Power distribution network hierarchical fault diagnosis method, device and equipment and storage medium

Abstract

This application relates to a hierarchical fault diagnosis method, apparatus, equipment, and storage medium for power distribution networks. It aims to improve the accuracy of fault diagnosis. The hierarchical fault diagnosis method for power distribution networks includes: acquiring real-time three-phase current signals of key nodes in the power distribution network under the current fault scenario; extracting fault window signals containing fault information based on the real-time three-phase current signals; converting the fault window signals from one-dimensional time-domain signals into two-dimensional images using rasterization technology; and inputting the two-dimensional image into a pre-trained hierarchical deep convolutional neural network model to obtain the fault type and phase classification results, fault section identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model.

Description

Technical Field

[0001] This invention relates to the field of power distribution network fault diagnosis technology, and in particular to a method, apparatus, equipment and storage medium for hierarchical fault diagnosis of power distribution networks. Background Technology

[0002] Existing technologies typically employ machine learning methods based on manual feature extraction for distribution network fault diagnosis. These methods first transform the acquired one-dimensional time-domain signal using signal processing techniques such as wavelet transform, S-transform, or Hilbert-Huang transform. Based on expert experience, specific fault feature vectors, such as energy, entropy, and amplitude, are manually extracted using rules. These extracted features are then input into traditional classifier models such as support vector machines, artificial neural networks, or extreme learning machines for training and testing. Finally, the model output determines the fault type or approximate location. However, these methods still have limitations in practical applications: the signal processing and feature extraction processes rely on expert experience, the quality of feature selection directly affects the diagnostic effect, and the computational complexity is high; traditional classifier models have limited generalization ability, making it difficult to guarantee diagnostic accuracy when faced with the complex and variable fault waveforms of distribution networks. Summary of the Invention

[0003] This invention provides a method, apparatus, device, and storage medium for hierarchical fault diagnosis of power distribution networks, in order to solve the problems in the prior art where fault feature information is lost due to reliance on manual feature extraction, and the traditional classifier model has limited generalization ability, resulting in low diagnostic accuracy when facing complex and ever-changing fault waveforms in power distribution networks.

[0004] The first aspect of this invention provides a hierarchical fault diagnosis method for a distribution network, comprising: acquiring real-time three-phase current signals of key nodes in the distribution network under the current fault scenario; extracting fault window signals containing fault information based on the real-time three-phase current signals; converting the fault window signals from a one-dimensional time-domain signal into a two-dimensional image using rasterization technology; inputting the two-dimensional image into a pre-trained hierarchical deep convolutional neural network model to obtain fault type and phase classification results, fault section identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model, wherein the hierarchical deep convolutional neural network model is trained based on training samples generated in a dynamic simulation environment that considers the output fluctuation of distributed power sources and the uncertainty of load demand, and the dynamic simulation environment uses a probability distribution model to quantify the uncertainty.

[0005] In one feasible implementation, before acquiring the real-time three-phase current signals of key nodes in the distribution network under the current fault scenario, the method further includes constructing a hierarchical deep convolutional neural network model: determining the probability distribution model corresponding to the output fluctuations and load demand randomness of distributed power sources in the distribution network; based on the probability distribution model, combined with a real-time digital simulator, constructing a distribution network simulation environment capable of simulating the dynamic changes in distributed power source output and load demand; based on the distribution network simulation environment, generating simulated fault scenarios covering multiple preset fault types, fault locations, and fault parameters, and collecting the original three-phase current signals of key nodes in the distribution network under each simulated fault scenario as training samples; preprocessing and rasterizing the training samples to generate corresponding two-dimensional image training sets; constructing an initial neural network model, and training the neural network model using the two-dimensional image training sets until the model converges, thus obtaining a hierarchical deep convolutional neural network model.

[0006] In one feasible implementation, acquiring the real-time three-phase current signal of a key node in the distribution network under the current fault scenario includes: deploying phasor measurement units at key nodes in the distribution network, wherein the key nodes include at least one of bus outlet, distributed power grid connection point, feeder sectionalizing switch, and load access point; configuring the sampling frequency and acquisition duration of the phasor measurement units, wherein the sampling frequency is set to a sampling rate capable of capturing high-frequency details of the fault transient signal, and the acquisition duration covers at least the continuous time period before and after the fault occurs; monitoring the operating status of the distribution network in real time, and when a fault is detected, triggering the phasor measurement units to synchronously acquire three-phase current signals according to the configured sampling parameters as real-time three-phase current signals.

[0007] In one feasible implementation, the step of extracting a fault window signal containing fault information based on the real-time three-phase current signal includes: determining the fault occurrence time based on the waveform characteristics or abrupt change point detection of the real-time three-phase current signal; using the fault time as a reference, extracting a pre-fault period signal of a preset length forward and a fault period signal of a preset length backward to form a continuous data window containing fault information; and filtering or denoising the three-phase current signal within the continuous data window to obtain the fault window signal.

[0008] In one feasible implementation, the step of using rasterization technology to convert the fault window signal from a one-dimensional time-domain signal into a two-dimensional image includes: normalizing the fault window signal, mapping the normalized signal sample points to a two-dimensional pixel coordinate system, wherein the time series index of the signal corresponds to the pixel coordinates in the width direction of the two-dimensional image, the normalized amplitude of the signal corresponds to the pixel coordinates in the height direction of the two-dimensional image, and each sample point corresponds to a pixel coordinate point; using the Bressenham line algorithm, sequentially connecting the pixel coordinate points corresponding to adjacent sample points to construct a continuous signal trajectory; assigning a preset grayscale intensity or channel value to the pixel points corresponding to the continuous signal trajectory to generate a two-dimensional image.

[0009] In one feasible implementation, the step of inputting the two-dimensional image into a pre-trained hierarchical deep convolutional neural network model to obtain the fault type and phase classification results, fault segment identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model includes: Two-dimensional images corresponding to real-time three-phase current signals collected from a single node in a key node of the distribution network are input into the first-level network of the hierarchical deep convolutional neural network model. The first-level network extracts image features through convolutional layers, pooling layers, and fully connected layers, and outputs the fault type and phase classification results through a softmax function. The fault type and phase classification include single-phase grounding, two-phase grounding, three-phase grounding, or no-fault scenarios. Two-dimensional images corresponding to real-time three-phase current signals collected from multiple nodes in the key node of the distribution network are fused along the channel dimension to generate a fused image. The fused image is input into the second-level network of the hierarchical deep convolutional neural network model. The second-level network, based on the shared features extracted by the first-level network, performs feature fusion through multiple fully connected layers and outputs the fault section identification results through a softmax function. The fused image is input into the third-level network of the hierarchical deep convolutional neural network model. The third-level network shares the feature extraction modules of the first-level and second-level networks, and its output layer uses a linear activation function to directly output the fault distance regression estimation results between the fault point and the reference node.

[0010] In one feasible implementation, training the neural network model using the two-dimensional image training set until the model converges to obtain a hierarchical deep convolutional neural network model includes: training the neural network model using an optimizer with an adaptive learning rate, and setting the corresponding initial learning rate, batch size, and number of iterations; for fault type and phase classification tasks and fault segment identification tasks, using the category cross-entropy loss function to calculate the difference between the predicted probability and the true label; for fault distance regression estimation tasks, using the mean squared error loss function to calculate the squared difference between the predicted distance and the true distance; dividing the two-dimensional image training set into a training set and a validation set, calculating the model's performance index on the validation set after each training iteration cycle, and selecting the model with the best performance on the validation set as the converged hierarchical deep convolutional neural network model based on the performance index.

[0011] A second aspect of the present invention provides a hierarchical fault diagnosis device for a distribution network, comprising: an acquisition module for acquiring real-time three-phase current signals of key nodes in the distribution network under the current fault scenario; an extraction module for extracting fault window signals containing fault information based on the real-time three-phase current signals; a conversion module for converting the fault window signals from a one-dimensional time-domain signal into a two-dimensional image using rasterization technology; and an output module for inputting the two-dimensional image into a pre-trained hierarchical deep convolutional neural network model to obtain fault type and phase classification results, fault section identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model, wherein the hierarchical deep convolutional neural network model is trained based on training samples generated in a dynamic simulation environment that considers the output fluctuations of distributed power sources and the uncertainty of load demand, and the dynamic simulation environment uses a probability distribution model to quantify the uncertainty.

[0012] In one feasible implementation, the hierarchical fault diagnosis of the distribution network includes: a construction module for determining the probability distribution model corresponding to the output fluctuations and load demand randomness of distributed power sources in the distribution network; based on the probability distribution model, combined with a real-time digital simulator, constructing a distribution network simulation environment capable of simulating the dynamic changes in the output of distributed power sources and load demand; based on the distribution network simulation environment, generating simulated fault scenarios covering multiple preset fault types, fault locations, and fault parameters, and collecting the original three-phase current signals of key nodes in the distribution network under each simulated fault scenario as training samples; preprocessing and rasterizing the training samples to generate corresponding two-dimensional image training sets; constructing an initial neural network model, and training the neural network model using the two-dimensional image training set until the model converges, obtaining a hierarchical deep convolutional neural network model.

[0013] In one feasible implementation, the acquisition module is specifically used to: deploy phasor measurement units at key nodes of the distribution network, the key nodes including at least one of bus outlet, distributed power grid connection point, feeder sectionalizing switch, and load access point; configure the sampling frequency and acquisition duration of the phasor measurement units, the sampling frequency being set to a sampling rate capable of capturing high-frequency details of fault transient signals, and the acquisition duration covering at least the continuous time period before and after the fault occurs; monitor the operating status of the distribution network in real time, and when a fault is detected, trigger the phasor measurement units to synchronously acquire three-phase current signals according to the configured sampling parameters, as real-time three-phase current signals.

[0014] In one feasible implementation, the extraction module is specifically used to: determine the fault occurrence time based on the waveform characteristics or abrupt change point detection of the real-time three-phase current signal; based on the fault time, extract a pre-fault period signal of a preset length forward and a fault period signal of a preset length backward to form a continuous data window containing fault information; and perform filtering or noise reduction processing on the three-phase current signal in the continuous data window to obtain the fault window signal.

[0015] In one feasible implementation, the conversion module is specifically used to: normalize the fault window signal, map the normalized signal sample points to a two-dimensional pixel coordinate system, wherein the time series index of the signal corresponds to the pixel coordinates in the width direction of the two-dimensional image, the normalized amplitude of the signal corresponds to the pixel coordinates in the height direction of the two-dimensional image, and each sample point corresponds to a pixel coordinate point; using the Bressenham line algorithm, sequentially connect the pixel coordinate points corresponding to adjacent sample points to construct a continuous signal trajectory; assign a preset grayscale intensity or channel value to the pixel points corresponding to the continuous signal trajectory to generate a two-dimensional image.

[0016] In one feasible implementation, the output module is specifically used for: inputting a two-dimensional image corresponding to the real-time three-phase current signal collected by a single node in a key node of the distribution network into the first-level network of the hierarchical deep convolutional neural network model. The first-level network extracts image features through convolutional layers, pooling layers, and fully connected layers, and outputs the fault type and phase classification results through a softmax function. The fault type and phase include single-phase grounding, two-phase grounding, three-phase grounding, or no-fault scenarios. The module also fuses the two-dimensional images corresponding to the real-time three-phase current signals collected by multiple nodes in the key node of the distribution network along the channel dimension to generate a fused image. The module inputs the fused image into the second-level network of the hierarchical deep convolutional neural network model. Based on the shared features extracted by the first-level network, the second-level network performs feature fusion through multiple fully connected layers and outputs the fault segment identification result through a softmax function. Finally, the module inputs the fused image into the third-level network of the hierarchical deep convolutional neural network model. The third-level network shares the feature extraction modules of the first-level and second-level networks, and its output layer uses a linear activation function to directly output the fault distance regression estimation result between the fault point and the reference node.

[0017] In one feasible implementation, the building module is specifically used for: training the neural network model using an optimizer with an adaptive learning rate, and setting the corresponding initial learning rate, batch size, and number of iterations; for fault type and phase classification tasks and fault segment identification tasks, using the category cross-entropy loss function to calculate the difference between the predicted probability and the true label; for fault distance regression estimation tasks, using the mean squared error loss function to calculate the squared difference between the predicted distance and the true distance; dividing the two-dimensional image training set into a training set and a validation set, and after each training iteration cycle, calculating the model's performance index on the validation set, and selecting the model with the best performance on the validation set as the converged hierarchical deep convolutional neural network model based on the performance index.

[0018] A third aspect of the present invention provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to perform the above-described distribution network hierarchical fault diagnosis method.

[0019] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described distribution network hierarchical fault diagnosis method.

[0020] The technical solution provided by this invention involves acquiring real-time three-phase current signals of key nodes in a distribution network under the current fault scenario; extracting fault window signals containing fault information based on the real-time three-phase current signals; converting the fault window signals from one-dimensional time-domain signals into two-dimensional images using rasterization technology; and inputting the two-dimensional images into a pre-trained hierarchical deep convolutional neural network model to obtain the fault type and phase classification results, fault section identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model. The hierarchical deep convolutional neural network model is trained based on training samples generated in a dynamic simulation environment that considers the output fluctuations of distributed power sources and the uncertainty of load demand. The dynamic simulation environment uses a probability distribution model to quantify the uncertainty. In this embodiment of the invention, by employing rasterization technology to convert one-dimensional current signals into two-dimensional images, the deep convolutional neural network can extract richer and more comprehensive transient fault features. This effectively overcomes the problems of feature loss and poor adaptability caused by the reliance on manual feature extraction in traditional methods, significantly improving the accuracy of fault classification. Simultaneously, by introducing a probability distribution model to quantify the output fluctuations of distributed power sources and the uncertainty of load demand, and combining this with a dynamic simulation environment to generate massive training samples covering complex operating conditions, the model can maintain high diagnostic robustness in distributed power source access scenarios, effectively solving the problem of decreased diagnostic accuracy under the interference of uncertain factors in traditional methods. Furthermore, by using a hierarchical deep convolutional neural network model to achieve integrated collaborative diagnosis of fault type classification, fault segment identification, and fault distance regression, information fragmentation and error propagation between multiple tasks are reduced. While reducing dependence on the density of measurement equipment, high-precision fault location is achieved, thereby comprehensively improving the accuracy, robustness, and engineering practicality of distribution network fault diagnosis. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of one embodiment of the distribution network hierarchical fault diagnosis method in this invention;

[0023] Figure 2 This is a schematic diagram of another embodiment of the distribution network hierarchical fault diagnosis method in this invention;

[0024] Figure 3 This is a schematic diagram of one embodiment of the power distribution network hierarchical fault diagnosis device in this invention;

[0025] Figure 4 This is a schematic diagram of another embodiment of the power distribution network hierarchical fault diagnosis device in this invention;

[0026] Figure 5 This is a schematic diagram of one embodiment of the electronic device in this invention;

[0027] Figure 6 This is a comparison chart of the accuracy rates of embodiments of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will now be described with reference to the accompanying drawings. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0029] The terms “comprising,” “including,” or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0030] The terms “first,” “second,” etc., are used only to distinguish one entity or operation from another, and should not be construed as indicating or implying relative importance, nor as requiring or implying any such actual relationship or order between these entities or operations.

[0031] It is understood that the executing entity of this invention can be a distribution network hierarchical fault diagnosis device, a terminal, or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0032] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the distribution network hierarchical fault diagnosis method in this invention includes:

[0033] 101. Obtain the real-time three-phase current signal of key nodes in the distribution network under the current fault scenario;

[0034] When a fault occurs in the distribution network, phasor measurement units deployed at key nodes are triggered in real time to synchronously collect fault electrical quantity information. The selection of these key nodes is based on a comprehensive assessment of the distribution network topology, distributed generation connection locations, and load distribution characteristics. These typically include substation busbar outlets, distributed generation grid connection points, important feeder sectionalizing switches, and high-capacity load connection points, ensuring comprehensive coverage and observability of the fault current propagation path. The phasor measurement units continuously monitor the distribution network's operating status. When a sudden current change exceeds a preset trigger threshold or a protection device action signal is received, a fault is determined, and the data acquisition process is immediately initiated. During acquisition, the phasor measurement units synchronously sample the three-phase current at a pre-configured high sampling rate, generating high-fidelity raw current waveform data. The acquired real-time three-phase current signals completely record the transient process at the moment of fault occurrence, containing rich fault characteristic information.

[0035] 102. Extract fault window signals containing fault information based on real-time three-phase current signals;

[0036] After receiving the real-time three-phase current signals uploaded by each key node, the multi-source data is first time-aligned and synchronized to eliminate timing inconsistencies caused by communication delays or clock deviations. Subsequently, waveform feature analysis or abrupt change point detection algorithms are used to preprocess the current signals of each phase to pinpoint the start time of the fault. The abrupt change point detection algorithm identifies the transient start point of the fault based on features such as the instantaneous amplitude change rate of the current signal or the maximum value of the wavelet transform modulus, ensuring the accuracy of fault location. Using the pinpointed fault time as a reference, a pre-fault period signal containing the steady-state operating state before the fault is extracted, and a fault period signal containing the complete transient process after the fault is extracted, together forming the fault window signal. This fault window signal covers the complete electrical change process before and after the fault, retaining the normal operating conditions before the fault as a reference while fully recording the transient evolution information after the fault. To further improve signal quality, the extracted fault window signal can be filtered to remove high-frequency noise and power frequency harmonic interference while retaining effective information within the fault characteristic frequency band.

[0037] 103. The fault window signal is converted from a one-dimensional time-domain signal into a two-dimensional image using rasterization technology;

[0038] The fault window signal is normalized to map the amplitude of the three-phase current to a unified numerical range, eliminating the dimensional influence of current amplitude differences under different operating conditions while maintaining the relative shape of the signal waveform. After normalization, a mapping relationship between signal sample points and a two-dimensional pixel coordinate system is constructed: the time series index corresponds to the pixel coordinates in the width direction of the two-dimensional image, and the normalized current amplitude corresponds to the pixel coordinates in the height direction of the two-dimensional image, with each sample point mapped to a pixel on the image plane. To achieve a graphical representation of the continuous signal waveform, a linear rasterization algorithm is used to sequentially connect the pixel coordinates corresponding to adjacent sample points, constructing a continuous signal trajectory curve. This rasterization algorithm generates discretized lines on the pixel grid that are closest to the original signal waveform through point-by-point judgment and approximation, fully preserving the time-domain variation characteristics and amplitude details of the one-dimensional signal. For three-phase current signals, the three-phase waveforms can be mapped to the three color channels of an RGB image to generate a color two-dimensional image, allowing the waveform characteristics of different phases to be independently expressed in the image. The resulting standard-sized two-dimensional image retains both the temporal characteristics of the original signal and the structured properties of the image data, making it easy to input into a deep convolutional neural network for automatic feature extraction and learning.

[0039] 104. Input the two-dimensional image into the pre-trained hierarchical deep convolutional neural network model to obtain the fault type and phase classification results, fault section identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model. The hierarchical deep convolutional neural network model is trained based on training samples generated by a dynamic simulation environment that considers the output fluctuation of distributed power sources and the uncertainty of load demand. The dynamic simulation environment uses a probability distribution model to quantify uncertainty.

[0040] Two-dimensional images are input into a pre-trained hierarchical deep convolutional neural network model. The model extracts deep fault features from the image layer by layer through its multi-level network structure and outputs multi-dimensional diagnostic results. The hierarchical deep convolutional neural network model adopts a three-level progressive architecture design: The first-level network focuses on fault type and phase identification. By extracting features from the current image collected by a single node and combining the output of the fault category probability distribution by the softmax classification layer, it determines the specific type of the current fault, such as single-phase grounding, two-phase grounding, three-phase grounding, or no fault. The second-level network is responsible for fault segment localization. It receives multi-channel image input after channel fusion of current images collected by multiple key nodes. Based on the shared bottom-level features, it performs multi-point feature fusion through a multi-layer fully connected network and outputs the probability distribution of the fault segment, achieving effective definition of the fault location area. The third-level network undertakes the task of accurate fault distance estimation. It also takes the multi-node fused image as input and shares the feature extraction module of the previous network. Its output layer uses a linear activation function to directly regress and calculate the spatial distance between the fault point and the reference node. The entire diagnostic process requires no manual intervention. The model automatically completes the entire process from feature extraction to multi-task decision-making, and finally outputs a complete diagnostic result including fault type, fault segment, and fault distance.

[0041] In this embodiment of the invention, by employing rasterization technology to convert one-dimensional current signals into two-dimensional images, the deep convolutional neural network can extract richer and more comprehensive transient fault features. This effectively overcomes the problems of feature loss and poor adaptability caused by the reliance on manual feature extraction in traditional methods, significantly improving the accuracy of fault classification. Simultaneously, by introducing a probability distribution model to quantify the output fluctuations of distributed power sources and the uncertainty of load demand, and combining this with a dynamic simulation environment to generate massive training samples covering complex operating conditions, the model can maintain high diagnostic robustness in distributed power source access scenarios, effectively solving the problem of decreased diagnostic accuracy under the interference of uncertain factors in traditional methods. Furthermore, by using a hierarchical deep convolutional neural network model to achieve integrated collaborative diagnosis of fault type classification, fault segment identification, and fault distance regression, information fragmentation and error propagation between multiple tasks are reduced. While reducing dependence on the density of measurement equipment, high-precision fault location is achieved, thereby comprehensively improving the accuracy, robustness, and engineering practicality of distribution network fault diagnosis.

[0042] Please see Figure 2 Another embodiment of the distribution network hierarchical fault diagnosis method in this invention includes:

[0043] 201. Construct a hierarchical deep convolutional neural network model;

[0044] Determine the probability distribution model corresponding to the output fluctuations and load demand randomness of distributed generation in the distribution network;

[0045] Based on a probability distribution model and combined with a real-time digital simulator, a distribution network simulation environment capable of simulating the dynamic changes in distributed power output and load demand is constructed. Based on the distribution network simulation environment, simulated fault scenarios covering various preset fault types, fault locations, and fault parameters are generated, and the original three-phase current signals of key nodes in the distribution network under each simulated fault scenario are collected as training samples. The training samples are preprocessed and rasterized to generate corresponding two-dimensional image training sets. An initial neural network model is constructed, and the neural network model is trained using the two-dimensional image training set until the model converges, resulting in a hierarchical deep convolutional neural network model.

[0046] The neural network model is trained using a 2D image training set until convergence, resulting in a hierarchical deep convolutional neural network model. This process includes: training the neural network model using an optimizer with an adaptive learning rate, and setting the initial learning rate, batch size, and number of iterations; for fault type and phase classification tasks and fault segment identification tasks, using the category cross-entropy loss function to calculate the difference between the predicted probability and the true label; for fault distance regression estimation tasks, using the mean squared error loss function to calculate the squared difference between the predicted distance and the true distance; dividing the 2D image training set into a training set and a validation set, calculating the model's performance metrics on the validation set after each training iteration, and selecting the model with the best performance on the validation set as the converged hierarchical deep convolutional neural network model based on the performance metrics.

[0047] To address the random fluctuation characteristics of distributed power generation output in the distribution network, different types of probability distribution models are used for quantitative description: For wind power output, the shape and scale parameters of the Weibull probability density function are fitted based on historical wind speed observation data. This distribution simulates the random variation of wind speed within the cut-in to cut-out wind speed range, thus transforming it into the fluctuation characteristics of wind power output. The output fluctuation amplitude is generally controlled within ±10% of the rated power of 500kW. For photovoltaic (PV) power output, a Gaussian probability density function is used to quantify its uncertainty. The mean is set to the rated power of the PV power source, and the variance is calculated based on historical irradiance monitoring data to characterize the output dispersion caused by random irradiance fluctuations. The PV output fluctuation amplitude is also generally controlled within ±10% of the rated power. For load demand, a Gaussian probability density function is also used for quantification. The mean is taken as the rated power value of the load at each node, and the variance is fitted through historical load statistics of the distribution network to describe the random fluctuation characteristics of the load over time. The load fluctuation amplitude is generally set to ±15%.

[0048] The mathematical expression for the Weibull distribution is:

[0049] (1)

[0050] in, Wind speed; The shape parameter characterizes the steepness of the wind speed distribution; is a scale parameter, representing the characteristic value of wind speed.

[0051] The mathematical expression for the Gaussian distribution is:

[0052]

[0053] in, For photovoltaic power output or load demand; This is the average value, set as the rated power of the photovoltaic power source or the rated power value of the node load; The standard deviation is calculated based on historical irradiance or load statistics and is used to characterize the degree of fluctuation in output or load.

[0054] The aforementioned quantified uncertainties in wind power output, photovoltaic power output, and load demand are integrated into a dynamic simulation model of the distribution network built on a real-time digital simulator. This simulation model is based on a standard distribution network topology and incorporates multiple types of distributed power sources (for example, the simulation model is built on the IEEE 13-node standard distribution network with a rated voltage of 4.16kV, including one substation voltage regulator, multiple types of overhead and underground distribution lines, mixed loads, and two sets of parallel capacitor banks; simultaneously, a 300kW hydropower source is connected to bus 680, a 500kW wind power source is connected to bus 633, and a 300kW photovoltaic power source is connected to bus 645). Through the hardware-in-the-loop simulation capability of the real-time digital simulator, dynamic interaction and real-time updates of uncertainties are achieved, constructing a distribution network simulation environment that can realistically reflect the dynamic changes in distributed power output and load demand. Based on this dynamic simulation environment, simulated fault scenarios covering various preset fault types, fault locations, and fault parameters were designed, including typical fault types such as single-phase grounding faults, two-phase grounding faults, and three-phase grounding faults. Multiple random fault points distributed across each busbar node and key feeder locations were selected, and different values ​​for fault resistance and fault initiation angle were set (for example, 21 random fault points were selected, including all busbar nodes and key locations at 500ft intervals between busbars 650-632, 632-671, 671-680, and 684-652, with fault resistance set at 0.00). The fault initiation angle is randomly selected within the range of 1Ω to 10Ω and 0° to 360°. At the same time, the uncertainty quantification model of wind power, photovoltaic output and load demand is activated to realize the random combination of fault parameters and dynamic operating conditions, generating complex fault scenarios that closely resemble actual operation. The phasor measurement unit deployed at key nodes synchronously collects the original three-phase current signals as training samples according to the preset sampling parameters. The collected training samples are preprocessed, including extracting the fault window signal containing the pre-fault period and the fault period, and normalizing the current signal amplitude to eliminate the influence of dimensions.

[0055] The mathematical expression for the normalization method is:

[0056]

[0057] in, This represents the amplitude of the original current signal. This is the normalized amplitude of the current signal; and These represent the minimum and maximum values ​​of the current signal within the fault window, respectively.

[0058] The preprocessed one-dimensional time-domain signal is transformed into a standard-sized two-dimensional image using rasterization technology, such as a standard-sized two-dimensional RGB image of 224×224×3. Specifically, the normalized signal sample points are mapped to a two-dimensional pixel coordinate system using the Bressenham line algorithm to construct a continuous signal trajectory. Different color channels are assigned to the current signals of different phases to generate a two-dimensional image training set adapted to the input of the deep learning model (for example, three types of datasets are finally constructed: the fault classification dataset uses fault period signal images of 650 PMUs at the substation bus, totaling 5600 images (700 images / class × 8 classes); the fault section identification dataset uses fused images of pre-fault + fault period signals of 4 PMUs, generating 800 samples for each section, totaling 6400 images; and the fault distance estimation dataset generates 800 samples for each fault location, totaling 16800 images). Finally, a three-level progressive hierarchical deep convolutional neural network model architecture is constructed, corresponding to the three core tasks of fault type and phase classification, fault segment identification, and fault distance regression estimation. Each level of the network shares a basic convolutional module while optimizing the output layer design for different tasks (for example, the first-level network CNN1 contains 14 layers, where the core feature extraction module consists of 6 alternating 2D convolutional layers and 5 max-pooling layers; the convolutional layers use 3×3 kernels, and the number of filters is 8, 16, 32, 64, 128, and 256 respectively; each convolutional layer is followed by a batch normalization layer and a ReLU activation function; the second-level network CNN2 contains 19 convolutional layers and 6 fully connected layers; after flattening the feature maps, 6 additional cascaded fully connected layers are added, with output dimensions of 512, 256, 128, 64, 32, and 16 respectively; the third...). The CNN3 deep convolutional neural network shares the same network structure as CNN2 (with only one neuron in the output layer using a linear activation function). It trains the model using a generated 2D image training set. During training, an adaptive learning rate optimizer is used, with the initial learning rate, batch size, and number of iterations set. For classification tasks, the class cross-entropy loss function is used to calculate the difference between the predicted probability and the true label; for regression tasks, the mean squared error loss function is used to calculate the squared difference between the predicted distance and the true distance. The 2D image training set is divided into a training set and a validation set (e.g., randomly divided into training and test sets in a 9:1 ratio, with 10% reserved in the training set for validation). After each training iteration, the model's performance metrics on the validation set are calculated, and the model with the best performance on the validation set is selected as the converged hierarchical deep convolutional neural network model based on these metrics.

[0059] For example, in order to generate training samples that are comprehensive and closely resemble actual operating conditions, a simulation scenario that fully covers actual fault conditions is designed in the IEEE 13-node distribution network model based on the constructed dynamic simulation environment. The specific settings are as follows:

[0060] Fault types: Includes 7 typical faults and 1 fault-free scenario. The fault types cover single-phase ground faults (AG, BG, CG), two-phase ground faults (ABG, BCG, CAG), and three-phase ground faults (ABCG), covering common fault forms in distribution networks.

[0061] Fault location: 21 random fault points were selected, including all bus nodes, key locations at 500ft intervals between bus 650-632, 632-671, 671-680 and 684-652, to ensure full coverage of the fault section.

[0062] Fault parameter uncertainty: The fault resistance is selected as a random value in the range of 0.001Ω to 10Ω using a uniform probability density function (PDF). This distribution can cover the entire scenario from minor faults (high resistance) to severe faults (low resistance). 0.001Ω corresponds to metallic short circuit faults, and 10Ω corresponds to complex faults such as high-resistance grounding. The fault initiation angle is randomly selected in the range of 0° to 360° to simulate the randomness of the fault occurring at any time within the power frequency cycle.

[0063] Dynamic operating condition coupling: During the simulation process, the uncertainty quantification model of wind power, photovoltaic power output and load demand is activated simultaneously to realize the random combination of fault parameters and dynamic operating conditions, and generate complex fault scenarios that are close to reality.

[0064] Using the batch simulation function of the RTDS real-time digital simulator, all the above-mentioned fault scenarios and dynamic operating conditions are triggered. The deployed PMU synchronously acquires the original three-phase current signals according to the preset acquisition parameters (10kHz sampling frequency, 167 samples / cycle, two-cycle acquisition duration). For the acquired signals of each scenario, preprocessing, rasterization conversion and size standardization operations are performed in sequence to generate corresponding two-dimensional image samples, and label information is added to each sample: fault type label (8 types), fault section label (8 sections), and fault distance label (distance between the actual fault point and the substation, unit: km).

[0065] Finally, three types of datasets were constructed, each adapted for training a three-level CNN model:

[0066] Fault classification dataset: Only fault cycle signal images of 650 PMUs at the substation bus are used, totaling 5600 images (700 images / category × 8 categories), covering all fault types and dynamic operating condition combinations.

[0067] Fault segment identification dataset: The dataset uses pre-fault + fault cycle signal fusion images from 4 PMUs, generating 800 samples per segment (200 fault cases × 4 PMU signal fusion), for a total of 6400 images (800 images / segment × 8 segments).

[0068] Fault distance estimation dataset: Consistent with the acquisition logic of the segment identification dataset, 800 samples are generated for each fault location, for a total of 16,800 samples (800 samples / location × 21 locations), ensuring sample diversity for the regression task.

[0069] The three datasets were randomly divided into training and test sets in a 9:1 ratio, with 10% reserved in the training set as a validation set to monitor overfitting during model training. The training set was used for model parameter learning, the validation set for hyperparameter tuning and model selection, and the test set for evaluating model generalization performance. This division ratio ensures sufficient training data while effectively validating the model's performance in unseen scenarios. All image samples had already undergone normalization and size standardization (224×224×3 or 224×224×12) in the second step and require no additional preprocessing; they could be directly input into the CNN model.

[0070] A three-level CNN model was built using Python 3.9.0, employing the TensorFlow framework and Keras library. The training hardware consisted of an Apple M1 chip (8-core CPU, 7-core GPU, 8GB RAM) to balance training efficiency and computational power requirements. The training parameters were uniformly set as follows:

[0071] Optimizer: The Adam optimizer is used with a learning rate of 0.001. This optimizer can converge quickly and avoid local optima through an adaptive learning rate adjustment strategy. The learning rate of 0.001 balances convergence speed and parameter update stability.

[0072] Loss function: Classification tasks (CNN1, CNN2) use the class cross-entropy loss function to measure the difference between the predicted class probability and the true label; regression tasks (CNN3) use the mean squared error (MSE) loss function to minimize the squared difference between the predicted distance and the true distance.

[0073] Training batch size and number of iterations: Set the batch size to 10, which strikes a balance between memory usage and gradient estimation accuracy; set the number of iterations to 100 to ensure that the model fully learns the data features.

[0074] Training monitoring and model saving: After each training iteration, the performance metrics of the model on the validation set are calculated (accuracy for classification tasks and MAE for regression tasks). If the performance is better than the historical best value, the current model parameters are saved. Finally, the model with the best performance on the validation set is selected as the final training model to avoid the decline in generalization ability caused by overfitting.

[0075] 202. Obtain real-time three-phase current signals of key nodes in the distribution network under the current fault scenario;

[0076] Phasor measurement units are deployed at key nodes in the distribution network. These key nodes include at least one of the following: bus outlet, distributed generation grid connection point, feeder sectionalizing switch, and load connection point. The sampling frequency and acquisition duration of the phasor measurement units are configured. The sampling frequency is set to a sampling rate capable of capturing high-frequency details of fault transient signals, and the acquisition duration covers at least the continuous time period before and after the fault occurs. The operating status of the distribution network is monitored in real time. When a fault is detected, the phasor measurement units are triggered to synchronously acquire three-phase current signals according to the configured sampling parameters, which are then used as real-time three-phase current signals.

[0077] Phasor measurement units (PMUs) are deployed at key nodes in the distribution network. The selection of these key nodes comprehensively considers the distribution network topology, distributed generation access locations, load distribution characteristics, and fault propagation paths. Specifically, these nodes include at least one of the following: substation bus outlets, distributed generation grid connection points, feeder sectionalizing switches, and important load access points (e.g., deploying PMUs at key locations such as bus 650, 633, 645, and 680). This ensures comprehensive capture of fault current signals and redundancy in the observability of the distribution network. For each deployed PMU, the sampling frequency and acquisition duration parameters are uniformly configured. The sampling frequency is set to a sampling rate capable of capturing high-frequency details of fault transient signals (e.g., a sampling frequency of 10kHz, i.e., acquiring 10,000 data points per second, with 167 sampling points per cycle). This ensures complete recording of instantaneous current surges and high-frequency transient components during faults. The acquisition duration covers at least the continuous time period before and after the fault, including a complete pre-fault cycle to provide a normal operation baseline, and a complete fault cycle to record the fault transient evolution process. Under normal operating conditions, each phasor measurement unit continuously monitors the distribution network's operating status, calculating the instantaneous values ​​and rates of change of the three-phase current in real time. When a sudden current change exceeds a preset trigger threshold or a fault trigger signal is received from a protection device, a fault is determined to have occurred in the current scenario. The data acquisition process is immediately initiated, triggering all deployed phasor measurement units to perform high-precision synchronous acquisition according to uniformly configured sampling parameters, obtaining multi-node three-phase current signals under the same time reference. After acquisition, each phasor measurement unit adds a precise timestamp to the acquired current data and uploads the real-time three-phase current signal to the distribution network master station or edge computing node via the distribution network communication network.

[0078] 203. Extract fault window signals containing fault information based on real-time three-phase current signals;

[0079] Based on the waveform characteristics or abrupt change points of the real-time three-phase current signal, the fault occurrence time is calibrated; using the fault time as a reference, a pre-fault period signal of a preset length is extracted forward, and a fault period signal of a preset length is extracted backward to form a continuous data window containing fault information; the three-phase current signal within the continuous data window is filtered or noise-reduced to obtain the fault window signal.

[0080] After receiving the real-time three-phase current signals uploaded by each key node, the data of each node is first time-synchronized to ensure that the multi-source data is based on a unified time reference, eliminating timing inconsistencies caused by communication delays or clock deviations. Subsequently, waveform feature analysis is performed on the current signals of each phase, and a sudden change point detection algorithm is used to accurately pinpoint the onset time of the fault. This algorithm identifies the transient onset point of the fault based on indicators such as the instantaneous amplitude change rate of the current signal, the maximum value of the wavelet transform modulus, or the zero-crossing characteristic of the signal derivative, ensuring the accuracy of fault location and noise immunity. Using the calibrated fault time as a reference, a pre-fault period signal of a preset length is extracted forward. This pre-fault period signal reflects the normal operating state of the distribution network before the fault occurs, providing a reference for the current waveform characteristics under steady-state conditions. Simultaneously, a fault period signal of a preset length is extracted backward. This fault period signal completely records the transient evolution process after the fault occurs, including key fault information such as the amplitude change, phase shift, and attenuation characteristics of the fault current (for example, one power frequency cycle is taken for both the pre-fault period and the fault period, i.e., 167 sampling points each). The pre-fault period signal and the fault period signal together constitute a continuous fault window signal, covering the complete electrical change process before and after the fault occurs. To improve the purity and signal-to-noise ratio of the fault window signal, the three-phase current signals within the extracted continuous data window are filtered or denoised. A digital bandpass filter is used to retain the effective signal components within the fault characteristic frequency band, while filtering out high-frequency noise interference and power frequency harmonic components, ultimately obtaining a high-quality, high-fidelity fault window signal.

[0081] 204. The fault window signal is converted from a one-dimensional time-domain signal into a two-dimensional image using rasterization technology;

[0082] The fault window signal is normalized, and the normalized signal sample points are mapped to a two-dimensional pixel coordinate system. The time series index of the signal corresponds to the pixel coordinates in the width direction of the two-dimensional image, and the normalized amplitude of the signal corresponds to the pixel coordinates in the height direction of the two-dimensional image. Each sample point corresponds to a pixel coordinate point. The Bressenham line algorithm is used to connect the pixel coordinate points corresponding to adjacent sample points in sequence to construct a continuous signal trajectory. The pixels corresponding to the continuous signal trajectory are assigned preset grayscale intensity or channel value to generate a two-dimensional image.

[0083] The fault window signal is normalized to linearly map the amplitude of the three-phase current to a unified numerical range. Specifically, the minimum and maximum values ​​of the current signal within the fault window are calculated, and the original amplitude of each sampling point is subtracted from the minimum value and then divided by the range to obtain the normalized dimensionless value. This process eliminates the dimensional influence caused by the difference in current amplitude under different fault scenarios, while fully preserving the relative shape and timing characteristics of the signal waveform. After normalization, a mapping relationship between signal sample points and a two-dimensional pixel coordinate system is constructed. The time series index of the signal corresponds to the pixel coordinates in the width direction of the two-dimensional image, and the normalized current amplitude corresponds to the pixel coordinates in the height direction of the two-dimensional image. Each sample point is mapped to a pixel coordinate point on the image plane, realizing the coordinate transformation from one-dimensional time space to two-dimensional pixel space. To achieve a graphical representation of continuous signal waveforms, the Bressenham line algorithm is employed to sequentially connect the pixel coordinates of adjacent sample points. This algorithm generates a discretized straight line trajectory on the pixel grid that most closely approximates the original signal waveform through point-by-point judgment and error accumulation calculation. This effectively avoids waveform breakage and distortion during the conversion process, fully preserving the time-domain variation characteristics and amplitude details of the one-dimensional signal. For three-phase current signals, the signal trajectories of phases A, B, and C can be mapped to the red, green, and blue channels of an RGB image, respectively, generating a color two-dimensional image. This allows the waveform features of different phases to be independently expressed and distinguished in the image, facilitating the subsequent extraction of independent features and inter-phase correlation features of each phase current from the multi-channel image by the convolutional neural network. The final generated two-dimensional image undergoes size standardization, adjusting it to a uniform size suitable for the input layer of the deep learning model (e.g., a fixed size of 224×224×3) to ensure the consistency of the image data format.

[0084] 205. Input the two-dimensional image into the hierarchical deep convolutional neural network model to obtain the fault type and phase classification results, fault segment identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model.

[0085] An optimizer with an adaptive learning rate is used to train the neural network model, and the corresponding initial learning rate, batch size, and number of iterations are set. For fault type and phase classification tasks and fault segment identification tasks, the category cross-entropy loss function is used to calculate the difference between the predicted probability and the true label. For fault distance regression estimation tasks, the mean squared error loss function is used to calculate the squared difference between the predicted distance and the true distance. The two-dimensional image training set is divided into a training set and a validation set. After each training iteration cycle, the performance index of the model on the validation set is calculated, and the model with the best performance on the validation set is selected as the converged hierarchical deep convolutional neural network model based on the performance index.

[0086] A standard-sized two-dimensional image generated through rasterization is input into a pre-trained hierarchical deep convolutional neural network model. The model extracts deep fault features from the image layer by layer through its three-level progressive network architecture and outputs multi-dimensional diagnostic results. First, the two-dimensional image corresponding to the real-time three-phase current signal collected by a single key node is input into the first-level network. This level network automatically learns and extracts shallow edge features and deep semantic features from the input image through the alternating stacking of multiple convolutional layers and pooling layers. The convolutional layers use small-sized convolutional kernels to focus on local waveform features, and the pooling layers compress the feature map size and retain key information through downsampling operations. The number of filters is increased layer by layer to achieve progressive extraction from simple time-domain features to complex fault features. Each convolutional layer is followed by a batch normalization layer and a nonlinear activation function to accelerate convergence and enhance the model's expressive power. Finally, the extracted feature map is flattened and input into a fully connected layer for feature integration. The softmax function outputs the probability distribution of each category, and the category with the highest probability is selected as the fault type and phase classification result. The fault types include single-phase grounding, two-phase grounding, three-phase grounding, or no-fault scenarios.

[0087] Specifically, the two-dimensional image corresponding to the real-time three-phase current signal collected from a single key node (650 on the substation busbar) is input into the first-level network CNN1, which is specifically used for fault type and phase classification. The input image size is 224×224×3. Before input, the same preprocessing method as the VGG16 model is used: the RGB mean of all images in the training set is calculated, and the mean of the corresponding channel is subtracted from the pixel value of each input image to eliminate feature interference caused by illumination differences, while maintaining the integrity of the original signal features.

[0088] The first-level network CNN1 consists of 14 layers. The core feature extraction module is composed of alternating 6 2D convolutional layers and 5 max-pooling layers, while the decision output module consists of 1 fully connected layer and 1 output layer. The specific design is as follows:

[0089] The convolutional layers use 3×3 kernels, a size that reduces parameter computation while focusing on local fault features. The number of filters in each convolutional layer is 8, 16, 32, 64, 128, and 256 respectively, progressively increasing the number of filters to extract features from shallow temporal domains to deep fault domains. Each convolutional layer is followed by a batch normalization (BN) layer and a ReLU activation function. The BN layer normalizes the input data within a batch, addressing the gradient explosion problem, while the ReLU activation function... Nonlinearity is introduced to enhance the model's ability to fit complex fault features; padding is used during convolution to ensure that the output feature map size is consistent with the input and to avoid loss of edge features.

[0090] The pooling layer uses a 2×2 max pooling kernel to downsample the feature map output by the convolutional layer, preserving key features while compressing the data dimension and reducing computational complexity.

[0091] The fully connected layer flattens the 3×3×256 feature map output from the previous pooling layer into a 2304-dimensional vector, achieving global feature integration. The output layer has 8 neurons, corresponding to 7 fault types (single-phase grounding AG / BG / CG, two-phase grounding ABG / BCG / CAG, three-phase grounding ABCG) and 1 fault-free scenario. The softmax function is used to generate probability values ​​for each type. The mathematical expression of the softmax function is:

[0092]

[0093] in, For the first The probability of a class The output value of the fully connected layer is used to select the category with the highest probability as the fault classification result. x j represents the output value of the fully connected layer, j represents the category index of the summation traversal, used to sum the exponent terms of all categories (0~8), j∈[0,8] corresponds to 7 types of faults (single-phase ground AG / BG / CG, two-phase ground ABG / BCG / CAG, three-phase ground ABCG) + 1 type of no fault + 1 type of normal operation, for a total of 9 classification scenarios.

[0094] Secondly, the two-dimensional images corresponding to the real-time three-phase current signals collected from multiple key nodes are stitched and fused along the channel dimension to generate a multi-channel fused image. This fused image integrates fault information from different spatial locations, which can more comprehensively reflect the propagation characteristics and spatial distribution of the fault in the distribution network. The fused image is then input into the second-level network, which shares a feature extraction module similar to the first-level network, but adds multiple series fully connected layers after the feature map is flattened. Through the deep fully connected network, the fine fusion and interaction of features from multiple measurement points are achieved. Finally, the fault segment identification result is output through the softmax function to determine the specific segment range where the fault occurred.

[0095] Specifically, the two-dimensional images corresponding to the real-time three-phase current signals collected from multiple key nodes are input into the second-level network CNN2, which is specifically used for fault section identification. The input data for the second-level network consists of two-dimensional images corresponding to the signals collected by phasor measurement units deployed at four key locations (buses 650, 633, 645, and 680). Each phasor measurement unit provides a converted image of the three-phase current signal during the pre-fault period and the fault period, with an image size of 224×224×3. The images from the four phasor measurement units under the same fault scenario are stitched together along the channel axis to form a fused image of 224×224×12, achieving complementarity of fault information from multiple measurement points and improving the robustness of section identification. The input preprocessing method is consistent with that of the first-level network CNN1, calculating the RGB mean of all images in the training set and performing a mean subtraction operation to ensure data distribution consistency.

[0096] The second-level network, CNN2, consists of 19 convolutional layers and 6 fully connected layers. The structure of the core feature extraction module (kernel size, number of filters, BN layers, ReLU activation, and max-pooling layer settings) is identical to that of the first-level network, CNN1. The main difference lies in the decision output module. After the feature map undergoes alternating processing by convolutional and pooling layers, the 3×3×256 feature map output from the previous pooling layer is flattened into a 2304-dimensional vector. After flattening, six cascaded fully connected layers are added, all using the ReLU activation function, with output dimensions of 512, 256, 128, 64, 32, and 16 respectively. Deep fully connected layers achieve fine fusion of multi-point features, adapting to the complex classification requirements of fault segment identification. The output layer has 8 neurons, corresponding to the 8 fault segments in the power distribution network. A softmax function is used to output the probability value of each segment; the segment with the highest probability is the identification result.

[0097] Finally, the same multi-channel fused image is input into the third-level network. This network shares the feature extraction modules of the first two levels to maintain consistency in feature extraction, but its output layer is specifically designed for regression tasks. It uses a linear activation function instead of the softmax function to directly output the predicted spatial distance between the fault point and the reference node without probability transformation. During model training, the third-level network uses the actual fault distance as a label and optimizes parameters by minimizing the mean square error between the predicted and true values ​​to ensure the accuracy of distance estimation. The entire diagnostic process requires no manual intervention. The hierarchical deep convolutional neural network model automatically completes the entire process from image input to multi-task output, ultimately outputting a complete diagnostic result including fault type and phase, fault segment, and fault distance.

[0098] Specifically, the input data of the third-level network CNN3 is exactly the same as that of CNN2, namely 224×224×12 multi-PMU image fusion data. The preprocessing method is kept consistent to ensure the continuity of the previous feature extraction logic. Since fault distance estimation is a regression task rather than a classification task, the model only optimizes the output layer. The rest of the network structure (convolutional layer, pooling layer, fully connected layer) is exactly the same as that of CNN2, reducing the complexity of model design.

[0099] The output layer has only one neuron, using a linear activation function instead of the softmax function to directly output the predicted distance (in km) between the fault point and the substation, without requiring probability transformation. During model training, the actual fault distance is used as the label, and parameter optimization is achieved by minimizing the error between the predicted and true values. Error evaluation metrics include mean absolute error (MAE), root mean square error (RMSE), and percentage error, with the following mathematical expressions:

[0100]

[0101] (2) Among them, among them For the sample size, For the first The true fault distance of each sample To predict the distance to the fault, L is the total length of the branch where the fault is located. Multiple index constraints are used to ensure the accuracy of the distance estimation.

[0102] (3) N a L represents the total number of samples involved in the calculation. n表示 The true fault distance for the nth sample; Error is the percentage error.

[0103] In this embodiment of the invention, a dynamic simulation environment considering the uncertainties of distributed power sources and loads is constructed to generate a massive amount of training samples covering complex operating conditions. Combined with rasterization technology, one-dimensional current signals are transformed into two-dimensional images, enabling a hierarchical deep convolutional neural network to automatically extract complete transient fault features. This effectively overcomes the problems of feature loss and poor adaptability caused by traditional methods relying on manual feature extraction, significantly improving the robustness of fault diagnosis in distributed power source access scenarios. Simultaneously, a three-level progressive network architecture is used to achieve integrated collaborative diagnosis of fault type classification, fault segment identification, and fault distance regression, reducing information fragmentation and error propagation between multiple tasks. This reduces dependence on the density of measurement equipment while achieving high-precision fault location, comprehensively improving the accuracy, robustness, and engineering practicality of distribution network fault diagnosis.

[0104] Implementation Case:

[0105] This embodiment uses an IEEE 13-node active distribution network with distributed generation as the test object to verify the performance and stability of the proposed hierarchical fault diagnosis method. Five different training-test set partition ratios (90-10, 80-20, 70-30, 60-40, 50-50) were set up to conduct fault classification and fault segment identification tests. The accuracy comparison results are as follows: Figure 6 As shown.

[0106] As shown in the figure, the fault classification accuracy remained between 98.2% and 99.4% under all division ratios, and the fault segment identification accuracy remained between 98.4% and 99.3%. Even under the most stringent 50-50 division ratio, the accuracy of the two tasks still reached 98.3% and 98.5% respectively, without significant decline. Overall, the accuracy fluctuation range under different division ratios was less than 1.2%, which fully verified the strong robustness and stability of this method when the data sample allocation changes, and it can adapt to the scenario of limited sample size and uneven data distribution in actual engineering.

[0107] The distribution network hierarchical fault diagnosis method in the embodiments of the present invention has been described above. The distribution network hierarchical fault diagnosis device in the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 3 One embodiment of the distribution network hierarchical fault diagnosis device of the present invention includes:

[0108] The acquisition module 301 is used to acquire the real-time three-phase current signal of the key node of the distribution network under the current fault scenario.

[0109] Extraction module 302 is used to extract fault window signals containing fault information based on real-time three-phase current signals;

[0110] Conversion module 303 is used to convert the fault window signal from a one-dimensional time domain signal into a two-dimensional image using rasterization technology;

[0111] The output module 304 is used to input a two-dimensional image into a pre-trained hierarchical deep convolutional neural network model to obtain the fault type and phase classification results, fault segment identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model. The hierarchical deep convolutional neural network model is trained based on training samples generated by a dynamic simulation environment that considers the output fluctuation of distributed power sources and the uncertainty of load demand. The dynamic simulation environment uses a probability distribution model to quantify uncertainty.

[0112] In this embodiment of the invention, by employing rasterization technology to convert one-dimensional current signals into two-dimensional images, the deep convolutional neural network can extract richer and more comprehensive transient fault features. This effectively overcomes the problems of feature loss and poor adaptability caused by the reliance on manual feature extraction in traditional methods, significantly improving the accuracy of fault classification. Simultaneously, by introducing a probability distribution model to quantify the output fluctuations of distributed power sources and the uncertainty of load demand, and combining this with a dynamic simulation environment to generate massive training samples covering complex operating conditions, the model can maintain high diagnostic robustness in distributed power source access scenarios, effectively solving the problem of decreased diagnostic accuracy under the interference of uncertain factors in traditional methods. Furthermore, by using a hierarchical deep convolutional neural network model to achieve integrated collaborative diagnosis of fault type classification, fault segment identification, and fault distance regression, information fragmentation and error propagation between multiple tasks are reduced. While reducing dependence on the density of measurement equipment, high-precision fault location is achieved, thereby comprehensively improving the accuracy, robustness, and engineering practicality of distribution network fault diagnosis.

[0113] Please see Figure 4 Another embodiment of the distribution network hierarchical fault diagnosis device in this invention includes:

[0114] The acquisition module 301 is used to acquire the real-time three-phase current signal of the key node of the distribution network under the current fault scenario.

[0115] Extraction module 302 is used to extract fault window signals containing fault information based on real-time three-phase current signals;

[0116] Conversion module 303 is used to convert the fault window signal from a one-dimensional time domain signal into a two-dimensional image using rasterization technology;

[0117] The output module 304 is used to input a two-dimensional image into a pre-trained hierarchical deep convolutional neural network model to obtain the fault type and phase classification results, fault segment identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model. The hierarchical deep convolutional neural network model is trained based on training samples generated by a dynamic simulation environment that considers the output fluctuation of distributed power sources and the uncertainty of load demand. The dynamic simulation environment uses a probability distribution model to quantify uncertainty.

[0118] Optionally, hierarchical fault diagnosis of the distribution network includes:

[0119] Module 305 is used to determine the probability distribution model corresponding to the output fluctuations and load demand randomness of distributed generation in the distribution network. Based on the probability distribution model, combined with a real-time digital simulator, a distribution network simulation environment capable of simulating the dynamic changes in distributed generation output and load demand is constructed. Based on the distribution network simulation environment, simulation fault scenarios covering multiple preset fault types, fault locations, and fault parameters are generated, and the original three-phase current signals of key nodes in the distribution network under each simulation fault scenario are collected as training samples. The training samples are preprocessed and rasterized to generate corresponding two-dimensional image training sets. An initial neural network model is constructed, and the neural network model is trained using the two-dimensional image training set until the model converges, resulting in a hierarchical deep convolutional neural network model.

[0120] Optionally, the acquisition module 301 can be specifically used for:

[0121] Phasor measurement units are deployed at key nodes in the distribution network. These key nodes include at least one of the following: bus outlet, distributed power generation grid connection point, feeder sectionalizing switch, and load connection point. The sampling frequency and acquisition duration of the phasor measurement units are configured. The sampling frequency is set to a sampling rate capable of capturing high-frequency details of fault transient signals, and the acquisition duration covers at least the continuous time period before and after the fault occurs. The operating status of the distribution network is monitored in real time. When a fault is detected, the phasor measurement units are triggered to synchronously acquire three-phase current signals according to the configured sampling parameters, which are then used as real-time three-phase current signals.

[0122] Optionally, the extraction module 302 can be specifically used for:

[0123] Based on the waveform characteristics or abrupt change points of the real-time three-phase current signal, the fault occurrence time is calibrated; using the fault time as a reference, a pre-fault period signal of a preset length is extracted forward, and a fault period signal of a preset length is extracted backward to form a continuous data window containing fault information; the three-phase current signal within the continuous data window is filtered or noise-reduced to obtain the fault window signal.

[0124] Optionally, the conversion module 303 can be specifically used for:

[0125] The fault window signal is normalized, and the normalized signal sample points are mapped to a two-dimensional pixel coordinate system. The time series index of the signal corresponds to the pixel coordinates in the width direction of the two-dimensional image, and the normalized amplitude of the signal corresponds to the pixel coordinates in the height direction of the two-dimensional image. Each sample point corresponds to a pixel coordinate point. The Bressenham line algorithm is used to connect the pixel coordinate points corresponding to adjacent sample points in sequence to construct a continuous signal trajectory. The pixels corresponding to the continuous signal trajectory are assigned preset grayscale intensity or channel value to generate a two-dimensional image.

[0126] Optionally, output module 304 can be specifically used for:

[0127] Two-dimensional images corresponding to real-time three-phase current signals collected from a single node in a key node of the distribution network are input into the first-level network of a hierarchical deep convolutional neural network model. The first-level network extracts image features through convolutional layers, pooling layers, and fully connected layers, and outputs the fault type and phase classification results through a softmax function. The fault type and phase classification include single-phase grounding, two-phase grounding, three-phase grounding, or no-fault scenarios. Two-dimensional images corresponding to real-time three-phase current signals collected from multiple nodes in the key node of the distribution network are fused along the channel dimension to generate a fused image. The fused image is input into the second-level network of the hierarchical deep convolutional neural network model. The second-level network, based on the shared features extracted by the first-level network, performs feature fusion through multiple fully connected layers and outputs the fault section identification results through a softmax function. The fused image is input into the third-level network of the hierarchical deep convolutional neural network model. The third-level network shares the feature extraction modules of the first and second-level networks, and its output layer uses a linear activation function to directly output the fault distance regression estimation results between the fault point and the reference node.

[0128] Optionally, building module 305 can be specifically used for:

[0129] An optimizer with an adaptive learning rate is used to train the neural network model, and the corresponding initial learning rate, batch size, and number of iterations are set. For fault type and phase classification tasks and fault segment identification tasks, the category cross-entropy loss function is used to calculate the difference between the predicted probability and the true label. For fault distance regression estimation tasks, the mean squared error loss function is used to calculate the squared difference between the predicted distance and the true distance. The two-dimensional image training set is divided into a training set and a validation set. After each training iteration cycle, the performance index of the model on the validation set is calculated, and the model with the best performance on the validation set is selected as the converged hierarchical deep convolutional neural network model based on the performance index.

[0130] In this embodiment of the invention, a dynamic simulation environment considering the uncertainties of distributed power sources and loads is constructed to generate a massive amount of training samples covering complex operating conditions. Combined with rasterization technology, one-dimensional current signals are transformed into two-dimensional images, enabling a hierarchical deep convolutional neural network to automatically extract complete transient fault features. This effectively overcomes the problems of feature loss and poor adaptability caused by traditional methods relying on manual feature extraction, significantly improving the robustness of fault diagnosis in distributed power source access scenarios. Simultaneously, a three-level progressive network architecture is used to achieve integrated collaborative diagnosis of fault type classification, fault segment identification, and fault distance regression, reducing information fragmentation and error propagation between multiple tasks. This reduces dependence on the density of measurement equipment while achieving high-precision fault location, comprehensively improving the accuracy, robustness, and engineering practicality of distribution network fault diagnosis.

[0131] above Figure 3 and Figure 4 The distribution network hierarchical fault diagnosis device in this embodiment of the invention is described in detail from the perspective of modular functional entities. The electronic equipment in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0132] See Figure 5 As shown, the electronic device includes a processor 500 and a memory 501. The memory 501 stores machine-executable instructions that can be executed by the processor 500. The processor 500 executes the machine-executable instructions to implement the above-mentioned distribution network hierarchical fault diagnosis method.

[0133] Furthermore, Figure 5 The electronic device shown also includes a bus 502 and a communication interface 503. The processor 500, the communication interface 503 and the memory 501 are connected via the bus 502.

[0134] The memory 501 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 503 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 502 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0135] The processor 500 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 500 or by instructions in software form. The processor 500 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 501. The processor 500 reads the information in memory 501 and, in conjunction with its hardware, completes the method steps of the aforementioned embodiment.

[0136] The present invention also provides an electronic device, the computer device including a memory and a processor, the memory storing computer-readable instructions, which, when executed by the processor, cause the processor to perform the steps of the power distribution network hierarchical fault diagnosis method in the above embodiments.

[0137] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the distribution network hierarchical fault diagnosis method.

[0138] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0139] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0140] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hierarchical fault diagnosis method for a power distribution network, characterized in that, The hierarchical fault diagnosis method for the power distribution network includes: Acquire real-time three-phase current signals of key nodes in the distribution network under the current fault scenario; Based on the real-time three-phase current signal, extract the fault window signal containing fault information; The fault window signal is converted from a one-dimensional time-domain signal into a two-dimensional image using rasterization technology. The two-dimensional image is input into a pre-trained hierarchical deep convolutional neural network model to obtain the fault type and phase classification results, fault section identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model. The hierarchical deep convolutional neural network model is trained based on training samples generated in a dynamic simulation environment that considers the output fluctuation of distributed power sources and the uncertainty of load demand. The dynamic simulation environment uses a probability distribution model to quantify the uncertainty.

2. The distribution network hierarchical fault diagnosis method according to claim 1, characterized in that, Before acquiring the real-time three-phase current signals of key nodes in the distribution network under the current fault scenario, the following steps are also included: Determine the probability distribution model corresponding to the output fluctuations and load demand randomness of distributed generation in the distribution network; Based on the aforementioned probability distribution model, and combined with a real-time digital simulator, a power distribution network simulation environment capable of simulating the dynamic changes in distributed power output and load demand is constructed. Based on the power distribution network simulation environment, simulated fault scenarios covering multiple preset fault types, fault locations and fault parameters are generated, and the original three-phase current signals of key nodes in the power distribution network under each simulated fault scenario are collected as training samples. The training samples are preprocessed and rasterized to generate a corresponding two-dimensional image training set; An initial neural network model is constructed, and the neural network model is trained using the two-dimensional image training set until the model converges, resulting in a hierarchical deep convolutional neural network model.

3. The distribution network hierarchical fault diagnosis method according to claim 1, characterized in that, The acquisition of real-time three-phase current signals of key nodes in the distribution network under the current fault scenario includes: Phasor measurement units are deployed at key nodes of the distribution network. The key nodes include at least one of the following: bus outlet, distributed power grid connection point, feeder sectionalizing switch, and load connection point. Configure the sampling frequency and acquisition duration of the phasor measurement unit. The sampling frequency is set to a sampling rate that can capture high-frequency details of the fault transient signal, and the acquisition duration covers at least the continuous time period before and after the fault occurs. The system monitors the operation status of the power distribution network in real time. When a fault is detected, the phasor measurement unit is triggered to synchronously collect three-phase current signals according to the configured sampling parameters, which are then used as real-time three-phase current signals.

4. The distribution network hierarchical fault diagnosis method according to claim 1, characterized in that, The step of extracting a fault window signal containing fault information based on the real-time three-phase current signal includes: The fault occurrence time is determined based on the waveform characteristics or abrupt change points of the real-time three-phase current signal. Based on the fault time, a pre-fault cycle signal of a preset length is extracted forward, and a fault cycle signal of a preset length is extracted backward to form a continuous data window containing fault information. The three-phase current signal within the continuous data window is filtered or denoised to obtain the fault window signal.

5. The distribution network hierarchical fault diagnosis method according to claim 1, characterized in that, The step of using rasterization technology to convert the fault window signal from a one-dimensional time-domain signal into a two-dimensional image includes: The fault window signal is normalized, and the normalized signal sample points are mapped to a two-dimensional pixel coordinate system. The time series index of the signal corresponds to the pixel coordinates in the width direction of the two-dimensional image, and the normalized amplitude of the signal corresponds to the pixel coordinates in the height direction of the two-dimensional image. Each sample point corresponds to a pixel coordinate point. The Bressenham line algorithm is used to connect the pixel coordinates of adjacent sample points in sequence to construct a continuous signal trajectory. The pixels corresponding to the continuous signal trajectory are assigned preset grayscale intensity or channel value to generate a two-dimensional image.

6. The distribution network hierarchical fault diagnosis method according to claim 1, characterized in that, The step of inputting the two-dimensional image into a pre-trained hierarchical deep convolutional neural network model to obtain the fault type and phase classification results, fault segment identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model includes: The two-dimensional image corresponding to the real-time three-phase current signal collected by a single node in the key node of the distribution network is input into the first-level network of the hierarchical deep convolutional neural network model. The first-level network extracts image features through convolutional layers, pooling layers and fully connected layers, and outputs the fault type and phase classification results through the softmax function. The fault type and phase include single-phase grounding, two-phase grounding, three-phase grounding or no fault scenario. Two-dimensional images corresponding to real-time three-phase current signals collected from multiple nodes in key nodes of the distribution network are fused along the channel dimension to generate a fused image. The fused image is input into the second-level network of the hierarchical deep convolutional neural network model. Based on the shared features extracted by the first-level network, the second-level network performs feature fusion through multiple fully connected layers and outputs the fault segment identification result through the softmax function. The fused image is input into the third-level network of the hierarchical deep convolutional neural network model. The third-level network shares the feature extraction modules of the first-level and second-level networks. Its output layer uses a linear activation function to directly output the fault distance regression estimation result between the fault point and the reference node.

7. The distribution network hierarchical fault diagnosis method according to claim 2, characterized in that, The step of training the neural network model using the two-dimensional image training set until the model converges to obtain a hierarchical deep convolutional neural network model includes: The neural network model is trained using an optimizer with an adaptive learning rate, and the initial learning rate, batch size, and number of iterations are set accordingly. For fault type and phase classification tasks and fault segment identification tasks, the category cross-entropy loss function is used to calculate the difference between the predicted probability and the true label; for fault distance regression estimation tasks, the mean squared error loss function is used to calculate the squared difference between the predicted distance and the true distance. The two-dimensional image training set is divided into a training set and a validation set. After each training iteration, the performance index of the model on the validation set is calculated, and the model with the best performance on the validation set is selected as the converged hierarchical deep convolutional neural network model based on the performance index.

8. A distribution network hierarchical fault diagnosis device, characterized in that, The distribution network hierarchical fault diagnosis device includes: The acquisition module is used to acquire real-time three-phase current signals of key nodes in the distribution network under the current fault scenario; The extraction module is used to extract a fault window signal containing fault information based on the real-time three-phase current signal. The conversion module is used to convert the fault window signal from a one-dimensional time-domain signal into a two-dimensional image using rasterization technology; The output module is used to input the two-dimensional image into a pre-trained hierarchical deep convolutional neural network model to obtain the fault type and phase classification results, fault segment identification results, and fault distance regression estimation results output by the hierarchical deep convolutional neural network model. The hierarchical deep convolutional neural network model is trained based on training samples generated in a dynamic simulation environment that considers the output fluctuation of distributed power sources and the uncertainty of load demand. The dynamic simulation environment uses a probability distribution model to quantify the uncertainty.

9. An electronic device, characterized in that, The electronic device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the electronic device to execute the power distribution network hierarchical fault diagnosis method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements the distribution network hierarchical fault diagnosis method as described in any one of claims 1-7.

Images