A prototype meta-learning-based distribution network single-phase ground fault line selection method and system

Abstract

A single-phase ground fault line selection method and system based on prototype meta-learning are provided, zero sequence voltage and feeder zero sequence current sequences are intercepted at the time of fault; manual feature vectors and deep feature vectors of the feeder zero sequence voltage and zero sequence current are extracted, and a fault embedding vector is generated through gate fusion; the process of obtaining the fault embedding vector is constituted as an embedding function, a binary classification meta task is constructed by using a fault waveform library in an offline stage, and the embedding function is trained through meta learning to obtain the ability of fault discrimination and classification based on prototypes; in online application, an initial prototype is established by using a sample of a target distribution network, the fault embedding vectors of each feeder are compared with each prototype, and whether a fault occurs is judged according to the distance. The present application can effectively overcome the problems of distribution network topology change, sample scarcity and class imbalance, has high line selection accuracy, can quickly adapt to new scenes with only a few samples, and has outstanding generalization ability and engineering practicability.

Description

Technical Field

[0001] This invention belongs to the field of power system distribution network relay protection and fault diagnosis technology, and particularly relates to a method and system for selecting the line of single-phase grounding fault in distribution network based on prototype meta-learning. Background Technology

[0002] In distribution networks with ungrounded neutral points, grounded via arc suppression coils, or grounded via high resistance, single-phase grounding faults account for more than 80% of all faults. Due to the high impedance of the fault circuit, the weak fault current (especially when grounded via high resistance), the strong nonlinearity of the arcing process, and the frequent changes in system operation modes and line topology adjustments, accurately and reliably selecting the faulty line has always been a core challenge for distribution network protection.

[0003] The existing mainstream line selection methods can be categorized into the following types, but all of them have obvious limitations: Steady-state electrical quantity methods include zero-sequence current amplitude comparison, group amplitude and phase comparison, and zero-sequence active power direction method. These methods rely on the steady-state components of the fault, but the steady-state zero-sequence current is extremely small when a single-phase grounding occurs in a non-effectively grounded system, and is affected by the compensation current of the arc suppression coil. The difference in steady-state characteristics between healthy and faulty lines is blurred, and these methods almost fail when there is high resistance grounding.

[0004] Transient signal analysis methods utilize transient zero-sequence current and voltage signals from one to several power frequency cycles after a fault, extracting features through the first half-wave direction, wavelet packet energy, transient zero-sequence current polarity comparison, and S-transform. The criteria for these methods are mostly based on manually set fixed thresholds or simple rules. When the distribution network structure changes (e.g., line switching), the neutral grounding method is adjusted, the transition resistance changes significantly, or the initial fault angle becomes random, the time-frequency distribution of the transient signal changes significantly. The original thresholds and rules easily become invalid, and these methods lack self-learning and self-adjustment capabilities.

[0005] Traditional machine learning methods utilize support vector machines (SVMs) and feedforward neural networks to classify extracted feature vectors (see the existing technology CN113092934B, which discloses a single-phase grounding fault determination method and system based on clustering and LSTM, employing Fourier harmonic features + LSTM network training for line selection). This method requires a large number of labeled fault samples covering various operating conditions for training. However, fault waveform data in actual distribution networks is scarce and extremely unevenly distributed; when new lines are added to the power grid or the operating mode undergoes significant changes, the original training samples no longer become representative, requiring the model to collect massive amounts of new data and retrain entirely, resulting in extremely high transfer costs and severely insufficient generalization ability.

[0006] Deep learning end-to-end methods: These methods directly use the original sampled sequences as input and automatically extract and classify features using convolutional neural networks (see the prior art CN118655416A, which discloses an intelligent line selection method for single-phase grounding faults in distribution networks with transfer learning capabilities, disclosing a transient time-domain feature + residual neural network fusion network and transfer learning adaptation; see the prior art (Du Y et al., IEEE PESGM 2019-2020 "Single Line-to-Ground Faulted Line Detection…Based on Feature Fusion Framework"), which discloses a one-dimensional CNN feature fusion framework for line selection in resonant grounding systems). These methods also rely on large-scale training datasets and have difficulties in generalizing across distribution network topologies and adapting to small samples. They also have weak physical interpretability and are susceptible to noise interference.

[0007] Meta-learning, as a "learning-to-learn" paradigm, enables models to quickly adapt to new tasks with very few new samples by training on a large number of tasks (see the prior art CN119782938A, which discloses a method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion, introducing metric meta-learning and class prototype calculation into high-resistance grounding fault identification, but the object is fault / normal binary classification identification rather than multi-feeder fault selection, and it lacks dual-channel voltage-current input, manual feature and deep feature gating fusion, N-feeder embedding vector and positive and negative prototype distance comparison, and bus fault discrimination logic). This characteristic perfectly matches the reality of variable operating conditions and scarce fault samples in distribution networks. However, directly applying meta-learning to one-dimensional transient waveform line selection faces challenges: the transferable general knowledge contained in a single one-dimensional current or voltage sequence is limited, making it difficult to support the meta-learner to quickly form an accurate discrimination prototype in new scenarios. In addition, in reality, fault samples are far fewer than non-fault samples, the classes are severely imbalanced, and traditional classifiers tend to favor the majority class, leading to high false negatives.

[0008] Therefore, there is an urgent need for a binary classification line selection scheme that can extract stable, rich, and multi-dimensional physical and data-driven fusion features from one-dimensional signals, and on this basis, achieve rapid cross-scenario generalization through a meta-learning framework, while naturally handling the problem of imbalanced samples. Summary of the Invention

[0009] This invention aims to solve the technical challenges of existing technologies in single-phase ground fault location in non-effectively grounded distribution networks. These challenges arise under complex conditions such as changes in distribution network topology, switching of neutral grounding methods, wide variations in transition resistance, random fault initial angles, and a limited number of available historical fault samples. These conditions result in significantly reduced accuracy, the need for massive retraining of models, and poor adaptability. The invention provides a method and system that requires only a single-ended one-dimensional transient signal to achieve high-precision and high-reliability fault location under multiple operating conditions and limited sample sizes, and can continuously evolve online.

[0010] The present invention adopts the following technical solution.

[0011] This invention proposes a method for selecting the line of a single-phase grounding fault in a distribution network based on prototype meta-learning, comprising the following steps: The zero-sequence voltage of the distribution network bus and the zero-sequence current of each feeder are collected in real time. When the sudden change in zero-sequence voltage or the effective value of power frequency exceeds the preset threshold, a single-phase ground fault is determined to have occurred. The fault start time is locked, and the zero-sequence voltage sequence of the bus and the zero-sequence current sequence of each feeder are extracted within a preset time window after the fault. For each feeder, based on the feeder zero-sequence current sequence and the bus zero-sequence voltage sequence, statistical and physical features are extracted from four dimensions: time domain, frequency domain, time-frequency domain, and nonlinearity. The extracted statistical and physical features are concatenated and normalized to obtain the hand-crafted feature vector of the feeder. A dual-channel one-dimensional deep feature extraction network is constructed, and the deep feature vectors of the zero-sequence current sequence and the zero-sequence voltage sequence of each feeder are obtained using the dual-channel one-dimensional deep feature extraction network; the deep feature vectors represent the abstract temporal features in the zero-sequence current sequence and the zero-sequence voltage sequence of the bus. A gated fusion unit is used to fuse the manual feature vector with the deep feature vector to obtain the fault embedding vector of the feeder; The process of obtaining fault embedding vectors constitutes an embedding function. In the offline stage, a fault waveform library including various distribution network operating states is used to construct a binary classification meta-task. Through meta-learning training, the embedding function acquires the ability to perform fault discrimination and classification based on prototypes, and the parameters of the embedding function are updated. When applied online, an initial prototype is established and stored using samples from the target distribution network. When a real fault occurs, the zero-sequence current sequence of each feeder and the zero-sequence voltage sequence of the bus are mapped to a fault embedding vector through the updated embedding function. The distance to each prototype is calculated, and the faulty line or bus is determined based on the distance comparison results.

[0012] More preferably, the method for extracting statistical and physical features includes: For a given length of Given a sequence, define the mean, standard deviation, effective value, mean absolute value, and maximum absolute value, and perform a fast Fourier transform on the sequence to obtain the amplitude spectrum and corresponding frequencies; The time-domain features are calculated based on the mean, standard deviation, effective value, mean absolute value, maximum absolute value, amplitude spectrum and corresponding frequency, including peak factor, impulse factor, kurtosis, skewness, waveform factor and margin factor; Frequency domain characteristics include spectral centroid, frequency variance, 5th harmonic amplitude percentage, and 7th harmonic amplitude percentage; Time-frequency domain features include wavelet packet energy entropy calculated after performing a three-level decomposition of the sequence using db4 wavelet packets; Nonlinear features include sample entropy calculated under a given embedding dimension and a given similarity tolerance.

[0013] More preferably, the structure of the dual-channel one-dimensional deep feature extraction network is as follows: the bus zero-sequence voltage sequence and the zero-sequence current sequence of each feeder. Input layer, shape is The first channel is the feeder zero-sequence current sequence, and the second channel is the bus zero-sequence voltage sequence. Three cascaded one-dimensional convolutional blocks, each containing a one-dimensional convolutional layer, batch normalization, ReLU activation, and max pooling; A bidirectional long short-term memory network is used to perform temporal modeling of the time series output of convolution; From the attention pooling layer, the attention weights of the outputs of the bidirectional long short-term memory network at each time step are calculated and weighted summed to obtain the context vector; A fully connected projection layer maps the context vector to a deep feature vector.

[0014] More preferably, the processing of the gating fusion unit includes: The handcrafted feature vectors are mapped through a fully connected layer to obtain the mapping result; the deep feature vectors and the mapping result are concatenated and input into a fully connected layer, and activated by Sigmoid to obtain the gated vector; the fault embedding vector is calculated based on the gated vector, the handcrafted feature vectors, and the deep feature vectors.

[0015] More preferably, the process of obtaining the fault embedding vector constitutes an embedding function. In the offline stage, a binary classification meta-task is constructed using a fault waveform library encompassing various distribution network operating states. Through meta-learning training, the embedding function acquires the ability to classify faults based on prototypes, and the parameters of the embedding function are updated. Specific steps include: The process of obtaining fault embedding vectors constitutes the embedding function; in the offline stage, a fault waveform library including various distribution network operating states is used to construct a binary classification meta-task; in the support set and query set of each meta-task, positive class samples correspond to faulty feeders, and negative class samples correspond to non-faulty feeders, and the number of positive and negative samples in the support set is equal; positive class prototypes and negative class prototypes are calculated on the support set respectively, and the classification probability and cross-entropy loss are calculated on the query set based on the Euclidean distance between the query sample embedding and the two class prototypes; the embedding function parameters are updated through gradient descent, so that the embedding function can quickly establish fault / non-fault discrimination prototypes with only a small number of samples; The parameters of the embedding function are updated with cross-entropy loss as the optimization objective.

[0016] More preferably, the operating status of the distribution network includes: Changes in neutral grounding method, changes in line commissioning / decommissioning combination, changes in transition resistance value, changes in initial phase angle of fault, and changes in fault location are one or more of the following combinations.

[0017] More preferably, the specific method for determining the faulty line or bus is as follows: For each feeder, calculate the Euclidean distance between its fault embedding vector and the positive class prototype as the first distance, and the Euclidean distance with the negative class prototype as the second distance. If the first distance is not greater than the second distance, the feeder is determined to be a faulty line; If the first distance of all feeders is greater than the second distance, then it is determined to be a bus fault.

[0018] More preferably, the method further includes an online adaptive update step: When new operating conditions occur in the distribution network, the model is updated using newly emerging fault and non-fault samples through one of the following methods: The embedding vector of the new sample is subjected to an exponential moving average with the original corresponding prototype, and the positive or negative prototype is updated. Alternatively, on a minimal support-query set composed of new samples, the embedding function can be fine-tuned with gradient descent for a set number of steps at a set learning rate, and then the positive and negative class prototypes can be recalculated and updated.

[0019] This invention also proposes a distribution network single-phase grounding fault selection system based on prototype meta-learning, including a signal acquisition and fault detection module, a multi-dimensional manual feature extraction module, a deep feature extraction module, a gating fusion module, a meta-learning training engine, and an online prototype library and selection decision module: The signal acquisition and fault detection module is used to acquire the bus zero-sequence voltage and the zero-sequence current of each feeder in real time, and to capture fault window data during disturbances. A multidimensional manual feature extraction module is used to extract time-domain, frequency-domain, time-frequency-domain, and nonlinear features from the window data to form a manual feature vector. The deep feature extraction module incorporates a dual-channel one-dimensional convolution, a bidirectional LSTM, and a self-attention network to extract deep feature vectors from the window data. The gated fusion module is used to adaptively weight and fuse handmade feature vectors with deep feature vectors to output a fault embedding vector; The meta-learning training engine is used to construct a binary classification equalization meta-task using a library of simulated or historical fault waveforms, and to train the parameters of the deep feature extraction module and the gated fusion module through a prototype network loss. An online prototype library and route selection decision module are used to store positive and negative prototypes, and to calculate the distance between each feeder embedding and the prototype in real-time faults, and output the fault line identification results.

[0020] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0021] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps according to the method described above.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Meta-learning provides breakthrough generalization ability and naturally overcomes sample imbalance: It adopts a binary classification prototype network and forces the number of positive and negative samples to be equal in each meta-task, so that the model is not affected by the imbalance of the original data in terms of mechanism. Under completely unseen new topology, only 1 to 5 fault samples and an equal number of non-fault samples are needed to quickly establish an accurate prototype, with a line selection accuracy of over 98%, and the recall rate of fault classes is not affected by sparse samples.

[0023] (2) Multidimensional fusion features are robust and reliable, laying a solid foundation for meta-learning: handmade features provide clear physical interpretable statistics, deep features capture subtle differences that are difficult to describe by hand, and gated fusion enables the embedded vectors to remain separable between classes when faced with high resistance, electric arcs, and strong noise, thus ensuring the stability and accuracy of prototype computation under conditions of few samples.

[0024] (3) Supports online low-cost evolution, greatly reducing the complexity of operation and maintenance: After a new line is put into use or a rare fault occurs, only a small number of samples are needed to achieve rapid adaptation. There is no need to readjust the threshold or retrain the model. The engineering implementation is simple and the practicality is good.

[0025] (4) Strong high-resistance fault detection capability: A scenario with a transition resistance of up to several thousand ohms was specifically included in the meta-task construction stage, forcing the model to form a tight intra-class cluster under high-resistance conditions. Tests showed that the line selection accuracy can still reach more than 96% when a single-phase grounding of 3000Ω is achieved, far exceeding the traditional steady-state and transient methods. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method for selecting a single-phase grounding fault in a distribution network based on prototype meta-learning, according to the present invention. Figure 2 This is a schematic diagram of single-phase grounding fault selection in a 10kV non-effectively grounded distribution network according to the present invention; Figure 3 This is a summary diagram of fault simulation in the 10kV non-effectively grounded distribution network simulation model of the present invention; Figure 4 This is a schematic diagram of the overall process of the method of the present invention; Figure 5 This is a schematic diagram of the multi-dimensional feature extraction and fusion structure of the present invention; Figure 6 This is the deep feature extraction network structure diagram of the present invention (dual-channel 1D-CNN+Bi-LSTM+attention). Figure 7 This is a schematic diagram of the meta-training task construction based on a binary classification prototype network according to the present invention; Figure 8 This is a flowchart of the online fault selection and adaptive update process of the present invention; Figure 9 This is a schematic diagram of the binary classifier learning training curve of the present invention; Figure 10 This is a schematic diagram of the visualization results of the t-SNE of the training embedding vectors of the present invention; Figure 11 This is a schematic diagram of the confusion matrix obtained after testing according to the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0028] like Figure 1 As shown, this invention proposes a method for selecting the line of a single-phase grounding fault in a distribution network based on prototype meta-learning, which specifically includes the following steps: The zero-sequence voltage of the distribution network bus and the zero-sequence current of each feeder are collected in real time. When the sudden change in zero-sequence voltage or the effective value of power frequency exceeds the preset threshold, a single-phase ground fault is determined to have occurred. The fault start time is locked, and the zero-sequence voltage sequence of the bus and the zero-sequence current sequence of each feeder are extracted within a preset time window after the fault. For each feeder, based on the feeder zero-sequence current sequence and the bus zero-sequence voltage sequence, statistical and physical features are extracted from four dimensions: time domain, frequency domain, time-frequency domain, and nonlinearity. The extracted statistical and physical features are concatenated and normalized to obtain the hand-crafted feature vector of the feeder. The methods for extracting statistical and physical features include: For a given length of Given a sequence, define the mean, standard deviation, effective value, mean absolute value, and maximum absolute value, and perform a fast Fourier transform on the sequence to obtain the amplitude spectrum and corresponding frequencies; The time-domain features are calculated based on the mean, standard deviation, effective value, mean absolute value, maximum absolute value, amplitude spectrum and corresponding frequency, including peak factor, impulse factor, kurtosis, skewness, waveform factor and margin factor; Frequency domain characteristics include spectral centroid, frequency variance, 5th harmonic amplitude percentage, and 7th harmonic amplitude percentage; Time-frequency domain features include wavelet packet energy entropy calculated after performing a three-level decomposition of the sequence using db4 wavelet packets; Nonlinear features include sample entropy calculated under a given embedding dimension and a given similarity tolerance.

[0029] A dual-channel one-dimensional deep feature extraction network is constructed, and the deep feature vectors of the zero-sequence current sequence and the zero-sequence voltage sequence of each feeder are obtained using the dual-channel one-dimensional deep feature extraction network; the deep feature vectors represent the abstract temporal features in the zero-sequence current sequence and the zero-sequence voltage sequence of the bus. The structure of the dual-channel one-dimensional deep feature extraction network is as follows: the bus zero-sequence voltage sequence and the zero-sequence current sequence of each feeder. Input layer, shape is The first channel is the feeder zero-sequence current sequence, and the second channel is the bus zero-sequence voltage sequence. Three cascaded one-dimensional convolutional blocks, each containing a one-dimensional convolutional layer, batch normalization, ReLU activation, and max pooling; A bidirectional long short-term memory network is used to perform temporal modeling of the time series output of convolution; From the attention pooling layer, the attention weights of the outputs of the bidirectional long short-term memory network at each time step are calculated and weighted summed to obtain the context vector; A fully connected projection layer maps the context vector to a deep feature vector.

[0030] A gated fusion unit is used to fuse the manual feature vector with the deep feature vector to obtain the fault embedding vector of the feeder; The processing of the gated fusion unit includes: The handcrafted feature vectors are mapped through a fully connected layer to obtain the mapping result; the deep feature vectors and the mapping result are concatenated and input into a fully connected layer, and activated by Sigmoid to obtain the gated vector; the fault embedding vector is calculated based on the gated vector, the handcrafted feature vectors, and the deep feature vectors.

[0031] The process of obtaining fault embedding vectors constitutes the embedding function; in the offline stage, a fault waveform library including various distribution network operating states is used to construct a binary classification meta-task; in the support set and query set of each meta-task, positive class samples correspond to faulty feeders, and negative class samples correspond to non-faulty feeders, and the number of positive and negative samples in the support set is equal; positive class prototypes and negative class prototypes are calculated on the support set respectively, and the classification probability and cross-entropy loss are calculated on the query set based on the Euclidean distance between the query sample embedding and the two class prototypes; the embedding function parameters are updated through gradient descent, so that the embedding function can quickly establish fault / non-fault discrimination prototypes with only a small number of samples; The parameters of the embedding function are updated with cross-entropy loss as the optimization objective.

[0032] The operating status of the power distribution network includes: Changes in neutral grounding method, changes in line commissioning / decommissioning combination, changes in transition resistance value, changes in initial phase angle of fault, and changes in fault location are one or more of the following combinations.

[0033] The process of obtaining fault embedding vectors constitutes an embedding function. In the offline stage, a fault waveform library including various distribution network operating states is used to construct a binary classification meta-task. Through meta-learning training, the embedding function acquires the ability to perform fault discrimination and classification based on prototypes, and the parameters of the embedding function are updated. The specific method for identifying a faulty line or bus is as follows: For each feeder, calculate the Euclidean distance between its fault embedding vector and the positive class prototype as the first distance, and the Euclidean distance with the negative class prototype as the second distance. If the first distance is not greater than the second distance, the feeder is determined to be a faulty line; If the first distance of all feeders is greater than the second distance, then it is determined to be a bus fault.

[0034] When applied online, an initial prototype is established and stored using samples from the target distribution network. When a real fault occurs, the zero-sequence current sequence of each feeder and the zero-sequence voltage sequence of the bus are mapped to a fault embedding vector through the updated embedding function. The distance to each prototype is calculated, and the faulty line or bus is determined based on the distance comparison results.

[0035] The method also includes an online adaptive update step: When new operating conditions occur in the distribution network, the model is updated using newly emerging fault and non-fault samples through one of the following methods: The embedding vector of the new sample is subjected to an exponential moving average with the original corresponding prototype, and the positive or negative prototype is updated. Alternatively, on a minimal support-query set composed of new samples, the embedding function can be fine-tuned with gradient descent for a set number of steps at a set learning rate, and then the positive and negative class prototypes can be recalculated and updated.

[0036] This invention also proposes a single-phase grounding fault selection system for distribution networks based on multi-dimensional feature fusion and prototype meta-learning, including a signal acquisition and fault detection module, a multi-dimensional manual feature extraction module, a deep feature extraction module, a gated fusion module, a meta-learning training engine, and an online prototype library and selection decision module, characterized in that: The signal acquisition and fault detection module is used to acquire the bus zero-sequence voltage and the zero-sequence current of each feeder in real time, and to capture fault window data during disturbances. A multidimensional manual feature extraction module is used to extract time-domain, frequency-domain, time-frequency-domain, and nonlinear features from the window data to form a manual feature vector. The deep feature extraction module incorporates a dual-channel one-dimensional convolution, a bidirectional LSTM, and a self-attention network to extract deep feature vectors from the window data. The gated fusion module is used to adaptively weight and fuse handmade feature vectors with deep feature vectors to output a fault embedding vector; The meta-learning training engine is used to construct a binary classification equalization meta-task using a library of simulated or historical fault waveforms, and to train the parameters of the deep feature extraction module and the gated fusion module through a prototype network loss. An online prototype library and route selection decision module are used to store positive and negative prototypes, and to calculate the distance between each feeder embedding and the prototype in real-time faults, and output the fault line identification results.

[0037] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0038] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps according to the method described above.

[0039] Example 1 The technical solution adopted in this invention is a method for selecting single-phase grounding faults in distribution networks based on gated fusion of multidimensional handcrafted features and deep features, and binary classification prototype network meta-learning. Its core innovation lies in: using binary classification prototype network meta-learning as the overall framework, training an embedding function composed of multidimensional feature extraction and gated fusion, enabling it to learn the meta-ability to "establish fault / non-fault discrimination prototypes with very few support samples" on a large number of differentiated distribution network fault tasks; combined with the adaptive fusion of multidimensional handcrafted features with clear physical meaning and abstract features from deep learning, ensuring the robustness and discriminative power of the embedding space; and overcoming the class imbalance problem by forcing equal numbers of positive and negative class samples in the meta-task construction.

[0040] The method specifically includes the following steps: Step 1: Fault Initiation Detection and Multi-Channel One-Dimensional Signal Acquisition. Real-time acquisition of bus zero-sequence voltage. and zero-sequence current of each feeder , This refers to the total number of feeders. When the zero-sequence voltage surge or the effective value of the power frequency exceeds the setting threshold (e.g., 15% of the phase voltage), a single-phase ground fault is determined to have occurred in the distribution network, and the fault initiation time is locked. .by Based on the baseline, the sampling data of b sampling points before the fault, Lb sampling points after the fault, and T being the power frequency cycle (preferably 3 power frequency cycles, for a total of L=600 sampling points) are extracted to form the zero-sequence current sequence of each feeder. With bus zero-sequence voltage sequence , , The sequence length is given.

[0041] Step 2: Multi-dimensional manual feature extraction. Zero-sequence current sequence for each feeder. and bus zero-sequence voltage sequence Statistical and physical features are extracted from four dimensions: time domain, frequency domain, time-frequency domain, and nonlinearity, transforming one-dimensional signals into robust multi-dimensional representations of changing operating conditions. Twelve features are extracted for each signal type, with clearly defined mathematical expressions and variables for each feature as follows.

[0042] Let the one-dimensional discrete sequence to be analyzed be... , , Let be the sequence length. Definition: Mean:

[0043] Standard deviation:

[0044] Valid values:

[0045] Average absolute value:

[0046] Maximum absolute value:

[0047] For sequence Perform a Fast Fourier Transform (FFT) to obtain its amplitude spectrum. , Due to the symmetry of the real sequence, only the first few steps are taken. Each point corresponds to a positive frequency component:

[0048] in The sampling frequency.

[0049] List of statistical and physical characteristics (12 dimensions in total): (1) Temporal characteristics (6 dimensions) Peak factor The ratio of the signal peak value to the effective value reflects the intensity of the impact component.

[0050]

[0051] Pulse factor The ratio of the signal peak value to the average absolute value is sensitive to impulse pulses.

[0052]

[0053] cliff The ratio of the fourth central moment of a sequence to the fourth power of its standard deviation measures the steepness of the distribution.

[0054]

[0055] Skewness The ratio of the third central moment to the cube of the standard deviation reflects the asymmetry of the distribution.

[0056]

[0057] Waveform factor The ratio of the effective value to the average absolute value reflects the flatness of the waveform.

[0058]

[0059] margin factor The ratio of the signal peak value to the square root amplitude squared represents the stationarity and margin.

[0060]

[0061] (2) Frequency domain characteristics (4 dimensions) Spectral centroid : The center frequency that describes the energy distribution of the spectrum.

[0062]

[0063] Frequency variance : Measures the degree to which spectral energy is dispersed around the center of gravity.

[0064]

[0065] 5th harmonic amplitude percentage In a 50Hz system, the 5th harmonic corresponds to 250Hz.

[0066]

[0067] 7th harmonic amplitude percentage The 7th harmonic corresponds to 350Hz.

[0068]

[0069] (3) Time-frequency domain features (1-dimensional) Wavelet packet energy entropy Three-level wavelet packet decomposition using the db4 wavelet basis yielded eight sub-band signals. , Calculate the energy of each subband:

[0070] Normalized energy ratio Then the wavelet packet energy entropy is:

[0071] in It is a very small positive number (e.g.) ), to avoid taking the logarithm of zero.

[0072] (4) Nonlinear characteristics (1-dimensional) Sample Entropy Given the embedding dimension Similarity tolerance : Build dimensional vector sum 3D vector, calculate the number of matching pairs and .

[0073] Sample entropy:

[0074] when At that time, let the sample entropy be .

[0075] The above summarizes the 12-dimensional statistical and physical characteristics of each type of signal. (The feeder...) The current statistics and physical feature vectors are concatenated with the common voltage statistics and physical feature vectors to obtain a 24-dimensional feature, which is then normalized using Min-Max. The interval forms the final handcrafted feature vector of the feeder. .

[0076] Step 3: Adaptive Deep Temporal Feature Extraction. A dual-channel one-dimensional deep feature extraction network is constructed to automatically learn abstract temporal features from the original current and voltage sequences, obtaining deep feature vectors. The network structure is as follows: Input layer: Shape is The first channel is a current sequence, and the second channel is a voltage sequence.

[0077] One-dimensional convolution stack: three convolutional blocks, each containing Conv1D (kernel size 3, channel numbers 64, 128, and 256 respectively) + BatchNorm + ReLU + MaxPool (pooling size 2). Output size. .

[0078] Bidirectional LSTM: Two-layer Bi-LSTM, 128 hidden units, output shape .

[0079] Self-attention pooling layer: The attention weights of the LSTM output at each time step are calculated and summed to obtain a 256-dimensional context vector. This self-attention pooling layer captures long-range temporal dependencies in transient signals, automatically focusing on the most discriminative moment in the transient process.

[0080] Projection layer: Fully connected layer outputs a 128-dimensional depth feature vector. .

[0081] Step 4: Heterogeneous Feature Gated Fusion. Handcrafted features and deep features come from different feature spaces and are complementary. Adaptive integration is achieved using a gated fusion unit:

[0082] in, , For the Sigmoid function, For the dimension matching matrix (the depth feature here is already 128 dimensions), (Can be a unit array) Represents element-wise multiplication. Gated vector. This allows the network to dynamically adjust its dependence on handcrafted and deep features based on signal quality, thereby obtaining robust fault embedding vectors. .

[0083] Step 5: Constructing a meta-learning line selection model based on a binary classification prototype network. For each fault event, the sample consisting of a feeder current and bus voltage is mapped to an embedding vector and assigned a binary classification label. : This indicates that the feeder line is faulty. This indicates that the feeder is a non-faulty line (including all non-faulty conditions, including bus faults). The complete calculation process from steps 1 to 4 constitutes the embedded function. A binary classification prototype network is adopted as the meta-learning framework, giving it... The meta-capability of "quickly building faulty and non-faulty prototypes with minimal support samples" fundamentally solves the problem of sample imbalance.

[0084] Meta-task construction: A series of meta-tasks are randomly selected from a fault waveform library (which can be generated by electromagnetic transient simulation) covering a large number of distribution network operation changes. Each meta-task is an independent binary classification learning task, containing a support set. and a query set The support set contains, respectively, One positive class (fault) sample and There are negative class (non-fault) samples, among which It is a very small positive integer, usually ranging from 1 to 5; the query set also contains Positive class and One negative class sample, Typically, 15 samples are used. The samples come from different fault files and different feeders, covering various distribution network operating states, ensuring the diversity of the tasks. This forced class equalization sampling method ensures that the number of positive and negative class samples is completely equal in each learning task, and the model will not be biased due to the scarcity of fault samples in the original data.

[0085] Prototype computation: for meta-tasks Positive class prototype and negative class prototype The values ​​are obtained from the embedding mean of the samples in their respective support sets:

[0086] If a certain class of samples is not supported (in very rare cases), the embedding mean of that class of samples in the query set can be used instead.

[0087] Query classification probability: query sample The posterior probability of belonging to the positive class is obtained by Softmax of the negative Euclidean distance between the embedding and the two prototypes:

[0088] Meta-objective function: For a batch of meta-tasks, the optimization objective is to minimize the sum of the cross-entropy losses of all query samples.

[0089] Optimizing parameters on massive meta-tasks using stochastic gradient descent This process enables the embedded network to learn to map fault samples and normal samples under any distribution network condition to two compact regions in the embedded space, and the two regions are far enough apart.

[0090] Step 6: Online Line Selection Deployment and Rapid Adaptation with Few Samples. Initial Prototype Establishment: Before the target distribution network is put into operation, a very small number of typical samples (e.g., 5 each) are generated for both fault-related and non-fault-related classes using a calibrated simulation model or early waveform recordings. These samples are then trained... Calculate the embeddings and take the mean to obtain the initial positive class prototype. and negative class prototype Store it in the online prototype library.

[0091] Real-time fault location: In the event of an actual fault, signals from each feeder are detected through... Mapped to embedding vector Calculate its relationship with and The Euclidean distance is obtained. and .like If the feeder is faulty, then the feeder is considered a faulty line; if all feeders are faulty... All greater than If the fault is detected, it is determined to be a busbar fault. To further improve reliability, the distance ratio of each feeder can be calculated. Compare the ratio of the minimum distance to the second minimum distance. When the value is below a preset threshold (e.g., 0.5), it indicates that the feeder is significantly biased towards a certain category, and the decision confidence level is high.

[0092] Online meta-adaptation: When new operating conditions occur in the distribution network (such as the addition of new feeders, adjustment of arc suppression coil taps, or the emergence of unseen high-resistance fault modes), model evolution can be completed by using only 1 to 2 newly emerging confirmed positive and negative samples through prototype incremental updates (exponential moving average) or by making very few fine-tuning steps on the embedded network, without the need for full historical data or large-scale retraining.

[0093] Example 2 This embodiment proposes a distribution network single-phase grounding fault selection system based on prototype meta-learning, including a signal acquisition and fault detection module, a multi-dimensional manual feature extraction module, a deep feature extraction module, a gating fusion module, a meta-learning training engine, and an online prototype library and selection decision module, wherein: The signal acquisition and fault detection module is used to acquire the bus zero-sequence voltage and the zero-sequence current of each feeder in real time. This module continuously monitors the instantaneous value or RMS value of the bus zero-sequence voltage. When a sudden change in zero-sequence voltage or the RMS value exceeds a preset trigger threshold, it determines that a single-phase ground fault has occurred in the distribution network and locks in the fault initiation time. Subsequently, the module extracts the bus zero-sequence voltage sequence and the zero-sequence current sequence of each feeder within a preset time window after the fault, using this as the raw data for subsequent analysis.

[0094] A multi-dimensional manual feature extraction module, connected to the signal acquisition and fault detection module, is used to receive the fault window data. For each feeder, this module extracts statistical and physical features from four dimensions: time domain, frequency domain, time-frequency domain, and nonlinearity, based on its zero-sequence current sequence and bus zero-sequence voltage sequence. Specifically, time-domain features include peak factor, impulse factor, kurtosis, skewness, waveform factor, and margin factor; frequency-domain features include spectral centroid, frequency variance, 5th harmonic amplitude percentage, and 7th harmonic amplitude percentage; time-frequency domain features are wavelet packet energy entropy; and nonlinear features are sample entropy. This module concatenates and normalizes the extracted current and voltage features to form the manual feature vector for that feeder.

[0095] A deep feature extraction module, connected to the signal acquisition and fault detection module, is used to receive the fault window data. This module incorporates a dual-channel one-dimensional deep neural network, using the feeder zero-sequence current sequence as the first channel input and the bus zero-sequence voltage sequence as the second channel input. The network sequentially includes: multiple cascaded one-dimensional convolutional blocks, each containing a one-dimensional convolutional layer, a batch normalization layer, an activation function layer, and a pooling layer; a bidirectional long short-term memory network for temporal modeling of the convolutional output time series; a self-attention pooling layer for calculating attention weights on the output of the bidirectional long short-term memory network at each time step and performing a weighted summation to obtain a context vector; and a fully connected projection layer for mapping the context vector to a deep feature vector.

[0096] The gated fusion module, connected to both the multidimensional manual feature extraction module and the deep feature extraction module, receives the manual feature vector and the deep feature vector. This module first maps the manual feature vector to the same dimension as the deep feature vector using a fully connected layer; then, it concatenates the mapped manual and deep feature vectors and inputs the concatenated vectors into another fully connected layer, processing them with a sigmoid activation function to obtain a gated vector; finally, it performs an element-wise weighted summation of the manual and deep feature vectors based on the gated vector to obtain the fault embedding vector for the feeder.

[0097] A meta-learning training engine, connected to the deep feature extraction module and the gated fusion module, is used to train the parameters of these two modules offline. This engine constructs a series of binary classification meta-tasks using a simulated or historical fault waveform library. Each meta-task includes a support set and a query set; positive class samples correspond to fault feedlines, and negative class samples correspond to non-fault feedlines, with an equal number of positive and negative samples in the support set. Positive and negative prototypes are calculated on the support set of each meta-task, and classification probability and cross-entropy loss are calculated on the query set based on the Euclidean distance between the fault embedding vector of the query sample and the two prototypes. Using the cross-entropy loss as the optimization objective, the parameters of the deep feature extraction module and the gated fusion module are updated through gradient descent, enabling the jointly constructed embedding function to acquire the ability to classify faults based on prototypes.

[0098] An online prototype library and route selection decision module, connected to the gated fusion module, are used to store positive and negative prototypes and make route selection decisions in the event of a real-time fault. Before the target distribution network is put into operation or during its first operation, this module uses a small number of labeled samples to establish and store initial positive and negative prototypes. When a real fault occurs, the zero-sequence current sequence of each feeder and the zero-sequence voltage sequence of the busbar are mapped to fault embedding vectors through a trained deep feature extraction module and a gated fusion module. The Euclidean distances to the positive and negative prototypes are calculated respectively. For each feeder, if the distance between its fault embedding vector and the positive prototype is less than the distance to the negative prototype, the feeder is determined to be a faulty line; if the distances between the fault embedding vectors of all feeders and the positive prototypes are greater than the distances to the negative prototypes, the busbar is determined to be a fault.

[0099] The data flow between modules is as follows: the output of the signal acquisition and fault detection module is simultaneously sent to the multidimensional manual feature extraction module and the deep feature extraction module; the outputs of the multidimensional manual feature extraction module and the deep feature extraction module are both sent to the gate control fusion module; the output of the gate control fusion module is sent to the meta-learning training engine for offline training, and to the online prototype library and the route selection decision module for online route selection.

[0100] Example 3 This embodiment proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the following steps: Real-time acquisition of zero-sequence voltage of the distribution network bus and zero-sequence current of each feeder; when a single-phase ground fault is detected, the zero-sequence voltage sequence of the bus and the zero-sequence current sequence of each feeder within a preset time window after the fault are extracted. For each feeder, its zero-sequence current sequence and the bus zero-sequence voltage sequence are used as a pair of inputs. Statistical and physical features are extracted from four dimensions: time domain, frequency domain, time-frequency domain, and nonlinearity. The extracted features are concatenated and normalized to obtain the handmade feature vector of the feeder. A one-dimensional deep neural network is constructed, and the current sequence and voltage sequence are used as dual-channel inputs to extract its deep time-series features, thereby obtaining the deep feature vector of the feeder. A gated fusion unit is used to adaptively weight and fuse the handmade feature vector and the deep feature vector to obtain the fault embedding vector of the feeder. When applied online, an initial prototype is established and stored using samples from the target distribution network. When a real fault occurs, the voltage and current sequences of each feeder are mapped to a fault embedding vector through the embedding function. The distance to each prototype is calculated, and the faulty line or bus is determined based on the distance comparison results.

[0101] Preferably, the computer-readable storage medium includes, but is not limited to, read-only memory, random access memory, erasable programmable read-only memory, flash memory, hard disk, optical disk, or solid-state drive. When the computer program is executed, it may also implement the additional steps described in any of the dependent method claims above.

[0102] Example 4 This embodiment proposes a terminal, including a processor and a storage medium, wherein: The storage medium is used to store instructions, which include a computer program for executing the above-described method for selecting single-phase grounding faults in distribution networks based on multi-dimensional feature fusion and prototype meta-learning. The processor is communicatively connected to the storage medium and is used to read and execute the instructions stored in the storage medium to achieve the following operations: real-time acquisition of zero-sequence voltage of the distribution network bus and zero-sequence current of each feeder and fault detection; multi-dimensional manual feature extraction and deep feature extraction of fault data; generation of fault embedding vector through gating fusion; and online line selection decision using the embedding function obtained by meta-learning training.

[0103] Preferably, the terminal is one of an embedded terminal, an industrial control computer, a power distribution automation terminal, or a substation monitoring system. The terminal may also include a communication interface for data interaction with a higher-level dispatching system or operation and maintenance platform, and for reporting fault location results and alarm information.

[0104] Example 5 The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0105] Taking a typical 10kV neutral-point grounded distribution network with an arc-suppression coil as an example, the busbar has 6 feeders, including cables, overhead lines, and mixed lines, such as... Figure 2 As shown. A system was built in MATLAB / Simulink as follows. Figure 2 The simulation model of the consumption coil grounded (non-effective grounded) system shown is shown. Figure 2There are six feeders in total, each with several branch lines, such as L11 and L12. Cable lines and overhead lines make up the feeders. The simulation addresses single-phase grounding faults with varying transition resistances occurring on different feeders, at different fault locations, and with different initial phase angles. The fault location varies from f0 to f8, the initial phase angle from 0° to 330°, and the transition resistance from 0.1Ω to 15000Ω (including metallic grounding, arc grounding of several hundred ohms, and 15000Ω extreme high-resistance grounding). This comprehensively covers various grounding conditions that may occur in actual distribution networks, and specifically includes an extreme high-resistance scenario to verify the invention's tolerance to high noise and weak fault signals. A summary table of the simulated single-phase grounding faults under various conditions is shown below. Figure 3 As shown, a total of 975 sample files were obtained. Sampling frequency kHz, truncating a 60ms window after the fault, i.e., the sequence length. The file is named like "3_xxx.txt", and the first field represents the faulty feeder number (0 indicates a bus fault). The file is divided into training, validation, and test sets in a ratio of 7:1.5:1.5.

[0106] A schematic diagram of the overall process of the method of the present invention is shown below. Figure 4 As shown, it mainly includes: Step 1: Fault Signal Extraction and Sample Construction. After reading each fault file, locate the fault initiation time, extract the zero-sequence voltage of the bus and the zero-sequence current of the 6 feeders, select 3 power frequency cycles, and collect sampling data from a total of L=600 sampling points to form the zero-sequence current sequence of each feeder. With bus zero-sequence voltage sequence , , For the sequence length. For each feeder. Construct a sample: voltage sequence and the first The current sequence of each feeder is paired. If the file indicates a fault in feeder 3, then that sample is labeled 1, and the other feeder samples are labeled 0; if it indicates a bus fault, all samples are labeled 0. This process is repeated for each file. Of the samples, the overall positive sample ratio is approximately 1 / 6, indicating a significant imbalance.

[0107] Step 2: Detailed calculation of manual features. For example... Figure 5 As shown, 12-dimensional features are calculated for each sample's current and voltage sequences: peak factor, impulse factor, kurtosis, skewness, waveform factor, margin factor, spectral centroid, frequency variance, proportion of 5th and 7th harmonics, wavelet packet energy entropy, and sample entropy. These are concatenated into a 24-dimensional array and, after Min-Max normalization, used as handcrafted features. The calculation process is detailed in the technical solution.

[0108] Steps 3-4: Deep features and fusion. For example... Figure 5 and Figure 6 As shown, voltage and current sequences are fed into a dual-channel deep network to output 128-dimensional deep features; handcrafted features are also mapped to 128 dimensions through a fully connected layer and input together with the deep features into a gated fusion unit to adaptively generate a 128-dimensional embedding vector.

[0109] Step 5: Training the binary classification prototype network meta-network. (e.g.) Figure 7 As shown, during meta-training, each meta-task is randomly selected from the training set. positive samples and A negative sample constitutes the support set, and then... positive samples and The query set consists of negative samples. The optimizer uses Adam, with a learning rate of... The model employs cosine annealing scheduling, undergoing 150 training rounds, and incorporates an early stopping mechanism (terminating training if accuracy does not improve after 30 rounds). Because each task is perfectly balanced between positive and negative samples, the model does not favor the negative class, but instead learns the essential measure of faulty and non-faulty samples.

[0110] Step 6: Online Testing and Results. The flowchart for online fault location and adaptive update is as follows: Figure 8 As shown. During the training process, the binary classifier's learning training curve is as follows. Figure 9 As shown. After training, the visualization results of the embedding vector t-SNE are as follows. Figure 10 As shown, positive and negative prototypes are calculated using all training samples. On the test set, the distance to each prototype of the two classes is calculated for each sample, and the one with the closer distance is used as the prediction. The confusion matrix obtained after testing is shown below. Figure 11 As shown, the test set accuracy reached 100%, with a recall (sensitivity) of 97.8% for faulty classes and a specificity of 99.2% for non-faulty classes. The confusion matrix indicates extremely low false positives and false negatives. The total number of supported normal samples was 734, and the total number of supported faulty samples was 124. The precision, l, and f1-score were all 1, demonstrating the excellent performance of the method of this invention.

[0111] When a new feeder is added to the distribution network, only two new fault samples and a small number of normal samples need to be generated using the simulation model. The accuracy of line selection can be restored to over 98% through prototype updates without any retraining, which fully verifies the powerful generalization and online evolution capabilities of the meta-learning framework.

[0112] In summary, this invention aims to address the technical challenges of significantly reduced accuracy and high adaptation costs of existing fault selection methods under complex conditions such as distribution network topology changes, neutral grounding mode switching, wide fluctuations in transition resistance (including high-resistance faults), random fault initial angles, and extremely limited and imbalanced historical fault samples. It provides a method that achieves high-precision fault selection under multiple operating conditions, limited samples, and strong imbalance, relying solely on one-dimensional transient signals from each feeder end, and can continuously evolve online. This invention discloses a complete technical solution based on binary classification prototype network meta-learning and supported by multi-dimensional feature gating fusion, and elaborates on the mechanism for solving the sample imbalance problem through class-balanced meta-task design, fully demonstrating the essential difference and significant progress compared to existing technologies.

[0113] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0114] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0115] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0116] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for selecting the fault location of a single-phase grounding fault in a distribution network based on prototype meta-learning, characterized in that, Includes the following steps: Real-time acquisition of zero-sequence voltage of the distribution network bus and zero-sequence current of each feeder. When the sudden change in zero-sequence voltage or the effective value of power frequency exceeds the corresponding preset threshold, a single-phase ground fault is determined to have occurred. The fault start time is locked, and the zero-sequence voltage sequence of the bus and the zero-sequence current sequence of each feeder are extracted within a preset time window after the fault. For each feeder, based on the feeder zero-sequence current sequence and the bus zero-sequence voltage sequence, statistical and physical features are extracted from four dimensions: time domain, frequency domain, time-frequency domain, and nonlinearity. The extracted statistical and physical features are concatenated and normalized to obtain the hand-crafted feature vector of the feeder. A dual-channel one-dimensional deep feature extraction network is constructed, and the deep feature vectors of the zero-sequence current sequence and the zero-sequence voltage sequence of each feeder are obtained using the dual-channel one-dimensional deep feature extraction network; the deep feature vectors represent the abstract temporal features in the zero-sequence current sequence and the zero-sequence voltage sequence of the bus. A gated fusion unit is used to fuse the manual feature vector with the deep feature vector to obtain the fault embedding vector of the feeder; The process of obtaining fault embedding vectors constitutes an embedding function. In the offline stage, a fault waveform library including various distribution network operating states is used to construct a binary classification meta-task. Through meta-learning training, the embedding function acquires the ability to perform fault discrimination and classification based on prototypes, and the parameters of the embedding function are updated. When applied online, an initial prototype is established and stored using samples from the target distribution network. When a real fault occurs, the zero-sequence current sequence of each feeder and the zero-sequence voltage sequence of the bus are mapped to a fault embedding vector through the updated embedding function. The distance to each prototype is calculated, and the faulty line or bus is determined based on the distance comparison results.

2. The method for selecting a single-phase grounding fault in a distribution network based on prototype meta-learning according to claim 1, characterized in that: The methods for extracting statistical and physical features include: For a given length of Given a sequence, define the mean, standard deviation, effective value, mean absolute value, and maximum absolute value, and perform a fast Fourier transform on the sequence to obtain the amplitude spectrum and corresponding frequencies; The time-domain features are calculated based on the mean, standard deviation, effective value, mean absolute value, maximum absolute value, amplitude spectrum and corresponding frequency, including peak factor, impulse factor, kurtosis, skewness, waveform factor and margin factor; Frequency domain characteristics include spectral centroid, frequency variance, 5th harmonic amplitude percentage, and 7th harmonic amplitude percentage; Time-frequency domain features include wavelet packet energy entropy calculated after performing a three-level decomposition of the sequence using db4 wavelet packets; Nonlinear features include sample entropy calculated under a given embedding dimension and a given similarity tolerance.

3. The method for selecting a single-phase grounding fault in a distribution network based on prototype meta-learning according to claim 1, characterized in that: The structure of the dual-channel one-dimensional deep feature extraction network is as follows: the bus zero-sequence voltage sequence and the zero-sequence current sequence of each feeder. Input layer, shape is The first channel is the feeder zero-sequence current sequence, and the second channel is the bus zero-sequence voltage sequence. Three cascaded one-dimensional convolutional blocks, each containing a one-dimensional convolutional layer, batch normalization, ReLU activation, and max pooling; A bidirectional long short-term memory network is used to perform temporal modeling of the time series output of convolution; From the attention pooling layer, the attention weights of the outputs of the bidirectional long short-term memory network at each time step are calculated and weighted summed to obtain the context vector; A fully connected projection layer maps the context vector to a deep feature vector.

4. The method for selecting a single-phase grounding fault in a distribution network based on prototype meta-learning according to claim 1, characterized in that: The processing of the gated fusion unit includes: The handcrafted feature vectors are mapped through a fully connected layer to obtain the mapping result; the deep feature vectors and the mapping result are concatenated and input into a fully connected layer, and activated by Sigmoid to obtain the gated vector; the fault embedding vector is calculated based on the gated vector, the handcrafted feature vectors, and the deep feature vectors.

5. The method for selecting a single-phase grounding fault in a distribution network based on prototype meta-learning according to claim 1, characterized in that: The process of obtaining fault embedding vectors constitutes an embedding function. In the offline phase, a binary classification meta-task is constructed using a fault waveform library encompassing various distribution network operating states. Meta-learning training enables the embedding function to acquire the ability to classify faults based on prototypes, and the parameters of the embedding function are updated. Specific steps include: The process of obtaining fault embedding vectors constitutes the embedding function; in the offline stage, a fault waveform library including various distribution network operating states is used to construct a binary classification meta-task; in the support set and query set of each meta-task, positive class samples correspond to faulty feeders, and negative class samples correspond to non-faulty feeders, and the number of positive and negative samples in the support set is equal; positive class prototypes and negative class prototypes are calculated on the support set respectively, and the classification probability and cross-entropy loss are calculated on the query set based on the Euclidean distance between the query sample embedding and the two class prototypes; the embedding function parameters are updated through gradient descent, so that the embedding function can quickly establish fault / non-fault discrimination prototypes with only a small number of samples; The parameters of the embedding function are updated with cross-entropy loss as the optimization objective.

6. The method for selecting a single-phase grounding fault in a distribution network based on prototype meta-learning according to claim 5, characterized in that: The operating status of the power distribution network includes: Changes in neutral grounding method, changes in line commissioning / decommissioning combination, changes in transition resistance value, changes in initial phase angle of fault, and changes in fault location are one or more of the following combinations.

7. The method for selecting a single-phase grounding fault in a distribution network based on prototype meta-learning according to claim 1, characterized in that: The specific method for identifying a faulty line or bus is as follows: For each feeder, calculate the Euclidean distance between its fault embedding vector and the positive class prototype as the first distance, and the Euclidean distance with the negative class prototype as the second distance. If the first distance is not greater than the second distance, the feeder is determined to be a faulty line; If the first distance of all feeders is greater than the second distance, then it is determined to be a bus fault.

8. The method for selecting a single-phase grounding fault in a distribution network based on prototype meta-learning according to claim 1, characterized in that: The method also includes an online adaptive update step: When new operating conditions occur in the distribution network, the model is updated using newly emerging fault and non-fault samples through one of the following methods: The embedding vector of the new sample is subjected to an exponential moving average with the original corresponding prototype, and the positive or negative prototype is updated. Alternatively, on a minimal support-query set composed of new samples, the embedding function can be fine-tuned with gradient descent for a set number of steps at a set learning rate, and then the positive and negative class prototypes can be recalculated and updated.

9. A distribution network single-phase grounding fault selection system based on prototype meta-learning using the method of any one of claims 1-8, comprising a signal acquisition and fault detection module, a multi-dimensional manual feature extraction module, a deep feature extraction module, a gating fusion module, a meta-learning training engine, and an online prototype library and selection decision module, characterized in that: The signal acquisition and fault detection module is used to acquire the bus zero-sequence voltage and the zero-sequence current of each feeder in real time, and to capture fault window data during disturbances. A multidimensional manual feature extraction module is used to extract time-domain, frequency-domain, time-frequency-domain, and nonlinear features from the window data to form a manual feature vector. The deep feature extraction module incorporates a dual-channel one-dimensional convolution, a bidirectional LSTM, and a self-attention network to extract deep feature vectors from the window data. The gated fusion module is used to adaptively weight and fuse handmade feature vectors with deep feature vectors to output a fault embedding vector; The meta-learning training engine is used to construct a binary classification equalization meta-task using a library of simulated or historical fault waveforms, and to train the parameters of the deep feature extraction module and the gated fusion module through a prototype network loss. An online prototype library and route selection decision module are used to store positive and negative prototypes, and to calculate the distance between each feeder embedding and the prototype in real-time faults, and output the fault line identification results.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 8.

11. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-8.

Images