loss threshold is obtained by referring to a table based on the transformer's rated capacity and core material type. The value range of the preset iron loss threshold is 5kW to 50kW, with a preferred value of 20kW. The preset copper loss threshold is calculated based on the transformer's rated capacity and load rate. The value range of the preset copper loss threshold is 10kW to 100kW, with a preferred value of 40kW. The preset difference threshold ranges from 10% to 20%, with a preferred value of 15%.

[0037] The recommended percentage reduction in load power is determined based on the ratio of the loss difference to the preset difference threshold. When the ratio of the loss difference to the preset difference threshold is between 1.0 and 1.5, the recommended percentage reduction in load power is 5%; when the ratio of the loss difference to the preset difference threshold is between 1.5 and 2.0, the recommended percentage reduction in load power is 10%; and when the ratio of the loss difference to the preset difference threshold is greater than 2.0, the recommended percentage reduction in load power is 15%.

[0038] Furthermore, the SFG mesh generation algorithm is combined with the multilevel fast multipole method to accelerate the boundary element solution. The steps of the multilevel fast multipole method include: decomposing the final mesh cell set into an octree according to spatial location; approximating the interaction between distant mesh cells using a multipole expansion method; and directly calculating the interaction between close mesh cells. The multilevel fast multipole method reduces the computational complexity of boundary element matrix-vector multiplication from... Reduce to The This represents the number of grid cells in the final grid cell set.

[0039] Furthermore, the CNN feature extraction layer employs a frequency domain mapping mechanism for feature extraction. The steps of the frequency domain mapping mechanism include: performing a Fast Fourier Transform on the feature maps output by the first max pooling layer, the second max pooling layer, and the third max pooling layer to convert spatial domain features into frequency domain features; performing point-by-point multiplication operations on the frequency domain features in the frequency domain; learning the nonlinear transformation rules of the spectral coefficients; and using an Inverse Fast Fourier Transform to restore the frequency domain features to time domain features and inputting them into the BiLSTM time-dependent layer.

[0040] Furthermore, the temperature distribution data in step S02 includes temperature data at the winding hot spot location, temperature data at the center of the core column, temperature data at the core yoke location, temperature data at the inner surface of the box wall, temperature data at the upper oil temperature location, and temperature data at the lower oil temperature location.

[0041] Furthermore, in step S04, the voltage sensor is installed at the high-voltage side bushing outlet and the low-voltage side bushing outlet of the transformer, and the current sensor is installed at the high-voltage side bushing outlet, the low-voltage side bushing outlet, and the neutral point grounding terminal of the transformer.

[0042] Furthermore, in step S06, the cooling enhancement command is used to start a backup cooling fan or increase the oil pump speed. When the loss difference is between 1.0 and 1.5 times the preset difference threshold, the cooling enhancement command is used to start one backup cooling fan; when the loss difference is between 1.5 and 2.0 times the preset difference threshold, the cooling enhancement command is used to start two backup cooling fans; when the loss difference is greater than 2.0 times the preset difference threshold, the cooling enhancement command is used to start all backup cooling fans and increase the oil pump speed to 1.2 times the rated speed.

[0043] The frequency domain mapping mechanism is independent of grid resolution. For transformer structures with different geometric dimensions, the spatial scale parameters of the input data are used as additional input features to the physical constraint model to achieve cross-scale prediction. The dropout rate of the first dropout layer is 0.3, and the dropout rate of the second dropout layer is 0.3. This dropout rate is used to randomly drop neurons during training to prevent overfitting. The first moment estimation decay rate of the adaptive moment estimation optimization algorithm is 0.9, the second moment estimation decay rate is 0.999, and the numerical stability term is... .

[0044] Optionally, the present invention also provides a transformer loss detection system based on CNN-BiLSTM and SFG implemented by a computer. The computer is provided with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they execute the above-mentioned transformer loss detection method based on CNN-BiLSTM and SFG.

[0045] The specific implementation methods of the above steps are described in detail below.

[0046] The specific implementation of step S01 involves first obtaining the core geometry of the converter transformer from the transformer design documents and manufacturing specifications, including parameters such as the core column diameter, core yoke height, and core lamination thickness; obtaining the winding geometry, including parameters such as inner diameter, outer diameter, number of turns, and wire gauge; and obtaining the electromagnetic shielding structure dimensions, including shielding layer thickness and location parameters. These geometric parameters are the foundation for establishing an accurate finite element model. Next, the core magnetization curve is extracted from the technical data sheet provided by the core material manufacturer. This curve describes the nonlinear relationship between magnetic field strength and magnetic induction intensity. Simultaneously, the core loss curve is extracted, reflecting the iron loss per unit mass under different magnetic induction intensities and frequencies. These two curves are key data for simulating the magnetic properties and loss characteristics of the core. Then, the DC resistance of the winding is measured using a resistance tester, and the inductance of the winding is measured using an inductance tester. These two parameters determine the impedance characteristics of the winding. Finally, the rated voltage, rated current, operating frequency, and short-circuit impedance are obtained from the transformer nameplate. These parameters define the normal operating range of the transformer. Based on all the above parameters, a three-dimensional geometric model of the converter transformer was constructed in the finite element simulation software. Magnetization and loss curves and material properties were assigned to the core region, resistance and inductance parameters were assigned to the winding region, and conductivity and permeability properties of the corresponding material were assigned to the shielding structure, thus completing the establishment of the finite element simulation model of the converter transformer. An external circuit module was set up in this simulation model, containing components such as a three-phase voltage source, line impedance, and load impedance. The circuit and electromagnetic field were jointly solved using circuit node equations and field-circuit coupling methods. A three-phase AC input voltage source with an amplitude equal to the rated voltage was applied at the voltage source. To simulate the characteristics of the transformer under different operating conditions, the following conditions were set: no-load condition (secondary side open circuit with only core excitation), short-circuit condition (secondary side short-circuit winding current reaching several times the rated current), and rated load condition (normal operation condition with rated voltage applied to the primary side and rated load connected to the secondary side). Harmonic excitation conditions are superimposed on the input voltage source. Harmonic pollution generated during converter operation is simulated by superimposing multiple harmonic components of the 2nd to 13th orders onto the fundamental voltage. The amplitude of each harmonic component is set to 3% to 8% of the fundamental voltage amplitude. The harmonic phase angle is calculated and determined based on the converter's firing angle and modulation method. This harmonic excitation can reflect the impact of non-normal waveforms existing in the actual power grid on transformer losses. A DC current component is injected into the transformer neutral point to set a DC bias condition. The amplitude of the DC current component is taken as 5% to 15% of the rated current value, preferably 10%. This DC bias condition simulates the influence of geomagnetic induced current or the DC transmission system on the transformer's magnetic circuit, causing the core to operate in an asymmetric magnetization state, thereby increasing iron losses.

[0047] The specific implementation of step S02 involves using an adaptive sparse mesh generation algorithm, namely the SFG mesh generation algorithm, to intelligently mesh the finite element simulation model of the converter transformer. This algorithm first performs initial coarse mesh generation on the entire simulation model region to obtain an initial set of mesh elements. The initial mesh elements are tetrahedral or hexahedral elements. The element size is set to one-tenth of the core column diameter in the core region, one-fifth of the winding radial thickness in the winding region, and relatively larger sizes in the oil and air regions to reduce the total number of elements. The magnetic field gradient value is calculated for each initial mesh element in the initial set. The magnetic field gradient value is obtained by spatially differencing the magnetic flux density vector of the element node; this value reflects the intensity of the magnetic field change within the element. All initial mesh element magnetic field gradient values ​​are sorted in descending order, and the magnetic field gradient values ​​corresponding to the top 30% of the sorted values ​​are selected as the magnetic field gradient threshold, with a value range of 0.5. Up to 5.0 The preferred value is 2.0. The adaptive selection of this threshold ensures mesh refinement in regions with the most drastic magnetic field changes. Initial mesh cells with magnetic field gradient values ​​greater than the magnetic field gradient threshold are selected as high-gradient mesh cells. These high-gradient mesh cells undergo a first mesh refinement process, subdividing each high-gradient mesh cell into eight sub-cells to obtain the first-stage refined mesh cell set. This refinement process specifically increases the mesh density in regions with large magnetic field gradients, thereby improving the accuracy of field quantity calculations in these regions. The eddy current density value is calculated for each first-stage refined mesh cell in the first-stage refined mesh cell set. The eddy current density value is calculated using Ohm's law based on the induced electric field intensity and material conductivity within the cell, and this value characterizes the strength of the eddy current effect within the cell. The eddy current density values ​​of all first-stage refined mesh cells are sorted in descending order, and the eddy current density value corresponding to the top 25% of the sorted values ​​is selected as the eddy current density threshold. The threshold value range is [insert range here]. to The preferred value is This threshold ensures a sufficiently fine mesh in the eddy current concentration region. First-stage mesh elements with eddy current density values ​​greater than the eddy current density threshold are selected as high-eddy current mesh elements. These high-eddy current mesh elements undergo a second mesh refinement process, with each high-eddy current mesh element further subdivided into 8 sub-elements to obtain the final mesh element set. This second refinement achieves locally extremely fine meshes in the skin effect region near the surface of metal structural components such as box walls, clamps, and tension plates. In the final mesh element set, surface boundary impedance conditions are applied to the surfaces of the box wall structure, clamping structure, and pull plate structure. These conditions are obtained by acquiring the material conductivity values ​​of the box wall structure, clamping structure, and pull plate structure, as well as their corresponding magnetic permeability values. Based on the operating frequency, the skin depth values ​​of the box wall structure, clamping structure, and pull plate structure are calculated respectively. The skin depth value is calculated based on the attenuation theory of electromagnetic waves in conductive media. The skin depth value, conductivity value, and magnetic permeability value are combined to calculate the surface impedance values ​​of the box wall, clamping structure, and pull plate. In the simulation model, these impedance values ​​are applied to the corresponding surfaces as boundary conditions. This surface impedance boundary condition equivalently simulates the eddy current loss inside the metal structure without requiring meshing of the internal structure, greatly reducing the number of meshes and the amount of computation. A transient solver was used to solve the time-domain model of the converter transformer finite element simulation model after mesh generation. The transient solver, based on the time-stepping method, solved the partial differential equations of Maxwell's equations at each time step. The time step size was set to one-hundredth of the period corresponding to the operating frequency to ensure accurate sampling of the time-domain waveforms. During the solution process, three-phase voltage waveforms, three-phase current waveforms, magnetic field distribution data, iron loss data, copper loss data, and temperature distribution data were acquired under different operating conditions. The temperature distribution data included temperature data at winding hotspot locations, core column center locations, core yoke locations, inner surface locations of the tank wall, upper oil temperature locations, and lower oil temperature locations. Time synchronization processing was performed on the three-phase voltage and current waveforms. This processing extracted the timestamps of both types of data, and linear interpolation was used to fill in misaligned data points, ensuring that the sampling times of the three-phase voltage and current waveforms corresponded perfectly, eliminating time offsets caused by differences in the solution order of different physical quantities.The synchronized data is standardized using a channel-by-channel standardization approach. The mean and standard deviation of the A-phase, B-phase, and C-phase voltage waveforms are calculated separately. Similarly, the mean and standard deviation of the A-phase, B-phase, and C-phase current waveforms are calculated separately. The standardized voltage waveforms are obtained by subtracting the mean voltage of the corresponding phase from the voltage waveform data of each phase and then dividing by the standard deviation of the corresponding phase. The standardized current waveforms are obtained by subtracting the mean current of the corresponding phase from the current waveform data of each phase and then dividing by the standard deviation of the corresponding phase. This standardization process eliminates the differences in the dimensions and numerical ranges of the data from different channels, making subsequent neural network training more stable and convergent. The processed data is divided into fixed-length time-series sample sequences using a sliding window method. The window length is set to 1000 sampling points, and the sliding step is set to 500 sampling points. The data is divided from standardized A-phase voltage waveform data, standardized B-phase voltage waveform data, standardized C-phase voltage waveform data, standardized A-phase current waveform data, standardized B-phase current waveform data, and standardized C-phase current waveform data according to the window length and sliding step. Each fixed-length time-series sample sequence contains six-dimensional time-series data at 1000 consecutive time points. This sliding window segmentation method increases the number of training samples by overlapping sampling while ensuring the integrity of the sample time sequence.

[0048] The specific implementation of step S03 involves constructing a physical constraint model. This model structure includes an input layer, a convolutional neural network feature extraction layer (CNN feature extraction layer), a bidirectional long short-term memory network temporal dependency layer (BiLSTM temporal dependency layer), a physical constraint fusion layer, and a multi-task output layer. The input layer receives a fixed-length time-series sample sequence, which contains six-dimensional time-series data consisting of standardized A-phase voltage waveform data, standardized B-phase voltage waveform data, standardized C-phase voltage waveform data, standardized A-phase current waveform data, standardized B-phase current waveform data, and standardized C-phase current waveform data. The data shape is 1000 time steps multiplied by 6 feature channels. The CNN feature extraction layer includes a first convolutional sub-layer, a second convolutional sub-layer, and a third convolutional sub-layer. The first convolutional sub-layer contains 32 first convolutional kernels with a kernel size of 3, the second convolutional sub-layer contains 32 second convolutional kernels with a kernel size of 5, and the third convolutional sub-layer contains 32 third convolutional kernels with a kernel size of 7. The three convolutional kernels of different sizes extract short-term local features, medium-term periodic features, and long-term trend features from the time-series data, respectively. The first, second, and third convolutional sub-layers are connected to the first batch normalization layer, the second batch normalization layer, and the third batch normalization layer, respectively, to accelerate training convergence and improve the model's generalization ability. The batch normalization layer is connected to the first max pooling layer, the second max pooling layer, and the third max pooling layer, respectively, to reduce the feature map size and extract the main features. The CNN feature extraction layer employs a frequency domain mapping mechanism for feature extraction. This mechanism performs Fast Fourier Transform on the feature maps output by the first, second, and third max pooling layers to convert spatial domain features into frequency domain features. In the frequency domain, point-by-point multiplication is performed on the frequency domain features to learn the nonlinear transformation rules of the spectral coefficients in order to extract the harmonic components and frequency characteristics in the voltage and current waveforms. Inverse Fast Fourier Transform is then used to restore the frequency domain features to time domain features and input them into the BiLSTM time-dependent layer. This frequency domain mapping mechanism can effectively capture frequency domain features under harmonic excitation conditions. The BiLSTM temporal dependency layer includes a first bidirectional long short-term memory (BSSM) layer and a second bidirectional BSSM layer. The first BSSM layer contains 128 first hidden units, and the second BSSM layer contains 128 second hidden units. The BSSM layer processes temporal data simultaneously through recurrent neural networks in both forward and backward directions, enabling it to capture long-term dependencies and contextual information in the temporal data. The first BSSM layer is followed by a first dropout layer, and the second BSSM layer is followed by a second dropout layer. The dropout rate of each dropout layer is set to 0.3, and neurons are randomly dropped during training to prevent overfitting.The physical constraint fusion layer embeds constraints from Maxwell's equations, the heat conduction equation, and Joule's law. This layer integrates physical laws as soft constraints into the neural network training process. The Maxwell's equations constraint is calculated by automatically differentiating the hidden state features output from the second bidirectional long short-term memory unit layer to obtain the spatial gradient of the predicted magnetic field value. The curl of the predicted magnetic field value and the curl-current residual between the predicted current density are calculated, and the square of the curl-current residual is used as the Maxwell's equations constraint. This constraint ensures that the predicted magnetic field distribution satisfies Ampere's circuital law. The heat conduction equation constraint is calculated by calculating the heat source density distribution based on the predicted iron loss and predicted copper loss results output from the loss prediction branch. The temperature prediction branch outputs the winding hot spot temperature value, the core column center temperature value, and the iron... Spatial gradient calculations are performed on the temperature values ​​at the yoke position, the inner surface of the tank wall, the upper oil temperature position, and the lower oil temperature position. The thermal conduction residuals satisfying the thermal conduction equation are calculated, and the square of the thermal conduction residuals is used as a constraint term in the thermal conduction equation. This constraint ensures that the temperature distribution prediction conforms to Fourier's law of thermal conduction. The Joule's law constraint term is calculated by calculating the equivalent current density based on the X-direction, Y-direction, and Z-direction component values ​​of the magnetic field output from the magnetic field prediction branch, calculating the theoretical loss value corresponding to the equivalent current density according to Joule's law, and calculating the Joule's law residual between the theoretical loss value and the predicted copper loss result output from the loss prediction branch. The square of the Joule's law residual is used as a Joule's law constraint term, which ensures that the loss prediction and current distribution satisfy the Joule thermal principle. The multi-task output layer includes a magnetic field prediction branch, a loss prediction branch, and a temperature prediction branch. The magnetic field prediction branch outputs the X-direction component value, Y-direction component value, and Z-direction component value of the magnetic field to fully describe the three-dimensional spatial magnetic field vector. The loss prediction branch outputs the predicted iron loss result and the predicted copper loss result. The temperature prediction branch outputs the temperature value of the winding hot spot location, the temperature value of the core column center location, the temperature value of the core yoke location, the temperature value of the inner surface of the box wall location, the temperature value of the upper oil temperature location, and the temperature value of the lower oil temperature location.A training dataset is established using a fixed-length time-series sample sequence as input data and magnetic field distribution data, iron loss data, copper loss data, and temperature distribution data as output target data. The steps for establishing the training dataset include extracting three-phase voltage waveform data and three-phase current waveform data from the solution results of the converter transformer finite element simulation model as input samples, and extracting magnetic field distribution data, iron loss data, copper loss data, and temperature distribution data as output samples. The input and output samples are time-stamp aligned to ensure correct causal relationships. The input and output samples are standardized to a mean of 0 and a standard deviation of 1. The standardized input samples are divided into sequences of 1000 sampling points each, and the corresponding output samples are divided synchronously to maintain a one-to-one correspondence between input and output. The segmented input and output samples are divided into training and validation sets in an 8:2 ratio. Noise following a normal distribution is added to the input samples in the training set for data augmentation. The noise amplitude is 5% of the standard deviation of the input samples. This data augmentation improves the robustness of the model to measurement noise. A physically constrained model is trained on the training dataset to obtain a trained physically constrained model. The training steps include inputting the training set into the physically constrained model for forward propagation to obtain the predicted output value, calculating the mean squared error loss between the predicted output value and the true output value, calculating the physical consistency loss of the predicted output value violating Maxwell's equations, and calculating the coupling consistency loss of the predicted output value violating the heat conduction equation and Joule's law. The mean squared error loss, physical consistency loss, and coupling consistency loss are weighted and summed with weight coefficients of 0.5, 0.3, and 0.2 to obtain the total loss. This weight allocation emphasizes both data fitting accuracy and physical law constraints. An adaptive moment estimation optimization algorithm, namely the Adam optimization algorithm, is used to backpropagate the total loss to update the model parameters. This optimization algorithm combines first-order moment estimation and second-order moment estimation to adaptively adjust the learning rate of each parameter. The decay rate of the first moment estimation is set to 0.9, the decay rate of the second moment estimation is set to 0.999, and the numerical stability term is set to... The initial learning rate was set to 0.001. When the validation set loss did not decrease for 10 consecutive rounds, the learning rate was reduced to 0.5 times the original learning rate to fine-tune the parameters. When the validation set loss did not decrease for 20 consecutive rounds, the model was considered to have learned sufficiently, and training was stopped, resulting in a physically constrained model. Training employed a phased training strategy. This strategy included: a first training phase training only the magnetic field prediction branch, fixing the parameters of the loss prediction and temperature prediction branches, with 50 training rounds to prioritize learning the basic laws of electromagnetic fields; a second training phase jointly training the magnetic field and loss prediction branches, fixing the parameters of the temperature prediction branch, with 50 training rounds to establish the mapping relationship between the magnetic field and loss; and a third training phase jointly training the magnetic field, loss, and temperature prediction branches, with 100 training rounds to learn the complete electromagnetic-thermal coupling relationship. This phased strategy progressively improved training stability and final accuracy by training from easy to difficult. The training employs a course-based learning strategy. This strategy involves initially training with only steady-state operating condition data, which is data under rated load conditions without harmonic excitation or DC bias. In the middle stage of training, load variation operating condition data is added, which is data where the load varies between 50% and 100% of the rated load. In the later stage of training, harmonic excitation operating condition data and DC bias data are added. This course-based learning strategy mimics the human learning process from simple to complex, allowing the model to gradually adapt to complex operating conditions.

[0049] The specific implementation of step S04 involves installing voltage and current sensors at the high-voltage side bushing outlet, low-voltage side bushing outlet, and neutral point grounding terminal of the transformer. The voltage sensor uses a capacitive voltage divider type voltage transformer to obtain the voltage signals of the high-voltage and low-voltage sides. The current sensor uses a through-core Rogowski coil current transformer to measure the three-phase current and neutral point current in a non-invasive manner. The sensor signals are converted from analog to digital by a data acquisition card, and the sampling frequency is set to 100 times the operating frequency to satisfy the Nyquist sampling theorem and capture higher harmonic components. The measured waveforms of three-phase voltage and three-phase current during transformer operation are acquired in real time. The acquisition process is continuous, forming a continuous time-series data stream. The acquired measured waveforms of three-phase voltage and three-phase current are subjected to the same time synchronization and standardization processing as in step S02. They are divided into a fixed-length time-series sample sequence of 1000 sampling points using the same sliding window method. The processed measured waveforms of three-phase voltage and three-phase current are input into the training of the physical constraint model. The input process transmits real-time data to the deployed neural network inference engine through the application programming interface.

[0050] The specific implementation of step S05 involves training the physical constraint model, receiving real-time input data, and performing forward propagation calculations. The data is passed through the input layer to the CNN feature extraction layer to extract local features. A BiLSTM temporal dependency layer captures temporal dependencies. A physical constraint fusion layer ensures the prediction results conform to physical laws. Finally, a multi-task output layer outputs the predicted magnetic field distribution, predicted iron loss, predicted copper loss, and predicted temperature distribution results. The absolute value of the difference between the predicted iron loss result and a preset iron loss threshold is calculated. This absolute value is divided by the preset iron loss threshold to obtain the iron loss difference, which is a dimensionless relative deviation index. The preset iron loss threshold is obtained by looking up a table based on the transformer's rated capacity and core material type, with a value range of 5kW to 50kW, preferably 20kW. Similarly, the absolute value of the difference between the predicted copper loss result and a preset copper loss threshold is calculated. This absolute value is divided by the preset copper loss threshold to obtain the copper loss difference. The preset copper loss threshold is calculated based on the transformer's rated capacity and load rate, with a value range of 10kW to 100kW, preferably 40kW. The larger value between iron loss difference and copper loss difference is selected as the loss difference degree. This selection method ensures that a response can be made to any type of loss exceeding the standard.

[0051] The specific implementation of step S06 involves comparing the loss difference degree with a preset difference threshold. The preset difference threshold ranges from 10% to 20%, with a preferred value of 15%. When the loss difference degree exceeds the preset difference threshold, it is determined that the transformer's current operating loss exceeds the normal range, posing a risk of overheating. In this case, a load adjustment command is sent to the transformer monitoring system, or a cooling enhancement command is sent to the transformer cooling control system. The load adjustment command includes a suggested percentage reduction in load power. This percentage is determined based on the ratio of the loss difference degree to the preset difference threshold. When the ratio is between 1.0 and 1.5, the suggested percentage reduction in load power is 5%; when the ratio is between 1.5 and 2.0, the suggested percentage reduction in load power is 10%; and when the ratio is greater than 2.0, the suggested percentage reduction in load power is 15%. This tiered adjustment strategy flexibly adjusts the load according to the degree of loss exceeding the standard to avoid excessive reduction in power supply capacity. The cooling enhancement command is used to activate the backup cooling fan or increase the oil pump speed. When the loss difference is between 1.0 and 1.5 times the preset difference threshold, the cooling enhancement command is used to activate one backup cooling fan to increase heat dissipation capacity. When the loss difference is between 1.5 and 2.0 times the preset difference threshold, the cooling enhancement command is used to activate two backup cooling fans to further enhance the cooling effect. When the loss difference is greater than 2.0 times the preset difference threshold, the cooling enhancement command is used to activate all backup cooling fans and increase the oil pump speed to 1.2 times the rated speed to maximize cooling capacity and prevent equipment damage. This tiered cooling strategy avoids energy waste and equipment wear caused by long-term high-load operation of the cooling system while ensuring equipment safety.

[0052] Specifically, the principle of this invention is as follows: The reason this invention can solve the technical problem lies in its use of a deep learning architecture that combines data-driven and physical constraints. First, a high-precision finite element model of the converter transformer is performed using the SFG adaptive sparse mesh partitioning algorithm. Mesh refinement is applied in key areas with high magnetic field gradients and high eddy current densities. Surface boundary impedance conditions are applied to the surfaces of structural components such as tank walls, clamps, and tie plates to simulate the electromagnetic field distribution and loss characteristics under complex operating conditions such as harmonic excitation and DC bias, generating a high-fidelity training dataset containing voltage and current waveforms and corresponding loss distributions. The subsequently constructed physical constraint model combines CNN multi-scale convolution kernels with BiLSTM bidirectional memory mechanisms. This allows for the extraction of spatial features representing the effects of harmonics and bias from transient waveforms, and also captures the dynamic dependence of losses over time. The physical constraint fusion layer uses the residuals between the predicted values ​​and Maxwell's equations, the heat conduction equation, and Joule's law as additional loss terms. This ensures that the model not only fits the data during optimization but also satisfies physical laws, guaranteeing the inherent consistency between the predicted magnetic field, losses, and temperature. Thus, the goal of accurately predicting the internal loss state of the transformer based solely on measured voltage and current waveforms is achieved.

[0053] The following provides a specific embodiment 1 of the present invention. The specific implementation methods of steps S01, S04 and S05 in this embodiment 1 are the same as those described above, and will not be repeated in detail here. The specific implementation methods of other steps are described in detail below.

[0054] In the specific implementation of step S02, the formula for calculating the magnetic field gradient value is as follows: ; In the formula, This is the magnetic field gradient value, which is dimensionless. For the magnetic field in Components of direction, in units of It is obtained by solving the finite element simulation model; For the magnetic field in Components of direction, in units of It is obtained by solving the finite element simulation model; For the magnetic field in Components of direction, in units of It is obtained by solving the finite element simulation model; magnetic field Component pairs Partial derivatives of coordinates, in units of ; magnetic field Component pairs Partial derivatives of coordinates, in units of ; magnetic field Component pairs Partial derivatives of coordinates, in units of ; The reference magnetic field gradient value is expressed in units of . The experience value is 2.0.

[0055] The formula for calculating eddy current density is as follows: ; In the formula, This is the eddy current density value, which is dimensionless. Electrical conductivity of structural component materials, in units of The information is obtained by looking up the table based on the material type of the structural component. Angular frequency, unit: It is calculated from the operating frequency value, and the calculation formula is: ,in This is the operating frequency value, in units of The operating frequency value obtained in step S01 is used to determine the frequency. The peak value of the local magnetic flux density is expressed in units of 1000 m / s. It is determined by the maximum value of the magnetic field distribution data within the grid cell; The feature size of the mesh cell, in units of The calculation formula is: ,in The volume of the grid cell is expressed in units of 1. The geometric dimensions of the elements after finite element mesh generation are calculated. For reference eddy current density values, the units are... Experience value .

[0056] The surface impedance value is calculated using the skin depth theory. The formula for calculating the surface impedance value of the box wall is as follows: ; In the formula, This is the surface impedance value of the enclosure wall, in units of... ; The imaginary unit satisfies ; The electrical conductivity value of the material of the box wall structural components, in units of The information is obtained by referring to a table based on the type of container wall material. The skin depth value of the material of the box wall structural component, in units of Its calculation formula is ,in The magnetic permeability value of the material of the box wall structural component, in units of The value is obtained by referring to a table based on the type of box wall material. The formula for calculating the surface resistance value of the clamping component is expressed as follows: ; In the formula, The surface resistance value of the clamping component, in units of ; The electrical conductivity value of the clamping structural component material, in units of... The information is obtained by referring to a table based on the type of clamping material. The skin depth value of the clamping structural component material, in units of Its calculation formula is ,in The magnetic permeability value of the clamping structural component material, in units of The value is obtained by referring to a table based on the material type of the clamping component. The formula for calculating the surface resistance value of the pull plate is expressed as follows: ; In the formula, This is the surface impedance value of the plate, in units of... ; The electrical conductivity value of the material for the tension plate structural component is given in units of 1000 kJ / m². The information is obtained by referring to a table based on the type of sheet material. The skin depth value of the material for the tension plate structural component, in units of... Its calculation formula is ,in The magnetic permeability value of the material for the tension plate structural component is given in units of 1000 ppm. The value can be obtained by referring to a table based on the material type of the sheet.

[0057] In the standardization process, the standardization formula for phase A voltage waveform data is expressed as follows: ; In the formula, The voltage waveform data for phase A is standardized and dimensionless. This is the voltage waveform data for phase A, in units of... ; This is the average voltage of phase A, in units of... ; The standard deviation of phase A voltage is given in units of 1000 kJ / m². The standardized formula for B-phase voltage waveform data is expressed as follows: ; In the formula, The B-phase voltage waveform data is standardized and dimensionless. This is the voltage waveform data for phase B, in units of... ; This is the average voltage of phase B, in units of... ; This represents the standard deviation of phase B voltage, in units of... The standardized formula for C-phase voltage waveform data is as follows: ; In the formula, The C-phase voltage waveform data is standardized and dimensionless. This is the C-phase voltage waveform data, in units of ; This is the average voltage of phase C, in units of... ; The standard deviation of phase C voltage is given in units of 1000 kJ / m². The standardized formula for the A-phase current waveform data is as follows: ; In the formula, The A-phase current waveform data is standardized and dimensionless. This is the waveform data of phase A current, in units of... ; This is the average current of phase A, in units of... ; The standard deviation of phase A current is given in units of 1000 m / s. The standardized formula for B-phase current waveform data is expressed as follows: ; In the formula, The B-phase current waveform data is standardized and dimensionless. This is the waveform data for phase B current, in units of... ; This is the average value of phase B current, in units of... ; The standard deviation of phase B current is given in units of 1000 m / s. The standardized formula for C-phase current waveform data is as follows: ; In the formula, The C-phase current waveform data is standardized and dimensionless. This is the C-phase current waveform data, in units of... ; This is the average C-phase current, in units of ; The standard deviation of the C-phase current is given in units of 1000 m / s. .

[0058] In the specific implementation of step S03, the formula for calculating the iron loss difference is as follows: ; In the formula, The degree of iron loss variation is dimensionless. To predict iron loss results, the unit is... ; The preset iron loss threshold is in units of .

[0059] The formula for calculating the copper loss difference is as follows: ; In the formula, The difference in copper loss is dimensionless. To predict copper loss results, the unit is... ; This is the preset copper loss threshold, in units of .

[0060] The formula for selecting the loss difference is expressed as follows: ; In the formula, The degree of loss difference is dimensionless; This is the function for finding the maximum value.

[0061] The formula for calculating the total loss during the training of the physical constraint model is as follows: ; In the formula, Total loss, dimensionless; The mean squared error loss is dimensionless. The loss is a dimensionless loss of physical consistency. The coefficients 0.5, 0.3, and 0.2 are the coupling consistency loss, which is dimensionless; the coefficients 0.5, 0.3, and 0.2 are the weighting coefficients, which are also dimensionless.

[0062] The formula for calculating the constraint terms of Maxwell's equations is as follows: ; In the formula, These are constraint terms in Maxwell's equations, and are dimensionless. For curl operator; This is a vector of predicted magnetic field values, in units of... Its weight is , , , respectively representing the predicted magnetic field values ​​in Directional components, Directional components, Directional components, all in units ; The value of free permeability is . The unit is ; This is a vector of predicted current density values, in units of... ; Let L be the L2 norm of the vector; For reference magnetic flux density, the average value of the magnetic flux density inside the core under rated operating conditions is taken, with units of . The experience value is 1.5; For reference length, the core column diameter is used, in units of... .

[0063] The formula for calculating the constraint terms of the heat conduction equation is as follows: ; In the formula, These are constraint terms in the heat conduction equation and are dimensionless. For divergence operators; Thermal conductivity, in units of The default value is 0.12 for transformer oil, 40 for core material, and 380 for winding copper. This is a temperature prediction value, in units of... ; This represents the heat source density distribution, in units of... Its calculation formula is ,in This refers to the core volume, in units of... The core geometry obtained in step S01 is used for calculation. The winding volume is expressed in units of 1000 liters. The winding geometry obtained in step S01 is used for calculation. For reference heat source density, the average loss density under rated load conditions is used, with units of [unit missing]. Experience value .

[0064] The formula for calculating the constraint term of Joule's law is as follows: ; In the formula, This is a dimensionless constraint term for Joule's law. This is the theoretical loss value, in units of Its calculation formula is ,in Equivalent current density, in units of Its calculation formula is , The conductivity of the copper winding is expressed in units of... Experience value , The integral volume domain of the winding is given in units of . ,and same; For reference loss values, the rated copper loss value under rated load conditions is used, in units of... The experience value is 40.

[0065] In a specific implementation of step S06, it is recommended that the relationship between the percentage reduction in load power and the ratio of loss difference be expressed as a piecewise function, as shown in the following formula: ; In the formula, The recommended reduction in load power is dimensionless. The preset difference threshold is dimensionless and ranges from 10% to 20%, with a preferred value of 15%.

[0066] To better understand and implement this invention, the following is a specific application scenario of Example 2: A technical team faced the problem of abnormal loss fluctuations in an operating converter transformer. The transformer has a rated capacity of 150MVA, a rated voltage of 220kV on the high-voltage side, a rated voltage of 35kV on the low-voltage side, and a rated frequency of 50Hz. Traditional monitoring methods can only detect anomalies through temperature sensors, but cannot accurately locate the source and distribution characteristics of the loss, resulting in a lack of basis for operation and maintenance decisions. The technical team decided to adopt a transformer loss detection method based on CNN-BiLSTM and SFG, combining multiphysics field coupled simulation with deep learning to achieve accurate prediction and real-time monitoring of the transformer's internal losses. The detection process of the entire embodiment is as follows: Figure 1 As shown.

[0067] The implementation process begins by obtaining key parameters from the transformer design documents. After obtaining complete parameters, a finite element simulation is established. The 3D model includes the complete geometry of the core, high and low voltage windings, tank, and internal structural components. During modeling, a stepped lamination method is used for the core joints, and a multi-conductor domain modeling method is used for the windings. Each layer of windings has an independent conductor domain with series constraints. The tank is filled with transformer oil, with a dielectric constant of 2.2 and a breakdown field strength of 15kV / mm. An external circuit is set in the model: an ideal voltage source is connected to the high-voltage side, a variable load impedance is connected to the low-voltage side, and the neutral point is grounded via a 10Ω grounding resistor. Simulation conditions include: no-load condition, where the high-voltage side is subjected to a rated voltage of 220kV and the low-voltage side is open-circuited; short-circuit condition, where the low-voltage side is short-circuited and the high-voltage side is energized with a voltage increase until the short-circuit current reaches 10 times the rated current; and rated load condition, where the load impedance is set to bring the low-voltage side current to the rated value of 2476A.

[0068] To simulate the actual operating characteristics of the converter transformer, harmonic components are superimposed on the input voltage source. The amplitude of the 2nd harmonic is set to 5% of the fundamental amplitude, with a phase angle of 45°. The amplitude of the 3rd harmonic is set to 6% of the fundamental amplitude, with a phase angle of 15°. The amplitude of the 5th harmonic is set to 5.5% of the fundamental amplitude, with a phase angle of 30°. The amplitude of the 7th harmonic is set to 4% of the fundamental amplitude, with a phase angle of 20°. The amplitude of the 11th harmonic is set to 3.5% of the fundamental amplitude, with a phase angle of 10°. The amplitude of the 13th harmonic is set to 3% of the fundamental amplitude, with a phase angle of 5°. The phase angle of each harmonic component is calculated and determined based on the 15° firing angle of the 12-pulse converter. The DC bias condition is achieved by injecting a DC current component at the neutral point, with the DC current amplitude set to 10% of the rated current, i.e., 247.6A, and the injection direction is positive. In addition, 10 leakage flux points are selected on the transformer's simulation model, such as... Figure 2 As shown.

[0069] Mesh generation was performed using the SFG adaptive sparse mesh generation algorithm. The initial mesh generation yielded 358,600 tetrahedral elements with an average element size of 35 mm. After calculating the magnetic field gradient value of each element, all elements were sorted in descending order of gradient value. The top 30% corresponded to a magnetic field gradient threshold of 2.0 T / m, identifying 107,580 high-gradient mesh elements. These elements were mainly distributed at the winding ends, core corner regions, and air gap edges. A first refinement process was performed on these high-gradient mesh elements, refining the element size to 15 mm, increasing the total number of elements to 485,200. The eddy current density value of each element after the first refinement was calculated, and all elements were sorted in descending order of eddy current density. The top 25% corresponded to an eddy current density threshold of [missing value]. A / A total of 121,300 high eddy current mesh elements were identified, mainly concentrated on the inner wall of the oil tank, the surface of the core clamping parts, and the edge area of ​​the tie plate. The high eddy current mesh elements were further refined by refining the element size to 8 mm, resulting in a final total of 627,800 mesh elements. The final mesh achieved a high-density distribution in the key electromagnetic region and maintained a reasonable mesh density ratio in the low gradient region.

[0070] Surface boundary impedance conditions are applied to the final mesh. The conductivity of the tank wall material is... S / m, relative permeability is 180, absolute permeability is calculated as follows: H / m, with a skin depth of 3.85mm at an operating frequency of 50Hz, the surface impedance of the tank wall is calculated using the skin depth formula. Ω. The electrical conductivity of the core clamping material is... S / m, relative permeability 1200, absolute permeability H / m, skin depth value is 1.83mm, and surface resistance of the clamping part is [value missing]. Ω. The plate material is non-magnetic steel with an electrical conductivity of [value missing]. S / m, relative permeability 1.02, absolute permeability H / m, skin depth value is 10.6mm, and surface resistance of the pull plate is [value missing]. Ω. The calculated surface impedance values ​​are applied to the surfaces of the corresponding structural components to achieve an equivalent simulation of eddy current losses in thin-walled structures.

[0071] A transient solver was used to solve the simulation model with a time step of 0.2 ms. The simulation duration covered five power frequency cycles (100 ms), and each operating condition was repeated three times to obtain stable results. After solving, the three-phase voltage and current waveforms were extracted. The peak value of the phase A voltage waveform was 179.6 kV, and the effective value was 127.0 kV, containing significant second and third harmonic components. The phase B voltage lagged phase A by 120°, with a peak value of 179.8 kV. The phase C voltage lagged phase B by 120°, with a peak value of 179.5 kV. The three-phase current waveform data showed that the peak value of the phase A current was 3510 A, and the effective value was 2481 A, with a fifth harmonic content of 5.2%. The peak value of the phase B current was 3505 A, and the peak value of the phase C current was 3515 A. The magnetic field distribution data was extracted, and the magnetic field strength in the winding end region reached 12500 A / m, corresponding to a magnetic induction intensity of 1.85 T, which is close to the core saturation region. The magnetic flux density at the center of the core yoke is 1.72 T, and at the center of the core column it is 1.68 T. The leakage magnetic field is mainly distributed between the winding and the tank wall, with a maximum leakage magnetic flux density of 0.158 T.

[0072] The loss data extraction results show that the total core loss is 22.8kW, of which hysteresis loss accounts for 65% (14.82kW) and eddy current loss accounts for 35% (7.98kW). Copper loss data shows that the high-voltage winding copper loss is 38.5kW, the low-voltage winding copper loss is 12.7kW, and the total copper loss is 51.2kW. Among the eddy current losses of structural components, the tank wall eddy current loss is 3.2kW, the core clamp eddy current loss is 1.8kW, the tie plate eddy current loss is 0.6kW, and the stray loss totals 5.6kW. The total transformer loss reaches 79.6kW, a deviation of 1.4% compared to the design expectation of 78.5kW. Temperature distribution data was obtained through thermal field coupling simulation. The temperature at the winding hot spot is 98℃, located in the 15th segment of the inner layer of the high-voltage winding. The core column center temperature is 87℃, and the core yoke temperature is 82℃. The surface temperature inside the oil tank is 76℃, the upper oil temperature is 85℃, the lower oil temperature is 68℃, and the oil temperature gradient is 17℃.

[0073] The data preprocessing stage begins with time synchronization. Voltage data is sampled at 10kHz, current data at 10kHz, and magnetic field and temperature data at 5kHz, resulting in timestamp misalignment. The timestamp sequences of each data stream are extracted, and linear interpolation is used to increase the sampling frequency of the magnetic field and temperature data to 10kHz, ensuring complete alignment of all data on the time axis. Standardization is performed channel-by-channel. The calculated mean of phase A voltage is 0.12kV, with a standard deviation of 126.8kV, resulting in a standardized range of -1.42 to 1.42. The mean of phase B voltage is -0.08kV, with a standard deviation of 127.1kV. The mean of phase C voltage is 0.05kV, with a standard deviation of 126.5kV. The mean of phase A current is 0.35A, with a standard deviation of 2478A, resulting in a standardized range of -1.42 to 1.42. The mean of phase B current is -0.28A, with a standard deviation of 2476A. The mean current of phase C is 0.18A, and the standard deviation is 2482A. Standardized waveform data are obtained by subtracting the mean from the data for each phase and then dividing by the standard deviation.

[0074] A sliding window approach was used to segment the standardized data into fixed-length time-series sample sequences. The window length was set to 1000 sampling points, corresponding to a time window of 0.1 s, covering 5 power frequency cycles, which can fully capture power frequency and harmonic characteristics. The sliding step size was set to 500 sampling points, with an overlap rate of 50% between adjacent windows, ensuring both the number of samples and the continuity of the time series. Segmenting the 100 ms simulation data according to these parameters yielded 196 time-series sample sequences. Each sample contained 6-dimensional input data: standardized A-phase voltage, standardized B-phase voltage, standardized C-phase voltage, standardized A-phase current, standardized B-phase current, and standardized C-phase current. The corresponding output target data included 10 dimensions: magnetic field X-direction component, magnetic field Y-direction component, magnetic field Z-direction component, predicted iron loss, predicted copper loss, winding hot spot temperature, core column center temperature, core yoke temperature, inner surface temperature of the casing wall, and upper oil temperature.

[0075] During the training dataset setup, 196 time-series samples were divided in an 8:2 ratio, with 157 samples used as the training set and 39 samples as the validation set. Gaussian noise with a mean of 0 and a standard deviation of 5% of the sample standard deviation was added to the training set input samples for data augmentation, expanding the training set to 471 samples. A phased training strategy was employed. In the first phase, only the magnetic field prediction branch was trained, with fixed parameters for loss and temperature branches. The loss function only included the mean squared error loss for magnetic field prediction and the Maxwell's equations constraint terms, with weighting coefficients of 0.7 and 0.3 respectively, and a learning rate of 0.001. After 50 training rounds, the mean absolute error of the magnetic field prediction validation set decreased to 0.048T. In the second phase, the magnetic field and loss branches were trained jointly, with fixed temperature branch parameters. The loss function included the mean squared error loss for both magnetic field and loss branches, Maxwell's constraint terms, and Joule's law constraint terms, with weighting coefficients of 0.4, 0.3, 0.2, and 0.1 respectively. After 50 training rounds, the mean relative error of the loss prediction validation set decreased to 2.8%. In the third training phase, all branches are trained together. The loss function includes the mean squared error loss of the three prediction tasks and three types of physical constraint terms, with weight coefficients of 0.3, 0.2, 0.3, 0.1, 0.05, and 0.05, respectively. After 100 training rounds, the mean absolute error of the temperature prediction validation set is reduced to 0.95℃.

[0076] The trained physical constraint model is deployed to the transformer field monitoring system. Voltage sensors (capacitive voltage divider type, 0-250kV, 0.2 accuracy, 10kHz sampling frequency) are installed at the A-phase, B-phase, and C-phase bushing outlets on the high-voltage side of the transformer. Current sensors (Rogowski coil type, 0-5000A, 0.5 accuracy, 10kHz sampling frequency) are installed at the high-voltage side three-phase bushing outlets, the low-voltage side three-phase bushing outlets, and the neutral grounding terminal. Sensor output signals are transmitted to the data acquisition unit via shielded cables. The acquisition unit performs 16-bit ADC conversion on the analog signals, and the converted digital signals are transmitted to the edge computing server via optical fiber.

[0077] During on-site operation, the system collects three-phase voltage and three-phase current waveform data in real time. At a certain moment, the measured peak value of phase A voltage was 181.2kV, with an effective value of 128.1kV; the peak value of phase B voltage was 180.8kV; and the peak value of phase C voltage was 181.5kV. The measured peak value of phase A current was 3680A, with an effective value of 2602A, significantly higher than the rated value. The peak value of phase B current was 3675A, and the peak value of phase C current was 3685A. After time synchronization and standardization, the waveform data was divided into time-series samples of 1000 sampling points and input into the trained physical constraint model for inference. The model output predicted magnetic field distribution results, showing that the magnetic field strength at the winding end reached 13800A / m, corresponding to a magnetic induction intensity of 1.92T, entering the deep saturation region of the iron core. The predicted iron loss result was 28.6kW, an increase of 25.4% compared to the rated operating condition. The predicted copper loss result was 68.3kW, an increase of 33.4% compared to the rated operating condition. The predicted temperature distribution results show that the winding hot spot temperature is 112℃, which is close to the 105℃ limit of insulation temperature resistance class A. The core column center temperature is 96℃, the core yoke temperature is 91℃, and the upper oil temperature is 94℃.

[0078] Based on on-site measurement data and consulting the transformer nameplate and design documents, the preset iron loss threshold for the 150MVA rated capacity of this transformer is 20kW, and the preset copper loss threshold is 45kW. The iron loss difference was calculated; the absolute value of the difference between the predicted iron loss of 28.6kW and the preset iron loss threshold of 20kW is 8.6kW. Dividing this by the preset iron loss threshold of 20kW yields an iron loss difference of 0.43, or 43%. The copper loss difference was calculated; the absolute value of the difference between the predicted copper loss of 68.3kW and the preset copper loss threshold of 45kW is 23.3kW. Dividing this by the preset copper loss threshold of 45kW yields a copper loss difference of 0.518, or 51.8%. The larger value between the iron loss difference of 43% and the copper loss difference of 51.8% was selected, resulting in a loss difference of 51.8%. With the preset difference threshold set at 15%, the loss difference of 51.8% significantly exceeds the preset difference threshold of 15%, triggering the alarm mechanism.

[0079] The calculated ratio of the loss difference to the preset difference threshold was 3.45, which is greater than 2.0. Based on the control strategy, the recommended load power reduction percentage was determined to be 15%, i.e., reducing the load from the current 2602A to 2212A. Simultaneously, a cooling enhancement command was sent to the transformer cooling control system. Since the ratio was greater than 2.0, the command instructed the activation of all three standby cooling fans and the increase of the oil pump speed from the rated speed of 1450 r / min to 1740 r / min, an increase of 20%. After the cooling system responded, the upper oil temperature dropped from 94℃ to 87℃ within 15 minutes, and the winding hot spot temperature dropped from 112℃ to 101℃, returning to a safe range. The load adjustment command was sent to the dispatch center. After coordination, within 30 minutes, 15% of the transformer load was transferred to other lines, the current dropped to 2210A, the predicted copper loss dropped to 48.5kW, the predicted iron loss dropped to 22.3kW, and the loss difference decreased to 7.8%, below the preset difference threshold of 15%. The system then returned to normal operation.

[0080] To verify the accuracy of the model's predictions, the technical team installed a fiber optic temperature measurement system inside the transformer. Fiber optic temperature sensors were placed at six locations: the winding hotspot, the center of the core column, the core yoke, the inner surface of the oil tank, the upper oil temperature, and the lower oil temperature. The measurement accuracy was 0.5℃. Table 1 shows a comparison between the measured data and the model's prediction results.

[0081] Table 1 Comparison of temperature prediction results and measured values

[0082] As shown in Table 1, the absolute errors of temperature prediction at all six locations are within 2℃, with an average absolute error of 1.17℃, meeting the requirements for engineering applications. Regarding loss prediction accuracy, the total loss was calculated to be 97.8kW by measuring the transformer's input and output power using a power analyzer. The model predicted a sum of iron and copper losses of 96.9kW, with a relative error of 0.92%, verifying the model's high-precision prediction capability.

[0083] The magnitude of the leakage magnetic field at 10 points at a certain moment was calculated using the fast calculation methods of CNN-BiLSTM and SFG, and the magnitude was compared with the magnitude calculated using finite element simulation software. The comparison results are shown in Table 2.

[0084] Table 2 Comparison of Leakage Magnetic Fields between the Two Methods

[0085] This method significantly improves the efficiency and accuracy of transformer leakage magnetic field calculation, and greatly shortens the calculation time compared to traditional finite element simulation. Furthermore, compared to other calculation methods, this model exhibits superior performance in nonlinear feature modeling and also achieves faster calculation speeds. In engineering applications, voltage and current time-series data can be collected in real time by sensors and input into the model for prediction, thereby obtaining the magnetic field, loss, and temperature distribution at corresponding moments, achieving near real-time transformer condition monitoring.

[0086] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A transformer loss detection method based on CNN-BiLSTM and SFG, characterized in that, To obtain the parameters of the converter transformer, a finite element simulation model of the converter transformer was established. An external circuit was set up to apply an input voltage source to simulate various operating conditions. The SFG mesh generation algorithm was used to mesh the finite element simulation model of the converter transformer and solve for three-phase voltage waveform data, three-phase current waveform data, magnetic field distribution data, iron loss data, copper loss data, and temperature distribution data. The three-phase voltage waveform data and three-phase current waveform data were time-synchronized and standardized, and then divided into fixed-length time-series sample sequences using a sliding window method. A physical constraint model was constructed, using the fixed-length time-series sample sequences as input data, along with the magnetic field distribution data... Iron loss data, copper loss data, and temperature distribution data are used as output target data to establish a training dataset for training to obtain a trained physical constraint model. The measured waveforms of three-phase voltage and three-phase current during transformer operation are collected in real time and input into the trained physical constraint model. The output of the physical constraint model includes predicted magnetic field distribution, predicted iron loss, predicted copper loss, and predicted temperature distribution. The iron loss difference between the predicted iron loss result and the preset iron loss threshold and the copper loss difference between the predicted copper loss result and the preset copper loss threshold are calculated, and the larger value is selected as the loss difference. When the loss difference is greater than the preset difference threshold, a load adjustment command or a cooling enhancement command is sent.

2. The method of claim 1, wherein, The SFG mesh generation algorithm is an adaptive sparse mesh generation algorithm. After performing initial mesh generation on the finite element simulation model of the converter transformer to obtain an initial mesh element set, the magnetic field gradient value of each initial mesh element is calculated, and initial mesh elements with magnetic field gradient values ​​greater than the magnetic field gradient threshold are selected for the first mesh refinement process. Then, the eddy current density value of the first refined mesh element is calculated, and the first refined mesh elements with eddy current density values ​​greater than the eddy current density threshold are selected for the second mesh refinement process to obtain the final mesh element set.

3. The method according to claim 2, characterized in that, The magnetic field gradient threshold is obtained by sorting the magnetic field gradient values ​​of all initial grid cells in the initial grid cell set in descending order and then selecting the magnetic field gradient value corresponding to the top 30% position after sorting as the magnetic field gradient threshold.

4. The method according to claim 3, characterized in that, The eddy current density threshold is obtained by sorting the eddy current density values ​​of all the first-time encrypted grid cells in the first-time encrypted grid cell set in descending order and then selecting the eddy current density value corresponding to the top 25% position after sorting as the eddy current density threshold.

5. The method according to claim 4, characterized in that, In the final mesh element set, surface boundary impedance conditions are applied to the surfaces of the box wall structure, clamping structure, and pull plate structure. Specifically, the surface boundary impedance conditions are applied by calculating the skin depth values ​​of the box wall structure, clamping structure, and pull plate structure materials based on the operating frequency value, and then combining the material conductivity and magnetic permeability values ​​to calculate the surface impedance values ​​of the box wall, clamping structure, and pull plate, and applying them to the corresponding structural component surfaces.

6. The method according to claim 5, characterized in that, Harmonic excitation conditions are set by superimposing 2nd, 3rd, 5th, 7th, 11th, and 13th harmonic components on the input voltage source, and DC bias conditions are set by injecting DC current components into the neutral point of the converter transformer finite element simulation model.

7. The method according to claim 6, characterized in that, The time synchronization process specifically involves extracting the timestamps of the three-phase voltage waveform data and the three-phase current waveform data, and then using linear interpolation to fill in the data points with misaligned timestamps so that the sampling times of the three-phase voltage waveform data and the three-phase current waveform data are completely corresponding.

8. The method of claim 7, wherein, The standardization process adopts a channel-by-channel standardization method. The mean and standard deviation of the A-phase voltage waveform data, B-phase voltage waveform data, and C-phase voltage waveform data of the three-phase voltage waveform data are calculated respectively. The mean and standard deviation of the A-phase current waveform data, B-phase current waveform data, and C-phase current waveform data of the three-phase current waveform data are calculated respectively. The standardized waveform data is obtained by subtracting the mean of the corresponding phase from the waveform data of each phase and dividing by the standard deviation of the corresponding phase.

9. The method of claim 8, wherein, The sliding window method has a window length of 1000 sampling points and a sliding step size of 500 sampling points. Fixed-length time-series sample sequences are obtained by segmenting the standardized A-phase voltage waveform data, standardized B-phase voltage waveform data, standardized C-phase voltage waveform data, standardized A-phase current waveform data, standardized B-phase current waveform data, and standardized C-phase current waveform data according to the window length and sliding step size.

10. The method according to claim 9, characterized in that, The structure of the physical constraint model includes an input layer, a CNN feature extraction layer, a BiLSTM temporal dependency layer, a physical constraint fusion layer, and a multi-task output layer. The input layer receives a fixed-length temporal sample sequence, the CNN feature extraction layer extracts local pattern features, the BiLSTM temporal dependency layer learns temporal dependencies, the physical constraint fusion layer embeds constraints from Maxwell's equations, the heat conduction equation, and Joule's law, and the multi-task output layer includes a magnetic field prediction branch, a loss prediction branch, and a temperature prediction branch.

Images