Methods, devices and non-volatile storage media for identifying high-resistance faults in power distribution networks

Abstract

This invention discloses a method, device, and non-volatile storage medium for identifying high-resistance faults in distribution networks. The method includes: acquiring the original zero-sequence current waveform of the distribution network line at the current moment; extracting features from the original zero-sequence current waveform to obtain an actual feature sequence; constructing an objective function with the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence, wherein the predicted feature sequence is obtained based on multiple historical feature sequences and an initial time-varying causal matrix; optimizing the initial time-varying causal matrix based on the objective function to obtain a target time-varying causal matrix; and determining the high-resistance fault identification result of the distribution network line based on the target time-varying causal matrix, wherein the high-resistance fault identification result is used to indicate whether a high-resistance fault has occurred in the distribution network line at the current moment. This invention solves the technical problem of low accuracy in identifying high-resistance faults in distribution networks in related technologies.

Description

Technical Field

[0001] This invention relates to the field of smart grids, and more specifically, to a method, apparatus, and non-volatile storage medium for identifying high-resistance faults in distribution networks. Background Technology

[0002] With the widespread integration of distributed power sources and the increasing complexity of distribution network structures, high-impedance faults (HIFs) have become a major challenge affecting the reliability and safety of distribution networks. These faults are characterized by a high transition resistance between the fault point and the ground, resulting in relatively small fault currents that are difficult to detect effectively by protection devices based on sudden current detection. This is especially true in distribution network systems with ungrounded neutral points or those using arc suppression coils, where minute changes in zero-sequence current are even more difficult to detect, significantly reducing the accuracy of high-impedance fault identification.

[0003] In relevant high-resistance fault detection technologies, many rely on direct analysis of the zero-sequence current waveform or simple frequency domain analysis to attempt to distinguish between normal operation and fault states. However, such methods often neglect the dynamic characteristics of the zero-sequence current waveform and its interaction with other electrical parameters. Especially under non-fault conditions, the zero-sequence current may fluctuate due to load imbalance, transformer errors, and other reasons, further increasing the difficulty of high-resistance fault identification and resulting in low accuracy in high-resistance fault identification in distribution networks.

[0004] There is currently no effective solution to the problem of low accuracy in identifying high-resistance faults in distribution networks in the aforementioned related technologies. Summary of the Invention

[0005] This invention provides a method, apparatus, and non-volatile storage medium for identifying high-resistance faults in power distribution networks, thereby at least addressing the technical problem of low accuracy in identifying high-resistance faults in power distribution networks in related technologies.

[0006] According to one aspect of the present invention, a method for identifying high-resistance faults in a distribution network is provided, comprising: acquiring the original zero-sequence current waveform of a distribution network line at the current moment; extracting features from the original zero-sequence current waveform to obtain an actual feature sequence, wherein the actual feature sequence includes the actual feature values ​​of multiple features at the current moment; constructing an objective function with the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence, wherein the predicted feature sequence is obtained based on multiple historical feature sequences and an initial time-varying causal matrix, and the elements in the initial time-varying causal matrix are used to indicate the contribution intensity of the historical feature value of one feature to the predicted feature value of another feature; optimizing the initial time-varying causal matrix based on the objective function to obtain a target time-varying causal matrix; and determining the high-resistance fault identification result of the distribution network line based on the target time-varying causal matrix, wherein the high-resistance fault identification result is used to indicate whether a high-resistance fault has occurred in the distribution network line at the current moment.

[0007] According to another aspect of the present invention, a high-resistance fault identification device for a distribution network is also provided, comprising: a current waveform acquisition module for acquiring the original zero-sequence current waveform of the distribution network line at the current moment; a feature extraction module for extracting features from the original zero-sequence current waveform to obtain an actual feature sequence, wherein the actual feature sequence includes the actual feature values ​​of multiple features at the current moment; an objective function construction module for constructing an objective function with the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence, wherein the predicted feature sequence is obtained based on multiple historical feature sequences and an initial time-varying causal matrix, and the elements in the initial time-varying causal matrix are used to indicate the contribution intensity of the historical feature value of one feature to the predicted feature value of another feature; an optimization module for optimizing the initial time-varying causal matrix based on the objective function to obtain a target time-varying causal matrix; and a fault identification module for determining the high-resistance fault identification result of the distribution network line based on the target time-varying causal matrix, wherein the high-resistance fault identification result is used to indicate whether a high-resistance fault has occurred in the distribution network line at the current moment.

[0008] According to another aspect of the present invention, a non-volatile storage medium is also provided, which stores multiple instructions, any one of which is adapted to be loaded by a processor for a method for identifying high-resistance faults in a power distribution network.

[0009] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store 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 implement any one of the high-resistance fault identification methods for power distribution networks.

[0010] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of any one of the distribution network high-resistance fault identification methods.

[0011] In this embodiment of the invention, the original zero-sequence current waveform of the distribution network line at the current moment is obtained; features are extracted from the original zero-sequence current waveform to obtain an actual feature sequence, wherein the actual feature sequence includes the actual feature values ​​of multiple features at the current moment; an objective function is constructed with the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence, wherein the predicted feature sequence is obtained based on multiple historical feature sequences and an initial time-varying causal matrix, and the elements in the initial time-varying causal matrix are used to indicate the contribution strength of the historical feature value of one feature to the predicted feature value of another feature; based on the objective function, the initial time-varying causal matrix is ​​processed... The optimization process yields the target time-varying causal matrix. Based on this matrix, the high-resistance fault identification result for the distribution network line is determined. This result indicates whether a high-resistance fault has occurred in the distribution network line at the current moment. This process constructs and optimizes the time-varying causal matrix to minimize the error between the actual and predicted feature sequences, enabling accurate detection and identification of high-resistance faults at the current moment. This improves the accuracy of high-resistance fault identification in the distribution network and overcomes the limitations of related technologies in handling non-stationary, time-varying signals. Ultimately, it solves the technical problem of low accuracy in high-resistance fault identification in the distribution network. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0013] Figure 1 This is a flowchart of a high-resistance fault identification method for a distribution network according to an embodiment of the present invention;

[0014] Figure 2 This is a schematic diagram of a high-resistance fault identification device for a power distribution network according to an embodiment of the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0016] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0017] According to an embodiment of the present invention, a method for identifying high-resistance faults in a distribution network is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0018] Figure 1 This is a flowchart of a high-resistance fault identification method for distribution networks according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0019] Step S102: Obtain the original zero-sequence current waveform of the distribution network line at the current moment.

[0020] Optionally, the raw zero-sequence current waveform is obtained by continuously collecting the zero-sequence current of the distribution network lines during the current moment and a period of time in between (which can be denoted as the current sampling period). Zero-sequence current refers to the vector sum of the three-phase currents in the distribution network lines. It is theoretically zero during normal operation, but it increases significantly during ground faults or asymmetrical faults. The raw zero-sequence current waveform can be collected in real time using current transformers (CTs) or zero-sequence current sensors, reflecting the changes in zero-sequence current in the distribution network during a period prior to the current moment. This waveform data contains transient and steady-state information of the current signal after the fault occurs, serving as the basis for subsequent feature extraction and fault analysis.

[0021] Step S104: Extract features from the original zero-sequence current waveform to obtain the actual feature sequence, wherein the actual feature sequence includes the actual feature values ​​of multiple features at the current time.

[0022] Optionally, after acquiring the original zero-sequence current waveform, feature extraction is required to process the data more efficiently and extract fault features. Feature extraction involves converting complex waveform data into a series of values ​​representing specific attributes of the signal. These attributes may include, but are not limited to, frequency domain features (such as the amplitude and phase of the fundamental and higher harmonics), time domain features (such as peak value, mean, RMS value, wavelet transform coefficients, etc.), or statistical features (such as skewness, kurtosis, etc.). For example, for 10 sampling points in the original zero-sequence current waveform corresponding to the current moment, feature extraction is performed on these 10 sampling points to obtain the corresponding actual feature sequence. The actual feature sequence is generated through the feature extraction step and contains the specific value of each feature at the current moment, providing direct input for subsequent prediction and analysis.

[0023] Step S106: With the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence, construct an objective function. The predicted feature sequence is obtained based on multiple historical feature sequences and an initial time-varying causal matrix. The elements in the initial time-varying causal matrix are used to indicate the contribution strength of the historical feature value of one feature to the predicted feature value of another feature.

[0024] Optionally, constructing the objective function is the core of the optimization process. In this embodiment, the predicted feature sequence is calculated based on historical feature sequences and a time-varying causal matrix, attempting to predict the value of each feature at the current moment using past data. The objective function aims to minimize the error between the actual feature sequence and the predicted feature sequence; this error can be quantified as the loss component of the objective function. Furthermore, the objective function may include a regularization term to control model complexity and prevent overfitting, ensuring the sparsity and stability of model parameters. This step ensures that the optimization process has a clear objective: to find the time-varying causal relationship pattern that can most accurately predict the current feature value.

[0025] Optionally, multiple historical feature sequences can be obtained by grouping the sampling points included in the zero-sequence current waveform of a historical period into multiple groups of sampling points, and then extracting features from each group of sampling points. For example, for 1000 sampling points in the zero-sequence current waveform of a historical period, the 1000 sampling points can be divided into 100 groups of 10, and features can be extracted from each group to obtain 100 historical feature sequences.

[0026] In an optional embodiment, before constructing the objective function with the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence, the method further includes obtaining the predicted feature sequence as follows: ;in, Indicates the predicted feature sequence; This represents the initial time-varying causal matrix, and the elements in the initial time-varying causal matrix. Representation of features Historical eigenvalues ​​of features The contribution strength of the predicted eigenvalues; Let T represent the historical information matrix, which includes multiple historical feature sequences. T represents the initial sampling time of the first historical feature sequence among the multiple historical feature sequences, and t represents the current time. This represents the initial time-varying feature extraction matrix, where each column of the initial time-varying feature extraction matrix... This is used to learn how to optimally linearly combine multiple historical feature sequences to extract the most relevant data for predicting the next feature. The most effective time-domain dynamic component of the feature, where k is any one of the multiple features.

[0027] Optionally, before formally constructing the objective function, the feature values ​​at the current moment are predicted based on historical data. This step is achieved by combining the historical information matrix, the initial time-varying causal matrix, and the initial time-varying feature extraction matrix. The process of generating the predicted feature sequence reflects the dynamic predictive relationship between system variables and is the basis for subsequent error calculation and model optimization. This is a historical information matrix, containing information from time [time]. arrive The observed values ​​(i.e., feature values) of all N features are the complete historical basis for the model's predictions. The time-varying feature extraction matrix can be understood as a set of dynamic finite impulse response (FIR) filters. Each column Responsible for learning how to learn the past of all features The information at time step 1 is optimally linearly combined to extract the information for predicting the 1st time step 2. The most effective temporal dynamic component of the feature. This is the time-varying causal relationship matrix, which is the core of the model. It quantifies the causal relationship at time t. ,Depend on The predictive ability among the extracted dynamic components of features. Its elements Represents characteristics Historical information for predicting features The strength of the contribution of the current value is used to capture the dynamic causal relationship between variables within the framework of Granger causality. Indicates the model at time... For all A vector of predicted values ​​for each feature.

[0028] The predicted feature sequence obtained through the above method means that at time t, the predicted feature value is calculated by multiplying the historical information matrix, the time-varying causality matrix, and the feature extraction matrix. The dynamic adjustment of the time-varying causality matrix and the feature extraction matrix ensures that the model can accurately capture the non-stationary and time-varying characteristics of the signal, thereby improving the accuracy of prediction and the ability to identify high-impedance faults. This prediction process provides the foundation for subsequent construction of the objective function and parameter optimization.

[0029] In an optional embodiment, the objective function is constructed with the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence. This includes constructing the objective function as follows: ;in, The value of the objective function is used to quantify the error between the actual feature sequence and the predicted feature sequence. This represents the first regularization parameter; This represents the second regularization parameter; This represents the prediction error term (L) in the objective function. L1 represents the sparse term in the objective function, and N represents the total number of features included in the actual feature sequence. This represents the stable term (L2) in the objective function.

[0030] Optionally, constructing the objective function is a crucial step in optimizing model parameters. Its core objective is to minimize the difference between the actual feature sequence and the predicted feature sequence, while introducing a regularization term to ensure the sparsity and stability of the model. The objective function is constructed using a comprehensive optimization framework combining Elastic Net Regularization, which ensures the model's predictive ability and robustness when handling complex, non-stationary signals. In this objective function, This represents the prediction error term. The primary goal of the model is accurate prediction, and this term ensures that the model output... Compared with the true value The deviation should be as small as possible. The sparse terms of L1 are composed of... In power systems, not all electrical characteristics have a direct, strong causal relationship; the purpose of this term is to select variables that penalize all non-zero causal coefficients, forcing those with weak contributions to prediction to be excluded. The coefficients become zero, thus achieving sparsity in the model. This not only prevents overfitting but also makes the final causal graph clearer and more physically interpretable. The stabilizing term of L2 is... Control. This item has two main functions: first, to prevent the model parameter values ​​from becoming too large, thereby improving generalization ability; second, to make the objective function a strictly convex function during the optimization process, which is crucial for ensuring the stable convergence of the iterative algorithm.

[0031] It should be noted that the objective function in this embodiment reflects a comprehensive consideration of model performance. The prediction error term L ensures the good agreement between the model output and the actual data, while the sparsity term L1 and the stability term L2 control the sparsity and stability of the model parameters, respectively. This comprehensive strategy helps to find the optimal balance between accurate prediction, parameter simplification, and model robustness, especially when dealing with non-stationary, time-varying, and weak signals such as high-resistance faults in power systems, significantly improving the accuracy and robustness of fault detection. By minimizing the value of the objective function, the model can learn the variables that truly play a role in prediction and the key causal relationships between them, thereby more accurately identifying abnormal states in the system.

[0032] Step S108: Based on the objective function, optimize the initial time-varying causal matrix to obtain the target time-varying causal matrix.

[0033] Optionally, the optimization process involves iteratively updating the elements of the initial time-varying causal matrix to minimize the value of the objective function. Gradient descent can be used, but is not limited to, for iterative optimization of the time-varying causal matrix. In each iteration, the parameters in the matrix are adjusted according to the gradient of the objective function, gradually approximating the causal relationships that most accurately predict the feature values ​​at the current time. The optimized target time-varying causal matrix contains the true, sparse, and stable causal contribution strength among the features within the system at a specific time point, serving as an important basis for fault identification.

[0034] In one optional embodiment, the initial time-varying causal matrix is ​​optimized based on the objective function to obtain the target time-varying causal matrix. This includes: updating the initial time-varying feature extraction matrix using gradient descent with the objective function as the minimum value; wherein each column of the initial time-varying feature extraction matrix is ​​used to learn how to optimally linearly combine multiple historical feature sequences to extract the most effective time-domain dynamic component for predicting any feature; performing a Taylor expansion on the prediction error term in the objective function to obtain the processed function; updating the elements in the initial time-varying causal matrix using a sub-gradient update rule to obtain the updated time-varying causal matrix; repeating the above operations until a predetermined termination condition is reached; and using the updated time-varying causal matrix obtained when the predetermined termination condition is reached as the target time-varying causal matrix.

[0035] Optionally, firstly, for the objective function, the initial time-varying feature extraction matrix is ​​optimized using gradient descent. Gradient descent is an iterative method for solving minimization problems. It determines the direction of the next parameter update by calculating the gradient of the objective function (i.e., the direction and rate of change of the function). In this process, each column of the initial time-varying feature extraction matrix at each time point is responsible for learning how to linearly combine historical feature sequences to extract the most influential time-domain dynamic components for predicting any feature at the current time. By minimizing the objective function value, gradient descent ensures that the parameters in the optimized time-varying feature extraction matrix are gradually adjusted to the optimal state to most effectively predict the value of each feature. To simplify the prediction error term in the objective function, Taylor expansion can be used to approximate it. Taylor expansion simplifies the processing of nonlinear or complex functions by approximating the derivative information of the function at a certain point using a polynomial. In this step, by performing a first-order Taylor expansion of the prediction error term L in the objective function at the current parameter value, a processed approximate function can be obtained. This approximate function not only preserves the main features of L but also makes its differentiation and optimization process more direct and efficient. Subsequently, the optimization objective shifts to updating the time-varying causal matrix. Because the objective function includes an L1 regularization term, it becomes a non-smooth function, which makes gradient descent, a common technique, difficult to handle. To address this, a subgradient update rule is used instead of gradient descent. The subgradient refers to the lower bound slope of the function at non-smooth points, guiding the direction of parameter updates. The subgradient update rule indicates the optimization path for non-smooth functions, especially when dealing with L1 regularization terms. It uses subgradient theory to find the appropriate direction and step size for parameter updates in each iteration, achieving sparsity and optimization of parameter values. This rule is particularly suitable for objective functions containing non-smooth terms such as absolute values ​​or step functions. Through subgradient updates, the elements in the time-varying causal matrix can be gradually adjusted, ultimately resulting in a sparse time-varying causal matrix that captures the key dynamic causal relationships between variables within the system. The optimization process is iterative, requiring the repetition of the gradient descent and subgradient update steps until a termination condition is met. The termination condition can be at least one of the following: reaching a predetermined number of iterations, the change in the objective function value being less than a certain threshold, or the parameter update magnitude no longer being significant. This process ensures that the model can learn sufficiently and gradually approach the optimal state. When the above optimization iterations reach the termination condition, the time-varying causal matrix obtained from the last update is determined as the target time-varying causal matrix. This matrix contains the strength of the optimal, sparse dynamic causal relationships among the variables within the system at the current moment, i.e., those causal connections that are truly effective and stable for feature prediction. The target time-varying causal matrix is ​​a key output for fault identification. By comparing it with the actual feature sequence, it can be used to determine whether a high-resistance fault has occurred in the distribution network, as well as the specific nature and location of the fault.

[0036] Optionally, due to the presence of the L1 norm term, the objective function becomes a non-smooth function and cannot be solved directly using standard gradient descent. An online update algorithm combining Taylor expansion and subgradient theory can be used. Specifically, considering... The update does not involve non-smooth terms, so the learning rate can be used directly. Gradient descent method: , where gradient It can be analytically derived using the chain rule. In updating C(t), first, the objective function regarding... The smoothing term L at the current point Perform a first-order Taylor expansion: Through this approximation, the original optimization problem is decoupled to a problem for each element. Subsequently, subgradient theory is introduced to handle... item.

[0037] In one optional embodiment, the elements in the initial time-varying causal matrix are updated using a subgradient update rule to obtain an updated time-varying causal matrix. This includes: updating the elements in the initial time-varying causal matrix using a subgradient update rule in the following manner to obtain the updated time-varying causal matrix:

[0038] ;

[0039] in, This represents the first regularization parameter in the objective function; L represents the prediction error term in the objective function. This represents the second regularization parameter in the objective function; Represents any element in the initial time-varying causal matrix; This represents the element corresponding to any element in the previous time-varying causal matrix, where the previous time-varying causal matrix is ​​the time-varying causal matrix corresponding to the previous feature sequence of the actual feature sequence.

[0040] Optional, Let the prediction error be at the previous time t-1. For coefficients The gradient or partial derivative; this value represents The extent to which minute changes affect prediction errors serves as a signal guiding the direction and magnitude of necessary corrections; The L1 regularization parameter acts as a sparsity threshold if the gradient is less than 1. In this case, the causal relationship is considered insignificant and is forcibly set to zero; The L2 regularization parameter acts as an update stability and scaling factor, controlling the magnitude of the non-zero coefficients and ensuring numerical stability during the update process. The application of the subgradient update rule in the above formula is reflected in the... The update includes processing of the L1 regularization term. Since the L1 term is non-smooth, gradient descent cannot be used directly; instead, the concept of subgradient is employed. Subgradient is an optimization technique for non-smooth functions, providing a reasonable descent direction at points of non-smoothness. Here, the sign function is used as the subgradient of the L1 regularization term, according to... The positive or negative value determines the direction of the update and the severity of the punishment.

[0041] By comprehensively considering prediction error, sparsity, stability, and a penalty term for time continuity, this gradient update rule effectively optimizes the elements in the time-varying causal matrix. In each iteration, The update not only reduces the current prediction error, but also... and The complexity of the model and the smoothness of parameter changes were controlled. This iterative optimization ensures that the model can capture the instantaneous dynamic changes of system characteristics while maintaining good sparsity and stability, thus performing excellently when handling time-varying and non-stationary signals. Finally, when the optimization process reaches the predetermined termination condition, the updated time-varying causal matrix is ​​used as the target time-varying causal matrix for subsequent fault identification processes. This matrix contains the dynamic causal relationship patterns between key variables within the system, providing crucial technical support for the accurate identification of non-stationary high-resistivity faults. Through this refined matrix adjustment, the model can better adapt to the continuous changes in signals within the power system, improving its ability to detect complex fault scenarios.

[0042] Optionally, according to Fermat's Lemma for convex optimization, a point is an optimal solution if and only if 0 belongs to the second differential of the objective function at that point. Therefore, the optimality condition can be obtained: ,in, Is the absolute value function at its optimal point? The subgradient. By analyzing the subgradient. By discussing the three cases (>0, <0, =0), the closed update rule of the above form (i.e.) can be derived. ).

[0043] Optionally, to prevent numerical overflow caused by sudden data mutations leading to excessively large gradients during training or prediction (i.e., during the updating of each element in the initial time-varying causal matrix using the subgradient update rule), norm clipping is performed after each gradient calculation. That is, a threshold is set. If the L2 norm of the gradient Then the gradient will be scaled to This provides the ultimate guarantee for the stability of the model.

[0044] Step S110: Based on the target time-varying causal matrix, determine the high-resistance fault identification result of the distribution network line, wherein the high-resistance fault identification result is used to indicate whether a high-resistance fault has occurred in the distribution network line at the current moment.

[0045] Optionally, after obtaining the optimized target time-varying causal matrix, this step aims to evaluate the difference between this matrix and the system under normal operating conditions to identify whether a high-resistance fault has occurred. High-resistance fault identification can be performed by combining dynamic information from the target time-varying causal matrix with reference to historical data to ensure the accuracy and reliability of fault identification.

[0046] It should be noted that steps S102 to S110 constitute a closed-loop fault detection process. From data acquisition, feature extraction, model optimization to fault diagnosis, each step is closely linked and works together to achieve the final high-resistance fault identification result. By dynamically capturing and quantifying the complex time-varying causal relationships between variables within the system, the method of this embodiment can overcome the limitations of related fault detection technologies when processing non-stationary signals, providing an innovative solution for real-time and accurate fault detection.

[0047] In one optional embodiment, determining the high-resistance fault identification result of the distribution network line based on the target time-varying causal matrix includes: correcting the elements in the target time-varying causal matrix from the time dimension to obtain the target causal stability matrix, wherein the elements in the target causal stability matrix represent the stability measure of any two features in the time series; determining the deviation between the target causal stability matrix and the preset benchmark causal stability matrix; and determining the high-resistance fault identification result based on the deviation.

[0048] Optionally, firstly, the elements of the target time-varying causal matrix are modified from a time perspective to generate a target causal stability matrix. This process typically involves statistical analysis of the strength of time-varying causal relationships, ensuring that only those causal relationships exhibiting stable and significant trends over time are retained. This helps filter out spurious associations caused by noise, system fluctuations, or other transient factors. Each element in the target causal stability matrix can be represented as the ratio of the average strength of the corresponding element within a time window to the variability of its strength over a short period (similar to the signal-to-noise ratio). The purpose of calculating this matrix is ​​to quantify the long-term consistency and short-term stability of each causal relationship. In this way, truly stable and persistent causal links can be selected, even if their strength or direction may change significantly during high-impedance faults. Secondly, the target causal stability matrix is ​​compared with a pre-defined baseline causal stability matrix. The baseline matrix can be trained using a large amount of historical normal operation data and represents the typical causal structure of a power system under fault-free conditions. By calculating the deviation between the target matrix and the baseline matrix, the degree of difference between the current system state and the normal operating state can be quantified. The calculation of deviation may involve distance metrics between matrix elements. For example, Euclidean distance, Manhattan distance, or more complex distance metrics such as Earth Mover's Distance can be used to assess the similarity or difference between two matrices. Finally, based on the calculated deviation, a high-resistance fault is determined. For example, a high deviation indicates a significant change in the target causal stability matrix compared to the baseline matrix, which may be due to a high-resistance fault in the system. Specific operations include, but are not limited to, setting a threshold; when the deviation exceeds this threshold, the system is determined to be in a high-resistance fault state. Optionally, the deviation can also be used to assess the severity or type of the fault. For example, the magnitude of the deviation may be positively correlated with the fault intensity, and a high deviation may point to a specific fault type or location. This approach provides a systematic way to detect and locate high-resistance faults by analyzing changes in the causal relationships between variables within the system, offering higher sensitivity and specificity compared to methods based solely on a single electrical feature.

[0049] In the above approach, a more stable and reliable causal stability matrix is ​​extracted from the target time-varying causal matrix and compared with the baseline under normal conditions, thus achieving effective identification of high-resistance faults. This fault detection method based on dynamic causal relationship analysis can provide a deeper understanding of the internal state changes of the power system, offering strong support for the safe operation and intelligent maintenance of the power grid.

[0050] In one optional embodiment, the elements in the target time-varying causal matrix are modified from the time dimension to obtain the target causal stability matrix, including: modifying the elements in the target time-varying causal matrix to obtain the target causal stability matrix in the following manner: ;in, Represents any element in the target causal stability matrix; Represents any element in the target time-varying causality matrix; The mean of any element and the corresponding elements in multiple historical time-varying causal matrices is given. Multiple historical time-varying causal matrices correspond one-to-one with multiple historical feature sequences. is the standard deviation of any element and the corresponding element in multiple historical time-varying causal matrices; This represents a preset positive constant. It takes into account the elements in the target time-varying causality matrix. The sequence may contain noise. To extract stable and reliable causal relationships, a causal stability measure of the form described above was designed. The underlying idea is that a reliable causal relationship should have high average strength and low volatility. Among these, It is a standard deviation function in the time dimension, which measures the fluctuation or variation of the strength of causal relationship over time. A low value indicates a stable relationship. A small positive constant (e.g., 0.1) is used to avoid division by zero. This also contributes to the stability measure when the standard deviation is zero (i.e., the causal relationship has constant strength) and when the standard deviation is very close to zero. This measure is similar to the signal-to-noise ratio, effectively filtering out accidental strong associations. Ultimately, a stable causal matrix can be computed for a set of data, i.e., the target causal stability matrix. .

[0051] Optionally, by Average strength of the table with standard deviation The ratio of is used as an element in the target causal stability matrix. Essentially, this involves calculating the signal-to-noise ratio (SNR) of each causal relationship. A higher SNR indicates a more stable and significant causal relationship over time, making it more reliable. Therefore, the target causal stability matrix can more accurately identify core causal relationships that change significantly under high-resistance faults while maintaining a certain level of stability. The target causal stability matrix not only enhances the accuracy of fault identification but also increases the robustness of the method. In high-resistance fault detection, a reliable causal relationship should exhibit significant changes in strength while not being entirely affected by random noise. By quantifying the temporal stability and strength of causal relationships, we can more effectively filter out internal system connections that truly change under fault conditions, thereby improving the reliability of diagnostic results.

[0052] In summary, this correction process generates a target causal stability matrix through statistical analysis of the target time-varying causal matrix. This matrix contains a comprehensive evaluation of the stability and strength of causal relationships within the system. This matrix is ​​a crucial input for subsequent fault identification and location, helping to accurately identify key changes under high-resistance fault conditions and providing strong data support for real-time monitoring and maintenance of power systems.

[0053] In one optional embodiment, the high-resistance fault identification result is determined based on the deviation degree, including: if the deviation degree is greater than a preset deviation threshold, the high-resistance fault identification result is determined to be that the distribution network line has a high-resistance fault at the current moment; or if the deviation degree is less than or equal to the preset deviation threshold, the high-resistance fault identification result is determined to be that the distribution network line has not experienced a high-resistance fault at the current moment.

[0054] Optionally, after constructing the target causal stability matrix, the next step is to compare and analyze it with a preset benchmark causal stability matrix. By calculating the deviation between the two, it is determined whether a high-resistance fault has occurred in the distribution network at the current moment. The specific judgment logic is based on comparing the magnitude of the deviation with a preset deviation threshold, and falls into two categories: if the calculated deviation exceeds the preset deviation threshold, it is determined that a high-resistance fault may exist in the distribution network at the current moment. The deviation is a quantifiable measure of the difference between the target causal stability matrix and the benchmark stability matrix, reflecting the degree of change in the dynamic causal structure within the system. Under normal operating conditions, the causal structure of the system is relatively stable, and the deviation should remain at a low level. However, when a high-resistance fault occurs, some causal relationships within the system suddenly change, leading to a significant deviation between the benchmark and target causal stability matrices. Therefore, once the deviation exceeds the preset threshold, it can be determined that a high-resistance fault is very likely to have occurred at the current moment. Conversely, if the calculated deviation is lower than or equal to the preset deviation threshold, it indicates that the difference between the target causal stability matrix and the benchmark matrix is ​​small, and the dynamic causal relationships within the system have not changed significantly. In this scenario, the model will determine that the distribution network lines are currently in a normal state and no high-resistance fault has occurred. This judgment is based on the assumption that, under normal operating conditions, the causal structure of the system should be stable and predictable, consistent with historical patterns under normal conditions. Therefore, when the deviation remains at a low level, the system can be considered to have not yet been affected by a high-resistance fault. The deviation can be calculated by comparing the differences between corresponding elements in two matrices and then aggregating these differences into a single value. Aggregation methods can include calculating the average of the differences between all elements, or employing more complex techniques such as calculating Euclidean distance, Hamming distance, or Jaccard similarity coefficients. The selection of the judgment threshold can be obtained through extensive experimental verification to ensure that the model can effectively identify high-resistance faults in real power systems while avoiding excessive false alarms.

[0055] By comparing the deviation between the target causal stability matrix and the benchmark causal stability matrix, and combining this with a set threshold, real-time monitoring and rapid response to high-resistance faults can be achieved. The core advantage of this method lies in its transcendence of overcurrent detection in related technologies; instead, it identifies faults by capturing and quantifying the dynamic causal relationship changes between multidimensional electrical characteristics within the power system. Since high-resistance faults are typically accompanied by complex and subtle changes in electrical characteristics, methods in related technologies often struggle to detect them. However, analysis based on dynamic causal relationships can monitor these changes more meticulously and promptly identify potential safety hazards. Furthermore, setting deviation thresholds allows for flexible adjustment of monitoring strategies under different operating environments to adapt to the diversity and complexity of the power grid, improving overall detection efficiency and system safety.

[0056] Through the above steps S102 to S110, the time-varying causal matrix can be constructed and optimized to minimize the error between the actual feature sequence and the predicted feature sequence. This enables accurate detection and identification of high-resistance faults occurring at the current moment, thereby improving the accuracy of high-resistance fault identification in the distribution network and overcoming the limitations of related technologies in processing non-stationary and time-varying signals. This also solves the technical problem of low accuracy in high-resistance fault identification in the distribution network.

[0057] Currently, the primary challenge facing high-impedance fault identification technology in practical applications stems from the fundamental mismatch between the physical characteristics of the fault signal itself and traditional analysis tools. The arc generated by a high-impedance fault exhibits significant intermittency and randomness, causing its electrical signal to display typical non-stationary and time-varying characteristics. This inherent property directly contradicts the theoretical foundation of many classical signal processing methods. For example, analysis tools based on steady-state assumptions, such as Fourier transforms, cannot effectively capture the complete dynamic evolution of the fault, thus losing crucial transient information.

[0058] Secondly, even more advanced model-based statistical analysis methods are still limited by their theoretical assumptions. Take classic Granger causality analysis as an example; its theoretical validity strictly depends on the assumption of time series stationarity. When these methods are directly applied to inherently non-stationary high-resistivity fault signals, serious model mismatch problems inevitably occur, leading to unreliable or even completely erroneous diagnostic conclusions. This disconnect between theoretical foundations and practical applications is one of the key reasons why related technologies frequently fail in the face of high-resistivity faults.

[0059] To address the time-varying nature of signals, while some advanced models capable of tracking dynamic changes, such as time-varying vector autoregression (VAR) models, have been developed, they introduce new challenges: poor redundancy and interpretability of results. Due to the general lack of effective sparsity constraint mechanisms, these models identify a large number of redundant and weak correlations between variables in the system. This "dense" output results cause key fault characteristics to be submerged in massive amounts of irrelevant information, making it difficult to provide a clear physical interpretation of the diagnostic results and making the model more susceptible to random noise, thus affecting the robustness of the diagnosis. Therefore, the relevant technical system lacks a perfect solution that can simultaneously handle non-stationarity and ensure both sparsity and interpretability of results.

[0060] To address the aforementioned problems, and based on the above embodiments and optional embodiments, this invention proposes an optional implementation method for identifying high-resistance faults in distribution networks, the method comprising:

[0061] Step S1: Problem definition and fault identification model framework, specifically including:

[0062] To address the non-stationarity and time-varying nature of high-impedance fault signals, this paper proposes a Time-Varying Sparse Causal Regression (TSCR) model that dynamically captures the correlation patterns between variables within the system, abandoning static models used in related techniques. The core idea of ​​this model is to transform the fault identification problem into a process of quantifying changes in the dynamic behavior patterns of the system.

[0063] First, the original zero-sequence current waveform of the univariate model is transformed into a multivariate time series (MTS) through windowing and feature extraction (such as harmonic analysis, wavelet energy, statistical moments, etc.). Where M is the number of time points (windows) and N is the feature dimension.

[0064] An end-to-end time-varying prediction model was constructed, the core of which is to predict the feature value at the current time t by utilizing information from the past T time points of all features. This model consists of a time-varying causal matrix. and time-varying weight matrix Driven by the common factors, its prediction function is defined in the following form: The interpretation of each variable is the same as in the previous embodiments, and will not be repeated here.

[0065] Step S2: Optimize the objective function, specifically including: In order for the model to identify key, non-redundant causal relationships, an objective function incorporating Elastic Net regularization was designed. The formula for this objective function is as follows: The interpretation of each variable is the same as in the previous embodiments, and will not be repeated here.

[0066] Step S3: Sparse time-varying parameter update algorithm, specifically including:

[0067] Due to the presence of the L1 norm term, the objective function becomes a non-smooth function and cannot be solved directly using standard gradient descent. An online update algorithm combining Taylor expansion and subgradient theory can be used, specifically including:

[0068] S31, Update:

[0069] The update does not involve non-smooth terms, so the learning rate can be used directly. Gradient descent method: .

[0070] S32, Update of C(t):

[0071] This is the key innovation of the algorithm. First, the objective function regarding... The smoothing term L at the current point Perform a first-order Taylor expansion: Through this approximation, the original optimization problem is decoupled to each element.

[0072] This is an independent optimization subproblem. Subsequently, subgradient theory is introduced to handle it. According to Fermat's lemma for convex optimization, a point is an optimal solution if and only if 0 belongs to the second differential of the objective function at that point. Therefore, the optimality condition can be obtained: ,in, Is the absolute value function at its optimal point? The subgradient. By analyzing the subgradient. By discussing the three cases (>0, <0, =0), the following closed-form update rule can be derived: The interpretation of each variable is the same as in the previous embodiment, and will not be repeated here. The above formula achieves parameter updating and sparsification in each iteration through a soft thresholding method.

[0073] S33, Gradient clipping:

[0074] To prevent numerical overflow caused by sudden data mutations leading to excessively large gradients during training or prediction (i.e., when updating each element in the initial time-varying causality matrix using a subgradient update rule), norm clipping is performed after each gradient calculation. That is, a threshold is set. If the L2 norm of the gradient Then the gradient will be scaled to This provides the ultimate guarantee for the stability of the model.

[0075] Step S4: Fault diagnosis process, specifically including:

[0076] S41, Causal Stability Quantification:

[0077] The sequence may contain noise. To extract stable and reliable causal relationships, a causal stability measure was designed. The idea is that a reliable causal relationship should have a high average strength and low volatility, among which, The interpretation of each variable is the same as in the previous embodiments, and will not be repeated here.

[0078] This metric, similar to signal-to-noise ratio, effectively filters out accidental strong correlations. Ultimately, a stable causal matrix can be calculated for a dataset. .

[0079] S42, Fault Identification:

[0080] This method is based on a core assumption: the occurrence of a failure will significantly alter the causal structure within the system.

[0081] Training phase: The model is trained using a large amount of normal training data (normal_train_data), and a baseline causal stability matrix representing the health state is calculated and stored. .

[0082] Testing phase: For the test data new_data, calculate its target causal stability matrix. .

[0083] Fault score calculation: through calculation and The difference is used to quantify the degree of deviation of the system (i.e., the degree of deviation): ,in, It is a causal stability index calculated from new data samples being tested; it represents the "fingerprint" of the dynamic relationships of the current state of the system. It is a causal stability index calculated from a large set of normal, healthy operating data. It serves as a "baseline" or "reference fingerprint" for comparison with all new data.

[0084] The high-resistance fault identification result, Prediction, is obtained through the following decision-making method: , where Threshold represents the preset deviation threshold.

[0085] It should be noted that this embodiment innovatively proposes an identification method based on time-varying sparse causal relationships to address the challenges of non-stationary, time-varying, and weak features in high-impedance fault signals in power systems. The core of this method lies in transforming the fault identification problem from traditional signal feature extraction into the quantification and comparison of the dynamic causal structure between multi-dimensional electrical features within the system. By constructing an end-to-end time-varying prediction model, this embodiment's method can dynamically capture and sparsely quantify Granger causal relationships between features, thereby effectively addressing the non-stationary characteristics of high-impedance fault signals and solving the fundamental problem that static or stationary assumption models fail in such cases. This embodiment also proposes a novel parameter update algorithm capable of effectively solving non-smooth sparse optimization problems within an online time-varying framework. This algorithm cleverly combines Taylor expansion and subgradient theory to provide a closed-form solution for updating causal matrix parameters including L1 regularization terms, thus overcoming the difficulty of introducing sparse constraints in most time-varying models. This design enables the model to automatically select variables, filter out redundant relationships, and extract key fault features while processing non-stationary data, significantly enhancing the model's interpretability and noise resistance.

[0086] In summary, this embodiment achieves high-precision and robust identification of high-resistance faults by uniformly handling non-stationarity and sparsity, and introducing a causal stability metric to further screen reliable connections. Its parameter update algorithm has a closed-form solution, ensuring computational efficiency and providing potential for online monitoring applications. Gradient pruning and L2 regularization ensure numerical stability, making the method more reliable in engineering practice. This embodiment provides a novel and more physically interpretable technical path for detecting highly concealed and high-risk high-resistance faults, and has significant application value for improving the safety level of power grid operation and intelligent operation and maintenance capabilities.

[0087] This embodiment also provides a high-resistance fault identification device for power distribution networks. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the terms "module" and "device" can refer to a combination of software and / or hardware that performs a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0088] According to an embodiment of the present invention, an apparatus embodiment for implementing the above-described high-resistance fault identification method for power distribution networks is also provided. Figure 2 This is a schematic diagram of the structure of a high-resistance fault identification device for a power distribution network according to an embodiment of the present invention, as shown below. Figure 2 As shown, the above-mentioned high-resistance fault identification device for power distribution networks includes: a current waveform acquisition module 200, a feature extraction module 202, an objective function construction module 204, an optimization module 206, and a fault identification module 208, wherein:

[0089] The current waveform acquisition module 200 is used to acquire the original zero-sequence current waveform of the distribution network line at the current moment.

[0090] The feature extraction module 202 is connected to the current waveform acquisition module 200 and is used to extract features from the original zero-sequence current waveform to obtain an actual feature sequence, wherein the actual feature sequence includes the actual feature values ​​of multiple features at the current time.

[0091] The objective function construction module 204, connected to the feature extraction module 202, is used to construct an objective function with the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence. The predicted feature sequence is obtained based on multiple historical feature sequences and an initial time-varying causal matrix. The elements in the initial time-varying causal matrix are used to indicate the contribution strength of the historical feature value of one feature to the predicted feature value of another feature.

[0092] Optimization module 206, connected to objective function construction module 204, is used to optimize the initial time-varying causal matrix based on the objective function to obtain the target time-varying causal matrix;

[0093] The fault identification module 208, connected to the optimization module 206, is used to determine the high-resistance fault identification result of the distribution network line based on the target time-varying causal matrix. The high-resistance fault identification result is used to indicate whether a high-resistance fault has occurred in the distribution network line at the current moment.

[0094] It should be noted that the above modules can be implemented by software or hardware. For example, for the latter, it can be implemented in the following ways: the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0095] It should be noted that the current waveform acquisition module 200, feature extraction module 202, objective function construction module 204, optimization module 206, and fault identification module 208 mentioned above correspond to steps S102 to S110 in the embodiments. The instances and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules, as part of the device, can run on a computer terminal.

[0096] It should be noted that the optional or preferred implementation methods of this embodiment can be found in the relevant descriptions in the embodiments, and will not be repeated here.

[0097] The aforementioned high-resistance fault identification device for power distribution networks may also include a processor and a memory. The aforementioned current waveform acquisition module 200, feature extraction module 202, objective function construction module 204, optimization module 206, fault identification module 208, etc., are all stored in the memory as program modules, and the processor executes the aforementioned program modules stored in the memory to realize the corresponding functions.

[0098] The processor contains a core that retrieves the corresponding program modules from memory. One or more cores may be configured. Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory includes at least one memory chip.

[0099] According to an embodiment of this application, an embodiment of a non-volatile storage medium is also provided. Optionally, in this embodiment, the non-volatile storage medium includes a stored program, wherein, when the program runs, it controls the device containing the non-volatile storage medium to execute any of the aforementioned high-resistance fault identification methods for power distribution networks.

[0100] Optionally, in this embodiment, the non-volatile storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals, and the non-volatile storage medium includes stored programs.

[0101] Optionally, a program may be used to control the device containing the non-volatile storage medium to execute any of the steps of the above-mentioned high-resistance fault identification method for power distribution networks during program execution.

[0102] According to an embodiment of this application, an embodiment of a processor is also provided. Optionally, in this embodiment, the processor is used to run a program, wherein the program executes any of the above-described high-resistance fault identification methods for power distribution networks.

[0103] According to an embodiment of this application, an embodiment of a computer program product is also provided, which, when executed on a data processing device, is adapted to execute a program that initializes the steps of the distribution network high-resistance fault identification method having any of the above-described steps.

[0104] This invention provides an electronic device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of any of the above-described methods for identifying high-resistance faults in power distribution networks.

[0105] The order of the above embodiments of the present invention is merely for description and does not represent the superiority or inferiority of the embodiments.

[0106] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0107] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of modules described above can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between modules, and may be electrical or other forms.

[0108] The modules described above as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0109] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0110] If the aforementioned integrated modules are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable non-volatile storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a non-volatile storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned non-volatile storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0111] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying high-resistance faults in a distribution network, characterized in that, include: Obtain the original zero-sequence current waveform of the distribution network line at the current moment; The original zero-sequence current waveform is subjected to feature extraction to obtain an actual feature sequence, wherein the actual feature sequence includes the actual feature values ​​of multiple features at the current time. An objective function is constructed with the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence. The predicted feature sequence is obtained based on multiple historical feature sequences and an initial time-varying causal matrix. The elements in the initial time-varying causal matrix are used to indicate the contribution strength of the historical feature value of one feature to the predicted feature value of another feature. Based on the objective function, the initial time-varying causal matrix is ​​optimized to obtain the target time-varying causal matrix; Based on the target time-varying causal matrix, the high-resistance fault identification result of the distribution network line is determined, including: correcting the elements in the target time-varying causal matrix from the time dimension to obtain a target causal stability matrix, wherein the elements in the target causal stability matrix represent the stability measure of any two features in the time series; determining the deviation between the target causal stability matrix and a preset benchmark causal stability matrix; and determining the high-resistance fault identification result based on the deviation, wherein the high-resistance fault identification result is used to indicate whether a high-resistance fault has occurred in the distribution network line at the current time.

2. The method according to claim 1, characterized in that, Before constructing the objective function with the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence, the method further includes: The predicted feature sequence is obtained as follows: ； in, This represents the predicted feature sequence; This represents the initial time-varying causal matrix, and the elements in the initial time-varying causal matrix are... Representation of features Historical eigenvalues ​​of features The contribution strength of the predicted eigenvalues; The historical information matrix is ​​represented, wherein the historical information matrix includes the plurality of historical feature sequences, T represents the total number of sampling times corresponding to the plurality of historical feature sequences, and t represents the current time; This represents the initial time-varying feature extraction matrix, where each column of the initial time-varying feature extraction matrix... This is used to learn how to optimally linearly combine the multiple historical feature sequences to extract the predictive value for the first... The most effective temporal dynamic component of the features, where k is any one of the features.

3. The method according to claim 2, characterized in that, The objective function, which aims to minimize the error between the actual feature sequence and the corresponding predicted feature sequence, is constructed as follows: The objective function is constructed as follows, with the goal of minimizing the error between the actual feature sequence and the predicted feature sequence: ； in, The value of the objective function is used to quantify the error between the actual feature sequence and the predicted feature sequence. This represents the first regularization parameter; This represents the second regularization parameter; This represents the prediction error term in the objective function; The term N represents the sparse term in the objective function, and N represents the total number of the multiple features included in the actual feature sequence. This represents the stable term in the objective function.

4. The method according to claim 1, characterized in that, The optimization of the initial time-varying causal matrix based on the objective function to obtain the target time-varying causal matrix includes: With the objective function value being minimized, the gradient descent method is used to update the initial time-varying feature extraction matrix. Each column of the initial time-varying feature extraction matrix is ​​used to learn how to optimally linearly combine the multiple historical feature sequences to extract the most effective time-domain dynamic component for predicting any feature. Taylor expansion is performed on the prediction error term in the objective function to obtain the processed function; The elements in the initial time-varying causal matrix are updated using a subgradient update rule to obtain the updated time-varying causal matrix; Repeat the above operation until the predetermined termination condition is met; The updated time-varying causal matrix obtained when the predetermined termination condition is met is taken as the target time-varying causal matrix.

5. The method according to claim 4, characterized in that, The step of updating the elements in the initial time-varying causal matrix using a subgradient update rule to obtain the updated time-varying causal matrix includes: Using the aforementioned subgradient update rule, the elements in the initial time-varying causal matrix are updated in the following manner to obtain the updated time-varying causal matrix: ； in, L represents the first regularization parameter in the objective function; L represents the prediction error term in the objective function. This represents the second regularization parameter in the objective function; Represents any element in the initial time-varying causal matrix; This represents the element corresponding to any of the elements in the previous time-varying causal matrix, where the previous time-varying causal matrix is ​​the time-varying causal matrix corresponding to the previous feature sequence of the actual feature sequence.

6. The method according to claim 1, characterized in that, The step of correcting the elements in the target time-varying causal matrix from the time dimension to obtain the target causal stability matrix includes: The target causal stability matrix is ​​obtained by modifying the elements in the target time-varying causal matrix as follows: ； in, Represents any element in the target causal stability matrix; Represents any element in the target time-varying causal matrix; The mean of any element and the corresponding elements in multiple historical time-varying causal matrices is given, wherein the multiple historical time-varying causal matrices correspond one-to-one with the multiple historical feature sequences; The standard deviation of any element and the corresponding element in the plurality of historical time-varying causal matrices; This indicates a preset positive integer.

7. The method according to claim 1, characterized in that, The determination of the high-resistance fault identification result based on the deviation includes: If the deviation is greater than a preset deviation threshold, the high-resistance fault identification result is determined to be that a high-resistance fault has occurred in the distribution network line at the current moment; or If the deviation is less than or equal to the preset deviation threshold, the high-resistance fault identification result is determined to be that the distribution network line has not experienced a high-resistance fault at the current time.

8. A high-resistance fault identification device for power distribution networks, characterized in that, include: The current waveform acquisition module is used to acquire the original zero-sequence current waveform of the distribution network line at the current moment; The feature extraction module is used to extract features from the original zero-sequence current waveform to obtain an actual feature sequence, wherein the actual feature sequence includes the actual feature values ​​of multiple features at the current time. The objective function construction module is used to construct an objective function with the goal of minimizing the error between the actual feature sequence and the corresponding predicted feature sequence. The predicted feature sequence is obtained based on multiple historical feature sequences and an initial time-varying causal matrix. The elements in the initial time-varying causal matrix are used to indicate the contribution strength of the historical feature value of one feature to the predicted feature value of another feature. An optimization module is used to optimize the initial time-varying causal matrix based on the objective function to obtain a target time-varying causal matrix; The fault identification module is used to determine the high-resistance fault identification result of the distribution network line based on the target time-varying causal matrix, including: correcting the elements in the target time-varying causal matrix from the time dimension to obtain a target causal stability matrix, wherein the elements in the target causal stability matrix represent the stability measure of any two features in the time series; determining the deviation between the target causal stability matrix and a preset benchmark causal stability matrix; and determining the high-resistance fault identification result based on the deviation, wherein the high-resistance fault identification result is used to indicate whether a high-resistance fault has occurred in the distribution network line at the current time.

9. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores multiple instructions, which are adapted to be loaded by a processor and executed by the distribution network high-resistance fault identification method according to any one of claims 1 to 7.

Images