Power distribution network equipment fault root cause analysis data modeling method

Abstract

The invention discloses a power distribution network equipment fault root cause analysis data modeling method, relates to the technical field of power distribution network fault analysis, and aims to solve the technical problems of low fault root cause analysis accuracy and slow response under high-frequency topology dramatic change and edge terminal computing power fluctuation of an existing power distribution network. Comprising the following steps: S1, acquiring equipment operation data and real-time topology information by a heterogeneous edge terminal of a power distribution network, constructing a dynamic lightweight time sequence diagram model with online self-adaptive capability according to an edge terminal computing power high-frequency dynamic topology dramatic change scene, and outputting a fault root cause coarse positioning result; s2, the heterogeneous edge terminal uploads topology increment information and a coarse positioning result in a grading manner based on topology change intensity, wherein the topology increment information is structured data of local topology change of the power distribution network; and S3, the cloud platform updates the global time sequence diagram according to the topology change intensity in a grading manner. The method has the advantages of improving the fault root cause analysis accuracy and the response speed.

Description

Technical Field

[0001] This invention relates to the field of power distribution network fault analysis technology, and more specifically, to a data modeling method for root cause analysis of power distribution network equipment faults. Background Technology

[0002] With the large-scale integration of new power equipment such as distributed photovoltaics and user-side energy storage, urban power distribution networks have formed a complex topology with multiple interactions between "source-grid-load-storage". Typical scenarios, such as power distribution networks in high-density residential areas and industrial parks, exhibit high-frequency dynamic and drastic changes in their topology due to factors such as equipment switching, load fluctuations, and changes in the output of distributed power sources. At the same time, the heterogeneous edge terminals deployed in the power distribution network (including local terminals in substations, distribution area monitoring terminals, and distributed power source supporting terminals) have significant differences in computing power, making it difficult for traditional root cause analysis models to adapt to the dual challenges of "high-frequency topology drastic changes + heterogeneous computing power fluctuations".

[0003] Existing root cause analysis methods for distribution network faults mostly employ fixed-structure models, lacking online adaptability to topology changes and computing power fluctuations. On the one hand, the rigid model structure cannot quickly adjust the feature extraction dimensions when the topology changes drastically, leading to the omission of key fault features and a significant decrease in analysis accuracy. On the other hand, the data transmission and model updates between edge terminals and cloud platforms lack a hierarchical collaborative mechanism. Uploading all data results in transmission delays, and fixed-frequency updates cannot match dynamic topology changes, ultimately causing a lag in the response to fault root cause tracing, which is difficult to meet the needs of rapid fault handling in distribution networks. In view of this, we propose a data modeling method for root cause analysis of distribution network equipment faults. Summary of the Invention

[0004] The purpose of this invention is to provide a data modeling method for root cause analysis of faults in distribution network equipment, so as to solve the technical problems of low accuracy and slow response of root cause analysis under the conditions of high-frequency topological changes and fluctuations in computing power of edge terminals in existing distribution networks.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a data modeling method for root cause analysis of faults in power distribution network equipment, comprising the following steps: S1. The heterogeneous edge terminal of the distribution network acquires equipment operation data and real-time topology information. For the scenario of high-frequency dynamic topology changes in edge terminal computing power, a dynamic lightweight time sequence graph model with online adaptive capability is constructed to output coarse fault root cause localization results. S2. The heterogeneous edge terminal uploads topology incremental information and coarse positioning results based on the intensity of topology changes. The topology incremental information is structured data of local changes in the distribution network topology. S3: The cloud platform updates the global time series graph according to the intensity of topological changes, and combines reinforcement learning and fault propagation physical rules to conduct fine-grained source tracing of fault root causes and generate accurate fault root cause analysis results. S4. The cloud platform aggregates the model parameters of each heterogeneous edge terminal through a hierarchical lightweight federated learning mechanism, generates dynamic optimization rules and distributes them. The heterogeneous edge terminals realize online model iteration based on the optimization rules, forming a closed loop of continuous fault root cause analysis in edge-cloud collaboration.

[0006] Preferably, the construction of a dynamic lightweight time series graph model with online adaptive capabilities in S1 specifically includes: Heterogeneous edge terminals monitor the frequency of topology changes in the distribution network and their own computing power fluctuations in real time, and establish a two-dimensional adaptation model of "topology change frequency - computing power fluctuations". When the frequency of topology changes exceeds a preset threshold, the online model distillation mechanism is triggered. This mechanism dynamically freezes non-core network layers and retains the topology-sensitive feature extraction module to achieve real-time adjustment of the model structure. The topology-sensitive feature extraction module designs exclusive feature mapping rules for the connection relationships of distribution network lines and the associated attributes of equipment to ensure the effective extraction of key features when the topology changes drastically. To quantify the two-dimensional adaptation relationship between "topology change frequency and computing power fluctuation", a model pruning ratio calculation algorithm is introduced, and the specific formula is as follows: ; in, The pruning ratio for the model, For the frequency of topology changes in the distribution network, The reference threshold for topological change frequency. For topological change threshold, Real-time computing power for edge terminals This represents the historical average computing power of edge terminals. This represents the minimum available computing power for edge terminals.

[0007] Preferably, S1 further includes a model adaptation verification mechanism: After the dynamic lightweight time series graph model outputs coarse localization results, the current model's fit is scored by comparing the analysis accuracy under similar historical topology scenarios. If the score is lower than the preset threshold, the model parameter fine-tuning will be automatically initiated, and the weight allocation strategy of the time sequence graph attention mechanism will be adjusted to prioritize strengthening the feature weights that are strongly correlated with the current topology. To quantify the model fitness score and the adjustment range of attention weights, a fitness score algorithm and an attention weight update formula are introduced. Model fit score calculation formula: ; The formula for updating attention weights in a time series graph is as follows: ; in, Score the model fit. This is the topological feature vector of the current scene. For the first Topological feature vectors of similar historical scenarios, For the first Accuracy of model analysis for similar historical scenarios. For the current and the first Cosine similarity of topological features of historical scenes Nodes in the timing graph With nodes The original attention weights, For the adjusted attention weights, For nodes With nodes The topological correlation coefficient, This is the weight adjustment coefficient; the lower the fit, the larger the adjustment coefficient.

[0008] Preferably, the step S2 of uploading topology incremental information based on the intensity of topology change specifically includes: The heterogeneous edge terminal classifies the intensity of topology changes into three levels: normal change, significant change, and drastic change, based on the number of nodes involved in the topology change, the type of key equipment, and the scope of the fault impact. For routine changes, upload incremental topology information in batches according to a preset cycle; For significant changes, upload information on key change nodes in real time; In the event of drastic changes, a compressed transmission protocol is used to prioritize uploading information on the nodes with the highest fault correlation and changes in connection relationships. The topological incremental information includes a change level identifier, node attribute change records, and edge relationship evolution sequence, and is accompanied by a dynamic timestamp to ensure temporal consistency. To quantify the intensity level of topological change, a change intensity index calculation algorithm is introduced, the specific formula of which is as follows: ; in, This is the topological change intensity index. This represents the number of nodes involved in the topology change. This represents the total number of nodes within the monitoring range of the edge terminal. This refers to the number of critical devices involved in the topology change. The number of users affected by the topology change. This represents the total number of users within the monitoring range of the edge terminal. , , These are the weighting coefficients for each dimension.

[0009] Preferably, step S2 further includes incremental information verification preprocessing: Before uploading incremental topology information, the heterogeneous edge terminal performs consistency verification on the incremental data through locally cached historical topology snapshots, eliminating duplicate or invalid change records. The incremental information that passes the verification is feature-encoded, and the unstructured information such as device type and connection status is converted into numerical feature vectors to reduce the amount of data transmitted. To quantify the results of incremental data consistency verification, a data validity coefficient calculation algorithm is introduced, the specific formula of which is as follows: ; During the feature encoding stage, the device type is encoded using one-hot encoding. The connection status is encoded in binary. Finally, the numerical feature vector ; in, The data validity coefficient is... For node attribute change records in incremental data, This is the set of node attributes that are allowed to change in a historical topology snapshot. This records the changes in edge relationships within the incremental data. This is the set of edge relationships that are allowed to change in the historical topology snapshot. A function to calculate the number of elements in a set. For the one-hot encoded vector of the device type, The binary encoded vector of the connection state. This is the final numerical feature vector.

[0010] Preferably, the step S3 of updating the global time series graph according to the intensity of topological changes specifically includes: After receiving the topology increment information, the cloud platform initiates the corresponding update strategy based on the level of change intensity: Regular changes are handled using batch asynchronous updates, significant changes using incremental synchronous updates, and dramatic changes using priority queue updates. During the update process, distribution network physical constraint rules are introduced to correct the global timing diagram. These physical constraint rules include power flow correlation constraints derived from Kirchhoff's laws and equipment capacity limitation constraints, ensuring that the updated global timing diagram conforms to the operation rules of the distribution network. To quantify the impact of physical constraints on the global timing graph, a power flow consistency verification algorithm is introduced, with the specific formula as follows: Based on Kirchhoff's current law, the nodes in the global timing graph Verify the trend ; like Based on equipment capacity constraints Correct the node current value, and use the corrected formula: , ; in, For tidal current deviation rate, For inflow node The sum of the currents, outflow node The sum of the currents, This is the corrected inflow node current. This is the corrected outflow node current. For nodes Actual load power For nodes Rated capacity.

[0011] Preferably, the detailed source tracing of fault root causes in S3, combining reinforcement learning and fault propagation physics rules, specifically includes: The cloud platform uses the updated global time series graph as its environment to construct a reinforcement learning agent that integrates physical rules. The state space of the agent includes node physical parameters and topological features, and the action space is the search direction for the root cause tracing path. A hybrid reward function is set up, which includes path length reward, physical rule matching degree reward and root cause feature similarity reward, wherein the physical rule matching degree reward is used to penalize the tracing path that does not conform to the physical law of fault propagation; Through the interaction and iteration between the intelligent agent and the environment, the shortest tracing path that conforms to physical rules is generated to locate the root cause device and key propagation nodes of the fault. To quantify the hybrid reward function, a multi-dimensional reward calculation algorithm is introduced, the specific formula of which is as follows: ; The agent updates the action selection probability using the policy gradient algorithm. The policy update formula is as follows: ; in, For mixed reward values, This represents the length of the tracing path currently being searched by the agent. This is the shortest historical tracing path length. The physical rule matching degree is calculated from the power flow deviation rate. , For root cause feature similarity, , , As a reward weight, The gradient of the objective function of the policy network. The parameter set for the reinforcement learning policy network; In the state Select action The probability, As a discount factor, For the next state The value function of .

[0012] Preferably, the model parameters of each heterogeneous edge terminal aggregated through a hierarchical lightweight federated learning mechanism in S4 specifically include: The cloud platform divides federated learning levels according to the power distribution network partitions. Edge terminals within the same partition form sub-federations. Each sub-federation first performs local parameter aggregation to generate partition-level feature parameters. The cloud platform aggregates feature parameters at each partition level to form global optimization parameters. During the aggregation process, a dynamic adjustment strategy for partition weights is adopted, assigning higher aggregation weights to partitions with high accuracy in recent fault analysis. The feature parameters include only the core parameters of the time series graph attention matrix and the fault feature mapping coefficients, and the parameter size is further compressed through feature dimensionality reduction technology. To quantify the dimensionality reduction effect of partition aggregation weights and parameters, a hierarchical aggregation weight algorithm and a dimensionality reduction error calculation formula are introduced, as follows: Partition aggregation weight calculation: ; in, For the first The aggregate weight of each partition, For the first Accuracy of recent fault analysis for each partition. For the first Number of edge terminals within each partition This represents the total number of distribution network zones. Partition-level feature parameters Aggregate into global parameters , ; in, For the first The region-level feature parameters of each partition. Parameters are optimized globally; Feature dimensionality reduction error calculation: ; in, For feature dimensionality reduction error, Original parameter dimensions Dimensionality of parameters after dimensionality reduction The first parameter is the covariance matrix of the parameter. Each feature value.

[0013] Preferably, step S4 further includes a privacy-preserving dynamic adaptation mechanism: The injection intensity of differential privacy noise is dynamically adjusted according to the privacy sensitivity level of different partitions, and the privacy sensitivity level is positively correlated with the noise intensity. Before parameter aggregation, federated distillation is performed on the partition-level feature parameters to extract common feature patterns from the parameters and remove personalized parameter details that may leak partition topology privacy, thus ensuring data privacy and security during the aggregation process. To quantify the intensity of differential privacy noise, a noise injection algorithm is introduced, the specific formula of which is as follows: ; Partition parameters after noise injection : ; in, The privacy sensitivity level of a partition is calculated based on the percentage of user data within that partition. , The number of users is affected by the partitioning. Total number of users The standard deviation of Gaussian noise. Noise figure To increase the probability of privacy leaks, For differential privacy budgeting, For the first The original region-level feature parameters of each partition. These are the partition feature parameters after noise injection. With a mean of 0 and a variance of The Gaussian distribution.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs a dynamic lightweight time series graph model with online adaptive capabilities and combines it with an edge-cloud hierarchical collaboration mechanism to significantly improve the accuracy and response speed of fault root cause analysis in scenarios of high-frequency topological changes and computing power fluctuations, effectively solving the core problems of poor adaptability and analysis lag in existing methods.

[0015] 2. This invention also solves the problem of balancing data transmission efficiency and effectiveness in hierarchical collaboration by using a topology incremental information hierarchical uploading and validity verification mechanism to ensure real-time transmission of key data while eliminating invalid and redundant data and reducing transmission resource consumption.

[0016] 3. This invention also solves the contradiction between improving model performance and protecting privacy during collaborative optimization by using a hierarchical lightweight federated learning and privacy protection dynamic adaptation mechanism. This mechanism aggregates high-quality global model parameters and enables continuous iterative optimization of the model while ensuring the privacy of partition topology and the security of user data. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0018] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0019] Example 1, such as Figure 1 As shown, this invention provides a data modeling method for root cause analysis of faults in power distribution network equipment, comprising the following steps: S1. The heterogeneous edge terminal of the distribution network acquires equipment operation data and real-time topology information. For the scenario of high-frequency dynamic topology changes in edge terminal computing power, a dynamic lightweight time sequence graph model with online adaptive capability is constructed to output coarse fault root cause localization results. S2. Heterogeneous edge terminals upload topology incremental information and coarse positioning results based on the intensity of topology changes. The topology incremental information is structured data of local changes in the distribution network topology. S3: The cloud platform updates the global time series graph according to the intensity of topological changes, and combines reinforcement learning and fault propagation physical rules to conduct fine-grained source tracing of fault root causes and generate accurate fault root cause analysis results. S4. The cloud platform aggregates model parameters from various heterogeneous edge terminals through a layered lightweight federated learning mechanism, generates dynamic optimization rules, and distributes them. The heterogeneous edge terminals implement online model iteration based on the optimization rules, forming a closed loop of continuous fault root cause analysis in edge-cloud collaboration.

[0020] In an embodiment of the present invention, S1, constructing a dynamic lightweight time series graph model with online adaptive capabilities, specifically includes: Heterogeneous edge terminals monitor the frequency of topology changes in the distribution network and their own computing power fluctuations in real time, and establish a two-dimensional adaptation model of "topology change frequency - computing power fluctuations". When the frequency of topology changes exceeds a preset threshold, the online model distillation mechanism is triggered. This mechanism dynamically freezes non-core network layers and retains the topology-sensitive feature extraction module to achieve real-time adjustment of the model structure. The topology-sensitive feature extraction module designs exclusive feature mapping rules for the connection relationships of distribution network lines and the associated attributes of equipment to ensure the effective extraction of key features when the topology changes drastically. To quantify the two-dimensional adaptation relationship between "topology change frequency and computing power fluctuation", a model pruning ratio calculation algorithm is introduced, and the specific formula is as follows: ; in, The pruning ratio for the model is [0.3, 0.8]. A larger value indicates a higher degree of model compression. This refers to the frequency of topology changes in the distribution network, measured in "times / minute," reflecting the frequency of dynamic topology changes. This serves as the baseline threshold for topological change frequency, with a default of 2 times per minute, and is the dividing point between normal and significant topological changes. This is the threshold for topological changes, with a default of 8 times per minute, serving as the dividing line between significant changes and dramatic changes. It provides real-time computing power for edge terminals, reflecting the terminal's current computing capabilities; The historical average computing power of edge terminals serves as a reference benchmark for computing power fluctuations. This represents the minimum available computing power for an edge terminal, and the minimum computing power limit that a terminal can run. This formula is the logic for calculating the model pruning ratio in two dimensions: "topology change frequency - computing power fluctuation": when the topology change frequency is at the baseline threshold... With the threshold of dramatic change In between, the pruning ratio increases linearly with frequency to adapt to high-frequency topology changes through higher compression ratios; when the real-time computing power of the edge terminal is at its lowest available computing power... Compared with historical average computing power In the meantime, the pruning ratio decreases linearly with increasing computing power to retain more model features when computing power is sufficient; in other scenarios, the default pruning ratio is used to balance model compression and feature preservation. By dynamically adjusting the pruning ratio in two dimensions, the inference delay caused by feature redundancy in the model during high-frequency topology changes is avoided, and the problem of low-computing-power terminals being unable to run due to overly complex models is also prevented. This ensures that heterogeneous edge terminals can efficiently output coarse localization results of fault roots under different topology and computing power scenarios, balancing model lightweighting and analysis effectiveness.

[0021] In an embodiment of the present invention, S1 further includes a model adaptation verification mechanism: After the dynamic lightweight time series graph model outputs coarse localization results, the current model's fit is scored by comparing the analysis accuracy under similar historical topology scenarios. If the score is lower than the preset threshold, the model parameter fine-tuning will be automatically initiated, and the weight allocation strategy of the time sequence graph attention mechanism will be adjusted to prioritize strengthening the feature weights that are strongly correlated with the current topology. To quantify the model fitness score and the adjustment range of attention weights, a fitness score algorithm and an attention weight update formula are introduced. Model fit score calculation formula: ; The accuracy of historical scene analysis is weighted and summed by the cosine similarity of topological features between the current scene and similar historical scenes to obtain the current model fitness score. The higher the similarity of the historical scene, the greater its contribution to the accuracy of the current score, ensuring that the score can reflect the model's ability to adapt to the current topological scene. The formula for updating attention weights in a time series graph is as follows: ; in, The model fitness score is given, with a value of [0,1]. The larger the value, the better the model is adapted to the current scene. This is the current scene topology feature vector, which includes topology dimension information such as the number of nodes, connection relationships, and device types. For the first Topological feature vectors of similar historical scenarios; For the first Accuracy of model analysis for similar historical scenarios; For the current and the first The cosine similarity of the topological features of each historical scene takes a value of [0,1]. The larger the value, the higher the scene similarity. Nodes in the timing graph With nodes The original attention weights reflect the initial importance of the association between the two nodes; The adjusted attention weights are the final weights after strengthening the topological association features; For nodes With nodes The topological association coefficient is output by the topological sensitive feature extraction module and takes the value [0,1]. The larger the value, the stronger the topological association between the two nodes. This is the weight adjustment coefficient; the lower the fit, the larger the adjustment coefficient. The weight adjustment coefficient is determined by the fit score. The lower the fit, the larger the adjustment coefficient, which is then combined with the topological correlation coefficient. Original attention weights The model is amplified by prioritizing the enhancement of feature weights that are strongly correlated with the current topology, thereby achieving fine-tuning of model parameters. The model's adaptability is quantitatively assessed by using an adaptability score, avoiding analysis errors caused by insufficient adaptation when the model changes topology. The attention weight is fine-tuned by combining the topology correlation coefficient to ensure that the model can dynamically focus on the key features of the current topology, further improving the accuracy of coarse fault location and providing a reliable foundation for subsequent fine-grained tracing in the cloud.

[0022] In an embodiment of the present invention, S2, which involves uploading topology incremental information based on the intensity of topology change, specifically includes: Based on the number of nodes involved in the topology change, the type of key equipment, and the scope of the fault impact, heterogeneous edge terminals classify the intensity of topology changes into three levels: normal change, significant change, and dramatic change. For routine changes, upload incremental topology information in batches according to a preset cycle; For significant changes, upload information on key change nodes in real time; In the event of drastic changes, a compressed transmission protocol is used to prioritize uploading information on the nodes with the highest fault correlation and changes in connection relationships. The topology incremental information includes change level identifiers, node attribute change records, and edge relationship evolution sequences, and is accompanied by dynamic timestamps to ensure temporal consistency. To quantify the intensity level of topological change, a change intensity index calculation algorithm is introduced, the specific formula of which is as follows: ; in, This is the topological change intensity index, with values ​​ranging from [0,1]. A larger value indicates a greater intensity and impact of the topological change. This refers to the number of nodes involved in the topology change, i.e., the total number of nodes added, deleted, or whose status changed in this topology change. The total number of nodes within the monitoring range of the edge terminal is used as the basis for calculating the proportion of nodes. The number of critical equipment involved in the topology change; critical equipment includes core equipment such as distribution transformers, switches, and reactive power compensation devices. This refers to the number of users affected by the topology change, i.e., the total number of users who may experience power supply anomalies due to the topology change. The total number of users within the monitoring range of the edge terminal is used as the basis for calculating the impact on the user ratio; , , The weight coefficients for each dimension correspond to the percentage of nodes, the percentage of key devices, and the percentage of users affected, respectively, ensuring that the sum of the weights is 1 to guarantee the rationality of the index calculation. The formula calculates the topology change intensity index using three dimensions: "node quantity ratio," "critical equipment ratio," and "user impact ratio due to failure," combined with preset weights. The weight of critical equipment is the highest at 0.5, because changes in critical equipment (such as distribution transformers and switches) have the greatest impact on fault analysis; the weight of the number of nodes and the weight of the affected users are 0.3 and 0.2 respectively, which comprehensively reflect the scope and degree of impact of topology changes. Finally, the change level is divided according to the index value, which provides a basis for hierarchical uploading. By quantifying the intensity of topology changes in multiple dimensions, the system enables hierarchical uploading of incremental topology information. Regular changes are uploaded in batches to reduce transmission resource consumption, while significant changes and dramatic shifts are uploaded first to ensure real-time performance. This approach avoids the waste of resources from full uploads and prevents analysis delays caused by the postponement of key change information, thus balancing transmission efficiency and analysis timeliness.

[0023] In an embodiment of the present invention, S2 further includes incremental information verification preprocessing: Before uploading incremental topology information, heterogeneous edge terminals perform consistency checks on incremental data using locally cached historical topology snapshots, eliminating duplicate or invalid change records. The incremental information that passes the verification is feature-encoded, and the unstructured information such as device type and connection status is converted into numerical feature vectors to reduce the amount of data transmitted. To quantify the results of incremental data consistency verification, a data validity coefficient calculation algorithm is introduced, the specific formula of which is as follows: ; During the feature encoding stage, the device type is encoded using one-hot encoding. The connection status is encoded in binary. Finally, the numerical feature vector ; in, The data validity coefficient takes values ​​in the range [0,1]. The data is then determined to be valid. This is a record of node attribute changes in incremental data, including node ID, changed attribute type, and values ​​before and after the change; This is the set of node attributes that are allowed to change in historical topology snapshots, such as dynamically changing attributes like running status, load rate, and temperature. This is a record of edge relationship changes in incremental data, including edge ID, change type (e.g., closure / opening), and change time; It is the set of edge relationships that are allowed to change in the historical topology snapshot, such as dynamic relationships like line connection status and power flow direction; Functions for calculating the number of elements in a set, such as This indicates the number of valid records in the node attribute change log; This is the unique thermal encoding vector for the equipment type, such as [1,0,0] for distribution transformers and [0,1,0] for switches; This is a binary encoded vector representing the connection status, where 1 indicates a closed connection and 0 indicates a closed connection. To obtain the final numerical feature vector, device type, connection status, and change intensity index are integrated to provide standardized input for the model; The effectiveness coefficient is calculated by combining two dimensions: "efficiency of node attribute changes" and "efficiency of edge relationship changes" with weights. Node attribute changes have a higher weight because node attributes (such as running status and load rate) have a more direct impact on fault analysis; efficiency is achieved by calculating the proportion of the intersection of incremental data and historical snapshots that allow changes, ensuring that only logically valid data is retained; Convert unstructured device types and connection states into numerical vectors. , ), and incorporate the intensity of change index To form a unified feature vector This provides standardized input for subsequent model feature extraction; By validating data, duplicate and invalid incremental data are eliminated, preventing invalid data from consuming transmission resources or interfering with subsequent analysis. Unstructured information is converted into numerical feature vectors to ensure consistency in data understanding between edge terminals and cloud models, while reducing data transmission volume and improving the efficiency and accuracy of model feature extraction.

[0024] In an embodiment of the present invention, S3 specifically includes updating the global time series graph according to the intensity of topological changes: After receiving the topology increment information, the cloud platform initiates the corresponding update strategy based on the level of change intensity: Regular changes are handled using batch asynchronous updates, significant changes using incremental synchronous updates, and dramatic changes using priority queue updates. During the update process, distribution network physical constraint rules are introduced to correct the global timing diagram. The physical constraint rules include power flow correlation constraints derived from Kirchhoff's laws and equipment capacity limit constraints, to ensure that the updated global timing diagram conforms to the operation rules of the distribution network. To quantify the impact of physical constraints on the global timing graph, a power flow consistency verification algorithm is introduced, with the specific formula as follows: Based on Kirchhoff's current law, the nodes in the global timing graph Verify the trend ; like Based on equipment capacity constraints Correct the node current value, and use the corrected formula: , ; The corrected current value is used to update the node attributes of the global timing graph to ensure that subsequent source analysis conforms to physical laws; in, The current deviation rate, expressed as a percentage, reflects the degree of deviation between the nodal current and Kirchhoff's laws. For inflow node The sum of currents, expressed in amperes (A); outflow node The sum of currents, expressed in amperes (A); The corrected inflow node current is ensured to comply with Kirchhoff's laws; The corrected outflow node current and the corrected inflow current satisfy a conservation relationship; For nodes The actual load power is expressed in kilovolt-amperes (kVA). For nodes The rated capacity, i.e. the maximum allowable load power of the equipment, is expressed in kilovolt-amperes (kVA). Based on Kirchhoff's current law (the sum of the currents flowing into a node equals the sum of the currents flowing out), calculate the relative deviation rate between the inflow and outflow currents. Quantify the consistency deviation of node power flow in the global time sequence graph; When the deviation rate exceeds the threshold of 5%, the inflow and outflow currents are reversed according to the magnitude of the deviation rate. The inflow current is reduced by half the deviation rate, and the outflow current is increased by the same proportion to ensure that the corrected current conforms to Kirchhoff's laws, while also taking into account equipment capacity constraints. To avoid the current exceeding the rated capacity of the equipment after correction; By verifying and correcting power flow consistency, we ensure that the global time sequence diagram conforms to the physical operation law of the distribution network and avoids power flow data contradictions caused by topology updates. By combining equipment capacity constraints to correct the current, we prevent the model from performing source analysis based on data that does not conform to physical rules. This ensures the accuracy of fine-grained source tracing of faults from the bottom up and avoids analysis results that violate the principles of the power system.

[0025] In an embodiment of the present invention, S3, which combines reinforcement learning and fault propagation physics rules to perform fine-grained source tracing of fault causes, specifically includes: The cloud platform uses the updated global time series graph as its environment to build a reinforcement learning agent that integrates physical rules. The agent's state space includes node physical parameters and topological features, while its action space is the search direction for root cause tracing paths. A hybrid reward function is set up, which includes path length reward, physical rule matching degree reward and root cause feature similarity reward. The physical rule matching degree reward is used to penalize the tracing path that does not conform to the physical law of fault propagation. Through the interaction and iteration between the intelligent agent and the environment, the shortest tracing path that conforms to physical rules is generated to locate the root cause device and key propagation nodes of the fault. To quantify the hybrid reward function, a multi-dimensional reward calculation algorithm is introduced, the specific formula of which is as follows: ; The reward for the reinforcement learning agent is calculated from three dimensions: path length, physical rule matching degree, and root cause feature similarity. The physical rule matching degree has the highest weight of 0.5 to ensure that the tracing path conforms to the laws of electrical physics; the path length has a weight of 0.3 to encourage the agent to search for short paths to improve efficiency; and the root cause feature similarity has a weight of 0.2 to guide the agent to focus on nodes similar to historical root cause features, balancing efficiency and accuracy. The agent updates the action selection probability using the policy gradient algorithm. The policy update formula is as follows: ; in, The reward value is a mixed value; the larger the value, the better the agent's current tracing action. This is the length of the tracing path currently being searched by the agent, i.e., the number of nodes contained in the path; The shortest historical tracing path length is used as the benchmark for path length rewards; The physical rule matching degree is calculated from the power flow deviation rate. The value ranges from [0,1], and the larger the value, the higher the degree to which the path conforms to physical rules; Root cause feature similarity, which is the cosine similarity between the current node feature and the historical root cause features, taking values ​​[0,1]. The larger the value, the higher the probability that the node is the root cause. , , The reward weights correspond to path length, physical rule matching degree, and root cause feature similarity, respectively, and the sum of the weights is 1. The gradient of the objective function of the policy network is used to update the policy parameters; The parameter set for the reinforcement learning policy network; In the state Select action The probability of the action selection strategy of the agent; This is a discount factor, with a default value of 0.9, used to weigh the importance of current rewards against future rewards; For the next state The value function is derived from the time-series graph attention weights. Weighted calculation reflects the potential value of the next state; The policy gradient algorithm is used to calculate the "current reward + next state value function". "With the goal of updating the agent's action selection probabilities." This enables the agent to gradually learn a tracing strategy that conforms to physical rules, has the shortest path, and matches the root cause. By using a multi-dimensional hybrid reward function, the reinforcement learning agent is guided to simultaneously consider physical rules, path efficiency, and root cause matching during the source tracing process, avoiding problems such as "short path but violating physical rules" or "compliant with rules but redundant path". Combined with the policy gradient algorithm, the action policy is continuously optimized, enabling the agent to adapt to different topological scenarios and output accurate fault root causes and propagation paths, providing clear fault handling basis for operation and maintenance personnel.

[0026] In an embodiment of the present invention, the aggregation of model parameters of various heterogeneous edge terminals through a hierarchical lightweight federated learning mechanism in S4 specifically includes: The cloud platform divides federated learning levels according to the power distribution network partitions. Edge terminals within the same partition form sub-federations. Each sub-federation first performs local parameter aggregation to generate partition-level feature parameters. The cloud platform aggregates feature parameters at each partition level to form global optimization parameters. During the aggregation process, a dynamic adjustment strategy for partition weights is adopted, assigning higher aggregation weights to partitions with high accuracy in recent fault analysis. The feature parameters only include the core parameters of the time series graph attention matrix and the fault feature mapping coefficients, and the parameter size is further compressed through feature dimensionality reduction techniques; To quantify the dimensionality reduction effect of partition aggregation weights and parameters, a hierarchical aggregation weight algorithm and a dimensionality reduction error calculation formula are introduced, as follows: Partition aggregation weight calculation: ; in, For the first The aggregate weight of each partition takes a value of [0,1], and the sum of the weights of all partitions is 1. For the first The accuracy of recent (e.g., 7 days) fault analysis for each partition reflects the quality of the partition data and the model; For the first The number of edge terminals within a partition reflects the size of the partition data: This represents the total number of distribution network partitions, i.e., the number of partitions participating in federated learning. Partition-level feature parameters Aggregate into global parameters , ; in, For the first The region-level feature parameters of each partition are determined by the attention weights of the terminal models within that partition. composition; The global optimization parameters are obtained by aggregating the parameters of each partition according to their weights. The weight of each partition in the global aggregation is calculated based on the product of "recent analysis accuracy of the partition" and "number of terminals in the partition". The higher the accuracy and the more terminals a partition has, the greater its weight, ensuring that the global parameters can absorb more features of high-quality and large-scale data. Feature dimensionality reduction error calculation: ; in, The feature reduction error is represented by a value of [0,1]. A smaller value indicates less feature loss during the dimensionality reduction process. The original parameter dimension, i.e., the number of feature dimensions of the global parameters before dimensionality reduction; The dimension of the parameters after dimensionality reduction, i.e. the number of principal components retained by PCA; The first parameter is the covariance matrix of the parameter. Each eigenvalue reflects the eigenvalue importance of the principal component. Principal component analysis is used to calculate the proportion of parameter eigenvalues ​​before and after dimensionality reduction, thus obtaining the dimensionality reduction error. The feature loss during the dimensionality reduction process is quantified to ensure that the parameters retain core features after dimensionality reduction, and to provide a basis for adjusting the pruning ratio of the edge model. By dynamically adjusting partition weights, global parameters can prioritize the fusion of partition features with high analysis accuracy and large data scale, thereby improving the analytical capabilities of the global model. Combined with PCA dimensionality reduction and error control, while compressing parameter scale to reduce transmission and storage costs, core features are not lost, providing high-quality global optimization parameters for edge terminal model iteration and strengthening edge-cloud collaboration.

[0027] In an embodiment of the present invention, S4 further includes a privacy-preserving dynamic adaptation mechanism: The injection intensity of differential privacy noise is dynamically adjusted according to the privacy sensitivity level of different partitions, and the privacy sensitivity level is positively correlated with the noise intensity. Before parameter aggregation, federated distillation is performed on the partition-level feature parameters to extract common feature patterns from the parameters and remove personalized parameter details that may leak partition topology privacy, thus ensuring data privacy and security during the aggregation process. To quantify the intensity of differential privacy noise, a noise injection algorithm is introduced, the specific formula of which is as follows: ; Partition parameters after noise injection : ; in, The privacy sensitivity level of the partition, with a value range of [0,1], is calculated based on the proportion of user data within the partition. , The number of users is affected by the partitioning. This represents the total number of users. A larger value indicates that the partition data involves more user privacy and a stronger need for privacy protection. The standard deviation of Gaussian noise is represented by a value that indicates a higher level of injected noise and stronger privacy protection, but also a corresponding increase in the risk of parameter distortion. This is the noise figure, with a default value of 1.0. It is used to fine-tune the mapping relationship between noise intensity and privacy level to adapt to the privacy control requirements of different power distribution networks. The probability of privacy leakage is set to 1e-5 (0.001%) by default, which means that the upper limit of the probability of inferring the original privacy data by adding noise parameters is used to ensure that privacy risks are controllable. The differential privacy budget, with a default value of 0.1, is used to measure the strength of privacy protection. The smaller the value, the more stringent the privacy protection. It is the core control indicator of the differential privacy framework. For the first The original region-level feature parameters of each partition are composed of key features of the edge terminal model within the partition (such as attention weights and fault mapping coefficients), and may contain privacy-sensitive information. These are the partition feature parameters after noise injection, used to participate in hierarchical federated aggregation, thus avoiding the leakage of original privacy data; With a mean of 0 and a variance of The Gaussian distribution of the probability density function has a bell-shaped shape and can generate continuous and symmetrical random noise, which is suitable for scenarios that protect parameter privacy. Based on "Partition Privacy Sensitivity Level" "as the core variable, combined with the classic "privacy budget" concept in the field of differential privacy" - Leakage probability "Framework, calculate the standard deviation of Gaussian noise" Among them, the privacy sensitivity level directly determines the noise intensity. The higher the level, the larger the noise standard deviation, thereby ensuring the parameter privacy of highly sensitive areas (such as areas containing a large amount of user electricity data); at the same time, the overall privacy risk is controlled by privacy budget and leakage probability to avoid excessive noise causing parameter failure. It conforms to a Gaussian distribution Random noise injection partition original parameters Generate noisy parameters The symmetry and continuity of the Gaussian distribution ensure that noise interference on parameter features is uniform, thus masking the private details of the original parameters while preserving the core distribution features of the parameters to the greatest extent, without affecting the effectiveness of subsequent global aggregation. By dynamically matching "privacy level - noise intensity", differentiated privacy protection is achieved for different zones. This satisfies the strict privacy requirements of highly sensitive areas (such as densely populated urban core areas) while avoiding parameter failure caused by excessive noise in low-sensitivity areas (such as remote transformer areas). Combining the characteristics of Gaussian noise, the core features of parameters are preserved to the greatest extent while protecting data privacy. This ensures that the model still has accurate fault root cause analysis capabilities after global federated aggregation, thus resolving the contradiction between data privacy protection and model collaborative optimization in the distribution network.

[0028] In embodiments of the present invention, a full-process adaptive optimization step is also included: A dynamic evaluation module was built to monitor the edge coarse localization accuracy, cloud fine source tracing response time, and federated learning parameter aggregation efficiency in real time, forming a multi-dimensional evaluation index system. Based on a multi-dimensional evaluation index system, an index correlation analysis model is constructed to quantify the mapping relationship between edge-side model pruning ratio, topology incremental upload cycle, federated learning aggregation frequency and various evaluation indicators. The core parameters are automatically adjusted based on the mapping relationship: when the coarse localization accuracy on the edge side is lower than the preset threshold, the model pruning ratio is reduced to retain more feature extraction modules; when the cloud fine-line tracing response time exceeds the threshold, the topology incremental upload cycle is shortened and the cloud update strategy is optimized; when the efficiency of federated learning parameter aggregation decreases, the partition aggregation weight calculation rules are adjusted. It provides a visual operation and maintenance interface, which supports operation and maintenance personnel to manually intervene and optimize parameters based on the actual operating conditions of the distribution network. The interface has a built-in library of optimization scheme recommendations for typical topology scenarios (such as dense access of distributed power sources and high load scenarios in urban core areas). Operation and maintenance personnel can directly call the recommended schemes or customize parameter thresholds without modifying the underlying code of the model. Regularly generate full-process optimization reports, recording the history of parameter adjustments, the trend of evaluation index changes, and the fluctuation of the accuracy of root cause analysis of failures, providing data support for subsequent model iteration and operation and maintenance strategy optimization.

[0029] To quantify the adjustment range of the end-to-end adaptive optimization, a dynamic parameter adjustment algorithm is introduced, with the specific formula as follows: Let the actual value of the evaluation indicator be The threshold value of the indicator is The indicator weight is (With coarse positioning accuracy weighted at 0.4, source tracing response time weighted at 0.3, and aggregation efficiency weighted at 0.3), the parameter adjustment coefficients are... : ; when When (the indicator does not meet the standard), Parameters according to Adjustments (such as pruning ratio) Upload cycle );when When the indicators are met, To maintain stable or slightly optimize parameters in order to balance performance and resource consumption; in, The actual values ​​of the full-process evaluation indicators can be respectively identified as the three major indicators: "edge-side coarse positioning accuracy", "cloud-based fine-grained source tracing response time" and "federated learning parameter aggregation efficiency". The preset thresholds for evaluating the indicators are set by the operation and maintenance personnel according to the operation and maintenance needs of the distribution network (such as a coarse positioning accuracy threshold of 90% and a response time threshold of 200ms), which serve as the benchmark for judging whether the indicators meet the standards. To evaluate the weights of the indicators, coarse location accuracy is weighted at 0.4, source tracing response time at 0.3, and aggregation efficiency at 0.3. The weight allocation is based on the degree of influence of each indicator on the root cause analysis of the fault. Accuracy directly determines the reliability of the analysis results, so it has the highest weight. This is a parameter adjustment coefficient, ranging from [0.8, 1.5]. A value greater than 1 indicates that parameter adjustment needs to be strengthened (e.g., ...). This indicates a 20% reduction in pruning ratio; a value less than 1 indicates that parameters need to be simplified (e.g., ...). This indicates that the upload cycle has been shortened by 10%. The adjusted pruning ratio for the model is based on the original pruning ratio. and Calculate and retain more model features to improve accuracy; This is the adjusted topology incremental upload cycle. hour Shorten and improve the real-time performance of data transmission to accelerate the source tracing response; By dynamically adjusting coefficients, the system achieves adaptive optimization of parameters throughout the entire process, avoiding parameter mismatch issues caused by changes in distribution network topology and load fluctuations, and ensuring that the model is always in the optimal operating state. Combined with a visual operation and maintenance interface and a typical scenario recommendation library, it not only lowers the technical threshold for operation and maintenance personnel (without modifying the underlying code) but also improves the efficiency and accuracy of parameter adjustment. Regularly generated optimization reports provide data support for subsequent model iterations and operation and maintenance strategy optimization, promoting the continuous evolution of the distribution network fault root cause analysis system and adapting to long-term power grid upgrade needs (such as the access of new power electronic equipment and user load growth).

[0030] The embodiments disclosed in this invention are preferred embodiments, but are not limited thereto. Those skilled in the art can easily understand the spirit of this invention based on the above embodiments and make different extensions and variations, but as long as they do not depart from the spirit of this invention, they are all within the protection scope of this invention.

Claims

1. A data modeling method for root cause analysis of faults in power distribution network equipment, characterized in that, Includes the following steps: S1. The heterogeneous edge terminal of the distribution network acquires equipment operation data and real-time topology information. For the scenario of high-frequency dynamic topology changes in edge terminal computing power, a dynamic lightweight time sequence graph model with online adaptive capability is constructed to output coarse fault root cause localization results. S2. The heterogeneous edge terminal uploads topology incremental information and coarse positioning results based on the intensity of topology changes. The topology incremental information is structured data of local changes in the distribution network topology. S3: The cloud platform updates the global time series graph according to the intensity of topological changes, and combines reinforcement learning and fault propagation physical rules to conduct fine-grained source tracing of fault root causes and generate accurate fault root cause analysis results. S4. The cloud platform aggregates the model parameters of each heterogeneous edge terminal through a hierarchical lightweight federated learning mechanism, generates dynamic optimization rules and distributes them. The heterogeneous edge terminals realize online model iteration based on the optimization rules, forming a closed loop of continuous fault root cause analysis in edge-cloud collaboration.

2. The data modeling method for root cause analysis of power distribution network equipment faults according to claim 1, characterized in that, The construction of a dynamic lightweight time series graph model with online adaptive capabilities in S1 specifically includes: Heterogeneous edge terminals monitor the frequency of topology changes in the distribution network and their own computing power fluctuations in real time, and establish a two-dimensional adaptation model of "topology change frequency - computing power fluctuations". When the frequency of topology changes exceeds a preset threshold, the online model distillation mechanism is triggered. This mechanism dynamically freezes non-core network layers and retains the topology-sensitive feature extraction module to achieve real-time adjustment of the model structure. The topology-sensitive feature extraction module designs exclusive feature mapping rules for the connection relationships of distribution network lines and the associated attributes of equipment to ensure the effective extraction of key features when the topology changes drastically. To quantify the two-dimensional adaptation relationship between "topology change frequency and computing power fluctuation", a model pruning ratio calculation algorithm is introduced, and the specific formula is as follows: ； in, The pruning ratio for the model, For the frequency of topology changes in the distribution network, The reference threshold for topological change frequency. For topological change threshold, Real-time computing power for edge terminals This represents the historical average computing power of edge terminals. This represents the minimum available computing power for edge terminals.

3. The data modeling method for root cause analysis of power distribution network equipment faults according to claim 2, characterized in that, S1 also includes a model adaptation verification mechanism: After the dynamic lightweight time series graph model outputs coarse localization results, the current model's fit is scored by comparing the analysis accuracy under similar historical topology scenarios. If the score is lower than the preset threshold, the model parameter fine-tuning will be automatically initiated, and the weight allocation strategy of the time sequence graph attention mechanism will be adjusted to prioritize strengthening the feature weights that are strongly correlated with the current topology. To quantify the model fitness score and the adjustment range of attention weights, a fitness score algorithm and an attention weight update formula are introduced. Model fit score calculation formula: ； The formula for updating attention weights in a time series graph is as follows: ； in, Score the model fit. This is the topological feature vector of the current scene. For the first Topological feature vectors of similar historical scenarios, For the first Accuracy of model analysis for similar historical scenarios. For the current and the first Cosine similarity of topological features of historical scenes Nodes in the timing graph With nodes The original attention weights, For the adjusted attention weights, For nodes With nodes The topological correlation coefficient, This is the weight adjustment coefficient; the lower the fit, the larger the adjustment coefficient.

4. The data modeling method for root cause analysis of power distribution network equipment faults according to claim 1, characterized in that, The specific aspects of uploading topology incremental information based on the intensity of topology change in S2 include: The heterogeneous edge terminal classifies the intensity of topology changes into three levels: normal change, significant change, and drastic change, based on the number of nodes involved in the topology change, the type of key equipment, and the scope of the fault impact. For routine changes, upload incremental topology information in batches according to a preset cycle; For significant changes, upload information on key change nodes in real time; In the event of drastic changes, a compressed transmission protocol is used to prioritize uploading information on the nodes with the highest fault correlation and changes in connection relationships. The topological incremental information includes a change level identifier, node attribute change records, and edge relationship evolution sequence, and is accompanied by a dynamic timestamp to ensure temporal consistency. To quantify the intensity level of topological change, a change intensity index calculation algorithm is introduced, the specific formula of which is as follows: ； in, This is the topological change intensity index. This represents the number of nodes involved in the topology change. This represents the total number of nodes within the monitoring range of the edge terminal. This refers to the number of critical devices involved in the topology change. The number of users affected by the topology change. This represents the total number of users within the monitoring range of the edge terminal. , , These are the weighting coefficients for each dimension.

5. The data modeling method for root cause analysis of power distribution network equipment faults according to claim 4, characterized in that, S2 also includes incremental information verification preprocessing: Before uploading incremental topology information, the heterogeneous edge terminal performs consistency verification on the incremental data through locally cached historical topology snapshots, eliminating duplicate or invalid change records. The incremental information that passes the verification is feature-encoded, and the unstructured information such as device type and connection status is converted into numerical feature vectors to reduce the amount of data transmitted. To quantify the results of incremental data consistency verification, a data validity coefficient calculation algorithm is introduced, the specific formula of which is as follows: ； During the feature encoding stage, the device type is encoded using one-hot encoding. The connection status is encoded in binary. Finally, the numerical feature vector ; in, The data validity coefficient is... For node attribute change records in incremental data, This is the set of node attributes that are allowed to change in a historical topology snapshot. This records the changes in edge relationships within the incremental data. This is the set of edge relationships that are allowed to change in the historical topology snapshot. A function to calculate the number of elements in a set. For the one-hot encoded vector of the device type, The binary encoded vector of the connection state. This is the final numerical feature vector.

6. The data modeling method for root cause analysis of power distribution network equipment faults according to claim 1, characterized in that, The step of updating the global time series graph according to the intensity of topological change in S3 specifically includes: After receiving the topology increment information, the cloud platform initiates the corresponding update strategy based on the level of change intensity: Regular changes are handled using batch asynchronous updates, significant changes using incremental synchronous updates, and dramatic changes using priority queue updates. During the update process, distribution network physical constraint rules are introduced to correct the global timing diagram. These physical constraint rules include power flow correlation constraints derived from Kirchhoff's laws and equipment capacity limitation constraints, ensuring that the updated global timing diagram conforms to the operation rules of the distribution network. To quantify the impact of physical constraints on the global timing graph, a power flow consistency verification algorithm is introduced, with the specific formula as follows: Based on Kirchhoff's current law, the nodes in the global timing graph Verify the trend ； like Based on equipment capacity constraints Correct the node current value, and use the corrected formula: ， ； in, For tidal current deviation rate, For inflow node The sum of the currents, outflow node The sum of the currents, This is the corrected inflow node current. This is the corrected outflow node current. For nodes Actual load power For nodes Rated capacity.

7. The data modeling method for root cause analysis of power distribution network equipment faults according to claim 6, characterized in that, The detailed source tracing of fault causes in S3, which combines reinforcement learning with fault propagation physics rules, specifically includes: The cloud platform uses the updated global time series graph as its environment to construct a reinforcement learning agent that integrates physical rules. The state space of the agent includes node physical parameters and topological features, and the action space is the search direction for the root cause tracing path. A hybrid reward function is set up, which includes path length reward, physical rule matching degree reward and root cause feature similarity reward, wherein the physical rule matching degree reward is used to penalize the tracing path that does not conform to the physical law of fault propagation; Through the interaction and iteration between the intelligent agent and the environment, the shortest tracing path that conforms to physical rules is generated to locate the root cause device and key propagation nodes of the fault. To quantify the hybrid reward function, a multi-dimensional reward calculation algorithm is introduced, the specific formula of which is as follows: ； The agent updates the action selection probability using the policy gradient algorithm. The policy update formula is as follows: ； in, For mixed reward values, This represents the length of the tracing path currently being searched by the agent. This is the shortest historical tracing path length. The physical rule matching degree is calculated from the power flow deviation rate. , For root cause feature similarity, , , As a reward weight, The gradient of the objective function of the policy network. The parameter set for the reinforcement learning policy network; In the state Select action The probability, As a discount factor, For the next state The value function of .

8. The data modeling method for root cause analysis of power distribution network equipment faults according to claim 1, characterized in that, The model parameters of each heterogeneous edge terminal aggregated through the hierarchical lightweight federated learning mechanism in S4 specifically include: The cloud platform divides federated learning levels according to the power distribution network partitions. Edge terminals within the same partition form sub-federations. Each sub-federation first performs local parameter aggregation to generate partition-level feature parameters. The cloud platform aggregates feature parameters at each partition level to form global optimization parameters. During the aggregation process, a dynamic adjustment strategy for partition weights is adopted, assigning higher aggregation weights to partitions with high accuracy in recent fault analysis. The feature parameters include only the core parameters of the time series graph attention matrix and the fault feature mapping coefficients, and the parameter size is further compressed through feature dimensionality reduction technology. To quantify the dimensionality reduction effect of partition aggregation weights and parameters, a hierarchical aggregation weight algorithm and a dimensionality reduction error calculation formula are introduced, as follows: Partition aggregation weight calculation: ； in, For the first The aggregate weight of each partition, For the first Accuracy of recent fault analysis for each partition. For the first Number of edge terminals within each partition This represents the total number of distribution network zones. Partition-level feature parameters Aggregate into global parameters , ; in, For the first The region-level feature parameters of each partition. Parameters are optimized globally; Feature dimensionality reduction error calculation: ； in, For feature dimensionality reduction error, Original parameter dimensions Dimensionality of parameters after dimensionality reduction The first parameter is the covariance matrix of the parameter. Each feature value.

9. A data modeling method for root cause analysis of faults in power distribution network equipment according to claim 8, characterized in that, S4 also includes a privacy-preserving dynamic adaptation mechanism: The injection intensity of differential privacy noise is dynamically adjusted according to the privacy sensitivity level of different partitions, and the privacy sensitivity level is positively correlated with the noise intensity. Before parameter aggregation, federated distillation is performed on the partition-level feature parameters to extract common feature patterns from the parameters and remove personalized parameter details that may leak partition topology privacy, thus ensuring data privacy and security during the aggregation process. To quantify the intensity of differential privacy noise, a noise injection algorithm is introduced, the specific formula of which is as follows: ； Partition parameters after noise injection : ; in, The privacy sensitivity level of a partition is calculated based on the percentage of user data within that partition. , The number of users is affected by the partitioning. Total number of users The standard deviation of Gaussian noise. Noise figure To increase the probability of privacy leaks, For differential privacy budgeting, For the first The original region-level feature parameters of each partition. These are the partition feature parameters after noise injection. With a mean of 0 and a variance of The Gaussian distribution.

Images