Power distribution network fault distance measurement scheme recommendation method and system

Abstract

The invention relates to the technical field of power distribution network fault diagnosis, and discloses a power distribution network fault distance measurement scheme recommendation method and system, and the method comprises the steps: collecting multi-dimensional data when a power distribution network fault occurs, and carrying out the preprocessing of the multi-dimensional data, so as to construct a training data set; constructing a deep learning recommendation model, and training the deep learning recommendation model by using the training data set; extracting image spatial features from the fault electrical quantity data based on a convolutional neural network, and sequentially performing nonlinear dimensionality reduction and adaptive redundancy elimination processing on the image spatial features; extracting a multi-scale gray dispersion entropy feature from the fault electrical quantity data; fusing the image spatial features after redundancy elimination with the multi-scale gray dispersion entropy features, weighting the fused features, and finally outputting the probability of adaptation with each candidate ranging scheme; and collecting real-time fault data, inputting the real-time fault data into the trained deep learning recommendation model, and outputting a recommended distance measurement scheme. Therefore, the recommendation accuracy and engineering practicability of the distance measurement scheme are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution network fault diagnosis, and in particular to a power distribution network fault distance measurement scheme recommendation method and system. BACKGROUND

[0002] The current power distribution network structure presents an increasingly complex development trend. In the combined scenarios of different fault types (such as single-phase grounding, phase-to-phase short circuit), different topological structures (such as radial, loop network) and different operating conditions (such as load fluctuation, distributed power access), the adaptability of a single fault distance measurement scheme is significantly insufficient, and problems such as large distance measurement error and low reliability are prone to occur. The traditional fault distance measurement scheme relies on manual experience selection, and has inherent defects such as slow response speed and strong subjectivity, which cannot meet the timeliness requirements of real-time fault disposal of the power distribution network; the existing automatic recommendation methods are mostly designed based on fixed rule bases, and the dynamic correlation between multi-dimensional fault characteristics is not fully considered, making it difficult to comprehensively cover complex and variable actual fault scenarios, resulting in obvious short boards in the accuracy and engineering practicability of the distance measurement scheme recommendation.

[0003] Further analysis of the existing technical bottlenecks: in the fault characteristic processing link, the traditional dimension reduction method mostly uses Euclidean distance to measure the similarity of high-dimensional characteristics, but the fault characteristics of the power distribution network have significant high-dimensionality and nonlinearity, and the Euclidean distance is easily affected by the curse of dimensionality, resulting in distortion of the feature structure in the dimension reduction process and loss of key fault information; in terms of feature redundancy, existing methods mostly rely on simple linear screening or experience threshold setting, making it difficult to accurately identify core features that are strongly adaptable to fault distance measurement schemes, and a large amount of redundant information will increase the model calculation burden and interfere with the recommendation decision; in terms of time series feature extraction, traditional methods (such as mean, variance, and ordinary entropy) can only depict the surface statistical characteristics of time series data, and cannot accurately capture the nonlinear dynamic complexity in the fault transient process, and lack the ability to integrate multi-time scale features, making it difficult to fully reflect the short-term fluctuations and long-term trend rules of fault development. These technical defects collectively result in insufficient adaptability and accuracy of existing recommendation methods in complex fault scenarios, and cannot provide reliable scheme support for fault disposal of the power distribution network. SUMMARY

[0004] The core purpose of the present application is to solve the technical problems of low selection efficiency and poor scene adaptability of traditional power distribution network fault distance measurement schemes, and to provide a power distribution network fault distance measurement scheme recommendation method and system, which improves the accuracy and engineering practicability of the distance measurement scheme recommendation through innovative feature processing and extraction methods.

[0005] According to a first aspect of the present application, a power distribution network fault distance measurement scheme recommendation method is provided, comprising the following steps:

[0006] Collect multi-dimensional data when a fault occurs in a power distribution network, and preprocess the multi-dimensional data to construct a training data set; the multi-dimensional data includes fault electrical quantity data, power distribution network topology data, and historical fault handling data;

[0007] Construct a deep learning recommendation model and train it using the training data set; the deep learning recommendation model extracts and fuses multi-modal features in the following way: based on a convolutional neural network, image spatial features are extracted from fault electrical quantity data, and the image spatial features are sequentially subjected to nonlinear dimension reduction and adaptive redundancy removal processing; multi-scale gray scatter entropy features for quantifying time series dynamic complexity are extracted from the fault electrical quantity data; the image spatial features after redundancy removal are fused with the multi-scale gray scatter entropy features, and the fused features are weighted through an attention mechanism, and finally the adaptation probability of each candidate ranging scheme is output;

[0008] When a fault occurs in a power distribution network, real-time fault data is collected and input into the trained deep learning recommendation model to output a recommended ranging scheme.

[0009] Further, the nonlinear dimension reduction of the image spatial features is specifically: the SKLAN pseudo distance based on the neighborhood symmetric KL divergence is used to replace the Euclidean distance, and the dimension reduction is realized by combining the multidimensional scaling (MDS) technology;

[0010] The adaptive redundancy removal processing of the image spatial features is specifically: the selective principal component layer (SPCL) method is used to filter out redundant features by calculating the feature amplitude and setting an adaptive threshold.

[0011] Further, the calculation of the SKLAN pseudo distance in the nonlinear feature dimension reduction includes:

[0012] Neighborhood probability distribution construction: for any two feature points i and j, the neighborhood probability distribution of the two is calculated based on the Gaussian kernel function.

[0013] SKLAN pseudo distance calculation: the difference between the neighborhood probability distributions of the two feature points is quantified by symmetric KL divergence to obtain the SKLAN pseudo distance, and the calculation formula is:

[0014]

[0015] wherein, : the probability of feature point as a neighborhood point; : the probability of feature point as a neighborhood point; : the probability of feature point as a neighborhood point; : the probability of feature point and The SKLAN pseudo-distance reflects the strength of the local structural correlation between the two in the original high-dimensional space;

[0016] MDS dimensionality reduction optimization: Using the SKLAN pseudo-distance matrix as the input to the MDS, the mapping from high-dimensional features to low-dimensional space is achieved by minimizing the stress function. The formula for minimizing the stress function is as follows:

[0017]

[0018] Among them, Stress: the stress function value of MDS dimensionality reduction, used to quantify the structural differences before and after dimensionality reduction of high-dimensional features of distribution network faults; Feature points in the lower-dimensional space after dimensionality reduction and The Euclidean distance is a low-dimensional mapping of the original SKLAN pseudo-distance; : Feature point index in the distribution network fault feature set, used to traverse all feature point pairs; : Feature point index in the distribution network fault feature set, and Form feature point pairs.

[0019] Furthermore, the Selective Principal Component Layer (SPCL) method is used to achieve adaptive redundancy removal, including:

[0020] Feature amplitude calculation: For the dimension-reduced feature matrix, calculate the feature amplitude of each feature sample;

[0021] Determining the screening threshold: Calculate the number of key features to be retained based on the feature screening threshold ratio r. The characteristic amplitudes of all samples are expanded into one-dimensional vectors, sorted in ascending order, and the nth vector is selected. One value is used as the filtering threshold;

[0022] Redundant feature filtering: For each sample, retain features with amplitudes greater than or equal to the threshold and filter out redundant features.

[0023] Furthermore, the specific process of extracting the multi-scale gray-level scattering entropy feature from the fault electrical quantity data for quantifying the temporal dynamic complexity is as follows:

[0024] Calculating the multi-scale gray-scale scattering entropy (MGDE) of fault electrical quantity time-series data includes:

[0025] Coarse-grained sequence generation: The original time series is coarse-grained according to the scale factor s to generate multi-scale sequences;

[0026] Multi-scale GDE calculation: Calculate the gray-level scatter entropy (GDE) for the sequence at each scale, and finally concatenate the gray-level scatter entropy (GDE) values ​​of all scales to obtain the multi-scale gray-level scatter entropy feature vector.

[0027] Furthermore, the calculation of gray-scale scattering entropy (GDE) includes:

[0028] The standardized time-series data is mapped to discrete class sequences, and the phase space is reconstructed to generate a scattering pattern.

[0029] Construct the transition probability matrix between scattering patterns and convert it into a grayscale matrix;

[0030] Calculate the probability of occurrence of each gray value in the gray-scale matrix. , Based on the probability The gray-level scattering entropy is then calculated using the Shannon entropy formula:

[0031]

[0032]

[0033] in, Distribution network fault timing data The grayscale distribution entropy; This represents the probability of grayscale value t appearing. Number of categories; Embedded dimension; Time delay; Transition step size; The grayscale matrix of the first Line number The elements of the column are transition probabilities. Mapped grayscale value; Number : Counting function, counting the values ​​in a grayscale matrix The total number of elements, i.e., grayscale value. Number of times it appears.

[0034] Furthermore, the weighting of the fused features through an attention mechanism includes:

[0035] The channel attention module assigns weights to different dimensions of the fused features, and then uses these weights to weight the fused features. The weights are calculated as follows:

[0036]

[0037]

[0038] in, To fuse the feature matrix, the image spatial features after redundancy removal are used. With multi-scale entropy feature matrix Obtained by concatenation along the feature dimension; : Sample size; : The dimension of the spatial features of the image after redundancy removal; Dimensions of multi-scale gray-scale scattering entropy features; : Weight matrix of the first layer of the attention module; : Weight matrix of the second layer of the attention module; ReLU activation function; : Transpose of the fused feature matrix; The first-level bias term of the attention module; : Second-layer bias term of the attention module; Attention-weighted feature matrix; : by weight vector The constructed diagonal matrix.

[0039] Furthermore, the training and optimization of the deep learning recommendation model includes:

[0040] A multi-dimensional fault feature dataset is constructed based on the structure of "fault electrical quantity characteristics - topological characteristics - operating condition characteristics - optimal ranging scheme label". The training set, validation set, and test set are further divided into... Proportional division;

[0041] The model is trained using historical fault handling data and corresponding optimal ranging scheme label data. The model's generalization ability is improved by using cross-entropy loss function and Adam optimizer, and cross-validation and regularization are applied.

[0042] Furthermore, the method also includes: updating the training dataset and iteratively optimizing the deep learning recommendation model based on the actual performance of the recommended ranging scheme, specifically including:

[0043] If the ranging error of the recommended solution exceeds the preset threshold, the sample label of the fault will be updated to the solution with the best actual effect, and the model will be incrementally trained periodically using the accumulated feedback data.

[0044] According to a second aspect of the present invention, a distribution network fault location scheme recommendation system is also provided, for implementing the distribution network fault location scheme recommendation method as described in the first aspect, the system comprising:

[0045] The fault data acquisition module is used to collect real-time fault data of the power distribution network.

[0046] The ranging scheme recommendation module inputs the real-time fault data of the power distribution network into the trained deep learning recommendation model, and extracts and fuses multimodal features in the following ways: extracting image spatial features from the fault electrical quantity data based on a convolutional neural network, and performing nonlinear dimensionality reduction and adaptive redundancy removal on the image spatial features in sequence; extracting multi-scale gray-level scatter entropy features from the fault electrical quantity data to quantify the temporal dynamic complexity; fusing the redundancy-removed image spatial features with the multi-scale gray-level scatter entropy features, and weighting the fused features through an attention mechanism, finally outputting the recommended ranging scheme;

[0047] The update and iteration optimization module is used to update the training dataset and iteratively optimize the deep learning recommendation model based on the actual performance of the recommended ranging scheme.

[0048] Compared with the prior art, the present invention achieves the following beneficial effects:

[0049] 1. To address the problem that Euclidean distance in traditional dimensionality reduction methods can easily lead to distortion of high-dimensional feature structures, this invention proposes a pseudo-distance (SKLAN) based on neighborhood symmetric KL divergence to replace Euclidean distance. Combined with multidimensional scaling (MDS) technology, it achieves accurate dimensionality reduction. While reducing feature dimensions and computational complexity, it preserves the local structural correlation of high-dimensional features of distribution network faults to the greatest extent, providing high-quality basic data for subsequent feature processing.

[0050] 2. To address the shortcomings of existing feature redundancy removal methods in accurately selecting core features, this invention proposes a Selective Principal Component Layer (SPCL) method. By quantifying the importance of feature amplitude and adaptively setting the screening threshold, it efficiently filters out invalid and redundant information that is poorly adapted to fault location schemes, retains key features, and strengthens the correlation between features and recommendation targets while simplifying the model structure, thereby improving the accuracy of recommendation decisions.

[0051] 3. To address the problem that traditional time-series feature extraction methods cannot characterize the nonlinear dynamic complexity of faults, this invention proposes Gray-scale Spread Entropy (GDE). By constructing a spread pattern transition probability matrix and converting it into a gray-scale matrix, the local dynamic complexity of fault time series is quantified, accurately capturing the pattern change patterns in the fault transient process, and providing more discriminative time-series feature support for solution recommendation.

[0052] 4. To address the limitation that single-scale features cannot fully reflect the multi-scale characteristics of fault time series, this invention proposes Multi-scale Gray-scale Dispersion Entropy (MGDE). By coarse-grained processing, sequences at different time scales are generated and the corresponding GDE values ​​are calculated. This integrates the short-term fluctuations and long-term trend features of fault time series, achieving a comprehensive characterization of fault dynamic changes and further improving the model's adaptability to different fault scenarios.

[0053] 5. This invention designs an end-to-end deep learning recommendation model consisting of "data preprocessing - multimodal feature extraction - feature fusion - scheme recommendation - feedback iteration" to achieve accurate matching between distribution network fault scenarios and optimal ranging schemes, improve recommendation efficiency and reliability, and meet the engineering requirements for real-time fault handling in distribution networks. Attached Figure Description

[0054] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0055] Figure 1 This is a flowchart illustrating a method for recommending fault location schemes in a power distribution network, as provided in an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of the feature extraction and fusion structure of the deep learning recommendation model provided in this embodiment of the invention;

[0057] Figure 3 This is a schematic diagram of the SKLAN-MDS feature dimensionality reduction process provided in an embodiment of the present invention;

[0058] Figure 4 This is a schematic diagram of the MGDE temporal feature extraction process provided in an embodiment of the present invention;

[0059] Figure 5 This is a schematic diagram of a distribution network fault location scheme recommendation system provided in Embodiment 3 of the present invention;

[0060] Figure 6 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. Detailed Implementation

[0061] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0062] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0063] The acquisition, storage, use, and processing of data in this application comply with relevant national laws and regulations. It should be noted that existing industry solutions such as software, components, or models may be mentioned in the embodiments of this application. These should be considered exemplary and intended only to illustrate the feasibility of implementing the technical solution of this application, but do not imply that the applicant has already used or necessarily used such a solution.

[0064] Example 1

[0065] Figure 1 This is a flowchart illustrating a recommended method for fault location in a power distribution network according to an embodiment of the present invention. Figure 1 As shown, a recommended method 100 for fault location in a distribution network includes the following steps:

[0066] S1: Collect multi-dimensional data when a distribution network fault occurs, and preprocess the multi-dimensional data to construct a training dataset; the multi-dimensional data includes: fault electrical quantity data, distribution network topology data, and historical fault handling data;

[0067] Step S1 is used to perform data acquisition and preprocessing. It includes the following steps:

[0068] Step S1.1: Multi-dimensional data collection

[0069] Collect multi-dimensional data when a distribution network fault occurs, including fault electrical quantity data (voltage, current, and power time-series data, with a sampling frequency of 5kHz~20kHz to ensure the capture of transient fault characteristics), distribution network topology data (node ​​connection relationships, line resistance / inductance / capacitance parameters, location and capacity of distributed power sources), and historical fault handling data (historical fault type, fault occurrence time, corresponding ranging scheme, actual ranging results and errors, and handling completion time).

[0070] Step S1.2: Multidimensional data preprocessing

[0071] The steps for preprocessing the collected raw data are as follows:

[0072] Step S1.2.1: Data Cleaning

[0073] use The criteria filter out outliers in electrical quantity data (such as instantaneous pulse values ​​caused by lightning strikes) and use cubic spline interpolation to complete missing time-series data (such as local data loss caused by communication interruptions) to ensure data continuity.

[0074] Step S1.2.2: Data Conversion

[0075] One-dimensional electrical quantity time-series data (such as voltage instantaneous value sequences) is converted into a two-dimensional feature map, specifically through the sliding window method (with a window length set to...). 1 sampling point, window overlap rate set to 1. A two-dimensional matrix of "time-amplitude" is constructed to present the local fluctuation characteristics of electrical quantities in the form of an image, providing an appropriate input for subsequent image feature extraction.

[0076] Step S1.2.3: Normalization Process

[0077] Min-Max normalization is used to map electrical quantity data of different dimensions (such as voltage in kV and current in A) to the [0,1] interval to avoid interference from the difference in dimensions on feature extraction.

[0078] Step S1.2.4: Dataset Construction

[0079] A multi-dimensional fault feature dataset is constructed based on the structure of "fault characteristics (electrical quantity characteristics) - topology characteristics - operating condition characteristics - optimal ranging scheme label". The training set, validation set, and test set are further divided into... The proportions are divided, and the consistency of the distribution of fault types, topology, and operating conditions in each set is ensured.

[0080] S2: Construct a deep learning recommendation model and train it using the training dataset; the deep learning recommendation model extracts and fuses multimodal features in the following ways: extracting image spatial features from fault electrical quantity data based on convolutional neural networks, and performing nonlinear dimensionality reduction and adaptive redundancy removal on the image spatial features in sequence; extracting multi-scale gray-level scatter entropy features from fault electrical quantity data to quantify temporal dynamic complexity; fusing the redundancy-removed image spatial features with the multi-scale gray-level scatter entropy features, and weighting the fused features through an attention mechanism, finally outputting the compatibility probability with each candidate ranging scheme;

[0081] Figure 2 This is a schematic diagram of the feature extraction and fusion structure of the deep learning recommendation model provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the SKLAN-MDS feature dimensionality reduction process provided in an embodiment of the present invention; Figure 4This is a schematic diagram of the MGDE temporal feature extraction process provided in an embodiment of the present invention. Figure 2 , Figure 3 and Figure 4 As shown, step S2 is used to construct a deep learning recommendation model and perform model training and optimization. Specifically, the nonlinear dimensionality reduction of image spatial features involves replacing Euclidean distance with SKLAN pseudo-distance based on neighborhood symmetric KL divergence, combined with multidimensional scaling (MDS) technology to achieve dimensionality reduction. The adaptive redundancy removal of image spatial features involves using the Selective Principal Component Layer (SPCL) method, calculating feature amplitude and setting an adaptive threshold to filter out redundant features. The specific steps include:

[0082] Step S2.1: Extract image features

[0083] A convolutional neural network (CNN) is used as the base network. This network contains three convolutional layers (with kernel sizes of 3×3, 3×3, and 5×5, and the number of kernels being 32, 64, and 128, respectively), two max-pooling layers (with a pooling kernel size of 2×2 and a stride of 2), and one flattening layer. These layers are used to extract local key features (such as fault transient spikes and voltage / current phase abrupt change regions) from the two-dimensional feature map of fault electrical quantities. The specific process is as follows:

[0084] Step S2.1.1: Feature Dimensionality Reduction

[0085] Traditional Multidimensional Scaling (MDS) uses Euclidean distance to measure the similarity of high-dimensional data. However, the fault characteristics of distribution networks are high-dimensional and nonlinear, and Euclidean distance is susceptible to the curse of dimensionality, leading to distortion in dimensionality reduction. This paper proposes a pseudo-distance (SKLAN) based on neighborhood symmetric KL divergence to replace Euclidean distance, which, combined with MDS, achieves accurate dimensionality reduction.

[0086] Furthermore, the calculation of the SKLAN pseudo-distance in nonlinear feature dimensionality reduction includes: Neighborhood probability distribution construction: For any two feature points i and j, the neighborhood probability distribution extracted by the CNN is calculated based on the Gaussian kernel function for the high-dimensional image features; SKLAN pseudo-distance calculation: The difference between the neighborhood probability distributions of the two feature points is quantified using a symmetric KL divergence measure to obtain the SKLAN pseudo-distance; MDS dimensionality reduction optimization: The SKLAN pseudo-distance matrix is ​​used as the input to the MDS, and the high-dimensional features are mapped to a low-dimensional space by minimizing the stress function. Specific steps include:

[0087] Step S2.1.1.1: Construction of Neighborhood Probability Distribution

[0088] High-dimensional image features extracted by CNN (dimension set to) For any two feature points and The neighborhood probability distribution of the two is calculated based on the Gaussian kernel function. and ( (For any other point in the feature set)

[0089]

[0090] In the high-dimensional image features of distribution network faults, feature points With feature points The probability of a neighboring point is used to characterize a feature point. The local neighborhood structure distribution. The first feature in the image of a distribution network fault The high-dimensional feature vector of each feature point contains local key features of the two-dimensional feature map of fault electrical quantities (such as fault transient spikes and phase change region features). The first feature in the image of a distribution network fault The high-dimensional feature vector of each feature point is the feature point. Features of potential neighboring points. : Feature vector and The Euclidean distance is used to measure the similarity between two fault feature points. The first feature in the image of a distribution network fault The high-dimensional feature vectors of the feature points are used to construct the feature points. The global neighborhood reference. : Feature set except The index of any other feature point outside the specified range is used to traverse and calculate feature points. It is associated with the neighborhood of all points that are not its own feature points. Feature points The Gaussian kernel bandwidth is used to adjust the smoothness of the neighborhood probability distribution, adapting to the distribution characteristics of fault features in the distribution network. This is controlled by perplexity and satisfies the entropy constraint formula:

[0091]

[0092] perplexity: Neighborhood complexity control parameter, ranging from 20 to 50, is used to ensure that the neighborhood contains enough effective feature points of distribution network faults, balancing the accuracy of local structure characterization and computational efficiency. In practice, a grid search can be performed on the validation set (e.g., trying different values ​​between 20 and 50 with a step size of 5) to select the perplexity value that enables the dimensionality-reduced features to achieve the best performance in subsequent tasks (such as classification). For typical distribution network fault datasets, a perplexity value around 30 usually yields relatively stable results. Feature points The entropy value of the neighborhood probability distribution is used to constrain the rationality of the neighborhood distribution. The number of neighborhood points corresponding to perplexity, i.e., feature points. The number of feature points contained in the core neighborhood ensures that the neighborhood can effectively reflect the local correlation of fault characteristics in the distribution network.

[0093] Step S2.1.1.2: SKLAN pseudo-distance calculation

[0094] because and Reflecting feature points respectively and Given the local neighborhood structure, the difference between the two is quantified using the symmetric KL divergence to obtain the SKLAN pseudo-distance:

[0095]

[0096] Feature points in high-dimensional images of distribution network faults and The SKLAN pseudo-distance is used to quantify the differences in the local neighborhood structures of the two, providing a suitable similarity metric for subsequent MDS dimensionality reduction. Distribution network fault characteristics concentrated except The index of any feature point outside the range is used to traverse and calculate the difference in the neighborhood distribution of two feature points. Feature points With feature points Let be the probability of a neighboring point, and be the probability of a feature point. The core element of neighborhood distribution. Feature points With feature points Let be the probability of a neighboring point, and be the probability of a feature point. The core element of neighborhood distribution. Feature points and For the same neighborhood point The logarithm of the probability ratio is used to amplify the difference in the strength of the neighborhood association between the two. Feature points and For the same neighborhood point The logarithm of the probability ratio, together with the above logarithmic terms, constitutes a symmetric difference measure.

[0097] in, This is to avoid The resulting logarithmic undefined problem can be further simplified through approximate derivation as follows:

[0098]

[0099] Feature points and For neighboring points The probability difference is used to quantify the absolute difference in the strength of the neighborhood association between the two. Feature points and For neighboring points The probability logarithmic difference is used to reinforce the relative differences in neighborhood distributions.

[0100] This pseudo-distance satisfies symmetry. Nonnegativity It can effectively characterize the local structural differences of high-dimensional features, and its sensitivity to dimensional changes is lower than that of Euclidean distance.

[0101] Step S2.1.1.3: MDS dimensionality reduction optimization

[0102] Using the SKLAN pseudo-distance matrix as input to the MDS, the high-dimensional features are transformed into a low-dimensional space (dimension set to 1) by minimizing the stress function. Mapping between balancing computational cost and feature retention:

[0103]

[0104] Stress: The stress function value for MDS dimensionality reduction, used to quantify the structural differences of high-dimensional features of distribution network faults before and after dimensionality reduction. The smaller the value, the better the feature retains the original structure after dimensionality reduction. Feature points in high-dimensional features of distribution network faults and The SKLAN pseudo-distance reflects the strength of the local structural correlation between the two in the original high-dimensional space. Feature points in the lower-dimensional space after dimensionality reduction and The Euclidean distance is a low-dimensional mapping of the original SKLAN pseudo-distance. : Feature point index in the distribution network fault feature set, used to traverse all feature point pairs. : Feature point index in the distribution network fault feature set, and Constitute feature point pairs ( Ensure that each pair of feature points is calculated only once to avoid duplication. The sum of the squared differences between the SKLAN pseudo-distances and the low-dimensional Euclidean distances of all feature point pairs reflects the degree of deviation of the overall structure before and after dimensionality reduction. The sum of squares of the pseudo-distances of all feature points to the SKLAN is used as a normalization term to make the stress function value independent of the feature scale.

[0105] The low-dimensional feature vector is iteratively optimized using gradient descent to ensure that the Stress is less than 0.1, thus ensuring the structural consistency of the features after dimensionality reduction.

[0106] Step S2.1.2: Feature redundancy removal

[0107] Features after dimensionality reduction by SKLAN-MDS may still contain redundancy (such as linear correlation between different electrical quantity features). Therefore, Selective Principal Component Layer (SPCL) is proposed to filter out invalid information and retain key features that are highly adaptable to fault location schemes.

[0108] Furthermore, the Selective Principal Component Layer (SPCL) method is used to achieve adaptive redundancy removal, including: feature amplitude calculation: calculating the feature amplitude of each feature sample for the dimensionality-reduced feature matrix; and selection threshold determination: calculating the number of key features to be retained based on the feature selection threshold ratio r. The characteristic amplitudes of all samples are expanded into one-dimensional vectors, sorted in ascending order, and the nth vector is selected. The values ​​are used as the screening threshold; redundant feature filtering: for each sample, features with amplitudes greater than or equal to the threshold are retained, and redundant features are filtered out. Specifically, the steps are as follows:

[0109] Step S2.1.2.1: Calculation of characteristic amplitude

[0110] The dimensionality-reduced feature matrix ( For the sample size, (The feature dimension after dimensionality reduction), calculate each feature sample. The characteristic amplitude, quantifying the "importance" of the feature (the larger the amplitude, the higher the feature's ability to distinguish fault types / operating conditions):

[0111]

[0112] in, Each element corresponds to a sample The absolute value of the amplitude of a feature in a certain dimension.

[0113] Step S2.1.2.2: Determining the screening threshold

[0114] Set the feature filtering threshold ratio r (based on the complexity of distribution network fault features) As a preferred implementation, the initial value of r can be set to 0.7. Then, the model performance is evaluated on the validation set: if the validation set accuracy is more than 5 percentage points lower than the training set accuracy, overfitting is considered, and the value of r is increased by 0.05; if the model's single inference time on the validation set exceeds 50ms, the computational burden is considered too heavy, and the value of r is decreased by 0.05. The adjusted value of r should be maintained within the range of 0.6 to 0.8. The number of key features to be retained is calculated:

[0115]

[0116] The number of key features that need to be retained for distribution network fault characteristics, used to control the filtering scale of redundant features. Feature selection threshold ratio, value range The value is adaptively adjusted according to the complexity of the fault characteristics of the distribution network. When there are many fault types and complex topology, a higher value (such as 0.8) is taken to retain more key characteristics. The total number of samples in the distribution network fault feature dataset, covering samples of different fault types, topologies, and operating conditions. The feature dimension of each sample after dimensionality reduction by SKLAN-MDS is the number of features per sample after dimensionality reduction.

[0117] Characteristic amplitudes of all samples Expand into a one-dimensional vector Sort in ascending order and take the first one. One value is used as the filtering threshold:

[0118]

[0119] : No. The feature amplitude vector of a power distribution network fault sample, each element corresponds to the absolute value of the amplitude of a certain dimension feature of the sample, quantifying the feature's ability to distinguish fault scenarios. : A one-dimensional expanded vector of the feature amplitudes of all samples, which integrates the feature amplitude information of the entire dataset. :right The vectors sorted in ascending order are used to filter the amplitude thresholds of key features. Threshold: Feature filtering threshold, i.e., the amplitude threshold after sorting. The Each element value serves as the basis for determining whether a feature is a key feature. After sorting The Each element ensures that only the top-ranked amplitudes are retained. Key features.

[0120] Step S2.1.2.3: Redundant Feature Filtering

[0121] For each sample Features with amplitudes greater than or equal to a threshold are retained, while redundant features are filtered out to obtain the deredundant feature matrix. :

[0122]

[0123] The feature matrix of distribution network fault images after redundancy removal retains only the key features that significantly affect the adaptability of fault location schemes, with the dimension being the number of samples. Feature dimensions after redundancy removal. The first in the characteristic matrix The deredundant feature vector of the i-th sample, corresponding to the i-th Key image features of a power distribution network fault case. After dimensionality reduction by SKLAN-MDS, the first The original feature vector of each sample contains high-dimensional reduced features of the distribution network fault image (such as the spatial distribution features of fault transient pulses). Element-wise multiplication (Hadamard product) is used to multiply the original feature vector with the mask vector element by element to achieve zero-value filtering of redundant features. : No. The feature amplitude vector of each sample, where each element is the absolute value of the amplitude of the corresponding dimension feature, quantifies the ability of that dimension feature to distinguish between distribution network fault types / operating conditions. threshold: The feature screening threshold determined in step S2.1.2.2 is set based on the complexity of distribution network fault features and is used to define the critical value between key features and redundant features. threshold: a 0-1 mask vector, and... The dimensions are consistent, where an element of 1 indicates that the corresponding dimension feature is a key feature (amplitude meets the standard), and an element of 0 indicates that it is a redundant feature (amplitude does not meet the standard), which is used to accurately filter redundant information.

[0124] This step can filter out The redundant features reduce the computational load of the model while improving the correlation between the features and the ranging scheme.

[0125] Step S2.2: Extract temporal features

[0126] Electrical quantities during power distribution network faults (such as transient currents and voltage spikes) exhibit significant temporal dynamic characteristics, and traditional time-series feature extraction methods (such as mean and variance) struggle to capture their nonlinear variations. Step S2.2 proposes Gray-Scale Dispersion Entropy (GDE) and its multi-scale version (MGDE) to accurately quantify the complexity and dynamic characteristics of fault time sequences.

[0127] Furthermore, the specific process for extracting multi-scale entropy features from fault electrical quantity data to quantify the dynamic complexity of the time series is as follows: Calculating the multi-scale gray-level scattering entropy (MGDE) of the fault electrical quantity time series data includes: coarse-grained sequence generation: coarsening the original time series according to the scale factor s to generate a multi-scale sequence; multi-scale GDE calculation: calculating the gray-level scattering entropy (GDE) for each scale sequence, and finally concatenating the values ​​of the gray-level scattering entropy (GDE) of all scales to obtain the multi-scale gray-level scattering entropy feature vector. Specifically, calculating the gray-level scattering entropy (GDE) includes: mapping the standardized time series data to discrete category sequences and reconstructing the phase space to generate scattering patterns; constructing a transition probability matrix between scattering patterns and converting it into a gray-level matrix; and statistically analyzing the probability of occurrence of each gray value in the gray-level matrix. , Based on probability The gray-level scattering entropy is then calculated using the Shannon entropy formula. The specific steps include:

[0128] Step S2.2.1: Time series data preprocessing and mode conversion

[0129] Step S2.2.1.1 Category Mapping

[0130] The sequence after data standardization in step 1.3 Mapped to discrete category sequences This facilitates the subsequent construction of the distribution pattern:

[0131]

[0132] The first fault in the distribution network time sequence data The discrete category values ​​of each sampling point are used to convert continuous standardized data into a discrete form to adapt to the scattering pattern construction. The rounding function is used to convert the calculation result into an integer category, ensuring that each time series point corresponds to a unique discrete category. Number of categories, range of values Adjust the value according to the complexity of the fault timing in the distribution network (take a higher value, such as 10, when the fault transient fluctuations are severe and the patterns are rich, and take a lower value, such as 6, otherwise). The first fault in the distribution network timing sequence Standardized values ​​of each sampling point (mapped to) (Interval), which contains time-series data of electrical quantities such as voltage and current. Intermediate calculated values ​​for category mapping, which are scaled and offset to uniformly map standardized data to... The range should be such that it covers all values ​​after rounding. All categories. The discrete category sequence of distribution network fault time series has the same length as the original standardized sequence, providing discrete input for subsequent phase space reconstruction. The range of discrete category values ​​limits the category at each time point to between 1 and c, ensuring the uniformity and integrity of the category system.

[0133] Step S2.2.1.2: Phase space reconstruction and scattering mode generation

[0134] Based on Takens' embedding theorem, for class sequences Phase space reconstruction is performed to obtain the embedding vector:

[0135]

[0136] Category sequence The Middle The discrete class values ​​at each time point are the second element of the embedding vector, reflecting the delay. The subsequent temporal correlation. Category sequence The Middle The discrete category values ​​at each time point are the last element of the embedding vector, completely covering... Temporal correlations across multiple dimensions. The unique scattering pattern is obtained by embedding vector mapping, and each pattern corresponds to a dynamic association type of distribution network fault timing. : Index of the scattering pattern, range of values Used to uniquely identify each scattering pattern ( (Total number of all possible scattering patterns). The total number of scattering patterns, consisting of the number of categories. and embedding dimension The decision covers all possible dynamic correlation modes of distribution network fault timing. For the embedded dimension (set as) (to capture short-term dependencies in time series). The total number of embedded vectors, i.e., the number of vectors obtained after phase space reconstruction, ensures a comprehensive representation of the dynamic characteristics of distribution network fault timing. The formula is:

[0137]

[0138] For time delay (set to) (To avoid information redundancy). The total number of original sampling points for the fault time sequence of the distribution network provides the basis for calculating the total number of embedded vectors.

[0139] Step S2.2.2: Calculation of Gray-Scale Spread Entropy (GDE)

[0140] By constructing a grayscale matrix to quantize the transition relationship of the scattering pattern, and then calculating the GDE, the local dynamic complexity of the fault sequence is characterized:

[0141] Step S2.2.2.1: Construction of the transition probability matrix

[0142] Introducing transition step size (set as) (This captures the transitional relationships between non-adjacent patterns, overcoming the limitation of traditional LDE which only considers adjacent patterns) and calculates the scattering pattern. arrive Transition probability:

[0143]

[0144] Distribution patterns in the timing sequence of distribution network faults go through Step transition to distributed mode The probability is used to quantify the transition patterns between different dynamic modes. Transition step size, range of values It is used to capture the transitional relationship of non-adjacent scattering patterns (the larger the step size, the stronger the long-term temporal dependence it can reflect), which makes up for the limitation of traditional methods that only focus on adjacent patterns. The first fault sequence of the distribution network A distribution pattern represents a specific type of dynamic correlation (such as a certain combination of fluctuations in a fault transient). The first fault sequence of the distribution network Such a distribution pattern, for The target pattern after the transition. The mapping function from the embedded vector to the scattering pattern reconstructs the embedded vector from the phase space (e.g., ) is converted into the corresponding unique distribution pattern (e.g. ). : No. Each phase space embedding vector contains the fault time sequence of the distribution network. Time and delay Discrete category information. : No. Each phase space embedding vector, and interval Steps are used to capture non-adjacent transitions in a pattern. Number : A counting function used to count the indices of embedded vectors that meet certain conditions. The quantity. Simultaneously satisfying "the first Each embedding vector corresponds to a pattern And the Each embedding vector corresponds to a pattern index of " The total number. Representation pattern go through Step transition to The probability is used to construct the transition probability matrix. .

[0145] Step S2.2.2.2: Grayscale matrix conversion

[0146] Transition probability matrix The elements are linearly mapped to The grayscale matrix is ​​obtained by dividing the range. To achieve visualization and quantification of probability information:

[0147]

[0148] The grayscale matrix of the first Line number The elements of the column (grayscale values) are determined by the transition probability. The range of values ​​obtained through mapping round The rounding function converts the mapped non-integer result into an integer grayscale value, which conforms to the integer characteristics of image grayscale values. The formula for linearly mapping transition probability to grayscale value is used to scale the probability value. Range) mapped to grayscale value ( (Interval). Gray-scale matrix, dimension 1 The matrix elements are grayscale values. This enables the visualization and quantitative representation of the transition probability matrix. The definition of a grayscale matrix indicates that the matrix's first... Line number The elements of the column are The dimension is consistent with the transition probability matrix.

[0149] Step S2.2.2.3: GDE Calculation

[0150] Statistical analysis of the probability of occurrence of each gray value in a gray-scale matrix Calculate GDE based on Shannon entropy formula:

[0151]

[0152]

[0153] Distribution network fault timing data The gray-scale distribution entropy is the comprehensive category number. Embedded Dimension Time delay and transition step size The calculated entropy value is used to quantify the local dynamic complexity of the fault sequence (the larger the value, the more complex the sequence pattern). The original time-series data of distribution network faults (such as voltage and current transient sequences) are the input objects for GDE calculations. The number of categories (same as step S2.2.1.1) affects the number and granularity of the scattering patterns, which in turn affects the accuracy of GDE in characterizing complexity. Embedding dimension (same as step S2.2.1.2) determines the number of temporal points contained in each embedding vector in phase space reconstruction, affecting the ability to capture dynamic patterns. Time delay (same as step S2.2.1.2) controls the interval between time points in the embedding vector, affecting the representation of short-term temporal dependencies. Transition step size (same as step S2.2.2.1) determines the number of interval steps for pattern transition and affects the capture of non-adjacent pattern associations. Grayscale value The probability of occurrence in the grayscale matrix reflects the frequency of the mode transition relationship corresponding to the grayscale value in the fault timing sequence. Grayscale value, range of values This corresponds to the quantized value after the transition probability mapping. : Counting function, counting the values ​​in a grayscale matrix The total number of elements, i.e., grayscale value. Number of times it appears. The total number of elements in the grayscale matrix is ​​equal to the total number of scattering patterns. The square of is used as the normalized denominator for probability calculation (ensuring...). The sum is 1). The grayscale matrix of the first Line number The elements of the column (same as step 2.2.2.2) represent transition probabilities. The mapped grayscale value. : The logarithm of the probability of grayscale values ​​occurring, used in Shannon entropy calculation to amplify the contribution of low-probability events to the entropy value. Shannon entropy is the core calculation term, which summarizes the product of the probability and logarithmic probability of all gray values.

[0154] The larger the GDE value, the higher the dynamic complexity of the fault sequence (e.g., the GDE value of a phase-to-phase short circuit is usually higher than that of a single-phase ground fault).

[0155] Step S2.2.3: Calculation of Multiscale Gray-Level Spread Entropy (MGDE)

[0156] Single-scale GDEs struggle to capture the changing patterns of fault time series across different time scales (such as short-term transient fluctuations and long-term trends). Therefore, a coarse-grained MGDE is constructed to achieve multi-scale feature extraction.

[0157] Step S2.2.3.1: Coarse-grained sequence generation

[0158] For raw time series data According to scale factor (set as) It covers different time scales to ensure the coarsest-grained sequence length. Divide the sequence (with a minimum of 50 sequences and sufficient sequence structure information) into multiple non-overlapping subsequences, and calculate the mean of each subsequence as the coarse-grained sequence. Elements:

[0159]

[0160] : Scale factor is At that time, coarse-grained sequence The Each element reflects the local average characteristics of the fault sequence in the distribution network at this scale. Scale factor, range of values The time scale used to divide the original time series ( The smaller the focus, the more short-term details. The larger the value, the more it emphasizes long-term trends. The original time-series data of distribution network faults (such as voltage and current transient sequences) are input data for coarse-grained processing. : Scale factor is The generated coarse-grained sequence has a length of It is a simplified representation of the original time series at the corresponding scale. : Element index of the coarse-grained sequence, range of values Iterate through all the coarsened elements. Raw time series data The Each sampling point serves as the basic data unit for the timing of faults in the distribution network. : Scale factor is At that time, the first The summation of all sample points of non-overlapping subsequences provides the basis for calculating the mean of the subsequences. The normalization coefficient for mean calculation ensures that the coarse-grained elements reflect the average level of the corresponding subsequence. Raw time series data The total number of sampling points (same as step 2.2.1.2) determines the maximum possible length of the coarse-grained sequence. The floor function is used to determine the length of the coarse-grained sequence, avoiding errors caused by rounding down. Cannot be The problem of remainders caused by divisibility. : Scale factor is The length of the coarse-grained sequence is determined to ensure that all subsequences are non-overlapping complete fragments.

[0161] Step S2.2.3.2: Multi-scale GDE calculation

[0162] For each coarse-grained sequence Repeat steps 2.2.1 to 2.2.2 to obtain the GDE values ​​at the corresponding scale. By concatenating the GDE values ​​at all scales, we obtain the multi-scale gray-level scatter entropy feature vector:

[0163]

[0164] Distribution network fault timing data The multi-scale gray-level scattering entropy feature vector integrates complexity information at different time scales, providing comprehensive temporal features for ranging scheme recommendations. : Scale factor is Time coarsening sequence The grayscale distribution entropy value is used to quantify the local dynamic complexity of the fault time sequence at this scale. : Scale factor is The coarse-grained sequence at that time (same as step 2.2.3.1) is the input data for the GDE calculation at the corresponding scale. Total number of scales, range of values Typically, s=1, 2, ..., 10, i.e., S=10, can be selected to cover the complete range from the original scale to the maximum scale, and to cover multiple time scales from the short term to the long term, ensuring the comprehensiveness of the feature vector. The multi-scale gray-level scattering entropy feature vector is constructed by concatenating the GDE values ​​of each scale in ascending order of scale factor to form a one-dimensional feature vector.

[0165] MGDE can simultaneously characterize both short-term fluctuations (GDE corresponding to small-scale factors) and long-term trends (GDE corresponding to large-scale factors) of fault time series, providing more comprehensive time series feature support for subsequent scheme recommendations.

[0166] Step S2.3: Multimodal fusion

[0167] The redundant image features (dimensions) extracted in step S2.1 are then processed. The multi-scale gray-level scattering entropy features obtained in step S2.2, i.e., time series features (dimensions) Multimodal fusion is performed to strengthen the correlation between different features. Specific implementation details include:

[0168] Step S2.3.1: Feature concatenation and dimensionality unification

[0169] Image feature matrix With MGDE feature matrix By concatenating along the feature dimensions, a fused feature matrix is ​​obtained. .

[0170] The deredundant image feature matrix obtained in step 2.1.2.3 has a dimension of [missing information]. It contains key image features of distribution network faults (such as the spatial distribution of fault areas). : Multi-scale gray-level scattering entropy feature matrix, dimension is It includes the multi-scale dynamic complexity characteristics of the fault timing in the distribution network. : Sample size (same as step S2.1.2.2), i.e. the total number of distribution network fault cases, to ensure that the sample dimensions of the two feature matrices are consistent. Dimensions of redundant image features, and their range of values. This reflects the dimensionality of image features. : Dimensions of multi-scale gray-level scattering entropy features, and their value range , equal to the total number of scales The dimensional scale that reflects the temporal characteristics. The fused feature matrix obtained after feature concatenation has a dimension of [missing information]. It integrates all information from image features and temporal features.

[0171] Step S2.3.2: Weighted Attention Mechanism

[0172] A channel attention module is introduced to assign weights to different dimensions of the fused features, thereby strengthening the correlation between topology (such as line length and number of nodes), operating conditions (such as load factor and distributed power output), and fault type.

[0173]

[0174] Attention weight vector, with dimension 1 Each element corresponds to the importance weight of a certain dimension of the fusion feature (the larger the value, the stronger the correlation between the feature in that dimension and the ranging scheme). The sigmoid activation function maps weight values ​​to... The interval ensures the interpretability of the weights (representing the relative importance of feature dimensions). The first layer weight matrix of the attention module has dimensions of . This is used to map fused features from the original dimension to the hidden layer dimension. The second-layer weight matrix of the attention module has dimensions of [dimension number missing]. This is used to map hidden layer features back to the original feature dimensions and output a weight vector. The ReLU activation function introduces a non-linear transformation, enhancing the attention module's ability to capture feature associations. The transpose of the fused feature matrix, with dimension 1. It adapts to the matrix operation requirements of the attention module. The first-layer bias term of the attention module has the following dimensions: This is used to adjust the baseline values ​​of the hidden layer features. The second-layer bias term of the attention module has the following dimensions: This is used to adjust the baseline value of the weight vector. Attention hidden layer dimension, with values... This reduces computational complexity while preserving key feature association information. The attention-weighted feature matrix has the following dimensions: This strengthens the feature dimensions that are strongly correlated with the ranging scheme and suppresses redundant information. : by weight vector The constructed diagonal matrix has dimensions of This enables weighted scaling of each dimension of the fused features.

[0175] Step S2.3.3: Design of the Scheme Matching Layer

[0176] Build a fully connected layer at the model output, Mapped to the ranging scheme space (set as) One candidate solution, The fitness probability of each solution is output through the softmax function:

[0177]

[0178] The probability matrix for recommended solutions has the following dimensions: Each row represents a pair of distribution network fault samples. The fit probability distribution of each candidate ranging scheme. The softmax activation function converts the output of a fully connected layer into a probability distribution (the sum of the elements in each row is 1), making it easier to directly interpret the suitability of the solution. The weight matrix of the scheme matching layer, with dimension 1. This characterizes the correlation strength between attention-weighted features and various ranging schemes. : Bias terms of the scheme matching layer, with dimensions of This is used to adjust the baseline adaptation probability of each ranging scheme. The number of candidate ranging schemes, and their range. It is a mainstream ranging solution that covers common fault scenarios in power distribution networks. The attention-weighted feature matrix (same as step S2.3.2) is the input feature of the scheme matching layer.

[0179] Each row corresponds to a sample's scheme adaptation probability distribution, and finally an end-to-end recommendation model of "feature extraction - feature fusion - scheme matching" is constructed.

[0180] Step S2.4, Model Training and Optimization

[0181] A multi-dimensional fault feature dataset is constructed based on the structure of "fault electrical quantity characteristics - topological characteristics - operating condition characteristics - optimal ranging scheme label". The training set, validation set, and test set are further divided into... The proportions are divided; the model is trained using historical fault data and the corresponding optimal ranging scheme's labeled data to ensure the model's generalization ability under different fault scenarios:

[0182] Step S2.4.1: Loss Function Design

[0183] The cross-entropy loss function is used to measure the difference between the probability distribution of the scheme output by the model and the true label (one-hot encoding).

[0184] Step S2.4.2: Optimizer and Learning Rate Adjustment

[0185] The Adam optimizer was used (initial learning rate set to 100%). Weight decay is set to The learning rate is dynamically adjusted based on the validation set loss during training: if the validation set loss does not decrease for three consecutive epochs, the learning rate is halved.

[0186] Step S2.4.3: Cross-validation and regularization

[0187] Five-fold cross-validation is used to divide the training and validation sets to avoid overfitting; dropout is also introduced into the fully connected layer (dropout rate set to...). Randomly deactivate some neurons to improve the model's generalization ability.

[0188] Step S2.4.4: Model Convergence Determination

[0189] When the training set loss is less than 0.05, and the fluctuation of the validation set loss over 5 consecutive epochs is less than... When the model is deemed to have converged, training is stopped.

[0190] S3: When a fault occurs in the distribution network, real-time fault data is collected and input into the trained deep learning recommendation model to output a recommended ranging scheme.

[0191] Recommendation and feedback mechanism

[0192] Step S3.1: Real-time Recommendation

[0193] When a fault occurs in the distribution network, the system outputs an appropriate ranging solution according to the following procedure:

[0194] Step S3.1.1: Real-time data acquisition

[0195] Data is collected after a fault occurs through distribution network monitoring terminals (such as FTUs and DTUs). Electrical quantity data (voltage, current), current topology and operating conditions for each cycle.

[0196] Step S3.1.2: Feature Extraction and Input

[0197] The collected real-time fault data of the distribution network is processed according to steps S1 to S2.3 to extract image features and time-series features of the real-time faults and fuse them to obtain a real-time feature vector. .

[0198] Step S3.1.3: Solution Output

[0199] Will Input the trained model, output the adaptation probability of each ranging scheme, and select the top scheme in descending order of probability. Each option is presented as a recommendation, along with its confidence level (probability value), for the operation and maintenance personnel to make a decision.

[0200] Step 3.2 Feedback Iteration

[0201] Preferably, the method of the present invention further includes: updating the training dataset and iteratively optimizing the deep learning recommendation model based on the actual performance of the recommended ranging scheme, specifically including: if the ranging error of the recommended scheme is greater than a preset threshold, updating the sample label of the fault to the scheme with the best actual performance, and periodically using the accumulated feedback data to incrementally train the model. This constructs a closed-loop feedback mechanism to continuously optimize model performance, specifically including the following steps:

[0202] Step 3.2.1: Feedback Data Collection

[0203] Collect the execution results of the actual ranging scheme, including ranging error (such as the deviation between the actual fault location and the calculated location), handling time, and fault elimination effect.

[0204] Step S3.2.2: Tag Update

[0205] If the ranging error of the recommended solution is greater than a preset threshold (e.g., 1km), the label of the fault sample will be updated to the solution with the best actual effect.

[0206] Step S3.2.3: Model Iteration

[0207] For every 100-200 feedback data points accumulated, the model parameters are updated using incremental training (the learning rate is set to 1 / 10 of the initial value) to ensure that the model is continuously optimized as the actual scenario changes, and the recommendation accuracy improves by 3%-5% per year.

[0208] Example 2

[0209] According to a second aspect of the present invention, a distribution network fault location scheme recommendation system 200 is also provided, for implementing the distribution network fault location scheme recommendation method 100 as described in the first embodiment. The system 200 includes:

[0210] Fault data acquisition module 210 is used to collect real-time fault data of the power distribution network;

[0211] The ranging scheme recommendation module 220 inputs the real-time fault data of the power distribution network into the trained deep learning recommendation model, and extracts and fuses multimodal features in the following ways: extracting image spatial features from the fault electrical quantity data based on a convolutional neural network, and performing nonlinear dimensionality reduction and adaptive redundancy removal on the image spatial features in sequence; extracting multi-scale entropy features from the fault electrical quantity data to quantify the temporal dynamic complexity; fusing the redundancy-removed image spatial features with the multi-scale entropy features, and weighting the fused features through an attention mechanism, and finally outputting the recommended ranging scheme;

[0212] The update and iteration optimization module 230 is used to update the training dataset and iteratively optimize the deep learning recommendation model based on the actual performance of the recommended ranging scheme.

[0213] Those skilled in the art will understand that, for the sake of convenience and brevity, the various modules described in the system description can be referred to the working process of the corresponding steps in the foregoing method embodiments, and will not be repeated here.

[0214] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0215] Example 3

[0216] Figure 6 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0217] like Figure 6 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0218] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0219] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 executes the recommended method 100 for distribution network fault location scheme described above.

[0220] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0221] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A power distribution network fault location scheme recommendation method, characterized in that, The method comprises the following steps: Collecting multi-dimensional data when a fault occurs in a power distribution network, and preprocessing the multi-dimensional data to construct a training data set; The multi-dimensional data comprises fault electrical quantity data, power distribution network topology data and historical fault handling data; A deep learning recommendation model is constructed, and the deep learning recommendation model is trained using the training data set; the deep learning recommendation model extracts and fuses multi-modal features in the following manner: image spatial features are extracted from the fault electrical quantity data based on a convolutional neural network, and the image spatial features are sequentially subjected to nonlinear dimension reduction and adaptive redundancy removal processing; multi-scale gray scatter entropy features for quantifying time series dynamic complexity are extracted from the fault electrical quantity data; the image spatial features after redundancy removal are fused with the multi-scale gray scatter entropy features, and the fused features are weighted through an attention mechanism, and finally the adaptation probability of each candidate ranging scheme is outputted; When a fault occurs in the power distribution network, real-time fault data is collected and inputted into the trained deep learning recommendation model, and a recommended ranging scheme is outputted.

2. The power distribution network fault location scheme recommendation method of claim 1, wherein, The nonlinear dimension reduction of the image spatial features specifically comprises: using a SKLAN pseudo distance based on a neighborhood symmetric KL divergence to replace the Euclidean distance, and combining a multidimensional scaling technology to realize dimension reduction; The adaptive redundancy removal processing of the image spatial features specifically comprises: using a selective principal component layer method, calculating feature amplitudes and setting an adaptive threshold, and filtering out redundant features.

3. The power distribution network fault location scheme recommendation method of claim 2, wherein, The calculation of the SKLAN pseudo distance in the nonlinear feature dimension reduction comprises: Neighborhood probability distribution construction: for any two feature points i and j, the neighborhood probability distribution of the two feature points is calculated based on a Gaussian kernel function; SKLAN pseudo distance calculation: the difference between the neighborhood probability distributions of the two feature points is quantified through symmetric KL divergence to obtain the SKLAN pseudo distance, and the calculation formula is: wherein, : feature point with feature point as the neighborhood point; : feature point with feature point as the neighborhood point; : feature point and SKLAN pseudo-distance of the feature point in the power distribution network fault high-dimensional feature, reflecting the local structure correlation strength of the two in the original high-dimensional space; MDS dimension reduction optimization: the SKLAN pseudo distance matrix is taken as the MDS input, and the mapping from the high-dimensional features to the low-dimensional space is realized by minimizing the stress function, wherein the formula for minimizing the stress function is: Wherein, Stress: the stress function value of MDS dimension reduction, used to quantify the structural difference before and after the dimension reduction of the high-dimensional feature of power distribution network fault; : the Euclidean distance between the feature points in the low-dimensional space after dimension reduction and ; : the low-dimensional mapping result of the original SKLAN pseudo-distance; : the index of the feature points in the feature set of power distribution network fault, used to traverse all feature point pairs; : the index of the feature points in the feature set of power distribution network fault, and : the index of the feature points in the feature set of power distribution network fault, and 4. The power distribution network fault location scheme recommendation method of claim 2, wherein, The adaptive redundancy removal processing is realized by using a selective principal component layer method, which comprises: Feature amplitude calculation: the feature amplitude of each feature sample is calculated for the dimension-reduced feature matrix; Screening threshold determination: according to the feature screening threshold ratio r, the number of key features to be retained is calculated , and the feature amplitudes of all samples are expanded into one-dimensional vectors, sorted in ascending order, and the first value is taken as the screening threshold; Redundant feature filtering: for each sample, the features with an amplitude greater than or equal to the threshold are retained, and the redundant features are filtered out.

5. The power distribution network fault location scheme recommendation method of claim 1, wherein, The specific process of extracting the multi-scale gray scatter entropy features for quantifying the time series dynamic complexity from the fault electrical quantity data comprises: Calculating the multi-scale gray scatter entropy of the fault electrical quantity time series data, which comprises: Coarse-grained sequence generation: the original time series is subjected to coarse-grained processing according to a scale factor s to generate a multi-scale sequence; Multi-scale gray scatter entropy feature calculation: the gray scatter entropy of each scale sequence is calculated, and finally the values of the gray scatter entropies of all scales are spliced to obtain a multi-scale gray scatter entropy vector.

6. The power distribution network fault location scheme recommendation method of claim 5, wherein, The calculation of the gray scatter entropy comprises: Mapping the normalized time series data to a discrete category sequence, and performing phase space reconstruction to generate a scatter pattern; Constructing a transition probability matrix between the scatter patterns and converting it into a gray matrix; counting the occurrence probability of each gray value in the gray matrix , , calculating the gray dispersion entropy based on the probability and using the Shannon entropy formula wherein, : power distribution network fault timing data : gray scale scatter entropy : probability of occurrence of gray value t : number of categories : embedding dimension : time delay : transition step, value range 2~5 : element in the gray scale matrix, row i, column j, transition probability : element in the gray scale matrix, row i, column j, transition probability : element in the gray scale matrix, row i, column j, transition probability : mapped gray value : counting function, total number of elements in the gray scale matrix with value i, i.e., the number of times the gray value i appears : counting function, total number of elements in the gray scale matrix with value i, i.e., the number of times the gray value i appears : counting function, total number of elements in the gray scale matrix with value i, i.e., the number of times the gray value i appears 7. The power distribution network fault location scheme recommendation method of claim 6, wherein, The weighting of the fused features through the attention mechanism includes: The channel attention module is used to assign weights to different dimensions of the fused features, and then the fused features are weighted using the weights, wherein the weight calculation method is: wherein, is a fusion feature matrix, which is obtained by fusing the image spatial features after de-redundancy and the multi-scale gray scatter entropy matrix along the feature dimension to obtain; : the number of samples; : the dimension of the image spatial features after de-redundancy; : the dimension of the multi-scale gray scatter entropy features; : the first layer weight matrix of the attention module; : the second layer weight matrix of the attention module; : the ReLU activation function; : the transpose of the fusion feature matrix; : the first layer bias term of the attention module; : the second layer bias term of the attention module; : the feature matrix after attention weighting; : the diagonal matrix constructed by the weight vector .

8. The power distribution network fault location scheme recommendation method of claim 1, wherein, The training and optimization of the deep learning recommendation model includes: The multi-dimensional fault feature dataset is constructed in the structure of "fault electrical quantity feature-topology feature-working condition feature-optimal distance measurement scheme label", wherein the training set, the verification set and the test set are divided in the proportion of 6:2:

2. ​ The model is trained using historical fault handling data and corresponding label data of the optimal ranging scheme, and the cross-entropy loss function and Adam optimizer are used to improve the generalization ability of the model through cross-validation and regularization.

9. The power distribution network fault location scheme recommendation method of claim 1, wherein, The method further includes: updating the training data set and iteratively optimizing the deep learning recommendation model according to the actual execution effect of the recommended ranging scheme, specifically including: If the ranging error of the recommended scheme is greater than a preset threshold, the sample label of this fault is updated to the scheme with the best actual effect, and the model is incrementally trained using the accumulated feedback data at regular intervals.

10. A power distribution network fault location scheme recommendation system for implementing the power distribution network fault location scheme recommendation method according to any one of claims 1 to 9, characterized in that, The system includes: A fault data acquisition module is configured to acquire real-time fault data of a power distribution network; A ranging scheme recommendation module is configured to input the real-time fault data of the power distribution network into the trained deep learning recommendation model, extract and fuse multi-modal features through the following manner: image spatial features are extracted from fault electrical quantity data based on a convolutional neural network, and the image spatial features are sequentially subjected to nonlinear dimension reduction and adaptive redundancy removal processing; multi-scale gray scatter entropy features for quantifying time series dynamic complexity are extracted from the fault electrical quantity data; the image spatial features after redundancy removal are fused with the multi-scale gray scatter entropy features, and the fused features are weighted through an attention mechanism, and finally a recommended ranging scheme is outputted; An update and iterative optimization module is configured to update the training data set and iteratively optimize the deep learning recommendation model according to the actual execution effect of the recommended ranging scheme.

Images