A power distribution network fault diagnosis method based on multi-source data reconstruction and causal reasoning

Abstract

The application discloses a power distribution network fault diagnosis method based on multi-source data reconstruction and causal reasoning, and belongs to the technical field of electric power. The method aims at the problems of noise, missing and asynchrony of multi-source data of the power distribution network, insufficient physical consistency of fault characteristics, weak generalization ability of the diagnosis model and poor interpretability. Firstly, data repair and space-time alignment are realized through a physically constrained diffusion generative model. Then, node steady-state characteristics integrated with dynamic topology and electrical rules are extracted by using a physically guided space-time state network. Subsequently, fault relationship modeling is performed through physical semantic contrast learning and a dynamic hypergraph prototype network. Finally, integrated and interpretable diagnosis of fault detection, classification, positioning and cause explanation is realized based on a causal concept bottleneck reasoning framework, and the generalization ability of the model is improved by using a meta-learning mechanism. The application improves the accuracy, adaptability and decision transparency of the power distribution network fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of power technology, and in particular relates to a method for fault diagnosis of distribution networks based on multi-source data reconstruction and causal reasoning. Background Technology

[0002] With the continuous advancement of new power system construction, the form and operation mode of distribution networks are undergoing profound changes. The high proportion of distributed power sources such as wind turbines and photovoltaics has resulted in significant multi-source heterogeneity in distribution network operation data. This data originates from various acquisition devices, including synchronous phasor measurement units, monitoring and data acquisition systems, advanced measurement systems, and meteorological systems. In practical applications, due to factors such as equipment accuracy, communication latency, and environmental interference, these multi-source measurement data commonly suffer from prominent problems during acquisition, such as noise interference, missing key information fragments, and the inability to strictly synchronize timestamps between different sources. These data quality issues severely interfere with the complete extraction and effective fusion of fault features, posing fundamental difficulties for subsequent fault analysis and diagnosis.

[0003] In terms of fault feature extraction and characterization, existing methods for fault diagnosis in distribution networks often fail to organically combine the physical topology of the power grid, its inherent electrical physical laws, and the real-time operating status of the system. Many methods rely primarily on data-driven pattern recognition, resulting in extracted fault features that, while statistically diverse, lack sufficient physical consistency. In other words, the features are not strongly correlated with the actual physical properties such as current, voltage, power, and topological connections. This lack of physical consistency makes the feature representation insufficiently accurate in describing the transient process of the fault, easily leading to misjudgments in complex fault or strong interference scenarios.

[0004] Furthermore, at the level of fault relationship modeling and propagation analysis, most existing methods rely on a pre-defined, fixed distribution network topology to construct fault association models. However, in actual operation, distribution networks often undergo topology changes due to load switching, fault isolation, distributed generation switching, or network reconfiguration. Models relying on fixed topologies struggle to effectively reflect fault propagation paths and coupling relationships under such dynamic operating conditions, especially in depicting the complex electrical couplings and fault interactions between multiple feeders. This results in a significant decrease in the model's diagnostic performance and weak generalization ability in scenarios after topology changes.

[0005] Furthermore, most advanced fault diagnosis models, such as those based on deep learning, remain "black box" structures. While they can output a fault diagnosis result through complex nonlinear mappings, they completely lack the ability to explain the decision-making process. Maintenance personnel cannot know the basis for the model's specific judgment, such as the cause of the fault, its evolution path, and key evidence. This lack of interpretability makes it difficult for maintenance personnel to fully trust the diagnostic results and fails to provide direct, physically meaningful decision support for post-fault recovery operations or hazard mitigation, significantly limiting the practical application value and feasibility of these diagnostic technologies in intelligent operation and maintenance decision-making for distribution networks. Solving these interconnected challenges—including data quality, physical characteristics, model dynamic adaptability, and decision interpretability—is crucial to improving the level of distribution network fault diagnosis technology. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a power distribution network fault diagnosis method based on multi-source data reconstruction and causal reasoning, thereby resolving the issues present in the prior art.

[0007] Firstly, to achieve the above objectives, the present invention provides a method for fault diagnosis of power distribution networks based on multi-source data reconstruction and causal reasoning, comprising the following steps: Acquire multi-source heterogeneous operation data of the power distribution network; Based on the constraints of physical laws, the multi-source heterogeneous operating data is repaired and spatiotemporally aligned to generate a physically consistent data matrix. Based on the physically consistent data matrix, steady-state feature vectors of distribution network nodes are extracted by fusing dynamic topology and electrical physical laws into a state-space network. Based on the steady-state feature vector, fault relationship modeling is performed through physical semantic contrast learning and dynamic hypergraph to generate fault prototypes and enhanced features with fused topological structures. Based on the enhanced features and fault prototypes, fault diagnosis is performed through a causal conceptual bottleneck reasoning framework, and the output is a decision result that includes fault type, location and explanation.

[0008] Optionally, the process of repairing and spatiotemporally aligning the multi-source heterogeneous operational data includes: A conditional diffusion model is constructed, which adds noise to the data through a forward diffusion process and reconstructs the data through a reverse generation process; Kirchhoff's laws and power balance equations are embedded in the loss function of the reverse generation process as physical constraints. In the reverse generation process, latent variable interpolation and gating mechanisms are used to achieve time alignment of different source data; An iterative refinement strategy is adopted, and the data is corrected by a physical verification module after each generation step, outputting a physically consistent data matrix.

[0009] Optionally, the process of extracting the steady-state feature vectors of distribution network nodes by fusing dynamic topology and electrical physical laws into a state-space network includes: Based on the real-time operating status of the distribution network, calculate the dynamic coupling relationship between nodes to construct a dynamic spatiotemporal diagram; Construct a selective state-space model and generate the state transition matrix and input mapping matrix of the model based on the electrical and physical properties of the nodes. Based on the dynamic spatiotemporal graph and the parameterized selective state space model, the spatiotemporal evolution and fusion of the hidden states of nodes are carried out. The evolved node features are subject to regularization constraints based on the consistency of the power flow equations, and the steady-state feature vectors of the nodes are aggregated.

[0010] Optionally, the process of calculating the dynamic coupling relationships between nodes includes: The basic coupling strength is determined based on the admittance between nodes, real-time apparent power, and physical connection relationships. The basic coupling strength is modified by combining the time decay characteristics of fault propagation to generate a dynamic spatial weight matrix.

[0011] Optionally, the process of fault relationship modeling through physical semantic contrastive learning and dynamic hypergraphs includes: Build a multimodal physical descriptor for each node; Positive and negative sample pairs are constructed based on the consistency of fault type and electrical distance proximity, and the model is trained using a physically weighted contrastive loss function to obtain contrastive features; A hypergraph is dynamically constructed based on the aforementioned comparative features and real-time branch power, and the spectral embedding matrix of the hypergraph is extracted. By fusing the spectral embedding matrix with the feature space information, a dynamic fault prototype is generated.

[0012] Optionally, the process of fault diagnosis using the causal conceptual bottleneck reasoning framework includes: By using a structured concept discovery network, intermediate concept vectors with clear physical meaning are extracted from the enhanced features; Based on the intermediate concept vector, a causal directed acyclic graph between concepts and faults is constructed through differentiable structure learning; Based on the causal directed acyclic graph and the dynamic fault prototype, fault type classification and fault line location are performed.

[0013] Optionally, the process of extracting intermediate concept vectors with explicit physical meaning includes: Use multi-head attention mechanisms to separate independent latent concept dimensions from features; By using pseudo-concept labels extracted from measured data, the alignment of the latent concept dimension with the real physical semantics is supervised through concept alignment loss.

[0014] Optionally, the method further includes: inputting the intermediate concept vectors activated during the reasoning process into a natural language decoder to generate a structured text explanation describing the causes and evolution paths of the failure.

[0015] In a second aspect, the present invention also provides a computer terminal device, comprising: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the power distribution network fault diagnosis method based on multi-source data reconstruction and causal reasoning in the first aspect above.

[0016] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the power distribution network fault diagnosis method based on multi-source data reconstruction and causal reasoning in the first aspect described above.

[0017] Compared with the prior art, the present invention has the following advantages and technical effects: This invention provides a distribution network fault diagnosis method based on multi-source data reconstruction and causal reasoning. By constructing a physical constraint diffusion generation model, it explicitly embeds circuit physics laws during data reconstruction and utilizes a gated interpolation mechanism, effectively overcoming the challenges of feature extraction caused by noise interference, missing key information, and time asynchrony in multi-source data of the distribution network, ensuring high fidelity and physical authenticity of the input data. By proposing a physically guided spatiotemporal state network, it deeply integrates the dynamic topology of the distribution network with the real-time electrical physical evolution laws into feature encoding, solving the problem of lack of feature physical consistency caused by the failure of existing methods to organically combine topology and physical laws, and achieving accurate characterization of fault transient processes. The introduction of a dynamic hypergraph prototype network and meta-learning mechanism overcomes the limitations of relying on fixed topology modeling, enabling flexible capture of high-order fault propagation relationships such as multi-feeder coupling and rapid adaptation to topology changes, significantly enhancing the model's generalization ability in complex dynamic distribution network operating environments. Finally, by constructing a causal conceptual bottleneck reasoning framework, the decision-making barrier of the traditional fault diagnosis "black box" model is broken, realizing full-link interpretable reasoning from fault characteristics to physical concepts to diagnostic conclusions. This provides operation and maintenance personnel with intuitive analysis of fault causes and evolution paths, greatly improving the credibility and practical application value of fault diagnosis in intelligent operation and maintenance decision-making. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a power distribution network fault diagnosis method based on multi-source data reconstruction and causal reasoning, according to an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0021] like Figure 1 As shown, this embodiment provides a method for fault diagnosis of power distribution networks based on multi-source data reconstruction and causal reasoning, including: Acquire multi-source heterogeneous operation data of the power distribution network; Based on the constraints of physical laws, the multi-source heterogeneous operating data is repaired and spatiotemporally aligned to generate a physically consistent data matrix. Based on the physically consistent data matrix, steady-state feature vectors of distribution network nodes are extracted by fusing dynamic topology and electrical physical laws into a state-space network. Based on the steady-state feature vector, fault relationship modeling is performed through physical semantic contrast learning and dynamic hypergraph to generate fault prototypes and enhanced features with fused topological structures. Based on the enhanced features and fault prototypes, fault diagnosis is performed through a causal conceptual bottleneck reasoning framework, and the output is a decision result that includes fault type, location and explanation.

[0022] Furthermore, the process of repairing and spatiotemporally aligning the multi-source heterogeneous operational data includes: A conditional diffusion model is constructed, which adds noise to the data through a forward diffusion process and reconstructs the data through a reverse generation process; Kirchhoff's laws and power balance equations are embedded in the loss function of the reverse generation process as physical constraints. In the reverse generation process, latent variable interpolation and gating mechanisms are used to achieve time alignment of different source data; An iterative refinement strategy is adopted, and the data is corrected by a physical verification module after each generation step, outputting a physically consistent data matrix.

[0023] Specifically, the implementation process of this embodiment includes: Step 1: Processing and Fault Feature Reconstruction of Multi-Source Heterogeneous Data Voltage phasor data of the distribution network were collected using synchronous phasor measurement units (PMUs). V t Current phasor data i t Supervisory Control and Data Acquisition (SCADA) system collects active power. P With reactive power Q Switch status s t and alarm signals a t Advanced Measurement Systems (AMI) acquire load profiles l t Meteorological system collects temperature T t ,humidity H t Wind speed W t External environment data, etc.

[0024] To address issues of noise, missing data, and temporal asynchrony, a physical constraint generation framework based on a diffusion model is constructed. A conditional diffusion process between observed data and latent variables is established, and data repair and reconstruction are achieved through inverse generation. This involves multi-source data... X 0 is projected onto a unified dimension through an embedding layer and then concatenated to form a multivariate time series matrix. x 0.

[0025] The forward diffusion process is defined as follows: ; in, β t For noise scheduling parameters, x 0 represents a multivariate time series matrix.

[0026] To ensure that the reverse-generated data strictly follows the steady-state operation laws of the distribution network, Kirchhoff's laws and the power balance equation are integrated into the loss function, and the generation process is anchored to the nodal power observations: ; Where, λ kcl Kirchhoff's current law constraint coefficient; l 2. Kirchhoff's voltage law constraint coefficients; A This is the node-branch correlation matrix; B The basic loop matrix; To generate the branch current phasor vector in the data; Inject current phasor vectors into nodes; This is the branch voltage drop vector; P These are active power observations; Q The reactive power observation value of the node; V node The node voltage vector; It is the conjugate of the current vector.

[0027] Time alignment is achieved through latent variable interpolation and gating mechanisms: ; in, E t For time t The sine and cosine position encoding vectors; Gate(·) is a single-layer neural network with Sigmoid activation; Interp(·) is a resampling interpolation operation based on spline functions; x t This is the observation vector at the current moment; x tasync This is the raw data sequence acquired at non-aligned time points; z t This is the aligned fused feature vector.

[0028] A multi-step iterative refinement generation strategy is adopted, with the results of each step corrected by a physical verification module. This module is a Newton-Raphson power flow solver embedded in the generation process, which corrects the results by calculating the residuals between the generated data and the solutions to the power flow equations. ; in, This is the corrected hidden state vector; x t-1 This represents the original hidden state vector; α grad This represents the physical gradient step size; ▽ x Let be the gradient operator with respect to the variable.

[0029] The data matrix generated by the output diffusion model H gen Post-hoc normalization: ; in, H out To output the dataset; H gen Data matrix generated for diffusion model; m gen The mean vector of the data matrix generated for the diffusion model; s gen The standard deviation vector of the data matrix generated for the diffusion model.

[0030] Furthermore, the process of extracting the steady-state feature vectors of distribution network nodes by integrating dynamic topology and electrical physical laws into a state-space network includes: Based on the real-time operating status of the distribution network, calculate the dynamic coupling relationship between nodes to construct a dynamic spatiotemporal diagram; Construct a selective state-space model and generate the state transition matrix and input mapping matrix of the model based on the electrical and physical properties of the nodes. Based on the dynamic spatiotemporal graph and the parameterized selective state space model, the spatiotemporal evolution and fusion of the hidden states of nodes are carried out. The evolved node features are subject to regularization constraints based on the consistency of the power flow equations, and the steady-state feature vectors of the nodes are aggregated.

[0031] Furthermore, the process of calculating the dynamic coupling relationships between nodes includes: The basic coupling strength is determined based on the admittance between nodes, real-time apparent power, and physical connection relationships. The basic coupling strength is modified by combining the time decay characteristics of fault propagation to generate a dynamic spatial weight matrix.

[0032] Specifically, the implementation process of this embodiment includes: Step 2: Fault-aware topology physical feature encoding based on PG-STSN: Processed distribution network dataset H out ∈R T×N×D As input, where T Let N be the number of time steps, N be the number of distribution network nodes, and D be the feature dimension. A dynamic spatiotemporal graph is constructed based on the actual distribution network topology. G =( B , e s , e t Where B is the node set, e s For spatial edge sets, e t This is a temporal edge set. To quantize the dynamic coupling relationships between nodes, a dynamic spatial weight matrix is ​​defined. W s ( t )∈R N×N Its elements W ij ( t ) represents a node i and j At any moment t The coupling strength is calculated using the following formula: ; in, W ij ( t ) represents the dynamic edge weight; y ij For nodes i and j Admittance between; Norm (·) is a normalization function that ensures that branches with larger admittances have higher base weights. P ij ( t )for t Apparent power amplitude of the branch at any given time; P base This is a power reference value; α pf The power flow influence coefficient; β top These are the topology enhancement coefficients; I line ( i , j ) is the line existence indicator function if node i and j The value is 1 if a physical line exists, otherwise it is 0. f (Δ t ij ) is the fault propagation time decay function. c delay This is the fault propagation delay coupling coefficient.

[0033] A selective state-space model is used as the core of feature propagation for each node. i Establish the discrete-time state equations. Define the nodes. i exist t The observation vector at time is x i,t The hidden state vector is h i,t The output vector is o i,t To achieve physical-guided feature learning, a parameterized distribution network state-space matrix is ​​first constructed based on the nodal electrical characteristics: ; in, Embed vectors for node physical attributes; This is the normalized value of the average voltage deviation from adjacent nodes; The average phase angle at the nodes; The average apparent power of the node; The standard deviation of load volatility; The average current amplitude at the node; Estimate the proportion of line loss;U A , U B , U C , U D It is a learnable projection matrix.

[0034] Based on the parameter matrix of the above physical perception, the state-space evolution equation is: ; in, h i,t For nodes i exist t The hidden state vector at time t; x i,t For nodes i At any moment t Measurement data; A i ( t ) represents the state transition matrix; B i ( t Input mapping matrix; C i To output the mapping matrix; D i Through matrix.

[0035] Node features are updated through spatiotemporal fusion and physical attention mechanisms: ; in, N ( i ) is a node i The set of neighbors; Δ V ij With Δ i ij These represent the voltage magnitude difference and phase angle difference between nodes, respectively; SSM Θ (·) represents a parameterized state-space module; q i For query vector; K G This is the global pattern key matrix; V G is the global state value matrix; LayerNorm(·) is the layer normalization function.

[0036] To enhance the physical feasibility of feature representation, a regularization constraint based on the consistency of power flow equations is introduced: ; in, Y iThe row vectors corresponding to the nodal admittance matrices; , and They are respectively based on features z i，t The voltage, current, and power estimates obtained from decoding; l bal To balance the hyperparameters.

[0037] Then, the steady-state representation of the node is obtained through time aggregation and attribute fusion: ; in, z i The steady-state feature vector of the node; W out To output the projection weight matrix; Embed vectors for node physical attributes; b out This is the output bias vector.

[0038] Furthermore, the process of fault relationship modeling through physical semantic contrastive learning and dynamic hypergraphs includes: Build a multimodal physical descriptor for each node; Positive and negative sample pairs are constructed based on the consistency of fault type and electrical distance proximity, and the model is trained using a physically weighted contrastive loss function to obtain contrastive features; A hypergraph is dynamically constructed based on the aforementioned comparative features and real-time branch power, and the spectral embedding matrix of the hypergraph is extracted. By fusing the spectral embedding matrix with the feature space information, a dynamic fault prototype is generated.

[0039] Specifically, the implementation process of this embodiment includes: Step 3: Fault feature enhancement and hypergraph relationship modeling based on physical semantic comparison: Step 2 outputs the steady-state feature vector of the distribution network node. z i The input is fed into the physical-semantic fusion contrastive learning framework proposed in this step, which aims to explicitly model the structured similarity relationships between failure modes by introducing multimodal physical criteria.

[0040] First, a multimodal physical descriptor vector containing electrical attributes and historical behavior is constructed for each distribution network node. ,in t i Encode the node type. or i Historical failure frequency f i For voltage sag sensitivity, d iFor electrical zone identification, k i The load-meteorological response coupling coefficient. n i Topological criticality is defined. Based on this, a dynamic positive and negative sample pair construction mechanism is defined: For fault classification tasks, samples with "same fault type" but "different fault locations, different load levels, and different transition resistances" are regarded as strong positive sample pairs, aiming to learn the invariant characteristics of fault types; For fault location tasks, based on the power grid topology, fault samples of nodes with "electrically adjacent" nodes are regarded as weak positive sample pairs, aiming to construct a fault manifold space that maintains topological continuity; Negative sample pairs: Samples with different fault types or electrical distances exceeding a threshold are selected to form negative sample pairs, and hard examples are mined according to the physical deviation criterion.

[0041] To incorporate physical semantics into feature similarity metrics, a physically weighted contrastive loss function is designed: ; in, P t The set of positive sample pairs; N t For the set of negative sample pairs; The weights represent the physical semantic similarity. z i and z j All are node feature vectors; sim(·) is the similarity function; t This refers to temperature hyperparameters. l neg The penalty coefficient for negative samples; or margin λ is the similarity boundary threshold; phy The physical consistency coefficient; T phy (·) represents the physical semantic projection network; This is a multimodal physical descriptor vector.

[0042] To further capture the topological dependencies of failure modes, a dynamic hypergraph prototype network is constructed. In each training round, based on the current contrastive representation... With real-time branch power P ij ( t Dynamically generated hypergraph H t Its hyperedge weight is determined by both feature similarity and real-time electrical coupling strength: ; in, w h ( em , t ) represents the dynamic weight of the hyperedge; e m This is an indicator for the super-edge; c hyper The hypergraph fusion coefficient; P ij ( t ) represents the active power of the branch; max k | P ik ( t )| represents the maximum adjacent power of a node. For super-edge e m Middle node i The contrast feature vectors; For super-edge e m Middle node j The contrast feature vectors.

[0043] Selective eigenvalue decomposition is performed on the Laplacian matrix of the dynamic hypergraph to extract the spectral embedding matrix. U ( t Then, the spectral prototype vector of the fused topology is calculated. By fusing the spectral prototype and the feature space prototype through an attention mechanism, a multimodal fault prototype that simultaneously reflects topological constraints and feature distributions is obtained. P c ( t To improve the model's generalization ability in scenarios with scarce samples, a meta-learning fast adaptation mechanism is introduced, with the following parameter update rule: ; in, i meta These are the prototype parameters of the spectrum; These are the updated spectral prototype meta-parameters; c m The learning rate for the meta-spectral prototype; Gradient operator for elementary parameters; z * For node comparison feature matrix; P t The set of positive sample pairs; N t It is a set of negative sample pairs.

[0044] Furthermore, the process of fault diagnosis using the causal conceptual bottleneck reasoning framework includes: By using a structured concept discovery network, intermediate concept vectors with clear physical meaning are extracted from the enhanced features; Based on the intermediate concept vector, a causal directed acyclic graph between concepts and faults is constructed through differentiable structure learning; Based on the causal directed acyclic graph and the dynamic fault prototype, fault type classification and fault line location are performed.

[0045] Furthermore, the process of extracting intermediate concept vectors with clear physical meaning includes: Use multi-head attention mechanisms to separate independent latent concept dimensions from features; By using pseudo-concept labels extracted from measured data, the alignment of the latent concept dimension with the real physical semantics is supervised through concept alignment loss.

[0046] Specifically, the implementation process of this embodiment includes: Step 4: Distribution networks based on causal concepts can explain fault classification, line location, and accurate distance calculation. The steady-state feature vector of the distribution network node z * Dynamic fault prototype of distribution network P c ( t and spectral embedding matrix U ( t The input is fed into the causal concept bottleneck reasoning framework proposed in this step, aiming to achieve integrated decision-making for distribution network fault detection, classification, location, and interpretation by introducing a physically interpretable intermediate concept layer and a causal graph structure. First, an interpretable physical concept vector is defined for each distribution network node, with each dimension corresponding to an intermediate variable with a clear physical meaning. This is implemented through a structured concept discovery network, which employs a multi-head causal attention mechanism to separate independent and interpretable concept dimensions from the features. ; in, c i For distribution network nodes i Structured causal concept vector; Attn h (·) represents the attention function of the h-th concept discovery head; M h For physical guidance mask matrix; H This refers to the number of concept heads.

[0047] To supervise the training of the concept interpreter, this application employs a physics-based rule-based preprocessing algorithm to automatically extract pseudo-concept labels. ={Transition Resistance, Voltage Sag, Waveform Distortion, Traveling Wave Time Difference Characteristics, Fault Area Indication}, these labels are initially estimated from the raw data using simple physical models such as single-ended ranging, Fourier transform, and wavelet transform. These pseudo-concept labels are used to construct the alignment loss. This ensures that the intermediate features extracted by the neural network have clear physical semantics. Each concept dimension is regularized using a concept alignment loss to ensure that the concept is aligned with the electrical physical quantity. ; in, ψ k For concept interpreters; These are pseudo-concept labels extracted from measured data of the power distribution network. K This represents the total number of concept dimensions.

[0048] Based on the learned concept vectors, a differentiable causal structure learning framework is used to dynamically construct a causal graph between concepts and faults. The causal adjacency matrix is ​​optimized through structure learning under acyclic graph constraints. ; Where C is the conceptual matrix of all nodes in the distribution network; A Let be the causal adjacency matrix to be learned; h ( A ) is the acyclic penalty function; l sparse for L 1. Regularization coefficient; l dag This is the acyclicity penalty coefficient.

[0049] Analysis of power distribution network fault types: The distribution network fault type is determined using a Softmax classifier, which aggregates and maps the concept vectors of key nodes after propagation to a fault category probability distribution. ; in, This is a probability vector of fault types in the distribution network, where the fault types are {normal, A-phase grounding, B-phase grounding, C-phase grounding, AB-phase short circuit, BC-phase short circuit, CA-phase short circuit, and three-phase short circuit}. Each data point in the vector represents the probability of a certain fault type occurring. V critical This represents the set of key nodes selected through causal impact analysis; Pool(·) is the graph pooling aggregation function.

[0050] Analysis of faulty lines in the distribution network: Utilizing post-propagation characteristics and dynamic fault prototypes P c ( t and topological spectral embedding U ( t The matching and fusion processes are performed to output the probability of faulty lines in the distribution network. ; in, The confidence score for a faulty line in a distribution network represents the degree of certainty that a fault has occurred on a particular line; Causal Influence Score (CSI) I , U ( t )) for combining topological embedding The node causal influence score.

[0051] Calculation of fault distance in power distribution network: A subset of causal concepts related to traveling waves and impedance is selected and input into a lightweight physical ranging model to calculate the precise distance of distribution network faults. ; in, The precise distance from the measurement point to the fault point in the power distribution network; v wave• Δ t est This is a rough distance calculated based on the classical traveling wave method formula; For the concept vector quantum related to ranging; Θ line ( t () represents the physical parameters of the line; f physical (·) represents the parameterized physical ranging model.

[0052] Furthermore, the method also includes: inputting the intermediate concept vectors activated during the reasoning process into a natural language decoder to generate a structured text explanation describing the causes and evolution paths of the failure.

[0053] Specifically, the implementation process of this embodiment includes: Step 5: Joint loss analysis of fault detection, classification, and localization based on natural language causal interpretation: To analyze the joint losses of fault detection, classification, and location in distribution networks, a natural language causal explanation generator is introduced, using the activation concept sequence of the model. As input, generate a structured text explanation: ; Here, Decoder(·) is a decoder function based on Transformer; c active This refers to the set of concepts that are activated during the reasoning process.

[0054] Confidence assessment is achieved through the concept uncertainty propagation method, which simultaneously quantifies the uncertainty in the concept extraction and causal reasoning stages: ; in, U concept For the concept of posterior variance; Ucausal The uncertainty lies in the causal graph structure.

[0055] To enhance the system's generalization ability under novel fault or topology change scenarios, a meta-concept learning framework is introduced. This framework addresses meta-tasks. In the learning of shared concept representations and causal structure priors, the meta-update rule is: ; in, i Meta-parameters are learned for the concept; I will Learn meta-parameters for the updated concept; i i Specific parameters for the task; For task support set; L adapt To adapt to the loss; or meta The learning rate is a meta-concept.

[0056] The overall training employs an alternating optimization strategy, training the concept extraction layer and the causal reasoning layer in stages, and finally fine-tuning end-to-end using multi-task loss. ; in, The combined losses from fault detection, classification, and location are considered comprehensively for the distribution network; For sparsity regularization of causal graphs; Loss of causal consistency; l a Alignment loss weights; l s For sparse regularized weights; l c Weights are assigned to the consistency loss.

[0057] In this embodiment, a computer terminal device is provided, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-described method for fault diagnosis of power distribution networks based on multi-source data reconstruction and causal reasoning.

[0058] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described method for diagnosing power distribution network faults based on multi-source data reconstruction and causal reasoning.

[0059] This invention provides a distribution network fault diagnosis method based on multi-source data reconstruction and causal reasoning. By constructing a physical constraint diffusion generation model, it explicitly embeds circuit physics laws during data reconstruction and utilizes a gated interpolation mechanism, effectively overcoming the challenges of feature extraction caused by noise interference, missing key information, and time asynchrony in multi-source data of the distribution network, ensuring high fidelity and physical authenticity of the input data. By proposing a physically guided spatiotemporal state network, it deeply integrates the dynamic topology of the distribution network with the real-time electrical physical evolution laws into feature encoding, solving the problem of lack of feature physical consistency caused by the failure of existing methods to organically combine topology and physical laws, and achieving accurate characterization of fault transient processes. The introduction of a dynamic hypergraph prototype network and meta-learning mechanism overcomes the limitations of relying on fixed topology modeling, enabling flexible capture of high-order fault propagation relationships such as multi-feeder coupling and rapid adaptation to topology changes, significantly enhancing the model's generalization ability in complex dynamic distribution network operating environments. Finally, by constructing a causal conceptual bottleneck reasoning framework, the decision-making barrier of the traditional fault diagnosis "black box" model is broken, realizing full-link interpretable reasoning from fault characteristics to physical concepts to diagnostic conclusions. This provides operation and maintenance personnel with intuitive analysis of fault causes and evolution paths, greatly improving the credibility and practical application value of fault diagnosis in intelligent operation and maintenance decision-making.

[0060] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fault diagnosis of power distribution networks based on multi-source data reconstruction and causal reasoning, characterized in that, Includes the following steps: Acquire multi-source heterogeneous operation data of the power distribution network; Based on the constraints of physical laws, the multi-source heterogeneous operating data is repaired and spatiotemporally aligned to generate a physically consistent data matrix. Based on the physically consistent data matrix, steady-state feature vectors of distribution network nodes are extracted by fusing dynamic topology and electrical physical laws into a state-space network. Based on the steady-state feature vector, fault relationship modeling is performed through physical semantic contrast learning and dynamic hypergraph to generate fault prototypes and enhanced features with fused topological structures. Based on the enhanced features and fault prototypes, fault diagnosis is performed through a causal conceptual bottleneck reasoning framework, and the output is a decision result that includes fault type, location and explanation.

2. The method according to claim 1, characterized in that, The process of repairing and spatiotemporally aligning the multi-source heterogeneous operational data includes: A conditional diffusion model is constructed, which adds noise to the data through a forward diffusion process and reconstructs the data through a reverse generation process; Kirchhoff's laws and power balance equations are embedded in the loss function of the reverse generation process as physical constraints. In the reverse generation process, latent variable interpolation and gating mechanisms are used to achieve time alignment of different source data; An iterative refinement strategy is adopted, and the data is corrected by a physical verification module after each generation step, outputting a physically consistent data matrix.

3. The method according to claim 1, characterized in that, The process of extracting steady-state feature vectors of distribution network nodes by integrating dynamic topology and electrical physical laws into a state-space network includes: Based on the real-time operating status of the distribution network, calculate the dynamic coupling relationship between nodes to construct a dynamic spatiotemporal diagram; Construct a selective state-space model and generate the state transition matrix and input mapping matrix of the model based on the electrical and physical properties of the nodes. Based on the dynamic spatiotemporal graph and the parameterized selective state space model, the spatiotemporal evolution and fusion of the hidden states of nodes are carried out. The evolved node features are subject to regularization constraints based on the consistency of the power flow equations, and the steady-state feature vectors of the nodes are aggregated.

4. The method according to claim 3, characterized in that, The process of calculating the dynamic coupling relationship between nodes includes: The basic coupling strength is determined based on the admittance between nodes, real-time apparent power, and physical connection relationships. The basic coupling strength is modified by combining the time decay characteristics of fault propagation to generate a dynamic spatial weight matrix.

5. The method according to claim 1, characterized in that, The process of fault relationship modeling through physical semantic contrastive learning and dynamic hypergraphs includes: Build a multimodal physical descriptor for each node; Positive and negative sample pairs are constructed based on the consistency of fault type and electrical distance proximity, and the model is trained using a physically weighted contrastive loss function to obtain contrastive features; A hypergraph is dynamically constructed based on the aforementioned comparative features and real-time branch power, and the spectral embedding matrix of the hypergraph is extracted. By fusing the spectral embedding matrix with the feature space information, a dynamic fault prototype is generated.

6. The method according to claim 5, characterized in that, The process of fault diagnosis using the causal conceptual bottleneck reasoning framework includes: By using a structured concept discovery network, intermediate concept vectors with clear physical meaning are extracted from the enhanced features; Based on the intermediate concept vector, a causal directed acyclic graph between concepts and faults is constructed through differentiable structure learning; Based on the causal directed acyclic graph and the dynamic fault prototype, fault type classification and fault line location are performed.

7. The method according to claim 6, characterized in that, The process of extracting intermediate concept vectors with clear physical meaning includes: Use multi-head attention mechanisms to separate independent latent concept dimensions from features; By using pseudo-concept labels extracted from measured data, the alignment of the latent concept dimension with the real physical semantics is supervised through concept alignment loss.

8. The method according to claim 6, characterized in that, The method further includes: inputting the intermediate concept vectors activated during the reasoning process into a natural language decoder to generate a structured text explanation describing the causes and evolution paths of the fault.

9. A computer terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the method as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-8.

Images