Smart fault location method for distribution network with high proportion of distributed power supply

Abstract

This invention discloses an intelligent fault location method for distribution networks with a high proportion of distributed generation (DG) access, belonging to the field of power system fault diagnosis technology. The method includes: acquiring multi-source operation data of the distribution network and establishing a fault analysis model; extracting a fault feature set containing dynamic response characteristics of DG; constructing a graph structure model based on the distribution network topology; extracting spatial coupling features using a graph neural network and extracting dynamic evolution features of the fault using a temporal neural network with an attention mechanism; fusing the two types of features to identify the fault segment; estimating the initial fault distance by combining the local dynamic response information of the DG within the segment; performing consistency verification and residual correction according to physical constraint rules; and outputting the final location result. This invention significantly improves the accuracy of fault location in scenarios with a high proportion of DG access by explicitly modeling the differences in DG control modes, fusing spatiotemporal dual-stream features, and embedding dual verification based on physical constraints.

Description

Technical Field

[0001] This invention relates to the field of power system fault diagnosis and intelligent operation and maintenance technology, specifically to an intelligent fault location method for distribution networks with a high proportion of distributed power sources connected to the network. Background Technology

[0002] With the advancement of new power system construction, the penetration rate of inverter-type distributed power sources such as distributed photovoltaic and wind power in distribution networks continues to increase, and distribution networks are evolving from traditional single-source radial networks to active distribution networks with multiple sources and bidirectional power flow. This change leads to complex fault current distribution, and inverter-type distributed power sources have multiple control modes such as constant power, constant current, reactive power support, and low voltage ride-through. Their fault transient response differs significantly from that of traditional synchronous power sources, resulting in a substantial decrease in the accuracy of traditional fault location methods.

[0003] Among existing fault location methods, the mechanism analysis method based on impedance or traveling wave is greatly affected by the amplification of multiple power sources and line parameter errors; the section judgment method based on switch action or fault indicator can only achieve coarse-grained location; although the data-driven method based on artificial intelligence can handle nonlinear relationships, some methods fail to make full use of the spatial coupling characteristics of the distribution network topology, some methods do not systematically characterize the differentiated impact of multiple control modes of distributed power sources on fault characteristics, and generally lack an effective mechanism to deeply embed the physical laws of the power system into model training and inference, resulting in insufficient accuracy and robustness of location in scenarios with a high proportion of distributed power source access.

[0004] Therefore, there is an urgent need for an intelligent fault location method that can accurately perceive differences in distributed power control modes, deeply integrate spatiotemporal fault characteristics, and embed physical constraints. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent fault location method for high-proportion distributed power source access to distribution networks. By explicitly introducing distributed power source control mode identifiers, constructing a dual-stream deep fusion feature extraction architecture, and setting a physical constraint dual-layer verification mechanism, the accuracy of fault location in high-proportion distributed power source access scenarios is effectively improved.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent fault location method for high-proportion distributed power generation connected to distribution networks, comprising the following steps:

[0007] Acquire multi-source operation data of the distribution network and establish a fault analysis model to extract a fault feature set containing dynamic response characteristics of distributed power sources; the dynamic response characteristics include electrical quantity change characteristics that characterize the change in operating state before and after the fault, and control mode identifiers that distinguish distributed power source control strategies.

[0008] Based on the distribution network topology, a graph structure model is constructed. A graph neural network is used to extract spatial coupling features that characterize the association between node and branch faults, and a temporal neural network with attention mechanism is used to extract dynamic evolution features of faults. The spatial coupling features and the dynamic evolution features are fused together, and fault sections are identified based on the fused features to obtain candidate fault sections.

[0009] For the candidate fault section, the location of the fault point is estimated by combining the local dynamic response information of the distributed power source within the section, and the initial value of the fault distance is obtained.

[0010] Based on preset physical constraint rules, the candidate fault segments and initial fault distance values ​​are physically consistent, and residual corrections are performed on the initial fault distance values ​​that pass the verification, and the final fault location result is output.

[0011] Preferably, the dynamic response characteristics of the distributed power source are expressed as follows:

[0012]

[0013] In the formula, This represents the dynamic response feature vector of a distributed power source. This represents the change in active power before and after a distributed power source failure. This represents the change in reactive power before and after a distributed power source failure. This represents the change in output current before and after a distributed power source failure. This is an identifier for the distributed power supply control mode.

[0014] The control mode identifier is used to indicate the current control mode of the distributed power source, and the control mode includes at least one of the following: constant power control mode, constant current control mode, reactive power support control mode, and low voltage ride-through control mode.

[0015] Preferably, the step of extracting spatial coupling features representing the association between nodes and branch faults using graph neural networks includes:

[0016] The bus nodes, load nodes, distributed generation access nodes, measurement nodes and switch nodes in the distribution network are used as nodes in the graph structure model, and feeder branches, tie branches and transformer connection branches are used as edges. Attribute matrices are constructed for each node and edge.

[0017] The spatial coupling characteristics are obtained by propagating and aggregating local fault disturbance information along the distribution network topology through the message propagation mechanism of a multi-layer graph neural network; wherein, the inter-layer propagation process of the graph neural network is represented as follows:

[0018]

[0019] In the formula, For the first The feature matrix output by the layer after message propagation and feature transformation. The adjacency matrix after adding self-loops, For degree matrix, For the first Layer input feature matrix, For the first Layer weight matrix, This is the activation function.

[0020] Preferably, the extraction of fault dynamic evolution features using a temporal neural network with an attention mechanism includes:

[0021] Long Short-Term Memory (LSTM) networks or gated recurrent unit (GRU) networks are used to encode the time-series measurement data within a preset time window before and after the fault, thereby obtaining the hidden state vector at each time point.

[0022] The attention weights of the hidden state at each time step are calculated using an attention mechanism. The formula for calculating the attention weights is as follows:

[0023]

[0024] In the formula, for The hidden state vector at time t. , and For learnable parameters, for Attention weight at any given moment; The dot product operation maps the vector output by the tanh function to a scalar fraction. , used to measure the The importance of the hidden state at any given time for fault location tasks;

[0025] The fault dynamic evolution characteristics are obtained by weighting and summing the hidden state at each time step based on the attention weights.

[0026]

[0027] In the formula, This represents the feature vector of fault dynamic evolution. This represents the total number of time points. For time indexing.

[0028] Preferably, the step of fusing the spatial coupling features with the dynamic evolution features and identifying fault segments based on the fused features includes:

[0029] Spatial coupling feature vectors output by graph neural networks The dynamic evolution feature vector output by the temporal neural network Feature concatenation is performed to obtain a fused feature vector. ;

[0030] The fused feature vector is input into the classification layer, which outputs the probability distribution of each candidate segment as a fault segment, and selects the segment corresponding to the maximum probability or the segment whose probability exceeds a preset threshold as the candidate fault segment.

[0031] Preferably, estimating the fault location by combining the local dynamic response information of distributed power sources within the section includes:

[0032] Extract the voltage and current measurements at the beginning and end of the candidate fault section, and combine them with the line resistance, line reactance, line length of the candidate fault section, as well as the changes in active power and reactive power of the distributed power sources within the section, to construct a regression input feature vector.

[0033] The regression input feature vector is input into a pre-trained fault distance regression model, and the ratio of the distance from the fault point to the beginning of the segment to the total length of the segment is output as the initial value of the fault distance.

[0034] Preferably, the preset physical constraint rules include at least one of the following:

[0035] Topology reachability constraints are used to verify whether the candidate fault section is within the effective power supply path based on the real-time switch status and feeder connectivity of the distribution network.

[0036] Voltage sag distribution consistency constraint is used to verify the consistency between the actual voltage sag amplitude of each node and the theoretical voltage sag amplitude inverted based on the location results;

[0037] The consistency constraint of power flow direction change is used to verify the consistency between the actual power flow change direction of each branch and the power flow change direction derived from the positioning results.

[0038] Preferably, the preset physical constraint rules are also used for constructing the loss function during the model training phase, wherein the joint loss function is used to jointly optimize the fault segment identification model and the fault distance regression model during the model training phase; the joint loss function is expressed as:

[0039]

[0040] In the formula, For the joint loss function; The cross-entropy loss term is used to identify faulty sections; This is the mean squared error loss term for fault distance regression; This is the physical constraint loss term constructed based on the preset physical constraint rules; , , These are the corresponding weighting coefficients;

[0041] The physical constraint loss term It includes a topology reachability penalty, a voltage drop distribution consistency penalty, and a power flow direction change consistency penalty, which are used to quantify the degree to which the model output violates the physical constraint rules during training.

[0042] Preferably, the residual correction includes:

[0043] A residual learning model is constructed, which adopts a fully connected neural network. Its input features include the initial value of the fault distance, the line resistance, line reactance, line length of the candidate fault section, and the changes in active power and reactive power of the distributed power sources within the section.

[0044] The residual learning model is pre-trained using historical fault samples. During the training process, the difference between the actual fault distance corresponding to the initial fault distance value and the initial fault distance value is used as the residual label, so that the residual learning model learns the mapping relationship between the initial fault distance value and the actual fault distance.

[0045] During the inference phase, the initial value of the fault distance and the corresponding segment parameters of the current fault are input into the trained residual learning model, and the residual correction amount is output. The corrected fault distance is obtained by the following formula:

[0046]

[0047] In the formula, The initial value of the fault distance is... This is the corrected fault distance.

[0048] Preferably, the final fault location result includes at least one of the following: fault feeder number, fault section number, corrected fault distance, fault type discrimination result, and location confidence, and the final fault location result is sent to the distribution automation master station, fault repair terminal, or intelligent operation and maintenance platform.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] This invention explicitly introduces distributed power source control mode identifiers into the fault feature set, enabling the model to distinguish the differentiated fault responses of inverter-type distributed power sources under different control strategies, significantly enhancing its adaptability to scenarios with high proportions of distributed power source access. By constructing a dual-stream fusion architecture of graph neural network and temporal neural network with attention mechanism, it fully explores the correlation and propagation laws and key temporal evolution features of fault information in the topological space, achieving refined identification of fault sections and fault distances, resulting in higher positioning accuracy. By introducing physical constraint loss terms including topological reachability, voltage drop consistency, and power flow direction consistency constraints during the training phase, and performing multi-dimensional physical consistency verification and residual correction on the preliminary positioning results during the inference phase, it ensures that the positioning results conform to both data statistical laws and basic physical laws of the power system, effectively avoiding illogical misjudgments. Through a two-stage progressive positioning strategy of first identifying sections and then estimating distances, combined with residual learning compensation, an end-to-end complete solution is formed, which can directly support rapid emergency repair and intelligent self-healing decision-making in distribution networks, demonstrating high engineering practical value. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0053] Example 1

[0054] This embodiment provides an intelligent fault location method for distribution networks with a high proportion of distributed power sources. This method is applicable to active distribution networks containing inverter-type distributed power sources such as distributed photovoltaic, distributed wind power, and energy storage devices. The high proportion of distributed power sources leads to a more complex distribution of fault currents and bidirectional power flow in the distribution network. Traditional fault location methods based on single-source radial networks have significantly reduced adaptability in this scenario. This invention effectively solves the above problems by systematically characterizing the differentiated dynamic response features of distributed power sources under multiple control modes, constructing a deep learning model that integrates spatial coupling and dynamic evolution, and embedding a physical constraint two-layer verification mechanism.

[0055] Figure 1A flowchart illustrating the intelligent fault location method provided in this embodiment of the invention is shown. The steps of this method have a close logical progression: first, feature engineering provides sufficient information for the model; then, a dual-stream fusion architecture is used to accurately identify the faulty section; next, fine ranging is performed within the identified section; and finally, physical verification and residual correction ensure the reliability and accuracy of the output results. The method includes the following steps:

[0056] Step S1: Obtain multi-source operation data of the distribution network and establish a fault analysis model to extract a fault feature set containing the dynamic response characteristics of distributed power sources.

[0057] This step serves as the data foundation and feature engineering step for the entire localization method. Its purpose is to extract discriminative feature representations from multi-source heterogeneous data to provide high-quality input for subsequent deep learning models. In scenarios with a high proportion of distributed power supply access, the fault response behavior of distributed power supply differs significantly from that of traditional synchronous power supply. Therefore, this difference must be explicitly characterized at the feature level; otherwise, subsequent models will find it difficult to accurately perceive the evolution of fault features under different control modes.

[0058] Specifically, the following multi-source operational data are obtained from the distribution network energy management system, distribution automation system, and distributed power source monitoring system:

[0059] (1) Distribution network topology data, including bus information, feeder connection relationship, tie switch information, transformer connection relationship, and distributed power source access location and rated capacity information;

[0060] (2) Measurement data of distribution automation terminals, including three-phase voltage amplitude, three-phase current amplitude, active power, reactive power and voltage phase angle data at substation outlet, feeder section switch, feeder terminal unit, fault indicator installation point, distributed power grid connection point and key load node;

[0061] (3) Switch status data, including the real-time open / closed status of circuit breakers, sectionalizing switches, and tie switches;

[0062] (4) Distributed power source operation data, including the active power output, reactive power output, output current amplitude and current control mode information of each distributed power source.

[0063] Based on the obtained line impedance parameters, branch connection relationships, and distributed power source access status, a distribution network fault analysis model that takes into account the influence of multiple power sources is established to characterize the voltage, current, and power flow distribution patterns under fault conditions.

[0064] The collected multi-source heterogeneous measurement data are preprocessed as follows: First, a unified time reference is used to align the multi-source data, eliminating clock deviations between the acquisition devices and ensuring time consistency in subsequent time series analysis. Second, a sliding window method is used to extract electrical quantity data within a preset time window before and after the fault. The preset time window is preferably from 5 cycles before the fault to 10 cycles after the fault, to take into account both the reference value of steady-state characteristics before the fault and the discriminative value of transient characteristics after the fault. Third, the Laida criterion is used to identify and remove outliers, and cubic spline interpolation is used to repair missing data, ensuring data integrity and continuity. Finally, the features of each dimension are subjected to maximum and minimum normalization to eliminate dimensional differences between different electrical quantities and avoid adverse effects of numerical range differences on the training of deep learning models.

[0065] In the fault feature extraction stage, in addition to extracting conventional node voltage drop features, branch current mutation features, zero-sequence component features, negative-sequence component features, transient energy features, and harmonic features, this invention specifically extracts the dynamic response features of distributed generation. The fundamental reason for introducing this feature is that the response behavior of inverter-type distributed generation to the same fault differs significantly under different control modes. For example, distributed generation under constant power control mode tries to maintain a constant output power during a fault, while distributed generation under low voltage ride-through control mode prioritizes providing reactive power support to the grid. If the model cannot perceive this difference in control modes, it will be difficult to accurately analyze the source and distribution of fault current. Therefore, the dynamic response feature of distributed generation is defined as follows:

[0066]

[0067] In the formula:

[0068] , which represents the change in active power before and after a distributed power source failure;

[0069] , which is the change in reactive power;

[0070] , which is the change in output current;

[0071] This is a control mode identifier used to explicitly distinguish the control strategy currently being executed by the distributed power source;

[0072] In this context, the superscripts pre and fault represent the steady-state value before the fault and the value during the fault period, respectively.

[0073] Control mode identifier The mode is represented by a single-hot encoding method. For example, the constant power control mode is identified as [1,0,0,0], the constant current control mode is identified as [0,1,0,0], the reactive power support control mode is identified as [0,0,1,0], and the low voltage ride-through control mode is identified as [0,0,0,1].

[0074] The node voltage sag characteristics are represented as follows:

[0075] In the formula, in the formula, For nodes Voltage drop and They are nodes Voltage amplitude before and during the fault. This characteristic directly reflects the degree of impact of the fault on the voltage of each node in the distribution network and is the basis for judging the fault section.

[0076] The characteristics of sudden changes in branch current are represented as follows:

[0077]

[0078] In the formula, in the formula, branch road The rate of change of current, and Branch roads The current amplitude before and during the fault. This feature can effectively reflect the degree of abrupt change in fault current and helps to distinguish between faulty and non-faulty branches.

[0079] Combine all the above features into a complete fault feature vector. ,

[0080] For use in subsequent models; where, Characteristics of node voltage variation. Characteristics of branch current variation For the characteristics of the ordered components, Transient spectral characteristics, The characteristics of harmonic energy distribution This refers to the dynamic response characteristics of distributed power sources.

[0081] The fault feature set constructed in this step not only retains the effective information of traditional electrical quantity features, but also explicitly introduces the differences in control modes of distributed power sources, laying a feature foundation for the deep learning model to accurately perceive fault scenarios in step S2.

[0082] Step S2: Construct a graph structure model based on the distribution network topology, use graph neural networks to extract spatial coupling features that represent the association between node and branch faults, and use a temporal neural network with attention mechanism to extract fault dynamic evolution features; after fusing the two types of features, identify fault sections and obtain candidate fault sections.

[0083] This step is one of the core components of the present invention, and its design purpose is to solve the technical problem of complex spatial distribution of fault characteristics and difficulty in capturing the temporal evolution law in scenarios with a high proportion of distributed power access.

[0084] Traditional single-dimensional analysis methods—whether using only topology or only time-series data—are insufficient to fully depict the complete picture of faults. In distribution networks, faults manifest as spatial disturbances propagating along the topology and as transient processes evolving over time. These two aspects are coupled and complementary. Therefore, this step constructs a dual-stream extraction and fusion architecture for spatial coupling characteristics and dynamic evolution characteristics, enabling the model to understand fault behavior simultaneously from both spatial and temporal dimensions.

[0085] First, the distribution network is abstracted as a graph structure model G=(V,E). The node set V includes: bus nodes, load nodes, distributed generation access nodes, measurement nodes, and switch nodes. The edge set E includes: feeder branches, tie branches, and transformer connection branches.

[0086] The reason for using a graph structure instead of a vector sequence to model the distribution network is that the distribution network is essentially a network with a clear topological connection relationship. The electrical connection between nodes is limited by the physical connection. The graph structure can naturally express this relationship in a non-Euclidean space, while traditional feedforward neural networks cannot directly handle this topological constraint.

[0087] For each node Construct node attribute vectors ∈V This includes the three-phase voltage amplitude, three-phase current amplitude, injected active power, injected reactive power, voltage phase angle, and switch status of the node. If it is a distributed generation access node, it also includes the operating status of the distributed generation. This is for each edge. Construct edge attribute vectors for ∈E It includes information on line resistance, line reactance, line length, and branch type.

[0088] This embodiment employs a graph convolutional network to extract the spatial coupling features of the distribution network. The core advantage of a graph convolutional network lies in its message propagation mechanism: the features of each node are updated through weighted aggregation of the features of its neighboring nodes. After multiple layers are stacked, each node can perceive the state changes of its neighbors within a multi-hop range, thereby propagating local fault disturbance information along the topology to the entire network. The inter-layer propagation process of a graph convolutional network is represented as follows:

[0089]

[0090] In the formula: The adjacency matrix after adding self-loops, This is the original adjacency matrix. Using an identity matrix, introducing self-loops allows nodes to retain their own information when aggregating neighbor features, preventing their own state from being diluted during propagation; for degree matrix For the first Layer input feature matrix, initial input This is a matrix of node attributes. For the first Layer-learnable weight matrix; For non-linear activation functions, this embodiment uses the ReLU function.

[0091] go through After the layer graph convolution propagation, the global average pooling of the features of all nodes in the last layer is taken as the spatially coupled feature vector. :

[0092]

[0093] This process, through a multi-layered message propagation mechanism, fully propagates and aggregates the local voltage and current disturbance information at the fault location along the distribution network topology, enabling... It can characterize the deep spatial relationships between nodes, branches, and distributed power access points and surrounding networks. For example, when a feeder segment fails, the voltage drop characteristics of its upstream and downstream nodes will be automatically associated through a graph convolutional network, and the current output characteristics of the distributed power access point will also affect adjacent nodes during the aggregation process.

[0094] Meanwhile, this embodiment employs a Long Short-Term Memory (LSTM) network with an attention mechanism to extract the dynamic evolution features of the fault. The reason for introducing a temporal network is that a fault is a dynamic evolution process. The transient impact at the moment of the fault, the response delay of the distributed power supply control mode, and the action sequence of the protection device all unfold in the time dimension, and spatial features alone cannot capture this temporal information. The temporal feature sequences of each measurement node within the time window before and after the fault in step S1 are input into the LSTM network. The LSTM unit effectively captures the long-term and short-term dependencies of electrical quantities during the fault process through the synergistic effect of forget gates, input gates, and output gates. It can both remember the steady-state baseline before the fault and track the transient evolution after the fault.

[0095] To further highlight the contribution of critical fault moments, an attention mechanism is introduced. The rationale for this mechanism is that the contribution of each moment in the fault evolution process to the localization task is not equal—critical frames such as the sudden current change at the moment of fault occurrence and the switching of distributed power supply control modes contain richer discriminative information, while the information redundancy in the steady-state period is relatively high. By adaptively assigning higher weights to critical frames through the attention mechanism, the discriminative power of feature representation can be significantly improved. The formula for calculating the attention weight at each moment is as follows:

[0096]

[0097] In the formula, for The hidden state vector at time t. , and For learnable parameters, for Attention weight at any given moment; The dot product operation maps the vector output by the tanh function to a scalar fraction. , used to measure the The importance of the hidden state at any given time for fault location tasks;

[0098] The fault dynamic evolution characteristics are obtained by weighting and summing the hidden state at each time step based on the attention weights.

[0099]

[0100] In the formula, This represents the feature vector of fault dynamic evolution. This represents the total number of time points. For time indexing.

[0101] The attention mechanism enables the model to adaptively highlight key frames during the fault occurrence and transient processes—such as the instant of sudden change in fault current or the moment of switching of distributed power supply control modes—while suppressing interference from irrelevant or redundant moments.

[0102] Subsequently, spatial coupling features With dynamic evolutionary characteristics Feature concatenation and fusion are performed. The reason for using concatenation fusion instead of weighted fusion is that spatial and temporal features belong to different modalities; they are complementary rather than substitutable. The concatenation operation can completely preserve all the information from both feature streams, allowing the subsequent classification layer to learn the optimal combination weights. The fused feature vector is: ;

[0103] Will The input consists of a classification layer composed of a fully connected layer and a softmax function, and the output is the probability distribution of each candidate segment being a faulty segment:

[0104]

[0105] In the formula, ∈{1,2,…,M} represents the candidate fault segment number. This represents the total number of candidate segments; and The first The classification layer weight vector and bias term corresponding to each segment; select the segment corresponding to the maximum probability or the segment whose probability exceeds a preset threshold as the candidate fault segment, and output its corresponding confidence level. The preset threshold can be set to 0.5 for example.

[0106] Through this step, the model completes the end-to-end mapping from the original measurement data to the fault segment identification result. The output candidate fault segments limit the spatial range of the fine ranging in step S3, effectively narrowing the target domain for subsequent calculations.

[0107] Step S3: For the candidate fault section, combine the local dynamic response information of the distributed power source within the section to estimate the location of the fault point and obtain the initial value of the fault distance.

[0108] This step follows the candidate fault segments output from step S2, and performs fine-grained fault location within the defined segment range. The two-stage design of first identifying segments and then estimating distances has clear engineering rationale: if fault point location is performed directly across the entire network, the search space is too large and the mapping relationship between features and fault distances is highly nonlinear, making it difficult for the model to converge. However, by decomposing the localization task into two sub-tasks, "coarse-grained segment identification" and "fine-grained distance estimation," the objectives of each sub-task are more focused, the model complexity is significantly reduced, and the localization accuracy is improved.

[0109] For each candidate fault segment output in step S2, construct the fault distance regression input feature vector. The input features include:

[0110] Three-phase voltage amplitude at the beginning of the target candidate fault section and three-phase current amplitude ;

[0111] Three-phase voltage amplitude at the end of the target candidate fault section and three-phase current amplitude ;

[0112] The line resistance R, line reactance X, and line length L of the target candidate fault section;

[0113] Local dynamic response information of distributed power sources within the section and at adjacent nodes, specifically the change in active power. and reactive power change .

[0114] The local dynamic response information of distributed power sources is introduced here. The physical basis is that the path impedance of the fault current flowing through the grid connection point of the distributed power source is different depending on the location of the fault point, which leads to different voltage drop perceived by the distributed power source, and thus affects the change in its output power. The closer the distributed power source is to the fault point, the more significant its power change is usually. Therefore, the local power change implies the physical information of the relative position of the fault point. Using it as a regression feature can effectively improve the ranging accuracy.

[0115] The constructed regression input feature vector The input is fed into a pre-trained fault distance regression model. In this embodiment, the fault distance regression model employs a three-layer fully connected neural network. The first two layers use the ReLU activation function to enhance nonlinear expressive power, and the output layer uses the Sigmoid activation function to constrain the output within the range of 0 to 1. The regression model outputs a normalized distance ratio. ∈(0,1):

[0116]

[0117] Among them, The complete set of learnable parameters for the fault distance regression model, including the weight matrices and bias vectors of each layer in the network, is determined through training with historical fault samples. The nonlinear mapping function characterized by the fault distance regression model is usually implemented by a multi-layer fully connected neural network;

[0118] The initial value of the fault distance is:

[0119] In the formula, This represents the estimated distance from the fault point to the beginning of the target candidate fault segment. This step achieves a refined transition from segment-level localization to point-level localization, providing an initial estimation benchmark for the verification and correction in step S4.

[0120] Step S4: Based on the preset physical constraint rules, perform physical consistency verification on the candidate fault segment and the initial value of the fault distance, and perform residual correction on the initial value of the fault distance that passes the verification, and output the final fault location result.

[0121] The purpose of this step is to compensate for the inherent defects of pure data-driven models. Although deep learning models are outstanding in fitting complex nonlinear mappings, they are essentially statistical approximations of the training data distribution and lack an explicit understanding of the basic physical laws of power systems. Under certain extreme operating conditions or edge scenarios of data distribution, the model may output erroneous results that violate Kirchhoff's laws, violate topological connectivity, or do not meet the power flow distribution law. Therefore, it is necessary to introduce a physical constraint verification step after the model output to identify and correct unreasonable results.

[0122] This embodiment constructs a multi-dimensional physical constraint verification system, including the following three types of constraints:

[0123] Topology reachability constraint: Based on the real-time switch status and feeder connectivity of the distribution network, determine whether a candidate fault section is within the valid power supply path under the current operating mode. If a candidate fault section is in an islanded or out-of-service state due to a switch being open, the verification fails, and the candidate section is marked as unreliable. The physical basis of this constraint is that faults can only occur on energized sections electrically connected to the power source; short circuits or grounding faults cannot occur in islanded or out-of-service sections.

[0124] Voltage sag distribution consistency constraint: Calculate the actual voltage sag amplitude at each measurement node. Assuming the fault occurs at the estimated location, the theoretical voltage drop distribution is inverted using a forward-backward power flow calculation method. Calculate the root mean square error between the two:

[0125]

[0126] like If the value exceeds a preset threshold, such as 0.05, the verification fails. The physical basis of this constraint is that, given the fault location and fault type, the voltage drop distribution of each node in the distribution network is uniquely determined, and the location result output by the model should be able to reproduce the actual observed voltage distribution through circuit equations.

[0127] Consistency constraint for power flow direction changes: Monitor the changes in power flow direction of each branch before and after a fault, and compare them with the theoretical power flow change direction derived from the candidate fault location. Define the number of branches with inconsistent power flow direction changes as follows: ,like If the proportion exceeds a preset threshold, such as 10% of the total number of monitored branches, the verification will fail. The physical basis of this constraint is that after a fault occurs, the power flow direction changes upstream and downstream of the fault point in a definite manner—the current amplitude upstream of the fault point increases but the direction remains unchanged, while the current downstream of the fault point may reverse or disappear. The model output must be consistent with this pattern.

[0128] The positioning results that pass the above physical consistency verification proceed to the residual correction stage. The reason for setting up the residual correction stage is that although deep learning regression models generally have high ranging accuracy, they still have systematic biases under specific working conditions. For example, the estimation error for faults at the end of a section is often greater than that for faults in the middle of the section. Compensating for the systematic errors of the initial estimates using a residual learning model can further improve positioning accuracy. This embodiment constructs a residual learning model, which uses a lightweight fully connected neural network. Its input features include: the initial value of the fault distance. Line resistance of candidate fault sections Line reactance Line length and the change in active power of distributed power sources within the section. and reactive power change .

[0129] The residual learning model is pre-trained using historical fault samples; during training, the initial fault distance values ​​output by the fault distance regression model from the historical fault samples are used. Distance to the corresponding actual fault difference As residual labels, they enable the residual learning model to learn the mapping relationship between the initial fault distance and the actual fault distance.

[0130] During the inference phase, the initial value of the fault distance of the current fault is set. The corresponding segment parameters are input into the trained residual learning model, which then outputs the residual correction amount. The corrected fault distance is:

[0131]

[0132] In the formula, This is the final, precise location of the fault point. Through the verification and correction in this step, the final location result conforms to both data-driven statistical laws and follows the physical constraints of the power system, achieving a balance between statistical accuracy and physical rationality.

[0133] Step S5: Output the final fault location result.

[0134] This step serves as the output of the entire method. It standardizes and encapsulates the location information obtained through the progressive processing of the preceding steps and distributes it to various application terminals, thus completing a closed loop from data collection to decision support.

[0135] The final fault location result, after verification and correction, includes the fault feeder number, fault section number, and corrected fault distance. The fault type identification result and location confidence level are encapsulated into a standardized message and sent to the distribution automation master station, fault repair terminal, or intelligent operation and maintenance platform via a communication network to support rapid fault isolation, repair path planning, and system self-healing control decisions. The fault type identification result may be, for example, single-phase grounding, two-phase short circuit, or three-phase short circuit.

[0136] Example 2

[0137] The difference between this embodiment and Embodiment 1 lies in the use of a joint loss function to jointly optimize the fault segment identification model and the fault distance regression model during the model training phase. The purpose of setting up the joint loss function is that, in traditional phased training methods, the segment identification model and the distance regression model are optimized independently, lacking a unified constraint on physical laws. This may lead to physically inconsistent results output by the two models during the inference phase. By introducing a physical constraint loss term during the training phase, the model is guided by physical laws during parameter learning, and after training, it can output more physically consistent positioning results without additional post-processing.

[0138] The expression for the joint loss function is:

[0139]

[0140] The loss terms in the formula are defined as follows: Cross-entropy loss term for faulty sections:

[0141]

[0142] In the formula, The number of samples in the training batch. This represents the total number of candidate fault sections. For the first One-hot encoded labels for the real fault sections of each sample. The model predicts the first The sample belongs to the first The probability of each segment; the purpose of this loss term is to make the segment probability distribution output by the model approximate the true label, thus ensuring the accuracy of segment identification.

[0143] The mean squared error loss term for fault distance regression:

[0144]

[0145] In the formula, The ratio of distances predicted by the model. This is the ratio of the predicted distance to the actual distance. The purpose of this loss is to minimize the deviation between the predicted distance and the actual distance, thus ensuring ranging accuracy.

[0146] The physical constraint loss term is constructed based on preset physical constraint rules. Its function is to quantify the degree to which the model output violates physical laws during training, and to backpropagate this degree of violation as part of the optimization objective, thereby guiding the model parameters to converge in the direction that satisfies physical laws.

[0147] It consists of the following three penalty items:

[0148] Topological reachability penalty For candidate fault sections with high output probabilities from the model, if they are not in the current switching state, then the set of reachable sections is considered. If it occurs within the internal system, then punishment will be imposed:

[0149]

[0150] This penalty term forces the model to reduce the prediction probability for unreachable sections.

[0151] Voltage sag distribution consistency penalty The difference between the theoretical voltage drop distribution derived from the fault location predicted by the calculation model and the actual voltage drop distribution:

[0152]

[0153] This penalty term forces the model to output fault locations that can reproduce the actual voltage distribution through circuit equations.

[0154] Consistency penalty for changes in trend direction The number of branches whose power flow direction does not match the theoretical derivation:

[0155] In the formula, , , Here, K represents the weighting coefficient for each penalty item, and K represents the total number of branches. For indicator functions; , , In this embodiment, the corresponding weighting coefficients are set to 1.0, 1.0, and 0.5, respectively. The setting is slightly lower than the previous two in order to prioritize data fitting and use physical constraints as a supplement in the early stages of training, so as to avoid the model from having difficulty converging due to excessive physical constraints.

[0156] During training, the Adam optimizer is used to iteratively update the model parameters. By backpropagating the joint loss function during the training phase, the model parameters are forced to converge in the direction that satisfies topological reachability, voltage drop distribution consistency and power flow direction change consistency while fitting the distribution of historical fault data. This results in outputting positioning results that are more in line with the physical laws of the power system during the inference phase.

[0157] The remaining steps are the same as in Example 1, and will not be repeated here.

[0158] Example 3

[0159] This embodiment demonstrates the application process of the method of the present invention in a real power distribution network fault scenario, and verifies the logical connection and technical effect between each step through a complete fault location example.

[0160] Suppose a single-phase ground fault occurs in a 10kV active distribution network, which is connected to multiple distributed photovoltaic power generation systems. After the distribution automation system detects the fault, it triggers the execution of the method of this invention:

[0161] First, the system automatically collects three-phase voltage and current waveform data for a total of 15 cycles before and after the fault, while simultaneously acquiring data on the distribution network topology, line parameters, switch status, and operational data of each distributed power source. After data preprocessing, node voltage sag characteristics, branch current abrupt change characteristics, sequence component characteristics, and distributed power source dynamic response characteristics are extracted. Among these, two distributed photovoltaic systems near the fault time are in constant power control mode and low voltage ride-through control mode, respectively, and their feature vectors contain... These are encoded as [1,0,0,0] and [0,0,0,1], respectively. This difference in control mode is explicitly preserved through feature extraction in step S1, providing a key basis for the subsequent model to distinguish the fault response behaviors of the two distributed power sources.

[0162] Secondly, the distribution network topology is modeled as a graph structure, and a pre-trained graph convolutional network is used to extract spatial coupling features. Simultaneously, time-series data is input into an LSTM network with an attention mechanism to extract dynamic evolution features. The attention mechanism adaptively assigns higher weights to the 3rd to 6th cycles after the fault, which happens to be the stage where the distributed power source control mode response and fault current transient changes are most significant. This phenomenon verifies the effectiveness of the attention mechanism setting—the model can indeed automatically focus on the moment most discriminative for the localization task.

[0163] Then, the spatial coupling features and dynamic evolution features were fused and input into the classification layer, identifying the candidate fault segment as the third segment of feeder F2, i.e., between nodes 12 and 15, with a confidence level of 0.93. The high confidence level indicates that the model has sufficient confidence in determining that this segment is a fault segment.

[0164] Next, the voltage and current data at the beginning and end of the section, as well as the local active and reactive power changes of the distributed power sources within the section, are extracted and input into the fault distance regression model. The initial fault distance is obtained as approximately 1.2 km from node 12, and the total length of the section is 3.5 km. Since step S2 has narrowed the positioning range to a single section, the regression model in step S3 only needs to be estimated within a 3.5 km range, significantly reducing the search space and improving ranging accuracy.

[0165] Subsequently, the system performs a physical consistency check: the section is within the effective power supply path under the current switching state, so the topology reachability check passes; the root mean square error between the actual voltage drop distribution and the inverted distribution at each node is 0.03, which is less than the threshold of 0.05, so the voltage drop consistency check passes; the power flow direction change is consistent with the theoretical derivation, so the power flow direction consistency check passes. All three checks pass, indicating that the model's output positioning results are completely consistent with the physical laws of the power system. After the checks pass, the residual learning model outputs a correction amount. The corrected fault distance was obtained. =1.12km. The residual correction step further eliminated the systematic bias of the regression model;

[0166] Finally, the system outputs the fault location results: faulty feeder F2, faulty section number S203, fault distance 1.12km, fault type single-phase grounding, and location confidence level 0.93. This result is pushed to the mobile terminal of the emergency repair team in real time, guiding repair personnel to quickly reach the fault location.

[0167] This embodiment verifies that in complex scenarios with a high proportion of distributed power supply access, the steps of the present invention work together in a coordinated and progressive manner, ultimately realizing a complete processing link from multi-source data to accurate positioning results. The positioning error is within an acceptable range, demonstrating significant engineering application value.

[0168] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart fault location method for distribution network with high penetration of distributed generation, characterized in that, Includes the following steps: Acquire multi-source operation data of the distribution network and establish a fault analysis model to extract a fault feature set that includes the dynamic response characteristics of distributed power sources; The dynamic response characteristics include electrical quantity change characteristics that characterize changes in operating status before and after a fault, and control mode identifiers that distinguish distributed power supply control strategies. Based on the distribution network topology, a graph structure model is constructed. Graph neural networks are used to extract spatial coupling features that characterize the association between node and branch faults, and temporal neural networks with attention mechanisms are used to extract the dynamic evolution features of faults. The spatial coupling features and the dynamic evolution features are fused together, and the fault segments are identified based on the fused features to obtain candidate fault segments; For the candidate fault section, the location of the fault point is estimated by combining the local dynamic response information of the distributed power source within the section, and the initial value of the fault distance is obtained. Based on preset physical constraint rules, the candidate fault segments and initial fault distance values ​​are physically consistent, and residual corrections are performed on the initial fault distance values ​​that pass the verification, and the final fault location result is output.

2. The intelligent fault location method for distribution network with high penetration of distributed generation according to claim 1, characterized in that: The dynamic response characteristics of the distributed power source are expressed as follows: ; In the formula, This represents the dynamic response feature vector of a distributed power source. This represents the change in active power before and after a distributed power source failure. This represents the change in reactive power before and after a distributed power source failure. This represents the change in output current before and after a distributed power source failure. This is an identifier for the distributed power supply control mode. The control mode identifier is used to indicate the current control mode of the distributed power source, and the control mode includes at least one of the following: constant power control mode, constant current control mode, reactive power support control mode, and low voltage ride-through control mode.

3. The intelligent fault location method for high-proportion distributed power generation access to distribution networks according to claim 1, characterized in that: The method of extracting spatial coupling features representing the association between nodes and branch faults using graph neural networks includes: The bus nodes, load nodes, distributed generation access nodes, measurement nodes and switch nodes in the distribution network are used as nodes in the graph structure model, and feeder branches, tie branches and transformer connection branches are used as edges. Attribute matrices are constructed for each node and edge. The spatial coupling characteristics are obtained by propagating and aggregating local fault disturbance information along the distribution network topology through the message propagation mechanism of a multi-layer graph neural network; wherein, the inter-layer propagation process of the graph neural network is represented as follows: ; In the formula, For the first The feature matrix output by the layer after message propagation and feature transformation. The adjacency matrix after adding self-loops, For degree matrix, For the first Layer input feature matrix, For the first Layer weight matrix, This is the activation function.

4. The intelligent fault location method for distribution network with high penetration of distributed generation according to claim 1, characterized in that: The extraction of fault dynamic evolution features using a temporal neural network with an attention mechanism includes: Long Short-Term Memory (LSTM) networks or gated recurrent unit (GRU) networks are used to encode the time-series measurement data within a preset time window before and after the fault, thereby obtaining the hidden state vector at each time point. The attention weights of the hidden state at each time step are calculated using an attention mechanism. The formula for calculating the attention weights is as follows: ; In the formula, for The hidden state vector at time t. , and For learnable parameters, for Attention weight at any given moment; The dot product operation maps the vector output by the tanh function to a scalar fraction. , used to measure the The importance of the hidden state at any given time for fault location tasks; The fault dynamic evolution characteristics are obtained by weighting and summing the hidden state at each time step based on the attention weights. ; In the formula, This represents the feature vector of fault dynamic evolution. This represents the total number of time points. For time indexing.

5. The intelligent fault location method for distribution network with high penetration of distributed generation according to claim 1, characterized in that: The step of fusing the spatial coupling features with the dynamic evolution features, and identifying fault segments based on the fused features, includes: Coupling the spatial coupling feature vector output by the graph neural network with the dynamic evolution feature vector output by the temporal neural network Performing feature splicing to obtain a fusion feature vector ; The fused feature vector is input into the classification layer, which outputs the probability distribution of each candidate segment as a fault segment, and selects the segment corresponding to the maximum probability or the segment whose probability exceeds a preset threshold as the candidate fault segment.

6. The intelligent fault location method for distribution network with high penetration of distributed generation according to claim 1, characterized in that: The estimation of the fault location by combining the local dynamic response information of distributed power sources within the section includes: Extract the voltage and current measurements at the beginning and end of the candidate fault section, and combine them with the line resistance, line reactance, line length of the candidate fault section, as well as the changes in active power and reactive power of the distributed power sources within the section, to construct a regression input feature vector. The regression input feature vector is input into a pre-trained fault distance regression model, and the ratio of the distance from the fault point to the beginning of the segment to the total length of the segment is output as the initial value of the fault distance.

7. The intelligent fault location method for distribution network with high penetration of distributed generation according to claim 1, characterized in that: The preset physical constraint rules include at least one of the following: Topology reachability constraints are used to verify whether the candidate fault section is within the effective power supply path based on the real-time switch status and feeder connectivity of the distribution network. Voltage sag distribution consistency constraint is used to verify the consistency between the actual voltage sag amplitude of each node and the theoretical voltage sag amplitude inverted based on the location results; The consistency constraint of power flow direction change is used to verify the consistency between the actual power flow change direction of each branch and the power flow change direction derived from the positioning results.

8. The intelligent fault location method for distribution network with high penetration of distributed generation according to claim 7, characterized in that: The preset physical constraint rules are also used to construct the loss function during the model training phase. During the model training phase, a joint loss function is used to jointly optimize the fault segment identification model and the fault distance regression model. The joint loss function is expressed as follows: ; In the formula, is a joint loss function; is a cross-entropy loss term for fault section identification; is a mean square error loss term for fault distance regression; is a physical constraint loss term constructed based on the preset physical constraint rule; , , is a corresponding weight coefficient; the physical constraint loss term including a topological reachability penalty term, a voltage drop distribution consistency penalty term, and a power flow direction change consistency penalty term, for quantifying the degree to which the model output violates the physical constraint rules during the training process.

9. The intelligent fault location method for distribution network with high penetration of distributed generation according to claim 1, characterized in that: The residual correction includes: A residual learning model is constructed, which adopts a fully connected neural network. Its input features include the initial value of the fault distance, the line resistance, line reactance, line length of the candidate fault section, and the changes in active power and reactive power of the distributed power sources within the section. The residual learning model is pre-trained using historical fault samples. During the training process, the difference between the actual fault distance corresponding to the initial fault distance value and the initial fault distance value is used as the residual label, so that the residual learning model learns the mapping relationship between the initial fault distance value and the actual fault distance. During the inference phase, the initial value of the fault distance and the corresponding segment parameters of the current fault are input into the trained residual learning model, and the residual correction amount is output. The corrected fault distance is obtained by the following formula: ; In the formula, is the initial value of the fault distance, is the corrected fault distance.

10. The intelligent fault location method for high-proportion distributed power generation access to distribution networks according to claim 1, characterized in that: The final fault location result includes at least one of the following: fault feeder number, fault section number, corrected fault distance, fault type discrimination result, and location confidence. The final fault location result is then sent to the distribution automation master station, fault repair terminal, or intelligent operation and maintenance platform.

Images