Relay protection ai intelligent fault detection processing method and system under new power system

Abstract

This invention relates to the field of new energy power grid operation and maintenance technology, and in particular to a novel AI-powered intelligent fault detection and processing method and system for relay protection in power systems. This method significantly improves the timeliness, accuracy, and intelligence of fault handling in new energy distribution networks, meeting the requirements for safe and stable operation of new power systems. The method includes: collecting and preprocessing time-series data from the distribution network on the line side and the new energy equipment side, as well as real-time status data of circuit breakers, to obtain preprocessed power grid data; identifying the power grid connection relationships and levels based on matching algorithms and network topology breadth-first search technology, generating power grid topology data; distinguishing fault signals from new energy random fluctuation signals through a time-series fluctuation fault gating separation mechanism, and fusing time-series and spatial features to extract power grid fault features; and determining the propagation and correlation relationships of power grid fault features among topology nodes through a graph neural network classification mechanism to determine the faulty line number.

Description

Technical Field

[0001] This invention relates to the field of new energy power grid operation and maintenance technology, and in particular to a novel AI-based intelligent fault detection and processing method and system for relay protection in power systems. Background Technology

[0002] In the field of new energy power grid operation and maintenance, the accurate extraction of fault characteristics and efficient location of fault points in the fault handling of distribution networks with new energy access, especially when photovoltaic, wind power and other new energy equipment are connected, are of great significance. Currently, in this scenario, there is a core pain point that the fault modes are complex after the new energy access, and traditional methods are difficult to accurately extract features and locate faults. Specifically, the random fluctuations in the output of new energy cause fault signals to be easily confused with normal fluctuation signals. The interconnected topology of the distribution network makes the propagation path of fault signals complex. Traditional methods cannot effectively distinguish fault types and it is difficult to trace the source of faults.

[0003] In existing technologies, solutions to this pain point mainly fall into three categories: First, fault analysis methods based on manual rules, which use the experience of maintenance personnel to set current and voltage thresholds and fault judgment logic to complete feature identification and location; second, traditional machine learning solutions, which rely on manually designed feature engineering to classify and locate fault data through shallow models; and third, topology analysis-assisted location methods, which infer fault areas based on fixed power grid topology relationships and combined with single fault parameters.

[0004] However, solutions relying on manual rules are limited by the static nature of empirical thresholds and cannot adapt to the dynamic changes in fault modes caused by fluctuations in renewable energy output. They are prone to misjudgment in scenarios where fault signals and normal fluctuation signals are confused. Traditional machine learning solutions are limited by the limitations of manual feature engineering and cannot capture the nonlinear and non-stationary characteristics of faults after renewable energy access. They are also susceptible to voltage and current oscillations, resulting in large positioning errors. Simple topology analysis methods can only handle fixed topology relationships and cannot adapt to the dynamic adjustment of distribution network operation modes. They also cannot integrate high-dimensional fault time-series data with topology association information, leading to delayed fault identification, ambiguous positioning range, and inability to quickly locate the fault area. Ultimately, this affects the timeliness and accuracy of power grid fault handling and makes it difficult to meet the requirements for the safe and stable operation of renewable energy power grids. Summary of the Invention

[0005] To address one of the aforementioned problems in the prior art, this invention provides a novel AI-based intelligent fault detection and processing method and system for relay protection in power systems.

[0006] To achieve the above objectives, this invention provides a novel AI-powered intelligent fault detection and processing method for relay protection in power systems, comprising: S1: collecting and preprocessing three-phase current time-series data, three-phase voltage time-series data, and circuit breaker real-time status data of the distribution network at the line side and new energy equipment side to obtain preprocessed grid data; S2: based on the preprocessed grid data, identifying grid connection relationships and levels using a matching algorithm and network topology breadth-first search technology to generate grid topology data; S3: based on the preprocessed grid data and the grid topology data, distinguishing fault signals from new energy random fluctuation signals through a time-series fluctuation fault gating separation mechanism, and fusing time-series and spatial features to extract grid fault features; S4: mapping the grid fault features to the topology nodes of the grid topology data, and determining the propagation and correlation relationships of the grid fault features among the topology nodes through a graph neural network classification mechanism to determine the fault line number.

[0007] In addition, step S2 specifically includes: S21: Based on the real-time status data of the circuit breakers, identify the real-time opening and closing status of each circuit breaker in the distribution network to determine the actual on / off relationship of the electrical connection; S22: Use the matching algorithm to match the line connection relationship, and based on the network topology breadth-first search technology, traverse the topology nodes and branches of the distribution network under the actual on / off relationship; S23: Generate the power grid topology data based on the traversal results.

[0008] Furthermore, step S3 specifically includes: S31: Based on the preprocessed line-side time-series data and new energy equipment-side time-series data, filtering out new energy fluctuation time-series features through a time-series fluctuation fault separation gating mechanism, and extracting fault time-series basic features; S32: Based on the fault time-series basic features and the power grid topology data, generating spatiotemporal fusion fault features through a spatiotemporal coupling enhancement mechanism; S33: Optimizing the feature weights of the spatiotemporal fusion fault features to obtain the power grid fault features.

[0009] Furthermore, step S31 specifically includes: extracting local features from the line-side time-series data and the new energy equipment-side time-series data respectively to obtain line-side time-series features and new energy equipment-side time-series features; concatenating the line-side time-series features and the new energy equipment-side time-series features, and processing them through a multilayer perceptron to obtain the original time-series features; acquiring historical time-series features and the rated output power of the new energy equipment, and calculating the normal fluctuation feature suppression weight based on the original time-series features, the historical time-series features, and the rated output power of the new energy equipment; dynamically adjusting the original time-series features using the normal fluctuation feature suppression weight, and superimposing a time-series rated power constraint function to obtain the fault time-series basic features.

[0010] Furthermore, step S32 specifically includes: obtaining the grid node connection relationship matrix and the line impedance matrix based on the grid topology data; generating a topology constraint matrix based on the grid node connection relationship matrix and the line impedance matrix; processing the topology constraint matrix using a graph convolutional network to obtain topology spatial features; processing the fault time-series basic features across time periods using a long short-time memory network to obtain cross-time period features; and coupling the cross-time period features, the topology constraint matrix, and the topology spatial features to generate the spatiotemporal fusion fault features.

[0011] In addition, step S33 specifically includes: calculating the fault feature attention weights based on the spatiotemporal fusion fault features at the current time and the spatiotemporal fusion fault features at historical times through an attention mechanism; weighting the spatiotemporal fusion fault features using the fault feature attention weights, and simultaneously enhancing the features using a time-series enhancement function to generate the power grid fault features, wherein the time-series enhancement function is implemented by calculating the difference between the current features and the average values ​​of recent historical features and performing a nonlinear transformation.

[0012] In addition, step S4 specifically includes: S41: Based on the power grid fault characteristics and the power grid topology data, the fault characteristics are matched to the topology nodes through a fault topology mapping mechanism to obtain topology-based fault characteristics; S42: Based on the topology-based fault characteristics, the topology features are enhanced through a second-order topology attention classification mechanism to obtain fault classification features; S43: The fault classification features are mapped to the fault probability distribution of each line, and the line with the highest probability is selected as the fault line output.

[0013] Furthermore, step S41 specifically includes: normalizing the power grid fault features to obtain normalized power grid fault features, and obtaining a topology node attribute matrix based on the power grid topology data; concatenating the normalized power grid fault features with the topology node attribute matrix, and obtaining regularized node-level fault features after passing through a multilayer perceptron and random deactivation processing; and mapping the node-level fault features with the power grid node connection relationship matrix to obtain the topology-based fault features.

[0014] Furthermore, step S42 specifically includes: performing first-order graph convolution and second-order graph convolution operations sequentially on the topological fault features, and combining residual connections to obtain second-order graph convolution node features; calculating the attention weights between the second-order graph convolution node features and the power grid fault features, and performing weighted and nonlinear transformations on the features to obtain features enhanced by the feedforward network; and performing global average pooling and layer normalization on the features enhanced by the feedforward network to obtain the fault classification features.

[0015] In addition, the method further includes: S5: determining the faulty line based on the faulty line number, obtaining the rated parameters of the faulty line, using the rated parameters as initial values ​​for iteration, establishing and solving the topology node voltage equations according to the power grid equivalent circuit, performing iterative calculations and outputting the actual electrical parameters of the faulty line after convergence.

[0016] Furthermore, step S5 specifically includes: S51: Determine the faulty line according to the faulty line number output in step S4, obtain the rated output current of the new energy equipment of the faulty line, and set the rated output current as the initial value for iteration; S52: Construct the equivalent circuit of the power grid based on the power grid topology data; S53: Divide the voltage of the topology node into a superposition of normal and faulty components, and list the node voltage equations of the normal component network and the faulty component network based on the equivalent circuit; S54: Calculate the fault current according to the node voltage equations, and obtain the faulty component node voltage based on the self-impedance of the faulty node and the node voltage of the faulty node in the normal component network; S55: Superimpose the normal component node voltage and the faulty component node voltage to obtain the voltage calculation results of each node of the power grid after the fault; S56: Set the iteration criterion, and determine whether the calculation process of the fault current and the node voltage converges based on the iteration criterion; S57: If the convergence condition is met, output the actual electrical parameters of the faulty line.

[0017] Another aspect of this invention provides a novel AI-powered intelligent fault detection and processing system for relay protection in power systems, comprising: a data acquisition and preprocessing module for acquiring and preprocessing three-phase current time-series data, three-phase voltage time-series data, and real-time status data of circuit breakers on the line side and new energy equipment side of the distribution network to obtain preprocessed power grid data; a topology generation module for identifying the power grid connection relationships and levels based on the preprocessed power grid data using a matching algorithm and network topology breadth-first search technology to generate power grid topology data; a power grid fault feature determination module for distinguishing fault signals from new energy random fluctuation signals through a time-series fluctuation fault gating separation mechanism based on the preprocessed power grid data and the power grid topology data, and fusing time-series and spatial features to extract power grid fault features; and a fault line number determination module for mapping the power grid fault features to the topology nodes of the power grid topology data, and determining the propagation and correlation relationships of the power grid fault features among the topology nodes through a graph neural network classification mechanism to determine the fault line number.

[0018] The beneficial effects of this invention are reflected in the fact that, based on the novel AI-powered intelligent fault detection and processing method and system for relay protection in power systems provided by this invention, a systematic solution is offered to address the core technical problems of complex fault modes after the integration of new energy sources and the difficulty of accurately extracting features and locating faults using traditional methods. Firstly, the solution of this invention uses a time-series fluctuation-fault separation gating mechanism to dynamically suppress the interference of random power output fluctuations from new energy sources on fault signals, solving the problem of confusion between fault signals and normal operation fluctuations, and achieving accurate and robust extraction of fault features. Secondly, by synergistically fusing time-series evolution and topological spatial features, highly identifiable and robust grid fault features can be extracted, enabling fault features to carry both temporal information and spatial propagation relationships. This achieves accurate tracking of fault propagation paths and rapid and precise location of fault lines in complex interconnected power grids, improving the efficiency and reliability of fault handling. Furthermore, by accurately mapping fault features to dynamic topology nodes and utilizing a graph neural network classification mechanism, high-precision and rapid location of fault sources can be achieved in the dynamic topology. The solution of this invention can significantly improve the timeliness, accuracy, and intelligence level of fault handling in new energy distribution networks, meeting the requirements for the safe and stable operation of new power systems. Attached Figure Description

[0019] Figure 1 The flowchart of the novel AI-based intelligent fault detection and processing method for relay protection in power systems provided by this invention is shown below. Figure 2 This is a flowchart illustrating an example method for implementing step S2 provided by the present invention. Figure 3 A flowchart illustrating an example method for implementing step S3 provided by the present invention; Figure 4 This is a flowchart illustrating an example method for implementing step S4 provided by the present invention. Figure 5 This is a schematic diagram of the structure of the novel AI-powered intelligent fault detection and processing system for relay protection in a power system provided by the present invention. Detailed Implementation

[0020] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 This embodiment provides a novel AI-powered intelligent fault detection and processing method for relay protection in power systems, such as... Figure 1 As shown, it includes: S1: Collect and preprocess three-phase current time-series data, three-phase voltage time-series data, and circuit breaker real-time status data from the distribution network on both the line side and the renewable energy equipment side to obtain preprocessed grid data. Specifically, step S1 is a general step for data acquisition and preprocessing. The collected data includes: three-phase current time-series data, three-phase voltage time-series data, and circuit breaker real-time status data from the distribution network lines and renewable energy equipment. The preprocessing step can be performed through wavelet denoising, outlier removal, and format standardization to obtain preprocessed grid data, including: preprocessed three-phase current time-series data, three-phase voltage time-series data, circuit breaker real-time status data, and three-phase current time-series data and three-phase voltage time-series data from the renewable energy equipment side. By performing preprocessing operations such as wavelet denoising, outlier removal, and format standardization on the raw monitoring data, noise interference can be effectively filtered out, data heterogeneity can be eliminated, and high-quality, standardized input can be provided for subsequent AI models, thus laying the data foundation for accurate fault feature extraction and reliable fault location.

[0022] S2: Based on the preprocessed power grid data, the power grid connection relationships and hierarchy are identified using matching algorithms and breadth-first search (BFS) technology to generate power grid topology data. Matching algorithms can employ patterns such as KPM (Knuth Morris Pratt), while BFS is a dynamic topology modeling method based on breadth-first traversal principles, used to automatically discover and construct real-time electrical connection relationships within the power grid. This step uses the KPM matching algorithm and BFS technology to identify the connection relationships and location hierarchy between main branches and external equipment, generating a dynamically operable distributed renewable energy graph model for the distribution network, thereby obtaining the power grid topology data.

[0023] In one alternative implementation, such as Figure 2 As shown, step S2 may specifically include: S21: Based on the real-time status data of the circuit breakers, identify the real-time opening and closing status of each circuit breaker in the distribution network to determine the actual on / off relationship of the electrical connections.

[0024] S22: Use a matching algorithm to match the line connection relationships, and based on the breadth-first search technique of network topology, traverse the topology nodes and branches of the distribution network under the actual on / off relationship.

[0025] S23: Generate power grid topology data based on the traversal results.

[0026] Steps S21-S23 construct a power grid topology model based on the on / off information reflected in the real-time status data of circuit breakers from the preprocessed power grid data, utilizing pattern matching algorithms and graph traversal techniques. Specifically, string matching algorithms such as KMP are used to perform pattern recognition and association matching on equipment identifiers and line names in the power grid. Building upon this, a breadth-first search technique is used to traverse all electrically connected equipment and branches layer by layer, starting from known power sources or key topology nodes, based on the real-time circuit breaker status. This identifies the main lines, branch levels, and the access locations and connections of distributed renewable energy equipment outside the substations, ultimately generating a dynamically operable distributed renewable energy graph model as the power grid topology data. This model not only describes the static connections between equipment but also dynamically reflects the real-time changes in the power grid's operating mode by integrating the real-time status of circuit breakers, providing an accurate spatial structure foundation for subsequent fault feature extraction based on spatiotemporal information and precise localization based on graph neural networks.

[0027] S3: Based on preprocessed grid data and grid topology data, a time-series fluctuation fault gating separation mechanism is used to distinguish fault signals from random fluctuation signals from renewable energy sources. Temporal and spatial features are then fused to extract grid fault characteristics. Specifically, the time-series fluctuation fault gating separation mechanism can separate fault signals from random fluctuations in renewable energy sources based on dynamic gating weights and physical parameter constraints. Temporal and spatial features refer to two types of information attributes that respectively characterize the evolution of fault signals over time and their propagation path along the grid topology. This step uses preprocessed time-series data (voltage, current) and dynamically generated grid topology data as input. An AI feature extraction mechanism separates fault signals from random fluctuations in renewable energy sources, and fuses the temporal evolution patterns with topological spatial correlations to ultimately extract grid fault characteristics.

[0028] In one alternative implementation, such as Figure 3 As shown, step S3 may specifically include the following steps: S31: Based on the preprocessed line-side time-series data and new energy equipment-side time-series data, the new energy fluctuation time-series features are filtered out by the time-series fluctuation fault separation gating mechanism, and the basic fault time-series features are extracted. Specifically, the time-series fluctuation-fault separation gating mechanism can be used to process the data. This mechanism performs local feature extraction on the current and voltage time-series data of the line side and the new energy side respectively, thereby separating the fault features.

[0029] In an optional implementation, step S31 specifically includes: extracting local features from the line-side time-series data and the new energy equipment-side time-series data respectively to obtain line-side time-series features and new energy equipment-side time-series features; concatenating the line-side time-series features and new energy equipment-side time-series features, and processing them through a multilayer perceptron to obtain the original time-series features; acquiring historical time-series features and the rated output power of the new energy equipment; calculating the normal fluctuation feature suppression weight based on the original time-series features, historical time-series features, and the rated output power of the new energy equipment; dynamically adjusting the original time-series features using the normal fluctuation feature suppression weight, and superimposing a time-series rated power constraint function to obtain the fault time-series basic features. This step S31 integrates current features, historical features, and the rated parameters of the new energy equipment to generate a suppression weight, which is used to actively attenuate components judged to belong to the normal output fluctuations of new energy at the feature level, and at the same time constrains and calibrates the features through physical rated parameters, thereby effectively separating the fault time-series basic features.

[0030] In a specific example, the calculation method for extracting the basic fault timing features through the timing fluctuation-fault separation gating mechanism is as follows: First, extract the timing features on the line side and the timing features on the new energy side: in, Let t represent the line-side time-series characteristics at time t, where t is the time index. It is a convolutional layer. These are the preprocessed three-phase current time series data, preprocessed three-phase voltage time series data, and preprocessed three-phase current time series data at time t-1, respectively. The time-series characteristics of the new energy side at time t. These are the preprocessed three-phase current time-series data and the preprocessed three-phase voltage time-series data of the new energy equipment side at time t, respectively.

[0031] The time-series characteristics of the line side and the new energy source side are then concatenated and processed by a multilayer perceptron to calculate the original time-series characteristics: in, For splicing operations, It is a multilayer perceptron.

[0032] Then, based on the original time-series characteristics, historical time-series characteristics, and the rated output power of the new energy equipment, the normal fluctuation characteristic suppression weight is calculated as follows: in, The normal fluctuation characteristic suppression weight at time t, The original time series characteristics at time t, For the Sigmoid function, The original time series characteristics at time t-1, This is the rated output power matrix for new energy equipment.

[0033] Finally, the original time-series characteristics are dynamically adjusted using the normal fluctuation feature suppression weight, and a time-series rated power constraint function is superimposed to obtain the basic fault time-series characteristics. The calculation method is as follows: in, The basic characteristics of the fault timing at time t are: For Hadama accumulation, To sum element by element, For time-sequential rated power constraint function, To take the absolute value, This is the rated current matrix for new energy equipment.

[0034] As can be seen from the above specific implementation methods, the timing fluctuation-fault separation gating mechanism addresses the problem of confusion between fault and normal fluctuation signals. First, it extracts local features from the relevant timing data of the line side and the renewable energy side separately. This side-by-side extraction design fully adapts to the different operating characteristics of the line side and the renewable energy side, ensuring that the core features of each side are not interfered with each other. The extracted local features are then spliced ​​and processed by a multilayer perceptron to obtain the original timing features. Subsequently, the interference features caused by the normal power output fluctuations of renewable energy are suppressed by dynamically adjusting the gating weight. At the same time, the features are physically calibrated using the rated parameters of renewable energy equipment by combining the timing-based rated power constraint function. The introduction of this constraint function ensures that the feature extraction does not deviate from the basic physical laws of power grid operation, thereby obtaining the basic fault timing features and effectively solving the core pain point of difficulty in distinguishing between fluctuation and fault signals.

[0035] S32: Based on the basic characteristics of fault timing and power grid topology data, a spatiotemporal fused fault feature is generated through a spatiotemporal coupling enhancement mechanism. Building upon step S31, this step further fuses features using the spatiotemporal coupling enhancement mechanism. The spatiotemporal coupling enhancement mechanism can generate spatiotemporal fused fault features by associating the fused timing evolution features with the topological space.

[0036] In an optional implementation, step S32 specifically includes: obtaining the grid node connection relationship matrix and line impedance matrix based on the grid topology data; generating a topology constraint matrix based on the grid node connection relationship matrix and line impedance matrix; processing the topology constraint matrix using a graph convolutional network to obtain topology spatial features; processing the cross-time fault temporal characteristics using a long short-term memory network to obtain cross-time features; and coupling the cross-time features, the topology constraint matrix, and the topology spatial features to generate spatiotemporally fused fault features. Specifically, step S32 uses a graph convolutional network to perform deep encoding on the dynamic topology structure and extract topology spatial features containing connection and propagation relationships; simultaneously, it uses a long short-term memory network to capture the cross-time evolution law of fault temporal characteristics. Then, the temporal features and topology spatial features are coupled to generate spatiotemporally fused fault features, so that the fault features simultaneously carry temporal and spatial information.

[0037] In a specific example, spatiotemporal fused fault features are generated based on fault timing characteristics and power grid topology data through a spatiotemporal coupling enhancement mechanism. The specific calculation method is as follows: in, The topological constraint matrix is... For the Softmax function, This is the power grid node connection matrix in the power grid topology data obtained in step S2. This refers to the line impedance matrix in the power grid topology data obtained in step S2. For topological space features, For graph convolutional networks, The time-crossing feature at time t, For Long Short-Term Memory (LSTM) networks, The fault timing characteristics at time t-1, The spatiotemporal fusion fault characteristics at time t.

[0038] Based on the above specific implementation methods, it can be seen that the spatiotemporal coupling enhancement mechanism addresses the problem of complex fault propagation paths caused by the interconnection of distribution network topologies. First, it performs deep processing on the power grid topology data through graph convolutional networks to mine the inherent correlation between topology nodes and obtain topological spatial features. At the same time, it uses long short-term memory networks to process the basic features of fault time series across time periods to capture the temporal evolution law of fault features. Then, it deeply couples the topological spatial features with the time series features across time periods to obtain spatiotemporally fused fault features. This breaks through the limitations of traditional techniques that analyze time series or topology data separately. Through the synergistic fusion of time series and spatial features, it depicts the propagation law of fault signals in complex topologies, providing a foundation for the subsequent extraction of fault features.

[0039] S33: Perform feature weight optimization on spatiotemporal fusion fault features to obtain power grid fault features.

[0040] In an optional implementation, step S33 specifically includes: calculating fault feature attention weights based on the spatiotemporal fusion fault features at the current time and the spatiotemporal fusion fault features at historical times using an attention mechanism; weighting the spatiotemporal fusion fault features using the fault feature attention weights, and simultaneously enhancing the features using a time-series enhancement function to generate power grid fault features. The time-series enhancement function is implemented by calculating the difference between the current feature and the average value of recent historical features and performing a nonlinear transformation. This step S33 uses a feature weight optimization method based on an attention mechanism to dynamically weight and enhance the spatiotemporal fusion features, highlighting feature dimensions strongly correlated with faults, suppressing irrelevant or weakly correlated interference, and outputting refined power grid fault features.

[0041] In a specific example, the specific calculation method for power grid fault features is as follows, based on the optimized feature weight allocation according to spatiotemporal fusion fault features: in, The attention weights for fault features at time t. For attention mechanisms, Calculated for the mean. Let represent the spatiotemporal fusion fault characteristics of the k time steps preceding time t, where k is the length of the time series window, which can range from 3 to 5. Let be the characteristics of the power grid fault at time t. For timing enhancement functions, This is the ReLU function.

[0042] Based on the specific implementation methods described above, this step dynamically optimizes the weights of each feature based on the obtained spatiotemporal fusion fault features, strengthens the proportion of features strongly correlated with the fault, weakens irrelevant interference features, and finally obtains power grid fault features. This can further improve the identification and targeting of fault features and ensure that the extracted power grid fault features can reflect the core attributes of the fault.

[0043] In summary, the core technical challenges of fault feature extraction in distribution networks connected to renewable energy sources are due to the unique characteristics of these scenarios. Renewable energy output exhibits inherent random fluctuations, and the distribution network itself possesses a complex interconnected topology. These factors combine to create a high degree of overlap between fault signals and normal operation fluctuations at the feature level, making accurate differentiation difficult using conventional methods. Furthermore, the interconnected topology of the distribution network causes signals to propagate along multiple paths after a fault occurs, further exacerbating the complexity and ambiguity of fault features. This makes it difficult to guarantee the accuracy of fault feature extraction, directly impacting the reliability and timeliness of subsequent fault location. Therefore, step S3 addresses the two core challenges of confusing fault signals with random fluctuations and the complexity of fault propagation paths, providing high-quality and robust feature inputs for subsequent accurate location.

[0044] S4: Map the power grid fault features to the topology nodes of the power grid topology data, and determine the propagation and correlation of the power grid fault features among the topology nodes through a graph neural network classification mechanism to identify the faulty line number. Specifically, in the fault location scenario of a distribution network connected to new energy sources, there is a lack of dimensional matching relationship between fault features and power grid topology nodes, and the fault signal propagates along complex topological paths, making it difficult for traditional methods to effectively establish the correlation between features and topology nodes. At the same time, single-dimensional topology analysis cannot fully explore the indirect correlation between nodes, further reducing the accuracy and robustness of faulty line classification. The graph neural network classification mechanism of this invention can adopt the Graph Neural Network (GNN) framework, which is an algorithm framework based on deep learning for processing graph structure data. This step is the core link of fault location. Its goal is to map the power grid fault features extracted in step S3 to the actual physical topology of the power grid, and use the structural relationship modeling of the graph neural network to finally determine the specific line number where the fault occurred.

[0045] In one alternative implementation, such as Figure 4 As shown, step S4 specifically includes: S41: Based on power grid fault characteristics and power grid topology data, a fault topology mapping mechanism is used to match fault features to topology nodes, resulting in topologicalized fault features. Specifically, the fault topology mapping mechanism is used to perform dimensional transformation to align and associate abstract power grid fault features with physical topology nodes. Step S41 mainly aims to solve the mismatch problem caused by the mismatch between feature vectors and physical node dimensions, and direct input into the graph neural network, thus providing correct and standardized input for subsequent topology-based analysis.

[0046] In an optional implementation, step S41 specifically includes: normalizing the power grid fault features to obtain normalized power grid fault features, and obtaining the topology node attribute matrix based on the power grid topology data; concatenating the normalized power grid fault features with the topology node attribute matrix, and after passing through a multilayer perceptron and random deactivation processing, obtaining regularized node-level fault features; and mapping the node-level fault features with the power grid node connection relationship matrix to obtain topological fault features. Specifically, the fault-topology mapping mechanism addresses the problem of feature-node dimension mismatch by first performing layer normalization on the power grid fault features to eliminate dimensional differences, then concatenating the fault features with the topology node attribute matrix through a multilayer perceptron to achieve accurate matching between feature dimensions and the number of topology nodes, and then suppressing overfitting through random deactivation operation to finally obtain regularized node-level fault features. This solves the mismatch problem caused by directly inputting features into a GNN, ensuring that fault features can establish a one-to-one correspondence with topology nodes.

[0047] In a specific implementation example, based on power grid fault characteristics and power grid topology data, a fault-topology mapping mechanism is used to match characteristics with topology nodes. The specific calculation method for topology-based fault characteristics is as follows: in, These are the node-level fault characteristics after regularization. This is a random deactivation operation. For layer normalization, This is the topology node attribute matrix in the power grid topology data obtained in S2. This represents the random inactivation probability. Let be the topological fault characteristics at time t.

[0048] S42: Based on the topological fault features, topological features are enhanced through a second-order topological attention classification mechanism to obtain fault classification features. The second-order topological attention classification mechanism is a graph neural network decision-making method that integrates second-order graph convolution and attention weighting, used to enhance fault association features in complex topologies and achieve accurate classification. Step S42 is used to mine and enhance the topological fault features already mapped to nodes.

[0049] In an optional implementation, step S42 specifically includes: performing first-order graph convolution and second-order graph convolution operations sequentially on the topological fault features, and combining residual connections to obtain second-order graph convolution node features; calculating the attention weights between the second-order graph convolution node features and the power grid fault features, and performing weighted and nonlinear transformations on the features to obtain features enhanced by the feedforward network; and performing global average pooling and layer normalization on the features enhanced by the feedforward network to obtain fault classification features. Specifically, the second-order topological attention classification mechanism addresses the problem of insufficient topological correlation analysis. It first mines the direct correlation relationships between topological nodes through first-order graph convolution, retains the original topological fault features by combining residual connections, and then mines the indirect correlations between nodes through second-order graph convolution, further enhancing the expressive power of topological features. At the same time, it introduces an attention mechanism to calculate the correlation weights between nodes and fault features, and uses a feedforward neural network to realize cross-dimensional nonlinear interaction of features, ultimately obtaining fault classification features. This step combines second-order graph convolution with attention mechanisms, which overcomes the limitations of traditional GNNs that rely solely on single-layer topology analysis. It can more comprehensively characterize the propagation patterns of fault signals in complex topologies, thereby improving the accuracy and robustness of fault line classification.

[0050] In a specific implementation example, the calculation method for obtaining fault classification features by enhancing topological features through a second-order topological attention classification mechanism based on topological fault characteristics is as follows: in, The features of the second-order graph convolution nodes at time t are... For second-order graph convolution, For first-order graph convolution, This is a second-order power grid node connection matrix. Features enhanced by the feedforward network It is a feedforward neural network. The fault classification features at time t are... This is global average pooling.

[0051] S43: Map the fault classification features to the fault probability distribution of each line, and select the line with the highest probability as the faulty line output. Step S43 maps the fault classification features through a linear layer, and then normalizes them using the Softmax function so that the sum of the fault probabilities of each line is 1, thus obtaining the fault probability distribution of each line at time t; finally, by taking the index operation of the maximum value, the line number with the highest fault probability is determined and output as the faulty line number at the current time.

[0052] Based on the novel AI-powered intelligent fault detection and processing method for relay protection in power systems provided in this embodiment, a systematic solution is offered to address the core technical challenges of complex fault modes after the integration of new energy sources and the difficulty of accurately extracting features and locating faults using traditional methods. This embodiment first employs a time-series fluctuation-fault separation gating mechanism to dynamically suppress the interference of random power output fluctuations from new energy sources on fault signals, solving the problem of confusion between fault signals and normal operation fluctuations, and achieving accurate and robust extraction of fault features. Secondly, by collaboratively fusing time-series evolution and topological spatial features, highly identifiable and robust grid fault features can be extracted, ensuring that fault features carry both temporal information and spatial propagation relationships. This enables accurate tracking of fault propagation paths and rapid, precise location of fault lines in complex interconnected power grids, improving the efficiency and reliability of fault handling. Furthermore, by accurately mapping fault features to dynamic topology nodes and utilizing a graph neural network classification mechanism, high-precision and rapid location of fault sources can be achieved in the dynamic topology. This embodiment significantly improves the timeliness, accuracy, and intelligence of fault handling in new energy distribution networks, meeting the requirements for the safe and stable operation of new power systems.

[0053] In one optional embodiment, the novel AI-powered intelligent fault detection and processing method for relay protection in power systems of the present invention may further include: S5: Based on the faulty line number, determine the faulty line, obtain the rated parameters of the faulty line, use the rated parameters as the initial values ​​for iteration, establish and solve the topology node voltage equations according to the equivalent circuit of the power grid, perform iterative calculations, and output the actual electrical parameters of the faulty line after convergence. Following intelligent fault location, step S5 performs fault electrical quantity calculations for the determined faulty line to provide a quantitative basis for protection setting verification and analysis.

[0054] In one optional implementation, step S5 may specifically include the following steps: S51: Determine the faulty line based on the faulty line number output in step S4, obtain the rated output current of the new energy equipment on the faulty line, and set the rated output current as the initial value for iteration; to quantify the fault in detail, it is necessary to obtain the rated output current of the new energy equipment (such as photovoltaic and wind turbine inverters) on the line or related new energy equipment, and set this rated value as the initial value for iterative calculation, which provides a reasonable starting point for subsequent iterative solutions. Especially considering the fluctuating and uncertain characteristics of new energy power output, it can improve the reliability and efficiency of calculation convergence.

[0055] S52: Construct an equivalent circuit of the power grid based on the power grid topology data; Specifically, based on the dynamic power grid topology data generated in step S2, an equivalent circuit model of the power grid including elements such as system power supply, line impedance, load and new energy access can be constructed so as to establish an equivalent circuit for calculation based on the real-time dynamic topology and ensure that the calculation model is consistent with the actual operating structure of the current power grid.

[0056] S53: The voltage of the topology nodes is divided into a superposition of normal and fault components. Based on the equivalent circuit, the node voltage equations for the normal component network and the fault component network are listed separately. The superposition principle can be used to treat the post-fault grid state as a superposition of normal and fault components. Based on the equivalent circuit, the node voltage equations for the normal component network and the fault component network are listed separately. By separating the normal and fault components through the superposition principle, the complex fault calculation problem is decomposed into two relatively independent and easier-to-solve sub-problems, making the calculation more reliable.

[0057] S54: Calculate the fault current based on the node voltage equation, and obtain the fault component node voltage based on the self-impedance of the fault node and the node voltage of the fault node in the normal component network; by solving the fault component network, the fault current and fault component voltage are obtained, which can be used to quantitatively describe the impact of fault injection on the voltage of each node in the power grid.

[0058] S55: The normal component node voltage and the fault component node voltage are superimposed to obtain the voltage calculation results of each node in the power grid after the fault; by superimposing the normal component and the fault component to obtain the actual voltage after the fault, the complete voltage amplitude information of each point in the power grid under the fault state can be restored.

[0059] S56: Set an iterative criterion to determine whether the calculation process of fault current and node voltage has converged. This is achieved by setting a convergence criterion (such as the iterative change in voltage or current being less than a threshold). During the calculation, this criterion is used to determine whether the calculation results have reached stable convergence. If the convergence condition is met, the final determined fault current, node voltage, and other actual electrical parameters are output. By introducing an iterative criterion to control the calculation accuracy and convergence, it ensures that even under conditions of strong nonlinearity or initial value deviation, the calculation process can automatically adjust through iteration until a stable and accurate solution is obtained, guaranteeing the reliability and accuracy of the output results.

[0060] S57: If the convergence condition is met, output the actual electrical parameters of the faulty line.

[0061] Step S5 achieves automated and high-precision calculation of electrical parameters of faulty lines by setting reasonable initial values ​​for new energy sources, using dynamic topology modeling, applying the superposition principle to solve the problem step by step, and introducing iterative convergence control. This provides direct and reliable data support for the subsequent intelligent verification of relay protection settings.

[0062] This embodiment also provides a novel AI-powered intelligent fault detection and processing system for relay protection in power systems, used to perform the aforementioned novel AI-powered intelligent fault detection and processing for relay protection in power systems. The content already described in the method section will not be repeated here; only the system architecture will be explained. Figure 5 As shown, the novel AI-powered intelligent fault detection and processing system for relay protection in this embodiment includes: The acquisition and preprocessing module 501 is used to acquire and preprocess the three-phase current time-series data, three-phase voltage time-series data and circuit breaker real-time status data of the distribution network on the line side and the new energy equipment side to obtain the preprocessed power grid data.

[0063] The topology generation module 502 is used to identify the grid connection relationships and levels based on the preprocessed grid data, using matching algorithms and network topology breadth search technology, and generate grid topology structure data.

[0064] The power grid fault feature determination module 503 is used to distinguish fault signals from new energy random fluctuation signals based on preprocessed power grid data and power grid topology data, through a time-series fluctuation fault gating separation mechanism, and to extract power grid fault features by fusing time-series and spatial features.

[0065] The fault line number determination module 504 is used to map the power grid fault features to the topology nodes of the power grid topology data, and determine the propagation and correlation of the power grid fault features between the topology nodes through a graph neural network classification mechanism, so as to determine the fault line number.

[0066] Based on the novel AI-powered intelligent fault detection and processing system for relay protection in power systems provided in this embodiment, a systematic solution is offered to address the core technical challenges of complex fault modes after the integration of new energy sources and the difficulty of accurately extracting features and locating faults using traditional methods. This embodiment first employs a time-series fluctuation-fault separation gating mechanism to dynamically suppress the interference of random power output fluctuations from new energy sources on fault signals, solving the problem of confusion between fault signals and normal operation fluctuations, and achieving accurate and robust extraction of fault features. Secondly, by collaboratively fusing time-series evolution and topological spatial features, highly identifiable and robust grid fault features can be extracted, ensuring that fault features carry both temporal information and spatial propagation relationships. This enables accurate tracking of fault propagation paths and rapid, precise location of fault lines in complex interconnected power grids, improving the efficiency and reliability of fault handling. Furthermore, by accurately mapping fault features to dynamic topology nodes and utilizing a graph neural network classification mechanism, high-precision and rapid location of fault sources can be achieved in the dynamic topology. This embodiment's solution significantly improves the timeliness, accuracy, and intelligence of fault handling in new energy distribution networks, meeting the requirements for the safe and stable operation of new power systems.

[0067] In the description of the embodiments of the present invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "center," "top," "bottom," "top," "bottom," "inner," "outer," "inner side," and "outer side," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. "Inner side" refers to the interior or enclosed area or space. "Outer perimeter" refers to the area surrounding a specific component or specific area.

[0068] In the description of embodiments of the present invention, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of that feature. In the description of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0069] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "joining," and "assembly" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0070] In the description of embodiments of the present invention, specific features, structures, materials or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0071] In the description of the embodiments of the present invention, it should be understood that "-" and "~" represent a range between two numerical values, and this range includes the endpoints. For example, "AB" represents a range greater than or equal to A and less than or equal to B. "A~B" represents a range greater than or equal to A and less than or equal to B.

[0072] In the description of embodiments of the present invention, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A novel AI-powered intelligent fault detection and handling method for relay protection in power systems, characterized in that, include: S1: Collect and preprocess the three-phase current time-series data, three-phase voltage time-series data, and circuit breaker real-time status data of the distribution network on the line side and the new energy equipment side to obtain the preprocessed power grid data; S2: Based on the preprocessed power grid data, the power grid connection relationships and levels are identified using a matching algorithm and network topology breadth-first search technology, and power grid topology data is generated. S3: Based on the preprocessed power grid data and the power grid topology data, the fault signal and the random fluctuation signal of new energy are distinguished through the time-series fluctuation fault gating separation mechanism, and the time-series and spatial features are fused to extract the power grid fault features; S4: Map the power grid fault features to the topology nodes of the power grid topology data, and determine the propagation and correlation of the power grid fault features among the topology nodes through a graph neural network classification mechanism, so as to determine the fault line number.

2. The novel AI-based intelligent fault detection and processing method for relay protection in a power system according to claim 1, characterized in that, Step S2 specifically includes: S21: Based on the real-time status data of the circuit breaker, identify the real-time opening and closing status of each circuit breaker in the power distribution network to determine the actual on / off relationship of the electrical connection. S22: Use the matching algorithm to match the line connection relationship, and based on the network topology breadth-first search technology, traverse the topology nodes and branches of the distribution network under the actual on / off relationship; S23: Generate the power grid topology data based on the traversal results.

3. The novel AI-powered intelligent fault detection and processing method for relay protection in a power system according to claim 1, characterized in that, Step S3 specifically includes: S31: Based on the preprocessed line-side timing data and new energy equipment-side timing data, filter out the new energy fluctuation timing features through the timing fluctuation fault separation gating mechanism, and extract the basic fault timing features; S32: Based on the fault timing characteristics and the power grid topology data, generate spatiotemporal fused fault characteristics through a spatiotemporal coupling enhancement mechanism; S33: Perform feature weight optimization on the spatiotemporal fusion fault features to obtain the power grid fault features.

4. The novel AI-based intelligent fault detection and processing method for relay protection in a power system according to claim 3, characterized in that, Step S31 specifically includes: Local feature extraction is performed on the time-series data of the line side and the time-series data of the new energy equipment side respectively to obtain the time-series features of the line side and the time-series features of the new energy equipment side. The line-side timing features and the new energy-side timing features are concatenated and then processed by a multilayer perceptron to obtain the original timing features. Obtain historical time-series characteristics and rated output power of new energy equipment, and calculate normal fluctuation characteristic suppression weights based on the original time-series characteristics, the historical time-series characteristics, and the rated output power of new energy equipment; The original time-series characteristics are dynamically adjusted using the normal fluctuation feature suppression weights, and a time-series rated power constraint function is superimposed to obtain the fault time-series basic characteristics.

5. The novel AI-based intelligent fault detection and processing method for relay protection in a power system according to claim 3, characterized in that, Step S32 specifically includes: Based on the power grid topology data, obtain the power grid node connection matrix and the line impedance matrix; Based on the power grid node connection matrix and the line impedance matrix, a topology constraint matrix is ​​generated; The topological constraint matrix is ​​processed using a graph convolutional network to obtain topological spatial features; The fault time-series basic features across time points are processed using a long short-term memory network to obtain cross-time-point features; The spatiotemporal fusion fault features are generated by coupling the cross-time features, the topological constraint matrix, and the topological space features.

6. The novel AI-based intelligent fault detection and processing method for relay protection in a power system according to any one of claims 1 to 5, characterized in that, Step S4 specifically includes: S41: Based on the power grid fault characteristics and the power grid topology data, the fault characteristics are matched to the topology nodes through a fault topology mapping mechanism to obtain topology-based fault characteristics; S42: Based on the topological fault features, topological features are enhanced through a second-order topological attention classification mechanism to obtain fault classification features; S43: Map the fault classification features to the fault probability distribution of each line, and select the line with the highest probability as the fault line output.

7. The novel AI-based intelligent fault detection and processing method for relay protection in a power system according to claim 6, characterized in that, Step S41 specifically includes: The power grid fault characteristics are normalized to obtain normalized power grid fault characteristics, and the topology node attribute matrix is ​​obtained based on the power grid topology data. The normalized power grid fault features are concatenated with the topology node attribute matrix, and after passing through a multilayer perceptron and random deactivation processing, regularized node-level fault features are obtained. The node-level fault features are mapped to the power grid node connection matrix to obtain the topology-based fault features.

8. The novel AI-based intelligent fault detection and processing method for relay protection in a power system according to claim 6, characterized in that, Step S42 specifically includes: The topological fault features are sequentially subjected to first-order graph convolution and second-order graph convolution operations, and combined with residual connections to obtain second-order graph convolution node features. The attention weights between the features of the second-order graph convolutional nodes and the features of the power grid faults are calculated, and the features are weighted and nonlinearly transformed to obtain the features after the feedforward network is enhanced. The enhanced features of the feedforward network are subjected to global average pooling and layer normalization to obtain the fault classification features.

9. The novel AI-powered intelligent fault detection and processing method for relay protection in a power system as described in claim 1, characterized in that, The method further includes: S5: Determine the faulty line based on the faulty line number, obtain the rated parameters of the faulty line, use the rated parameters as the initial values ​​for iteration, establish and solve the topology node voltage equations according to the power grid equivalent circuit, perform iterative calculations and output the actual electrical parameters of the faulty line after convergence.

10. A novel AI-powered intelligent fault detection and processing system for relay protection in a power system, characterized in that, include: The data acquisition and preprocessing module is used to acquire and preprocess the three-phase current time-series data, three-phase voltage time-series data, and circuit breaker real-time status data of the distribution network on the line side and the new energy equipment side to obtain the preprocessed power grid data. The topology generation module is used to identify the power grid connection relationships and levels based on the preprocessed power grid data, using a matching algorithm and network topology breadth-first search technology, and generate power grid topology structure data. The power grid fault feature determination module is used to distinguish fault signals from new energy random fluctuation signals based on the preprocessed power grid data and the power grid topology data, through a time-series fluctuation fault gating separation mechanism, and to extract power grid fault features by fusing time-series and spatial features. The fault line number determination module is used to map the power grid fault features to the topology nodes of the power grid topology data, and determine the propagation and correlation of the power grid fault features among the topology nodes through a graph neural network classification mechanism, so as to determine the fault line number.

Images