Transformer excitation inrush risk identification method, training method and electronic equipment

Abstract

This application provides a method, training method, and electronic device for identifying transformer inrush current risks, mainly relating to the field of power system relay protection and control technology. The method includes: determining multiple target operating state data from various real-time operating state data of the target transformer; processing the multiple target operating state data based on an inrush current risk identification model to obtain a risk probability value corresponding to the inrush current; and determining the risk identification result for the inrush current based on the risk probability value and a dynamic identification threshold corresponding to the safety tolerance requirements of the target transformer.

Description

Technical Field

[0001] This application relates to the field of power system relay protection and control technology, and more specifically, to a method, training method and electronic equipment for identifying transformer inrush current risk. Background Technology

[0002] The inrush current generated when a transformer is closed under no-load conditions can reach several times the rated current. Its waveform characteristics are highly similar to those of the internal fault current, which can easily lead to maloperation of differential protection. This is a classic problem in the field of relay protection.

[0003] Traditional methods linearize factors such as residual magnetism and closing angle, neglecting the complex nonlinear coupling between harmonics, DC bias, and core saturation characteristics. This leads to a persistently high risk of misjudgment under complex operating conditions. Furthermore, existing data-driven methods lack physically interpretable decision logic, mostly prioritizing overall accuracy and ignoring the asymmetric risk control requirements of relay protection. Summary of the Invention

[0004] In view of this, this application provides a method for identifying transformer inrush current risks, a training method, and an electronic device.

[0005] One aspect of this application provides a method for identifying transformer inrush current risk, comprising: determining multiple target operating state data from multiple real-time operating state data of a target transformer; processing the multiple target operating state data based on an inrush current risk identification model to obtain a risk probability value corresponding to the inrush current; and determining a risk identification result for the inrush current based on the risk probability value and a dynamic identification threshold corresponding to the safety tolerance requirements of the target transformer.

[0006] One aspect of this application provides a method for training a transformer inrush current risk identification model, comprising: simulating a digital twin model of a target transformer to obtain multiple sample data corresponding to multiple target operating state data; using a Bayesian optimization algorithm to globally optimize the hyperparameters of the initial model to determine the optimal hyperparameter combination; and training the initial model using the multiple sample data based on the optimal hyperparameter combination to obtain the inrush current risk identification model.

[0007] Another aspect of this application provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method as described above.

[0008] Another aspect of this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the method described above.

[0009] Another aspect of this application provides a computer program product including computer-executable instructions that, when executed, implement the method described above.

[0010] This application provides a method for identifying transformer inrush current risk. It generates physically consistent samples using a digital twin model, solving the overfitting problem caused by the scarcity of real fault data. By selecting features with nonlinear coupling relationships as input, the model can effectively cope with complex operating conditions under multi-source disturbances, achieving accurate quantification of inrush current risk. Simultaneously, based on the safety tolerance requirements of different substations, adaptive calibration is performed using dynamic identification thresholds to ensure extremely low false negative rates for high-risk identification, significantly improving the accuracy, scenario adaptability, and engineering practicality of risk identification. Attached Figure Description

[0011] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0012] Figure 1 The schematic diagram illustrates a flowchart of a transformer inrush current risk identification method according to an embodiment of this application;

[0013] Figure 2 A flowchart illustrating a transformer inrush current risk identification model training method according to an embodiment of this application is shown schematically.

[0014] Figure 3 A block diagram of a transformer inrush current risk identification device according to an embodiment of this application is shown schematically;

[0015] Figure 4 A block diagram of a transformer inrush current risk identification model training device according to an embodiment of this application is shown schematically.

[0016] Figure 5 A block diagram of an electronic device suitable for implementing a transformer inrush current risk identification method according to an embodiment of this application is shown schematically. Detailed Implementation

[0017] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0018] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0019] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0020] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0021] Transformer inrush current is a classic technical challenge that has long been faced in the field of power system relay protection. When an unloaded transformer is closed, due to the magnetic saturation characteristics of the iron core, an inrush current with an amplitude several to tens of times higher than the rated current is generated. The waveform characteristics of this inrush current are highly similar to those of the internal fault current, which can easily lead to malfunction of differential protection and seriously threaten the safe and stable operation of the power grid.

[0022] The physical mechanisms and methods of related technologies, such as second harmonic braking and flux calculation, have inherent defects. These methods typically linearize factors such as residual magnetism and closing angle, neglecting the nonlinear coupling relationship between disturbances such as grid harmonics, DC bias, and voltage fluctuations and the magnetic saturation characteristics of the iron core. This makes it difficult to accurately distinguish between fault current and inrush current under complex grid conditions, resulting in a high risk of protection misjudgment.

[0023] Data-driven intelligent diagnostic methods that have emerged in recent years also face challenges. On the one hand, purely data-driven models are like black boxes, lacking physical interpretability in their decision-making logic and resulting in low trust in engineering applications. On the other hand, such models often prioritize overall accuracy as an optimization objective, neglecting the asymmetric risk control requirements in the relay protection field, where it is better to err on the side of caution than to fail to operate. Furthermore, traditional algorithms suffer from high computational overhead and are prone to getting trapped in local optima during hyperparameter optimization, making it difficult to meet the millisecond-level real-time control requirements of power systems.

[0024] In view of this, the embodiments of this application generate sample data through a digital twin model, enabling the initial model to obtain sufficient and physically consistent training data even when real fault samples are scarce, thus avoiding overfitting from the source. By selecting various target operating state data with nonlinear physical coupling relationships as input and processing them in conjunction with a model capable of learning such coupling relationships, the system can effectively cope with complex operating conditions under the combined action of multiple disturbances, achieving accurate quantitative assessment of excitation inrush current risk. Simultaneously, by using a dynamic identification threshold, the risk identification results can be adaptively calibrated according to the safety tolerance of different substations, ensuring an extremely low false negative rate in the identification of high-risk inrush current events, thereby significantly improving the accuracy, scenario adaptability, and engineering practicality of transformer excitation inrush current risk identification.

[0025] Specifically, embodiments of this application provide a transformer inrush current risk identification method based on physical-data dual-drive, wherein the method includes: determining multiple target operating state data from multiple real-time operating state data of the target transformer; processing the multiple target operating state data based on the inrush current risk identification model to obtain a risk probability value corresponding to the inrush current; and determining the risk identification result for the inrush current based on the risk probability value and a dynamic identification threshold corresponding to the safety tolerance requirements of the target transformer.

[0026] It should be noted that the transformer inrush current risk identification method determined in this application embodiment can be used in the field of power system relay protection and control technology. The transformer inrush current risk identification method determined in this application embodiment can also be used in any field other than power system relay protection and control technology, such as the field of condition monitoring and fault diagnosis of other industrial equipment based on physical-data dual-drive. The application field of the transformer inrush current risk identification method determined in this application embodiment is not limited.

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0028] Figure 1 The schematic diagram illustrates a flowchart of a transformer inrush current risk identification method according to an embodiment of this application.

[0029] like Figure 1 As shown, the transformer inrush current risk identification method includes operations S110 to S130.

[0030] In operation S110, various target operating status data are determined from multiple real-time operating status data of the target transformer.

[0031] In operation S120, various target operating state data are processed based on the inrush current risk identification model to obtain a risk probability value corresponding to the inrush current.

[0032] In operation S130, according to the risk probability value and the dynamic identification threshold corresponding to the safety tolerance requirement of the target transformer, the risk identification result for the inrush current is determined.

[0033] The target transformer can represent a specific power transformer for which inrush current risk identification is required. In one example, the transformer can be a single-phase or three-phase transformer, and its specifications and parameters such as core material, rated voltage, rated current, etc. do not affect the implementation of this method and are adaptable to transformers in various power systems.

[0034] The real-time operating state data can represent the operating-related data that can reflect the current working conditions of the target transformer collected in real time during the operation of the target transformer through various monitoring devices such as current transformers, voltage transformers, sensors, etc., including but not limited to voltage data, current data, opening records, closing records, core flux data, grid frequency data, etc.

[0035] The various target operating state data can represent the operating state data selected from the above various real-time operating state data that have a stronger influence on the inrush current and have a physical coupling relationship with each other, so as to be used as the input for subsequent model processing.

[0036] The physical coupling relationship represents the interaction relationship formed between various target operating state data based on physical laws such as electromagnetic induction and core hysteresis. For example, the change of a certain type of data will cause the change of the core magnetic saturation degree, which in turn affects another type of data related to magnetic flux, and finally jointly affects the peak value and duration of the inrush current.

[0037] [[ID=,18]]In this embodiment, through the monitoring system supporting the target transformer, various real-time operating state data during its operation are collected in real time. Based on the generation mechanism of the transformer inrush current, various target operating state data with a stronger influence on the inrush current are selected. At the same time, through physical analysis, it is confirmed that there is a physical coupling relationship between the selected target operating state data, providing physical constraints for subsequent data modeling and model training, and avoiding meaningless feature screening.

[0038] The inrush current risk identification model represents a model used to quantitatively evaluate the inrush current risk and output a risk probability value. This risk identification model can, based on the input target operating state data, obtain the probability that the inrush current corresponding to this data is a high risk through model operation.

[0039] The initial model refers to a basic model that has not been trained and lacks risk identification capabilities. In specific implementations, the initial model can be a machine learning model, such as a decision tree model or a neural network. It should be noted that this embodiment does not specifically limit the type of the initial model; as long as it can achieve classification or probability prediction, it is within the scope of protection of this application.

[0040] Multiple sample data represent the dataset used to train the initial model. The data type of the sample data is the same as the data type of the target running state data.

[0041] A digital twin model represents a simulation model that replicates the physical characteristics of a real transformer, built based on the actual physical parameters of the target transformer. This digital twin model can be used to simulate the transformer's operation, core hysteresis characteristics, and the generation process of inrush current.

[0042] In this embodiment, the digital twin model includes physical equations describing the hysteresis characteristics of the iron core and voltage balance equations coupled to these physical equations.

[0043] The physical equations describing the hysteresis characteristics of transformer cores are mathematical equations used to accurately simulate hysteresis phenomena in transformer cores, such as the nonlinear relationship between core magnetization and magnetic field strength. These physical equations can be used to reconstruct the physical process of core magnetic saturation, providing physical constraints for sample data and avoiding the generation of invalid samples that deviate from actual physical laws. Examples of these physical equations include the differential equations corresponding to the Jiles-Atherton (JA) hysteresis theory and the equations corresponding to the Preisach hysteresis model.

[0044] The voltage balance equation describes the balance between voltage, current, inductance, and other parameters in a transformer circuit. This voltage balance equation is coupled with the physical equation describing the hysteresis characteristics of the iron core to simulate the electromagnetic coupling during transformer operation, ensuring that the simulated sample data conforms to the actual operating conditions of the power grid.

[0045] In this embodiment, a digital twin model of the target transformer is constructed based on the physical equation describing the hysteresis characteristics of the iron core and the voltage balance equation coupled with the physical equation. By simulating and solving this digital twin model under different operating conditions, an enhanced sample dataset rich in physical laws is generated, covering multiple boundary conditions. This enhanced sample dataset includes multiple sample data corresponding to various target operating state data, in order to solve the problem of scarce and unevenly distributed inrush current fault samples in actual power grids.

[0046] The risk probability value represents the probability that the inrush current corresponding to the current target operating state data, as output by the excitation inrush current risk identification model, is a high-risk event. In a specific implementation, the risk probability value can be configured to range from 0 to 1. The closer the probability value is to 1, the greater the probability that the excitation inrush current is high-risk. The closer it is to 0, the greater the probability that the excitation inrush current is in a safe state.

[0047] In this embodiment, multiple types of generated sample data are input into an initial model to train it. This allows the initial model to learn the correspondence between target operating state data and inrush current risk, ultimately resulting in an inrush current risk identification model with risk identification capabilities. Multiple target operating state data selected in real-time are then input into the trained inrush current risk identification model. The model calculates and outputs the corresponding risk probability value, completing a quantitative assessment of the current inrush current risk.

[0048] The safe operation procedures refer to the rules and regulations formulated by the substation or power system where the target transformer is located for the safe operation of the transformer. These include the safety tolerance requirements for the risk of inrush current, such as the allowable rate of missed alarms and false alarms.

[0049] The dynamic identification threshold represents a risk probability decision boundary that is dynamically adjusted based on the safety operation procedures of the target transformer. This dynamic identification threshold can adaptively adjust according to the safety tolerance of the substation to meet the asymmetric risk control requirement of relay protection: "It is better to err on the side of caution than to fail to operate," ensuring that high-risk inrush currents are not missed.

[0050] The risk identification result represents the final output, which is the judgment result of the current excitation inrush current risk status.

[0051] In this embodiment, the safety tolerance of the target transformer is determined according to the safety operation procedures of the substation where it is located, and a corresponding dynamic identification threshold is determined based on this safety tolerance. The risk probability value output by the inrush current risk identification model is compared with the dynamic identification threshold. If the risk probability value exceeds the dynamic identification threshold, the current inrush current is determined to be a high-risk event, and the corresponding inrush current suppression strategy is triggered. If the risk probability value is less than the dynamic identification threshold, the current inrush current is determined to be a safe state, and no suppression strategy is triggered, thus completing the entire inrush current risk identification process.

[0052] Based on this, the embodiments of this application generate sample data through a digital twin model, enabling the initial model to obtain sufficient and physically consistent training data even when real fault samples are scarce, thus avoiding overfitting from the outset. By selecting various target operating state data with nonlinear physical coupling relationships as input and processing them in conjunction with a model capable of learning such coupling relationships, the system effectively addresses complex operating conditions under the combined effects of multiple disturbances, achieving accurate quantitative assessment of excitation inrush current risk. Simultaneously, a dynamic identification threshold allows the risk identification results to be adaptively calibrated according to the safety tolerance of different substations, ensuring an extremely low false negative rate in the identification of high-risk inrush current events. This significantly improves the accuracy, scenario adaptability, and engineering practicality of transformer excitation inrush current risk identification.

[0053] According to embodiments of this application, the various target operating status data include at least residual magnetism data, closing angle data, DC bias data, harmonic distortion data, and voltage jitter data.

[0054] According to embodiments of this application, various target operating status data are determined from multiple real-time operating status data of the target transformer, including: determining the tripping angle based on the historical tripping operation records of the target transformer, and estimating the residual magnetism data by combining the physical equation describing the hysteresis characteristics of the iron core; determining the closing angle data based on the zero-crossing detection of the grid voltage; extracting the DC component from the current data collected by the current transformer to obtain DC bias data; calculating the total harmonic distortion rate based on the voltage waveform or current waveform to obtain harmonic distortion data; and calculating the rate of change based on the effective voltage value of multiple consecutive cycles to obtain voltage jitter data.

[0055] Residual magnetism data represents the parameters corresponding to the residual magnetic flux in the transformer core after the transformer is switched off. When the transformer is switched on, if there is residual magnetism in the core, and the direction of the residual magnetism is consistent with the direction of the magnetic flux at the moment of switching on, it will cause the core to quickly enter a deep saturation state, thereby generating a large-amplitude inrush current. Therefore, residual magnetism data is one of the core data for characterizing the risk of inrush current.

[0056] In a specific implementation, since residual magnetism data cannot be directly measured, it can be determined as follows: Historical tripping operation records of the target transformer are obtained, and key information such as the grid voltage phase and core magnetic flux change at the moment of tripping is extracted. Subsequently, the tripping information is substituted into the physical equation describing the core hysteresis characteristics, and the residual magnetism data corresponding to the residual magnetic flux of the core after tripping is calculated through inverse solving. The value range corresponds to… Magnetic flux angle ensures the physical accuracy and real-time performance of residual magnetism data.

[0057] The closing angle data represents the phase angle of the grid voltage at the instant the transformer is switched on. This closing angle directly affects the rate of change of magnetic flux in the core at the moment of switching. When the closing angle is near the voltage zero-crossing point, the rate of change of magnetic flux is at its maximum, and the core is prone to saturation, generating inrush current. If the closing angle is near the voltage peak, the rate of change of magnetic flux is smaller, and the amplitude of the inrush current will also decrease accordingly. Therefore, the closing angle data directly determines the probability and magnitude of inrush current generation.

[0058] In a specific implementation, the closing angle data can be determined as follows: real-time acquisition of grid voltage waveform data of the target transformer, accurate determination of the voltage zero-crossing point through a zero-crossing detection algorithm; then, based on the voltage zero-crossing point, detection of the actual time of transformer closing operation, calculation of the angle between the closing time and the reference phase, obtaining the closing angle data, with a value range of 0~360°, directly reflecting the driving conditions of magnetic flux change in the core at the moment of closing.

[0059] DC bias data represents the parameter data corresponding to the DC component present in the transformer windings, mainly originating from DC interference in the power grid, rectifier equipment, etc. DC bias changes the magnetization point of the iron core, causing the hysteresis loop of the iron core to shift, reducing the saturation magnetic flux density of the iron core, making the iron core more likely to enter the saturation state, and thus inducing inrush current. The larger the amplitude of DC bias, the higher the risk of inrush current.

[0060] In a specific implementation, the DC bias data can be determined as follows: the operating current data of the transformer is collected in real time by a current transformer, and the DC component is separated from the AC current waveform by signal processing methods such as filtering and integration. The value of the DC component is used as the DC bias data to accurately reflect the influence of DC interference from the power grid on the magnetization of the iron core.

[0061] Harmonic distortion data represents the degree to which the voltage or current waveform deviates from a sinusoidal waveform during transformer operation, and is typically characterized by the total harmonic distortion rate (THC). Because harmonic distortion causes nonlinear distortion in the magnetization process of the core, exacerbating magnetic saturation, and because harmonic components themselves are superimposed on the inrush current, increasing its amplitude and complexity, harmonic distortion data is a primary indicator of inrush current risk.

[0062] In a specific implementation, harmonic distortion data can be determined as follows: real-time acquisition of voltage or current waveform data of the transformer, decomposition of the fundamental component and each harmonic component in the waveform using signal analysis methods such as Fourier transform, calculation of total harmonic distortion rate, and use of this value as harmonic distortion data to quantify the influence of waveform distortion on core magnetization.

[0063] Voltage fluctuation data represents the degree of fluctuation in the effective voltage value over multiple consecutive cycles during transformer operation, reflecting the stability of the power grid voltage. Because voltage fluctuations cause transient fluctuations in the magnetic flux of the core, disrupting the core's steady-state magnetization, increasing the probability of the core entering saturation, and thus inducing inrush current, voltage fluctuation data is also a key indicator of inrush current risk.

[0064] In a specific implementation, voltage jitter data can be determined as follows: real-time acquisition of voltage data of the target transformer, calculation of the effective voltage value of multiple consecutive grid cycles (e.g., 10 cycles); then, calculation of the rate of change of these effective voltage values ​​(e.g., the difference between the maximum and minimum values ​​divided by the average value), and using this rate of change as voltage jitter data to reflect the influence of grid voltage stability on the transient fluctuation of core flux.

[0065] In actual implementation, the above five types of data can be selected from the various real-time operating status data of the target transformer as the target operating status data. These five-dimensional data together constitute a complete feature space describing the electromagnetic state of the transformer before closing, which is the key factor that determines the amplitude and waveform characteristics of the excitation inrush current.

[0066] It should be noted that the multiple target operating status data includes at least the five types of data mentioned above, but not only these five types. In practical applications, other operating status data that affect the inrush current can be added according to the specific type of transformer and the power grid operating environment, and all of these fall within the protection scope of this application.

[0067] Based on this, the embodiments of this application comprehensively and accurately reflect the correlation between the transformer's operating status and inrush current risk by selecting five types of data, including residual magnetism data and closing angle data. This provides a clear input direction for subsequent digital twin model simulation to generate sample data and for training the inrush current risk identification model, effectively improving the effectiveness of the sample data and the model's identification accuracy. Furthermore, the physical coupling relationship between the five types of data further strengthens the advantages of the physics-data dual-driven paradigm, enabling the model to better capture the physical mechanism of inrush current generation, reducing the risk of model overfitting, and laying a solid foundation for subsequent accurate identification of inrush current risk.

[0068] According to embodiments of this application, simulation of a digital twin model yields various sample data, including: sampling within a preset feature parameter space to generate multiple sets of sampling parameters; simulating the digital twin model using the multiple sets of sampling parameters to obtain the inrush current characteristic quantity corresponding to each set of sampling parameters; and labeling the sample data corresponding to each set of sampling parameters with risk category labels based on the comparison results between the inrush current characteristic quantity and the differential protection threshold, thereby obtaining various sample data.

[0069] The preset characteristic parameter space represents a pre-defined set of value ranges for all characteristic parameters that affect inrush current. In specific implementations, the range of the preset characteristic parameter space can be defined based on the actual operating conditions of the transformer, the power grid environment, and the physical mechanism to ensure coverage of all possible operating conditions at the boundary between high-risk and low-risk conditions, and to avoid missing key operating conditions in the sampling parameters.

[0070] In this embodiment, the multiple sets of sampling parameters include at least five dimensions corresponding to the aforementioned five-dimensional target operating state data: residual magnetism parameter, closing angle parameter, DC bias parameter, harmonic distortion parameter, and voltage jitter parameter.

[0071] In one specific embodiment, to generate a complete sample set covering the boundaries between high-risk and low-risk conditions, a Monte Carlo sampling method can be used to perform large-scale random sampling within the aforementioned five-dimensional preset feature parameter space, generating multiple sets of sampling parameters. Each set of sampling parameters represents a possible operating condition before closing the circuit breaker. These multiple sets of sampling parameters collectively cover the high-risk, low-risk, and boundary conditions within the preset feature parameter space.

[0072] In one example, for the target transformer, residual magnetism The sampling range can be set to Closing angle The sampling range can be set to DC bias The sampling range can be set to 0~50A; the sampling range for harmonic distortion rate (THD) can be set to 0%~20%, and voltage jitter... The sampling range can be set to The selection of the above range is based on a combination of historical operating data of the target transformer and industry-standard recommended extreme operating condition boundaries to ensure coverage of all possibilities from normal operation to deep saturation. Within this range, a total of 100,000 sets of sampling parameters were generated, which can be denoted as... It should be understood that 'i' represents the position, i.e., X. i This can represent the sampling parameters of the i-th group.

[0073] Inrush current characteristic quantities refer to physical quantities obtained through simulation that characterize the properties of inrush current. They are the core basis for judging the risk level of inrush current and include, but are not limited to, inrush current peak value, inrush current duration, and inrush current harmonic content. Among these, the inrush current peak value is the most critical characteristic quantity, directly reflecting the intensity of the inrush current.

[0074] In this embodiment, multiple sets of sampling parameters are input into the digital twin model one by one. For each set of sampling parameters, the digital twin model solves for the core magnetization and magnetic flux variation by simultaneously solving physical equations and voltage balance equations, thereby simulating the generation process of inrush current under this operating condition, and finally outputting the inrush current characteristic quantity corresponding to that set of sampling parameters. The above process is repeated to complete the simulation of all sampling parameters and obtain the inrush current characteristic quantity corresponding to each set of sampling parameters.

[0075] In a preferred embodiment, a digital twin model can be constructed based on the physical equations describing the hysteresis characteristics of the iron core and the voltage balance equations. The physical equations describing the hysteresis characteristics of the iron core can employ the Jiles-Atherton (JA) hysteresis theory model, which can accurately simulate the hysteresis loop of ferromagnetic materials. The differential equations of the JA hysteresis theory model can be expressed in differential form as the magnetization M interacts with the external magnetic field H.

[0076] In this preferred embodiment, for the target transformer, JA parameters, such as saturation magnetization, can be obtained by measuring the silicon steel sheets of its core using the Epstein square and fitting the data. Shape parameter a = 20 A / m, mean field parameter The hysteresis loss coefficient k = 50 A / m, and the reversibility coefficient c = 0.2. Simultaneously, the transformer's structural parameters, such as the effective cross-sectional area of ​​the core, are obtained. The primary winding has N=500 turns, an average magnetic circuit length l=5m, and a primary winding resistance R=0.1Ω.

[0077] In this preferred embodiment, the total magnetization M is decomposed into reversible components. and irreversible components And the hysteresis-free magnetization is described by the Langevin function. The differential form of the JA model and its equation for hysteresis-free magnetization are shown below:

[0078] (1);

[0079] (2);

[0080] in, Indicates hysteresis-free magnetization; The saturation magnetization is Indicates the effective magnetic field. , For shape parameters; Indicates irreversible components. ,in, Indicates the direction coefficient; Indicates reversible components. ,and .

[0081] To facilitate numerical solutions, the above system of equations is usually transformed into a form with time t as the independent variable. Because ,and It needs to be obtained from the circuit equations, so the JA hysteresis theory model can be combined with the voltage balance equations to form a set of differential equations about H and M.

[0082] Specifically, the voltage balance equations on the primary side of the target transformer are established as follows:

[0083] (3);

[0084] In the formula, This represents the power supply voltage, where the power supply voltage... , This refers to the closing angle parameter. U m The voltage amplitude can be determined based on the rated voltage of the target transformer. It is the power frequency angular frequency. R represents the primary winding resistance; N represents the number of turns in the primary winding; It is an iron core magnetic flux. A represents the effective cross-sectional area of ​​the iron core; B(t) represents the magnetic induction intensity. .

[0085] Due to the excitation current With magnetic field strength Related by Ampere's circuital law, such as Where l is the average magnetic path length of the iron core. Meanwhile, the DC bias component is also considered. The actual excitation current is Substituting this into the voltage balance equation, we can obtain a system of first-order differential equations concerning the state variables H(t) and M(t):

[0086] (4);

[0087] (5);

[0088] in, The JA model provides functions related to H and M.

[0089] For each set of sampling parameters X i ,according to , , Constructing the power supply voltage :

[0090] (6);

[0091] in, .

[0092] according to Set initial conditions, take , The fourth-order Runge-Kutta method was used to numerically integrate the above-mentioned first-order differential equation system. The simulation time was... The excitation inrush current waveform i(t) is obtained by solving the equation, and the peak current is extracted from it. For example, finding the maximum value of i(t) over the simulation duration, i.e. This peak value is the inrush current characteristic quantity corresponding to this set of sampling parameters.

[0093] Risk category labels represent identifiers used to label the risk level of sample data. In one example, risk category labels could include "High Risk" and "Safe". These risk category labels are used for supervised learning during model training, enabling the model to learn the correspondence between sampling parameters, excitation inrush current characteristics, and risk categories.

[0094] The differential protection threshold is a preset threshold value of the transformer differential protection device used to determine whether a fault has occurred. Its value can be set according to the transformer's rated parameters and protection requirements. When the inrush current characteristic quantity, such as the peak value, exceeds this threshold, it indicates that the inrush current intensity is large and may damage the transformer, representing a high-risk situation. When the peak value of the inrush current does not exceed this threshold, it indicates that the inrush current intensity is small and will not damage the transformer, representing a safe state.

[0095] Multiple sample data represent an augmented sample dataset that has been labeled with sampling parameters and risk category labels, which can be used to train the initial model and obtain the excitation inrush current risk identification model.

[0096] In this embodiment, the differential protection threshold of the target transformer can be determined based on the transformer's rated parameters and the substation's safe operation procedures. After determining the differential protection threshold, the inrush current characteristic quantity corresponding to each set of sampling parameters is compared with the differential protection threshold. If the inrush current characteristic quantity exceeds the differential protection threshold, the operating condition corresponding to that set of sampling parameters is determined to be high-risk, and the sample data corresponding to that set of sampling parameters is labeled "High Risk". If the inrush current characteristic quantity does not exceed the differential protection threshold, the operating condition corresponding to that set of sampling parameters is determined to be safe, and the sample data corresponding to that set of sampling parameters is labeled "Safe".

[0097] For example, in this embodiment, the differential protection threshold can be set to 20mA. For each set of sampling parameters, if If it is positive, it is labeled as "High Risk" (positive sample). .like If it is a negative sample, it is labeled as "Safe" (negative sample). .

[0098] In this embodiment, after labeling the sample data corresponding to all sampling parameters, the result includes feature vectors. and tags The dataset contains various sample data, which covers the inrush current results caused by various combinations of features such as residual magnetism, closing angle, and DC bias, and is used for training the subsequent initial model.

[0099] Based on this, the embodiments of this application construct a digital twin model by introducing JA hysteresis theory, and combine Monte Carlo sampling and numerical simulation to successfully generate a massive amount of physically enhanced samples covering various boundary conditions. This fundamentally solves the problem of scarce and unevenly distributed measured samples of high-risk excitation inrush current in power systems, provides data assurance for training high-precision, highly generalizable machine learning models, and ensures that the generated data inherently conforms to the physical laws of electromagnetic induction.

[0100] According to an embodiment of this application, an initial model is trained based on multiple sample data corresponding to multiple target operating state data to obtain an excitation inrush current risk identification model, including: performing feature space mapping on multiple sample data to obtain multiple feature data, wherein the multiple feature data includes original feature components and nonlinear coupled feature components; standardizing the multiple feature data to obtain multiple optimized sample data; and using the multiple optimized sample data to train the initial model to obtain the excitation inrush current risk identification model.

[0101] Multiple feature data represent the feature set obtained after feature space mapping, used for model training. It includes original feature components and nonlinear coupled feature components, which comprehensively reflect the feature information and physical coupling relationship of the sample data.

[0102] The original feature components represent the feature components that directly correspond to the operating state data of multiple targets. That is, the features are obtained by directly extracting the sampling parameters in the sample data, and are used to directly reflect the original information of the target operating state data.

[0103] Nonlinear coupling characteristic components represent characteristic components constructed mathematically based on the physical coupling relationship between the original characteristic components. They are used to capture the physical coupling relationship between the original characteristic components, reflect the mechanism by which multiple target operating state data work together on the excitation inrush current, and make up for the insufficiency of the original characteristic components which can only reflect single data information.

[0104] In this embodiment, original sampling parameters, such as residual magnetism parameters, closing angle parameters, DC bias parameters, harmonic distortion parameters, and voltage jitter parameters, are extracted from various sample data and used as original feature components to ensure that the original feature components can completely retain the original information of the target operating state data. Subsequently, based on the physical coupling relationship between the original feature components, such as phase modulation coupling and bias sensitization coupling, nonlinear coupled feature components are constructed through mathematical operations such as multiplication, ratio, and nonlinear transformation to capture the synergistic effect and mutual influence between the original feature components. The original feature components and the nonlinear coupled feature components are then integrated to obtain multiple feature data, providing more comprehensive and targeted input features for subsequent model training.

[0105] In this embodiment, after feature mapping, the resulting new feature data often have different dimensions and numerical ranges. To eliminate the influence of dimensions and make model training more stable and efficient, data standardization is required. In one example, Z-score standardization, Min-Max standardization, and other methods can be used to process various feature data. After standardization, all feature data are mapped to a distribution with a mean of 0 and a variance of 1, thereby obtaining various optimized sample data. This ensures that the data can adapt to the training requirements of the initial model and improve the model training effect.

[0106] In this embodiment, optimized sample data, after feature mapping and standardization, is used as input to train an initial machine learning model. In one example, the initial model can be a Lightweight Gradient Boosting Machine (LightGBM).

[0107] LightGBM is an efficient gradient boosting decision tree framework that uses a depth-constrained leaf-wise growth strategy to build decision trees. Compared with the traditional level-wise growth strategy, it has the advantages of fast training speed, small memory footprint, and high accuracy when dealing with high-dimensional and non-linear data.

[0108] In this embodiment, during the initial model training process, the model aims to minimize the loss function on the validation set and is iteratively optimized. To further improve the model's generalization ability and performance, key parameters of the model can be adjusted before or during training, incorporating hyperparameter optimization. The final trained model is an excitation inrush current risk identification model suitable for online risk identification.

[0109] Based on this, the embodiments of this application introduce nonlinear coupled feature components based on physical knowledge by mapping the original sample data to the feature space. This enables the model to perceive the interaction between features in advance, rather than relying entirely on black-box learning, thereby significantly improving the model's ability to fit complex physical laws and the transparency of the decision-making process.

[0110] According to embodiments of this application, feature space mapping is performed on various sample data to obtain various feature data, including: extracting residual magnetism feature data, closing angle feature data, DC bias feature data, harmonic distortion feature data, and voltage jitter feature data for each sample data to obtain original feature components; based on the residual magnetism feature data and closing angle feature data, a phase modulation coupling feature component characterizing the synergistic effect of the residual magnetism phase and the closing phase is obtained; based on the DC bias feature data and harmonic distortion feature data, a bias sensitization coupling feature component characterizing the amplification effect of the DC bias on harmonics is obtained; based on the DC bias feature data and voltage jitter feature data, a transient fluctuation coupling feature component characterizing the synergistic effect of the DC bias and voltage fluctuation is obtained; and the original feature components, phase modulation coupling feature components, bias sensitization coupling feature components, and transient fluctuation coupling feature components are combined to obtain various feature data.

[0111] In this embodiment, five original feature components are extracted from each type of sample data, namely: remanence feature data. Closing angle characteristic data DC bias characteristic data Harmonic distortion characteristic data (THD) and voltage jitter characteristic data .

[0112] In this embodiment, to enhance the model's perception of physical coupling relationships, it is necessary to construct nonlinear coupling feature components.

[0113] In one specific embodiment, the nonlinear coupling characteristic components include at least a phase modulation coupling characteristic component, a bias sensitization coupling characteristic component, and a transient fluctuation coupling characteristic component.

[0114] Phase modulation coupling refers to the synergistic effect between the residual magnetization phase and the closing phase. When the direction of the residual magnetization is consistent with the voltage phase direction at the closing moment, it leads to magnetic flux superposition, which can easily cause deep core saturation. To quantify this coupling effect, this embodiment constructs the following features: This coupling feature This characterizes the degree of matching between the residual magnetization phase and the closing phase. When the phase difference between the two is close to 0 or... hour, A value close to 1 indicates the strongest coupling; when the phase difference is close to 1, the coupling effect is strongest. hour, A value close to -1 indicates that the coupling effects cancel each other out.

[0115] Bias-sensitized coupling refers to the amplification of the effect of harmonic distortion on core saturation by the presence of DC bias. Without DC bias, a certain harmonic content may not be sufficient to cause deep saturation; however, with DC bias, the same harmonic content will induce more severe saturation. To quantify this amplification effect, this embodiment constructs the following features: The coupling feature C bs The sensitizing effect of DC bias on harmonics is demonstrated through a product form. When the DC bias I... dc When the harmonic distortion (THD) is large, even if the harmonic distortion (THD) is small, the coupling characteristic (C) is large. bs It may also increase significantly.

[0116] Transient fluctuation coupling refers to the synergistic effect between DC bias and voltage jitter. Voltage jitter reflects the transient stability of the system, and in the presence of DC bias, the disturbance effect of voltage jitter on magnetic flux establishment is amplified. To quantify this synergistic effect, this embodiment constructs the following features: This feature C vf The product form reflects the combined effect of DC bias and voltage fluctuation on inrush risk.

[0117] Finally, the five original feature components are combined with the three nonlinearly coupled feature components to form the final, higher-dimensional feature data vector: This vector serves as the input for subsequent model training.

[0118] In another implementation, the feature space can also be constructed based on the generalized flux response equation during the five-dimensional feature vector extraction process. This generalized flux response equation is an approximate analytical expression derived theoretically from the coupled solution results of the aforementioned JA hysteresis model and voltage balance equation. It explicitly reveals the nonlinear coupling mechanism between feature variables.

[0119] Specifically, total magnetic flux Modulated by the five-dimensional features, their approximate relationship is as follows:

[0120] (7);

[0121] In the formula, Residual magnetism, The closing angle, DC bias, Harmonic distortion rate, This is due to voltage fluctuations. Voltage amplitude, Where ω is the power frequency angular frequency, and h is the harmonic order. is a coefficient function of harmonic distortion rate, characterizing the contribution of harmonics to magnetic flux.

[0122] The above total flux equation explicitly reflects two key physical coupling mechanisms leading to deep core saturation: x1 and x2 in the equation are coupled through trigonometric functions, which this application quantifies as phase modulation coupling characteristics. The x3 and x5 terms (which affect the voltage amplitude) in the equation are expressed as a product. Coupling, which this application quantifies as transient fluctuation coupling. and bias sensitization coupling features .

[0123] It should be noted that this claim is limited to including "at least" the above three types of nonlinear coupling characteristic components, not only these three types. Other types of nonlinear coupling characteristic components may be added according to the actual physical coupling relationship, and all of them fall within the protection scope of this application.

[0124] Based on this, the embodiments of this application construct and introduce three nonlinear feature components with clear physical meanings: phase modulation coupling, bias sensitization coupling, and transient fluctuation coupling. This accurately captures the mechanism by which multiple target operating state data interact with the inrush current, overcoming the limitation of the original feature components reflecting only single data information, thus making the feature data more targeted and effective. Simultaneously, the combination of the original and coupled feature components makes the feature data more comprehensive, covering various influencing factors of inrush current generation. This allows the model to directly learn and utilize these key physical laws, rather than relying solely on statistical correlations in the data. This significantly improves the model's physical interpretability and generalization ability under complex operating conditions, ensuring a high degree of consistency between the model's decision logic and the principle of electromagnetic induction.

[0125] According to an embodiment of this application, an initial model is trained based on multiple sample data corresponding to multiple target operating state data to obtain an excitation inrush current risk identification model, including: using a Bayesian optimization algorithm to globally optimize the hyperparameters of the initial model and determine the optimal hyperparameter combination, wherein the hyperparameters include at least one of learning rate, tree depth, number of leaf nodes, feature sampling ratio, and regularization coefficient; based on the optimal hyperparameter combination, the initial model is retrained using multiple sample data to obtain the excitation inrush current risk identification model.

[0126] According to embodiments of this application, the selection of hyperparameters is crucial to the final performance of a machine learning model during training. Traditional grid search or random search methods are inefficient and struggle to guarantee finding the global optimum. Therefore, in this embodiment, a Bayesian optimization algorithm, such as the Tree-structured ParzenEstimator (TPE) algorithm in the Optuna framework, can be introduced to automatically optimize the hyperparameters of the LightGBM model globally.

[0127] In this embodiment, the hyperparameters to be optimized and their search space can be determined based on the model type and the characteristics of the sample data. In one example, the hyperparameters to be optimized include at least one or more of the following: learning rate, tree depth, number of leaf nodes, feature sampling ratio, and regularization coefficient.

[0128] In one specific implementation, the search space, such as the learning rate / step size, can be configured as follows: The maximum number of leaf nodes in a single tree can be configured as [16, 128], and the maximum depth of the tree can be configured as [5, 20].

[0129] In this embodiment, considering the zero-tolerance characteristic of the power system for missed high-risk samples, a weighted cross-entropy loss function can be used as the optimization objective to guide the search process to converge towards a direction that focuses more on high-risk samples. The loss function Loss is shown below:

[0130] (8);

[0131] Among them, y i =1 indicates a high-risk sample, y i =0 indicates a safe sample. Weighting coefficients. Set as the ratio of safe samples to high-risk samples to balance the sample imbalance problem.

[0132] In this embodiment, a probabilistic surrogate model between hyperparameters and model performance can be constructed based on a Bayesian optimization algorithm. Then, based on this probabilistic surrogate model, historical optimization information, such as tested hyperparameter combinations and their corresponding model performance, guides the selection of hyperparameters for the next iteration. Different hyperparameter combinations are iteratively tested, and the model performance corresponding to each combination is calculated, such as the prediction accuracy on the validation set and the F1 score. When the model performance of the hyperparameter combination reaches a preset optimal standard, or the number of iterations reaches a preset threshold, the optimization stops, and the hyperparameter combination at this point is determined to be the optimal hyperparameter combination.

[0133] In one specific implementation, the TPE algorithm can be used. Specifically, the historical hyperparameter evaluation results can be divided into two groups, "excellent" and "poor," and probability density functions l(x) and g(x) can be constructed for them respectively. The probability density functions are shown below:

[0134] (9);

[0135] Then, by maximizing expected improvement The next most promising hyperparameter combination is then selected for evaluation. In this way, the TPE algorithm can quickly locate the vicinity of the global optimum in a multidimensional nonlinear search space with only a few iterations, such as 200. Through iterative searching by the TPE algorithm, a set of hyperparameter combinations that minimizes the weighted loss function on the validation set is finally obtained; this is the optimal hyperparameter combination.

[0136] In this embodiment, after obtaining the optimal combination of hyperparameters, it is applied to the LightGBM model and retrained on the optimized sample data. This training process no longer involves hyperparameter search, but directly uses the optimal parameters to fit all the data, ultimately obtaining a high-performance excitation inrush current risk identification model that can be used for practical deployment, namely the Hyperparameter Optimized LightGBM (HO-LGBM) model.

[0137] It should be noted that the various sample data used in this step can be either the original sample data or optimized sample data after feature space mapping and standardization, neither of which will affect the implementation of this method.

[0138] Based on this, the embodiments of this application employ the TPE Bayesian optimization algorithm to replace traditional manual trial and error or grid search, achieving automated and globally optimal selection of model hyperparameters. This not only significantly improves the efficiency and scientific rigor of model hyperparameter tuning but also ensures that the final HO-LGBM model can achieve optimal generalization performance and identification accuracy when dealing with complex nonlinear coupling features.

[0139] According to an embodiment of this application, after obtaining the HO-LGBM model through the aforementioned training, a dynamic and adaptive decision threshold can also be applied to the model.

[0140] Multiple validation sample data represent validation datasets that are separated from multiple sample datasets and are specifically used to validate model performance and determine dynamic thresholds. They are independent of the training dataset used for model training, thus avoiding interference from training data in the threshold determination process and ensuring the generalizability of the threshold.

[0141] The predicted probability value represents the risk probability value of the inrush current corresponding to the verification sample after the inference of the inrush current risk identification model on the verification sample data. In one example, its value ranges from 0 to 1.

[0142] The precision-recall curve (PR curve) plots the relationship between model precision and recall at different classification thresholds, with recall on the horizontal axis and precision on the vertical axis. It visually reflects the model's classification performance at different thresholds. Precision measures how many samples the model classifies as high-risk are truly high-risk, while recall measures how many of those truly high-risk samples are successfully captured by the model.

[0143] In this embodiment, pre-divided multiple types of validation sample data are input into the trained inrush current risk identification model. The model infers for each validation sample data and outputs the corresponding risk probability value. At the same time, the pre-labeled risk category label of each validation sample data is extracted to ensure that the predicted probability value corresponds one-to-one with the risk category label, providing a complete data pair for subsequent PR curve construction.

[0144] In one specific implementation, each set of feature data in the verification dataset will be... The input is fed into the trained HO-LGBM model to obtain the risk probability value output by the model. At the same time, the true risk category label for each data set can be obtained from the validation set. In this context, 0 represents safe and 1 represents high risk.

[0145] In this embodiment, multiple consecutive risk probability thresholds are set (iterates stepwise from 0 to 1). For each threshold, samples with predicted probabilities greater than or equal to the threshold are classified as high-risk, and samples with probabilities less than the threshold are classified as safe. Then, based on the classification results and the true risk category labels, the precision and recall corresponding to each threshold are calculated. Finally, with recall on the horizontal axis and precision on the vertical axis, the precision-recall points corresponding to all thresholds are connected to form a complete PR curve, clearly demonstrating the model's classification performance under different thresholds.

[0146] In the aforementioned specific implementation, a series of predicted probabilities are obtained on the validation dataset. and risk category labels Then, all possible risk probability thresholds can be iterated, for example, from 0 to 1, with a step size of 0.01. For each candidate threshold... The validation dataset samples can be divided into high-risk predictions ( ) and predictive security ( The system is divided into two categories, and compared with the risk category label to calculate the precision and recall at that threshold. This yields the complete PR curve.

[0147] The safety tolerance parameter is set by the substation operator based on the importance of the transformer, representing the system's maximum tolerance limit for missed alarms. For example, for a critical hub transformer, a certain threshold might be set. This means that the model is required to have a recall rate of no less than 99.9% for high-risk surge events.

[0148] The harmonic mean (F1-Score) represents the harmonic average of precision and recall, comprehensively measuring the balance between these two metrics. A higher F1-Score indicates better model classification performance. Its calculation formula is:

[0149] (10).

[0150] In this embodiment, based on the safety operation procedures of the target transformer, the minimum recall value corresponding to the safety tolerance parameter is determined, such as a recall rate of no less than 99.9%, which serves as a constraint for threshold determination. Subsequently, all threshold intervals with recall rates greater than or equal to this constraint are selected from the PR curve. Within these intervals, the risk probability value that maximizes the F1-Score is found, and this value is determined as the dynamic identification threshold, thereby satisfying the safety tolerance requirements of different substations. This dynamic identification threshold can be fixed together with the trained HO-LGBM model for decision-making during the online inference phase. The mathematical expression for the dynamic identification threshold is as follows:

[0151] (11);

[0152] In the formula, Indicates the dynamic identification threshold. This represents the validation dataset. This indicates the level of safety tolerance.

[0153] Based on this, the embodiments of this application construct PR curves based on verification sample data and set recall rate constraints in conjunction with safety tolerance parameters, strictly ensuring zero or extremely low false negatives for high-risk inrush currents, which meets the core safety requirements of relay protection. Furthermore, by maximizing the F1-Score to determine the threshold, the accuracy of the model is considered while meeting safety constraints, thus improving the overall performance of risk identification.

[0154] This application also proposes a method for training a transformer inrush current risk identification model.

[0155] The following is combined Figure 2 The training method for the transformer inrush current risk identification model described above is explained in detail.

[0156] Figure 2 The flowchart illustrates a method for training a transformer inrush current risk identification model according to an embodiment of this application.

[0157] like Figure 2 As shown, the training method for the transformer inrush current risk identification model includes operations S210 to S230.

[0158] By operating S210, a digital twin model of the target transformer is simulated to obtain various sample data corresponding to various target operating status data.

[0159] In this embodiment, for a specific target transformer, a digital twin model is constructed, comprising physical equations describing the core hysteresis characteristics, such as the JA hysteresis model, and voltage balance equations coupled with the physical equations. Large-scale sampling is performed within a preset characteristic parameter space (including at least residual magnetism, closing angle, DC bias, harmonic distortion, and voltage jitter), and these sampled parameters are used to simulate the digital twin model, calculating the corresponding inrush current characteristics. The inrush current characteristics are compared with differential protection thresholds, and each set of sampled parameters is labeled with a risk category, thereby generating multiple sample data sets containing characteristic data and their corresponding risk labels.

[0160] In this embodiment, the generated sample data are randomly divided into training, validation, and test datasets according to a set ratio. For each sample in the training dataset, nonlinear coupling feature components with clear physical meaning, such as phase modulation coupling features, bias sensitization coupling features, and transient fluctuation coupling features, are constructed based on the original five-dimensional features. The original features are combined with the constructed features to obtain an enhanced feature vector. The mean and standard deviation of each feature on the training set are calculated, and Z-score standardization is performed on all features in the training, validation, and test sets to obtain various optimized sample data.

[0161] In operating S220, the Bayesian optimization algorithm is used to globally optimize the hyperparameters of the initial model and determine the optimal combination of hyperparameters.

[0162] In this embodiment, an initial model (such as LightGBM) is selected, and the hyperparameters to be optimized and their search range are defined, such as learning rate, tree depth, number of leaf nodes, feature sampling ratio, and regularization coefficient. Then, the TPE Bayesian optimization algorithm under the Optuna framework is used to perform an iterative search with the goal of minimizing the loss function on the validation set, and finally determine the optimal combination of hyperparameters for the model.

[0163] Specifically, considering the imbalance of sample classes, with high-risk samples typically far fewer than safe samples, weighted cross-entropy loss is used as the optimization objective. The TPE algorithm within the Optuna framework is employed for hyperparameter optimization, which involves: randomly sampling several hyperparameter combinations and evaluating their loss on the validation dataset. Iterative optimization is then performed, for example, with a total of 200 iterations. The iterative process includes: sorting the historical hyperparameter combinations by their loss values, selecting the top 15% as the "high-quality group" and the rest as the "low-quality group." Probability density functions are constructed for both the high-quality and low-quality groups. The next hyperparameter combination to be evaluated is selected by maximizing the expected value improvement. The model is then trained, the validation set loss is calculated, and historical observations are updated. After the iterations, the hyperparameter combination that minimizes the validation set loss is selected as the optimal hyperparameter combination.

[0164] In operating S230, based on the optimal hyperparameter combination, the initial model is trained using multiple sample data to obtain the excitation inrush current risk identification model.

[0165] In this embodiment, after obtaining the optimal hyperparameter combination, it is applied to an initial model such as the LightGBM model, and the model is retrained using the entire training dataset. The model employs a depth-constrained leaf-wise growth strategy to construct a decision tree, aiming to minimize the weighted cross-entropy loss, and iteratively generates an ensemble model. After training, the final HO-LGBM inrush current risk identification model is obtained.

[0166] In this embodiment, the trained HO-LGBM model is evaluated using a test dataset, and metrics such as accuracy, precision, recall, and F1-score are calculated. Simultaneously, the Shapley Additive Explanations (SHAP) method is employed to analyze feature importance, verifying whether the model successfully captures physical laws such as remanence dominance and phase modulation coupling, ensuring that the model's decision-making logic is consistent with the principles of electromagnetic induction.

[0167] Based on this, the embodiments of this application use a digital twin model constructed with JA hysteresis theory as the foundation, and generate physically enhanced samples covering boundary conditions through Monte Carlo sampling, fundamentally solving the problem of scarce real fault samples. By introducing the TPE Bayesian optimization algorithm, automated global optimization of model hyperparameters is achieved, avoiding the blindness and local optima problems of manual parameter tuning. Combining feature space mapping and dynamic thresholding strategies, the finally trained HO-LGBM model not only has high accuracy and high generalization ability, but its decision logic also has clear physical interpretability and can adapt to the asymmetric risk control requirements of different substations. This training method is universal and can be quickly deployed for any target transformer, providing it with accurate inrush current risk identification capabilities.

[0168] This application also proposes a transformer inrush current risk identification device based on physical-data dual-drive.

[0169] Figure 3 A block diagram of a transformer inrush current risk identification device according to an embodiment of this application is shown schematically.

[0170] like Figure 3 As shown, the transformer inrush current risk identification device 300 includes a first determining module 310, a second determining module 320, and a third determining module 330.

[0171] The first determining module 310 is used to determine multiple target operating status data from multiple real-time operating status data of the target transformer.

[0172] The second determining module 320 is used to process various target operating status data based on the excitation inrush current risk identification model to obtain the risk probability value corresponding to the excitation inrush current.

[0173] The third determination module 330 determines the risk identification result for inrush current based on the risk probability value and the dynamic identification threshold corresponding to the safety tolerance requirements of the target transformer.

[0174] Figure 4 A block diagram of a transformer inrush current risk identification model training device according to an embodiment of this application is shown schematically.

[0175] like Figure 4 As shown, the transformer inrush current risk identification model training device 400 includes a data generation module 410, a global optimization module 420, and a training module 430.

[0176] The data generation module 410 is used to simulate the digital twin model of the target transformer to obtain various sample data corresponding to various target operating status data.

[0177] The global optimization module 420 is used to perform global optimization of the hyperparameters of the initial model using the Bayesian optimization algorithm to determine the optimal combination of hyperparameters.

[0178] Training module 430 is used to train the initial model based on the optimal hyperparameter combination and using multiple sample data to obtain the excitation inrush current risk identification model.

[0179] Any one or more of the modules, submodules, units, and subunits according to the embodiments of this application, or at least part of the functions of any one or more of them, can be implemented in one module. Any one or more of the modules, submodules, units, and subunits according to the embodiments of this application can be implemented by dividing them into multiple modules. Any one or more of the modules, submodules, units, and subunits according to the embodiments of this application can be at least partially implemented as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or implemented by hardware or firmware in any other reasonable manner by integrating or packaging circuits, or implemented in any one of software, hardware, and firmware, or in a suitable combination of any of these. Alternatively, one or more of the modules, submodules, units, and subunits according to the embodiments of this application can be at least partially implemented as computer program modules, which, when run, can perform corresponding functions.

[0180] For example, any plurality of the first determining module 310, the second determining module 320, and the third determining module 330 can be combined into one module / unit / subunit, or any one of these modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least part of the functionality of one or more of these modules / units / subunits can be combined with at least part of the functionality of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of this application, at least one of the first determining module 310, the second determining module 320, and the third determining module 330 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the first determining module 310, the second determining module 320, and the third determining module 330 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0181] It should be noted that the transformer inrush current risk identification device part in the embodiments of this application corresponds to the transformer inrush current risk identification method in the embodiments of this application, and the transformer inrush current risk identification model training device part corresponds to the transformer inrush current risk identification model training method in the embodiments of this application. For a detailed description of the transformer inrush current risk identification device and the transformer inrush current risk identification model training device part, please refer to the transformer inrush current risk identification method and the transformer inrush current risk identification model training method, which will not be repeated here.

[0182] Figure 5 A block diagram of an electronic device suitable for implementing a transformer inrush current risk identification method according to an embodiment of this application is shown schematically. Figure 5 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0183] like Figure 5 As shown, an electronic device according to an embodiment of this application includes a processor 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory ROM 502 or a program loaded from a storage portion 508 into a random access memory RAM 503. The processor 501 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 501 may also include onboard memory for caching purposes. The processor 501 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0184] RAM 503 stores various programs and data required for the operation of the electronic device. Processor 501, ROM 502, and RAM 503 are interconnected via bus 504. Processor 501 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 502 and / or RAM 503. It should be noted that the programs may also be stored in one or more memories other than ROM 502 and RAM 503. Processor 501 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0185] According to embodiments of this application, the electronic device may further include an input / output (I / O) interface 505, which is also connected to a bus 504. The electronic device may also include one or more of the following components connected to the input / output (I / O) interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the input / output (I / O) interface 505 as needed. A removable medium 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 510 as needed so that computer programs read from it can be installed into the storage section 508 as needed.

[0186] According to embodiments of this application, the method flow according to embodiments of this application can be implemented as a computer software program. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the computer program is executed by processor 501, it performs the functions defined in the system of embodiments of this application. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0187] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0188] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0189] For example, according to embodiments of this application, a computer-readable storage medium may include the ROM 502 and / or RAM 503 described above and / or one or more memories other than ROM 502 and RAM 503.

[0190] Embodiments of this application also include a computer program product, which includes a computer program containing program code for performing the methods provided in the embodiments of this application. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the transformer inrush current risk identification method provided in the embodiments of this application.

[0191] When the computer program is executed by the processor 501, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0192] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 509, and / or installed from a removable medium 511. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0193] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0194] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations are not explicitly described in this application. In particular, without departing from the spirit and teachings of this application, the features described in the various embodiments of this application can be combined and / or combined in various ways. All such combinations and / or combinations fall within the scope of this application.

[0195] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A method for identifying transformer inrush current risk, wherein, The method includes: From various real-time operating status data of the target transformer, various target operating status data are determined. These various target operating status data have a stronger influence on the inrush current than other types of operating status data in the various real-time operating status data. Furthermore, these various target operating status data have a physical coupling relationship with each other. The inrush current risk identification model is used to process the various target operating state data to obtain the risk probability value corresponding to the inrush current. The inrush current risk identification model is obtained by training an initial model based on various sample data corresponding to the various target operating state data. The various sample data are obtained by simulating a digital twin model. The digital twin model includes physical equations describing the hysteresis characteristics of the iron core and voltage balance equations. Based on the risk probability value and the dynamic identification threshold corresponding to the safety tolerance requirement of the target transformer, the risk identification result for the inrush current is determined.

2. The method according to claim 1, wherein, The various target operating status data include at least residual magnetism data, closing angle data, DC bias data, harmonic distortion data, and voltage jitter data.

3. The method according to claim 1, wherein, The various sample data obtained by simulating the digital twin model include: Sampling is performed within a preset feature parameter space to generate multiple sets of sampling parameters. The multiple sets of sampling parameters include at least residual magnetism parameters, closing angle parameters, DC bias parameters, harmonic distortion parameters, and voltage jitter parameters. The digital twin model is simulated using the multiple sets of sampling parameters to obtain the excitation inrush current characteristic quantity corresponding to each set of sampling parameters; Based on the comparison results between the inrush characteristic quantity and the differential protection threshold, risk category labels are marked for the sample data corresponding to each group of sampling parameters to obtain the various sample data.

4. The method according to claim 1, wherein, The initial model is trained based on multiple sample data corresponding to the various target operating state data to obtain the excitation inrush current risk identification model, including: The various sample data are respectively mapped to feature space to obtain various feature data, wherein the various feature data include original feature components and nonlinear coupled feature components. The original feature components correspond to the various target running state data, and the nonlinear coupled feature components are constructed based on the physical coupling relationship between the original feature components. The various feature data are standardized to obtain various optimized sample data; The initial model is trained using the various optimized sample data to obtain the excitation inrush current risk identification model.

5. The method according to claim 4, wherein, The nonlinear coupling characteristic components include at least phase modulation coupling characteristic components, bias sensitization coupling characteristic components, and transient fluctuation coupling characteristic components; The various sample data are mapped into feature spaces to obtain various feature data, including: Extract the residual magnetism characteristic data, closing angle characteristic data, DC bias characteristic data, harmonic distortion characteristic data and voltage jitter characteristic data of each of the sample data to obtain the original characteristic components; Based on the residual magnetism characteristic data and the closing angle characteristic data, a phase modulation coupling characteristic component characterizing the synergistic effect of the residual magnetism phase and the closing phase is obtained; Based on the DC bias characteristic data and the harmonic distortion characteristic data, a bias-sensitized coupling characteristic component that characterizes the amplification effect of DC bias on harmonics is obtained. Based on the DC bias characteristic data and the voltage jitter characteristic data, a transient fluctuation coupling characteristic component characterizing the synergistic effect of DC bias and voltage fluctuation is obtained. The original feature components, the phase modulation coupling feature components, the bias sensitization coupling feature components, and the transient fluctuation coupling feature components are combined to obtain the various feature data.

6. The method according to claim 1, wherein, The initial model is trained based on multiple sample data corresponding to the various target operating state data to obtain the excitation inrush current risk identification model, including: The Bayesian optimization algorithm is used to globally optimize the hyperparameters of the initial model to determine the optimal combination of hyperparameters. The hyperparameters include at least one of the following: learning rate, tree depth, number of leaf nodes, feature sampling ratio, and regularization coefficient. Based on the optimal combination of hyperparameters, the initial model is trained using the various sample data to obtain the excitation inrush current risk identification model.

7. The method according to claim 6, wherein, The various sample data also include various verification sample data; The dynamic identification threshold is determined by the following method: Obtain the predicted probability values ​​of the excitation inrush current risk identification model for the various verification sample data, and the risk category label corresponding to each of the verification sample data; Based on the predicted probability value and the risk category label, a precision-recall curve is constructed; Using the safety tolerance parameter of the target transformer as the recall constraint, and utilizing the precision-recall curve, the risk probability value that maximizes the harmonic average of precision and recall is determined from multiple risk probability values ​​as the dynamic identification threshold.

8. The method according to claim 2, wherein, From various real-time operating status data of the target transformer, determine various target operating status data, including: The tripping angle is determined based on the historical tripping operation records of the target transformer, and the residual magnetism data is obtained by inversion estimation using the physical equations describing the hysteresis characteristics of the iron core. The closing angle data is determined based on the zero-crossing detection of the grid voltage. The DC component is extracted from the current data collected by the current transformer to obtain the DC bias data; The total harmonic distortion rate is calculated based on the voltage waveform or current waveform to obtain the harmonic distortion data. The voltage jitter data is obtained by calculating the rate of change based on the effective voltage value of multiple consecutive cycles.

9. A method for training a transformer inrush current risk identification model for use in any one of claims 1 to 8, wherein, include: The digital twin model of the target transformer is simulated to obtain various sample data corresponding to various target operating state data. The digital twin model includes physical equations describing the core hysteresis characteristics and voltage balance equations coupled with the physical equations. Each sample data includes characteristic data affecting the inrush current and its corresponding risk category label. The Bayesian optimization algorithm is used to globally optimize the hyperparameters of the initial model and determine the optimal combination of hyperparameters. Based on the optimal combination of hyperparameters, the initial model is trained using the various sample data to obtain the excitation inrush current risk identification model.

10. An electronic device, comprising: One or more processors; Storage device for storing one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 8 or the method according to claim 9.

Images