Power distribution network grounding fault positioning method and system based on CNN-Transformer multi-layer feature fusion

Abstract

The application discloses a power distribution network grounding fault positioning method and system based on CNN-Transformer multi-layer feature fusion, utilizes a wide-frequency sensing node to perform high-fidelity collection on a power distribution network grounding fault transient current traveling wave signal, reduces wave head attenuation and high-frequency feature distortion caused by traditional insufficient bandwidth, constructs an equivalent fault modulus signal through phase-mode transformation, converts multi-measurement-point fault modulus signals into time-frequency diagrams, utilizes CNN to extract local fault features such as fault wave heads, high-frequency energy concentration areas, reflected wave textures and frequency attenuation, utilizes Transformer to fuse global propagation features such as traveling wave arrival sequences between multi-measurement points, multi-time slice correlations, multi-branch topological relationships and multi-band energy changes, and finally realizes intelligent output of power distribution network grounding fault sections, fault distances in the sections, whole-network topological positions and positioning reliabilities, and improves the accuracy and reliability of grounding fault positioning in a complex multi-branch power distribution network.

Description

Technical Field

[0001] This invention belongs to the field of smart grid fault location technology, specifically relating to a method and system for locating grounding faults in distribution networks based on CNN-Transformer multi-layer feature fusion. Background Technology

[0002] As the end point of the power system, the distribution network connects the transmission system with end users. Its safe and stable operation directly affects the reliability of power supply and power quality, and accurate and rapid fault location is crucial to ensuring its reliable operation. However, due to the current multi-branch radial structure of distribution network lines, complex operating modes, and inconsistent parameters among lines, faults occur frequently, with single-phase grounding faults accounting for a relatively high proportion. In particular, high-resistance grounding faults and arcing grounding faults are characterized by small fault currents, weak transient signals, indistinct fault characteristics, and significant influence from the neutral point grounding method of the system, making them difficult to identify in a timely and accurate manner, and resulting in large location errors.

[0003] If grounding faults are not detected and addressed in a timely manner, they can easily trigger arcing overvoltages, leading to damage to expensive electrical equipment such as transformers or rotating motors. At the same time, the voltage to ground of non-faulty phases can rise, easily breaking down the insulation at weak points in the system, potentially inducing severe phase-to-phase short circuits and causing large-scale power outages. In addition, large step voltages may be generated, posing a threat of electric shock to pedestrians, while continuous arcing may ignite surrounding objects, causing fires and seriously threatening the safe and stable operation of the power system and personal safety.

[0004] Existing methods for fault location in power distribution networks and their problems:

[0005] (1) Fault location method based on traditional traveling wave ranging: The time it takes for the traveling wave of the high-frequency transient current or voltage signal generated instantaneously at the fault point to reach the detection terminal is measured, and the distance between the fault point and the detection terminal is calculated by combining the known traveling wave velocity. It can be divided into single-end method and double-end method according to the number of detection terminals. Since this method still depends on the arrival time of the first wave in the fault location process, it depends on the accuracy of the wavefront calibration. However, since the distribution network branches are complex and radial, the fault traveling wave will be refracted and reflected when it reaches the branch point, and the first wavefront is easy to be mixed with the refracted and reflected waves, resulting in a large wavefront calibration error. Moreover, when high-resistance grounding or arc grounding faults occur, the traveling wave signal is weak, which greatly increases the difficulty of wavefront calibration and leads to a large location error. At the same time, considering that in actual operation, the propagation speed of the traveling wave may fluctuate due to factors such as line environment, operating temperature or equipment aging, the location error is large.

[0006] (2) Fault localization method based on time-frequency graph and convolutional neural network: Usually, traveling wave transient signal is first collected, and the one-dimensional fault waveform is converted into a two-dimensional time-frequency graph using methods such as Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT). Then, this signal is denoised and purified before being input into the CNN model. The CNN automatically extracts local fault features from the time-frequency graph through convolutional layers, pooling layers and fully connected layers. After sample training, the CNN can learn the mapping relationship between time-frequency graph features and fault location, thereby outputting the fault segment or estimating the fault distance. However, these methods have the following shortcomings: First, when the bandwidth of traditional signal detection devices is insufficient or the performance of acquiring high-frequency transient signals is poor, the fault wavefront and high-frequency energy characteristics may be attenuated or distorted, resulting in a severely distorted time-frequency spectrum, making it impossible for CNNs to extract effective fault features. Second, when the ground fault is a high-resistance ground fault or an arc ground fault, it has characteristics such as small fault current, weak transient signals, strong environmental noise, and wavefront aliasing caused by multiple branches of the line. The energy amplitude characteristics in the time-frequency spectrum are weak, and CNNs are easily affected by noise interference. Moreover, if complex operating conditions are not covered during training, CNN fault localization may have significant errors. Third, ground faults require the simultaneous fusion of multi-level global fault features such as local wavefronts, high-frequency energy, propagation sequence of multiple measurement points, and multi-branch relationships within a continuous time period. A single CNN model has limited ability to fuse these multi-level features, thus failing to accurately locate ground faults in the distribution network. Summary of the Invention

[0007] This invention provides a method and system for locating ground faults in distribution networks based on CNN-Transformer multi-layer feature fusion, which can improve the accuracy and reliability of ground fault location in complex multi-dominated power grids.

[0008] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0009] A method for locating ground faults in distribution networks based on CNN-Transformer multi-layer feature fusion includes:

[0010] Acquire the three-phase transient current signals at each measuring point within a preset time period before and after a ground fault occurs in the distribution network;

[0011] The three-phase transient currents at each measuring point are processed to construct an equivalent fault modulus signal, which is then converted into a time-frequency diagram.

[0012] The CNN is used to extract local fault features of each measurement point based on the time-frequency map. Then, the Transformer module is used to fuse the local fault features of all measurement points to obtain global features. Finally, the fusion module is used to fuse the global fused features with the local fault features of all measurement points to obtain fused features.

[0013] Using the trained fault segment prediction model and normalized distance prediction model, the fault segment and the normalized distance of the fault within the segment are output respectively based on the fusion features.

[0014] Furthermore, the measuring points are distributed at the feeder outlets, main lines, branch lines, switch nodes, cable terminals, and overhead lines and cable connection points of the distribution network. The power frequency component, low-frequency transient component, and MHz-level high-frequency transient component in the line current at each measuring point are collected synchronously using a current sensor based on tunnel magnetoresistive properties.

[0015] Furthermore, the three-phase transient currents are processed to construct an equivalent fault modulus signal, including:

[0016] First, Clarke phase-mode transformation is performed on the three-phase transient currents to obtain the zero-mode current component and two line-mode current components.

[0017] Then, the two line-mode current components are all converted into an equivalent line-mode traveling wave current;

[0018] The zero-mode current component is then combined with the equivalent line-mode traveling wave current to form an equivalent fault mode signal.

[0019] Furthermore, wavelet transform is used to perform time-frequency analysis on the equivalent fault modulus signal to obtain a time-frequency diagram.

[0020] Furthermore, the Transformer is used to fuse the local fault characteristics of all measurement points, including:

[0021] First, the local fault characteristics of the measuring point are converted into measuring point feature units:

[0022]

[0023] In the formula: Indicates the first Each measurement point corresponds to a measurement point feature unit; The feature mapping matrix, The measurement point location is encoded to indicate its relative position within the feeder. This is a topology code used to represent the line attributes of the measurement point;

[0024] Then, all the feature units of the measuring points are used to construct a measuring point feature sequence. :

[0025]

[0026] Then, the feature sequence of the measurement points Inputting into the Transformer module allows it to learn the relationships between different measurement points through a self-attention mechanism:

[0027]

[0028] In the formula, This represents the global feature output by the Transformer module.

[0029] Furthermore, the fusion module is used to fuse the global fusion features with the local fault features of all measurement points, represented as:

[0030]

[0031] In the formula: Indicates will and To splice, To fuse the weight matrix, It is a non-linear activation function. This is a feature of fusion.

[0032] Furthermore, a fault segment prediction model is used, and the fault segment is output based on the fused features, including:

[0033] The distribution network lines are divided into several sections, and the length of each section and the position of its starting point in the distribution network topology coordinates are recorded.

[0034] The fused features are input into the fault segment prediction model to obtain the probability that the fault is located in each segment. The segment with the highest probability is selected as the fault segment output, and the highest probability is also used as the fault location confidence output.

[0035] Furthermore, a normalized distance prediction model is used, and the normalized distance of the fault within the segment is output based on the fused features, including:

[0036] The fused features are input into the fault segment prediction model to obtain the normalized distance of the fault location relative to the starting point in each segment; then, the normalized distance corresponding to the fault segment is selected from these distances, and the final location of the fault point in the distribution network topology coordinates is output based on the length of the fault segment and the position of its starting point in the distribution network topology coordinates.

[0037] A distribution network grounding fault location system based on CNN-Transformer multi-layer feature fusion includes a TMR sensing node, a data transmission module, and a cloud-based fault location platform;

[0038] The TMR sensing node is used to: collect three-phase transient current signals within a preset time period before and after a ground fault occurs in the power distribution network at each measuring point;

[0039] The data transmission module is used to transmit the three-phase transient current signals collected at each measuring point to the cloud-based fault location platform.

[0040] The cloud-based fault location platform is used for: (1) processing the three-phase transient current of each measuring point, constructing an equivalent fault modulus signal, and converting it into a time-frequency diagram; (2) using CNN and extracting local fault features of each measuring point based on the time-frequency diagram, and then using the Transformer module to fuse the local fault features of all measuring points to obtain global features; and then using the fusion module to fuse the global fusion features with the local fault features of all measuring points to obtain fusion features; (3) using the fault segment prediction model and the normalized distance prediction model, and outputting the fault segment and the normalized distance of the fault in the fault segment according to the fusion features respectively.

[0041] Compared with existing methods for locating grounding faults in power distribution networks, the present invention has at least the following advantages and beneficial effects:

[0042] (1) Improve the broadband sensing capability of ground fault transient signals.

[0043] This invention employs a broadband sensing node based on tunnel magnetoresistive (TMR) to acquire transient current signals from grounding faults in distribution networks. This enables high-fidelity acquisition of high-frequency transient components and weak traveling wave characteristics in the initial stage of a fault, reducing wavefront attenuation, phase shift, and high-frequency energy distortion problems caused by insufficient bandwidth in traditional detection devices. Consequently, subsequent time-frequency map construction and intelligent location models can obtain more complete and reliable raw fault data, improving the effectiveness of the basic data for fault location.

[0044] (2) Enhance the time-frequency expression capability of weak transient fault characteristics.

[0045] This invention converts transient fault signals acquired by broadband sensing nodes into multi-point time-frequency maps, enabling features such as fault wavefronts, high-frequency energy concentration areas, reflected wave textures, and frequency attenuation in ground fault signals to be simultaneously represented in both time and frequency dimensions. Compared to methods relying solely on time-domain waveform analysis, this approach is more suitable for characterizing non-stationary, broadband, and weak transient fault signals such as high-resistance ground faults and arcing ground faults, providing a clearer input representation for subsequent CNN extraction of local fault features.

[0046] (3) Improve the ability to automatically extract local fault features.

[0047] This invention utilizes CNN to extract local fault features from time-frequency maps at multiple measurement points. It can automatically identify fault wavefront abrupt changes, high-frequency energy distribution, reflected and refracted wave textures, and frequency attenuation characteristics corresponding to different fault distances. By extracting local time-frequency features through CNN, errors caused by manually selecting wavefronts, setting feature values, or relying on single time-domain features can be reduced, improving the ability to identify local fault features in weak fault signals and complex noise environments.

[0048] (4) Improve the ability to model the global propagation relationship of multiple measurement points.

[0049] This invention further introduces a Transformer module to perform global feature fusion on the arrival order, energy attenuation relationship, branch reflection relationship, and topological position relationship of traveling waves among multiple measuring points. Compared with the single CNN model, which mainly focuses on local texture features, the Transformer can learn the correlation weights between different measuring points. This allows the model to not only focus on the waveform changes of a single measuring point, but also to comprehensively analyze the propagation law of fault traveling waves in the main line and branch lines, thereby improving the fault location accuracy in complex multi-dominated power grids.

[0050] (5) Improve the ability to fuse local features and global features for local positioning.

[0051] This invention employs a CNN-Transformer multi-layer feature fusion structure, which comprehensively expresses the local wavefront details, high-frequency energy distribution, and reflected wave texture features extracted by CNN with the multi-point propagation order, energy attenuation, and topological correlation features obtained by fusion with Transformer. This fusion method avoids both the insufficient global correlation modeling capability of a single CNN model and the weakening of local wavefront details by relying solely on the global model. This makes the final fused features more suitable for subsequent fault segment classification and intra-segment distance regression, thereby improving the reliability of multi-dominant power grid grounding fault location.

[0052] (6) Reduce the risk of positioning misjudgment caused by single distance output in multi-branch lines.

[0053] This invention employs a joint location method combining "fault segment classification + intra-segment distance regression." It first identifies candidate fault segments where the fault point is located, then predicts the relative position of the fault point within that segment, and finally combines the segment length and the overall network topology coordinates to obtain the final fault location. This method avoids the misjudgment problem of "same distance but different actual location" that occurs when only a single line distance is output in a multi-branch distribution network, making the location results more consistent with the characteristics of the multi-branch topology of the distribution network.

[0054] In summary, this invention constructs a complete ground fault location method based on TMR broadband sensing, CNN local fault feature extraction, Transformer multi-point global propagation relationship fusion, and CNN-Transformer multi-layer feature fusion. It can improve the sensing quality, feature expression capability, and location reliability of ground fault transient signals in distribution networks under complex multi-branch lines, and provide technical support for shortening fault investigation time, reducing manual line inspection workload, and improving the operational reliability of distribution networks. Attached Figure Description

[0055] Figure 1 This is an overall flowchart of the method described in the embodiments of this application.

[0056] Figure 2 This is a detailed flowchart of the method described in the embodiments of this application.

[0057] Figure 3 This is a schematic diagram of the distribution of TMR sensing nodes in the power distribution network lines in the embodiments of this application.

[0058] Figure 4 This is a CNN structure diagram in the embodiments of this application.

[0059] Figure 5 This is a schematic diagram of the Transformer module performing global feature fusion in an embodiment of this application.

[0060] Figure 6 This is a structural diagram of the system described in the embodiments of this application. Detailed Implementation

[0061] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0062] Example 1

[0063] This embodiment provides a method for locating grounding faults in distribution networks based on CNN-Transformer multi-layer feature fusion, referencing... Figure 1 , 2 As shown, it includes the following steps:

[0064] Step 1: Deploy broadband magnetoresistive sensing nodes in the tunnel.

[0065] (1) At locations such as feeder outlets, main lines, branch lines, switch nodes, cable terminals, and overhead line-to-cable connection points in the distribution network, such as Figure 3As shown, a broadband sensing node based on tunnel magnetoresistive (TMR) is installed to collect three-phase transient current signals of the power distribution network. The TMR broadband sensing node includes a TMR magnetic sensing unit, a signal conditioning module, a high-speed analog-to-digital converter module, an edge computing module, a time synchronization module, a communication module, and a power supply module. The TMR broadband sensing node is used to synchronously collect the power frequency component, low-frequency transient component, and MHz-level high-frequency transient component in the power distribution network line current. In this implementation, the node sampling frequency can be set to 10 MHz to meet the requirements for collecting the initial transient traveling wave and high-frequency energy characteristics of ground faults. The time synchronization module can use BeiDou satellite time synchronization or a precision clock synchronization method to provide a unified time reference for the TMR broadband sensing node and add timestamps to the collected data. The spacing between each TMR node is determined based on the line length, branch location, transient signal attenuation, and on-site noise level; preferably, the distance between adjacent measuring points does not exceed 10 km.

[0066] (2) Broadband Sensing Principle of TMR Sensor: The core magnetic sensing structure of the TMR sensor chip consists of a single magnetic tunnel junction (MTJ) composed of a free layer, a barrier layer, and a pinned layer, and multiple MTJs form a bridge structure. Let the excitation voltage of the TMR chip be... The voltages at the two output terminals are respectively and . and The magnetic induction intensity at the chip location is The angle between the direction of magnetic induction intensity and the direction of magnetic sensitivity of the chip is Then the TMR chip output differential voltage is:

[0067]

[0068] In the formula: This is the differential output voltage; The sensitivity coefficient of the TMR chip, expressed in V / V / T, represents the output voltage value of the sensor under unit magnetic flux density and excitation source voltage.

[0069] In this invention, the TMR chip is mounted in a fixed orientation, such that the magnetic sensitivity direction is parallel to the direction of the magnetic flux density generated by the current-carrying conductor. Combined with the Biot-Savart law To obtain the current in the current-carrying conductor Relationship with the differential output voltage of the TMR chip:

[0070]

[0071] In the formula, This represents the shortest distance from the current-carrying conductor to the magnetically sensitive point of the TMR chip. is the vacuum permeability.

[0072] From this formula, it can be seen that when the TMR chip operates in the linear response range and the relative positions of the sensor and the current-carrying conductor are fixed, the primary current... With output differential voltage The relationship is approximately linear. Therefore, this invention can obtain the primary side current by measuring the differential voltage output by the TMR chip.

[0073] (3) Assume that the distribution network is laid out with There are three measuring points, each equipped with three sets of TMR sensor units, which collect the current of the three phases A, B, and C wires respectively. Each measuring point is denoted as:

[0074]

[0075] The three-phase currents collected by the TMR node are:

[0076]

[0077] In the formula, Indicates the measurement point number. These represent phases A, B, and C, respectively.

[0078] Step 2: The TMR sensor detects the fault trigger and records transient waveforms.

[0079] The TMR node continuously acquires three-phase current signals at a uniform sampling frequency, and the edge processing unit calculates current surges, high-frequency energy, or zero-sequence changes in real time. When a transient surge is detected exceeding a set threshold, a suspected ground fault is identified, and waveform recording is triggered. To preserve the complete transient process before and after the fault and avoid losing the initial waveform information by only recording data after the fault, the waveform recording window must include a pre-fault time window and a post-fault time window.

[0080]

[0081] In the formula, For the trigger time, This is a time limit to hold before a failure occurs. Retention time after a failure.

[0082] Step 3: Synchronize and process the collected three-phase current signals to construct an equivalent fault modulus signal.

[0083] Three-phase current signals collected from multiple TMR measurement points are uploaded to the edge terminal or main station after being marked with a unified timestamp. The system aligns the data measured at different measurement points according to the timestamps of each measurement point. Subsequently, the three-phase current at each measurement point is preprocessed, including: resampling at a unified sampling rate, truncating data according to the same fault time window, removing DC bias, eliminating outliers, and normalization. Finally, an equivalent single-channel fault modulus signal is constructed based on phase-mode transformation.

[0084] (1) Phase mode transformation:

[0085] For the Clarke phase-mode transformation is performed on the three-phase current at each measuring point:

[0086]

[0087] In the formula: For the first Zero-mode current component at each measuring point; The first Two line-mode current components at each measuring point. The zero-mode component is obtained as follows:

[0088]

[0089] (2) Constructing the equivalent line-mode traveling wave current components:

[0090] This invention further combines the two line-mode components into an equivalent line-mode traveling wave current:

[0091]

[0092] In the formula, For the first The equivalent line-mode traveling wave current at each measuring point.

[0093] (3) The zero-mode and the linear-mode signal are combined to form an equivalent fault mode signal:

[0094]

[0095] In the formula: For the first Each measurement point is ultimately input into the equivalent fault modulus signal of the subsequent model.

[0096] Through the above processing, the zero-mode characteristics of grounding faults and the line-mode characteristics of traveling wave propagation can be preserved at the same time, and the three-phase transient currents can be converted into a unified equivalent single-channel input signal, reducing the difficulty of subsequent model learning.

[0097] Step 4: Construct a multi-point single-modulus time-frequency diagram from the equivalent fault modulus signal.

[0098] Because the transient signal of a single-phase ground fault is non-stationary, has a wide bandwidth, and exhibits strong abrupt changes, it is difficult to fully reflect the time-frequency variation law of the fault traveling wave by analyzing it only in the time domain. Therefore, in this embodiment, after obtaining the equivalent fault modulus signal at each measuring point, the wavelet transform method is used to perform time-frequency analysis on it, converting the one-dimensional transient fault signal into a time-frequency energy map that can simultaneously characterize time variation and frequency distribution.

[0099] Specifically, for the first The equivalent fault modulus signal at each measurement point is processed by wavelet transform to obtain the time-frequency energy map corresponding to that measurement point:

[0100]

[0101] The time-frequency energy maps of multiple measurement points are stacked sequentially according to their spatial distribution information to form an equivalent fault time-frequency signal input for multiple measurement points:

[0102]

[0103] In the formula: These represent time and frequency, respectively, corresponding to the scaling factor and time shift factor in wavelet transform.

[0104] Step 5: Input the time-frequency graph into the convolutional neural network to extract local fault features.

[0105] like Figure 4 As shown, the Convolutional Neural Network (CNN) in this embodiment mainly includes an input layer, a combination of multiple convolutional layers, activation layers, and pooling layers, a fully connected layer, and an output layer. The core of a CNN is the convolution operation. Input time-frequency diagram of each measurement point The convolution operation formula is as follows:

[0106]

[0107] In the formula: For the first Layer output features, For convolution kernel, For bias, The activation function, where * represents convolution operation.

[0108] CNN output number Local fault characteristics of individual measuring points For all measuring points, we get:

[0109]

[0110] This embodiment uses a convolutional neural network to specifically extract local features of reflected waves and refracted waves, high-frequency transient energy concentration features, weak traveling wave features of high-resistance grounding and arc grounding, and frequency attenuation features corresponding to different fault distances in the fault transient state, thereby obtaining the local fault features of each measuring point.

[0111] Step 6: Input the fault local features extracted by the CNN into the Transformer and fuse them with the global propagation relationships. For example... Figure 6 As shown.

[0112] While CNNs can extract local features from a single measuring point, fault location in a distribution network depends not only on the waveform of a single measuring point but also on the arrival order, propagation path, branch reflection relationships, and topological positional relationships of multiple measuring points. Therefore, this embodiment inputs the local features of each measuring point output by the CNN into the Transformer module.

[0113] (1) First, the first The CNN local fault features of each measurement point are converted into measurement point feature units:

[0114]

[0115] In the formula: Indicates the first Each measuring point corresponds to a measuring point feature unit, which is used to represent the relative position of the measuring point in the feeder and the line topology attributes; This is the feature mapping matrix, belonging to the feature embedding layer before the Transformer input, used to convert the local features of the CNN at each measurement point into measurement point feature units of uniform dimension; The measurement point location is encoded to indicate its relative position within the feeder. It is a topology code used to represent the line attributes of the measuring point, such as the main line, branch line, cable segment, or overhead line segment.

[0116] All feature units at the measurement points constitute the Transformer input sequence:

[0117]

[0118] (2) The feature sequence of the measurement points Inputting into the Transformer module allows it to learn the relationships between different measurement points through a self-attention mechanism:

[0119]

[0120] In the formula, This represents the global fusion feature output by the Transformer.

[0121] Through a self-attention mechanism, the Transformer can automatically learn the correlation weights between different measurement points. For example, when a fault occurs between measurement points 2 and 3, the time-frequency plots of measurement points 2 and 3 will show a strong correlation, while measurement points far from the fault point may exhibit delayed arrival, energy attenuation, or more complex reflected waves. This embodiment uses the Transformer to automatically capture the correlation weights between different measurement points through a self-attention mechanism.

[0122] Global fusion features Instead of representing only the local waveform at a single measuring point, it represents the global characteristics of the fault response at multiple measuring points in the integrated distribution network. Specifically, this includes: the arrival order of traveling waves at different measuring points, the high-frequency energy attenuation relationship at different measuring points, the topological correlation between the main line and branch lines, the propagation laws of reflected and refracted waves among multiple measuring points, and the global time-frequency distribution pattern corresponding to different fault sections.

[0123] Step 7, multi-layer feature fusion.

[0124] To prevent the local wavehead detail information extracted by CNN from being weakened during the global fusion process of Transformer, this embodiment sets up a multi-layer feature fusion structure to fuse the local features of CNN with the global features of Transformer to obtain fused features. :

[0125]

[0126] In the formula: Indicates the characteristics of local faults It is then combined with global fusion features. It is a trainable weight parameter matrix. It is a non-linear activation function.

[0127] Fusion features It is not a single current waveform characteristic, but a comprehensive fault characterization formed by "wideband sensing information, phase-mode decoupling information, local time-frequency information, and global propagation information." Specifically, First, it includes high-fidelity fault transient broadband current information acquired by a TMR sensor, preserving high-frequency abrupt changes and weak traveling wave details in the initial traveling wave of a single-phase ground fault. Second, it includes zero-mode grounding characteristics and equivalent line-mode traveling wave characteristics obtained through phase-mode transformation, where the zero-mode component reflects the ground fault attributes and the line-mode component reflects the fault traveling wave propagation characteristics. Third, it includes locally sensitive features extracted from the time-frequency plot by a CNN, including wavefront abrupt changes, high-frequency energy concentration areas, reflected wave textures, and frequency attenuation features. Finally, it includes multi-measurement point global correlation features obtained by Transformer fusion, including the traveling wave arrival order, energy attenuation relationship, topological position relationship, and reflection and refraction relationship between the main line and branch lines among different measurement points. Therefore, the fused features... Simultaneously, it characterizes "whether the fault is grounded, how the traveling wave propagates, how the local waveform changes, and how the measurement points are related," serving as a comprehensive positioning basis for subsequent fault segment identification and fault distance regression.

[0128] Step 8: Fault segment classification and intra-segment distance regression.

[0129] This embodiment employs a joint location method combining "fault segment classification + intra-segment distance regression". Since distribution networks typically consist of main lines, branch lines, and multiple measurement points distributed across the network, if the model only outputs a distance value relative to the beginning of a line, erroneous judgments may occur on different branch lines where "the distance is the same but the actual location is different". Therefore, this invention preferably uses normalized intra-segment distance as the regression target; that is, the model does not directly output "how many meters", but instead first outputs the relative proportion of the fault point located within that segment.

[0130] (1) First, based on the distribution network topology, the installation locations of TMR measuring points, branch node locations, switch node locations, and line end locations, the entire distribution line is divided into several candidate fault sections. Each candidate fault section can be a line between two adjacent TMR measuring points, or a line between a TMR measuring point and a branch node, a branch node and the line end, or a main line node and a branch line node. All candidate fault sections are represented as follows:

[0131]

[0132] in, Indicates the first One candidate fault section, This represents the total number of candidate segments. For each candidate segment... Record the segment length in advance. The position of the segment's starting point in the network topology coordinates .

[0133] After TMR broadband sensing, phase mode transformation, equivalent fault modulus construction, time-frequency graph generation, CNN local feature extraction, and Transformer global feature fusion, the model obtains the fused features. This fusion feature They are simultaneously sent to the fault section classification branch and the intra-section distance regression branch.

[0134] (2) Fault section classification branch output predicts fault sections:

[0135]

[0136] in, Indicates the fault is located at the first Candidate segments The probability within the range. The model selects the segment with the highest probability as the predicted fault segment:

[0137]

[0138]

[0139] in, To predict the fault section number, To predict faulty sections.

[0140] (3) The normalized distance predicted by the regression branch within the segment is used to calculate the specific fault location. The normalized distance prediction values ​​for each candidate segment are as follows:

[0141]

[0142] in, This is the normalized distance prediction vector within the segment. The sigmoid function is used to limit the output to a certain value. Within the range, This is the weight matrix of the distance regression branches within the segment. This is a bias term. If there are multiple... If there are candidate segments, then:

[0143]

[0144] in, This indicates that the model predicts the fault point to be located at the th When within a candidate segment, the normalized distance from the fault point to the starting point of that segment.

[0145] During online location, only the predicted fault section is considered. Corresponding normalized distance Then, it is converted into the actual distance from the fault point to the starting point of the predicted section:

[0146]

[0147] in, To predict fault sections The actual length, This represents the distance from the fault point to the starting point of the section.

[0148] Finally, the location of the predicted segment's starting point in the network topology coordinates is considered. The final location of the fault point in the entire network is obtained:

[0149]

[0150] in, This indicates the final location of the fault point in the distribution network topology coordinates.

[0151] The final model output is:

[0152]

[0153] in, To predict faulty sections, is the distance from the fault point to the starting point of the predicted segment, and is the coordinate position of the fault point in the entire network topology. For location reliability.

[0154] Location confidence can be represented by the maximum probability in the fault segment classification branch:

[0155]

[0156] In this way, the present invention first uses segment classification to solve the problem of "which segment of the line is the fault", and then uses intra-segment distance regression to solve the problem of "the specific location of the fault within the segment", thereby avoiding the location error caused by using only a single distance regression in a multi-control power grid.

[0157] The CNN, Transformer module, fusion module, fault segment prediction model, and normalized distance prediction model used in this invention are first used to construct a training sample library using simulation data, historical fault waveform data, or field test data before being put into actual operation. Each training sample includes three-phase transient current signals collected from multiple measurement points, as well as the corresponding real fault segment and real fault distance.

[0158] During training, the three-phase transient currents collected from each measuring point are first preprocessed to extract fault transient features and generate feature maps or feature vectors suitable for neural network input. These features are then input into the CNN-Transformer model, enabling the model to learn the correspondence between fault signals from multiple measuring points and fault locations.

[0159] For each training sample, the label consists of two parts: the first part is the fault segment label, which tells the model which line segment the actual fault occurred in; the second part is the normalized distance label within the segment, which tells the model the relative location of the fault point within that segment.

[0160] Let the actual fault section length be... The actual distance from the fault point to the starting point of this section is Then the normalized distance within the segment is:

[0161]

[0162] in, The value range is from 0 to 1. For example, This indicates that the fault point is located at 30% of the distance from the start to the end of the section.

[0163] During model training, efforts are made to ensure that the classification results of fault sections are as close as possible to the actual fault sections, and to ensure that the normalized distance output by the model is as close as possible to the actual normalized distance. After training with a large number of samples under different fault locations, fault types, transition resistances, fault initial angles, line topologies, and noise conditions, the model is able to establish a mapping relationship between fault transient features and fault sections and distances within sections.

[0164] After training, the model parameters are fixed and used for online fault location. When a fault occurs in an actual line, the real-time acquired transient current signals from multiple measurement points are input into the trained model. The model first outputs the predicted fault section, then the normalized distance within that section. Finally, based on the actual length of the predicted fault section, the normalized distance is converted into the actual fault distance.

[0165]

[0166] in, To predict the actual length of the faulty section, The normalized distance output by the model. This represents the actual distance from the fault point to the starting point of the section.

[0167] Through the training method described above, the model can be used in this step to achieve joint prediction from transient fault signals at multiple measurement points to fault sections and specific locations within those sections, thereby completing the fault location in the distribution network.

[0168] Example 2

[0169] This embodiment provides a power distribution network grounding fault location system based on CNN-Transformer multi-layer feature fusion, such as... Figure 6 As shown, it includes a TMR sensing node, a data transmission module, and a cloud-based fault location platform, used to implement the power distribution network grounding fault location system based on CNN-Transformer multi-layer feature fusion as described in Example 1.

[0170] The TMR sensing node is used to: collect three-phase transient current signals within a preset time period before and after a ground fault occurs in the power distribution network at each measuring point;

[0171] The data transmission module is used to transmit the three-phase transient current signals collected at each measuring point to the cloud-based fault location platform.

[0172] The cloud-based fault location platform is used for: (1) processing the three-phase transient current of each measuring point, constructing an equivalent fault modulus signal, and converting it into a time-frequency diagram; (2) using CNN and extracting local fault features of each measuring point based on the time-frequency diagram, and then using the Transformer module to fuse the local fault features of all measuring points to obtain global features; and then using the fusion module to fuse the global fusion features with the local fault features of all measuring points to obtain fusion features; (3) using the fault segment prediction model and the normalized distance prediction model, and outputting the fault segment and the normalized distance of the fault in the fault segment according to the fusion features respectively.

[0173] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, these changes or improvements should fall within the scope of protection claimed in this application.

Claims

1. A method for locating grounding faults in distribution networks based on CNN-Transformer multi-layer feature fusion, characterized in that, include: Acquire the three-phase transient current signals at each measuring point within a preset time period before and after a ground fault occurs in the distribution network; The three-phase transient currents at each measuring point are processed to construct an equivalent fault modulus signal, which is then converted into a time-frequency diagram. The CNN is used to extract local fault features of each measurement point based on the time-frequency map. Then, the Transformer module is used to fuse the local fault features of all measurement points to obtain global features. Finally, the fusion module is used to fuse the global fused features with the local fault features of all measurement points to obtain fused features. Both the fault segment prediction model and the normalized distance prediction model are used to output the fault segment and the normalized distance of the fault within the fault segment, respectively, based on the fusion features.

2. The method for locating grounding faults in distribution networks based on CNN-Transformer multi-layer feature fusion according to claim 1, characterized in that, The measuring points are distributed at the feeder outlets, main lines, branch lines, switch nodes, cable terminals, and overhead lines and cable connection points of the distribution network. The power frequency component, low-frequency transient component, and MHz-level high-frequency transient component in the line current at each measuring point are collected synchronously using a current sensor based on tunnel magnetoresistive properties.

3. The method for locating grounding faults in distribution networks based on CNN-Transformer multi-layer feature fusion according to claim 1, characterized in that, The three-phase transient currents are processed to construct an equivalent fault modulus signal, including: First, Clarke phase-mode transformation is performed on the three-phase transient currents to obtain the zero-mode current component and two line-mode current components. Then, the two line-mode current components are all converted into an equivalent line-mode traveling wave current; The zero-mode current component is then combined with the equivalent line-mode traveling wave current to form an equivalent fault mode signal.

4. The method for locating grounding faults in distribution networks based on CNN-Transformer multi-layer feature fusion according to claim 1, characterized in that, Wavelet transform is used to perform time-frequency analysis on the equivalent fault modulus signal to obtain a time-frequency diagram.

5. The method for locating grounding faults in distribution networks based on CNN-Transformer multi-layer feature fusion according to claim 1, characterized in that, The Transformer is used to fuse the local fault characteristics of all measurement points, including: First, the local fault characteristics of the measuring point are converted into measuring point feature units: ； In the formula: Indicates the first Each measurement point corresponds to a measurement point feature unit; The feature mapping matrix, The measurement point location is encoded to indicate its relative position within the feeder. This is a topology code used to represent the line attributes of the measurement point; Then, all the feature units of the measuring points are used to construct a measuring point feature sequence. : ； Then, the feature sequence of the measurement points Inputting into the Transformer module allows it to learn the relationships between different measurement points through a self-attention mechanism: ； In the formula, This represents the global feature output by the Transformer module.

6. The method for locating grounding faults in distribution networks based on CNN-Transformer multi-layer feature fusion according to claim 1, characterized in that, The fusion module is used to fuse the global fusion features with the local fault features of all measurement points, represented as follows: ； In the formula: Indicates will and To splice, To fuse the weight matrix, It is a non-linear activation function. This is a feature of fusion.

7. The method for locating grounding faults in distribution networks based on CNN-Transformer multi-layer feature fusion according to claim 1, characterized in that, The fault segment prediction model is used, and the fault segments are output based on the fused features, including: The distribution network lines are divided into several sections, and the length of each section and the position of its starting point in the distribution network topology coordinates are recorded. The fused features are input into the fault segment prediction model to obtain the probability that the fault is located in each segment. The segment with the highest probability is selected as the fault segment output, and the highest probability is also used as the fault location confidence output.

8. The method for locating grounding faults in distribution networks based on CNN-Transformer multi-layer feature fusion according to claim 7, characterized in that, The normalized distance prediction model is used, and the normalized distance of the fault within the segment is output based on the fused features, including: The fused features are input into the fault segment prediction model to obtain the normalized distance of the fault location relative to the starting point in each segment; then, the normalized distance corresponding to the fault segment is selected from these distances, and the final location of the fault point in the distribution network topology coordinates is output based on the length of the fault segment and the position of its starting point in the distribution network topology coordinates.

9. A distribution network grounding fault location system based on CNN-Transformer multi-layer feature fusion, characterized in that, This includes TMR sensing nodes, data transmission modules, and a cloud-based fault location platform; The TMR sensing node is used to: collect three-phase transient current signals within a preset time period before and after a ground fault occurs in the power distribution network at each measuring point; The data transmission module is used to transmit the three-phase transient current signals collected at each measuring point to the cloud-based fault location platform. The cloud-based fault location platform is used for: (1) processing the three-phase transient current of each measuring point, constructing an equivalent fault modulus signal, and converting it into a time-frequency diagram; (2) using CNN and extracting local fault features of each measuring point based on the time-frequency diagram, and then using the Transformer module to fuse the local fault features of all measuring points to obtain global features; and then using the fusion module to fuse the global fusion features with the local fault features of all measuring points to obtain fusion features. (3) Using the fault segment prediction model and the normalized distance prediction model, the fault segment and the normalized distance of the fault within the fault segment are output respectively based on the fusion features.

Images