Fault feature and adaptability analysis method of transformer protection under new energy output scenario

Abstract

The application discloses a transformer protection fault feature and adaptability analysis method under a new energy output scene. Multi-source fault signals of a transformer under a large-scale new energy sending-out scene are acquired, a graph model with the multi-source fault signals as nodes is constructed, and spectral filtering is performed to obtain a graph signal feature matrix; mask operation is performed based on the similarity between fault samples in the graph signal feature matrix; the graph signal feature matrix and the mask feature are constructed into unified representation for mapping of a fault degree; the mean value of the fault degrees of all samples is used to obtain a protection action threshold value; a protection action request value is calculated, the threshold value and the request value are compared, and a transformer protection decision is output; a confidence index is defined by loss function quantization value, which is used for analyzing the adaptability of a protection mechanism based on the protection action threshold value, and the threshold value is optimized according to the adaptability analysis result. The application provides multi-dimensional and mutual verification technical support for significant evaluation of fault features and protection adaptability analysis.

Description

Technical Field

[0001] This invention belongs to the field of power system automation technology, specifically relating to a method for analyzing the characteristics and adaptability of transformer protection faults in new energy output scenarios. Background Technology

[0002] With the large-scale integration of renewable energy sources, the transient characteristics of transformer faults change due to the rapid response of the power control system. Fault currents exhibit complex characteristics such as high harmonic content and suppressed power frequency components, leading to a decrease in the adaptability of traditional protection methods. In-depth research into the impact of large-scale renewable energy integration on transformer operating characteristics, exploring the development process of fault characteristics under different fault conditions, and evaluating the adaptability of existing transformer protection configurations in new scenarios have become critical issues that urgently need to be addressed to ensure power grid security.

[0003] In recent years, data-driven technologies, represented by deep learning, have provided new approaches to state perception of complex systems. Some studies have attempted to learn fault characteristics and fluctuation patterns in historical data through neural networks. However, these methods often rely on manual feature engineering, making it difficult to fully mine the deep spatiotemporal correlation information contained in high-dimensional fault data; at the same time, the model training process has a slow convergence speed and is prone to getting trapped in local optima, limiting its application in real-time protection. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for analyzing the fault characteristics and adaptability of transformer protection. This method constructs a graphical model of multi-source fault signals in the transformer, extracts graphical signal features, and performs similarity analysis to map fault severity and make protection decisions. By comparing the protection action threshold with the requested value and using confidence indices to analyze the adaptability of the protection mechanism, the threshold is ultimately optimized to improve protection effectiveness.

[0005] The first aspect of this application discloses a fault characteristic and adaptive analysis method for transformer protection, adopting the following technical solution: To acquire multi-source fault signals of transformers in large-scale new energy transmission scenarios, construct an correlation graph model with multi-source fault signals as nodes, and perform spectral filtering analysis to obtain the graph signal feature matrix; A spectral domain signal evaluation network is constructed to perform nonlinear measurement on the graph signal feature matrix, calculate the dependency strength between nodes, and output an enhanced feature matrix. The contribution of the enhanced features to the transformer health status analysis results is evaluated, and the uncertainty of the feature combination on which the analysis results depend is quantified by a loss function. The similarity between transformer fault samples in the graph signal feature matrix is ​​calculated, and the fault features are masked according to the similarity to obtain masked features; a unified representation of the changing features of different fault samples is constructed based on the graph signal feature matrix and the masked features. A transformer fault severity assessment network is constructed to map the fault severity of the unified representation; the mean fault severity of all samples is calculated to obtain the transformer protection action threshold; the transformer fault severity assessment network is invoked to calculate the protection action request value of the transformer state based on the enhanced feature matrix; the threshold value is compared with the request value to output the transformer protection decision. A confidence index is defined by quantifying the loss function value, which is used to analyze the adaptability of the transformer protection mechanism based on the transformer protection action threshold; the transformer protection action threshold is optimized based on the adaptability analysis results.

[0006] Furthermore, the transformer multi-source fault signal is a characteristic quantity of the transformer's operating data during a fault, including current, voltage, harmonics, saturation magnetic flux density, permeability, and frequency; Each feature quantity serves as a node in the graph structure, forming a set of nodes in the relational graph model; the relational graph model is an undirected graph, and there are undirected edges between each pair of feature quantities.

[0007] The process of obtaining the feature matrix of the graph signal includes: A learnable spectral filter is invoked to extract structured information from the correlation graph model; The relational graph model includes a set of nodes, a set of edges, and a set of edge weights; the edge weight is the correlation strength between every two nodes in the graph structure, represented by the correlation between the nodes. Furthermore, the spectral domain signal evaluation network includes a feature mapping layer, a dependency evaluation layer, and an output layer; The feature mapping layer is used to extract the nonlinear coupling relationship between node features; the query vector and key vector of transformer fault features are calculated through nonlinear functions and trainable parameter matrices. The dependency evaluation layer is used to calculate an attention weight matrix based on the query vector and the key vector. Each element in the attention weight matrix... Quantized feature nodes right The strength of dependence; The output layer is used to fuse the graph signal feature matrix with the attention weight matrix to output an enhanced feature matrix.

[0008] Furthermore, the degree to which the evaluation enhancement features contribute to the transformer adaptability analysis results includes: The feature mapping of the transformer fault state is extracted from the enhanced feature matrix and fused with the feature weight vector and observation noise to form the transformer health status evaluation function; the transformer health status evaluation function is used to generate a health status evaluation value representing the transformer fault state. Define a Bayesian probability model for the health status evaluation value and introduce prior knowledge into the feature weight vector; when a new transformer state and health status evaluation value are obtained, update the feature weight vector according to the Bayesian probability model and obtain its posterior distribution, and generate a posterior mean vector. The posterior mean vector quantifies the contribution of the corresponding enhanced features to the transformer health status decision.

[0009] Furthermore, the assessment of the contribution of the enhancement features to the transformer adaptability analysis results also includes: The loss function is constructed by combining the feature map with its own transpose and posterior uncertainty matrix; The loss value output by the loss function evaluates and quantifies the cognitive risk brought about by the combination of features on which the current decision is based in real time.

[0010] Furthermore, the masking operation includes the following steps: Calculate the similarity distance metric between the fault samples, sort them in descending order of similarity, and assign corresponding labels to all samples; The labels are linearly weighted by weighting coefficients, and the normalized result is the output mask rate. The feature matrix of the graph signal is then randomly masked based on the mask rate to obtain the mask features.

[0011] Furthermore, the construction process of the unified representation includes: The image signal feature matrix and the mask feature are concatenated and then multiplied with a learnable weight to obtain a composite feature; The contributions of the mask feature and the graph signal feature matrix to the composite feature are calculated respectively. Based on their respective contributions, the mask features and the graph signal feature matrix are weighted to obtain the unified representation.

[0012] Furthermore, the calculation process for the contribution includes: The contribution of the graph signal feature matrix to the composite feature , calculated using exponential function normalization; The contribution of the mask feature to the composite feature Subtracting from 1 calculate.

[0013] Furthermore, the mapping of the fault degree includes: The unified characterization is input into the transformer fault severity assessment network, which outputs the fault severity. The inner layer of the transformer fault severity assessment network is a nonlinear activation function, and the outer layer is a non-negative activation function.

[0014] Furthermore, by comparing the threshold with the requested value, a transformer protection decision is output: When the requested value is greater than or equal to the threshold, the transformer protection operates; when the requested value is less than the threshold, the transformer protection does not operate.

[0015] Furthermore, the adaptability of the transformer protection mechanism is analyzed, including: A confidence index is defined based on the loss value output by the loss function, which is used to represent the adaptability of the transformer protection mechanism; the value of the confidence index is proportional to the confidence level of the transformer protection decision. When the confidence index value is less than the set standard, the optimal transformer protection action threshold is determined by iteratively optimizing the spectral domain signal evaluation network and the acquisition and mapping of the unified representation with the goal of minimizing the loss value.

[0016] The second aspect of this application discloses a transformer protection fault characteristic and adaptability analysis system, which implements the analysis method described in the first aspect of this application. The system includes: The graph model construction module is used to acquire multi-source fault signals of transformers in large-scale new energy transmission scenarios, construct an association graph model with multi-source fault signals as nodes, and perform spectral filtering analysis to obtain the graph signal feature matrix. The signal evaluation module is used to construct a spectral domain signal evaluation network, perform nonlinear measurement on the graph signal feature matrix, calculate the dependency strength between nodes, and output an enhanced feature matrix; evaluate the contribution of enhanced features to the transformer health status analysis results, and quantify the uncertainty of the feature combination on which the analysis results depend through a loss function; A unified representation processing module is used to calculate the similarity between transformer fault samples in the graph signal feature matrix, perform a masking operation on the fault features based on the similarity to obtain masked features, and construct a unified representation of the variation features of different fault samples based on the graph signal feature matrix and the masked features. The protection decision output module is used to construct a transformer fault severity assessment network, map the fault severity of the unified representation, calculate the mean fault severity of all samples to obtain the transformer protection action threshold, call the transformer fault severity assessment network, calculate the protection action request value of the transformer state based on the enhanced feature matrix, compare the threshold with the request value, and output the transformer protection decision. An adaptive analysis and optimization module is used to define confidence indices by quantifying values ​​through a loss function, and to analyze the adaptability of the transformer protection mechanism based on the transformer protection action threshold; and to optimize the transformer protection action threshold based on the adaptive analysis results.

[0017] The beneficial effects of this invention are that, compared with the prior art, 1. By constructing a correlation graph model with multi-source fault signals as nodes and implementing spectral filtering based on key frequency bands with saturation characteristics, it is possible to adaptively extract and enhance deep time-frequency domain features related to transformer faults, thereby effectively overcoming the problem of weak and nonlinear enhancement of fault features caused by the access of new energy sources, and providing more robust feature inputs for protection decisions.

[0018] 2. The constructed spectral domain signal evaluation network provides a loss function to guide the multi-timescale stochastic policy network in making transformer protection action decisions and adaptive analysis. It deeply integrates the transformer operation status analysis process with the protection action decision strategy, improving the adaptability and safety of transformers in dealing with complex faults under large-scale new energy output scenarios.

[0019] 3. Compared with existing transformer protection analysis methods, this invention combines graph domain reinforcement learning to achieve collaborative perception from graph domain correlation analysis to time domain multi-scale decision-making. This provides multi-dimensional and mutually verified technical support for the saliency assessment of fault characteristics and protection adaptability analysis, significantly improving the comprehensiveness and reliability of the analysis. Attached Figure Description

[0020] Figure 1 A flowchart of a transformer protection fault feature and adaptive analysis method based on temporal domain reinforcement learning in a large-scale new energy transmission scenario provided for an example; Figure 2 A logic block diagram of the transformer protection adaptability analysis method provided in the embodiments; Figure 3 A block diagram of a transformer structure provided for an embodiment. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0022] The causes of failures in "high-voltage and high-efficiency" power systems are difficult to analyze due to the unique coupling interactions and nonlinearities. This is especially true as the global energy structure transformation continues, with the widespread integration of large-scale renewable energy sources such as wind and solar power into the power system. This change presents new challenges to the safe and stable operation of critical power system equipment, particularly transformers. The randomness and intermittency of renewable energy generation, along with the extensive application of power electronic equipment, significantly alter the fault characteristics and transient processes of the power grid, making traditional transformer protection principles based on the power frequency sinusoidal assumption inadequately adaptable. The complexity of the system structure and the highly nonlinear nature of distributed energy sources, particularly the coupling interactions between different types of power sources, make the impact path on the system's dynamic response after fault disturbances uncertain, necessitating analysis of the resulting fault characteristics.

[0023] Currently, research and practice in transformer protection mainly rely on physical model-based analysis methods and data-driven methods based on fixed thresholds. These methods are typically built upon equivalent models of typical operating conditions such as symmetrical short circuits or core saturation, and protection criteria are set by studying the changes in conventional operating characteristics such as transformer saturation characteristics and differential current levels under large-scale renewable energy transmission scenarios. However, the integration of renewable energy leads to rich harmonic content in the system, weak fault characteristics, and complex evolution (covering the entire process of generation, development, and stabilization). This makes it difficult for traditional methods to accurately capture the essential characteristics of faults and effectively distinguish complex operating conditions such as internal faults, inrush currents, and external faults accompanied by saturation. Consequently, existing protection configurations are at risk of false tripping or failure to trip. The root cause of this limitation lies in the insufficient ability to extract and correlate multi-source, nonlinear, and time-varying fault information.

[0024] In recent years, data-driven technologies, represented by deep learning, have provided new approaches to state perception of complex systems. Some studies have attempted to learn fault characteristics and fluctuation patterns from historical data through neural networks. However, these methods often rely on manual feature engineering, making it difficult to fully mine the deep spatiotemporal correlation information contained in high-dimensional fault data. Furthermore, the model training process has slow convergence speed and is prone to getting trapped in local optima, limiting its application in real-time protection. Against this backdrop, this patent proposes an analysis method based on time-spectrum domain reinforcement learning. By constructing a correlation graph model and integrating graph signal processing and Bayesian inference, it can adaptively extract and enhance the time-frequency domain features of multi-source fault signals. It also utilizes a multi-timescale random policy network to dynamically evaluate the saliency of fault features, thus providing a more accurate and adaptive solution for analyzing the adaptability of existing transformer protection configurations in large-scale new energy access scenarios and locating the causes of their limitations. This achieves the goal of characterizing the threshold conditions of protection actions based on the features of the fault process using protection principles and artificial intelligence methods.

[0025] As an embodiment of this application, a specific implementation method for transformer protection fault feature and adaptive analysis based on time-spectrum domain reinforcement learning in large-scale new energy transmission scenarios is disclosed. The execution flow of the method embodiment is as follows: Figure 1 Logic block diagram reference Figure 2 .

[0026] The historical operation data for this implementation method comes from, for example, Figure 3 The transformer structure shown includes a 70km line, which includes one centralized new energy power source and one centralized traditional power source. The transformer adopts a Y / Y connection, with a voltage rating of 220 / 110kV and a rated capacity of 280MVA.

[0027] S1: As one implementation method of the embodiment, step 1: acquire the multi-source fault signals of the transformer under large-scale new energy transmission, including current, voltage, harmonics, saturation flux density, permeability and frequency; construct the correlation graph model with fault feature quantities as nodes; perform spectral filtering analysis based on graph Fourier transform to obtain the graph signal feature matrix.

[0028] Step 1.1: Collect transformer operating data during a fault and construct a multi-source signal. Construct a node set based on the elements in the multi-source signal. The multi-source signal is represented as follows: including current ,Voltage ,harmonic saturation magnetic flux density magnetic permeability and frequency It is necessary to clarify that the constructed sample data is the transformer's operating data during a fault, meaning it will trigger transformer protection actions. Each feature quantity serves as a node in the graph structure, thus forming a set of nodes. As an example of this embodiment, 3200 sets of transformer operation data were collected during faults, and each set of data contained the above 6 characteristic quantities.

[0029] In a further implementation, when initially constructing the graph structure, it is assumed that there are undirected edges between the six feature quantities. 3200 sets of runtime data were collected, categorized according to time windows. The 3200 sets of data are divided into 32 time windows, with 100 data points (i.e., 100 sets of data) in each time window. This creates K=32 undirected graph structure samples, with each node of the graph structure corresponding to 100 sets of running data.

[0030] Step 1.2: Analyze the multi-source signals of the transformer based on characteristic nodes. Specifically, calculate the correlation strength between every two characteristic nodes as the edge weight. Characteristic standardization eliminates differences in the dimensions and amplitudes of different characteristic signals. In this embodiment, the Pearson correlation coefficient is used to calculate the edge weights. For the node set... Any two nodes and The edge weight between the two The calculation formula is expressed as: ; in It is a feature node At any moment The characteristic values, It is a feature node At any moment The mean of the feature quantity within the time window. It is the length of the time window; Used to refer to feature nodes respectively and In this embodiment, the time window length can be set to 100.

[0031] Step 1.3: Set the feature nodes Combined with the edge set and edge weights, they form an adjacency matrix. A graphical model for characterizing the correlation of multi-source fault signals in transformers can be obtained. .

[0032] Furthermore, to extract harmonic features directly related to the transformer's saturation state, this application utilizes a learnable spectral filter to extract structured key information from the correlation graph model; the filter output is a graph signal feature matrix. The number of nodes is the same as the input graph, and the feature information on the nodes is updated. For example, this filtering process can be implemented using ChebyNet and ChebyNetII.

[0033] S2: As one implementation of the example, a spectral domain signal evaluation network for multi-source Bayesian inference is constructed, and the graph signal feature matrix generated in step 1 is nonlinearly measured to obtain the dependency strength of the nodes, thereby obtaining the enhanced feature matrix; Bayesian inference is constructed to obtain the immediate reward function. Step 2.1: The spectral domain signal evaluation network consists of a feature mapping layer, a dependency evaluation layer, and an output layer; The feature mapping layer is used to extract the nonlinear coupling relationship between node features; the query vector and key vector of transformer fault features are calculated using a nonlinear function and a trainable parameter matrix; the dependency evaluation layer is used to calculate the attention weight matrix based on the query vector and key vector, where each element in the attention weight matrix... Quantized feature nodes right The dependence strength; the output layer is used to fuse the graph signal feature matrix with the attention weight matrix to output the enhanced feature matrix.

[0034] In one embodiment, the graph signal feature matrix output by S1 Mapping to the Reproducing Kernel Hilbert Spaces (RKHS) facilitates the extraction of nonlinear coupling relationships between features of multi-source monitoring signals from the transformer. Specifically: Through the nonlinear Sigmoid function and the trainable parameter matrix , , Calculate the query vector representing the current state to be evaluated in the transformer fault characterization process. Key vectors representing fault feature templates and value vectors containing detailed feature information ,include: ; ; ; In the formula, It is a non-linear sigmoid mapping function.

[0035] Furthermore, the attention weight matrix is ​​calculated using a Gaussian kernel function. , where each element Quantized feature nodes right The strength of the dependency. For example, it can be calculated using a standard Gaussian kernel followed by Softmax normalization, expressed as: ; in and yes and The row vector, It is the kernel function bandwidth parameter. Characterizes the features under the current fault state. To what extent are the occurrences or changes influenced by characteristics? The impact. The higher the value, the stronger the coupling between the two features in the current fault.

[0036] Output the characteristic matrix of the enhanced transformer , This indicates that the enhanced features retain the structured information from the filtered version. It also incorporates aggregated information from the original node information. .

[0037] In this application, the filtering features focus on extracting global features, but the coupling relationships between individual features are not effectively represented. Therefore, this step maps the spectral-filtered features to the RKHS to address the nonlinear coupling problem between features of multi-source monitoring signals from transformers. The RKHS space facilitates refined fault analysis. In the original feature space, features of different fault modes are tightly coupled. By mapping to the high-dimensional RKHS using kernel functions, complex nonlinear relationships can be transformed into approximately linearly separable expressions, thus laying the foundation for measuring the strength of nonlinear dependencies between feature nodes.

[0038] Step 2.2: Construct a transformer health status evaluation function based on a Bayesian probability model to quantify the contribution of each enhanced feature to the final decision, and calculate the uncertainty of the contribution evaluation by the Bayesian model, so as to provide a basis for the final decision on the transformer adaptability analysis results.

[0039] From a convolutional network Extract higher-level feature maps Here, This represents the current state of the transformer, i.e., the fault state. Further, a transformer health status evaluation function is constructed: ; Indicates the current state of the transformer Health status evaluation value; This is a feature weight vector, where the magnitude of each element reflects the corresponding feature pair. The degree of contribution; To observe the noise, it is assumed to follow a mean of 0 and an accuracy of . The Gaussian distribution, i.e. .

[0040] Furthermore, we define a Bayesian probabilistic model: given a feature weight vector and transformer status Below, observed values Follows a Gaussian distribution: ; in, To measure the accuracy of observation noise, used to evaluate the observation data. The reliability, The larger the value, the more reliable the observation (the smaller the noise variance); It is an identity matrix.

[0041] Furthermore, for the unknown weight vector Introduce Gaussian priors to incorporate prior knowledge (such as manufacturer technical specifications and historical experience): ; in, Let be the prior mean vector. Let be the prior covariance matrix.

[0042] When new observation data is acquired Then, the weight vector is updated according to the Bayesian probability model, and its posterior distribution is obtained: ; Feature mapping for the real-time state of the transformer; and Let be the prior and posterior covariance matrices. This update expression means that the reduction in posterior uncertainty comes from the superposition of information provided by prior information and real-time observation data. The more data available and the more reliable it is, the lower the uncertainty.

[0043] The posterior mean vector is updated to posterior mean vector Each element directly quantifies the contribution of the corresponding enhancement feature to the transformer health status decision (i.e., the fault diagnosis result output based on the transformer's real-time status and health status evaluation value). The element with the largest absolute value corresponds to the feature that affects the final evaluation. The impact is significant. Step 2.3: Construct an analytical reward loss function for fault characteristics based on the evaluation results of Step 2.2.

[0044] Based on features Its own transpose and uncertainty matrix Combined, construct the loss function function value The cognitive risk arising from the feature combination upon which the current decision relies was assessed and quantified in real time; that is, the uncertainty of several high-weighted mapping features upon which the real-time fault diagnosis output depends. If the system over-relies on features with high uncertainty during analysis, then… This increases. As part of the optimization loss, it guides the decision network in step 3 to determine the adaptability of transformer protection.

[0045] S3: As one implementation method, a multi-time-scale random policy network is constructed, and the transformer fault features filtered in step 2 are parsed using multi-channel feature masks. The time-domain invariance of the sample mask feature extraction is used to obtain the mask rate; the process of constructing a transformer fault behavior mapping based on mask feature evaluation is to achieve a unified representation of the changing characteristics of different fault samples; a transformer fault degree assessment network is established, combined with the loss in step 2.3. Optimize network parameters to achieve the final transformer protection trigger threshold, so as to realize the transformer adaptability characteristic analysis and action threshold setting under large-scale new energy output scenarios.

[0046] Step 3.1: [Referring to the information obtained in Step 2.1] Calculate the similarity distance of features for all transformer fault samples. Its main function is to assess the characteristics of fault samples. The similarity between samples is used for masking operations on transformer fault features. The similarity distance is calculated as follows: ; Where A represents the total number of elements in the feature matrix of each sample. The label is a feature of the sample.

[0047] Based on similarity distance Sort the samples from highest to lowest to obtain the corresponding labels for all samples. .generally, The numerical settings are related to the ranking to ensure that high-ranking samples receive greater numerical weight.

[0048] The higher a sample ranks in the list, the greater the difference in its fault characteristics compared to most other samples, making it more "unique". In fault diagnosis, this uniqueness often means that the sample contains more typical, more extreme, or more informative fault modes, and is a key "invariant feature".

[0049] Calculate the mask rate based on the sample label: ; In the formula, the mask rate Through weighting coefficients For sorting labels Linear weighting and normalization directly quantify the relative importance of the samples within the overall system. In transformer protection effectiveness analysis, a higher value indicates that the patterns contained within the fault sample set are more discriminative and representative, guiding the protection system to focus on high-value characteristics, optimize fault discrimination thresholds and action logic, thereby achieving more accurate and reliable protection performance improvements under complex operating conditions.

[0050] Based on mask rate Feature matrix of graph signal Perform random masking to obtain mask features. .

[0051] Step 3.2: Based on the graph signal feature matrix and mask features Construct a unified characterization of the variation features of fault samples ; Calculate the mask rate in step 3.1 Based on this, by relying on weight Convolutional encoder Feature matrix of graph signal Perform random masking to obtain mask features. Based on this, the mask features and filtering features Determining composite features:

[0052] in For learnable weights, This represents the horizontal concatenation of two features. Further extraction of mask features is then performed. and filtering features Composite characteristics of transformer faults Contribution in and This provides refined feature data for adaptive protection analysis, thereby enhancing the adaptive adjustment capability of protection devices to complex fault scenarios. The method for obtaining this data is as follows: ; ; in, This indicates that two feature matrices are concatenated along their feature dimensions. and For learnable feature mapping weights, and As learnable weights, the matrix multiplication rules guarantee that... It is a scalar.

[0053] Furthermore, a unified characterization of the variation features of different fault samples can be obtained. Used to describe transformers: .

[0054] Step 3.3: Establish a transformer fault severity assessment network to provide a unified representation of the fault sample variation characteristics from Step 3.2. Map protection thresholds to construct a transformer protection action mechanism.

[0055] Establish a transformer fault severity assessment network to provide a unified representation of the changing characteristics of fault samples in step 3.2. Determine the degree of transformer failure The mapping. As shown below, the input features... Based on the evaluation network output , This indicates the severity of a fault that may occur during transformer operation:

[0056] in, The activation function ensures the non-negativity of the transformer fault severity. The activation function undergoes a non-linear transformation. and This represents the learned weights and biases used for mapping.

[0057] By calculating the transformer fault severity from K samples The average value is used to obtain the threshold for transformer protection operation. :

[0058] In practical applications, the enhanced transformer feature matrix from step 2.1 is used as a basis. Calculate the protection action request value for transformer status :

[0059] when When the transformer's protection trips; when At that time, the transformer's protection did not activate.

[0060] Step 3.4: Constructing a transformer protection adaptability analysis based on confidence index Based on the function value in step 2.3 Define confidence index . The closer the value is to 1, the stronger the characteristic matrix of the transformer enhanced in step 2.1. The higher the confidence level of the transformer protection decision, i.e., the more it relies on the threshold, the better. The more adaptable the transformer protection mechanism is, the better.

[0061] Conversely, it indicates that the transformer protection mechanism is not adaptable enough (e.g., (Value less than 0.7) requires... To achieve this goal, the optimal threshold for transformer protection action is determined by iteratively optimizing the parameters of the spectral domain signal evaluation network in step 2 and the multi-timescale stochastic strategy network in step 3. To enhance the adaptability of protection.

[0062] As an embodiment of this application, a transformer protection fault characteristic and adaptability analysis system for new energy output scenarios is disclosed. Employing the specific implementation method described in the above-mentioned embodiment, the system includes: The graph model construction module is used to acquire multi-source fault signals of transformers in large-scale new energy transmission scenarios, construct an association graph model with multi-source fault signals as nodes, and perform spectral filtering analysis to obtain the graph signal feature matrix. The signal evaluation module is used to construct a spectral domain signal evaluation network, perform nonlinear measurement on the graph signal feature matrix, calculate the dependency strength between nodes, and output an enhanced feature matrix; evaluate the contribution of enhanced features to the transformer health status analysis results, and quantify the uncertainty of the feature combination on which the analysis results depend through a loss function; A unified representation processing module is used to calculate the similarity between transformer fault samples in the graph signal feature matrix, perform a masking operation on the fault features based on the similarity to obtain masked features, and construct a unified representation of the variation features of different fault samples based on the graph signal feature matrix and the masked features. The protection decision output module is used to construct a transformer fault severity assessment network, map the fault severity of the unified representation, calculate the mean fault severity of all samples to obtain the transformer protection action threshold, call the transformer fault severity assessment network, calculate the protection action request value of the transformer state based on the enhanced feature matrix, compare the threshold with the request value, and output the transformer protection decision. An adaptive analysis and optimization module is used to define confidence indices by quantifying values ​​through a loss function, and to analyze the adaptability of the transformer protection mechanism based on the transformer protection action threshold; and to optimize the transformer protection action threshold based on the adaptive analysis results.

[0063] As an embodiment of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it employs the specific implementation described in the analysis method above.

[0064] As an embodiment of this application, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, employs the specific implementation described in the analysis method above.

[0065] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0066] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0067] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0068] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0069] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios, characterized in that, The method includes: To acquire multi-source fault signals of transformers in large-scale new energy transmission scenarios, construct an correlation graph model with multi-source fault signals as nodes, and perform spectral filtering analysis to obtain the graph signal feature matrix; A spectral domain signal evaluation network is constructed to perform nonlinear measurement on the graph signal feature matrix, calculate the dependency strength between nodes, and output an enhanced feature matrix. The contribution of the enhanced features to the transformer health status analysis results is evaluated, and the uncertainty of the feature combination on which the analysis results depend is quantified by a loss function. The similarity between transformer fault samples in the graph signal feature matrix is ​​calculated, and the fault features are masked according to the similarity to obtain masked features; a unified representation of the changing features of different fault samples is constructed based on the graph signal feature matrix and the masked features. A transformer fault severity assessment network is constructed to map the fault severity of the unified representation; the mean fault severity of all samples is calculated to obtain the transformer protection action threshold; the transformer fault severity assessment network is invoked to calculate the protection action request value of the transformer state based on the enhanced feature matrix; the threshold value is compared with the request value to output the transformer protection decision. A confidence index is defined by quantifying the loss function value, which is used to analyze the adaptability of the transformer protection mechanism based on the transformer protection action threshold; the transformer protection action threshold is optimized based on the adaptability analysis results.

2. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 1, characterized in that, The transformer multi-source fault signal is a characteristic quantity of the transformer's operating data during a fault, including current, voltage, harmonics, saturation magnetic flux density, permeability, and frequency. Each feature quantity serves as a node in the graph structure, forming a set of nodes in the relational graph model; the relational graph model is an undirected graph, and there are undirected edges between each pair of feature quantities.

3. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 1, characterized in that, The process of obtaining the feature matrix of the graph signal includes: A learnable spectral filter is invoked to extract structured information from the correlation graph model; The relational graph model includes a set of nodes, a set of edges, and a set of edge weights; the edge weight is the correlation strength between every two nodes in the graph structure, represented by the correlation between the nodes.

4. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 1, characterized in that, The spectral domain signal evaluation network includes a feature mapping layer, a dependency evaluation layer, and an output layer; The feature mapping layer is used to extract the nonlinear coupling relationship between node features; the query vector and key vector of transformer fault features are calculated through nonlinear functions and trainable parameter matrices. The dependency evaluation layer is used to calculate an attention weight matrix based on the query vector and the key vector. Each element in the attention weight matrix... Quantized feature nodes right The strength of dependence; The output layer is used to fuse the graph signal feature matrix with the attention weight matrix to output an enhanced feature matrix.

5. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 1, characterized in that, The degree to which the evaluation enhancement features contribute to the transformer adaptability analysis results includes: The feature mapping of the transformer fault state is extracted from the enhanced feature matrix and fused with the feature weight vector and observation noise to form the transformer health status evaluation function; the transformer health status evaluation function is used to generate a health status evaluation value representing the transformer fault state. Define a Bayesian probability model for the health status evaluation value and introduce prior knowledge into the feature weight vector; when a new transformer state and health status evaluation value are obtained, update the feature weight vector according to the Bayesian probability model and obtain its posterior distribution, and generate a posterior mean vector. The posterior mean vector quantifies the contribution of the corresponding enhanced features to the transformer health status decision.

6. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 5, characterized in that, The assessment of the contribution of enhanced features to the transformer adaptability analysis results also includes: The loss function is constructed by combining the feature map with its own transpose and posterior uncertainty matrix; The loss value output by the loss function evaluates and quantifies the cognitive risk brought about by the combination of features on which the current decision is based in real time.

7. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 1, characterized in that, The masking operation includes the following steps: Calculate the similarity distance metric between the fault samples, sort them in descending order of similarity, and assign corresponding labels to all samples; The labels are linearly weighted by weighting coefficients, and the normalized result is the output mask rate. The feature matrix of the graph signal is then randomly masked based on the mask rate to obtain the mask features.

8. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 1, characterized in that, The construction process of the unified representation includes: The image signal feature matrix and the mask feature are concatenated and then multiplied with a learnable weight to obtain a composite feature; The contributions of the mask feature and the graph signal feature matrix to the composite feature are calculated respectively. Based on their respective contributions, the mask features and the graph signal feature matrix are weighted to obtain the unified representation.

9. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 8, characterized in that, The calculation process for the contribution includes: The contribution of the graph signal feature matrix to the composite feature , calculated using exponential function normalization; The contribution of the mask feature to the composite feature Subtracting from 1 The calculation yielded the result.

10. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 1, characterized in that, The mapping of the degree of failure includes: The unified characterization is input into the transformer fault severity assessment network, which outputs the fault severity. The inner layer of the transformer fault severity assessment network is a nonlinear activation function, and the outer layer is a non-negative activation function.

11. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 1, characterized in that, Compare the threshold value with the requested value, and output a transformer protection decision: When the requested value is greater than or equal to the threshold, the transformer protection operates; when the requested value is less than the threshold, the transformer protection does not operate.

12. The method for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios according to claim 1, characterized in that, Analysis of the adaptability of transformer protection mechanisms includes: A confidence index is defined based on the loss value output by the loss function, which is used to represent the adaptability of the transformer protection mechanism; the value of the confidence index is proportional to the confidence level of the transformer protection decision. When the confidence index value is less than the set standard, the optimal transformer protection action threshold is determined by iteratively optimizing the spectral domain signal evaluation network and the acquisition and mapping of the unified representation with the goal of minimizing the loss value.

13. A system for analyzing the fault characteristics and adaptability of transformer protection in new energy output scenarios, which executes the analysis method as described in any one of claims 1-12, characterized in that, The system includes: The graph model construction module is used to acquire multi-source fault signals of transformers in large-scale new energy transmission scenarios, construct an association graph model with multi-source fault signals as nodes, and perform spectral filtering analysis to obtain the graph signal feature matrix. The signal evaluation module is used to construct a spectral domain signal evaluation network, perform nonlinear measurement on the graph signal feature matrix, calculate the dependency strength between nodes, and output an enhanced feature matrix; evaluate the contribution of enhanced features to the transformer health status analysis results, and quantify the uncertainty of the feature combination on which the analysis results depend through a loss function; A unified representation processing module is used to calculate the similarity between transformer fault samples in the graph signal feature matrix, perform a masking operation on the fault features based on the similarity to obtain masked features, and construct a unified representation of the variation features of different fault samples based on the graph signal feature matrix and the masked features. The protection decision output module is used to construct a transformer fault severity assessment network, map the fault severity of the unified representation, calculate the mean fault severity of all samples to obtain the transformer protection action threshold, call the transformer fault severity assessment network, calculate the protection action request value of the transformer state based on the enhanced feature matrix, compare the threshold with the request value, and output the transformer protection decision. An adaptive analysis and optimization module is used to define confidence indices by quantifying values ​​through a loss function, and to analyze the adaptability of the transformer protection mechanism based on the transformer protection action threshold; and to optimize the transformer protection action threshold based on the adaptive analysis results.

14. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the analysis method according to any one of claims 1-12.

15. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the analysis method according to any one of claims 1-12.

Images