Fault determination method and system for power distribution network containing distributed wind power plant, and medium
By decoupling and extracting features from transient traveling wave data, and using a neural network model to determine faults in the distribution network of distributed wind farms, the efficiency and accuracy problems of traditional diagnostic methods in fault diagnosis in distributed wind farms are solved, and efficient and accurate fault diagnosis is achieved.
Patent Information
- Application Number
- CN202511830812.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-06
- Publication Date
- 2026-03-03
AI Technical Summary
Traditional fault diagnosis methods for distribution networks are difficult to quickly and accurately identify faulty lines and fault points in distributed wind farms due to the large-scale grid connection of inverter-type resources, which weakens the electrical characteristics of faults and increases the complexity of fault response strategies.
By acquiring transient traveling wave data, decoupling is performed using Karenbauer transform. Zero-mode and line-mode feature vectors are extracted by combining CNN-LSTM and 1D-CNN neural network models. A fused feature vector is generated through interactive attention fusion. Finally, the trained neural network model is used to determine the fault line, type, and severity level.
It enables efficient and accurate fault location and detailed fault feature identification for distribution network faults, improving the efficiency and accuracy of fault diagnosis.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This application generally relates to the field of power system operation and maintenance technology. More specifically, this application relates to a method, system, medium, and equipment for determining faults in a distribution network including distributed wind farms. Background Technology
[0002] As the final link in the power system, the safe and stable operation of the distribution network is crucial for ensuring electricity supply for social production and residential life. When a line fault occurs, quickly and accurately identifying the faulty line and pinpointing the exact location of the fault is a prerequisite for rapid isolation and restoration of power supply.
[0003] Accurate fault diagnosis in distribution networks is the cornerstone of ensuring power supply reliability. Traditional diagnostic techniques are based on power system models dominated by synchronous generators, and their core relies on the fundamental characteristic that the system provides a large and clearly defined short-circuit current during a fault. However, with the large-scale grid connection of inverter-type resources, such as distributed wind farms, the transient characteristics of power systems have undergone a radical change, posing a severe challenge to the fault detection range and reliability of traditional diagnostic methods.
[0004] First, the electrical characteristics of faults are fundamentally weakened. Wind farms are connected to the grid via inverters. To protect the semiconductor devices inside the inverters, the inverter control system is designed to strictly limit its fault current output, typically not exceeding 1.2 times its rated value. This significantly reduces the short-circuit current level of the distribution network in areas with a high proportion of wind power integration. Traditional protection and diagnosis methods based on overcurrent principles are thus rendered "blind," severely limiting the wind farm's ability to perceive external faults and resulting in low accuracy.
[0005] Secondly, the complexity and proactivity of the fault response strategy. Unlike the passive response of synchronous generators, the fault response of inverters is entirely determined by their complex, software-based low-voltage ride-through control strategy. When an external grid fault causes a voltage drop at the grid connection point, the wind farm immediately triggers a series of coordinated response strategies to address the core power imbalance problem (i.e., the input mechanical power is much greater than the instantaneous output electrical power) and fulfill the support obligations stipulated in the grid guidelines. These response strategies mainly include: (1) Fast response of power electronics: executed by the inverter control system. This includes rapidly injecting inductive reactive current to support the grid voltage, actively reducing active power output, and converting excess energy into rotational kinetic energy for storage by controlling the generator torque.
[0006] (2) Coordinated response of mechanical system: executed by the mechanical system of the entire wind turbine. The variable pitch control will activate, and quickly reduce the mechanical power captured from the wind by adjusting the angle of the blades towards the wind.
[0007] Understandably, this series of complex and programmed active response strategies results in voltage and current waveforms during faults exhibiting highly nonlinear, time-varying, and multi-band harmonic characteristics, posing a significant challenge to traditional analysis methods. Summary of the Invention
[0008] In order to at least solve one or more of the technical problems mentioned above, this application proposes a method, system, medium and equipment for determining faults in a distribution network containing distributed wind farms in several aspects.
[0009] In the first aspect, the fault determination method for distribution networks containing distributed wind farms provided in this application includes the following steps: Acquire transient traveling wave data of each line in the distribution network connected to the distributed wind farm to obtain multiple transient traveling wave data. Decoupling operations are performed on multiple transient traveling wave data to generate zero-mode components and line-mode components for each transient traveling wave data. The zero-mode components of each transient traveling wave data are input into the first neural network model to generate the zero-mode feature vector of each transient traveling wave data. At the same time, the linear mode components of each transient traveling wave data are input into the second neural network model to generate the linear mode feature vector of each transient traveling wave data. The first neural network model includes a CNN-LSTM model, and the second neural network model includes a one-dimensional convolutional neural network 1D-CNN model. The zero-mode component and the linear-mode component of each transient traveling wave data are fused to obtain the fused feature vector of each transient traveling wave data. The fused feature vectors of each transient traveling wave data are input into a trained third neural network model to obtain the fault characteristics of the distribution network, wherein the fault characteristics include fault line identification, fault type and severity level.
[0010] In some examples, the training process of the trained neural network model includes: Multiple fault samples were obtained from multiple distribution networks, and the types of wind farms connected to each distribution network were different. Each fault sample is labeled with the fault line identifier, fault type, and severity level; Each fault sample, carrying the fault line identifier, fault type, and severity level, is input into the third neural network model to obtain the trained third neural network model.
[0011] In some examples, performing decoupling operations on the multiple transient traveling wave data includes: Using the Karenbauer transform, phase mode transformation is performed on multiple transient traveling wave data to generate the zero-mode component and line-mode component of each transient traveling wave data.
[0012] In some examples, after performing decoupling operations on the multiple transient traveling wave data respectively, the method further includes: Multi-scale decomposition is performed on the zero-mode and linear-mode components of each transient traveling wave data to generate multiple subsequences corresponding to the zero-mode and linear-mode components of each transient traveling wave data.
[0013] In some examples, before inputting the zero-mode components of each transient traveling wave data into the first neural network model and the zero-line mode components of each transient traveling wave data into the second neural network model, the method further includes: Multiple subsequences corresponding to the zero-mode components of each transient traveling wave data are spliced together. At the same time, multiple subsequences corresponding to the linear-mode components of each transient traveling wave data are spliced together to obtain the zero-mode components and linear-mode components of each transient traveling wave data after splicing.
[0014] In some examples, the step of inputting the zero-mode components of each transient traveling wave data into a first neural network model to generate zero-mode feature vectors for each transient traveling wave data, and simultaneously inputting the linear mode components of each transient traveling wave data into a second neural network model to generate linear mode feature vectors for each transient traveling wave data, includes: Using the first neural network model, the zero-mode feature vector of each transient traveling wave data is extracted from the zero-mode component after splicing each transient traveling wave data. Simultaneously, using the second neural network model, the linear mode feature vectors of each transient traveling wave data are extracted from the linear mode components after splicing the transient traveling wave data.
[0015] In some examples, the fused feature vectors of various transient traveling wave data are input into a trained third neural network model to obtain the fault characteristics of the distribution network, including: Based on the fused feature vectors of each transient traveling wave data, the third neural network model performs the following operations: Based on the fused feature vectors of each transient traveling wave data, the fault probability of each line is calculated respectively. Select the line with the highest probability of failure as the faulty line. The fault characteristics of the distribution network are determined based on the fused feature vector corresponding to the faulty line. Based on the fault characteristics, a corresponding response strategy is generated.
[0016] In the second aspect, the fault determination system for distribution networks including distributed wind farms provided in this application includes: The acquisition module is configured to acquire transient traveling wave data of each line of the distribution network connected to the distributed wind farm, and obtain multiple transient traveling wave data. The decoupling module is configured to perform decoupling operations on the multiple transient traveling wave data respectively, generating the zero-mode component and the line-mode component of each transient traveling wave data; The generation module is configured to input the zero-mode components of each transient traveling wave data into a first neural network model to generate zero-mode feature vectors of each transient traveling wave data, and simultaneously input the linear mode components of each transient traveling wave data into a second neural network model to generate linear mode feature vectors of each transient traveling wave data. The first neural network model includes a CNN-LSTM model, and the second neural network model includes a one-dimensional convolutional neural network 1D-CNN model. The fusion module is configured to fuse the zero-mode components and line-mode components of each transient traveling wave data to obtain the fused feature vector of each transient traveling wave data. The input module is configured to input the fused feature vector of each transient traveling wave data into a trained third neural network model to obtain the fault characteristics of the distribution network, wherein the fault characteristics include fault line identification, fault type and severity level.
[0017] In a third aspect, this application provides a computer-readable storage medium containing program instructions that, when executed by a processor, cause the method disclosed in the first aspect to be implemented.
[0018] In a fourth aspect, this application provides an electronic device, comprising: Processor; and A memory storing computer instructions that, when executed by the processor, cause the electronic device to perform the method disclosed in the first aspect above.
[0019] Compared with the prior art, this application has the following advantages: By decoupling the transient traveling wave data of each line in the distribution network, and using different neural network models to extract the zero-mode feature vector and line-mode feature vector from the decoupled transient traveling wave data, efficient and accurate fault line selection and fine identification of fault characteristics in the distribution network are achieved, improving efficiency and accuracy. Attached Figure Description
[0020] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, and the same or corresponding reference numerals denote the same or corresponding parts, wherein: Figure 1 This paper presents an exemplary flowchart of a method for determining faults in a distribution network containing distributed wind farms, as provided in an embodiment of this application. Figure 2 An exemplary framework diagram of the first neural network model provided in an embodiment of this application is shown; Figure 3 An exemplary framework diagram of the second neural network model provided in an embodiment of this application is shown; Figure 4 This paper presents an exemplary structural diagram of a power distribution network fault determination system containing distributed wind farms, provided in an embodiment of this application. Figure 5 An exemplary structural block diagram of an electronic device according to some embodiments of this application is shown. Detailed Implementation
[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0022] It should be understood that the terms "comprising" and "including" used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0023] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.
[0024] As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."
[0025] The specific embodiments of this application will now be described in detail with reference to the accompanying drawings.
[0026] like Figure 1As shown in the embodiments of this application, the method for determining faults in a distribution network containing distributed wind farms includes the following steps: S101: Acquire transient traveling wave data of each line in the distribution network connected to the distributed wind farm, and obtain multiple transient traveling wave data.
[0027] Understandably, a high-quality data source is a prerequisite for the validity of this application. In this step, three-phase data is acquired by deploying high-speed data acquisition cards based on field-programmable gate arrays at each line terminal. Microsecond-level clock synchronization using satellite timing services such as GPS or BeiDou ensures high fidelity in both time and amplitude of the acquired traveling wave signal. The high-speed data acquisition card uses a sampling rate of no less than 1MHz to ensure complete capture of the steep wavefront data of the high-resolution traveling wave signal, improving the fidelity of the traveling wave data. The high-speed data acquisition card monitors the line signal change rate in real time, and automatically extracts the complete traveling wave signal for a period before and after the fault when it exceeds the limit.
[0028] Understandably, due to training and practical application requirements, neural network models need a pre-emptive fault transient event capture mechanism before operation. The high-speed data acquisition card's data calculation module calculates the rate of change of voltage and current in real time. During normal operation, this value remains at a low level. Once a fault occurs, the arrival of a traveling wave signal causes this value to increase sharply. When the rate of change exceeds a preset threshold, the high-speed data acquisition card determines that a fault transient has occurred and automatically captures three-phase voltage and current waveform data before and after the trigger point, using a sampling rate of 2MHz for up to 1.024 milliseconds (in the example), totaling 2048 sampling points, to form a complete fault event sample.
[0029] S102 performs decoupling operations on multiple transient traveling wave data to generate the zero-mode component and line-mode component of each transient traveling wave data.
[0030] In some examples, this step specifically includes: Using the Karenbauer transform method, phase mode transformation operations are performed on multiple transient traveling wave data to generate zero-mode components and line-mode components for each line of the distribution network.
[0031] Understandably, due to electromagnetic coupling between the three phases, the original three-phase timing data are interconnected, and direct analysis would introduce errors. Phase-mode transformation is a linear mathematical decoupling method that converts the interconnected three-phase timing data into three mathematically independent modulus components. The zero-mode component primarily reflects the system's asymmetry and is sensitive to ground faults. The two line-mode components primarily reflect the electrical relationships between phases and are sensitive to phase-to-phase short-circuit faults. The phase-mode transformation matrix corresponding to the Karenbauer transformation is as follows: (1) In equation (1), i0 is the zero-mode component, i1 and i2 are the linear-mode components, and i A i B andi C These are the current values for phase A, phase B, and phase C, respectively.
[0032] In some examples, after performing decoupling operations on the multiple transient traveling wave data respectively, the method further includes: Multi-scale decomposition is performed on the zero-mode and linear-mode components of each transient traveling wave data to generate multiple subsequences corresponding to the zero-mode and linear-mode components of each transient traveling wave data.
[0033] In some examples, before inputting the zero-mode components of each transient traveling wave data into the first neural network model and the zero-line mode components of each transient traveling wave data into the second neural network model, the method further includes: Multiple subsequences corresponding to the zero-mode components of each transient traveling wave data are spliced together. At the same time, multiple subsequences corresponding to the linear-mode components of each transient traveling wave data are spliced together to obtain the zero-mode components and linear-mode components of each transient traveling wave data after splicing.
[0034] Understandably, multiple subsequences, comprising macroscopic, mesoscopic, and microscopic hierarchical subsequences, are concatenated after edge processing. This step involves using windows to extract sequences of different lengths from the data, guiding the neural network model to understand the traveling wave signal from different time scales. The 2048-point sequence, after multi-scale decomposition, generates three subsequences in parallel. Specifically, a macroscopic-scale sequence reflecting the overall envelope of the fault energy is obtained through RMS pooling with a large 128-point window; a mesoscopic-scale sequence reflecting the main energy oscillations is obtained through RMS pooling with a medium 32-point window; and the original sequence itself is used as the microscopic-scale sequence to preserve wavefront details. After edge processing to align their lengths, the three subsequences are concatenated along the feature dimension to form a more information-rich input tensor.
[0035] S103, the zero-mode components of each transient traveling wave data are input into the first neural network model to generate the zero-mode feature vector of each transient traveling wave data. At the same time, the linear mode components of each transient traveling wave data are input into the second neural network model to generate the linear mode feature vector of each transient traveling wave data. The first neural network model includes a CNN-LSTM model, and the second neural network model includes a one-dimensional convolutional neural network 1D-CNN model.
[0036] In some examples, this step specifically includes: Using the first neural network model, the zero-mode feature vector of each transient traveling wave data is extracted from the zero-mode component after splicing each transient traveling wave data. Simultaneously, using the second neural network model, the linear mode feature vectors of each transient traveling wave data are extracted from the linear mode components after splicing the transient traveling wave data.
[0037] Understandably, since zero-mode signals and line-mode signals have different fault response characteristics, the architecture of the neural network model will differ. First, a hybrid CNN-LSTM neural network model is used to extract local waveform features from the spliced zero-mode components. Then, an LSTM model is used to understand the temporal relationships between local features to determine directionality. A multi-layered one-dimensional convolutional neural network model (1D-CNN) is employed, allowing it to perform multi-level filtering through multiple convolutions. Because the ability of line-mode components to exhibit interphase faults is concentrated in certain frequency bands, the multi-layered one-dimensional convolutional network model (1D-CNN) can effectively capture energy characteristics. (Reference) Figure 2 The 1D-CNN consists of convolutional layers with 64 filters and 7 kernels, pooling layers with a size of 2, convolutional layers with 128 filters and 5 kernels, pooling layers with a size of 2, convolutional layers with 256 filters and 3 kernels, pooling layers with a size of 2, and finally global pooling to obtain 256-dimensional features. (Reference) Figure 3 The CNN-LSTM model contains convolutional layers with 32 filters and 5 kernels, and convolutional layers with 64 filters and 5 kernels. After passing through a pooling layer of size 2, the input is fed into the LSTM model to learn the time dependency. Its output will be transformed into a new feature vector by a fully connected layer with 128 neurons and an activation function of ReLU.
[0038] S104, fuse the zero-mode component and the linear-mode component of each transient traveling wave data to obtain the fused feature vector of each transient traveling wave data.
[0039] Specifically, the concatenated zero-mode feature vector and line-mode feature vector are input into the interactive attention fusion unit, so that the zero-mode feature vector and the line-mode feature vector are used to generate a highly collaborative fusion feature vector through bidirectional cross-channel query, attention weight calculation and weighted summation.
[0040] Understandably, this step improves upon traditional simple feature concatenation and fixed-weight fusion methods. In the first interaction, the zero-mode feature vector F0 is mapped to the query vector Q through three independently learned linear transformation layers. 0，Simultaneously, the linear model feature vector F1 is mapped to the key vector K1 and the value vector V1. Then, the dot product similarity between the query vector Q0 and the key vector K1 is calculated. To prevent the dot product from becoming too large and causing gradient vanishing, it is divided by the square root of the key vector dimension for scaling. The scaled result is normalized using the Softmax function to obtain the first set of attention weight scores. Finally, the weight scores are weighted and summed with V1 to obtain the new linear model vector F1'. The second interaction should follow the method of the first interaction, performing similar weight processing on the zero-mode component F0 to transform it into F0'. The final output is a high-cooperative feature vector composed of the four vectors F0, F1, F0', and F1'.
[0041] S105, the fused feature vectors of each transient traveling wave data are input into the trained third neural network model to obtain the fault characteristics of the distribution network, wherein the fault characteristics include fault line identification, fault type and severity level.
[0042] In some examples, step S105 specifically includes: Based on the fused feature vectors of each transient traveling wave data, the trained third neural network model performs the following operations: S201, calculate the fault probability of each line based on the fused feature vector of each line.
[0043] Understandably, the trained third neural network model utilizes its built-in fully connected decision network, which consists of three fully connected layers. The input layer receives a 768-dimensional fused feature vector output by the interactive attention module, followed by two hidden layers with 256 and 128 neurons respectively, using ReLU as the activation function. To prevent overfitting, a Dropout layer with a dropout rate of 0.5 is connected after each hidden layer, and the final output layer has the same number of neurons as the total number of lines. The raw score vector Z generated by this output layer is then fed into the Softmax function to calculate the failure probability of each line.
[0044] S202, select the line with the highest probability of failure as the faulty line.
[0045] S203, determine the fault characteristics of the distribution network based on the fusion feature vector corresponding to the faulty line.
[0046] Understandably, after the initial assessment of the faulty line, the process moves into a deeper analysis of the fault's nature. A 768-dimensional fused feature vector representing the faulty line's characteristics is input into an independent, pre-trained fault feature identification network. The core function of this network is to perform multi-task learning or multi-label classification. Its internal structure employs a multilayer perceptron (MLP), which maps the high-dimensional fused feature vector to different fault category and severity level spaces through multiple nonlinear transformations. In its design, the network's output layer is configured with multiple independent fault type classification heads for refined identification. Each fault type classification head is a multi-classifier, with the number of output nodes corresponding to predefined fault types. The output head uses a Softmax activation function to provide the probability of each fault type. Simultaneously, this fault feature identification network should also have a continuous numerical output quantifying the fault severity; the severity level assessment will influence the aggressiveness of the wind power system's LVRT control strategy. This step transforms the complex neural network model's judgment into a structured, human-machine-readable diagnostic report that directly guides subsequent decision-making.
[0047] S204. Generate a corresponding response strategy based on the fault characteristics.
[0048] It is understandable that the fault characteristics output in step S203 serve as the core premise for triggering the response strategy generation. The response strategy generation module has a built-in expert rule base. This expert rule base consists of a series of rules based on "if...then..." logic, which are pre-established based on power grid guidelines, relay protection experience, wind power control strategies, and extensive simulation analyses. For example, a rule might include "metallic phase-to-phase short circuit with a severity level greater than 8," triggering a response strategy of "emergency suppression of active power, enhanced reactive power priority LVRT, and prohibition of reclosing." The expert rule base is a key component of this application's intelligent decision-making mapping between refined fault characteristic identification and adaptive response strategy recommendations. Its establishment follows knowledge engineering principles, transforming the professional knowledge scattered across power grid operation procedures, relay protection experience, wind power grid connection guidelines, and extensive simulation analyses into structured "if...then..." rules executable by a computer program. This rule base is established and continuously improved outside of the offline training phase and is invoked during diagnosis to generate response strategies.
[0049] The table below shows a portion of the expert rule base used in this application. This is just an example; the number of rules will be further expanded based on system complexity and requirements during actual deployment.
[0050] Table 1 In some examples, the training process of a trained neural network model includes: Multiple fault samples were obtained from multiple distribution networks, and the types of wind farms connected to each distribution network were different. Each fault sample is labeled with the fault line identifier, fault type, and severity level; Each fault sample, carrying the fault line identifier, fault type, and severity level, is input into the third neural network model to obtain the trained third neural network model.
[0051] Specifically, the training process of the trained third neural network model is as follows: S301: Generate fault samples containing various wind power operating conditions through simulation software.
[0052] Understandably, to enable the model to deeply understand the fault characteristics of wind-powered systems, this step requires building a distribution network simulation model centered on distributed wind farms. This model must include detailed information about wind turbine generators and their low-voltage ride-through control strategies. Simulations generate fault samples covering various operating conditions. These samples encompass different fault types, including single-phase grounding, phase-to-phase short circuits, and three-phase short circuits; fault locations, including near-end, mid-range, and far-end faults; fault impedances, including metallic and high-impedance faults; fault occurrence times; and different operating states of the wind farm, including wind speed and power output.
[0053] S302: Accurately label each sample with faulty line tags and detailed fault characteristics.
[0054] Understandably, this invention employs a supervised learning paradigm, requiring high-quality labeled data. The labeling work mainly includes two levels: first, labeling each sample in the intelligent line selection model with the corresponding faulty line identifier; second, labeling the fault feature identification network with more detailed fault features, including fault types such as high-resistance grounding and phase-to-phase short circuits, and fault severity levels. The generation of fault severity level data is crucial in this step. It is not simply output directly from the simulation software, but rather constructed through a combination of "comprehensive evaluation" and "expert experience quantification." The specific process is as follows: For each fault sample generated by simulation, a series of key electrical indicators that can reflect the severity of the fault are extracted from the simulation results during the fault occurrence and transient process, including but not limited to: the minimum voltage drop depth, the peak transient current, the fault duration, the transient recovery time, the transient frequency deviation, the voltage drop area range, the zero-sequence current amplitude or the zero-sequence voltage amplitude, especially for the characteristics of ground faults.
[0055] Based on the knowledge of relay protection experts and power grid dispatching experts with many years of experience, as well as the national and local power grid operation guidelines and low voltage ride-through requirements, a comprehensive judgment is made on the various indicators extracted above.
[0056] The qualitative "severity" determined by experts is quantified onto a predefined integer scale of 1 to 10. This mapping is typically established using a lookup table or rule-based system.
[0057] In the initial stage of data generation, some samples need to be manually reviewed and verified to ensure that the quantified severity level labels match the actual physical impact of the fault. In this way, each simulation sample is assigned an accurate and physically meaningful fault severity level numerical label, thereby constructing the supervised learning data required for training.
[0058] S303: Divide the labeled dataset into training, validation and test sets proportionally.
[0059] Understandably, rigorous data partitioning is a prerequisite for ensuring scientific model training and fair evaluation. The training set is used for learning, the validation set for tuning, and the test set for final evaluation.
[0060] S304: Define the hyperparameter search space, including learning rate, λ value, etc., and start the optimizer.
[0061] Understandably, this step is the starting point for automating the search for the optimal model configuration. Hyperparameters are parameters that need to be manually set before model training, not parameters learned by the model itself. In this embodiment, the search space mainly includes: an initial learning rate between 1e-4 and 1e-2, physical constraint weights λ between 0.01 and 1.0, the number of filters in the CNN layer, the number of hidden units in the LSTM layer, and optimizer-related parameters, etc. The optimizedr used is preferably a Bayesian optimizer because it can intelligently approach the optimal solution with fewer trials than traditional grid search or random search.
[0062] S305: The optimizer generates a new set of hyperparameter configurations.
[0063] Understandably, this step is the core intelligent manifestation of Bayesian optimization. The optimizer internally maintains a surrogate model that learns and constructs a probabilistic prediction map about the relationship between hyperparameters and model performance based on the results of all completed trials. Then, through a sampling function, it intelligently proposes the next set of hyperparameter configurations that are most worthy of fine-tuning, balancing the choice between meticulously exploring known optimal regions and probing unknown regions with insufficient information.
[0064] S306: Train the model from scratch on the training set using a new hyperparameter configuration.
[0065] Understandably, for each set of hyperparameters generated by S305, an independent and complete model training is required. This process starts with a randomly initialized model, using only the training set data, and through multiple rounds of iterative learning, allows the model to adjust its internal weights based on the training data.
[0066] S307: Apply a hybrid loss function with physical constraints to update model weights during training.
[0067] Understandably, this step uses a hybrid loss function as the model training standard to improve the model's generalization ability across multiple scenarios. In the practical application of the loss function, the cross-entropy loss obtained from the prediction of faulty samples and the physical constraints obtained from the features output by the zero-mode channel of weighted healthy samples will be used to calculate the total loss, which will ultimately guide the Adam optimizer in training. In this embodiment, the training process is implemented based on the PyTorch deep learning framework. The training set prepared in S301 is used for model parameter learning. The Adam optimizer is used for training, with an initial learning rate set to 1e-3, a total training epochs set to 100, and an early stopping mechanism introduced to stop learning when the training requirements are met to prevent model overfitting. This step uses an innovative hybrid loss function, as shown in the following formula. Wherein, the cross-entropy loss... This is the primary optimization objective, serving to measure the difference between the model's predicted probabilities and the true labels. The loss function adjusts the model parameters using the predicted values from each output. This term is designed to make the probability of correctly predicting the class as close to 1 as possible during training, thereby improving the model's classification accuracy. Physical constraint regularization term. The model will penalize incorrect judgments when it receives data samples labeled as healthy. Since the zero-mode feature vector output from healthy data after passing through the zero-mode channel should have a magnitude extremely close to 0, a large output magnitude indicates that the model is misinterpreting the data. Therefore, a regularization term is used to calculate the magnitude. Depending on the length of the obstacle, a corresponding penalty will be imposed on the model. After training, the weights of the best-performing model are fixed, resulting in the final intelligent lane selection model that can be used for online diagnostics.
[0068] S308: After training is complete, Bayesian optimization evaluates the predictive performance of the current model on the validation set and proposes the next hyperparameter configuration based on the evaluation results, until the model performance configured by the hyperparameters meets the stopping condition.
[0069] Understandably, the Bayesian optimizer internally maintains a surrogate model that makes probabilistic predictions about the relationship between hyperparameters and model performance based on the loss function values from completed tests. During its continuous updating of these internal probabilistic predictions, when a preset number of searches is reached, the Bayesian optimizer will provide a set of optimal hyperparameters. This set of hyperparameters will then be used for subsequent validation set adjustments. S309: Repeat the optimization process until the preset stopping condition is met or the performance converges to obtain the optimal hyperparameter configuration.
[0070] Understandably, this step represents the final decision and loop control point in the neural network model training and optimization phase. The performance evaluation results from S308 are precisely fed back to the Bayesian optimizer in S305. The optimizer uses this information to update its internal surrogate model, thereby more accurately estimating the potential performance of different hyperparameter combinations. Then, the system determines whether preset stopping conditions have been met. These stopping conditions may include: reaching the maximum number of preset hyperparameter search attempts; the optimizer finding a good local optimum in the current search space; and the validation set performance meeting or exceeding the preset minimum requirements.
[0071] Understandably, if any of the above stopping conditions are met, the entire Bayesian optimization loop will terminate, and the set of hyperparameters that performed best across all trials will be output. If not, the Bayesian optimizer will intelligently generate the next set of hyperparameters that is most promising for improving model performance and instruct the system to return to step S305 to continue the next round of training and evaluation until the optimal intelligent line selection model hyperparameters are found.
[0072] S401: Define the hyperparameter search space of the fault feature identification model and start the optimizer.
[0073] Understandably, this step prepares the network for training the fault feature identification network. The hyperparameter search space will cover key parameters affecting model performance, including: the initial learning rate, the number of hidden layers, the number of neurons per layer, and the dropout rate. Given that the fault feature identification network may be structurally simpler than the intelligent line selection model, in addition to Bayesian optimization, optimizers such as grid search or random search can also be considered to systematically or randomly explore predefined combinations of parameters to find the optimal configuration.
[0074] S402: The optimizer generates a new set of hyperparameter configurations.
[0075] Understandably, the core task of a Bayesian optimizer is to generate a unique combination of hyperparameters based on a preset search space and strategy, and intelligently select the parameter combination most likely to improve performance.
[0076] S403: Using a new configuration, train a single-input multi-output fault feature identification model from scratch on the training set.
[0077] Understandably, for each set of hyperparameters generated by S402, it is necessary to start from a randomly initialized model state and train using a dataset in the training set that has already been labeled with "fault type" and "fault severity level". The training input here is a 768-dimensional fused feature vector of the same simulation samples, and the output is the fault type probability distribution and fault severity level value predicted by the model. The model learns the mapping relationship between the fused features and these two types of diagnostic information by continuously adjusting the internal weights and biases.
[0078] S404: Apply a multi-task loss function to update the weights of the fault feature identification model.
[0079] Understandably, since the fault feature identification network simultaneously undertakes two tasks—fault type classification and fault severity assessment—its optimization objective is a multi-task loss function. This loss function aims to minimize the prediction errors of both tasks simultaneously, thereby guiding the model to learn shared feature representations that are beneficial to both tasks. Specifically, the multi-task loss function... It can be defined as the loss of the classification task. Losses related to the return mission Weighted sum: (2) In formula (2): It is a weighting coefficient, whose function is to balance the importance of classification and regression tasks in the total loss. Cross-entropy loss for classifying fault types. For a given set of fault types... For a batch of samples, if there are C types of faults, the formula is: (3) In equation (3), Is the first The actual fault type of each sample is An indicator variable that is 1 if it is true, and 0 otherwise; It is the model's prediction of the first Each sample belongs to the fault type. The probability of correct classification. This loss function aims to maximize the probability of correct classification.
[0080] This is the mean squared error loss for fault severity level regression. For a given set of fault severity levels... The formula for a batch of samples is: (4) In equation (4), It is the first The true fault severity level quantification value for each sample It is the model's prediction of the first The fault severity level is quantified for each sample. The loss function aims to minimize the squared error between the predicted severity level and the true severity level.
[0081] The total error calculated by the multi-task loss function will be used to calculate the gradient through the backpropagation algorithm. The optimizer Adam will then use these gradients to iteratively update all the weight parameters of the fault feature identification network, thereby enabling it to simultaneously optimize the performance of both fault type classification and severity assessment tasks.
[0082] S405: After training is complete, evaluate the performance of the current fault feature identification model on the validation set and feed it back to the optimizer.
[0083] Understandably, this step is an objective measure of the model's performance under the current hyperparameter configuration. Using the currently trained model, predictions are made on the validation set—data not directly used during model training—and performance metrics such as total loss, fault type classification accuracy, and mean squared error of severity level regression are calculated. These metrics will serve as the basis for further hyperparameter adjustments by the optimizer in S301.
[0084] S406: Repeat the optimization process until the preset stopping condition is met or the performance converges to obtain the optimal hyperparameter configuration.
[0085] Understandably, the loops from S402 to S405 will continue until preset stopping conditions are met. These stopping conditions may include reaching the maximum number of iterations, the validation set performance no longer improving over multiple consecutive rounds (i.e., convergence), or finding a hyperparameter configuration that satisfies a specific performance threshold. Ultimately, the optimization process will output the hyperparameter combination of the fault feature identification model that performs best on the validation set.
[0086] S501: Select the optimal hyperparameter combination of the intelligent lane selection model and the fault feature identification model on the validation set from all experiments.
[0087] Understandably, this step summarizes all the experimental results from the training of the intelligent lane selection model and the fault feature identification model. The system will select the hyperparameter configuration that performs best on the validation set from its respective optimization records, and use it as the basis for the final model training.
[0088] S502: Using their respective optimal hyperparameters, perform a final full training on the complete training set for both models.
[0089] Understandably, to maximize the model's generalization ability and utilize all available labeled data, after determining the optimal hyperparameters, a final full training run will be performed on the intelligent lane selection model and the fault feature identification model using this set of hyperparameters. This training is typically conducted on a larger dataset to ensure that the model can fully learn the patterns in the data.
[0090] S503: On an independent test set, perform a one-time performance evaluation on the finally trained intelligent lane selection model and fault feature identification model respectively.
[0091] Understandably, this step is an important way to measure the true generalization ability of the two final models. The test set is data that the model has not encountered during the entire offline training process. By running and evaluating on this completely independent dataset, we can obtain the lane selection accuracy, recall, and F1 score of the intelligent lane selection model, as well as the fault type classification accuracy and mean squared error of the severity level regression of the fault feature identification model.
[0092] S504: Saves the weights of the two final models that meet the performance requirements and deploys them on the required terminals.
[0093] Understandably, this step represents the endpoint and final outcome of the entire offline training process. Only when the performance evaluation results meet the preset standards of route selection accuracy > 99.8%, fault type classification accuracy > 98%, and severity level assessment error < 5% will all learned weight parameters be serialized and saved as .pth format files. These files will then be loaded into the intelligent route selection and fault feature identification parts of the online diagnostic system, thereby providing real-time and efficient fault perception and response capabilities in actual operation.
[0094] As described above, this application decouples the transient traveling wave data of each line in the distribution network and uses different neural network models to extract the zero-mode feature vector and line-mode feature vector from the decoupled transient traveling wave data, thereby achieving rapid and accurate fault line selection and fine-grained fault feature identification for power grid faults. Furthermore, this application can transform the diagnostic results of the neural network model into specific and executable adaptive response strategy suggestions for wind power systems through an expert rule base. This not only significantly improves the autonomous decision-making ability and proactive support performance of distributed wind power systems in dealing with complex power grid faults, but also enables them, as passive participants, to actively maintain power grid stability, which is of great significance for improving the reliability, robustness, and intelligence level of the distribution network.
[0095] This application also provides a system for implementing the aforementioned method. This system can be a software system deployed on a main server, or an edge computing device integrating hardware and software. (See reference...) Figure 4 The specific details provided in this application include: The acquisition module is configured to acquire transient traveling wave data of each line of the distribution network connected to the distributed wind farm, and obtain multiple transient traveling wave data. The decoupling module is configured to perform decoupling operations on the multiple transient traveling wave data respectively, generating the zero-mode component and the line-mode component of each transient traveling wave data; The generation module is configured to input the zero-mode components of each transient traveling wave data into a first neural network model to generate zero-mode feature vectors of each transient traveling wave data, and simultaneously input the linear mode components of each transient traveling wave data into a second neural network model to generate linear mode feature vectors of each transient traveling wave data. The first neural network model includes a CNN-LSTM model, and the second neural network model includes a one-dimensional convolutional neural network 1D-CNN model. The fusion module is configured to fuse the zero-mode components and line-mode components of each transient traveling wave data to obtain the fused feature vector of each transient traveling wave data. The input module is configured to input the fused feature vector of each transient traveling wave data into a trained third neural network model to obtain the fault characteristics of the distribution network, wherein the fault characteristics include fault line identification, fault type and severity level.
[0096] On the other hand, embodiments of this application also provide an electronic device, see [link to relevant documentation]. Figure 5 , Figure 5 This is an exemplary structural block diagram of an electronic device according to an embodiment of this application, such as... Figure 5 As shown, the electronic device includes a processor and a memory, the memory storing computer instructions, and the processor executing the computer instructions to perform the method provided in this application.
[0097] Specifically, processor 601 may include a central processing unit (CPU) or a graphics processing unit (GPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application. Memory 602 may include memory for data or instructions. For example, memory 602 may be at least one of the following: a hard disk drive (HDD), read-only memory (ROM), random access memory (RAM), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, universal serial bus (USB) drive, or other physical / tangible memory storage device. Alternatively, memory 602 may include removable or non-removable (or fixed) media. Furthermore, memory 602 may be internal or external to the integrated gateway disaster recovery device. Memory 602 may be non-volatile solid-state memory. In other words, typically memory 602 includes a tangible (non-transitory) computer-readable storage medium (such as a memory device) encoded with executable instructions, wherein the stored executable instructions, when executed by processor 601 (e.g., by one or more processors), can implement the methods in the embodiments of this application.
[0098] In one example Figure 5 The illustrated electronic device may also include a communication interface 603 and a bus 610. The processor 601, memory 602, and communication interface 603 are connected via bus 610 and communicate with each other. Communication interface 603 is primarily used to enable communication between modules, devices, units, and / or equipment within the electronic device. Bus 610, including hardware, software, or both, couples components of the online data flow metering device together. For example, the bus may include at least one of the following: Accelerated Graphics Port (AGP) or other graphics bus, Enhanced Industry Standard Architecture (EISA) bus, Front Side Bus (FSB), HyperTransport (HT) interconnect, Industry Standard Architecture (ISA) bus, Infinite Bandwidth Interconnect, Low Pin Count (LPC) bus, memory bus, Microchannel Architecture (MCA) bus, Peripheral Component Interconnect (PCI) bus, PCI-Express (PCI-X) bus, Serial Advanced Technology Attachment (SATA) bus, Video Electronics Standards Association Local (VLB) bus, or other suitable buses. Bus 610 may include one or more buses. Although specific buses are described or illustrated in the embodiments of this application, any suitable bus or interconnection method may be considered in the embodiments of this application.
[0099] In another aspect, embodiments of this application also provide a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the aforementioned method. The computer-readable storage medium may be, for example, a classic computer-readable storage medium, such as a read-only memory (ROM), random access memory (RAM), disk storage media, optical storage media, flash memory, or other electrical, optical, or other physical / tangible memory storage devices.
[0100] In another aspect, embodiments of this application also provide a computer program product, which includes computer program instructions that, when executed by a processor, implement the method provided in embodiments of this application. This computer program product may be, for example, a software installation package, a plug-in compatible with a related software system, etc.
[0101] The flowcharts and / or block diagrams of the methods and systems of embodiments of this application have been described above by way of example, and related aspects have been described. It should be understood that each block or combination thereof in the flowcharts and / or block diagrams can be implemented by computer program instructions, by dedicated hardware performing a specified function or action, or by a combination of dedicated hardware and computer instructions. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc.; when implemented in software, it is a program or code segment used to perform the required task. The program or code segment can be stored in memory or transmitted over a transmission medium or communication link via data signals carried in a carrier wave. The code segment can be downloaded via a computer network such as the Internet, an intranet, etc.
[0102] It should be noted that in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover 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 limitations, 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.
[0103] While this application has shown and described numerous embodiments, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will arise for those skilled in the art without departing from the spirit and intent of this application. It should be understood that various alternatives to the embodiments of this application described herein may be employed in the practice of this application. The appended claims are intended to define the scope of protection of this application and therefore cover equivalents or alternatives within the scope of these claims.
Claims
1. A method for determining faults in a distribution network containing distributed wind farms, comprising: Acquire transient traveling wave data of each line in the distribution network connected to the distributed wind farm to obtain multiple transient traveling wave data. Decoupling operations are performed on multiple transient traveling wave data to generate zero-mode components and line-mode components for each transient traveling wave data. The zero-mode components of each transient traveling wave data are input into the first neural network model to generate the zero-mode feature vector of each transient traveling wave data. At the same time, the linear mode components of each transient traveling wave data are input into the second neural network model to generate the linear mode feature vector of each transient traveling wave data. The first neural network model includes a CNN-LSTM model, and the second neural network model includes a one-dimensional convolutional neural network 1D-CNN model. The zero-mode component and the linear-mode component of each transient traveling wave data are fused to obtain the fused feature vector of each transient traveling wave data. The fused feature vectors of each transient traveling wave data are input into a trained third neural network model to obtain the fault characteristics of the distribution network, wherein the fault characteristics include fault line identification, fault type and severity level.
2. The method for determining distribution network faults according to claim 1, characterized in that, The training process of the trained neural network model includes: Multiple fault samples were obtained from multiple distribution networks, and the types of wind farms connected to each distribution network were different. Each fault sample is labeled with the fault line identifier, fault type, and severity level. Each fault sample, carrying the fault line identifier, fault type, and severity level, is input into the third neural network model to obtain the trained third neural network model.
3. The method for determining distribution network faults according to claim 1, characterized in that, The decoupling operation performed on the multiple transient traveling wave data includes: Using the Karenbauer transform, phase mode transformation is performed on multiple transient traveling wave data to generate the zero-mode component and line-mode component of each transient traveling wave data.
4. The method for determining distribution network faults according to claim 3, characterized in that, After performing decoupling operations on the multiple transient traveling wave data respectively, the method further includes: Multi-scale decomposition is performed on the zero-mode and linear-mode components of each transient traveling wave data to generate multiple subsequences corresponding to the zero-mode and linear-mode components of each transient traveling wave data.
5. The method for determining distribution network faults according to claim 4, characterized in that, Before inputting the zero-mode components of each transient traveling wave data into the first neural network model and the zero-mode components of each transient traveling wave data into the second neural network model, the method further includes: Multiple subsequences corresponding to the zero-mode components of each transient traveling wave data are spliced together. At the same time, multiple subsequences corresponding to the linear-mode components of each transient traveling wave data are spliced together to obtain the zero-mode components and linear-mode components of each transient traveling wave data after splicing.
6. The method for determining distribution network faults according to claim 5, characterized in that, The step of inputting the zero-mode components of each transient traveling wave data into the first neural network model to generate the zero-mode feature vector of each transient traveling wave data, and simultaneously inputting the linear mode components of each transient traveling wave data into the second neural network model to generate the linear mode feature vector of each transient traveling wave data, includes: Using the first neural network model, the zero-mode feature vector of each transient traveling wave data is extracted from the zero-mode component after splicing each transient traveling wave data. Simultaneously, using the second neural network model, the linear mode feature vectors of each transient traveling wave data are extracted from the linear mode components after splicing the transient traveling wave data.
7. The method for determining distribution network faults according to claim 6, characterized in that, The fused feature vectors of various transient traveling wave data are input into a trained third neural network model to obtain the fault characteristics of the distribution network, including: Based on the fused feature vectors of each transient traveling wave data, the third neural network model performs the following operations: Based on the fused feature vectors of each transient traveling wave data, the fault probability of each line is calculated respectively. Select the line with the highest probability of failure as the faulty line. The fault characteristics of the distribution network are determined based on the fused feature vector corresponding to the faulty line. Based on the fault characteristics, a corresponding response strategy is generated.
8. A fault determination system for a distribution network containing distributed wind farms, characterized in that, include: The acquisition module is configured to acquire transient traveling wave data of each line of the distribution network connected to the distributed wind farm, and obtain multiple transient traveling wave data. The decoupling module is configured to perform decoupling operations on the multiple transient traveling wave data respectively, generating the zero-mode component and the line-mode component of each transient traveling wave data; The generation module is configured to input the zero-mode components of each transient traveling wave data into a first neural network model to generate zero-mode feature vectors of each transient traveling wave data, and simultaneously input the linear mode components of each transient traveling wave data into a second neural network model to generate linear mode feature vectors of each transient traveling wave data. The first neural network model includes a CNN-LSTM model, and the second neural network model includes a one-dimensional convolutional neural network 1D-CNN model. The fusion module is configured to fuse the zero-mode components and line-mode components of each transient traveling wave data to obtain the fused feature vector of each transient traveling wave data. The input module is configured to input the fused feature vector of each transient traveling wave data into a trained third neural network model to obtain the fault characteristics of the distribution network, wherein the fault characteristics include fault line identification, fault type and severity level.
9. A computer-readable storage medium, characterized in that, It includes program instructions that, when executed by a processor, cause the method according to any one of claims 1-7 to be implemented.
10. An electronic device, characterized in that, include: processor; as well as A memory storing computer instructions that, when executed by the processor, cause the electronic device to perform the method according to any one of claims 1-7.