Power distribution network fault identification method combining deep learning and attention

Abstract

The invention discloses a power distribution network fault identification method combining deep learning and attention, and relates to the technical field of power distribution network fault identification, and the method comprises the following steps: collecting real fault data according to a lightning monitoring system, combining a simulation platform to generate simulation fault data, and inputting the real fault data into a preset generative adversarial model, outputting enhanced fault data; real fault data, simulated fault data and enhanced fault data are mixed, and a training set and a test set are constructed; optimizing a preset deep learning network containing a fault loss function and an attention network through the training set, and obtaining a power distribution network fault recognition model; and evaluating the power distribution network fault identification model by using the test set to obtain a fault type identification result. The method is used for solving the problems of small data volume and insufficient recognition precision in power distribution network fault recognition.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power distribution network fault identification, and more particularly to a power distribution network fault identification method combining deep learning and attention. BACKGROUND

[0002] As a key link in the power system directly supplying power to users, the power distribution network is large in scale, complex in topology and changeable in operation environment, and has a direct impact on the production and life of users. The power distribution network is easily affected by human, weather and natural environment and the like, especially in the season with frequent thunderstorms, disaster events such as lightning strikes occur frequently, which easily causes power grid faults, reduces power supply reliability and limits safety and stability.

[0003] The core of ensuring the safe and stable operation of the power distribution network lies in realizing rapid and accurate identification of faults. However, the power distribution network is complex in topology and huge in information amount, and the cost of comprehensive data collection is high. The detection method based on traditional multi-source information often requires a large number of sensors to be laid out to obtain various electrical quantities, switching quantities and state quantity data, which not only increases the construction and operation cost, but also may exist data redundancy.

[0004] Through the deep learning method, complex features can be automatically learned from massive electrical quantities, switching quantities and state quantity data, and high-precision identification of the power distribution network fault is realized. At the same time, the method can reduce the dependence on the number of sensors under the premise of ensuring the identification performance, thereby reducing the data collection cost and providing reliable protection for the safe and stable operation of the power distribution network.

[0005] For example, the invention patent with the announcement number CN117851868B announces a multi-model propulsion type power distribution network fault identification method and system based on deep learning, the main steps of which are: based on the full waveform diagram of the power distribution network and the fault classification model, four common fault types can be identified, and for fault types with different fault characteristics, different subsequent processing is performed according to the characteristics of these faults, thereby realizing identification of multiple fault types and subsequent fault positioning. The circuit topology graph is traversed through horizontal and vertical sliding windows and layer sequence traversal strategies, and then the complex topology is split and pruned, avoiding full traversal of the complex topology graph or complete input of the complex topology graph, which makes the application have good performance even in the face of complex topology graphs, and the fault identification is accurate and high. Moreover, different topology graphs can also be used to generate images for identification in the same way, and the images required by the model have less difference on different topology graphs, and the model has good generalization.

[0006] For example, the invention patent announcement No. CN109239527A, a power distribution network fault identification method based on deep belief network, includes the following steps: step S1: collect the signal waveform data of the three-phase voltage, zero sequence voltage and three-phase current of the main transformer low-voltage bus under fault condition; Step S2: respectively pre-process the signal waveform data of the three-phase voltage and zero sequence voltage of the main transformer low-voltage bus and the three-phase current of the main transformer low-voltage side, and obtain the pre-processed waveform data; Step S3: four deep belief network models are constructed, each restricted Boltzmann machine is unsupervised trained, and the parameters of all layers are fine-tuned by using the error back propagation principle, and the trained deep belief network model is obtained; Step S4: the pre-processed waveform data is input into the four trained deep belief network models; and the output results of the four trained deep belief network models are combined, and compared with the pre-set fault type label, and the fault identification of the power distribution network is completed.

[0007] In the above disclosed technical solution, there are at least the following technical problems: the actual operation data of the power distribution network is limited, especially the fault data is scarce and unbalanced, which makes it difficult for the existing method to fully capture the fault of the power distribution network in the training and modeling process. The topology of the power distribution network is complex, and the operation condition is diverse, and the existing method relies heavily on data, which is difficult to maintain high-precision fault identification ability in the case of insufficient data. The existing method relies on simple feature extraction or rule judgment, which is difficult to effectively mine the nonlinear relationship hidden in multi-source data, resulting in inaccurate fault identification results, especially in extreme conditions such as lightning, false positives or false negatives are obvious.

[0008] In view of the above problems, the present application provides a solution. SUMMARY

[0009] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a power distribution network fault identification method combining deep learning and attention, which combines deep learning and attention to solve the problems of insufficient data and insufficient recognition accuracy in power distribution network fault identification.

[0010] To achieve the above purpose, the present application provides the following technical solutions: A power distribution network fault identification method combining deep learning and attention, comprising the following steps: collecting real fault data, generating simulated fault data in combination with a simulation platform, and performing data enhancement on the real fault data; mixing real fault data, simulated fault data and enhanced fault data, training a deep learning network and an attention network to obtain a power distribution network fault identification model; inputting electrical data into the power distribution network fault identification model to obtain a fault type identification result; wherein the data enhancement is obtained by training a generative adversarial model through an optimized composite loss; the composite loss is composed of an adversarial loss and an auxiliary classification loss, and is weighted and summed, aiming to simultaneously constrain the data distribution fidelity of the generated samples and the fault type classification accuracy.

[0011] In a preferred embodiment, the real fault data, simulated fault data and enhanced fault data include time points, three-phase current amplitudes, three-phase voltage amplitudes, zero sequence currents, lightning peak values and fault types.

[0012] In a preferred embodiment, the fault types include single-phase grounding, two-phase short circuit, two-phase grounding, three-phase short circuit and open circuit fault.

[0013] In a preferred embodiment, the data enhancement specifically comprises: inputting a random noise vector, a fault type label and a lightning feature parameter into a conditional generator to obtain generated fault data; inputting the real fault data and the generated fault data into a discriminator to output a discrimination result and a fault type classification result; constructing a generative adversarial model by calculating the composite loss of the discrimination result and the fault type classification result, and performing data enhancement on the real fault data; the composite loss is obtained by weighting and summing an adversarial loss and an auxiliary classification loss, wherein the adversarial loss is obtained by measuring the distribution difference between the generated samples and the real samples, and the auxiliary classification loss is obtained by measuring the consistency between the classifier prediction and the real fault type through cross-entropy.

[0014] In a preferred embodiment, the mixing of real fault data, simulated fault data and enhanced fault data includes aligning simulated fault data and enhanced fault data, specifically: mapping the real, simulated and enhanced fault data to the same feature space through a feature extraction network; optimizing the parameters of the feature extraction network by calculating and minimizing the maximum mean difference between simulated / enhanced data and real data in the feature space; and converting the simulated and enhanced fault data into aligned data that approximates the real fault data in feature distribution using the optimized network.

[0015] In a preferred embodiment, the training of the deep learning network and the attention network includes a multi-task focusing loss function, wherein the multi-task focusing loss function is obtained by weighted summation of classification loss, center loss and physical consistency loss; the physical consistency loss is obtained by weighted summation of three-phase voltage and current imbalance.

[0016] In a preferred embodiment, the training of the deep learning network and the attention network further includes a model testing step, specifically: predicting test data based on the power distribution network fault identification model to obtain predicted labels for fault types; calculating the accuracy, precision, recall, and F1 score of the fault types based on the predicted labels and the corresponding real labels; and using the average of the calculated accuracy, precision, recall, and F1 score as the model testing result.

[0017] The technical effects and advantages of the power distribution network fault identification method combining deep learning and attention proposed in this invention are as follows: This invention integrates real fault data collected by a lightning monitoring system, simulated fault data generated by a simulation platform, and enhanced data generated by a generative adversarial network, effectively solving the problems of insufficient data volume and limited sample diversity in distribution network faults. Through a pre-defined deep learning network and attention network, it helps to automatically extract key fault features and accurately model the temporal characteristics and spatial correlations of faults, effectively improving the accuracy and robustness of fault identification. By constructing training and testing sets and evaluating model performance, it helps to ensure the stability and repeatability of identification results. Through the above data fusion and deep learning combined with attention network processing methods, the problems of insufficient data volume and inadequate identification accuracy in distribution network fault identification are effectively solved. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a power distribution network fault identification method that combines deep learning and attention, provided as an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, Figure 1 This invention presents a method for fault identification in power distribution networks that combines deep learning and attention, comprising the following steps: S1: Based on the real fault data collected by the lightning monitoring system, simulated fault data is generated by combining the simulation platform, and the real fault data is input into the preset generative adversarial model to output enhanced fault data. S2, mixes real fault data, simulated fault data and augmented fault data to construct training and test sets; S3, by optimizing a pre-defined deep learning network and attention network containing a fault loss function through the training set, a power distribution network fault identification model is obtained; S4 uses the test set to evaluate the power distribution network fault identification model and obtain fault type identification results.

[0021] This embodiment effectively solves the problems of insufficient data volume and limited sample diversity in distribution network faults by fusing real fault data collected by a lightning monitoring system, simulated fault data generated by a simulation platform, and enhanced data generated by a generative adversarial network. Through a pre-defined deep learning network and attention network, it helps to automatically extract key fault features and accurately model the temporal characteristics and spatial correlations of faults, effectively improving the accuracy and robustness of fault identification. Constructing training and testing sets and evaluating model performance helps ensure the stability and repeatability of the identification results. The above data fusion and deep learning combined with attention network processing methods effectively solve the problems of insufficient data volume and inadequate identification accuracy in distribution network fault identification.

[0022] S1: Based on real fault data collected by the lightning monitoring system, simulated fault data is generated by combining the simulation platform, and the real fault data is input into the preset generative adversarial model to output enhanced fault data.

[0023] In this embodiment, the real fault data, simulated fault data, and enhanced fault data include time points, three-phase current amplitudes, three-phase voltage amplitudes, zero-sequence current, lightning peak values, and fault types.

[0024] It should be noted that the specific steps for collecting real fault data according to the lightning monitoring system are as follows: First, the detection station measures the time, direction, and relative signal strength of the ground flash return stroke and sends the raw measurement data to the central station in real time. The central station processes the signals from multiple detection stations through a pre-processor, and the analysis host performs location calculations to obtain the time point, location, peak lightning current, and polarity of the ground flash. Subsequently, this information is correlated with the distribution network measurement data or simulation calculation data to extract the three-phase current amplitude, three-phase voltage amplitude, and zero-sequence current at the moment the fault occurs. The fault type is then determined by combining the lightning peak value and distribution network topology analysis. Finally, the time point, three-phase current amplitude, three-phase voltage amplitude, zero-sequence current, lightning peak value, and fault type are compiled into samples to construct real fault data that can be used for deep learning training.

[0025] Furthermore, the specific steps for generating simulated fault data using the simulation platform are as follows: First, a distribution network model is established on the simulation platform, including node and line parameters, load and power generation curves, and lightning or short-term disturbance parameters are defined. Depending on the type of fault to be simulated (e.g., single-phase grounding, two-phase short circuit, two-phase grounding, three-phase short circuit, or line open circuit), an equivalent fault model is applied at a specified node or line location, and the fault occurrence time, duration, and lightning peak amplitude are set. Then, the distribution network state at the time of the fault is obtained through power flow calculation, and the three-phase current amplitude, three-phase voltage amplitude, and zero-sequence current are extracted, while the lightning peak value is recorded. Each sample retains only these key amplitude features and the corresponding fault type label, and the fault occurrence time is marked. To increase data diversity, random perturbations are applied to the load, fault location, and lightning peak value to generate simulated fault samples under different operating conditions. The final data includes the time point, three-phase current amplitude, three-phase voltage amplitude, zero-sequence current, lightning peak value, and fault type, which are used for training and testing the deep learning model.

[0026] In this embodiment, the fault types include single-phase grounding, two-phase short circuit, two-phase grounding, three-phase short circuit, and open circuit fault.

[0027] In this embodiment, the preset generative adversarial model is specifically constructed as follows: Input the random noise vector, fault type label, and lightning characteristic parameters into the condition generator to obtain generated simulated fault data; The real fault data and the generated simulated fault data are input into the discriminator, which outputs the realism discrimination result and the fault type classification result. Design a composite loss function and calculate the composite loss based on the authenticity judgment result and the fault type classification result; Based on the composite loss function, the parameters of the discriminator and generator are optimized through an alternating optimization strategy to obtain a generative adversarial model.

[0028] It should be noted that the lightning characteristic parameters are the time point, three-phase current amplitude, three-phase voltage amplitude, zero-sequence current, and lightning peak value. The input to the condition generator consists of two parts: a random noise vector and a condition vector. The random noise vector is randomly sampled from a Gaussian distribution, and the condition vector is a vector composed of a fault type label and lightning characteristic parameters. The fault type label is a one-hot label formed through one-hot encoding. Then, the random noise vector and the condition vector are concatenated along the feature dimension to form a joint input vector. Finally, the output value is limited to the range [-1, 1] through a fully connected layer, normalization, and a nonlinear activation function to obtain simulated fault data. This process can be expressed by the following formula:

[0029] In the formula, To simulate fault data samples, Let z be the neural network mapping function of the generator, z be the random noise vector, and c be the lightning feature vector.

[0030] The discriminator first extracts features from the input data through convolutional layers, combining normalization and non-linear activation functions to obtain a high-dimensional feature representation. Based on this, the discriminator employs a dual-output structure: on one hand, it outputs the authenticity judgment result through the Sigmoid activation function, with a value range of [0,1], used to distinguish whether the input data comes from real samples or generated samples; on the other hand, it outputs the fault type classification result through the Softmax activation function, with a value range of multi-class probability distribution, used to identify the fault type of the input data. The overall process is expressed by the formula:

[0031] In the formula, x represents the input data sample, which can be a real fault data sample or a generated simulated fault data sample. The overall output of the discriminator includes the results of the authenticity determination. Fault type classification results , and These are the neural network mapping functions used in the discriminator for authenticity determination and fault type classification, respectively. It is the Sigmoid activation function. For the Softmax function, These are the trainable parameters in the discriminator.

[0032] In this embodiment, the alternating optimization strategy refers to fixing the parameters of the discriminator and the generator respectively and updating them alternately during the training process of the generative adversarial model.

[0033] In this embodiment, the composite loss function includes adversarial loss and auxiliary classification loss.

[0034] It should be noted that the specific formula for adversarial loss is:

[0035] In the formula, To combat the loss, z is a random noise vector, which is the input to the generator, and y is the one-hot encoded fault type. For the expected value of the sample, The true probability output by the discriminator, ranging from 0 to 1. Data samples generated by the generator.

[0036] The specific formula for the auxiliary classification loss is:

[0037] In the formula, To aid in classification loss, For cross-entropy loss, This is the prediction function for the classifier.

[0038] The composite loss function is expressed by the formula:

[0039] In the formula, For composite loss function, For adversarial loss coefficient, The auxiliary classification loss coefficient, adversarial loss coefficient, and auxiliary classification coefficient are obtained based on gradient normalization during training.

[0040] The effects of the composite loss function technique are as follows: This approach not only ensures a high degree of similarity in distribution between the generated data and real fault data, making the data augmentation results more credible, but also effectively guides the generator to produce data samples that precisely correspond to specific fault conditions by introducing classification constraints for fault types. This significantly enhances the diversity and physical meaning of the generated data. This composite loss function, as the core optimization objective, drives the generative adversarial model to converge stably during training, ultimately generating high-quality, high-fidelity simulated fault data, providing a sufficient and reliable training foundation for subsequent fault identification models.

[0041] S2, mixes real fault data, simulated fault data and augmented fault data to construct training and test sets; In this embodiment, the construction of training and testing sets by mixing real fault data, simulated fault data, and enhanced fault data specifically involves: Denoising, filling missing values, and aligning the time axis are performed on real fault data, simulated fault data, and enhanced fault data to obtain standardized real fault data, simulated fault data, and enhanced fault data. Based on domain adaptation technology, the feature distributions of the standardized simulated fault data and enhanced fault data are aligned with the standardized real fault data to obtain simulated fault aligned data and enhanced fault aligned data. A quality scoring strategy is used to filter real fault data, simulated fault-aligned data, and enhanced fault-aligned data to obtain a high-quality synthetic sample set. The high-quality synthetic sample set is used to obtain the training set and the test set according to a preset ratio.

[0042] In this embodiment, the denoising, missing value filling, and time axis alignment of real fault data, simulated fault data, and enhanced fault data specifically involve: for time-series signals such as current and voltage, signal processing methods are used to remove noise interference. First, wavelet threshold denoising is performed on the three-phase current amplitude and three-phase voltage amplitude signals to filter out high-frequency random noise. Then, low-pass filtering is used on the zero-sequence current and lightning peak signals to smooth abnormal spikes while maintaining the main waveform characteristics. When some sampling points are detected as missing, linear interpolation is used to recover the missing values ​​for the three-phase current amplitude and three-phase voltage amplitude at continuous time points. For the zero-sequence current and lightning peak, interpolation based on the time series trend is used. The method fills in the gaps by supplementing the fault type based on the consistency of labels in adjacent time periods. Considering that the sampling frequency and start time of data from different sources may be inconsistent, a unified sampling period is used as a benchmark to resample all data and align the time points. For data with a high sampling frequency, a downsampling method is used to align with the benchmark time scale. For data with a low sampling frequency, an interpolation expansion method is used to fill in the intermediate points to ensure that the three-phase current, voltage, zero-sequence current and lightning peak value correspond consistently on the same time axis. After alignment, standardized real fault data, simulated fault data and enhanced fault data are obtained, providing consistent input for subsequent domain adaptation and quality screening.

[0043] In this embodiment, the domain adaptation technique specifically involves: First, feature extraction is performed on the standardized real fault data, simulated fault data, and enhanced fault data. Three-phase current amplitude, three-phase voltage amplitude, zero-sequence current, and lightning strike peak value are used as input features, and a convolutional neural network is used to extract time-series feature representations. Second, considering the deviation between the distribution of simulated and enhanced fault data and real fault data, the maximum mean difference (MMD) method is used to calculate the mean distance between different data sources in the regenerative kernel Hilbert space to measure the distribution difference. Then, based on the above measurement results, the feature extraction network parameters are iteratively optimized so that the simulated and enhanced fault data gradually approximate the real fault data in the feature space distribution, resulting in simulated fault aligned data and enhanced fault aligned data. Finally, the domain adaptation effect is evaluated by calculating the change in MMD distance before and after alignment. When the maximum mean difference between the simulated or enhanced data and the real data distribution is lower than a preset threshold, alignment is considered complete.

[0044] The maximum mean difference (MMD) is expressed by the formula:

[0045] In the formula, This is a collection of real fault data samples. To simulate or enhance a set of fault data samples, A collection of real fault data With simulated or enhanced fault data sample set The square of the distribution difference in the feature space, The number of real fault data samples. To simulate or enhance the number of fault data samples, For the i-th sample in the real fault data, To simulate or enhance the j-th sample in the fault data, For kernel function mapping, This represents the mean vector of the actual fault data in the feature space. To simulate or enhance the mean vector of fault data in the feature space, Let be the squared Euclidean norm of the vector.

[0046] It should be noted that the preset threshold is ,when At that time, the alignment is determined to be complete.

[0047] In this embodiment, the quality scoring strategy involves introducing multi-dimensional scoring indicators, including the proportion of missing values ​​and Mahalanobis distance, with the specific formula as follows:

[0048]

[0049]

[0050] In the formula, The proportion of missing values ​​is represented by x, where x is the feature vector. The Mahalanobis distance of the feature vector x. This is the vector of the mean values ​​of the real data samples. Let i be the overall quality score for the i-th sample. and These are the weighting coefficients. This represents the maximum acceptable value for the normalized Mahalanobis distance.

[0051] In this embodiment, obtaining the training set and test set according to the preset ratio means using 80% of the high-quality synthetic sample set as the training set and the remaining 20% ​​of the high-quality synthetic sample set as the test set.

[0052] S3, by optimizing a pre-defined deep learning network and attention network containing a fault loss function through the training set, a power distribution network fault identification model is obtained; In this embodiment, the process of optimizing a pre-defined deep learning network and attention network containing a fault loss function using a training set to obtain a distribution network fault identification model specifically involves: Input the training set, extract spatial-temporal features based on convolutional neural network features; Spatial-temporal features are weighted and processed through a channel attention network to output refined features along the channel dimension. The refined features are then recalibrated using a temporal attention network to obtain temporal focused features; A multi-task focusing loss function is designed, and the convolutional neural network, channel attention network, and temporal attention network are optimized based on the aforementioned temporal focusing features to obtain a power distribution network fault identification model.

[0053] It should be noted that the convolutional neural network extracts spatial-temporal features from the multi-channel current, voltage, and zero-sequence current amplitude sequences input to the training set. Its structure includes multiple one-dimensional and two-dimensional convolutional layers, capable of capturing local patterns and correlations between different channels and in the temporal dimension. The channel attention network globally aggregates each channel in both the temporal and feature dimensions to generate a global channel representation. This representation is then processed by a feedforward network and non-linear activation to generate channel attention weights. These weights are multiplied with the original features channel by channel to obtain channel-weighted refined features. The temporal attention network recalibrates these refined features in the temporal dimension. It aggregates the channel and feature dimensions at each time step to generate a global representation for that time step. This representation is then processed by a feedforward network or attention mechanism to generate temporal attention weights, which are then multiplied with the features sequentially in the temporal dimension to obtain temporally focused features targeting key time periods.

[0054] Through convolution operations, the network can automatically learn the current and voltage variation patterns, the coupling relationship between the three phases, and abnormal features such as lightning strikes from the original amplitude data, providing rich spatial and temporal feature representations for the subsequent attention network. The main function of the channel attention network is to identify the importance of each channel in different time periods, highlight the channel features that play a key role in fault identification, and suppress redundant or noisy information. By learning the channel attention weights, the network can generate refined feature representations, making the model more discriminative and discriminative in the channel dimension. The temporal attention network can focus on the most critical time period for fault identification, such as lightning transients and fault abrupt changes, generating time-domain focused features. This processing not only improves the model's sensitivity to key time points, but also enhances its ability to capture transient information in non-stationary power signals.

[0055] In this embodiment, the multi-task focusing loss function includes classification loss, center loss, and physical consistency loss.

[0056] In this embodiment, the multi-task focusing loss function is specifically formulated as follows: Classification loss formula:

[0057] In the formula, The loss function is for classification, where N is the number of training samples and C is the number of fault categories. For the true category indicator of sample i, if the category belongs to c, It is 1 if it is true, otherwise it is 0. Predict the probability that sample i belongs to class c for the model.

[0058] Central loss:

[0059] In the formula, Loss at the center Let i be the time-domain focused feature vector. For category The feature center vector, It is a Euclidean norm.

[0060] Physical consistency loss:

[0061] In the formula, For physical consistency loss, and These are weighting coefficients. Let i be the three-phase voltage imbalance. Let represent the three-phase current imbalance of sample i.

[0062] The specific formulas for three-phase voltage unbalance and three-phase current unbalance are as follows:

[0063]

[0064] In the formula, This refers to the three-phase voltage imbalance. , and This refers to the three-phase voltage amplitude. Let be the average of the three-phase voltages, and max be the maximum value. This refers to the three-phase current imbalance. , and This refers to the amplitude of the three-phase current. This is the average value of the three-phase current.

[0065] Multi-task focusing loss formula:

[0066] In the formula, To focus on loss in a multi-tasking manner, , and These are the weighting coefficients for classification loss, center loss, and physical consistency loss, respectively.

[0067] Classification loss measures the difference between the probability distribution of fault categories predicted by the model and the actual fault labels. By minimizing classification loss, the model can correctly identify different types of distribution network faults, improve the accuracy of fault identification, and ensure that the model has a high prediction accuracy for various faults. It is the core constraint for achieving accurate classification of distribution network faults. Centralization loss constrains samples of the same category to cluster near the corresponding category center in the feature space, while widening the distance between samples of different categories, enhancing feature discriminability, making the features of samples of the same category tightly clustered, and making the differences between the features of samples of different categories obvious, thereby improving the model's ability to distinguish similar fault types. Physical consistency loss constrains the model output to conform to the physical laws of the distribution network, ensuring that the model prediction results are not only statistically accurate, but also conform to the actual operating constraints of the power system, improving the reliability and robustness of the model.

[0068] S4 uses the test set to evaluate the power distribution network fault identification model and obtain fault type identification results.

[0069] In this embodiment, the step of using a test set to evaluate the distribution network fault identification model and obtain fault type identification results specifically involves: Based on the power distribution network fault identification model, the test set data is predicted to obtain the predicted label of the fault type. Based on the predicted labels and the corresponding real labels in the test set, calculate the overall accuracy, precision, recall, and F1 score for the fault type; The overall accuracy, precision, recall, and F1 score are averaged to obtain the fault type identification result.

[0070] In this embodiment, the overall accuracy, precision, recall, and F1 score of the fault type are calculated using TP, FP, FN, and TN. TP is the number of samples correctly identified as a fault of a certain type; FP is the number of samples incorrectly identified as faults of other types; FN is the number of samples incorrectly identified as faults of this type as other types; and TN is the number of samples correctly identified as faults of a different type. The specific formula is as follows:

[0071] In the formula, Accuracy represents the overall accuracy.

[0072]

[0073] In the formula, Precision represents the accuracy rate.

[0074]

[0075] In the formula, Recall is the recall rate.

[0076]

[0077] In the formula, F1 is the F1 score.

[0078] It should be noted that the model compares the output predicted labels with the real fault labels in the test set to form a confusion matrix. Based on the confusion matrix, the overall accuracy, precision, recall, and F1 score are calculated. Finally, the average of the above indicators is calculated. If the average value is greater than 0.85, the model is considered to have completed training.

[0079] In practical power distribution network fault identification scenarios, the trained power distribution network fault identification model is deployed. First, multi-channel current, voltage, and zero-sequence current amplitude sequences are acquired in real time through data acquisition devices during the operation of the power distribution network. This data serves as the input to the model. The model uses convolutional neural network features to extract features from the input data, automatically learning the current and voltage variation patterns, the coupling relationships between the three phases, and abnormal features such as lightning strikes, thus obtaining spatial-temporal features.

[0080] Next, the spatial-temporal features are input into the channel attention network, which globally aggregates each channel in the time and feature dimensions to generate a global channel representation. The representation is then fed forward and nonlinearly activated to generate channel attention weights. These weights are multiplied by the original features channel by channel to obtain channel-weighted refined features.

[0081] Then, the refined features are input into the temporal attention network, which recalibrates the refined features in the time dimension. By aggregating the channel and feature dimensions of each time step, a global representation of the time step is generated. Then, a feedforward network or attention mechanism is used to generate temporal attention weights, which are then multiplied with the features in the time dimension to obtain temporal focused features that focus on key time periods.

[0082] Finally, based on temporal focusing features, the model utilizes the parameters optimized during training and outputs the predicted fault type through a convolutional neural network, a channel attention network, and a temporal attention network constrained by a multi-task focusing loss function.

[0083] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0084] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0085] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0086] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0087] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0088] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying distribution network faults combining deep learning and attention, characterized in that, Includes the following steps: Collect real fault data, combine it with a simulation platform to generate simulated fault data, and then perform data augmentation on the real fault data; By mixing real fault data, simulated fault data, and augmented fault data, a deep learning network and an attention network are trained to obtain a power distribution network fault identification model. The electrical data is input into the power distribution network fault identification model to obtain the fault type identification results; The data augmentation is achieved by training an adversarial model with optimized composite loss; the composite loss is obtained by weighted summation of adversarial loss and auxiliary classification loss, aiming to simultaneously constrain the fidelity of the data distribution of the generated samples and the accuracy of fault type classification.

2. The power distribution network fault identification method combining deep learning and attention as described in claim 1, characterized in that, The real fault data, simulated fault data, and enhanced fault data include time points, three-phase current amplitudes, three-phase voltage amplitudes, zero-sequence current, lightning peak values, and fault types.

3. The power distribution network fault identification method combining deep learning and attention as described in claim 2, characterized in that, The fault types include single-phase grounding, two-phase short circuit, two-phase grounding, three-phase short circuit, and open circuit fault.

4. The power distribution network fault identification method combining deep learning and attention as described in claim 3, characterized in that, The data enhancement specifically includes: Input the random noise vector, fault type label, and lightning characteristic parameters into the condition generator to obtain the generated fault data; The real fault data and the generated fault data are input into the discriminator, which outputs the discrimination result and the fault type classification result. A generative adversarial model is constructed by calculating the composite loss of the discrimination result and the fault type classification result, and data augmentation is performed on the real fault data. The composite loss is obtained by weighted summation of adversarial loss and auxiliary classification loss. The adversarial loss is obtained by measuring the distribution difference between generated samples and real samples, and the auxiliary classification loss is obtained by measuring the consistency between the classifier prediction and the real fault type using cross-entropy.

5. The power distribution network fault identification method combining deep learning and attention as described in claim 4, characterized in that, The hybrid real fault data, simulated fault data, and enhanced fault data include aligned simulated fault data and enhanced fault data, specifically: The real, simulated, and enhanced fault data are mapped to the same feature space through a feature extraction network. The parameters of the feature extraction network are optimized by calculating and minimizing the maximum mean difference between simulated / augmented data and real data in the feature space. By using an optimized network, simulated and augmented fault data are transformed into aligned data that approximates real fault data in terms of feature distribution.

6. The power distribution network fault identification method combining deep learning and attention as described in claim 5, characterized in that, The training deep learning network and attention network include a multi-task focusing loss function, which is obtained by weighted summation of classification loss, center loss and physical consistency loss; The physical consistency loss is obtained by weighted summation of the three-phase voltage and current imbalance.

7. The power distribution network fault identification method combining deep learning and attention as described in claim 6, characterized in that, The training of the deep learning network and attention network also includes a model testing step, specifically: Based on the power distribution network fault identification model, the test data is predicted to obtain the predicted label of the fault type; Based on the predicted labels and the corresponding real labels, calculate the accuracy, precision, recall and F1 score of the fault type; The average of the accuracy, precision, recall, and F1 score is used as the model test result.

Images