Self-healing topology reconstruction method after fault unit isolation of modular high-voltage power quality equipment

Abstract

The present invention discloses a method for self-healing topology reconstruction after isolation of a faulty unit of modular high-voltage power quality equipment, comprising: collecting faulty unit identification information and real-time grid operation data, and fusing them into a comprehensive context representation vector; mapping the vector to a pre-dimensionalized topological feature space, performing interpolation calculations between anchor point topologies through multi-core radial basis functions, generating a main topology architecture, and then determining the sub-module connection relationship through a progressive optimization strategy to form a preliminary topology scheme for the edge; comparing the scheme with a standard topology template to extract differences, and using context-dependent coding and sparsity-based entropy coding to generate a compressed difference data packet; parsing the data packet on the cloud to reconstruct complete topology information, generate an optimized topology scheme, and send it to the edge for execution through incremental update instructions. The present invention realizes the transition from discrete topology selection to continuous topology generation, and the edge-cloud layered collaborative optimization significantly improves the reconstruction quality while ensuring millisecond-level response.

Description

Technical Field

[0001] The present invention belongs to the field of power quality optimization, and in particular to a self-healing topology reconstruction method after isolation of a fault unit of modular high-voltage power quality equipment. Background Art

[0002] As power systems continue to increase their requirements for power supply reliability and power quality, modular high-voltage power quality equipment is playing an increasingly important role in modern power grids. This type of equipment achieves flexible functional configuration and redundant protection through modular design. However, when certain modules fail, isolating the faulty unit and reconstructing the optimal operating topology within milliseconds is directly related to the stable operation and power supply continuity of the power grid. Traditional manual maintenance and preset switching solutions can no longer meet the urgent need for rapid self-healing capabilities in modern power grids. Therefore, research on intelligent self-healing topology reconstruction methods after faulty unit isolation has theoretical significance and practical value. This technological breakthrough will provide strong support for improving power grid resilience, ensuring the power supply of critical loads, and reducing power outage losses.

[0003] Current research focuses on table lookup methods based on preset topology libraries, rule-based reasoning methods based on expert systems, and global search methods based on optimization algorithms. The preset topology library method calculates the optimal topology configuration for various fault scenarios offline, and then directly retrieves the corresponding solution based on the fault type at runtime through table lookup, offering the advantage of fast response time. The expert system method transforms the empirical knowledge of domain experts into a set of if-then rules, implementing fault diagnosis and topology reconstruction decisions through inference mechanisms. Global search methods based on genetic algorithms, particle swarm algorithms, and other algorithms can solve topology optimization problems online and theoretically obtain the global optimal solution. In recent years, some research has begun exploring machine learning-based methods, such as using neural networks to predict the optimal topology configuration or employing reinforcement learning to train intelligent agents for topology decision-making. The rise of edge computing technology has also provided new ideas for rapid on-site processing. Some research has proposed distributed topology reconstruction frameworks that collaborate between edge and central ends.

[0004] However, existing technical solutions have two key technical bottlenecks. The first is the discrete limitation of topological representation: traditional methods regard topology reconstruction as a discrete optimization problem of selecting the most matching items from a limited number of preset solutions. This discretization process makes it impossible for the system to generate an intermediate state topology that is truly adapted to the current specific fault scenario. The second is the problem of computing granularity mismatch in edge-cloud collaboration: existing edge-cloud collaboration solutions mostly use a simple task splitting strategy. This coarse-grained division of labor leads to insufficient utilization of edge computing resources, which may cause incompatibility or performance loss in the actual execution of the optimization results, making it difficult to achieve true global optimization while ensuring real-time performance. Summary of the Invention

[0005] The purpose of the invention is to provide a self-healing topology reconstruction method after the fault unit of modular high-voltage power quality equipment is isolated, in order to solve at least one technical problem existing in the prior art.

[0006] The technical solution is a self-healing topology reconstruction method for modular high-voltage power quality equipment after faulty units are isolated, including:

[0007] Fault unit identification information and real-time grid operation data are collected and integrated into a comprehensive context representation vector. This vector is then mapped to a pre-dimensionality-reduced topological feature space and interpolated between anchor topologies to generate a preliminary edge topology solution.

[0008] Compare the preliminary topology solution at the edge with the pre-stored standard topology template, extract the structural differences and parameter deviations, generate a compressed difference data packet, and send it to the cloud;

[0009] The cloud parses and compresses the difference data packets and reconstructs the complete topology information based on the standard topology template to generate an optimized topology solution. Based on this, the adjustments are sent to the edge through incremental update instructions to perform the final topology reconstruction.

[0010] Beneficial effects: The present invention breaks through the limitations of the traditional discrete table lookup mode, gets rid of the limitations of discrete selection, realizes the transition from discrete topology selection to continuous topology generation, and improves the reconstruction quality while ensuring millisecond-level response. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 A flowchart of the steps of a method for self-healing topology reconstruction after isolation of a faulty unit of modular high-voltage power quality equipment provided in an embodiment of the present application.

[0012] Figure 2 A flowchart of the steps for fusing into a comprehensive context representation vector provided in an embodiment of the present application.

[0013] Figure 3 A flowchart of the steps of fusion encoding into a comprehensive context representation vector of a predetermined dimension provided in an embodiment of the present application.

[0014] Figure 4 A flowchart of the steps for generating a preliminary edge topology solution provided in an embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0016] It should be noted that to clearly illustrate the steps of this application, serial numbers are assigned to each step in the specification. These serial numbers are for illustrative purposes only and do not limit the order in which the steps must be executed. In actual operation, depending on the technical requirements of the specific implementation scenario, the steps may be executed in a different order than shown in the specification, and in some cases, parallel processing between steps may be implemented.

[0017] Research has found that when faced with complex faults or new fault modes, the standard topologies in the preset library often cannot be accurately matched, and only suboptimal approximations can be selected, which affects the reconstruction quality. Existing edge-cloud collaboration solutions often use a simple task splitting strategy, where the edge is responsible for data collection and preprocessing, while the cloud is responsible for topology optimization calculations. This coarse-grained division of labor leads to insufficient utilization of edge computing resources, and the cloud optimization process lacks accurate perception of the edge's real-time constraints.

[0018] like Figure 1 As shown, a method for self-healing topology reconstruction after isolation of a faulty unit of a modular high-voltage power quality device is provided, comprising the following steps:

[0019] Collect fault unit identification information and real-time grid operation data and fuse them into a comprehensive context representation vector;

[0020] Map the comprehensive context representation vector to the pre-dimensionality-reduced topological feature space, and perform interpolation calculations between the anchor point topologies to generate a preliminary edge-end topology solution;

[0021] Compare the preliminary topology solution at the edge with the pre-stored standard topology template, extract the structural differences and parameter deviations, generate a compressed difference data packet, and send it to the cloud;

[0022] The cloud parses and compresses the difference data packets and reconstructs the complete topology information based on the standard topology template to generate an optimized topology solution. Based on the optimized topology solution, the adjustments are sent to the edge through incremental update instructions to perform the final topology reconstruction.

[0023] Specifically, the fault unit identification information and real-time power grid operation data are collected, and a multi-channel feature extractor is used to separate the fault features from the power grid status features. The two are then fused into a low-dimensional comprehensive context representation vector through a lightweight context encoder to achieve a unified representation of the fault scenario and the operating environment. The comprehensive context representation vector is mapped to the pre-dimensionality-reduced topological feature space, and continuous interpolation calculations are performed between the anchor topologies through radial basis functions. The key main topology architecture is first determined within 5-10ms, and then the sub-module connection relationship is optimized within 10-20ms. Finally, a preliminary edge topology solution containing the topology structure and control parameters is generated. The preliminary edge topology solution is compared with the pre-stored standard topology template to extract structural differences and parameter deviations. Adaptive entropy coding technology is used to generate compressed difference data packets. The priority queue mechanism ensures that key information is transmitted first, while maintaining the edge's real-time monitoring of the power grid and the execution capability of the preliminary topology. The cloud parses and compresses differential data packets and reconstructs complete topology information based on standard topology templates. It uses high-precision power grid models for simulation verification and parameter optimization to generate an optimized topology solution that takes long-term stability into consideration. Adjustment instructions are sent to the edge through incremental updates, executing the final topology reconstruction and monitoring changes in key performance indicators in real time.

[0024] Exemplarily, the construction of a topological feature space includes: collecting a topological data set containing a variety of typical topological structures and their performance indicators; defining an adjacency matrix representation for each topological structure; using an autoencoder neural network to encode the high-dimensional adjacency matrix into a low-dimensional feature vector (i.e., anchor point coordinates), and training the network so that it can reconstruct the original topology based on the feature vector; finally, using the trained encoder part as the mapping model.

[0025] like Figure 2 As shown, according to one aspect of the present application, the fusion is a comprehensive context representation vector, including:

[0026] Collect fault unit identification information, generate an initial fault feature set, apply feature selection algorithms to eliminate redundant features, extract core fault indicators and generate a streamlined fault feature set;

[0027] Collect real-time grid operation data, generate an initial grid state feature set, perform time-frequency analysis, extract load characteristics and grid fluctuation patterns, and generate a simplified grid state feature set;

[0028] The simplified fault feature set and the simplified grid status feature set are fused and encoded into a comprehensive context representation vector of a predetermined dimension.

[0029] Specifically, fault unit identification information includes the fault unit ID, fault type code, fault timestamp, and associated measurement values; real-time grid operation data includes load power curves, voltage harmonic content, power factor, and system frequency. A lightweight context encoder is constructed and applied to fuse and encode a reduced fault feature set and a reduced grid state feature set into a 4- to 6-dimensional comprehensive context representation vector. The encoder uses a nonlinear transformation function designed based on power system expert knowledge to ensure that key information is preserved in a low-dimensional space.

[0030] like Figure 3 As shown, according to one aspect of the present application, the fusion encoding is a comprehensive context representation vector of a predetermined dimension, including:

[0031] Performing linear normalization processing on the simplified fault and simplified grid state feature sets to generate standardized fault features and standardized grid state features;

[0032] The standardized fault features are multiplied by a preset feature importance weight matrix to generate weighted fault features, and the standardized grid state features are transformed using a nonlinear transformation function constructed based on a typical characteristic curve of the power system to generate transformed grid features;

[0033] A cross-correlation analysis is performed on the weighted fault features and the transformed power grid features to generate a cross-correlation feature matrix, which is then subjected to dimensionality reduction and compression processing to generate a comprehensive context representation vector.

[0034] Specifically, the simplified fault feature set and the simplified grid state feature set are read, and each feature is normalized using a linear normalization function to generate standardized fault features and standardized grid state features. A feature importance weight matrix is ​​constructed, where the weight values ​​are preset based on expert knowledge and verified using historical data. The standardized fault features are read and multiplied by the corresponding parts of the weight matrix to generate weighted fault features. The standardized grid state features are read, and each grid feature is independently transformed using a set of nonlinear transformation functions designed for the characteristics of typical power system characteristic curves to generate transformed grid features. The weighted fault features and transformed grid features are read, and an adaptive feature fusion algorithm is used for cross-correlation analysis. This algorithm dynamically adjusts the correlation weights between fault features and grid state features to highlight the mutual influence between the two and generate a cross-correlation feature matrix. The cross-correlation feature matrix is ​​read, and principal component analysis is used for dimensionality reduction. The principal components with a cumulative contribution rate exceeding 95% are retained to obtain a preliminary reduced-dimensionality feature vector. Read the preliminary dimensionality reduction feature vector and compress the feature x through the nonlinear mapping function f(x)=tanh(αx), where α is an adaptive parameter that is dynamically adjusted according to the feature distribution to ensure that key information is retained in the low-dimensional space, and finally generate a 4-6 dimensional comprehensive context representation vector.

[0035] It should be noted that the tanh function was chosen because its S-shaped curve effectively simulates the saturation effect common in power systems, effectively normalizing the input eigenvalues ​​to the (-1, 1) range and enhancing the model's nonlinear expression capabilities. Similar polynomial and piecewise functions can be used to fit other power grid data with different characteristics.

[0036] like Figure 4 As shown, according to one aspect of the present application, generating a preliminary edge topology solution includes:

[0037] Input the comprehensive context representation vector into the pre-trained topological feature space mapping model to locate the coordinate position of the current fault scenario in the topological feature space;

[0038] Identify the anchor topology closest to the coordinate position and calculate the similarity weight between the coordinate position and each anchor topology;

[0039] Interpolate between anchor point topologies based on similarity weights to generate a main topology. Based on the main topology, determine the submodule connection relationship through a progressive optimization strategy to generate a submodule topology.

[0040] Integrate the main topology architecture and sub-module topology, configure control parameters, and form a preliminary edge topology solution, including the topology connection matrix and key control parameters.

[0041] Specifically, a multi-core radial basis function interpolation algorithm performs nonlinear interpolation between anchor point topologies based on similarity weights. This algorithm dynamically balances precision across different regions through adaptive kernel width adjustment, rapidly generating a master topology within 5-10ms. This addresses the issue of reduced precision in sparse regions of the feature space encountered by traditional interpolation methods. Based on the master topology, a progressive optimization strategy is employed to determine the submodule connectivity. This strategy decomposes the optimization problem into a series of low-complexity subproblems, generating submodule topologies through iterative solutions, effectively avoiding computational bottlenecks associated with global optimization.

[0042] According to one aspect of the present application, generating a main topology architecture includes:

[0043] Analyze the distribution density of anchor points around the coordinate position, generate a regional density index, adjust the kernel width based on the regional density index through the multi-kernel radial basis function, calculate the kernel function value between the coordinate position and the topology of each anchor point, and generate a kernel function value vector;

[0044] Combine the kernel function value vector and the similarity weight to calculate the normalized hybrid weight vector;

[0045] The anchor topologies are weightedly fused according to the normalized hybrid weight vector and combined with the structural consistency constraints to generate the main topology architecture.

[0046] Specifically, the coordinate position P and the anchor point topology set {T1, T2, ..., Tk} and their weights {w1, w2, ..., wk} in the topological feature space are read, and the density perception algorithm is used to analyze the distribution density of anchor points around P to generate the regional density index D. area . Read the regional density index and dynamically adjust the kernel width parameter σ of the radial basis function according to the density value. Use a larger kernel width in the sparse anchor point area to enhance generalization ability, and use a smaller kernel width in the dense anchor point area to improve accuracy. Generate an adaptive kernel width parameter set Σ={σ1,σ2,...,σk}. Read the anchor point topology set, weights and adaptive kernel width parameter set to construct a multi-kernel radial basis function: φ(P,Ti,σi)=exp(-||P-Ti|| 2 / (2σi 2 )); Calculate the kernel function value between point P and each anchor point, generating a kernel function value vector Φ = {φ1, φ2, ..., φk}. Read the kernel function value vector and weights, and combine the structural information of each anchor point topology to design a hybrid weight calculation formula: ωi = wi·φi / Σ(wj·φj); Generate a normalized hybrid weight vector Ω = {ω1, ω2, ..., ωk}. Read the anchor point topology set and the normalized hybrid weight vector, perform weighted fusion on the topology matrix, and introduce structural consistency constraints to ensure the physical feasibility of the generated topology, generating a preliminary interpolated topology structure. Read the interpolated topology structure, apply a topology correction algorithm to check and correct possible electrical connection conflicts, ensure that all electrical nodes meet Kirchhoff's law constraints, and ultimately generate a physically feasible main topology architecture.

[0047] According to one aspect of the present application, generating a submodule topology includes:

[0048] Divide the main topology into a set of interconnected functional sub-modules and generate a sub-module partitioning diagram;

[0049] Based on the current fault characteristics, an optimization priority is assigned to each submodule in the submodule partitioning diagram to generate a submodule priority queue. The highest priority submodule is extracted from the submodule priority queue and the topology optimization algorithm is applied sequentially to generate an optimized submodule set.

[0050] Optimize the boundary connections between adjacent sub-modules in the sub-module set, eliminate connection conflicts, and generate a sub-module boundary connection plan;

[0051] The optimized sub-module set is combined according to the sub-module boundary connection scheme to generate a sub-module topology.

[0052] Specifically, the main topology architecture is read, and a graph-theory-based decomposition algorithm is used to partition the overall topology into a set of interconnected functional submodules. This generates a submodule partition graph, which contains the boundary nodes and interconnection relationships of each submodule. The submodule partition graph and the comprehensive context representation vector are read, and optimization priorities are assigned to each submodule based on the current fault characteristics. Submodules on the critical path are given higher priorities, generating a submodule priority queue. Based on the submodule priority queue, the highest-priority submodules are sequentially extracted, and a lightweight topology optimization algorithm is applied to their internal connection structures. This algorithm uses a greedy strategy to find the local optimal connection solution in the shortest possible time, generating an optimized submodule set. The optimized submodule set is read, and an interface coordination algorithm is designed and applied to the boundary connections between adjacent submodules to resolve potential connection conflicts, ensuring seamless integration between submodules, and generating a submodule boundary connection solution. The optimized submodule set and submodule boundary connection solution are read, and a progressive splicing algorithm is used to combine the submodules according to their boundary connection relationships. Local fine-tuning is performed to ensure overall consistency, generating a complete submodule topology. Verify the consistency of the sub-module topology against the main topology architecture, detect and repair possible structural differences, ensure that the two are consistent in key nodes and connection paths, and ultimately form a unified sub-module topology.

[0053] According to one aspect of the present application, generating a compressed difference data packet includes:

[0054] Obtain the standard topology template corresponding to the current fault type and the preliminary topology plan of the edge end, compare the structural differences between the preliminary topology plan of the edge end and the standard topology template, and generate a topology structure difference matrix;

[0055] Compare the control parameters of the preliminary topology solution at the edge end with those of the standard topology template, calculate the parameter deviation value, and generate a parameter deviation vector;

[0056] The encoding strategy is adjusted according to the characteristics of the current fault scenario, and the topology structure difference matrix and parameter deviation vector are compressed into an initial difference data packet. The initial difference data packet is then compressed secondary through entropy coding combined with the sparsity characteristics of the power system topology connection to generate a compressed difference data packet.

[0057] Specifically, a context-dependent coding algorithm is used to dynamically adjust the coding strategy based on the characteristics of the current fault scenario. This algorithm predicts coding efficiency by analyzing historical fault patterns and selects the optimal compression method for different types of information. The topology structure difference matrix and parameter deviation vector are compressed into an initial difference data packet. Secondary optimization compression is applied to the initial difference data packet, and entropy coding is performed using the sparsity characteristics of the power system topology connections to generate the final compressed difference data packet. This reduces the amount of transmitted data by 50-70% compared to traditional methods. A priority transmission queue is established, and key decision-making information in the compressed difference data packet is assigned a high priority to ensure priority transmission in bandwidth-constrained situations. Simultaneously, a preliminary topology execution module is activated at the edge to provide real-time control of the power grid.

[0058] For example, standard topology templates are pre-stored in a database on the edge or cloud. These templates are optimized or suboptimal stable topologies for various typical faults (such as single-phase grounding, overcurrent, and overvoltage) derived through offline simulation optimization or based on historical operating experience.

[0059] According to one aspect of the present application, compressing the topology structure difference matrix and the parameter deviation vector into an initial difference data packet includes:

[0060] Combined with the comprehensive context representation vector, the optimal coding strategy for similar fault scenarios is retrieved from the historical coding pattern library to generate an initial coding strategy;

[0061] Dynamically adjust the initial coding strategy according to the current network bandwidth status and generate an adjusted coding strategy, including structural coding and parameter coding;

[0062] Applying structural coding, using a run-length coding algorithm on the topological structure difference matrix to generate coded structural difference data; applying parameter coding, using a differential pulse code modulation method on the parameter deviation vector to generate coded parameter deviation data;

[0063] The encoded structural difference data and the encoded parameter deviation data are organized according to a predetermined format to generate an initial difference data packet, including the structural difference and the parameter deviation.

[0064] Specifically, the topology difference matrix and parameter deviation vector are read, combined with the comprehensive context representation vector, and a similarity matching algorithm is used to retrieve the optimal encoding strategy for similar fault scenarios from the historical encoding pattern library to generate an initial encoding strategy. The initial encoding strategy and the current network bandwidth parameters are read, and a resource adaptive adjustment algorithm is applied to dynamically adjust the encoding compression ratio, improving the compression rate when bandwidth is limited and retaining more details when bandwidth is sufficient, generating an adjusted encoding strategy. The topology difference matrix and the structural encoding portion of the adjusted encoding strategy are read, and a structure-specific run-length encoding algorithm is applied to process the topology differences. This algorithm is optimized for the sparsity and localized clustering characteristics of power system topologies, generating encoded structural difference data. The parameter deviation vector and the parameter encoding portion of the adjusted encoding strategy are read, and differential pulse code modulation (DPCM) with adaptive quantization step size is used to efficiently encode the parameter deviations, generating encoded parameter deviation data. A coding efficiency evaluation metric is constructed based on the actual compression rate and information importance of each data segment during the encoding process. Based on the evaluation results, the encoding strategy is fine-tuned in real time to generate an optimized encoding parameter set. The encoded structural difference data and the encoded parameter deviation data are read, the encoded data blocks are organized according to a predetermined format and verification information is added to generate an initial difference data packet.

[0065] According to one aspect of the present application, generating a compressed difference data packet includes:

[0066] Identify structural patterns and sparse distribution characteristics in the initial difference data packets and generate data characteristic analysis reports;

[0067] Based on the data characteristics analysis report, the difference data is divided into high entropy area and low entropy area according to the size of information entropy, and coding strategies are applied to each area to generate partition coding schemes;

[0068] For structural differences, the run-length coding for sparse matrix regions in the partition coding scheme is adopted to generate structurally optimized coding data;

[0069] For parameter deviation, the context-adaptive arithmetic coding technology in the partition coding scheme is used to generate parameter-optimized coding data;

[0070] The structure-optimized coded data and the parameter-optimized coded data are combined into a structured difference data packet and compressed to generate a compressed difference data packet.

[0071] Specifically, the initial difference data packet is read, and a data analysis algorithm specialized for power topology is applied to identify structural patterns and sparse distribution characteristics in the topology difference data, generating a data characteristic analysis report. Based on the data characteristic analysis report, the difference data is divided into high-entropy and low-entropy regions according to the information entropy. Different encoding strategies are applied to different regions to generate a partitioned encoding scheme. The structural difference portion and partitioned encoding scheme in the initial difference data packet are read, and run-length encoding is applied to the sparse matrix region. The encoding dictionary is optimized by introducing domain knowledge of power topology connections to generate structurally optimized encoded data. The parameter deviation portion and partitioned encoding scheme in the initial difference data packet are read, and context-adaptive arithmetic coding technology is used to dynamically adjust the encoding model based on parameter type and historical distribution to generate parameter-optimized encoded data. The structurally optimized encoded data and parameter-optimized encoded data are read, and a hierarchical data packet structure is constructed. Key topology information is placed at the front of the packet to ensure priority decoding. A moderate amount of redundant information is added to enhance transmission reliability to generate a structured difference data packet. The final compression operation is performed on the structured difference data packet, and lightweight Huffman coding is applied to further compress the entire data packet. At the same time, index information is generated to facilitate fast parsing, and finally a compressed difference data packet is output, which reduces the amount of transmitted data by 50-70% compared with traditional methods.

[0072] According to one aspect of the present application, performing a final topology reconstruction includes:

[0073] The cloud receives compressed differential data packets, decodes them based on the standard topology template, reconstructs the preliminary topology plan at the edge, imports it into the power grid model, performs simulation verification, evaluates performance under various load conditions, and generates a performance evaluation report.

[0074] Based on the performance evaluation report, adjust the topology structure and control parameters to generate an optimized topology solution that takes long-term stability into consideration;

[0075] Compare the optimized topology with the reconstructed preliminary edge topology, extract key differences, and generate incremental update instructions to minimize the amount of transmitted data.

[0076] The incremental update instructions are sent to the edge, which receives and applies the adjustments, executing the final topology reconstruction. Simultaneously, the performance monitoring module is activated to collect and analyze system performance metrics in real time to verify the reconstruction results.

[0077] Specifically, the process for generating an optimized topology solution involves reading the performance evaluation report, extracting key performance indicators (such as voltage harmonic distortion, power factor, and dynamic response time), constructing a set of objective functions for the multi-objective optimization problem, and generating an optimization target model. Based on the optimization target model and current grid operating constraints, a set of topology optimization constraints, including electrical connection constraints, power flow constraints, and equipment capacity constraints, is constructed to generate a constraint set. The preliminary edge topology solution is read, its topological structural characteristics and parameter sensitivity are analyzed, and the key parameters and structural nodes that most impact system performance are identified. A sensitivity analysis report is generated. Based on the sensitivity analysis report, a parameter tuning strategy is designed, applying a particle swarm optimization algorithm to key control parameters. This algorithm dynamically adjusts inertia weights and learning factors to adapt to the nonlinear characteristics of the power system, generating an optimized control parameter set. The topological structure portion of the preliminary edge topology solution and the sensitivity analysis report are read, and a graph-based local structural optimization algorithm is applied to key topological nodes. Connections are fine-tuned while maintaining the stability of the main structure, generating an optimized topology. Combine the optimized control parameter set and the optimized topology structure to carry out system-level integration and final verification to ensure the coordination and consistency of parameters and structure, and finally generate a complete optimized topology solution. At the same time, the key decision paths in the optimization process are recorded to facilitate subsequent interpretation and improvement.

[0078] This embodiment realizes the conversion of topological representation from discrete to continuous, breaks through the discrete query mode of the traditional preset topology library, represents the topological structure as a continuous and differentiable feature space, realizes the interpolation generation capability, and improves the system's adaptability to unseen fault scenarios. This conversion essentially changes the idea of ​​topology reconstruction, from selection to generation. The topology reconstruction problem is decomposed into three levels: main topology architecture, sub-module topology, and parameter fine-tuning, and is reasonably distributed between the edge and the cloud according to time sensitivity, which ensures both system response speed and optimization quality. This embodiment does not rely on enumerating all possible fault conditions, but achieves generalization capabilities for unseen scenarios through feature space mapping. It is specially optimized for the topology reconstruction scenario of power quality equipment, taking into account millisecond-level response requirements and power system characteristics, and provides a domain-specific solution.

[0079] The following describes the relevant processes and details through several cases.

[0080] For example, when the modular high-voltage power quality equipment of a 110kV substation is in operation, an overcurrent fault occurs in its module M3, and the system needs to perform self-healing topology reconstruction. This method mainly includes the following stages:

[0081] Step S100: context perception and feature extraction, i.e. collecting information at the edge and performing unified representation.

[0082] Step S101: Faulty unit identification information and real-time grid operation data are collected and integrated into a comprehensive context representation vector. In this embodiment, the comprehensive context representation vector encodes both the discrete characteristics of the fault (e.g., type and location) and the continuous state of the grid environment (e.g., load and harmonics), providing a comprehensive, quantified scenario description for subsequent topology generation. Since the optimal topology reconstruction solution depends not only on the fault itself but also on the current grid operating conditions, integrating these two factors can avoid suboptimal decisions caused by incomplete information.

[0083] Specifically, this step can be broken down into:

[0084] Step S101-A: Collect fault unit identification information, generate an initial fault feature set, apply a feature selection algorithm to eliminate redundant features, extract core fault indicators to generate a simplified fault feature set; collect real-time power grid operation data, generate an initial power grid status feature set, perform time-frequency analysis, extract load characteristics and power grid fluctuation patterns to generate a simplified power grid status feature set.

[0085] Feature selection algorithms calculate the information content of features (such as information gain and the Gini coefficient) to eliminate redundant or irrelevant features, thereby reducing computational complexity and noise interference. Time-frequency analysis uses tools such as the Fourier transform to analyze the characteristics of a signal at different times and frequencies, extracting more representative features from raw waveform data.

[0086] The collected fault information may be as follows: fault unit ID M3, type code F_OVERCURRENT_02, current 348.5A, voltage 10.85kV. This forms the initial fault feature set F0. The collected power grid data may be as follows: load power curve, voltage harmonic content (THD_V) 3.2%, power factor 0.92, etc. This forms the initial power grid state feature set E0. An information gain-based feature selection algorithm is applied to F0, and a fast Fourier transform (FFT) analysis is performed on the power grid data E0 to obtain the reduced fault feature set F1 and the reduced power grid state feature set E1.

[0087] Step S101-B: Fusion-encode the simplified fault feature set and the simplified grid status feature set into a comprehensive context representation vector of a predetermined dimension. This step combines two feature sets with different sources and properties into a compact vector for easy input to the subsequent machine learning model.

[0088] It should be noted that in addition to the weighted fusion and nonlinear transformation methods described in this embodiment, other feature fusion technologies can also be used, such as directly concatenating two feature vectors and then performing dimensionality reduction through a pre-trained autoencoder, or using a model based on an attention mechanism to dynamically assign weights to different features.

[0089] Furthermore, the fusion encoding step includes:

[0090] Step S101-B1, linearly normalize the simplified fault and simplified grid state feature sets; multiply the standardized fault features by a preset feature importance weight matrix; transform the standardized grid state features using a nonlinear transformation function constructed for a typical characteristic curve of the power system; perform cross-correlation analysis on the weighted fault features and the transformed grid features, and perform dimensionality reduction and compression processing to generate a comprehensive context representation vector.

[0091] Specifically, the process is as follows:

[0092] The Z-score standardization method (X-μ) / σ is used to process the feature sets F1 and E1 to obtain F1 norm and E1 norm .

[0093] F1 norm Multiplying it with the weight matrix W_fault based on expert knowledge and historical data verification yields F1weighted. norm Apply nonlinear transformation functions such as tanh(α×E1 norm +β), where α and β are preset scaling and offset parameters. The tanh function is used to simulate the saturation effect commonly seen in power systems, and E1 is obtained. _trans .

[0094] Adaptive feature fusion algorithm is used to perform cross-correlation analysis and generate the cross-correlation matrix Ccross. Subsequently, principal component analysis (PCA) method is applied to reduce the dimension of Ccross, retaining the principal components with a cumulative contribution rate of more than 95%, and obtaining the preliminary dimension reduction vector C pca .

[0095] Through nonlinear mapping functions such as C=tanh(α compress ×C pca ) is finally compressed to obtain a comprehensive context representation vector C preferably with 4-6 dimensions. In this example, the final generated 4-dimensional vector C is [0.720, 0.622, 0.435, 0.143].

[0096] In another embodiment of the present application, the simplified fault feature set and the simplified grid status feature set are fused and encoded into a comprehensive context representation vector of a predetermined dimension, which can also be implemented by using a deep learning model based on a stacked autoencoder (SAE).

[0097] Specifically, an SAE model consisting of multiple encoder layers is constructed. The streamlined fault feature set F1 and the precisely monitored grid status feature set E1 are concatenated to form an initial high-dimensional input vector. This high-dimensional vector is input into the pre-trained SAE model. Through layer-by-layer nonlinear transformation and dimensionality reduction, the model automatically learns deep connections and abstract features within the data. Each encoder layer maps its input to a lower-dimensional representation, which serves as the input to the next layer. Ultimately, the output of the model's middlemost hidden layer (bottleneck layer) is the fused comprehensive context representation vector C. The dimensionality of this vector is predetermined by the number of neurons in the bottleneck layer; for example, it can be set to 5 dimensions. This approach enables end-to-end adaptive fusion of features, reducing reliance on manually designed fusion rules (such as weighting and nonlinear function selection) and potentially capturing more complex nonlinear relationships.

[0098] Step S200: edge end topology generation, that is, quickly generating a preliminary and executable topology reconstruction plan at the edge end.

[0099] Step S201 : Map the comprehensive context representation vector to a pre-dimensionality-reduced topological feature space, and perform interpolation calculations between anchor point topologies to generate a preliminary edge-end topology solution.

[0100] In this embodiment, the topological feature space is a pre-constructed multi-dimensional (8-dimensional in this example) continuous vector space, in which each point corresponds to a theoretical topological structure. In this feature space, pre-defined, representative standard topologies (such as star and ring) are represented as fixed points, serving as the basis for interpolation calculations, namely, anchor topologies.

[0101] This step transforms the topology reconstruction problem from a traditional discrete problem of selecting from a finite library to a continuous problem generated in a continuous space. This enables the system to create better intermediate or hybrid topologies that have never appeared in the preset library, improving its adaptability to unknown failure scenarios.

[0102] Specifically, this step includes:

[0103] Step S201-A: input the comprehensive context representation vector into the pre-trained topological feature space mapping model to locate the coordinate position of the current fault scenario in the topological feature space; identify the anchor point topology closest to the coordinate position, and calculate the similarity weight.

[0104] The vector C obtained in the previous step is input into a pre-trained neural network model to obtain its coordinates P in the 8-dimensional topological feature space. Then, by calculating the topology T of each anchor point in the space i The Euclidean distance between them is calculated by applying the Gaussian kernel function w i =exp(-||PTi || 2 / (2σ 2 )) to calculate the similarity weight w i .

[0105] Step S201-B, interpolate between anchor point topologies according to similarity weights to generate a main topology architecture, and based on this, determine the sub-module connection relationship through a progressive optimization strategy to generate a sub-module topology; and finally integrate to form a preliminary edge end topology solution.

[0106] Here, a hierarchical optimization strategy, focusing on the primary and secondary components first, is adopted. Interpolation is used to quickly determine the overall, critical main topology, ensuring a fast response (in this case, within 5-10ms). Then, within the constraints of the main architecture, a low-computation, progressive optimization method is used to determine the specific submodule connections, ensuring a precise solution.

[0107] Furthermore, the process of generating the main topology includes:

[0108] Step S201-B1, analyze the anchor point distribution density around the coordinate position, adjust the kernel width through the multi-kernel radial basis function, and calculate the kernel function value; calculate the normalized hybrid weight vector based on the similarity weight; perform weighted fusion on the anchor point topology according to the normalized hybrid weight vector to generate the main topology architecture.

[0109] Traditional RBF interpolation uses a fixed kernel width, and the accuracy will decrease in data sparse areas. This paper analyzes the anchor point density D area To dynamically adjust the kernel width σ adaptive In sparse anchor point areas, a larger kernel width is used to enhance generalization, while in dense areas, a smaller kernel width is used to improve accuracy. This adaptive mechanism improves the accuracy of interpolation generation.

[0110] The specific implementation process can be as follows:

[0111] Calculate the regional density index D area , and dynamically adjust the kernel width of each anchor point accordingly to obtain the adaptive kernel width parameter set Σ.

[0112] The kernel function value vector Φ is calculated using the multi-kernel radial basis function φ(P,Ti,σi).

[0113] Combined with the similarity weight w i and kernel function value φ i , and the normalized mixing weight vector Ω is calculated.

[0114] Perform weighted fusion T on the structural matrix of each anchor point topology main =∑(Ω i ×T i), while introducing structural consistency constraints (such as ensuring that nodes satisfy Kirchhoff's law), and finally generating a physically feasible main topology architecture T main .

[0115] In another embodiment of the present application, generating an initial topology solution for an edge terminal may also be implemented using a graph neural network.

[0116] Specifically, the generative model consists of an encoder and a GNN decoder. The encoder is responsible for mapping the comprehensive context representation vector C to a graph-level hidden state vector z. The GNN decoder, based on the hidden state vector z, autoregressively generates nodes and edges of the topological graph. For example, the decoder can first predict the type and number of new nodes, and then iteratively predict whether there are connecting edges between each pair of nodes and the properties of the edges (such as impedance, capacity, etc.). At each step of prediction, the GNN decoder takes into account the generated partial graph structure to ensure that the newly generated connections meet basic electrical constraints and topological connectivity. When the decoding process is completed, the complete edge-end preliminary topology solution T containing the topological structure and preliminary parameters is obtained. edge This approach treats topology generation as a unified graph generation task, which can better learn the inherent structural information of the topology.

[0117] Furthermore, the process of generating the submodule topology includes:

[0118] Step S201-B2 divides the main topology into a set of functional submodules; assigns optimization priorities to each submodule based on the current fault characteristics; applies a topology optimization algorithm to each submodule sequentially; eliminates boundary connection conflicts between adjacent submodules, and finally combines them to generate a submodule topology. This progressive optimization strategy decomposes the complex global optimization problem into a series of low-complexity local optimization subproblems. A priority queue ensures that the components with the greatest impact on the system are solved first, effectively avoiding the problem of excessive computation time caused by global search.

[0119] Optionally, the implementation process can be as follows:

[0120] Use the decomposition algorithm based on graph theory to transform T main It is divided into functional sub-modules such as power supply, voltage transformation, and filtering.

[0121] Fault correlation fault determined based on the comprehensive context representation vector C relevance and impact factor factor , calculate the optimization priority of each sub-module score .

[0122] The submodules are taken out from the priority queue one by one, and lightweight optimization methods such as greedy algorithms are applied to their internal connections.

[0123] For the optimized adjacent submodules, the interface coordination algorithm is applied to resolve the boundary connection conflicts and finally splice them into a complete submodule topology T sub .

[0124] The main topology architecture T main and submodule topology T sub Integrate and configure preliminary control parameters to form a preliminary edge topology solution T including topological connection matrix and control parameters edge The entire initial reconstruction process at the edge takes about 19.3ms, meeting the millisecond-level real-time requirements of the power system.

[0125] In addition to the greedy algorithm, you can also use the simulated annealing algorithm.

[0126] Specifically, for each submodule to be optimized taken from the priority queue, its internal connection scheme is regarded as a solution. The simulated annealing algorithm starts with an initial connection scheme and sets the initial temperature T start and a cooling rate α (for example, 0.98). In each iteration, the algorithm randomly perturbs the current solution (for example, by changing a switch state or a connection path) to generate a new solution. It then calculates the change ΔE in the objective function value (for example, a comprehensive evaluation value that takes into account both loss and efficiency) of the new solution. If ΔE is superior (for example, a higher evaluation value), the new solution is accepted. If ΔE is inferior, the algorithm accepts the inferior solution with a certain probability P = exp(-|ΔE| / T), where T is the current temperature. Thereafter, the temperature decreases according to the rule T = T * α. This probabilistic acceptance of bad solutions allows the algorithm to escape from local optimal solutions, thereby increasing the probability of finding the globally optimal internal submodule connection solution. Compared to the greedy strategy, this method increases the likelihood of finding a better solution.

[0127] Step S300: Edge-cloud collaboration and difference compression, that is, generating difference data packets at the edge and sending them to the cloud.

[0128] Step S301, compare the preliminary topology solution at the edge with the pre-stored standard topology template, extract the structural differences and parameter deviations, generate a compressed difference data packet and send it to the cloud. Since the amount of data directly transmitted for the complete topology solution is huge, it will cause network congestion and delay. edge and standard template T std , extracting only the difference information for transmission. This incremental / differential approach significantly reduces the amount of transmitted data (78.8% in this example) and is key to achieving efficient edge-cloud collaboration.

[0129] Specifically, the process of generating a compressed difference data packet includes:

[0130] Step S301-A: obtain a standard topology template corresponding to the current fault type, compare the structure differences to generate a topology structure difference matrix; compare the control parameters to calculate the parameter deviation value, and generate a parameter deviation vector.

[0131] Specifically, by edge With T std The topological connection matrix and the control parameter list are compared element by element to obtain the sparse structural difference matrix D struct and parameter deviation vector D param .

[0132] Step S301-B, adjust the encoding strategy according to the characteristics of the current fault scenario, compress the difference matrix and deviation vector into an initial difference data packet, and perform secondary compression on it through entropy coding combined with the sparsity characteristics of the power system topology connection to generate the final compressed difference data packet.

[0133] This step employs a two-stage compression strategy. The first stage dynamically adjusts the encoding strategy based on the scenario, achieving context-sensitive adaptive compression. The second stage leverages domain knowledge of the inherent sparsity of power topology, further improving the compression ratio through deep optimization using methods such as entropy coding.

[0134] Furthermore, the first stage compression process includes:

[0135] Step S301-B1, combined with the comprehensive context representation vector, retrieves the optimal coding strategy from the historical coding mode library; dynamically adjusts it according to the current network bandwidth conditions; applies structural coding (such as run-length coding) and parameter coding (such as differential pulse code modulation) to generate the initial difference data packet.

[0136] Optionally, the implementation process can be:

[0137] According to the similarity between vector C and historical scenarios, an initial encoding strategy is retrieved from the library.

[0138] According to the current network bandwidth B current With the optimal bandwidth B optimal ratio, dynamically adjust the compression ratio target.

[0139] For sparse D struct Applying structure-specific run-length encoding, D param The DPCM technology with adaptive quantization step size is used for encoding to obtain the initial data packet P0.

[0140] In another embodiment of the present application, the topology structure difference matrix and the parameter deviation vector are compressed into an initial difference data packet, and a dictionary-based LZW compression algorithm may be used.

[0141] Specifically, the topology structure difference matrix D_struct and the parameter deviation vector D_param are first serialized into a one-dimensional data stream. When processing this data stream, the LZW algorithm dynamically builds a string dictionary. The algorithm reads increasingly longer substrings from the input stream and continues to read the next character as long as the substring exists in the dictionary. When encountering a new string that is not in the dictionary, the algorithm outputs the code corresponding to the prefix of the new string (already in the dictionary) and adds the new string to the dictionary. In this way, recurring data patterns (corresponding to common connection patterns or parameter deviation patterns in power topology) can be represented by a single code, thereby achieving efficient compression. This method is relatively simple to implement and does not require the probability distribution of the data to be known in advance like entropy coding, making it very universal.

[0142] Specifically, a special coding dictionary is pre-built, which contains common and meaningful connection patterns or structural fragments in the power system topology. When the encoder encounters these pre-defined patterns, it can directly represent them with short codes in the dictionary instead of encoding long strings of 0s and 1s. The encoding format is (symbol, length, context), and context relevance is introduced. The improvement in coding efficiency is related to the pattern matching rate. This means that when encoding the same symbol, its context is taken into account. For example, a 0 representing a disconnection may be encoded differently if it appears between two power nodes than if it appears between two load nodes. Context-awareness makes the encoding more targeted. Therefore, complex and recurring topological structure fragments can be compressed into a single short code. The context is used to predict the next symbol, so that it can be encoded with fewer information bits, further improving compression efficiency.

[0143] Furthermore, the second stage compression process includes:

[0144] Step S301-B2, identify the structural pattern and sparse distribution characteristics in the initial difference data packet; divide the difference data into high entropy area and low entropy area according to the size of information entropy, and apply different encoding strategies to each area; finally combine and compress to generate the final compressed difference data packet.

[0145] Optionally, the implementation process can be:

[0146] Analyze the data characteristics of P0 and partition it by information entropy.

[0147] For the low-entropy structural difference part, run-length coding based on topological connection prior knowledge optimization is adopted; for the high-entropy parameter deviation part, context-adaptive arithmetic coding technology is adopted.

[0148] The encoded data is constructed into a hierarchical data packet, and lightweight Huffman coding is applied to the whole data packet, and finally a compressed difference data packet P with a size of only 22 bytes is obtained. final .

[0149] Establish a priority transmission mechanism to ensure that key decision-making information in data packets can be transmitted first even when bandwidth is limited.

[0150] Step S400: Cloud optimization and instruction issuance

[0151] This step corresponds to the final phase, where global optimization is performed in the cloud and the final reconstruction is performed.

[0152] In step S401, the cloud parses the compressed difference data packet and reconstructs the complete topology information in combination with the standard topology template to generate an optimized topology solution. Based on this, the adjustment is sent to the edge end through incremental update instructions to perform the final topology reconstruction.

[0153] Since the cloud has powerful computing resources and high-precision power grid models, it can simulate and verify the preliminary plan of the edge end and deeply optimize it, making up for the possible inconsideration (such as long-term stability) of the edge end due to the limited computing power. delta , only the adjustment amount is transmitted, further ensuring the efficiency of communication.

[0154] Specifically, this step includes:

[0155] In step S401-A, the cloud receives and decodes the compressed difference data packet, reconstructs the preliminary topology plan at the edge, imports it into the power grid model for simulation verification, and generates a performance evaluation report. Based on the performance evaluation report, the topology structure and control parameters are adjusted to generate an optimized topology plan that takes long-term stability into consideration.

[0156] Optionally, the implementation process can be:

[0157] Cloud Decoding P final , combined with T std Reconstruction of T edge .

[0158] T edge Import a high-precision power grid simulation model, evaluate its performance under various working conditions, and generate a performance evaluation report R_performance.

[0159] Based on the evaluation report, an optimization function F is constructed, which includes multiple objectives such as voltage harmonics, transmission efficiency, response time, and stability margin.

[0160] The particle swarm optimization (PSO) algorithm is applied to optimize key control parameters; the local structure optimization algorithm based on graph theory is applied to fine-tune key topological nodes.

[0161] Integrate the optimized parameters and structures to generate the final optimized topology solution T opt In this example, the cloud optimization process takes about 13.7ms.

[0162] In addition, in PSO, the dynamic inertia weight w(t) adopts a linear decreasing strategy, w(t)=w max -(w max -w min )×t / t max In the early stage of the algorithm (t is small), w(t) is large (close to w max , such as 0.9), so that particles can explore the entire solution space at a faster speed, which helps to conduct global search and avoid falling into local optimality. In the later stage of the algorithm (t is large), the w(t) value is small (close to w min , such as 0.4), which slows down the particle speed and enables fine local search in the found good area, helping the algorithm converge to the optimal solution.

[0163] The dynamic learning factors c1(t) and c2(t) adopt an asynchronous change strategy. c1(t) (the cognitive component, representing self-learning) decreases linearly, while c2(t) (the social component, representing learning from the group) increases linearly. In the early stages, a larger c1 and smaller c2 encourage particles to be more confident in their historical optimal positions and conduct diversified exploration. In the later stages, a smaller c1 and larger c2 encourage particles to move closer to the global optimal position, accelerating the algorithm's convergence.

[0164] Power system optimization problems often involve multiple local optima. Due to fixed parameters, standard PSO algorithms can easily fall into local extreme points that are not globally optimal in the early stages of the search. In the early stages of the algorithm, high inertia weights and strong self-awareness allow the particle swarm to search for possible optimal areas within the vast solution space. Later in the algorithm, low inertia weights and strong social learning allow the particle swarm to fine-tune itself within the discovered high-quality areas and find the optimal solution.

[0165] The PSO algorithm can also be implemented using a genetic algorithm. The process is as follows:

[0166] For each individual in the population (ie, each topology solution), its comprehensive performance index is calculated through a high-precision simulation model, and its fitness value is obtained according to the multi-objective optimization function F.

[0167] Based on the fitness value, a roulette wheel selection strategy is used to select the best individuals from the current population to enter the next generation. Individuals with higher fitness have a greater probability of being selected.

[0168] The selected individuals are paired and part of their chromosome segments are exchanged with a certain crossover probability (for example, 0.8) to produce new offspring individuals.

[0169] For each newly generated offspring individual, a gene position on its chromosome is randomly altered with a small mutation probability (e.g., 0.05) (e.g., changing the value of a control parameter or a connection state). Through continuous selection, crossover, and mutation, the overall fitness of the population will continue to improve, eventually converging to one or a set of globally optimal topologies, T_opt. This method has powerful global search capabilities and robustness.

[0170] Step S401-B, compare the optimized topology solution with the reconstructed preliminary topology solution at the edge, extract key differences, and generate incremental update instructions; send the incremental update instructions to the edge, which receives and applies adjustments to perform the final topology reconstruction.

[0171] Optionally, the implementation process can be:

[0172] Contrast T opt and T edge , extract the incremental update value Δp of 5 connection point changes and 8 control parameters, and generate the incremental update instruction I of only 8 bytes delta .

[0173] I delta Sent to the edge.

[0174] The edge receives and verifies the instructions, applies adjustments in priority order, and completes the final topology reconstruction in about 3.2ms.

[0175] Through the above steps, this embodiment achieves high-quality self-healing topology reconstruction within 35.8ms after the fault occurs. Test data shows that compared with the system voltage harmonics THD_V before the fault, the reconstructed system improved by 43.8%, and the stability margin SM improved by 47.7%, achieving excellent technical results. The present invention realizes the transition from discrete topology selection to continuous topology generation, and through edge-cloud layered collaborative optimization, significantly improves the reconstruction quality while ensuring millisecond-level response.

[0176] In a specific embodiment of the present application, during the operation of modular high-voltage power quality equipment in a 110kV substation, an overcurrent fault occurs in the equipment module M3, requiring self-healing topology reconstruction after isolating the faulty unit. The specific steps are:

[0177] Step 1: Collect fault unit identification information and real-time grid operation data and fuse them into a comprehensive context representation vector.

[0178] 1.1. The collected original fault signal data includes: fault unit ID: M3; fault type code: F _OVERCURRENT_02 ;Fault timestamp: 2025-05-18 14:30:25.123;Associated measurement values: current amplitude 348.5A, voltage amplitude 10.85kV;Generate the initial fault feature set F0=[348.5, 10.85, 2, 14.30], where the values ​​correspond to current, voltage, fault type code, and time characteristics respectively.

[0179] 1.2. The grid operation status data collected in parallel include: load power curve: P _load =[85.2, 87.1, 89.3, 91.5]MW; voltage harmonic content: THD _V =3.2%; Power factor: cos _φ =0.92; system frequency: f=50.02Hz; generate the initial grid state feature set E0=[85.2, 87.1, 89.3, 91.5, 3.2, 0.92, 50.02].

[0180] 1.3. Apply a feature selection algorithm based on information gain to the initial fault feature set F0, calculate the information gain value of each feature, and exclude features with low information content. After feature selection, extract the core fault indicators and generate a simplified fault feature set F1 = [3.485, 1.085, 2], which is the normalized current, voltage, and fault type.

[0181] 1.4. Perform time-frequency analysis on the initial grid state feature set E0, using a fast Fourier transform to extract the primary frequency components of load variations and identify grid fluctuation patterns. After extracting key features, a reduced grid state feature set E1 = [0.88, 0.032, 0.92] is generated, representing the normalized load level, harmonic content, and power factor.

[0182] 1.5. Standardize the Reduced Feature Set: F1 norm =(F1-μ_F) / σ_F=([3.485, 1.085, 2]-[3.0, 1.0, 1.5]) / [0.5, 0.2, 0.8]=[0.97, 0.425, 0.625]; E1 norm =(E1-μ_E) / σ_E=([0.88, 0.032, 0.92]-[0.8, 0.03, 0.9]) / [0.1, 0.01, 0.05]=[0.8, 0.2, 0.4]; where F1 norm is the standardized fault feature; E1 norm is the standardized grid state feature; μ is the mean; σ is the standard deviation. Multiply the standardized fault feature with the preset weight matrix: F1 weighted =W_fault×F1norm =[0.6, 0.3, 0.1]×[0.97, 0.425, 0.625]=0.6×0.97+0.3×0.425+0.1×0.625=0.582+0.1275+0.0625=0.772; where W _fault is the feature importance weight matrix; F1 weighted is the weighted fault feature. Apply the nonlinear transformation function designed for the power system characteristic curve to the standardized grid state feature: E1 _trans =tanh(α×E1 norm +β)=tanh(2.5×[0.8, 0.2, 0.4]+[0.1, 0.05, 0.08])=tanh([2.1, 0.55, 1.08])=[0.974, 0.502, 0.789]; where α is the scaling parameter 2.5; β is the offset parameter [0.1, 0.05, 0.08]; E1 _trans is the grid feature after transformation. Adaptive feature fusion algorithm is used for cross-correlation analysis:

[0183] C cross =λ1×(F1 weighted ΘE1 _trans )+λ2×(E1 _trans ΘF1 weighted T )=0.7×(0.772×[0.974, 0.502, 0.789])+0.3×([0.974, 0.502, 0.789]×0.772)=0.7×[0.752, 0.388, 0.609]+0.3×[0.752, 0.388, 0.609]=[0.752, 0.388, 0.609]; where Θ is the outer product operation; λ1 and λ2 are the fusion weight coefficients; C cross is the cross-correlation feature matrix, T is transposed. Apply principal component analysis to the cross-correlation feature matrix: First calculate the covariance matrix Σ=C cross T ×C cross ; Calculate the eigenvalues ​​and eigenvectors, select the principal components with a cumulative contribution rate of more than 95%; obtain the preliminary dimensionality reduction eigenvector C pca =[0.752, 0.609, 0.388]. Final compression is performed through a nonlinear mapping function: C = tanh(α compress ×C pca)=tanh(1.2×[0.752, 0.609, 0.388])=tanh([0.902, 0.731, 0.466])=[0.720, 0.622, 0.435]; after adding the time feature dimension, the final 4-dimensional comprehensive context representation vector C=[0.720, 0.622, 0.435, 0.143] is generated; where α compress is the compression parameter 1.2; C is the final comprehensive context representation vector.

[0184] Step 2: Map the comprehensive context representation vector to the topological feature space to generate a preliminary topology solution for the edge.

[0185] 2.1. Input the comprehensive context representation vector C = [0.720, 0.622, 0.435, 0.143] into the pretrained topological feature space mapping model. This model uses a neural network structure to map the 4-dimensional input to the 8-dimensional topological feature space. After forward propagation, the coordinate position of the current fault scenario in the topological feature space is located at P = [0.42, 0.38, 0.15, 0.67, 0.23, 0.54, 0.89, 0.31].

[0186] 2.2. Search in the topological feature space and identify the five anchor topologies closest to the coordinate position P: T1: star topology, coordinates [0.45, 0.35, 0.12, 0.65, 0.25, 0.52, 0.87, 0.29]; T2: ring topology, coordinates [0.38, 0.41, 0.18, 0.69, 0.21, 0.56, 0.91, 0.33]; T3: tree topology, coordinates [0.44, 0 .36, 0.14, 0.66, 0.24, 0.53, 0.88, 0.30]; T4: grid topology, coordinates [0.40, 0.40, 0.16, 0.68, 0.22, 0.55, 0.90, 0.32]; T5: hybrid topology, coordinates [0.43, 0.37, 0.13, 0.67, 0.23, 0.54, 0.89, 0.31]; calculate the similarity weight, using the Gaussian kernel function: w i =exp(-||PT i || 2 / (2σ 2 )), where σ=0.1.

[0187] 2.3. Calculate the anchor point distribution density around the coordinate position P: D area =(1 / K(K-1))∑∑||T i -T_j|| 2 =(1 / 20)×0.52=0.026; where K is the number of anchor points 5; D areais the regional density index. Dynamically adjust the kernel width according to the regional density index: σ adaptive (i)=σ _base ×(1+β×D area ×density factor (i))=0.1×(1+2×0.026×[1.2, 0.9, 1.1, 1.0, 0.8])=[0.1062, 0.1047, 0.1057, 0.1052, 0.1042]; where σ _base is the basic kernel width 0.1; β is the density adjustment coefficient 2; density factor is the local density factor of each anchor point; σ adaptive is the adaptive kernel width parameter set. Construct a multi-kernel radial basis function and calculate the kernel function value: φ i =exp(-||PT i || 2 / (2σ adaptive (i) 2 ));φ1=exp(-0.008 / (2×0.1062 2 ))=exp(-0.008 / 0.0225)=exp(-0.356)=0.700;φ2=exp(-0.012 / (2×0.1047 2 ))=exp(-0.012 / 0.0219)=exp(-0.548)=0.578;φ3=exp(-0.006 / (2×0.1057 2 ))=exp(-0.006 / 0.0223)=exp(-0.269)=0.764;φ4=exp(-0.010 / (2×0.1052 2 ))=exp(-0.010 / 0.0221)=exp(-0.452)=0.636;φ5=exp(-0.002 / (2×0.1042 2 ))=exp(-0.002 / 0.0217)=exp(-0.092)=0.912; where φ i is the kernel function value; the kernel function value vector Φ = [0.700, 0.578, 0.764, 0.636, 0.912]. Combined with the kernel function value vector and similarity weight: mixed _weight_i =w i ×φ i ; mixed _weight=[0.992×0.700, 0.988×0.578, 0.994×0.764, 0.990×0.636, 0.998×0.912]=[0.694, 0.571, 0.760, 0.630, 0.910]; sum_mixed=3.565; normalized mixing weight vector Ω=mixed _weight / sum_mixed=[0.195, 0.160, 0.213, 0.177, 0.255]; where w i is the similarity weight; Ω is the normalized mixed weight vector, mixed _weight is the mixed weight, sum_mixed is the sum of the mixed weights. The anchor topology set is weighted fused according to the normalized mixed weights: T main =∑(Ω i ×T i )=0.195×T1+0.160×T2+0.213×T3+0.177×T4+0.255×T5; At the same time, structural consistency constraints are introduced to ensure that the topology connection matrix meets the electrical constraints; generate the main topology architecture T main , the calculation takes 6.8ms.

[0188] 2.4. Use graph-based decomposition algorithm to divide the main topology architecture: Apply the connectivity analysis and modular decomposition method of the graph to divide the main topology architecture T main Divided into 4 functional sub-modules: power module M_power, transformer module M _trans , filter module M_filter, load module M_load; generate submodule division graph G sub , including the node set and boundary connection relationship of each module. According to the fault characteristics determined by the comprehensive context representation vector, each submodule is assigned an optimization priority: priority score =impact factor ×fault relevance ; Transformer module: priority score =0.9×0.8=0.72; Power module: priority score =0.8×0.7=0.56; filter module: priority score =0.6×0.5=0.30; Load module: priority score =0.4×0.3=0.12; Generate submodule priority queue Q_priority=[M _trans , M_power, M_filter, M_load]. Among them, impact factor is the impact factor, fault relevanceOptimize each submodule in turn based on the priority queue: Apply the greedy algorithm to optimize the internal connection of the transformer module, with the objective function optimization_obj = minimize(loss) + maximize(efficiency); find the optimal connection solution through local search, which takes 2.1ms; complete the optimization of other modules in turn to generate the optimized submodule set {M _transopt , M_power opt , M_filter opt , M_load opt}. Where minimize is the minimization function, loss is the loss, and efficiency is the efficiency. For the boundary connections between adjacent submodules, the interface coordination algorithm is applied: detect connection conflicts conflict_matrix, and for each conflict(i, j), calculate the adjustment cost cost(i, j) = performance_loss + complexity increase ; Select the adjustment plan with the minimum cost to resolve all connection conflicts; Generate the submodule boundary connection plan B_connect to ensure seamless integration. _loss For performance loss, complexity increase To increase the complexity, the optimization submodules are combined through a progressive splicing algorithm: starting from the highest priority module, the modules are connected in sequence according to the submodule boundary connection scheme B_connect; local fine-tuning optimization_adjustment = Δconnection × performance is performed during the connection process. _weight ; Ensure overall consistency; Generate complete submodule topology T sub The total optimization time is 12.5ms. Δconnection is the connection adjustment amount, performance _weight is the performance weight.

[0189] 2.5. Integrate the main topology architecture T main and submodule topology T sub , configure preliminary control parameters, including switch status, regulator setting value, protection parameters, etc. Form a complete preliminary topology solution for the edge end T edge , including an 8×8 topological connection matrix and 24 control parameters.

[0190] Step 3: Compare the preliminary topology plan at the edge with the standard topology template to generate a compressed difference data packet.

[0191] 3.1、Get the standard topology template T corresponding to the current overcurrent fault type std. The preliminary topology scheme T edge With the standard topology template T std Perform element-by-element comparison to identify the difference positions in the topological connection matrix. Generate the topological structure difference matrix D struct , is an 8×8 sparse matrix, where there are connection differences at 12 positions.

[0192] 3.2. Compare the control parameters of the initial topology scheme at the edge with those of the standard topology template and calculate the deviation of each parameter. Generate the parameter deviation vector D param =[0.05, -0.03, 0.08, -0.02, 0.06, 0.04, -0.01, 0.07], corresponding to the relative deviations of 8 key control parameters.

[0193] 3.3, combined with the comprehensive context representation vector C, the historical coding pattern library is retrieved through the similarity matching algorithm: similarity(C, histi )=cosine_similarity(C,C _histi )=C·C _histi / (||C||×||C _histi ||); Find the most similar historical failure scenario similarity max =0.87; get the corresponding optimal encoding strategy E i nit, including the structure compression rate target 0.35 and parameter quantization level 8bit. i is the historical coding mode, cosine_similarity is the cosine similarity, C _histi is the historical fault scenario vector. According to the current network bandwidth status B current =2.4Mbps, apply adaptive adjustment algorithm: bandwidth_ratio=B current / B optimal =2.4 / 4.0=0.6;compression_adjustment=1+(1-bandwidth_ratio)×sensitivity factor =1+(1-0.6)×0.5=1.2; adjusted compression_rate =initial _rate ×compression_adjustment=0.35×1.2=0.42; generate the adjusted encoding strategy E_adjusted, and the compression rate target is increased to 0.42. Where bandwidth_ratio is the bandwidth ratio, B optimal is the optimal bandwidth, compression_adjustment is the compression adjustment coefficient, sensitivity factoris the sensitivity factor, adjusted compression_rate is the compression ratio after adjustment, initial _rate is the initial compression rate. Apply structure-specific run-length encoding to the topological structure difference matrix: first, convert the sparse matrix into a one-dimensional sequence sparse_seq by row scanning; identify continuous 0-value segments and non-zero-value segments; encoding format: (run length , value); the original 64-byte matrix is ​​encoded into 18 bytes, and the structure coding compression rate is R struct =(64-18) / 64=0.719; Generate encoded structural difference data D struct _encoded. Use adaptive DPCM technology for parameter deviation vector: quantization_step=max(|D param |) / 2 bits =0.08 / 256=0.0003125; for eachparam i , calculate predicted i =param_{i-1}+trend factor ; Difference value diff i =param i -predicted i Quantitative value quant i =round(diff i / quantization_step); the original 40-byte vector is encoded into 15 bytes, and the parameter encoding compression rate R param =(40-15) / 40=0.625; Generate encoded parameter deviation data D param _encoded. Among them, quantization _step is the quantization step size, predicted i is the predicted value, trend factor is the trend factor, param i The encoded structural difference data and parameter deviation data are organized according to a predetermined format, and a checksum and header information are added to generate an initial difference data packet P0 with a total size of 43 bytes (including a 10-byte header information).

[0194] 3.4. Apply the power topology specialized analysis algorithm to identify the pattern in P0: sparsity calculation sparsity = zero_count / total_count = 45 / 64 = 0.703; local clustering calculation clustering_coeff = clustered_nonzero / total_nonzero = 14 / 19 = 0.737; periodic pattern detection period_strength = 0.23; generate data characteristic analysis report R_analysis, including sparsity, clustering and periodic characteristics. _count The number of zero elements, total _count is the total number of elements, clustered _nonzero is the number of non-zero elements in the collection, total _nonzero The total number of non-zero elements. Based on the data characteristic analysis report R_analysis, the data is partitioned according to information entropy: the high entropy area entropy>0.8, accounting for 35%; the low entropy area entropy≤0.8, accounting for 65%; the partition coding scheme P is formulated. _zone :Arithmetic coding is used in high entropy areas, and run-length coding is used in low entropy areas. Run-length coding is applied to structural differences: Based on the prior knowledge of power topology connections, the coding dictionary dict_topology is constructed; the coding format (symbol, length, context); context correlation is introduced _weight =0.3; encoding efficiency improved encoding_gain=1+context _weight ×pattern_match _rate =1+0.3×0.6=1.18; generate structure optimized coding data S opt , size 12 bytes. _rate is the pattern matching rate, symbol is the symbol, length is the length, and context is the context information. Context-adaptive arithmetic coding is used for the parameter deviation part: an adaptive probability model adaptive_model is constructed, with the initial probability distribution being uniform; for each symbol, the probability probability is updated. i =(count i +α) / (total_count+α×alphabet_size), where α=0.5 is the smoothing parameter; encoding length bits i =-log2(probability i );Generate parameter optimized encoding data P opt , size 8 bytes. iis the number of times the symbol appears, alphabet _size is the size of the symbol set, probability i = is the symbol probability. Construct a hierarchical structure: header (2byte) + priority_data (8byte) + structure_data (12byte) + parameter_data (8byte) + checksum (2byte); prioritize key topology information; add appropriate forward error correction code redundancy _rate =0.1; generate structured difference data packet P struct , size 32 bytes. The header is the head, priority _data For priority data, structure _data For structural data, parameter _data is the parameter data, checksum is the checksum. struct Apply lightweight Huffman coding: Statistical symbol frequency _table ;Build Huffman tree huffman _tree ; Generate the optimal encoding table encoding_table; finally compress to P final , size 22 bytes; total compression ratio R_total = (original 104 bytes - 22 bytes) / 104 bytes = 0.788; transmitted data is reduced by 78.8%.

[0195] 3.5. Establish a priority transmission mechanism, assigning high priority to the key topology decision information (8 bytes) within the 22-byte compressed difference data packet, ensuring priority transmission in bandwidth-constrained situations. Simultaneously, the edge device activates the preliminary topology execution module and begins real-time control of the power grid.

[0196] Step 4: The cloud parses and compresses the difference data packets, generates an optimized topology solution, and sends it.

[0197] 4.1、The cloud receives the 22-byte compressed difference data packet P final , parse according to the agreed decoding process. Perform Huffman decoding to recover the 32-byte structured data packet, and decode the structural difference and parameter deviation information respectively. Combined with the standard topology template T stored in the cloud std , completely restore the initial topology scheme T at the edge through the difference reconstruction algorithm edge .

[0198] 4.2. Reconstruct the preliminary topology scheme of the edge end T edgeA high-precision power grid simulation model was imported to simulate and verify various operating conditions. Test conditions included full-load operation, light-load operation, harmonic interference, and voltage sag. Key performance indicators evaluated included voltage quality (THD_V) of 2.8%, power transmission efficiency (η) of 94.5%, dynamic response time (t_response) of 42ms, and system stability margin (SM) of 15.2°. A performance evaluation report (R_performance) was generated.

[0199] 4.3. Extract key indicators from the performance evaluation report and construct a multi-objective optimization function: f1 = minimize(THD_V) = minimize(2.8%); f2 = maximize(η) = maximize(94.5%); f3 = minimize(t_response) = minimize(42ms); f4 = maximize(SM) = maximize(15.2°); comprehensive objective function F = w1×f1+w2×f2+w3×f3+w4×f4, weight vector W = [0.3, 0.25, 0.25, 0.2]. Construct a topology optimization constraint set: electrical connection constraint ∑I i n = ∑ I_out (Kirchhoff's current law); power flow constraint P_gen - P_load - P_loss = 0; equipment capacity constraint I_equipment ≤ I _rated ; Voltage range constraint 0.95pu≤V≤1.05pu; Stability constraint eigenvalue(A_system)<0. Where I i n is the input current, I_out is the output current, P_gen is the generated power, P_load is the load power, P_loss is the power loss, I_equipment is the equipment current, I _rated is the rated current, eigenvalue is the eigenvalue, and A_system is the system matrix. edge Parameter sensitivity: for each parameterp i , calculate the sensitivity coefficient sensitivity i =ΨF / Ψp i |_{p i =p i 0}; estimation: sensitivity by finite difference method i ≈(F(p i +Δp)-F(p i-Δp)) / (2Δp); identify the 8 most sensitive parameters and 5 key topological nodes; generate a sensitivity analysis report S_sensitivity. Where Ψ is the partial derivative, parameter is the parameter, and p i is the value of the i-th parameter, and Δp is the small change of the parameter. Apply the PSO algorithm to the key control parameters: initialize the particle swarm population_size=50; each particle represents a set of parameter configuration particle i =[p1, p2, ..., p8]; dynamic inertia weight w(t) = w max -(w max -w_min)×t / t max =0.9-0.4×t / 100; learning factor c1(t)=2.5-2×t / t max , c2(t)=0.5+2×t / t max ; Particle update equation v i (t+1)=w(t)×v i (t)+c1(t)×r1×(p_best i -x i (t))+c2(t)×r2×(g_best-x i (t)); position update x i (t+1)=x i (t)+v i (t+1); where w max is the maximum weight value, w_min is the minimum weight value, t max is the maximum number of iterations, r1 and r2 are different random factors, p_best i is the individual optimal position, and g_best is the global optimal position.

[0200] After 80 iterations, the optimization result is: global_best_fitness=0.913; the optimal control parameter set P_control opt=[0.87, 0.95, 0.78, 0.92, 0.89, 0.83, 0.96, 0.81]; fitness improvement = (initial_fitness-final_fitness) / initial_fitness = (0.672-0.913) / 0.672 = -0.359 (negative value indicates 35.9% performance improvement). Where final_fitness is the final fitness and initial_fitness is the initial fitness. Apply graph theory-based local structure optimization to key topological nodes: for each critical_node, analyze its connectivity_degree and betweenness_centrality; calculate the impact of the structural adjustment. score =α×connectivity_change+β×centrality_change, where α=0.6, β=0.4; fine-tune the connections of the three key nodes while maintaining the stability of the main structure; generate the optimized topology structure T structopt , 5 elements in the connection matrix change. Among them, critical_node is the key node, connectivity_change is the change in connectivity, and centrality_change is the change in centrality. Combined with the optimal control parameter set P_control opt and optimized topology T structopt Perform system-level integration: consistency check consistency_check = verify_electrical_laws(T structopt , P_control opt ); performance verification performance_verify = simulate_system(T structopt , P_control opt ); optimized performance indicators: THD_V = 1.8% (improved by 35.7%), η = 96.7% (increased by 2.3%), t_response = 28ms (shortened by 33.3%), SM = 18.9° (increased by 24.3%); generate a complete optimized topology solution T opt .

[0201] 4.4 Comparison of optimized topology solutions T opt and the reconstructed edge-end preliminary topology scheme T edge , extract key differences: topology adjustment: 5 connection points changed; parameter adjustment: incremental update value Δp=P of 8 control parameters opt -P edge=[0.05, 0.02, -0.03, 0.04, 0.01, 0.06, -0.02, 0.03]; where P edge It is the preliminary parameter set for the edge end; priority information: mark the top 3 items with the greatest impact as high priority; generate incremental update instructions I delta , using differential encoding format, the data packet size is only 8 bytes, including all necessary adjustment information.

[0202] 4.5. Incremental update instruction I delta The command is sent down to the edge via a high-priority channel. After receiving the command, the edge performs the final topology reconstruction according to the following process: verifying the integrity and legitimacy of the command; applying topology adjustments step by step in order of priority; monitoring changes in key performance indicators in real time; and uploading confirmation information after execution. The edge execution process takes 3.2ms, and the entire reconstruction process takes 35.8ms (19.3ms for initial reconstruction at the edge, 13.7ms for cloud optimization, and 3.2ms for execution), meeting real-time requirements.

[0203] The voltage harmonics THD_V of this embodiment was 3.2% before the fault, 2.8% during the initial reconstruction at the edge, and 1.8% after cloud optimization, with an improvement of 43.8%. The transmission efficiency η of this embodiment was 94.2% before the fault, 94.5% during the initial reconstruction at the edge, and 96.7% after cloud optimization, with an improvement of 2.7%. The response time t_response of this embodiment was 42ms during the initial reconstruction at the edge, and 28ms after cloud optimization, with an improvement of 33.3%. The stability margin SM of this embodiment was 12.8° before the fault, 15.2° during the initial reconstruction at the edge, and 18.9° after cloud optimization, with an improvement of 47.7%.

[0204] This embodiment achieves efficient 4D representation through the fusion of lightweight context encoders, preserving key information and reducing encoding time. Adaptive radial basis function interpolation improves interpolation accuracy and adaptability compared to traditional fixed kernel width methods. A layered optimization strategy reduces computational complexity and ensures global optimality. An adaptive coding strategy improves compression ratio and adapts to varying network conditions. Global cloud-based optimization improves comprehensive performance indicators, with significant multi-objective coordinated optimization results. A rapid edge response mechanism ensures initial reconstruction within 20ms after fault isolation, meeting the real-time requirements of the power system and reducing network load. A two-layer optimization strategy ensures that even in the event of a network outage, the edge can independently complete basic topology reconstruction. This embodiment transforms the topology selection problem into a continuous space interpolation problem, improving adaptability to unseen fault scenarios. Rather than simple task allocation, this framework utilizes a deep collaborative framework with bidirectional knowledge flow and model co-evolution, enabling deep edge-cloud collaboration. Sparsity coding and topology constraints tailored to the characteristics of the power system enable domain-specific optimization, outperforming general-purpose methods. A layered progressive strategy achieves the organic integration of real-time response and global optimization. This embodiment fully demonstrates the technical feasibility and advancement of the invention, providing an efficient, reliable, and practical solution for fault self-healing in modular high-voltage power quality equipment. While ensuring system real-time performance, it improves the quality and efficiency of topology reconstruction, demonstrating significant engineering application value.

[0205] The present invention completely breaks through the limitations of the traditional discrete table lookup mode by constructing a continuously differentiable topological feature space. The specific implementation is to encode the discrete preset topology library into anchor points in an 8-dimensional continuous feature space, and then use a multi-core radial basis function interpolation algorithm to calculate the similarity weight with each anchor point based on the coordinate position of the 4-dimensional comprehensive context representation vector of the current fault scenario in the feature space and perform continuous interpolation generation. This enables the system to generate an infinite number of intermediate state configurations between the preset topologies, completely breaking away from the limitations of discrete selection, achieving a topology generation capability that truly adapts to the current specific fault scenario, and improving adaptability to unseen fault combinations. A refined three-layer hierarchical topology progressive optimization architecture is constructed, and the topology reconstruction problem is scientifically decomposed according to time sensitivity and computational complexity. The edge side undertakes the real-time critical tasks of generating the main topology architecture and optimizing the sub-module topology, making full use of edge computing resources; the cloud side undertakes the parameter fine-tuning and structure fine-tuning tasks based on high-precision simulation, giving full play to its powerful computing power. Through context-related coding and sparsity-based entropy coding technology, efficient information transmission between the edge and the cloud is ensured. More importantly, cloud-side optimization does not involve recalculation, but rather incremental optimization based on the preliminary plan on the edge. Through difference extraction and incremental update mechanisms, the cloud can accurately perceive the real-time constraints on the edge. The optimization results are fully compatible with the edge execution environment, achieving true collaborative optimization.

[0206] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A self-healing topology reconstruction method after isolation of a faulty unit of a modular high-voltage power quality device, characterized in that: include: Collect fault unit identification information and real-time grid operation data and fuse them into a comprehensive context representation vector; Map it to the pre-dimensionality-reduced topological feature space, and perform interpolation calculations between the anchor point topologies to generate a preliminary topology solution for the edge end; Compare the preliminary topology solution at the edge with the pre-stored standard topology template, extract the structural differences and parameter deviations, generate a compressed difference data packet, and send it to the cloud; The cloud parses the compressed difference data packets and reconstructs the complete topology information based on the standard topology template to generate an optimized topology solution. Based on this solution, the adjustments are sent to the edge through incremental update instructions to perform the final topology reconstruction. Generate a preliminary edge topology plan, including: Input the comprehensive context representation vector into the pre-trained topological feature space mapping model to locate the coordinate position of the current fault scenario in the topological feature space; Identify the anchor topology closest to the coordinate position and calculate the similarity weight between the coordinate position and each anchor topology; Interpolate between anchor topologies based on similarity weights to generate the main topology architecture, and then use a progressive optimization strategy to determine the sub-module connection relationship and generate the sub-module topology; Integrate the main topology architecture and sub-module topology, configure control parameters, and form a preliminary edge topology solution; Generates a compressed difference packet, including: Obtain the standard topology template corresponding to the current fault type and the preliminary topology plan for the edge end. Compare structural differences and generate a topological structure difference matrix; Compare control parameters, calculate parameter deviation values, and generate parameter deviation vectors; Adjust the encoding strategy based on the characteristics of the current fault scenario, compress the topology structure difference matrix and parameter deviation vector into an initial difference data packet, and then perform secondary compression on it by combining the sparsity characteristics of the power system topology connection through entropy coding to generate a compressed difference data packet; Perform final topology reconstruction, including: The cloud receives compressed differential data packets, decodes them based on the standard topology template, reconstructs the preliminary topology plan at the edge, imports it into the power grid model, performs simulation verification, evaluates performance under various load conditions, and generates a performance evaluation report. Based on the performance evaluation report, adjust the topology structure and control parameters to generate an optimized topology solution that takes long-term stability into consideration; Compare the optimized topology with the reconstructed preliminary edge topology, extract key differences, and generate incremental update instructions; The incremental update instructions are sent to the edge, which receives and applies the adjustments and performs the final topology reconstruction.

2. The method according to claim 1, characterized in that Fusion into a comprehensive context representation vector, including: Collect fault unit identification information, generate an initial fault feature set, apply feature selection algorithms to eliminate redundant features, extract core fault indicators and generate a streamlined fault feature set; Collect real-time grid operation data, generate an initial grid state feature set, perform time-frequency analysis, extract load characteristics and grid fluctuation patterns, and generate a simplified grid state feature set; The simplified fault and simplified grid status feature sets are fused and encoded into a comprehensive context representation vector of predetermined dimension.

3. The method according to claim 2, characterized in that The fusion encoding is a comprehensive context representation vector of a predetermined dimension, including: Performing linear normalization processing on the simplified fault and simplified grid state feature sets to generate standardized fault features and standardized grid state features; The standardized fault features are multiplied by a preset feature importance weight matrix to generate weighted fault features, and the standardized grid state features are transformed using a nonlinear transformation function constructed based on a typical characteristic curve of the power system to generate transformed grid features; A cross-correlation analysis is performed on the weighted fault features and the transformed power grid features to generate a cross-correlation feature matrix, which is then subjected to dimensionality reduction and compression processing to generate a comprehensive context representation vector.

4. The method according to claim 1, wherein Generate the main topology architecture, including: Analyze the distribution density of anchor points around the coordinate position, generate a regional density index, and adjust the kernel width through the multi-kernel radial basis function accordingly. Calculate the kernel function value between the coordinate position and the topology of each anchor point, and generate a kernel function value vector. Combine the kernel function value vector and the similarity weight to calculate the normalized hybrid weight vector; The anchor topologies are weightedly fused according to the normalized hybrid weight vector and combined with the structural consistency constraints to generate the main topology architecture.

5. The method according to claim 1, wherein Generate submodule topology, including: Divide the main topology into a set of interconnected functional sub-modules and generate a sub-module partitioning diagram; Based on the current fault characteristics, an optimization priority is assigned to each submodule in the submodule partitioning diagram, and a submodule priority queue is generated. The highest priority submodule is extracted from the queue and the topology optimization algorithm is applied sequentially to generate an optimized submodule set. Optimize the boundary connections between adjacent sub-modules in the sub-module set, eliminate connection conflicts, and generate a sub-module boundary connection plan; The optimized sub-module set is combined according to the sub-module boundary connection scheme to generate a sub-module topology.

6. The method according to claim 1, characterized in that Compress the topology difference matrix and parameter deviation vector into an initial difference data packet, including: Combined with the comprehensive context representation vector, the optimal coding strategy for similar fault scenarios is retrieved from the historical coding pattern library to generate an initial coding strategy. This strategy is then dynamically adjusted based on the current network bandwidth to generate an adjusted coding strategy, including structural coding and parameter coding. Applying structural coding, using a run-length coding algorithm on the topological structure difference matrix to generate coded structural difference data; applying parameter coding, using a differential pulse code modulation method on the parameter deviation vector to generate coded parameter deviation data; The encoded structural difference data and the encoded parameter deviation data are organized according to a predetermined format to generate an initial difference data packet, including the structural difference and the parameter deviation.

7. The method according to claim 1, characterized in that Generates a compressed difference packet, including: Identify the structural patterns and sparse distribution characteristics in the initial difference data packets and generate a data characteristics analysis report; based on this, divide the difference data into high entropy areas and low entropy areas according to the size of the information entropy, apply encoding strategies to each area, and generate a partition encoding scheme; For structural differences, the run-length coding for sparse matrix regions in the partition coding scheme is adopted to generate structurally optimized coding data; For parameter deviation, the context-adaptive arithmetic coding technology in the partition coding scheme is used to generate parameter-optimized coding data; The structure-optimized coded data and the parameter-optimized coded data are combined into a structured difference data packet and compressed to generate a compressed difference data packet.

Images