Power system fault positioning method, device and system based on BiGRU-GCN structure and medium

Abstract

The invention discloses a BiGRU-GCN structure-based power system fault positioning method, device and system and a medium. The method comprises the steps of constructing a power system power distribution network model comprising a high-proportion inverter; preprocessing simulation data of the power distribution network model of the power system by adopting a sliding window technology; performing multi-scale feature extraction on the preprocessed data by using an encoder to obtain local features and global features; performing feature fusion and fault position reduction on the local features and the global features by using a decoder; and in combination with fault feature distribution and confidence degree evaluation output by the decoder, the fault occurrence position and confidence degree information are determined, sorting is carried out according to confidence degrees, and the position with the highest confidence degree is found out and is the fault position. According to the method, the advantages of the neural network in the aspect of data processing and the advantages of a specific model solving method are fully utilized, and the problem that after a high-proportion inverter is connected to a power distribution network, a fault section is difficult to accurately locate through traditional relay protection is solved.

Description

Technical Field

[0001] The present invention relates to the field of power system monitoring and data analysis, and in particular to a power system fault location method, device, system and medium based on a BiGRU-GCN structure. Background Art

[0002] In modern power systems, with the large-scale access of renewable energy, especially the access of a high proportion of inverters to distribution networks, many new challenges have been brought to the operation and protection of power systems.

[0003] The integration of a high-ratio inverter into the distribution network significantly alters its operating characteristics. The current output by the inverter has characteristics different from those of a traditional synchronous generator, and its waveform, frequency, and phase are influenced by the inverter's control strategy. When a fault occurs, the magnitude, direction, and duration of the fault current differ from those in traditional grid faults due to the inverter's control mechanism.

[0004] Traditional relay protection methods primarily determine fault location by comparing the forward and reverse directions of the fault current. However, in distribution networks with a high proportion of inverters, this relay protection method, which relies on forward and reverse current comparisons, faces significant challenges. First, the inverter's ability to rapidly adjust its output current renders the direction of the fault current less regular than in traditional power grids, making it difficult for traditional relay protection devices to accurately determine the direction of the fault current. Second, the fault current output by the inverter may contain significant harmonic components, which can interfere with the determination of the forward and reverse directions of the current, further reducing the reliability and accuracy of traditional relay protection.

[0005] With the increasing complexity of power systems and the new challenges posed by the high proportion of inverters installed, traditional fault diagnosis and protection methods are no longer able to meet practical needs. Deep learning technology, however, has shown great potential in processing complex data and complex system problems. Neural networks, with their powerful self-learning and adaptive capabilities, can extract valuable features and patterns from large amounts of complex data.

[0006] Graph convolutional networks (GCNs) can effectively utilize the topological structure of distribution networks, which is crucial for processing distribution network data with network-specific characteristics. By performing convolution operations on the distribution network graph, GCNs can fully exploit the connectivity between nodes and information related to fault propagation paths.

[0007] The encoder-decoder architecture based on the self-attention mechanism also offers unique advantages in processing time series data in power systems. In distribution networks with a high proportion of inverters, the time series characteristics of operating data such as voltage and current are more complex. The model can adaptively extract features based on the relationships between data elements, facilitating analysis of how these time series data change during faults. Summary of the Invention

[0008] In order to solve the above problems, the present invention proposes a power system fault location method, device, system and medium based on the BiGRU-GCN structure, aiming to fully utilize the advantages of neural networks in data processing and the advantages of specific model solving methods to solve the problem that traditional relay protection is difficult to accurately locate the fault section after a high proportion of inverters are connected to the distribution network.

[0009] A power system fault location method based on a BiGRU-GCN structure is performed based on a power system distribution network fault location neural network. The neural network includes an encoder and a decoder. The encoder includes a self-attention mechanism, a bidirectional gated recurrent unit (BiGRU), a convolutional neural network (CNN), and a residual block. The decoder includes a multi-layer graph convolutional network (GCN). The method includes the following steps:

[0010] Step 1: Construct a power system distribution network model including a high-ratio inverter;

[0011] Step 2: Use sliding window technology to preprocess the simulation data of the power system distribution network model;

[0012] Step 3: Based on the preprocessed data, use the encoder to extract multi-scale features of the data to obtain local features and global features;

[0013] Step 4: Use the decoder to fuse local features and global features and restore the fault location

[0014] The decoder uses a multi-layer graph convolutional network (GCN) to perform a dual-path deep fusion of the local and global features output by the encoder. This method restores local features in a fine-grained manner while integrating global features to capture the overall fault characteristics. The fused dual-path decoded data is then optimized through a neural network module to generate an accurate fault feature distribution, from which the confidence distribution of each node in the power system is calculated.

[0015] Step 5: Locate the fault area and output the results

[0016] Combine the fault feature distribution and confidence assessment output by the decoder to determine the location of the fault and confidence information, sort them according to the confidence level, and find the location with the highest confidence, which is the fault location.

[0017] Furthermore, step one of constructing a power system distribution network model including a high proportion of inverters includes: building a simulated power system model including a high proportion of inverters to truly simulate the actual distribution network operating environment; the model contains 33 nodes, and the data of each node includes phase voltage, active power and reactive power.

[0018] Furthermore, step 2 uses sliding window technology to pre-process the simulation data of the power system distribution network model, specifically including: performing sliding window segmentation on the power system operation data generated by simulation, and dividing the data blocks according to time series.

[0019] Furthermore, in step three, based on the preprocessed data, an encoder is used to extract multi-scale features from the data to obtain local features and global features, including: using a self-attention mechanism: adjusting the importance distribution of features in the input data, adaptively sensing the significance of multi-scale features, and prioritizing the extraction of key features related to the fault; extracting local features of the data through a convolutional neural network (CNN) to generate multi-scale enhanced features; using a bidirectional gated recurrent unit (BiGRU) to further mine the global temporal features of the data, capture the correlation before and after the fault occurs, and form a more comprehensive global feature.

[0020] A power system fault location device based on a BiGRU-GCN structure, comprising:

[0021] A distribution network model building module, used to build a power system distribution network model including a high-ratio inverter;

[0022] A data preprocessing module is used to preprocess the simulation data of the power system distribution network model using a sliding window technology;

[0023] The encoder is used to extract multi-scale features from the preprocessed data to obtain local features and global features;

[0024] The decoder uses a multi-layer graph convolutional network (GCN) to perform a dual-path deep fusion of the local and global features output by the encoder. This allows for fine-grained restoration of local features while integrating global features to capture overall fault characteristics. The fused dual-path decoded data is then optimized through a neural network module to generate an accurate fault feature distribution, from which the confidence distribution of each node in the power system is calculated.

[0025] The fault location module is used to combine the fault feature distribution and confidence assessment output by the decoder to determine the location of the fault and the confidence information, sort them according to the confidence level, and find the location with the highest confidence, which is the fault location.

[0026] Furthermore, the distribution network model construction module is specifically used to build a simulated power system model, including a high proportion of inverters, to truly simulate the actual distribution network operating environment; the model contains 33 nodes, and the data of each node includes phase voltage, active power and reactive power.

[0027] Furthermore, the data preprocessing module is specifically used to preprocess the simulation data of the power system distribution network model using a sliding window technology, specifically including: performing sliding window segmentation on the power system operation data generated by simulation, and dividing the data blocks according to time series.

[0028] Furthermore, the encoder is specifically used to perform multi-scale feature extraction on the data based on the preprocessed data to obtain local features and global features, including: using a self-attention mechanism: adjusting the importance distribution of features in the input data, adaptively sensing the significance of multi-scale features, and prioritizing the extraction of key features related to the fault; extracting local features of the data through a convolutional neural network (CNN) to generate multi-scale enhanced features; using a bidirectional gated recurrent unit (BiGRU) to further mine the global temporal features of the data, capture the correlation before and after the fault occurs, and form a more comprehensive global feature.

[0029] A power system fault location system based on a BiGRU-GCN structure, comprising: a computer-readable storage medium and a processor;

[0030] The computer-readable storage medium is used to store executable instructions;

[0031] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the power system fault location method based on the BiGRU-GCN structure.

[0032] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements the power system fault location method based on the BiGRU-GCN structure.

[0033] The present invention uses a data-driven neural network to partially mine the features in the distribution network operation data, achieving accurate and rapid positioning of fault sections where a high proportion of inverters are connected to the distribution network, thereby ensuring the safe and stable operation of the distribution network. The invention also uses artificial intelligence algorithms based on the original FTU configuration measurement data of the distribution network, and obtains a model through big data training, achieving a fault location accuracy of 99 percent. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of the neural network diagnostic framework for fault location in the power system distribution network of the present invention;

[0035] Figure 2This is a schematic diagram of the structure of an improved IEEE-33 node model according to an embodiment of the present invention;

[0036] Figure 3 This is a diagram of the confusion matrix of the neural network test results. DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0038] like Figure 1 As shown in the figure, the present invention proposes a new neural network diagnostic framework for fault location in power distribution networks. The structural design of the neural network combines the self-attention mechanism, bidirectional gated recurrent unit BiGRU, convolutional neural network CNN and residual block composed of encoder, and multi-layer graph convolutional network GCN composed of decoder, so as to efficiently process the complex relationship of power system data. The encoder first extracts local time features through CNN, then uses bidirectional GRU (BiGRU) to capture the front and back dependencies of the time series, and dynamically focuses on key features through the self-attention mechanism. The introduction of residual blocks effectively alleviates the gradient vanishing problem caused by network deepening, while improving the stability of training and the generalization ability of the model.

[0039] In the decoder, the model uses a multi-layer GCN to model global topological features. The adjacency matrix describes the network structure and analyzes the coupling relationships between nodes, such as the mutual influence between voltages, currents, and active and reactive power at adjacent nodes. Compared to traditional tuning methods based on mathematical modeling, this design eliminates the complexity of manual formula derivation and directly utilizes GCN to aggregate neighborhood information, automatically learning and characterizing the coupling relationships between nodes from the data. This significantly improves the accuracy and efficiency of modeling the structural characteristics of the power system.

[0040] Each part will be described in detail below.

[0041] 1. Encoder part

[0042] Self-attention mechanism: The self-attention mechanism captures long-range dependencies in the input data by weightedly aggregating information about each element in the input sequence. In this paper, the self-attention mechanism uses multi-head self-attention to simultaneously calculate different attention scores and aggregate them to capture multiple relationships among input features. The relationship between each feature and other features is dynamically adjusted during this process, enhancing the model's ability to model global dependencies, especially when processing time series data or long-range dependencies.

[0043] The self-attention formula is as follows:

[0044]

[0045] Q is the query matrix;

[0046] K is the bond matrix;

[0047] V is the value matrix;

[0048] d k is the dimension of the key.

[0049] Convolutional Neural Network (CNN) Layer: The convolution layer is used to extract local features and perform hierarchical processing on the input data. By stacking multiple convolution layers, more abstract high-dimensional features can be gradually extracted from the original input. The CNN layer performs a convolution operation on the input through a filter, effectively capturing local information in the spatial structure, and performs dimensionality reduction on the input to improve computational efficiency. The convolution operation can capture the local pattern of the input data and is particularly suitable for signal processing and image processing tasks. In the present invention, the CNN layer is designed as a multi-layer stacked structure, combined with the maximum pooling operation, which can enhance the feature extraction capability and reduce overfitting.

[0050] The formula for the CNN layer is as follows:

[0051]

[0052] x is the input image or feature map,

[0053] ω is the convolution kernel,

[0054] y is the output after the convolution operation.

[0055] Bidirectional Gated Recurrent Unit (BiGRU) layer: The bidirectional GRU layer enhances the model's learning ability for time series data by combining forward and backward time series information. A GRU (Gated Recurrent Unit) is a recurrent neural network (RNN) with a gating mechanism that effectively captures long-term dependencies in time series data by controlling the flow of information. The bidirectional GRU leverages both forward and backward time information, allowing the network to simultaneously process past and future contextual information, thereby improving the model's understanding of temporal context. This bidirectional information flow is crucial for modeling time series data, especially in complex domains such as power systems.

[0056] The formula for GRU is as follows:

[0057] (1) Reset gate

[0058] The reset gate calculation formula is as follows:

[0059] r t =σ(W r ·[h t-1 ,x t ])

[0060] Where r t is the reset gate at time t; σ is the sigmoid function, which converts the data into a value in the range of [0,1], thus acting as a gating signal; W r To control the degree of retention of input information at each position at time t-1.

[0061] The candidate hidden layer states for the reset gate are:

[0062]

[0063] When r t When it approaches zero, the model will hide the past information h t-1 Discard, leaving only the current input information. t When it approaches 1, it is considered that the past hidden information is useful and is added to the current information.

[0064] Mainly includes input information x at time t t and the hidden layer h at time t-1 t-1 Retention information.

[0065] (2) Update Gate

[0066] The update gate calculation formula is as follows:

[0067] z t =σ(W z ·[h t-1 ,x t ])

[0068] Where z t is the update gate at time t; W z Used to control the degree of retention of new and old information input at time t.

[0069] The final hidden state of the updated gate is:

[0070]

[0071] z t The range is [0,1]. The closer the gate signal is to 1, the more past data is remembered; and the closer it is to 0, the more current data is retained.

[0072] In bidirectional GRU, there are two GRUs, one processing sequence data from left to right and the other from right to left.

[0073]

[0074] is the output of the forward GRU at time t,

[0075] is the output of the reverse GRU at time t.

[0076] Projection layer (not marked in the figure, located in the last step of BiGRU in the encoder, its function is dimensional compatibility, mainly for dimensional transformation and maintaining dimensional matching): The function of the projection layer is to map the output of the bidirectional GRU layer back to the target feature dimension of the model so as to be compatible with subsequent layers. In the present invention, the projection layer converts the high-dimensional features output by BiGRU into the dimensions required by the model, thereby ensuring the smooth progress of subsequent operations. The projection operation maps the features through linear transformation, which not only maintains the richness of the input information, but also avoids information loss, ensuring that the model has sufficient expressive power when processing subsequent tasks.

[0077] The residual block is a commonly used structure in deep learning models (the residual parameters of the present invention connect the original input and the output of the encoder). Its core idea is to alleviate the common gradient vanishing or gradient exploding problems in deep neural networks by introducing a shortcut connection. In the present invention, the residual block helps the network to learn effective feature representations more easily and speeds up the training process by directly adding the input features to the output features. The role of the residual block is to allow the network to skip certain layers during forward propagation, so that information can flow more directly to the next layer. This direct information flow method helps the model maintain more stable performance during training, especially in very deep network structures. Through residual connections, the network can retain more original information, thereby reducing the performance degradation when the network depth increases. In the present invention, the use of residual blocks improves the training efficiency of the model and helps the model maintain efficient learning ability with increasing number of layers. The residual connection makes the function of each layer more independent, and the gradient of the network can be directly transmitted through this path, thereby effectively avoiding the occurrence of network overfitting and gradient disappearance problems.

[0078] Layer normalization is a normalization technique commonly used in neural networks to improve model stability and accelerate the training process. Unlike batch normalization, layer normalization performs normalization on the feature dimension of a single sample, rather than on the batch dimension. This makes layer normalization particularly suitable for processing sequence data with variable input lengths or recurrent neural networks (RNN) tasks. In the present invention, layer normalization is applied to the output of each layer of the network (CNN, BiGRU, and after the self-attention mechanism) to ensure that the feature distribution of each layer remains stable. By normalizing the output of each layer, layer normalization reduces the deviation in feature distribution between different layers during training, thereby increasing the convergence rate of training. Layer normalization helps the model avoid training instability caused by uneven feature distribution, and by maintaining the mean of each layer output to 0 and the variance to 1, it further enhances the performance of the network. The main function of layer normalization is to improve the training efficiency and stability of the network, especially when using deep networks, it can effectively reduce the impact of changes in data distribution. Layer normalization also helps the network maintain high adaptability to features when facing different types of input data, ensuring the robustness of the model.

[0079] 2. Decoder part

[0080] Graph Convolutional Network (GCN) layer: The GCN layer is used to process graph structured data and aggregate the relationships between nodes through the adjacency matrix. In the present invention, GCN is used to process complex relationships such as voltage, current, active and reactive power between nodes in the power system. GCN transmits information between nodes through the adjacency matrix, allowing nodes to update their own features based on the information of their neighbors. Unlike traditional mathematical methods based on physical models, GCN can learn the coupling relationship between nodes without explicit modeling by aggregating neighbor node information, which gives it a unique advantage in processing complex structured data. By stacking multiple GCN layers, the model can capture deeper graph structure information and further improve the expressive power of the model.

[0081] The layer normalization formula is as follows:

[0082]

[0083] H (l) is the node feature matrix of the lth layer, H (0) is the initial input feature,

[0084] is the normalized adjacency matrix (A is the adjacency matrix, D is the degree matrix),

[0085] W (l) is the weight matrix of layer l,

[0086] σ is the ReLU activation function.

[0087] 3. Output layer

[0088] Output layer: After the decoder is processed by multiple layers of GCN, the output layer transforms the final features for use in specific downstream tasks. The output layer completes the model's prediction task by mapping the processed features to the target category or output dimension of the regression task. In the case of fault classification, the output layer is responsible for converting the decoder features into class labels that represent potential fault types in the power system.

[0089] The following example introduces a new power system fault location method based on CNN-BiGRU-GCN.

[0090] First, build a distribution network model. This case is based on the improved IEEE-33 node model. Figure 2 As shown in the figure, this case includes a high-ratio inverter in the new distribution network. To simulate the power flow and current magnitude of the fault while avoiding the problem of slow simulation data generation, the high-ratio inverter used in this example is replaced with a limited power source, which can obtain the same power flow and current trend data.

[0091] Power system relay protection incorporates four principles: reliability, selectivity, speed, and sensitivity. In power system distribution networks, reliability and selectivity are particularly challenging to achieve with a high proportion of inverters. In distribution networks with a high proportion of inverters and distributed generation, reverse current issues can complicate relay protection settings. The specific issues can be analyzed as follows:

[0092] 1. Current direction variability: Due to the widespread integration of distributed power sources (such as photovoltaic, wind power, and energy storage systems), the current direction during a fault may be opposite to that under traditional single-source power supply. The presence of reverse current makes it difficult for traditional directional protection devices to accurately determine the fault direction, complicating their setting.

[0093] 2. Diversified power sources and uncertainty in fault current contributions: Distributed power sources come in a variety of configurations, including synchronous generators and inverters. Inverter-based power sources, in particular, have limited fault current output and exhibit significant dynamic characteristics, making it difficult to provide sufficient short-circuit current for protection device setting. Furthermore, the uncertainty in the current contribution of different power sources during a fault makes it difficult to set protection settings across the entire protection range.

[0094] 3. Insufficient protection coverage, unable to protect the entire line: Traditional overcurrent protection schemes rely on distinct short-circuit current characteristics and fixed setting values. However, in the case of multiple power sources connected to the grid, the uneven distribution of fault currents makes it difficult to ensure the sensitivity of the protection device. This can lead to protection failure on some lines and failure to provide full-length protection.

[0095] 4. Problems of misoperation and refusal to operate

[0096] Misoperation: In a distribution network with multiple power sources connected, the protection device in the non-fault area may misjudge the fault due to the reverse current or harmonic components of other power sources, triggering false operation; Refusal to operate: When a fault occurs, the protection device may not be able to quickly and accurately distinguish between normal load current and fault current, especially when the fault current amplitude is small or the dynamic characteristics change drastically, resulting in refusal to operate.

[0097] To increase the model's complexity and verify its ability to address the previously mentioned issues, this model adopts the sampling rate of a real FTU, sampling 24 points per cycle at 50 Hz. In Simulink, the sampling rate is adjusted from 1% to 99%. A three-phase short-circuit fault is simulated for 3 seconds per fault. The resulting CSV file contains 165 columns (Phase A voltage, Phase B voltage, Phase C voltage, and PQ for 33 nodes).

[0098] Divide the data set. Since there are many rows of data, in the present invention, in order to improve the generalization ability of the model and the learning effect of time series data, the sliding window technology is used to pre-process the power system simulation data. Specifically, by setting the sliding window parameters to a window size of 24 and a step size of 12, the 165-dimensional data sampled per cycle is divided into sliding windows, where each window contains data from 24 consecutive time points, and there is an overlap of 12 data points between adjacent windows. This sliding window technology effectively increases the diversity of the data, allowing the model to capture the time dependency and change trend of the data during training, thereby improving the ability to identify complex fault modes of the power system. In addition, the sliding window method also reduces the memory overflow problem that may be caused by excessive data volume while ensuring the integrity of the data time series structure, providing more abundant and efficient data input for subsequent model training.

[0099] The above model was used for training. The training environment was Windows system, Intel Core i9-13950HX CPU, Nividia4090Laptop GPU, 128g memory, pytorch version 1.12.1, CUDA version 11.3, 400 training arguments, and the following results were obtained in the validation set, as shown below: Figure 3 shown.

[0100] In a comparison of the present invention with other baseline methods (as shown in Table 1), the results further validate the significant advantages of the proposed method. The precision, recall, and F1 scores of the present invention were 99.08%, 99.06%, and 99.06%, respectively, significantly outperforming the baseline methods BiGRU+GCN (97.83%, 97.77%, and 97.77%) and GCN+Transformer (90.06%, 89.06%, and 86.99%). These comparative results demonstrate that the present invention outperforms existing methods in the task of power system fault detection.

[0101] Table 1

[0102] Accuracy Recall F1 score The method proposed by the present invention 99.08 99.06 99.06 BiGRU+GCN 97.83 97.77 97.77 GCN+Transformer 90.06 89.。06 86.99

[0103] The high performance of the present invention is due to many aspects of technological innovation. First, the sliding window data enhancement technology effectively expands the diversity of training data, especially when dealing with the complexity of high-frequency sampling power system signals, it provides an efficient data processing mechanism. Secondly, the bidirectional GRU and self-attention mechanism in the encoder-decoder structure not only capture the long-range dependencies of time series data, but also focus on key patterns through dynamic feature weighting, thereby improving the model's adaptability to time-varying signals. In addition, the multi-layer GCN in the decoder models the coupling relationship between nodes through the adjacency matrix, which can automatically learn the complex interactions between different nodes, thereby showing excellent results in the spatial distribution modeling of power systems.

[0104] This paper also builds a highly realistic simulation environment, generating real-world power system signals in Simulink at a 50Hz frequency and 24 points / cycle sampling rate, covering simulated scenarios with fault depths ranging from 1% to 99%. This rigorous testing environment ensures the model's applicability and reliability in real-world scenarios, further demonstrating the engineering value of this invention.

[0105] In summary, this invention, through comprehensive technological innovation and rigorous testing and verification, demonstrates its leading position in the field of power system fault detection. Compared with existing methods, this invention significantly improves the accuracy and efficiency of fault detection.

[0106] The embodiment of the present invention further provides a power system fault location device based on a BiGRU-GCN structure, comprising:

[0107] A distribution network model building module, used to build a power system distribution network model including a high-ratio inverter;

[0108] A data preprocessing module is used to preprocess the simulation data of the power system distribution network model using a sliding window technology;

[0109] The encoder is used to extract multi-scale features from the preprocessed data to obtain local features and global features;

[0110] The decoder uses a multi-layer graph convolutional network (GCN) to perform a dual-path deep fusion of the local and global features output by the encoder. This allows for fine-grained restoration of local features while integrating global features to capture overall fault characteristics. The fused dual-path decoded data is then optimized through a neural network module to generate an accurate fault feature distribution, from which the confidence distribution of each node in the power system is calculated.

[0111] The fault location module is used to combine the fault feature distribution and confidence assessment output by the decoder to determine the location of the fault and the confidence information, sort them according to the confidence level, and find the location with the highest confidence, which is the fault location.

[0112] Another embodiment of the present invention provides a power system fault location system based on a BiGRU-GCN structure, comprising: a computer-readable storage medium and a processor;

[0113] The computer-readable storage medium is used to store executable instructions;

[0114] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the power system fault location method based on the BiGRU-GCN structure.

[0115] Another embodiment of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the power system fault location method based on the BiGRU-GCN structure is implemented.

[0116] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0117] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0118] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0120] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, persons skilled in the art will understand that modifications or equivalent substitutions may be made to the specific embodiments of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention are intended to be covered by the claims.

Claims

1. A power system fault location method based on BiGRU-GCN structure, characterized in that: The method is based on a power system distribution network fault location neural network, wherein the neural network includes an encoder and a decoder. The encoder includes a self-attention mechanism, a bidirectional gated recurrent unit (BiGRU), a convolutional neural network (CNN), and a residual block. The decoder includes a multi-layer graph convolutional network (GCN). The method includes the following steps: Step 1: Construct a power system distribution network model including a high-ratio inverter; Step 2: Use sliding window technology to preprocess the simulation data of the power system distribution network model; Step 3: Based on the preprocessed data, use the encoder to extract multi-scale features of the data to obtain local features and global features; Step 4: Use the decoder to fuse local and global features and restore the fault location. The decoder uses a multi-layer graph convolutional network (GCN) to perform dual-path deep fusion of the local and global features output by the encoder: on the one hand, local features are restored in a fine-grained manner, and on the other hand, global features are integrated to capture the overall fault characteristics. The fused dual-path decoded data is optimized by the neural network module to generate an accurate fault feature distribution, and the confidence distribution of each node in the power system is calculated from it. Step 5: Locate the fault area and output the results Combine the fault feature distribution and confidence assessment output by the decoder to determine the location of the fault and confidence information, sort them according to the confidence level, and find the location with the highest confidence, which is the fault location.

2. The power system fault location method based on the BiGRU-GCN structure according to claim 1, characterized in that: Step 1: Constructing a power system distribution network model including a high proportion of inverters includes: building a simulated power system model including a high proportion of inverters to truly simulate the actual distribution network operating environment; the model contains 33 nodes, and the data of each node includes phase voltage, active power and reactive power.

3. The power system fault location method based on the BiGRU-GCN structure according to claim 1, characterized in that: Step 2 uses sliding window technology to preprocess the simulation data of the power system distribution network model, specifically including: performing sliding window segmentation on the power system operation data generated by simulation and dividing the data blocks according to time series.

4. The power system fault location method based on the BiGRU-GCN structure according to claim 1, characterized in that: In step three, based on the preprocessed data, an encoder is used to extract multi-scale features from the data to obtain local features and global features, including: using a self-attention mechanism to adjust the importance distribution of features in the input data, adaptively sense the significance of multi-scale features, and prioritize the extraction of key features related to the fault; extracting local features of the data through a convolutional neural network (CNN) to generate multi-scale enhanced features; and using a bidirectional gated recurrent unit (BiGRU) to further explore the global temporal features of the data, capture the correlation before and after the fault occurs, and form a more comprehensive global feature.

5. A power system fault location device based on BiGRU-GCN structure, characterized in that: include: A distribution network model building module, used to build a power system distribution network model including a high-ratio inverter; A data preprocessing module is used to preprocess the simulation data of the power system distribution network model using a sliding window technology; The encoder is used to extract multi-scale features from the preprocessed data to obtain local features and global features; The decoder uses a multi-layer graph convolutional network (GCN) to perform a dual-path deep fusion of the local and global features output by the encoder. This allows for fine-grained restoration of local features while integrating global features to capture overall fault characteristics. The fused dual-path decoded data is then optimized through a neural network module to generate an accurate fault feature distribution, from which the confidence distribution of each node in the power system is calculated. The fault location module is used to combine the fault feature distribution and confidence assessment output by the decoder to determine the location of the fault and the confidence information, sort them according to the confidence level, and find the location with the highest confidence, which is the fault location.

6. The power system fault location device based on the BiGRU-GCN structure according to claim 5, characterized in that: The distribution network model construction module is specifically used to build a simulated power system model, including a high proportion of inverters, to truly simulate the actual distribution network operating environment; the model contains 33 nodes, and the data of each node includes phase voltage, active power and reactive power.

7. The power system fault location device based on the BiGRU-GCN structure according to claim 5, characterized in that: The data preprocessing module is specifically used to preprocess the simulation data of the power system distribution network model using the sliding window technology, specifically including: performing sliding window segmentation on the power system operation data generated by simulation and dividing the data blocks according to time series.

8. The power system fault location device based on the BiGRU-GCN structure according to claim 5, characterized in that: The encoder is specifically used to extract multi-scale features from preprocessed data to obtain local and global features, including: using a self-attention mechanism to adjust the importance distribution of features in the input data, adaptively sense the significance of multi-scale features, and prioritize the extraction of key features related to faults; extracting local features of data through a convolutional neural network (CNN) to generate multi-scale enhanced features; and using a bidirectional gated recurrent unit (BiGRU) to further explore the global temporal features of the data, capture the correlation before and after the fault occurs, and form more comprehensive global features.

9. A power system fault location system based on a BiGRU-GCN structure, comprising: Computer-readable storage medium and processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the power system fault location method based on the BiGRU-GCN structure according to any one of claims 1 to 4.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the power system fault location method based on the BiGRU-GCN structure according to any one of claims 1 to 4 is implemented.

Images