A transformer precise prediction method and system considering electro-accelerated aging

Abstract

The application provides a transformer precise prediction method and system considering electro-accelerated aging; in view of the defects of mechanism loss and time domain out of step of the traditional prediction model under the high-frequency current conversion working condition, the application obtains transformer broadband electrical signals and multi-source state data, and extracts electrical shock characteristics; based on the degradation mechanism of molecular thermodynamics, the electrical shock is mapped to an activation energy disturbance factor, and a theoretical gas production rate is deduced; time delay compensation factors are used to align the time domain differences, a fusion prediction network containing physical mechanism residual constraints is constructed, iterative training is carried out, and a prediction result is output; finally, based on the prediction result, a load reduction operation instruction is issued by the bottom layer measurement and control equipment. An interpretable method process from macroscopic electrical transient to microscopic insulation degradation is realized, the model error divergence caused by multi-scale time lag is overcome, a transformer active safety protection closed loop is constructed, and the prediction accuracy and system robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of insulation condition monitoring and protection of power system equipment, and in particular to a method and system for accurate prediction of transformer insulation condition considering electro-induced accelerated aging. Background Technology

[0002] As the core equipment for energy transmission in the power system, the insulation health of transformers directly determines the overall safety of the power grid operation. With the high proportion of large-scale new energy sources such as offshore wind power and photovoltaic converters being connected to the grid, the high-voltage side of transformers is frequently subjected to high-frequency pulse width modulation (PWM) transient voltage impacts generated by the front-end inverters. Such high-frequency alternating electric fields, accompanied by extremely high voltage change rates, can induce strong electromechanical fatigue and accelerated aging inside the oil-paper insulation, leading to the risk of early failure of transformer insulation.

[0003] Currently, the prediction of transformer insulation aging mainly relies on dissolved gas analysis (DGA) in transformer oil and pure data-driven deep learning models (such as long short-term memory networks or graph neural networks). However, pure data-driven models are data black boxes that lack physical and chemical mechanism constraints. When faced with high-frequency sudden changes in new energy operating conditions and small sample insulation deterioration data, they are prone to prediction divergence or overfitting.

[0004] To address the engineering problem of the lack of physical interpretability in data-driven models, existing technologies have proposed a prediction framework that integrates physical information. For example, the paper "Research on Prediction Method of Dissolved Gas in Transformer Oil Based on Physical Information Neural Network Model" discloses a prediction method of dissolved gas in transformer oil based on physical information neural network (PINN). It constructs an oil-paper insulation experimental platform with typical defect models, establishes a macroscopic correlation model between dissolved gas and partial discharge energy, and embeds the macroscopic correlation model into a time-series prediction network for training, in an attempt to improve the prediction accuracy of dissolved gas concentration changes in transformer oil.

[0005] However, after in-depth engineering practice and microscopic theoretical analysis, the above methods and existing similar prediction methods still have the following technical shortcomings when facing complex high-frequency converter operating conditions: The above method only establishes the correlation between gas concentration and partial discharge energy, without exploring in depth the fatigue failure mechanism of the chemical bonds of the microscopic molecular chain of insulating polymer caused by the electromechanical stress induced by the high steepness of the high-frequency pulse width modulation voltage. Under real working conditions, the work done by the transient high-frequency accumulated electromechanical force will significantly weaken the basic thermal activation energy inside the insulating material, leading to the nonlinear collapse of the molecular activation energy barrier that triggers the cracking reaction. Since no activation energy perturbation mapping logic based on molecular thermodynamics is introduced, it is impossible to quantify the dynamic catalytic effect of the high-frequency alternating electric field on the theoretical gas production rate of the characteristic gas bottom layer. As a result, the physical equations in the comparison file lack real microscopic dynamic support, and there is still a large prediction error under harsh electromagnetic transient working conditions. Meanwhile, the transient high-frequency electrical shocks to transformers often occur on a millisecond timescale, while the process of insulating polymers cracking and releasing characteristic gases, dissolving and fully diffusing into the gas chamber of the online chromatographic monitoring probe is a slow diffusion process with a time lag of several hours. When constructing the loss function of the physical information neural network, it forces the electrical input features and gas output features under the same acquisition time section to be synchronously constrained, ignoring the cross-scale time domain lag phenomenon. This leads to serious distortion of the physical residual terms during network training, which can easily cause the gradient descent path to get stuck in a local dead loop. As a result, the prediction model provided by this method has lag and missed alarm defects when capturing early failure conditions. Furthermore, existing transformer condition prediction technologies are limited to forward inference of algorithms in the cloud or local area, taking the aging condition prediction results or gas concentration curves as the final output nodes. They fail to form an active safety defense system with the underlying control architecture. When the prediction model finds that the transformer is facing irreversible accelerated aging or even insulation breakdown risk, the existing methods cannot link the front-end new energy grid-connected inverters or substation relay protection and control devices to perform peak shaving and load reduction intervention actions. They can only passively output alarms and wait for manual power outage maintenance, lacking the engineering value of timely blocking the malignant deterioration of insulation. In summary, given the shortcomings of existing technologies such as lack of microscopic mechanisms, misalignment in the time domain, and failure to consider physical control, there is an urgent need for a novel method and system for accurate prediction of transformers that takes into account electro-induced accelerated aging, so as to achieve high-fidelity simulation of insulation aging state and active safety protection under complex operating conditions. Summary of the Invention

[0006] The main objective of this invention is to provide a method and system for accurate prediction of transformers that takes into account electro-induced accelerated aging, thereby solving the three major problems mentioned above.

[0007] To solve the aforementioned engineering problems, the technical solution adopted in this invention is: a method for accurately predicting transformer aging considering electro-induced accelerated aging, comprising the following steps: S1. Obtain the effective wideband electrical quantity signal of the transformer. Simultaneously, real-time monitoring data vectors of dissolved gas concentration in transformer oil are collected. Historical concentration data series and real-time top oil temperature Generate synchronized timestamp data blocks ; S2, for valid wideband electrical quantity signals Feature extraction is performed to construct electrical shock feature vectors. , to transform the electrical shock feature vector Physical calibration weight matrix and from the synchronized timestamp data block Real-time top oil temperature extracted Combined to determine the equivalent cumulative electromechanical energy And based on the equivalent cumulative electromechanical energy With real-time top oil temperature Calculate activation energy perturbation factor ; S3, draw basic thermal activation energy According to the activation energy perturbation factor With basic thermal activation energy Determine the effective activation energy parameter The effective activation energy parameter and real-time top oil temperature Substituting into the reaction rate formula, the theoretical gas production rate of the characteristic gas is calculated. ; S4, will synchronize the timestamp data block Historical concentration data sequences extracted and electrical shock eigenvectors Input to the initial fusion prediction model China-Israel output network predicts gas production rate The network will predict the gas production rate. With characteristic gas theoretical gas production rate Time delay compensation factor Perform time-delay alignment subtraction to determine the residual terms of the physical equations. Based on the residual terms of the physical equation Construct the total loss function For the initial fusion prediction model Perform iterative training to obtain a fusion prediction model And utilize fusion prediction models Output insulation aging state prediction set ; S5. Set up the insulation aging state prediction dataset With the preset state threshold matrix Comparison to generate tiered early warning signals and respond to tiered early warning signals Send derating operation control signals to external control equipment .

[0008] In the preferred embodiment, step S1 involves obtaining the effective wideband electrical quantity signal of the transformer. The preceding includes: Acquire the instantaneous three-phase voltage sequence on the high-voltage side of the transformer. With the instantaneous current value sequence The sequence of instantaneous three-phase voltage values With the instantaneous current value sequence Circular writing to the rolling buffer queue middle; Based on the instantaneous value sequence of three-phase voltage Calculate the rate of change of voltage And based on the instantaneous current value sequence Determine the abrupt change in the effective value of the current. ; Determining the rate of voltage change Breakthrough in sensitivity coefficient triggered by voltage steepness With power frequency cycle The dynamic threshold of the constraint, or the amount of sudden change in the effective value of the current. Breakthrough in current-limited sensitivity coefficient RMS value of rated current When the constraint threshold is met, generate a trigger timestamp. ; Response trigger timestamp The wideband recording device is controlled to switch to high-frequency sampling mode, and the trigger timestamp is used. From the rolling buffer queue Bidirectional addressing is used to extract signal segments and splice them together to form the original abrupt change data segment. .

[0009] In the preferred embodiment, step S1 further includes: Calling a band-stop filter matrix with a preset stopband frequency range Using a band-stop filter matrix For the original mutation data segment Filtering is performed to obtain a filtered wideband signal. ; Determine the filtered wideband signal Integral energy and original mutation data segment The percentage of total integral energy ; In determining the energy percentage Meets the energy shielding threshold Upon confirmation of a severe electrical surge that caused insulation aging, the filtered broadband signal was... Assigned as a valid wideband electrical quantity signal .

[0010] In the preferred embodiment, step S2 includes: Rainflow counting algorithm is used to analyze effective broadband electrical signals. Perform cyclic statistics to generate a three-dimensional distribution matrix. And the time-frequency domain multi-resolution decomposition method is used to analyze the effective broadband electrical quantity signal. Perform analysis to obtain the signal energy spectrum ; Combined with three-dimensional distribution matrix and signal energy spectrum Dimensionless physical representation parameters of each dimension are extracted, and tensor concatenation and normalization are performed on these parameters to construct an electrical impact feature vector. ; Determining the electromechanical coupling constant based on the dielectric constant of insulating materials According to the physical calibration weight matrix In the electrical shock eigenvector Component weights for each dimension, electrical shock feature vector Transient eigenvalues ​​of each dimension and electromechanical coupling constant The equivalent accumulated electromechanical force energy is determined by nonlinear mapping of the high-frequency pulse energy accumulation damage mechanism. .

[0011] In the preferred embodiment, steps S2 and S3 include: Introducing the reaction rate theory of polymer material aging, the equivalent accumulated electromechanical energy is... With real-time top oil temperature The resulting molecular free energy exponential decay term is equivalent to the accumulated electromechanical energy. Joint settlement is conducted to determine the activation energy perturbation factor at the microscopic level. ; activation energy perturbation factor As a compensating energy for the breaking of microscopic chemical bonds, it performs barrier offset calculations for underlying chemical reactions to determine the effective activation energy parameter. ; Based on effective activation energy parameter With real-time top oil temperature Determine the probability term reflecting the relative height of the molecular reaction barrier, and combine it with the pre-exponential factor constant characterizing the frequency of microscopic molecular collisions. The theoretical gas production rate of the characteristic gas was calculated based on nonlinear reaction kinetic mapping. .

[0012] In the preferred embodiment, step S4 includes: Predicting gas production rate using networks With synchronized timestamp data blocks Real-time monitoring data vector extracted from Perform error comparison to determine the data prediction error term at the data-driven level. ; Based on the set time delay compensation factor Predicting gas production rate using a network Implement time window rolling alignment to match the network-predicted gas production rate after delay compensation correction with the theoretical gas production rate of the characteristic gas. Execution mechanism residual calculation, extraction of physical equation residual terms ; For data prediction error term Assign data prediction weight parameters and for the residual terms of the physical equation Assign physical residual weight parameters The total loss function is determined through feature aggregation and settlement. .

[0013] In the preferred embodiment, step S4 further includes: An adaptive weighted gradient descent strategy is adopted, based on the total loss function. Determine the initial fusion prediction model Network layer parameter matrix The gradient; In the early stages of network training, the physical residual weight parameters are dynamically amplified based on the gradient variance. This forces the initial fusion prediction model It conforms to the molecular thermodynamic degradation boundary; In the later stages of network training, the predicted weight parameters are dynamically amplified based on how closely the predicted values ​​approximate the true labels. ; The network layer parameter matrix is ​​updated through continuous iteration. until the total loss function Once the preset convergence conditions are met, the network structure is locked to generate the trained fusion prediction model. .

[0014] In the preferred embodiment, step S5 includes: From the set of insulation aging condition predictions The remaining insulation life was quantitatively assessed by analyzing the data. and the evolution trend of characteristic gas concentration ; In determining the quantitative assessment value of the remaining insulation life Falling below state threshold matrix The set lower limit of safe lifespan, or the determination of the evolution trend of characteristic gas concentration. When the slope of change exceeds the mutation tolerance, a graded early warning signal corresponding to the risk level is generated based on the number and severity of the dimensions exceeding the limit. .

[0015] In the preferred embodiment, step S5 further includes: In determining the graded early warning signal When the specific risk level identifier carried reaches the set physical intervention threshold, a derated operation control signal with a defined timing control cycle and decrement magnitude is generated. ; The derated operation control signaling is transmitted via the substation industrial control bus. The signal is directed to the wind and solar inverter group control system, triggering the intervention of the inverter switching transistor's pulse width modulation duty cycle to reduce outgoing power, or linking the substation circuit breaker to cut off sensitive load branches that cause harmonic penetration.

[0016] The present invention also provides a transformer accuracy prediction system that takes into account electro-induced accelerated aging, comprising: Intelligent edge waveform recording array for acquiring effective wideband electrical quantity signals Real-time monitoring data vector Historical concentration data series and real-time top oil temperature And combine them to generate synchronized timestamp data blocks. ; A distributed feature processing engine for valid wideband electrical quantity signals. Feature extraction is performed to construct electrical shock feature vectors. Combined with the physical calibration weight matrix and from the synchronized timestamp data block Real-time top oil temperature extracted Determine the equivalent cumulative electromechanical force energy And based on the equivalent cumulative electromechanical energy With real-time top oil temperature Determine the activation energy perturbation factor The distributed feature solving engine is also used to calculate based on the activation energy perturbation factor. With basic thermal activation energy Determine the effective activation energy parameter The effective activation energy parameter and real-time top oil temperature Substituting into the reaction kinetic model, the theoretical gas production rate of the characteristic gas is calculated. ; The cloud-based intelligent network inference computing center is used to retrieve synchronized timestamp data blocks. Historical concentration data sequences extracted and electrical shock eigenvectors Input to the initial fusion prediction model China-Israel output network predicts gas production rate Using time delay compensation factor Predicting gas production rate using networks With characteristic gas theoretical gas production rate Perform time-domain alignment to determine the residual terms of the physical equations. By including the residual terms of the physical equations Total loss function For the initial fusion prediction model Perform iterative training to obtain a fusion prediction model And utilize fusion prediction models Output insulation aging state prediction set ; Active safety protection controller, used to collect insulation aging state prediction sets With state threshold matrix Perform limit-crossing comparisons to generate tiered early warning signals. and respond to tiered early warning signals Send derating operation control signals to external control equipment .

[0017] This invention provides a method and system for accurate prediction of transformer aging considering electro-induced accelerated aging, which has the following advantages compared to the prior art: This invention maps the analytical parameters of the macroscopic transient electrical waveform on the high-voltage side of the transformer in real time to the physicochemical characteristics of the microscopic chain breaking process of the insulating polymer matrix, and establishes a calculation logic for the gas generation rate of the characteristic gas theory with physical logic support. By using the extracted electrical impact feature vector and the molecular thermodynamic degradation mechanism, the activation energy perturbation factor used to offset the chemical reaction barrier of the main chain is calculated, realizing the interpretable connection from the work done by high-frequency electromagnetic transient stress to the change of microscopic bond breaking free energy. It breaks the heavy dependence of traditional black-box deep learning models on large-scale high-quality fault label samples. Even if the transformer faces the high-frequency alternating pulse impact of the inverter and the composite stress deterioration region accompanied by local high humidity micro water intrusion, the prediction system can still maintain the accuracy of state inference and global generalization ability by relying on the physical constraints of the underlying mechanism. Meanwhile, in response to the time-domain out-of-sync phenomenon between the electrical transient response caused by high-frequency inverter and the chromatographic diffusion process of dissolved gases in oil, this invention introduces a time delay compensation mechanism when constructing the total loss function of the physical information neural network. It implements a sliding time window rolling alignment between the network's forward prediction flow and the expected chemical reaction kinetics, eliminating the residual distortion problem caused by the forced synchronous calculation in traditional physical neural networks, correcting the gradient distortion of the network parameter update path in the backpropagation algorithm, reducing the ineffective consumption of edge computing resources in the model iteration process, accelerating global training convergence, and improving the time-domain capture response efficiency of early insulation failure conditions of equipment while ensuring prediction stability. Furthermore, this invention breaks through the limitations of traditional state prediction technology, which can only output predicted gas concentration curves or static alarm scores in a single direction. By directly converting the insulation aging state prediction set output by the fusion prediction model into physical defense commands that can be recognized and transmitted by the industrial control bus, when the transformer insulation faces the risk of aging inflection point or sudden change, it can link the front-end new energy grid-connected inverter group control system or the field relay protection and control device to adjust the pulse width modulation duty cycle of the switching transistor to implement dynamic power peak shaving, or directly cut off the malicious harmonic load branch. This enables the construction of autonomous protection functions before the insulation undergoes physical breakdown, and enhances the robustness of the transformer system in preventing false alarms and missed alarms in complex grid-connected defect environments. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of a method for accurate prediction of transformers that takes into account electro-induced accelerated aging, according to the present invention. Figure 2 This is a schematic diagram of a transformer precision prediction system that takes into account electro-induced accelerated aging, according to the present invention. Detailed Implementation

[0019] Example 1 like Figures 1 to 2 As shown, a method for accurate prediction of transformer aging considering electro-induced accelerated aging includes the following steps: S1. Obtain the effective wideband electrical quantity signal of the transformer. Simultaneously, real-time monitoring data vectors of dissolved gas concentration in transformer oil are collected. Historical concentration data series and real-time top oil temperature Generate synchronized timestamp data blocks ; S2, for valid wideband electrical quantity signals Feature extraction is performed to construct electrical shock feature vectors. , to transform the electrical shock feature vector Physical calibration weight matrix and from the synchronized timestamp data block Real-time top oil temperature extracted Combined to determine the equivalent cumulative electromechanical energy And based on the equivalent cumulative electromechanical energy With real-time top oil temperature Calculate activation energy perturbation factor ; S3, draw basic thermal activation energy According to the activation energy perturbation factor With basic thermal activation energy Determine the effective activation energy parameter The effective activation energy parameter and real-time top oil temperature Substituting into the reaction rate formula, the theoretical gas production rate of the characteristic gas is calculated. ; S4, will synchronize the timestamp data block Historical concentration data sequences extracted and electrical shock eigenvectors Input to the initial fusion prediction model China-Israel output network predicts gas production rate The network will predict the gas production rate. With characteristic gas theoretical gas production rate Time delay compensation factor Perform time-delay alignment subtraction to determine the residual terms of the physical equations. Based on the residual terms of the physical equation Construct the total loss function For the initial fusion prediction model Perform iterative training to obtain a fusion prediction model And utilize fusion prediction models Output insulation aging state prediction set ; S5. Set up the insulation aging state prediction dataset With the preset state threshold matrix Comparison to generate tiered early warning signals and respond to tiered early warning signals Send derating operation control signals to external control equipment .

[0020] In the preferred embodiment, step S1 includes the following steps: S11. Configure basic monitoring parameters after system initialization. After the system is powered on, the preset rated voltage RMS value is obtained. Rated current effective value Compared with conventional chromatographic sampling step size Initialize the broadband waveform recording device on the edge side of the substation, as well as the online oil chromatography monitoring device, temperature sensor, and load monitoring sensor configured on the transformer body; based on the obtained rated voltage RMS value and the effective value of rated current Allocate and activate the rolling buffer queue for temporarily storing electrical transient data. ; S12. Execute an adaptive wideband waveform recording triggering mechanism based on transient changes. The instantaneous three-phase voltage sequence on the high-voltage side of the transformer was acquired using a set base sampling rate of 1kHz. With the instantaneous current value sequence And write it cyclically to the rolling buffer queue. In the middle; real-time calculation of three-phase voltage instantaneous value sequence Rate of change of voltage between two adjacent sampling points and the sequence of instantaneous current values Corresponding change in the effective value of current Then, determine whether any of the following transient mutation conditions are met: Condition 1 is ,in The set voltage steepness trigger sensitivity coefficient, It is the power frequency cycle; Condition 2 is ,in The set current over-limit sensitivity coefficient; When any of the conditions are met, the system generates a trigger timestamp. It also controls the wideband recording device to adaptively switch the sampling channel to a 20kHz high-frequency sampling mode; furthermore, it utilizes trigger timestamps... From the rolling buffer queue Bidirectional addressing from front to back, extracting Signal segments from the first two power frequency cycles to the last eight power frequency cycles are spliced ​​together to form the original abrupt change data segment. ; S13. Implement an anti-interference frequency band filtering mechanism for abnormal fluctuations. Call the pre-configured band-stop filter matrix The stopband frequency range of this matrix is ​​configured as follows: The corresponding known characteristic switching frequency band of the inverter, such as Using a band-stop filter matrix For the original mutation data segment Filtering is performed to remove conventional PWM switching noise, resulting in a filtered wideband signal. Subsequently, the filtered wideband signal is calculated. Integral energy and original mutation data segment The ratio of the total integral energy yields the energy percentage. Comparing energy percentages With the preset energy shielding threshold ,like If this transient change is determined to be a real, severe electrical shock that causes insulation aging, the filtered broadband signal will be... Assigned as a valid wideband electrical quantity signal ;like If so, it is determined to be a normal normal fluctuation and the current data segment is discarded; S14. Simultaneously execute multi-source auxiliary operation data and acquire chromatographic data. Receive trigger timestamp At the trigger timestamp At the precise moment indicated, a command is simultaneously sent to the online oil chromatography monitoring device to trigger the acquisition of real-time monitoring data vectors of dissolved gas concentration in the transformer oil. It includes the concentrations of various characteristic gas components. Synchronously, based on the trigger timestamp Address the historical database and extract the historical concentration data sequence for the corresponding time window. It also calls the temperature sensor and load monitoring sensor to obtain the current top oil temperature. With ambient temperature Ultimately, the acquired effective broadband electrical quantity signal will be... Real-time monitoring data vector Historical concentration data series Top oil temperature With ambient temperature Perform structured splicing to generate synchronized timestamp data blocks. .

[0021] In the preferred embodiment, step S2 includes the following steps: S21. Extracting transient features and combining multidimensional vectors using edge hardware rainflow counting. First, receive valid wideband electrical quantity signals. The data is then imported into the physical state analysis unit deployed in the FPGA chip at the edge node; the rainflow counting algorithm is invoked to analyze the effective broadband electrical quantity signals. The included alternating current electric field stress is statistically analyzed cyclically to extract the stress amplitude, mean stress, and stress cycle number, which represent the local transient stress intensity, and to generate a three-dimensional distribution matrix. ; Synchronous invocation of S-transform for effective wideband electrical quantity signals Perform time-frequency domain multi-resolution decomposition to obtain the signal energy spectrum. Combined with the three-dimensional distribution matrix Included data and signal energy spectrum Calculate the spectral energy entropy, which characterizes the degree of energy dispersion in the frequency band. Steepness Peak duration Energy spectral density Harmonic distortion rate Slope of distortion rate change And directly extract the equivalent stress cycle number from it. ; Finally, the spectral energy entropy Steepness Peak duration Energy spectral density Harmonic distortion rate Slope of distortion rate change and the number of equivalent stress cycles Tensor splicing and normalization are performed according to a fixed sequence to construct a 7-dimensional electrical impact feature vector. ; S22. Calibrate feature weight coefficients through offline comparative experiments. Before the system goes online, the initial set of weight parameters is retrieved from the local knowledge base; this set is obtained through an offline controlled aging experiment using a high-voltage pulse generator combined with a miniature transformer oil tank; during the test, the corresponding 7-dimensional electrical impact feature vector is applied independently. The waveform data for each dimension were analyzed, and the changes in the breaking threshold energy of the chemical bonds in the insulating oil molecules were observed using Raman spectroscopy. After obtaining the gas generation rate increments independently excited by each characteristic parameter, a multivariate nonlinear regression algorithm was used to optimize the fit of the observed data, and the weights of seven components representing the actual contribution of degradation were calculated. During the online operation phase, seven component weights are extracted and combined with a 7-dimensional electrical impact feature vector. Align the dimensional data sequence with the baseline to generate a physical calibration weight matrix. ; S23. Calculate the cumulative energy and then the activation energy perturbation factor based on molecular thermodynamics theory. Obtaining 7-dimensional electrical shock feature vectors Physical calibration weight matrix and real-time top oil temperature Based on the high-frequency pulse cumulative energy damage mechanism, macroscopic electrical parameters are converted into the equivalent cumulative electromechanical force energy of the medium caused by a single transient impact. The constructed mathematical expression is: ; in, Physical calibration weight matrix The first in Each component weight; 7-dimensional electrical shock feature vector The first in One transient eigenvalue; The electromechanical coupling constant is determined based on the dielectric constant of the insulating material; Subsequently, the Eyring reaction rate theory of polymer material aging is introduced, and the electromechanical work performed by the applied high-frequency transient alternating electric field is subtracted directly from the macroscopic activation free energy of the polymer material. Based on this physical mechanism, a microscopic activation energy perturbation factor is constructed. The transformation equation: ; in, The standard Boltzmann constant; The input is the real-time top oil temperature; The external characteristic matrix of the power grid is converted into the activation energy attenuation deviation.

[0022] In the preferred embodiment, step S3 includes the following steps: S31. Introducing activation energy perturbation factors to correct the underlying chemical barrier. Retrieve pre-pre-defined basic thermal activation energy from the system knowledge base Characterize the initial energy barrier required for molecular chain breakage in transformer oil-paper insulation material under pure thermodynamic stress; obtain the activation energy perturbation factor. Based on the catalytic degradation mechanism of transient physical energy in the material aging process, the activation energy perturbation factor is... As a compensating energy for the breaking of microscopic chemical bonds, it effectively reduces the potential barrier height of the original chemical reaction; based on this, a modified effective activation energy parameter is constructed. Its defined mathematical expression is: ; in, This refers to the equivalent microscopic activation energy barrier that insulating materials actually face after being subjected to high-frequency electrical shocks and environmental thermal stress. S32, Combine real-time oil temperature calculation to determine the theoretical gas production rate of characteristic gases Parse synchronized timestamp data blocks Extract the real-time top oil temperature at the current accurate moment. Receive effective activation energy parameter Introducing the Arrhenius reaction rate physical model, the real-time top oil temperature is... and effective activation energy parameters Substituting these values ​​into the improved chemical kinetic equations, the theoretical generation rate of specific characteristic gases, namely the aging characteristic gases that characterize discharge and thermal defects, is calculated. The specific formula is as follows: ; in, The theoretical gas production rate of the characteristic gas at the target time point; The pre-exponential factor constant is used to characterize the frequency of microscopic molecular collisions; The standard Boltzmann constant; Obtain the accurate expected value of gas production rate after considering the high-frequency impact spectrum coupling effect.

[0023] In the preferred embodiment, step S4 includes the following steps: S41. Construct a physical information neural network to calculate delay compensation loss. Obtain the theoretical gas production rate of the characteristic gas 7-dimensional electrical shock feature vector and synchronized timestamp data blocks ; Parse the synchronized timestamp data block Extract historical concentration data sequences With real-time monitoring data vector Establish an initial fusion prediction model. Historical concentration data sequence With 7-dimensional electrical impact eigenvector Input to an initial fusion prediction model such as an LSTM network The data-driven layer outputs a network that predicts the gas production rate. ; To address the time delay issue arising from the breakage of chemical bonds in insulating materials due to high-frequency physical shocks, leading to their dissolution into transformer oil, a time delay compensation factor is constructed. ;Predict gas production rate using network With real-time monitoring data vector Comparative calculations yielded the data prediction error term. Simultaneously, a time delay compensation factor is introduced. The network with time-aligned predictions will generate gas rates. With characteristic gas theoretical gas production rate Perform the difference calculation to obtain the residual term. Finally, the prediction error term is combined with the data. With the residual terms of the physical equation Construct the total loss function The specific formula is as follows: ; in, Predict weight parameters for the data; For physical residual weight parameters; S42. Optimize model parameters using an adaptive weight gradient descent strategy. Obtain the initial fusion prediction model Network layer parameter matrix and load the total loss function. ; When using the Adam gradient descent algorithm to analyze the network layer parameter matrix During the iterative optimization process, an adaptive weight allocation strategy is invoked to dynamically adjust the total loss function. Data prediction weight parameters and physical residual weight parameters ; Specifically, in the early stages of network training, the physical residual weight parameters are amplified based on the statistical variance of gradient backpropagation. The numerical value forces the initial fusion prediction model Following the molecular thermodynamic degradation boundary, overfitting of the neural network under highly fluctuating data is suppressed; in the later stages of training when the loss function tends to stabilize, the first weight, i.e., the data prediction weight parameter, is adaptively amplified based on the degree of approximation between the predicted value and the true label. Refine the fitting of nonlinear mapping relationships; continuously iterate and update the network layer parameter matrix. until the total loss function The convergence value is lower than the preset minimum threshold. Once the convergence condition is met, the updated network parameters are locked, and the trained fusion prediction model is obtained. ; S43. Predicting multidimensional insulation aging state using a training fusion model. During the real-time online prediction phase, the trained fusion prediction model is received and activated. ; from the obtained synchronization timestamp data block Extract the real-time monitoring data vector at the current moment from the parsing. Top oil temperature and the effective value of the load current ; extract the real-time monitoring data vector Top oil temperature Effective value of load current and 7-dimensional electrical impact feature vector Combined into a high-dimensional real-time operating condition matrix ; High-dimensional real-time operating condition matrix Inject the fusion prediction model that has been trained In the forward propagation network, the model calculates the expected values ​​of future states through internally fixed nonlinear weights and physical parameter boundaries, and finally outputs the characteristic gas concentration evolution trend within a preset time step, such as 24 hours. Predicted equivalent degree of polymerization of insulating paper and quantitative assessment value of remaining insulation life Insulation aging state prediction set .

[0024] In the preferred embodiment, step S5 includes the following steps: S51. Compare multidimensional aging prediction indicators to generate graded early warning signals. Obtain the insulation aging state prediction set And analyze the evolution trend of characteristic gas concentration within a preset time step. Predicted equivalent degree of polymerization of insulating paper and quantitative assessment value of remaining insulation life ; Retrieve the pre-set state threshold matrix from the system's local knowledge base This matrix contains multi-level safety baselines configured for different characteristic parameters; it also analyzes the evolution trend of characteristic gas concentrations. Predicted equivalent degree of polymerization of insulating paper and quantitative assessment value of remaining insulation life respectively with the state threshold matrix The corresponding baseline is compared; when the remaining insulation life is quantified... Falling below the set lower limit of safe lifespan, or the evolution trend of characteristic gas concentration. When the slope of change exceeds the mutation tolerance, a corresponding graded early warning signal is generated based on the number and severity of the specific dimensions exceeding the limit. ; S52. Trigger underlying hardware collaborative control actions based on the warning signal. Receive graded early warning signals The system analyzes the specific risk level identifier carried by the signal; when the risk level identifier is determined to reach a pre-set physical intervention threshold, such as a high-risk action zone, a degraded operation control signaling with specific execution timing constraints is generated for that risk level. ; Subsequently, the derating operation control signaling is transmitted via the substation industrial control bus. The forced directional transmission to the relay protection and control device on the transformer's incoming line side and the front-end wind and solar inverter group control system triggers physical peak shaving and load reduction actions, cutting off sensitive load branches that cause severe high-frequency penetration fluctuations.

[0025] In a preferred embodiment, based on the same inventive concept, this application provides a transformer accuracy prediction system that takes into account electro-induced accelerated aging, specifically including: The intelligent edge waveform recording array includes a wideband synchronous waveform recording device with an embedded AD7606 high-precision analog-to-digital converter chip and a high-speed physical memory hardware board. This array acquires basic electrical signals from the high-voltage side of the transformer in real time and continuously calculates waveform abrupt change parameters. When a transient over-limit is detected, a high-frequency sampling mode is triggered to capture the original abrupt change waveform. Subsequently, conventional switching noise is removed by a band-stop filter, and the effective energy ratio is verified before outputting an effective wideband electrical signal. Simultaneously, based on a unified trigger timestamp, an oil chromatography monitoring device is synchronously scheduled to acquire gas chromatography information, and an environmental sensor is linked to collect temperature parameters. Finally, the data is spliced ​​together and output as a synchronous timestamp data block. The distributed feature calculation engine uses Xilinx Kintex-7 series physical programmable gate array chips to build the core logic computing power base. The engine receives the aforementioned effective broadband electrical signals, calls the underlying rainflow counting module to statistically analyze the cyclic distribution of electric field stress, and extracts multi-dimensional transient physical feature vectors by combining time-frequency transformation algorithms. Then, it loads locally calibrated weight parameters and directly converts the macroscopic electrical shock into accumulated electromechanical energy based on molecular thermodynamics. It further deduces the microscopic activation energy perturbation to offset the basic thermal activation energy, and finally calculates the theoretical gas production rate of the feature gas with spectral coupling correction boundary. The cloud-based intelligent network simulation computing center uses an Inspur NF5280M6 digital processing server equipped with an NVIDIA A100 tensor computing chip to perform physical information fusion simulation. The center receives the preceding feature vector and the mechanism constraint boundary, and uses the forward simulation network to predict the gas production rate. By introducing a time delay compensation mechanism, the data prediction error terms and physical equation residual terms are aligned, and an adaptive weight strategy is used to iteratively optimize the model network parameters. In the online simulation phase, the trained fusion model is used to process multi-dimensional real-time operating condition features and accurately outputs a set of insulation aging state predictions that include future gas evolution trends and remaining lifetime assessments. The active safety protection controller adopts the core component of the Schneider Modicon M580 series programmable automation controller. This controller acquires and parses the insulation aging state prediction set, and calls the internal hardware comparator to check the multi-dimensional state safety baseline. When the predicted index falls below the safety lower limit or shows an abnormal sudden trend, the internal logic trigger deflects to generate a graded early warning signal. Subsequently, it sends a derated operation control signal to the external substation control bus, directly driving the front-end new energy inverter group control system to reduce the outgoing line power or physically disconnect specific harmonic load branches, thereby blocking the trend of insulation aging deterioration and completing the system control closed loop.

[0026] Example 2 This embodiment selects a 50kHz high-frequency isolation converter in an offshore wind power booster station as the actual application scenario. It focuses on the high-frequency switching drive conditions unique to this scenario, such as 76% relative humidity, high salt spray intrusion, and extreme voltage change rate. Traditional models are prone to prediction divergence when faced with early failure data caused by water molecule clustering effects. This embodiment introduces a dynamic amplification mechanism of the electromechanical coupling constant of water molecules to quantify the depth of chemical barrier destruction by composite stress.

[0027] S1. Perform adaptive trigger acquisition and anti-interference processing of multi-source operational data. The system's acquisition unit's direct memory access channel receives basic electrical quantity signals from the high-voltage side of the transformer. The input condition includes a drive waveform with a fundamental frequency of 50kHz, corresponding to a conventional sampling step size. The time interval is set to 20 microseconds. The basic monitoring parameters on the high-voltage side include the RMS value of the rated voltage. Rated current effective value and power frequency cycle The relative humidity of the microenvironment is transmitted in real time by the environmental humidity sensor. The standard is 76%; First, after the system is powered on, the preset rated voltage RMS value is retrieved through the microcontroller's configuration bus. RMS value of rated current Initialize the broadband waveform recording device on the edge side of the substation; based on the obtained effective value of the rated voltage. RMS value of rated current Allocate and activate the rolling buffer queue in high-speed non-volatile dynamic random access memory. ; Secondly, the edge recording card acquires the instantaneous three-phase voltage sequence on the high-voltage side of the transformer at a base sampling rate of 1kHz. With the instantaneous current value sequence The rolling loop writes to the rolling buffer queue. The microprocessor's floating-point unit dynamically calculates the rate of voltage change between two adjacent sampling points in real time. and the sudden change in the effective value of the current The floating-point computing unit detects transient electromagnetic penetration on the high-voltage side caused by the inverter's high-frequency switching-on, and calculates the current voltage change rate. It is 5.2 kV per microsecond; Next, the logic determination unit performs boundary condition verification to determine the current rate of voltage change. Breakthrough in voltage steepness trigger sensitivity coefficient With power frequency cycle The dynamic threshold of the common constraint satisfies trigger condition one. At this point, the logic level changes, and a trigger timestamp is automatically generated. And control the wideband recording device to adaptively switch the sampling channel to 20kHz high-frequency sampling mode; Then, using the trigger timestamp From the rolling buffer queue Bidirectional addressing from front to back to extract the trigger timestamp Signal segments from the first two power frequency cycles to the last eight power frequency cycles are spliced ​​together to form the original abrupt change data segment. ; Next, the pre-configured band-stop filter matrix is ​​invoked. Within the configured stopband frequency range Internal to the original mutation data segment Filtering is performed to remove the inherent pulse width modulation switching noise of the new energy grid-connected inverter, resulting in a filtered wideband signal. Further calculation of the filtered wideband signal Integral energy and original mutation data segment The ratio of the total integral energy yields the energy percentage. Due to the proportion of energy Breaking through the preset energy shielding threshold The system determines that the current impact is a severe electrical shock and will filter the wideband signal. Assigned to valid wideband electrical quantity signal ; Finally, at the trigger timestamp During the identification process, the online oil chromatography monitoring device located below the synchronously dispatched transformer is triggered to collect real-time monitoring data vectors of dissolved gas concentrations in the oil. The characteristic gas component concentrations at the current time point are obtained. The concentration of acetylene characteristic gaseous components, a characteristic product of discharge. It is 5.0 microliters per liter; synchronously based on the trigger timestamp. Address the historical database and extract the historical concentration data sequence for the corresponding time window. It also calls a platinum resistance temperature sensor to obtain the current top oil temperature. With ambient temperature Ultimately, this will effectively transmit wideband electrical signals. Real-time monitoring data vector Historical concentration data series Top oil temperature With ambient temperature Perform structured splicing to generate synchronized timestamp data blocks. .

[0028] S2. Extract time-frequency features of broadband electrical quantities and construct a dimensionality-reduced mapping model. To address the high-frequency transient characteristics of offshore wind power, the physical calibration weight matrix is ​​pre-determined in the offline register. The seven component weights The specific configuration value is The maximum weighting coefficient of 0.25 represents the steepness characteristic, used to capture the mechanical shear destructive force of extremely high voltage abrupt changes on the material chain; simultaneously, it is designed for a relative humidity of 76%. When water is introduced into the environment, the water molecules that have penetrated the matrix undergo polarization dipole reversal under a high-frequency alternating electric field. This process reduces the electromechanical coupling constant of the system. The default constant calibration value under normal drying conditions was dynamically increased to... joule; First, the edge computing card acquires valid broadband electrical quantity signals. The data is input into the physical state analysis unit in the FPGA chip; the rainflow counting algorithm is then called to analyze the effective broadband electrical quantity signals. The included alternating current electric field stresses are statistically analyzed, and the stress amplitude, mean stress, and stress cycle number representing the local transient stress intensity are extracted to generate a three-dimensional distribution matrix. Synchronous invocation of S-transform for effective wideband electrical quantity signals Perform time-frequency domain multi-resolution decomposition to obtain the signal energy spectrum. Combined with the three-dimensional distribution matrix Included data and signal energy spectrum The solution module extracts dimensionless physical characterization parameters of each dimension in parallel: spectrum, energy, and entropy. Steepness Peak duration Energy spectral density Harmonic distortion rate Slope of distortion rate change Equivalent stress cycle number The extracted physical waveform parameters are subjected to tensor concatenation and normalization to construct a 7-dimensional electrical impact feature vector. ; Secondly, extract the electrical impact feature vector. And retrieve the physical calibration weight matrix The corresponding weight parameters are used in formula (1) to substitute the waveform characteristics of each dimension and the weight parameters into the cumulative calculation model to obtain the equivalent cumulative electromechanical force energy. : ; The floating-point register outputs the current accumulated electromechanical energy. joule; Then, retrieve the standard Boltzmann constant. and from the synchronized timestamp data block The current real-time top oil temperature is extracted from the analysis. The thermodynamic energy of the current baseline thermally activated molecule is obtained as follows: Joule; for equivalent cumulative electromechanical energy The dimensionless ratio coefficient is determined to be 0.373 by quotienting it with thermodynamic energy. Next, the classical theory of the Eyring reaction rate for polymer dielectric aging is introduced, and formula (2) from Example 1 is used to calculate the activation energy perturbation factor. : ; Right now joule; With the relative humidity of the local air gap The activation energy perturbation factor at the microscopic level increases from 0% in a dry state to 76% in a high-humidity state. It exhibits a significant nonlinear multiplication, with its value ranging from... Joule increased to joule; Specifically, relative humidity The improvement in dielectric constant leads to a large-scale penetration and adsorption of water molecules into the amorphous interface region of the oil-paper insulation; the high-frequency directional polarization of water molecules not only macroscopically increases the dielectric constant and dielectric loss angle, but also forms charge-rich bands in microscopic defects, amplifying the electromechanical forces excited by the local transient alternating electric field, leading to an increase in the electromechanical coupling constant. A significant increase; accompanied by the work done by the overlapping electric field stress with a steep rise edge, although the characteristic sum of squares is fixed, the introduction of a negative exponential decay term that conforms to the change law of molecular thermodynamic potential barrier in the physical mapping model causes the destructive kinetic energy transferred to the polymer backbone to break the original linear change; confirming the existence of a nonlinear mutual catalytic effect between high-frequency electrical shock and high humidity in a harsh environment.

[0029] S3. Construct an improved chemical kinetic model to calculate the theoretical gas production rate. Retrieve the high-energy bond-breaking physical benchmark of epoxy insulating molecules solidified in the non-volatile memory space of the microprocessor, including the fundamental thermal activation energy required for the cleavage of carbon-hydrogen bonds in the polymer matrix backbone. Static calibration value Joule, a characteristic pre-exponential factor constant used to characterize the frequency of microscopic collisions of reacting molecules. Set to per hour microliters per liter; First, the underlying floating-point arithmetic unit receives the microscopic activation energy perturbation factor. Joule, referring to formula (3) in Example 1, substitutes the basic thermal activation energy and activation energy perturbation factor, and performs the energy barrier offset subtraction calculation: ; Lock and output the effective activation energy parameter after considering the high-frequency electro-induced degradation effect. joule; Secondly, parse the synchronization timestamp data block latched in step S1. Extract the current real-time top oil temperature The calculation unit calculates the effective activation energy parameter. Perform calculations with the thermodynamic fundamental energy calculated in step S2, i.e. The relative height of the molecular reaction barrier after electro-induced recombination catalysis was characterized. Then, the standard Arrhenius reaction rate theory framework is introduced, and formula (4) from Example 1 is used. The relative height of the potential barrier is substituted to perform the expected kinetics calculation of micro-aging, and the theoretical gas production rate of the characteristic gas at the target time point is calculated. : ; The current exponential collision fragmentation probability term is obtained as follows: The pre-exponential factor constant Perform a multiplication operation with this probability term, that is microliters per liter per hour; For horizontal comparison, if relative humidity For a 0% dry isolation zone, the corresponding effective activation energy parameter is... for Joule calculated the relative height of the barrier to be 24.033, and the theoretical gas production rate he derived was only... microliters per liter per hour; Overcoming the prediction errors caused by the inability of traditional aging models to quantify the microscopic catalytic mechanism of transient spectrum on the accelerated breaking of chemical bonds, through rigorous derivation, the distortion characteristics of macroscopic power grids are directly projected and mapped into the quantitative decay of the underlying molecular structure barrier, thus establishing accurate underlying data for the data-driven layer.

[0030] S4. Training time delay compensation physical information network predicts insulation aging status. Given the significant macroscopic time-domain hysteresis effect in the electromechanical impact fracture of the resin backbone under the combined high humidity conditions of offshore wind power, the generation of characteristic products, and their complete dissolution and diffusion into the gas cell of the online chromatographic probe, a time delay compensation factor is introduced into the neural network model of the system. The engineering calibration value is preferably set to 4.5 hours; at the same time, in order to balance the relationship between data approximation and physical laws, data prediction weight parameters are configured during model initialization. Physical residual weighting parameters ; First, the cloud-based intelligent network simulation computing center obtains the theoretical gas production rate of the characteristic gas output in step S3. and the 7-dimensional electrical impact feature vector extracted in step S2. Parallel parsing of synchronized timestamp data blocks Extract historical concentration data sequences And a real-time monitoring data vector containing the current concentrations of high-risk gaseous components. The system loads the initial fusion prediction model. In the forward propagation tensor computation flow, the historical concentration data sequence is... and 7-dimensional electrical impact eigenvector Input the Long Short-Term Memory (LSTM) network layer, calculate and predict the current real-time network-predicted gas production rate. microliters per liter per hour; Secondly, the calculation module will use the network to predict the gas production rate. With real-time monitoring data vector By iterating through the mean squared error loss function, the data prediction error term at the pure data-driven level can be obtained. ; Then, the data processing core call is set to a time delay compensation factor of 4.5 hours. Perform time window rolling alignment to match the network-predicted gas production rate after delay compensation with the theoretical gas production rate of the characteristic gas input from the physical mechanism layer. Perform the difference operation to extract the residual terms of the physical equations. ; Next, we apply formula (5) from Example 1 and substitute the data prediction weight parameters. Physical residual weight parameters And the two component loss values, and then execute the total loss function. Adaptive aggregation solution: Solving for the product term is... and The two terms are added together, and the tensor processing core accurately outputs the current total loss function. ; Finally, the system employs an adaptive weight gradient descent strategy to optimize model parameters; the backpropagation algorithm is based on the total loss function. Numerical calculation of parameter matrices for each network layer The gradient; in the early stages of training, the system automatically amplifies the physical residual weight parameters based on the gradient variance. Forced initial fusion prediction model The model aligns with molecular thermodynamic expectations; it adjusts the network layer parameter matrix through continuous gradient alternation. until the total loss function Iterative decay to a preset minimum threshold The following describes the fusion prediction model trained with the network structure locked. ; During the online real-time simulation phase, the system retrieves the current effective value of the load current. Real-time monitoring data vector and top oil temperature , and the 7-dimensional electrical shock feature vector Tensor splicing is performed to construct a high-dimensional real-time working condition matrix. Input it into the trained fusion prediction model In the forward propagation chain, the final predicted output is a multi-dimensional set of insulation aging state predictions. ; It innovatively introduces a time delay compensation mechanism to solve the problem of time-domain step loss in multi-scale time steps. Through physical mechanisms and purely data-driven adaptive hedging, it constrains the blind fitting behavior of the model when facing offshore wind power scenarios with drastic fluctuations and scarce historical samples, thereby improving the generalization ability of state inference.

[0031] S5. Generate an early warning command based on the aging prediction results to trigger the underlying hardware control. We introduce pure data-driven models such as pure graph neural networks or long short-term memory networks disclosed in the comparative literature as the traditional baseline algorithm. This traditional method completely severs the causal chain of "electrochemical barrier collapse" when performing loss optimization, and lacks the residual term of the physical equation. Rigid constraints; First, the active safety protection controller obtains the insulation aging state prediction set issued in step S4. It was then unpacked and disassembled to extract the current quantitative assessment value of the remaining insulation life. Year; Secondly, the controller calls the state threshold matrix stored in the local EEPROM. The comparator determines the current quantitative assessment value of the remaining insulation life. The annual value has fallen below the set minimum safe lifespan of 3 years, and the trend of characteristic gas concentration evolution... When the inflection point growth rate exceeds the mutation threshold, the hardware logic gate flips, and the controller immediately triggers the generation of a red high-risk level warning signal. ; Then, the edge linkage module responds to the hierarchical early warning signal. Based on the high-risk identifier and the preset defense strategy, a derated operation control signaling with a defined timing control period and decrement range is generated. ; Next, the system transmits the derating operation control signaling through the industrial control bus deployed at the substation site. The wind and solar inverter group control system is directed to the front end of the isolation converter to control the pulse width modulation duty cycle of the switching transistor, cut off the sensitive load branch that causes serious harmonic penetration, limit the overall output power to 30% of the rated value, block the vicious process of epoxy insulation evolving into breakdown, and achieve a closed loop of technical control. The system collected 1200 hours of real fault evolution data of offshore wind power step-up transformers under high humidity, low water, and premature death conditions. The algorithm in this application was compared with the traditional baseline algorithm, and the following comparison table was output:

[0032] Examining the data in the comparison matrix table above, it can be determined that the traditional Baseline algorithm suffers a prediction error rate soaring to 38.6% when facing the high humidity and micro-water change conditions in offshore wind power scenarios, and generates a capture delay of up to 148 hours. The fundamental reason for this technical defect is that pure graph neural networks or long short-term memory networks are essentially black-box models based on probability distribution assumptions. When water molecule polarization triggers accelerated electrochemical aging, leading to the explosive generation of characteristic gases in a short period of time, the historical chromatographic time series does not contain statistical samples of such extreme nonlinear changes. Furthermore, due to the lack of effective microscopic activation energy parameters... With the introduction of feedforward, traditional networks can only passively fine-tune weights several days after the apparent gas production occurs, resulting in a huge false alarm rate. This scheme is based on the Eyring molecular thermodynamic degradation mechanism and incorporates the 7-dimensional electrical shock characteristic vector. The energy barrier hedging value is calculated in real time to determine the upper limit of the theoretical gas production rate of the characteristic gas; in the parameter update in step S4, the residual terms of the physical equation are... By constructing rigid requirements in the hyperparameter search space, blindly overfitting gradients that deviate from thermodynamic causality will be penalized. As a result, the system achieved high-precision convergence in just 120 iterations, keeping the error within 4.1%, breaking the dependence of black-box models on large-scale high-quality labeled samples, and demonstrating the engineering and technical advantages of active protection for power grid equipment.

[0033] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for accurate prediction of transformer aging considering electro-induced accelerated aging, characterized in that, Includes the following steps: S1. Obtain the effective wideband electrical quantity signal of the transformer. Simultaneously, real-time monitoring data vectors of dissolved gas concentration in transformer oil are collected. Historical concentration data series and real-time top oil temperature Generate synchronized timestamp data blocks ; S2, for valid wideband electrical quantity signals Feature extraction is performed to construct electrical shock feature vectors. , to transform the electrical shock feature vector Physical calibration weight matrix and from the synchronized timestamp data block Real-time top oil temperature extracted Combined to determine the equivalent cumulative electromechanical energy And based on the equivalent cumulative electromechanical energy With real-time top oil temperature Calculate activation energy perturbation factor ; S3, Retrieve basic thermal activation energy According to the activation energy perturbation factor With basic thermal activation energy Determine the effective activation energy parameter The effective activation energy parameter and real-time top oil temperature Substituting into the reaction rate formula, the theoretical gas production rate of the characteristic gas is calculated. ; S4, will synchronize the timestamp data block Historical concentration data sequences extracted and electrical shock eigenvectors Input to the initial fusion prediction model China-Israel output network predicts gas production rate The network will predict the gas production rate. Compared with the theoretical gas production rate of characteristic gases Time delay compensation factor Perform time-delay alignment subtraction to determine the residual terms of the physical equations. Based on the residual terms of the physical equation Construct the total loss function For the initial fusion prediction model Perform iterative training to obtain a fusion prediction model And utilize fusion prediction models Output insulation aging state prediction set ; S5. Set up the insulation aging state prediction dataset With the preset state threshold matrix Comparison to generate tiered early warning signals and respond to tiered early warning signals. Send derating operation control signals to external control equipment .

2. The method for accurate prediction of transformers considering electro-induced accelerated aging according to claim 1, characterized in that, In step S1, the effective wideband electrical quantity signal of the transformer is obtained. The preceding includes: Acquire the instantaneous three-phase voltage sequence on the high-voltage side of the transformer. With the instantaneous current value sequence The sequence of instantaneous three-phase voltage values With the instantaneous current value sequence Circular writing to the rolling buffer queue middle; Based on the instantaneous value sequence of three-phase voltage Calculate the rate of change of voltage And based on the instantaneous current value sequence Determine the abrupt change in the effective value of the current. ; Determining the rate of voltage change Breakthrough in sensitivity coefficient triggered by voltage steepness With power frequency cycle The dynamic threshold of the constraint, or the amount of sudden change in the effective value of the current. Breakthrough in current-limited sensitivity coefficient RMS value of rated current When the constraint threshold is met, generate a trigger timestamp. ; Response trigger timestamp The wideband recording device is controlled to switch to high-frequency sampling mode, and the trigger timestamp is used. From the rolling buffer queue Bidirectional addressing is used to extract signal segments and splice them together to form the original abrupt change data segment. .

3. The method for accurate prediction of transformers considering electro-induced accelerated aging according to claim 2, characterized in that, Step S1 also includes: Calling a band-stop filter matrix with a preset stopband frequency range Using a band-stop filter matrix For the original mutation data segment Filtering is performed to obtain a filtered wideband signal. ; Determine the filtered wideband signal Integral energy and original mutation data segment Total integral energy percentage ; In determining the energy percentage Meets the energy shielding threshold Upon confirmation of a severe electrical surge that caused insulation aging, the filtered broadband signal was... Assigned as a valid wideband electrical quantity signal .

4. The method for accurate prediction of transformers considering electro-induced accelerated aging according to claim 1, characterized in that, Step S2 includes: Rainflow counting algorithm is used to analyze effective broadband electrical signals. Perform cyclic statistics to generate a three-dimensional distribution matrix. And the time-frequency domain multi-resolution decomposition method is used to analyze the effective broadband electrical quantity signal. Perform analysis to obtain the signal energy spectrum ; Combined with three-dimensional distribution matrix and signal energy spectrum Dimensionless physical representation parameters of each dimension are extracted, and tensor concatenation and normalization are performed on these parameters to construct an electrical impact feature vector. ; Determining the electromechanical coupling constant based on the dielectric constant of insulating materials According to the physical calibration weight matrix In the electrical shock eigenvector Component weights for each dimension, electrical shock feature vector Transient eigenvalues ​​of each dimension and electromechanical coupling constant The equivalent accumulated electromechanical force energy is determined by nonlinear mapping of the high-frequency pulse energy accumulation damage mechanism. .

5. The method for accurate prediction of transformers considering electro-induced accelerated aging according to claim 1, characterized in that, Steps S2 and S3 include: Introducing the reaction rate theory of polymer material aging, the equivalent accumulated electromechanical energy is... With real-time top oil temperature The resulting molecular free energy exponential decay term is equivalent to the accumulated electromechanical energy. Joint settlement is conducted to determine the activation energy perturbation factor at the microscopic level. ; activation energy perturbation factor As a compensating energy for the breaking of microscopic chemical bonds, it performs barrier offset calculations for underlying chemical reactions to determine the effective activation energy parameter. ; Based on the effective activation energy parameter With real-time top oil temperature Determine the probability term reflecting the relative height of the molecular reaction barrier, and combine it with the pre-exponential factor constant characterizing the frequency of microscopic molecular collisions. The theoretical gas production rate of the characteristic gas was calculated based on nonlinear reaction kinetic mapping. .

6. The method for accurate prediction of transformers considering electro-induced accelerated aging according to claim 1, characterized in that, Step S4 includes: Predicting gas production rate via network With synchronized timestamp data blocks Real-time monitoring data vector extracted from Perform error comparison to determine the data prediction error term at the data-driven level. ; Based on the set time delay compensation factor Predicting gas production rate using a network Implement time window rolling alignment to match the network-predicted gas production rate after delay compensation correction with the theoretical gas production rate of the characteristic gas. Execution mechanism residual calculation, extraction of physical equation residual terms ; For data prediction error term Assign data prediction weight parameters and for the residual terms of the physical equation Assign physical residual weight parameters The total loss function is determined through feature aggregation settlement. .

7. The method for accurate prediction of transformers considering electro-induced accelerated aging according to claim 6, characterized in that, Step S4 also includes: An adaptive weighted gradient descent strategy is adopted, based on the total loss function. Determine the initial fusion prediction model Network layer parameter matrix The gradient; In the early stages of network training, the physical residual weight parameters are dynamically amplified based on the gradient variance. This forces the initial fusion prediction model It conforms to the molecular thermodynamic degradation boundary; In the later stages of network training, the predicted weight parameters are dynamically amplified based on how closely the predicted values ​​approximate the true labels. ; The network layer parameter matrix is ​​updated through continuous iteration. until the total loss function Once the preset convergence conditions are met, the network structure is locked to generate the trained fusion prediction model. .

8. The method for accurate prediction of transformers considering electro-induced accelerated aging according to claim 1, characterized in that, Step S5 includes: From the set of insulation aging condition predictions The remaining insulation life was quantitatively assessed by analyzing the data. and the evolution trend of characteristic gas concentration ; In determining the quantitative assessment value of the remaining insulation life Falling below state threshold matrix The set lower limit of safe lifespan, or the determination of the evolution trend of characteristic gas concentration. When the slope of change exceeds the mutation tolerance, a graded early warning signal corresponding to the risk level is generated based on the number and severity of the dimensions exceeding the limit. .

9. The method for accurate prediction of transformers considering electro-induced accelerated aging according to claim 8, characterized in that, Step S5 also includes: In determining the graded early warning signal When the specific risk level identifier carried reaches the set physical intervention threshold, a derated operation control signal with a defined timing control cycle and decrement magnitude is generated. ; The derated operation control signaling is transmitted via the substation industrial control bus. The signal is directed to the wind and solar inverter group control system, triggering the intervention of the inverter switching transistor's pulse width modulation duty cycle to reduce outgoing power, or linking the substation circuit breaker to cut off sensitive load branches that cause harmonic penetration.

10. A transformer accuracy prediction system considering electro-induced accelerated aging, characterized in that, include: Intelligent edge waveform recording array for acquiring effective wideband electrical quantity signals Real-time monitoring data vector Historical concentration data series and real-time top oil temperature And combine them to generate synchronized timestamp data blocks. ; A distributed feature processing engine for valid wideband electrical quantity signals. Feature extraction is performed to construct electrical shock feature vectors. Combined with the physical calibration weight matrix and from the synchronized timestamp data block Real-time top oil temperature extracted Determine the equivalent cumulative electromechanical energy And based on the equivalent cumulative electromechanical energy With real-time top oil temperature Determine the activation energy perturbation factor ; The distributed feature solving engine is also used based on the activation energy perturbation factor. With basic thermal activation energy Determine the effective activation energy parameter The effective activation energy parameter and real-time top oil temperature Substituting into the reaction kinetic model, the theoretical gas production rate of the characteristic gas is calculated. ; The cloud-based intelligent network inference computing center is used to retrieve synchronized timestamp data blocks. Historical concentration data sequences extracted and electrical shock eigenvectors Input to the initial fusion prediction model China-Israel output network predicts gas production rate Using time delay compensation factor Predicting gas production rate via network Compared with the theoretical gas production rate of characteristic gases Perform time-domain alignment to determine the residual terms of the physical equations. By including the residual terms of the physical equations Total loss function For the initial fusion prediction model Perform iterative training to obtain a fusion prediction model And utilize fusion prediction models Output insulation aging state prediction set ; Active safety protection controller, used to collect insulation aging state prediction sets With state threshold matrix Perform limit-crossing comparisons to generate tiered early warning signals. and respond to tiered early warning signals. Send derating operation control signals to external control equipment .

Images