A Distribution Network Topology Identification Method Based on Structural Similarity Index

By constructing a topology identification method based on structural similarity index, the accuracy and robustness issues of dynamic topology identification in distribution networks are solved, achieving efficient and reliable topology tracking in complex environments and meeting the needs of real-time monitoring and scheduling.

CN123068973APending Publication Date: 2026-11-10SICHUAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611026479.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-10
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies cannot meet the requirements of real-time monitoring and operation scheduling when facing the dynamic time-varying characteristics caused by distributed power grid connection and distribution network switching and reconfiguration. Furthermore, data-driven and mathematical statistical inference methods lack the accuracy and robustness of topology identification in complex and sparse monitoring environments.

Method used

By constructing a candidate topology set that includes the physical constraints of the distribution network, the structural similarity index (SSI) is used to calculate the voltage and current distribution of each candidate topology in the topology library. Real-time electrical parameters are obtained by combining high-resolution measurement equipment, a spatiotemporal observation matrix is ​​constructed, missing data is filled in and normalized, the structural similarity of the topology is calculated, and the topology with the highest SSI score is selected as the actual operating topology.

Benefits of technology

It achieves high-precision and robust topology identification in extremely harsh environments, reduces dependence on high-end computing hardware, improves the flexibility and reliability of the system in dynamic tracking of complex power distribution network structures, and meets the computational efficiency requirements for online monitoring and scheduling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068973A_ABST
    Figure CN123068973A_ABST
Patent Text Reader

Abstract

This invention relates to the field of power system distribution network parameter estimation, and discloses a distribution network topology identification method based on structural similarity index. The method includes: acquiring real-time load data and inputting it into a candidate topology set to perform power flow simulation, generating a baseline state vector for each candidate topology; acquiring real-time electrical parameters and constructing an original observation matrix; copying and tiling the baseline state vector along the time axis to construct a simulation reference matrix; performing normalization preprocessing on the original observation matrix and the simulation reference matrix; and using the structural similarity index to calculate the structural similarity between the simulation reference matrix and the preprocessed original observation matrix for each candidate topology, selecting the candidate topology with the highest structural similarity index score as the actual operating topology of the current distribution network. The advantages of this invention are that it eliminates the need for complex model training processes and effectively solves the problem of difficulty in effectively identifying incomplete measurement data under extreme noise and high-proportion missing conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power system distribution network parameter estimation, and in particular to a distribution network topology identification method based on structural similarity index. Background Technology

[0002] To cope with the large-scale grid connection of distributed generation sources and the increasingly frequent switching and reconfiguration operations in distribution networks, the topology of active distribution networks exhibits highly dynamic and time-varying characteristics. Traditional static topology analysis capabilities can no longer meet the needs of real-time monitoring and operation scheduling of modern distribution networks. In order to accurately track these complex network dynamic changes, dynamic topology identification technology for distribution networks has become a core support for maintaining the safe and efficient operation of the power grid.

[0003] However, the actual operating environment of modern power distribution networks is extremely complex. Low-level data acquisition and monitoring systems commonly face challenges such as incomplete measurement data (e.g., communication packet loss), high-frequency measurement noise interference, and strong uncertainties caused by drastic power fluctuations on both the source and load sides. These harsh operating conditions place a heavy burden on traditional optimization algorithms based on physical solvers or data-driven neural network models, leading to complex model building costs and high offline pre-training overhead. When faced with severe data gaps or extreme measurement noise, the topology identification accuracy of existing algorithms often experiences a precipitous drop.

[0004] To address the technical challenge of achieving high-precision and robust topology identification in active distribution networks under complex and sparse monitoring environments, existing technologies mainly employ two mainstream data processing schemes, both of which suffer from significant technical drawbacks:

[0005] Existing technical solution 1: Artificial intelligence learning technology driven by massive data. Such solutions (such as deep learning architectures, graph neural networks, etc.) naturally have the ability to encode complex nonlinear mapping relationships, and are expected to achieve real-time inference of network topology without relying on explicit physical prior information.

[0006] However, the core mechanism of this scheme inevitably relies heavily on massive amounts of historical measurement data for extremely time-consuming offline pre-training. Even with the introduction of generative models such as generative adversarial networks for data augmentation to compensate for scarce data, it is difficult to achieve high-precision topology identification accuracy in large-scale distribution network structures. On the contrary, the increasing model complexity leads to an unbearable training burden. More critically, due to physical limitations, measurement data in actual power grids are inevitably accompanied by varying degrees of noise pollution and communication packet loss, which directly causes a sharp deterioration in the performance of AI learning models that are highly dependent on data quality. Furthermore, due to the inherent "black box" nature of neural network models, they lack provable robustness and physical interpretability under extreme operating conditions, making it difficult for this scheme to truly meet engineering reliability requirements in distribution networks with sparse monitoring equipment and complex operating environments.

[0007] Existing technical solution two: Mathematical statistical inference and analytical graph model technology. In order to avoid the excessive reliance of data-driven models on massive amounts of data and heavy offline pre-training, some existing technologies have turned to using pure graph models or statistical analytical models (such as Markov random fields, multinomial time structure learning, etc.) for topological estimation without physical priors.

[0008] However, these solutions also face insurmountable application bottlenecks. First, solutions based on pure graph models require extremely complex graph node association construction. When the algorithm needs to be applied across different distribution network physical structures, it faces extremely high algorithm migration and reconstruction costs. Second, models based on analytical inference (such as Markov random fields combined with iterative screening) often exhibit extremely poor robustness when faced with unpredictable measurement noise at the distribution network level. Although some solutions attempt to introduce multinomial time structure learning algorithms to improve the anti-interference capability of analytical methods, when continuous data packet loss or measurement interruptions occur in the underlying communication network, these algorithms, which strictly rely on the consistency of statistical data, cannot guarantee the absolute reliability of the identification conclusions. Existing technologies still face intractable problems such as the "high training cost, non-explainable black box, and extreme sensitivity to dirty data" of data-driven algorithms and the "high network migration cost and poor resistance to continuous packet loss and extreme noise" of statistical analytical graph models. They are unable to meet the stringent requirements of modern distribution networks for accuracy, real-time performance, and high robustness in extreme measurement environments. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a distribution network topology identification method based on structural similarity index.

[0010] The objective of this invention is achieved through the following technical solution: a distribution network topology identification method based on structural similarity index, the method comprising the following steps:

[0011] S1. Obtain real-time load data and input it into the candidate topology set to perform power flow simulation. Calculate the full network voltage and current distribution of each topology in the topology library. Based on the physical distribution of sensors, extract the measurement data of the corresponding n monitoring locations from the simulation results and generate the reference state vector of each candidate topology.

[0012] S2. Using high-resolution on-site measurement equipment, real-time electrical parameters of w time frames are acquired at a uniform frequency within a continuous observation time window. The measurement data of n monitoring locations are arranged according to the spatiotemporal dimension to construct a w×n-dimensional original observation matrix.

[0013] S3. Copy and tile the reference state vector w times along the time axis to construct a w×n dimensional simulation reference matrix that matches the dimension of the original observation matrix. Complete the missing data in the original observation matrix. Perform normalization preprocessing on the original observation matrix and the simulation reference matrix to map the heterogeneous measurement data to a unified numerical range.

[0014] S4. Calculate the structural similarity between the simulated reference matrix and the preprocessed original observation matrix of each candidate topology using the structural similarity index. The structural similarity index is quantitatively evaluated from three dimensions: mean deviation, fluctuation variance, and spatial cocorrelation. Select the candidate topology with the highest structural similarity index score as the actual operating topology of the current distribution network.

[0015] Specifically, the step of extracting measurement data corresponding to n monitoring locations from the simulation results based on the physical distribution of sensors and generating a baseline state vector for each candidate topology includes:

[0016] By utilizing the mapping relationship between the distribution network topology map and the equipment ledger in the topology library, the node voltage vector or branch current vector in the distribution network power flow calculation result corresponding to the real-time load distribution data is spatially filtered according to the physical node index and branch index of the actual deployed high-resolution measurement equipment, and combined to form a 1×n-dimensional static electrical feature vector, which serves as the initial physical reference benchmark for the candidate topology.

[0017] Specifically, in S3, completing the missing data in the original observation matrix includes:

[0018] When it is detected that the measurement data of a certain monitoring position in the w×n-dimensional original observation matrix is ​​lost or invalid in a specific time frame, the measurement values ​​of the effective time frames before and after the monitoring position on the time axis are extracted, and the arithmetic mean is calculated for interpolation and filling.

[0019] If multiple consecutive frames of data are missing on the timeline, the system first attempts to interpolate and fill the missing data by combining the measurement values ​​of the adjacent valid time frames before and after the monitoring location on the timeline. If the filling fails, the column data corresponding to the measurement point is removed, and the corresponding column data in the simulation reference matrix is ​​also removed.

[0020] Specifically, the normalization preprocessing of the original observation matrix and the simulation reference matrix includes:

[0021] A per-unit scaling algorithm is used to independently scale the voltage observation column and the current observation column in the w×n-dimensional original observation matrix and the simulation reference matrix, respectively, to eliminate the differences in the inherent magnitudes of different physical quantities.

[0022] Specifically, the calculation of the structural similarity between the simulated reference matrix and the preprocessed original observation matrix of each candidate topology using the structural similarity index includes:

[0023] The mean comparison function, variance comparison function, and covariance comparison function between the original observation matrix and the simulated reference matrix are calculated respectively to quantify the features of the three dimensions of mean deviation, fluctuation variance, and spatial cocorrelation.

[0024] By adaptively adjusting the weighting factors of the electrical characteristics of the distribution network, the mean comparison function, variance comparison function, and covariance comparison function are multiplied and fused to obtain the comprehensive score of the structural similarity index.

[0025] Specifically, after selecting the candidate topology with the highest structural similarity index score as the actual operating topology of the current distribution network, a security verification mechanism based on discrimination margin is also included, specifically including:

[0026] Extract the optimal candidate topology that obtains the maximum structural similarity index score and the second-best candidate topology that obtains the second-largest structural similarity index score, and calculate the score difference between the two as the discrimination margin.

[0027] When the discrimination margin is less than the preset safety margin threshold, the output of a single, definitive automated identification conclusion is paused, and instead a set of suspected topologies with low confidence alarms is output to prompt manual intervention for verification.

[0028] Specifically, the construction method of the candidate topology set includes:

[0029] Obtain the physical infrastructure ledger and connectivity parameters of the distribution network at the bottom layer, and abstract the actual distribution network into a mathematical graph theory model G=(V, E), where the node set V represents the bus, transformer and load connection point, the edge set E represents the physical feeder, and the edge set is further subdivided into uncontrollable static line segments and controllable switching equipment.

[0030] Driven by the graph theory spanning tree algorithm and with the radial operation constraints of the distribution network as a prerequisite, all combinations of switch states that satisfy the characteristics of connectivity and no loops are extracted in the graph model space to generate an initial radial topology library.

[0031] The initial radial topology library is pruned by applying multiple operational boundary constraints, including system global connectivity and island avoidance constraints, power supply capacity range constraints of upstream power sources, node voltage over-limit constraints, and branch thermal capacity constraints, ultimately generating a candidate topology set.

[0032] The present invention has the following advantages:

[0033] 1. This invention completely abandons the reliance of traditional data-driven models on massive historical datasets and cumbersome offline pre-training processes in its system architecture. By constructing a candidate topology set containing distribution network physical constraints, it innovatively achieves a training-free physical state matching mechanism. The system obtains the baseline electrical parameters of each feasible topology through a pre-set power flow simulator, completely avoiding the high costs of data collection, cleaning, labeling, and model iteration updates associated with deep learning and graph neural network algorithms. This not only effectively overcomes the "cold start" problem faced by black-box machine learning models in newly built power grids lacking historical measurement data, but also enables the algorithm to be used immediately without pre-training, greatly improving the system's flexibility and implementation efficiency when facing expansion and migration of different network structures.

[0034] 2. This invention introduces the structural similarity index from the field of image evaluation into the dynamic topology analysis of power distribution networks, and makes adaptive adjustments: 1. By replicating and tiling the reference vector along the time axis (solving the mismatch in dimension and properties between static references and dynamic observations); 2. By standardizing heterogeneous measurements column by column (eliminating the asymmetric masking of brightness and contrast terms by differences in voltage / current dimensions and magnitudes); 3. By maintaining consistent spatiotemporal interpolation (correcting matrix holes caused by communication packet loss), a comparable evaluation space with matching dimensions, dimensionlessness, and complete structure is constructed from the bottom layer. On this basis, a physical correspondence of "topology physical change - matrix mean / variance / covariance drift" is established and adaptively weighted according to the electrical characteristics of the power distribution network, transforming the original similarity operator oriented towards visual perception into a topology identification metric with clear physical discriminative significance. By quantifying the structural similarity between the simulated reference matrix and the real-time observation matrix, accurate matching of the actual operating physical topology is achieved. Unlike traditional sensitivity evaluation systems based on point-to-point absolute numerical errors (such as mean square error, MSE), the SSI operator of this invention can transform abstract physical changes in power grid topology into mathematically measurable matrix structure drift from multiple dimensions such as the spatial distribution, fluctuation trend, and correlation of electrical characteristics. This topology identification method based on structural similarity comparison enables the system to directly receive the topology association context containing the entire electrical picture during deduction, providing an extremely intuitive and accurate evaluation criterion for the dynamic tracking of complex distribution networks.

[0035] 3. This invention fully utilizes the global structural correlation of electrical parameters in the spatiotemporal dimension of the distribution network to suppress abnormal interference from local data, exhibiting extremely high robustness in the face of extremely harsh engineering measurement conditions. When large-scale packet loss (up to 50% data loss) occurs in the underlying communication of a real distribution network or when measurement equipment is subjected to extreme continuous high-frequency noise pollution, because the matching mechanism of this method focuses on the overall topological characteristics of the data matrix rather than local independent values, these local, isolated "dirty data" cannot easily destroy the global physical structural correlation. Numerical verification shows that the system can still maintain an identification success rate of over 88% even in the face of extreme data corruption scenarios, completely solving the fatal flaw of traditional learning algorithms that highly depend on data quality and suffer a precipitous drop in identification accuracy under harsh environments, greatly improving the security and reliability of the system in real complex communication environments;

[0036] 4. This invention provides training-free operation and high disturbance rejection capabilities while achieving excellent computational efficiency to meet the requirements of online monitoring and scheduling of distribution networks. Although some end-to-end deep learning baseline models show certain speed advantages in the online inference stage, the pre-training-free matching architecture of this invention completely eliminates the computational consumption and time barriers of massive parameter optimization. When performing real-time topology tracing, the system replaces the complex iterative solution process of nonlinear physical equations or the deep forward propagation process of deep neural networks with simplified and highly vectorized SSI matrix operations. This lightweight computational mechanism not only reduces the system's dependence on high-end computing hardware (such as high-end GPUs), but its extremely short identification response time perfectly matches the stringent time constraints of real-time scheduling of active distribution networks, demonstrating extremely high engineering practical value in dealing with the state estimation of modern complex power grids. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the power distribution network topology identification method of the present invention;

[0038] Figure 2 Generate diagrams for the IEEE 33-node distribution network topology library;

[0039] Figure 3 This is a schematic diagram of an SSI-based topology identification method.

[0040] Figure 4 A graph illustrating the accuracy of IEEE 33-node distribution network topology identification.

[0041] Figure 5 A schematic diagram illustrating the safety mechanism and alarm process for identifying margin and preventing false alarms. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0043] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0044] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0045] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0046] like Figures 1 to 5 As shown, a distribution network topology identification method based on structural similarity index is proposed, which includes the following steps:

[0047] S1. Acquire real-time load data and input it into the candidate topology set to perform power flow simulation. Calculate the overall network voltage and current distribution for each topology in the topology library. Based on the physical distribution of sensors, extract the measurement data of the corresponding n monitoring locations from the simulation results to generate the baseline state vector for each candidate topology. This embodiment aims to establish a physical baseline covering all possible operations of the distribution network without relying on historical data training. Specifically, it includes:

[0048] S11. Physical Network Schematic Analysis and Initial Radial Topology Library Construction: Obtain the physical infrastructure ledger and connectivity parameters of the distribution network's underlying structure, and directly acquire all possible topologies of the distribution network to form a topology library, such as... Figure 2As shown, the IEEE 33-node network can generate various topologies under the opening and closing coordination of tie switches. All topology sets form a topology library, which abstracts the actual distribution network into a mathematical graph theory model G=(V, E), where the node set V represents the bus, transformer and load connection point, and the edge set E represents the physical feeder. The edge set is further subdivided into uncontrollable static line segments and controllable switching equipment.

[0049] To avoid the computational explosion caused by fully enumerating all switch states, the system is driven by a built-in graph theory spanning tree algorithm, based on the radial operation constraint of the distribution network (the total number of closed electrical branches n). l Strictly equal to the total number of charged nodes n b minus one, n l =n b -1) As a prerequisite for network search, extract all combinations of switch states that satisfy the characteristics of connectivity and no loops in the graph model space to generate an initial radial topology library.

[0050] S12. Operational Boundary Constraint Pruning and Candidate Topology Set Generation: Although the configurations in the initial radial topology library all satisfy the characteristics of a loop-free network structure, deep filtering is still required in conjunction with the safety boundaries of the actual operation of the distribution network. Multiple operational boundary constraints are applied to the initial radial topology library for pruning. These multiple operational boundary constraints include system global connectivity and islanding avoidance constraints, power supply capacity range constraints of upstream power sources, node voltage over-limit boundary constraints, and branch thermal capacity limits. Through the above hard physical and electrical rules, dangerous configurations that may cause line overload or voltage over-limit under actual operating conditions are completely eliminated, and finally a candidate set of N absolutely safe and engineering feasible network topologies is extracted.

[0051] S13. Real-time load mapping and baseline power flow calculation: Receive the latest low-frequency load measurement data of the distribution network and input it as boundary conditions into each feasible topology in the candidate topology set generated by the above pruning; call the underlying power system power flow calculation engine to perform deterministic baseline power flow calculation one by one to obtain the theoretical electrical distribution of the entire network for each topology under the current load level.

[0052] S14. Spatial Measurement Point Orientation Extraction and Reference Vector Generation: Based on the full-network power flow simulation results, and strictly comparing with the geographical topology node indices of n high-resolution physical monitoring devices (such as PMUs or AMIs) deployed in the actual distribution network, voltage and current phasor data at corresponding locations are extracted. Through this spatial dimensionality reduction mapping, a static initial state vector of dimension 1×n is generated for each candidate topology, serving as a physical reference benchmark reflecting the current operating characteristics of the topology.

[0053] S2. Using high-resolution on-site measurement equipment (including but not limited to PMU or current transformer AMI), real-time electrical parameters (voltage, current) are acquired at a uniform frequency for w time frames within a continuous observation time window. The measurement data from n monitoring locations are arranged according to the spatiotemporal dimension to construct a w×n dimensional original observation matrix. The rows of the matrix represent the spatial distribution of electrical parameters at different monitoring points at a specific moment, and the columns represent the dynamic evolution trajectory of each monitoring point within the time window. This matrix construction method can completely capture the spatiotemporal coupling characteristics of the distribution network under load fluctuations or disturbances, providing standardized data input for subsequent topology identification using structured feature analysis theory.

[0054] In this embodiment, this step aims to utilize high-frequency synchronous measurement technology to extract a dynamic operational profile reflecting load fluctuations and potential topological changes from the real physical power grid, and transform it into a standardized two-dimensional matrix structure suitable for subsequent image-based feature analysis. Specifically, it includes:

[0055] S21. High-frequency real-time electrical parameter synchronous acquisition: Through miniature synchronous phasor measurement units (PMUs) or advanced measurement systems (AMIs) smart meters deployed at key nodes of the distribution network, real-time electrical operation profile data is synchronously acquired at a uniform high-frequency sampling rate within a pre-set continuous observation time window. The acquired data strictly corresponds to the n monitored physical measurement points extracted in step S1, specifically including high-precision timestamps, node three-phase voltage amplitudes and phase angles, and current phasors of key branches. This extremely high time resolution synchronous sampling mechanism ensures absolute alignment of measurement data scattered across different geographical locations on the time scale, thereby accurately recording the distribution network's true response to external disturbances.

[0056] S22. Spatiotemporal Dynamic Data Regularization and Matrix Construction: The discrete electrical measurement data collected synchronously over w consecutive time frames within the observation time window are rigorously structured and integrated according to two dimensions: spatial distribution and temporal evolution. Specifically, a w×n original observation matrix x is constructed. In this matrix, the row dimension (w rows in total) represents a snapshot of the electrical state of all network measurement points at each specific observation instant in the continuous time series; the column dimension (n columns in total) represents the dynamic evolution trajectory of each specific physical measurement point within the entire time window. This two-dimensional matrix construction method successfully transforms the one-dimensional implicit time series signal into explicit two-dimensional digital matrix features, thereby fully capturing the spatiotemporal coupling characteristics of the distribution network's underlying layer under disturbance, and providing a standardized data format input for subsequent seamless integration with structural similarity evaluation algorithms.

[0057] S3. Copy and tile the reference state vector w times along the time axis to construct a w×n dimensional simulation reference matrix that matches the dimension of the original observation matrix. Complete the missing data in the original observation matrix. Perform normalization preprocessing on the original observation matrix and the simulation reference matrix to map the heterogeneous measurement data to a unified numerical range.

[0058] In this embodiment, the initial reference vector generated by S1 is essentially a snapshot of the steady-state electrical parameters of a single section under ideal operating conditions (dimension 1×n), while the real-time observation matrix obtained by S2 contains the dynamic evolution process at multiple moments under real physical and communication environments, as well as unavoidable data packet loss (dimension w×n), and is also mixed with multi-source heterogeneous measurement signals such as voltage and current with different physical dimensions and huge numerical differences. These two types of data have inherent logical gaps and strong incomparability in terms of time dimension, structural continuity, and mathematical scale. In addition, if the image structure similarity index (SSI) is directly applied without processing, it will be impossible to calculate or will produce distorted conclusions due to dimension mismatch, inconsistent dimensions, and matrix gaps; therefore, this step must perform strict multi-dimensional data alignment and low-level feature cleaning. Therefore, to ensure that the subsequent structural similarity index can accurately extract the purely spatial distribution feature displacements caused by topological physical evolution, this step must perform rigorous multidimensional data alignment and low-level feature cleaning. This aims to construct a dimensionally matched, structurally complete, and dimensionless standardized two-dimensional evaluation space for the measured flow data and physical simulation benchmarks. Specifically, this includes:

[0059] S31. Time-Axis Tiling and Dimensional Expansion of the Dynamic Reference: Addressing the obstacle that image SSI requires two compared objects to have the same dimension, and that the dimensions and properties of the physical reference (1×n static) and real-time observation (w×n dynamic) do not match, this step involves tiling the 1×n-dimensional static initial state vector w times along the time axis. Specifically, the 1×n-dimensional static initial state vector generated in S14 is tiled w times along the time axis to construct a w×n-dimensional simulation reference matrix that perfectly matches the dimension of the real-time observation matrix, achieving absolute alignment between the simulation data and the measured data in the time dimension.

[0060] S32. Spatiotemporal Joint Interpolation Completion of Missing Data: Considering that the image is a dense, complete grid with no missing pixels, while real-time measurements inevitably have data gaps due to communication packet loss, which will disrupt the matrix integrity upon which the structural similarity index window statistics depend; to repair this gap while maintaining strict consistency in the dimensions of the two matrices (to maintain comparability), this step first uses the mean of adjacent frames for linear interpolation. If multiple consecutive frames are missing, causing interpolation failure, the corresponding columns in both the observation matrix and the simulation reference matrix are simultaneously removed. Therefore, once a data gap due to communication packet loss is detected in the real-time observation matrix, the effective measurement mean of the adjacent time frames before and after the missing node is immediately extracted for linear interpolation completion, thereby repairing the data breakpoint and ensuring the integrity of the observation matrix during subsequent structural feature extraction. For the missing value x of the j-th measurement point in matrix x at time t... t,j The interpolation calculation formula is:

[0061] ;

[0062] Where, x t-1,j and x t+1,j These are the measurement values ​​of the nearest valid time frames before and after the missing point for that measurement point. If multiple consecutive frames of data are missing on the time axis, the system first attempts to interpolate and fill the missing data by combining the measurement values ​​of the adjacent valid time frames before and after the monitoring location on the time axis. If the interpolation fails, the column data corresponding to that measurement point is removed, and the corresponding column data in the simulation reference matrix is ​​also removed.

[0063] S33. Per-unit normalization of heterogeneous measurement data: Image pixels are homogeneous, common-scale data, while voltage and current are heterogeneous measurements with different dimensions and vastly different magnitudes. If directly substituted into the brightness / contrast comparison term, high-amplitude voltage will asymmetrically mask the structural features of low-amplitude current, causing similarity distortion. To overcome this obstacle, this step uses max-min normalization to perform independent numerical scaling column-wise. This maps heterogeneous data with different physical dimensions and huge numerical magnitude differences to a unified [0,1] calculation interval, completely eliminating the asymmetric masking of low-amplitude current data features by high-amplitude voltage data. The mathematical expression for this processing is:

[0064] ;

[0065] Where, x* t,j These are the normalized electrical parameter values; x t,j The values ​​are the original measurements; xmaxj and xminj are the maximum and minimum values ​​recorded at the j-th monitoring position within the entire continuous observation time window w, respectively. The corresponding simulation reference matrix also undergoes the same scaling transformation to maintain mathematical alignment.

[0066] S4. Calculate the structural similarity between the simulated reference matrix and the preprocessed original observation matrix of each candidate topology using the structural similarity index. The structural similarity index is quantitatively evaluated from three dimensions: mean deviation, fluctuation variance, and spatial cocorrelation. Select the candidate topology with the highest structural similarity index score as the actual operating topology of the current distribution network.

[0067] In this example, the observation matrix x is compared with the simulation reference matrix y of the i-th topology. i The specific formula calculation steps for feature evaluation are as follows:

[0068] S41. Calculation of Multidimensional Mathematical Feature Comparison Function: Calculate the brightness comparison function l(x, y) reflecting the deviation of the electrical operating mean. i The contrast comparison function c(x, y) reflects the intensity of dynamic fluctuations. i ), and the structural comparison function s(x, y) that reflects the coordinated changes in spatial distribution morphology. i The calculation formula is as follows:

[0069] ;

[0070] ;

[0071] ;

[0072] ;

[0073] ;

[0074] ;

[0075] S42 and SSI Index Comprehensive Evaluation: The comparison functions of the above three dimensions are multiplied and fused to obtain the final comprehensive structural similarity score.

[0076] ;

[0077] In the formula, and Let x represent the measured observation matrix and y represent the simulated reference matrix, respectively. i The mean; and For each, the corresponding variance; Let x and y be matrices i The covariance between them; c1, c2 and c3 are small constants added to ensure the stability of numerical calculations.

[0078] S43. Maximum Score Optimization and Deterministic Conclusion Output: Traverse and compare the candidate topology set containing N feasible topologies, and compare the comprehensive SSI scores of all configuration schemes. (Reference) Figure 3 The diagram shows the topology identification process. The configuration scheme that obtains the maximum SSI score is selected as the identification result of the actual operating topology of the current distribution network.

[0079] After selecting the candidate topology with the highest structural similarity index score as the actual operating topology of the current distribution network, a security verification mechanism based on discrimination margin is also included, specifically including:

[0080] Extract the optimal candidate topology that obtains the maximum structural similarity index score and the second-best candidate topology that obtains the second-largest structural similarity index score, and calculate the score difference between the two as the discrimination margin.

[0081] When the discrimination margin is less than the preset safety margin threshold, the output of a single, definitive automated identification conclusion is paused, and instead a set of suspected topologies with low confidence alarms is output to prompt manual intervention for verification.

[0082] While the Structural Similarity Index (SSI) has achieved success in image quality assessment, it is designed for homogeneous, bounded, dense, and scale-uniform two-dimensional pixel images. Directly applying it to distribution network topology identification faces several insurmountable technical obstacles: First, the comparison objects are fundamentally mismatched in dimension and nature—the physical benchmark obtained from power flow simulation is a single-section 1×n steady-state electrical vector, while field measurements are w×n dynamic matrices with time evolution; the time dimension is misaligned, and there is an essential difference between "static benchmark and dynamic observation." Second, the measurements are heterogeneous physical quantities such as voltage and current, with different dimensions and vastly different numerical magnitudes; directly substituting them into pixel-based data would be problematic. When comparing brightness / contrast based on homogeneity, high-amplitude components will systematically mask the structural features of low-amplitude components; third, real-time measurements commonly suffer from packet loss, and data gaps in the observation matrix will destroy the matrix integrity upon which SSI window statistics depend, making it impossible to calculate; fourth, image SSI measures visual perceptual similarity, and its value does not naturally correspond to differences in distribution network topology, so simply taking the maximum value lacks physical basis for discrimination; fifth, image SSI always outputs a unique optimal match, and when there are electrically highly similar "twin" topologies or when measurements are extremely degraded, it will give incorrect definitive conclusions with false confidence levels, without any ability to reject or alarm.

[0083] Specific Implementation Case 1: Robustness Identification and Analysis of Distribution Network Topology under Extreme Measurement Environments;

[0084] This demonstrates how the system, through its "multi-dimensional spatiotemporal observation matrix construction and SSI feature evaluation" architecture, dynamically filters local data anomalies and achieves high-precision topology tracking when facing non-ideal input conditions such as extreme measurement noise, drastic load fluctuations, and severe communication packet loss. The specific steps are as follows:

[0085] Basic physical framework construction and topology set initialization

[0086] Network model loading: The system loads the standard IEEE 33-node distribution network model, which contains 33 electrical nodes and 37 lines (including 5 tie lines).

[0087] Candidate Topology Set Generation: By simulating the switching of different segment switches and tie switches within the network, the system extracts 50 independent topologies that satisfy electrical operation constraints. These 50 independent structures are packaged and integrated to form a pre-built baseline candidate topology set, which serves as the enumeration space for subsequent similarity matching.

[0088] Real-world dynamic simulation and measurement data preprocessing

[0089] Dynamic disturbance injection: To simulate the complex operating conditions of a real power grid, the system superimposes two types of network dynamic disturbances at the measurement input layer of the physical grid structure: measurement noise (denoted as d1) and load fluctuations (denoted as d2). Among them, the system uses d2 to specifically simulate the dynamic differences caused by the time resolution mismatch between the underlying low-frequency load curve and the high-frequency real-time measurement sampling.

[0090] Packet Loss Tolerance and Data Interpolation: To address the simulated communication packet loss phenomenon (data missing), the system triggers a data preprocessing mechanism to extract the average measurement values ​​of adjacent time frames before and after the data breakpoint and automatically interpolate to fill the gap. This fault tolerance mechanism ensures the spatiotemporal integrity of the two-dimensional observation matrix used for subsequent SSI calculations.

[0091] SSI-based topological feature matching and robustness verification

[0092] Quantitative Evaluation and Chart Generation: The system calculated and evaluated the topology identification accuracy under different data missing rates and various combinations of (d1, d2) perturbations. The characteristic curves of identification accuracy under different network dynamics are shown below. Figure 4 As shown.

[0093] Extreme operating condition immunity test: according to Figure 2The system demonstrated strong fault tolerance in the quantitative data feedback: even under the harsh verification conditions of injecting extreme dynamic disturbances (measurement noise d1=80%, load fluctuation d2=100%) and superimposed with a measurement data missing rate of up to 50%, the matching mechanism driven by SSI remained robust, and the system's topology identification accuracy remained at a high level of 88%.

[0094] Performance boundary detection: The system continuously increases the data corruption intensity for stress testing. The results show that only when the data missing rate exceeds the critical threshold of 70% does the global structured features of the measurement matrix at the bottom layer suffer irreversible substantial damage, and the identification performance shows a significant step drop.

[0095] The final inference and verification conclusions are synthesized.

[0096] Core Mechanism Conclusion: Test results confirm that the global structural similarity matching mechanism based on multi-node electrical profiles proposed in this system can naturally filter and suppress local random measurement noise from the underlying principle. The system demonstrates strong physical immunity and industrial-grade robustness against frequent severe communication packet loss in distribution networks without relying on a large training sample.

[0097] Specific Implementation Case 2: Security Mechanism and Collaborative Identification Analysis for Misjudgment Prevention in Complex Distribution Network Topologies Based on Discrimination Margin

[0098] This demonstrates how, when faced with an expanded power distribution network, interference topologies with extremely similar electrical characteristics, or extremely severe data corruption conditions, the system uses "discrimination margin" verification to trigger a false positive security mechanism. This breaks the black box of single-point blind decision-making, achieving industrial-grade high reliability with conservative fault tolerance and manual collaborative verification. The specific steps are as follows:

[0099] Loading complex physical network structures and extracting high-similarity candidate sets;

[0100] Network Loading and Local Reconfiguration Simulation and Topology Set Construction: The system loads the IEEE 69-node distribution network physical model and underlying equipment ledger. For adjacent controllable switches at the network end or with extremely close electrical distances, the system executes reconfiguration operation commands to swap their open and closed states, generating the corresponding global network connectivity state. Based on the above local minor reconfiguration operations, the system extracts and constructs a candidate topology set. This set not only includes conventional topologies but also includes "twin" interference topologies with highly similar electrical spatial characteristics, which serve as an extreme verification space for subsequently triggering the anti-false positive security mechanism.

[0101] Extreme damage condition injection and full SSI score vector generation

[0102] Bad data corruption simulation: Inject extreme boundary conditions at the measurement input end, namely continuous high-frequency non-Gaussian noise and cliff-like data loss caused by group outages of communication links, so as to simulate the communication damage scenario under typhoon or extreme ice-snow disaster. Global similarity traversal: The system extracts the preprocessed real-time observation matrix according to the aforementioned steps S1-S4, and conducts a three-dimensional (mean deviation, fluctuation variance, spatial co-correlation) comprehensive cross comparison with all dynamic simulation reference matrices in the 69-node candidate topology set, so as to obtain a global SSI score vector containing the structural similarity of all candidate schemes. Quantitative extraction of identification margin and joint verification of safety threshold, extreme value extraction and margin calculation: Combined with Figure 5 the schematic diagram of the misjudgment-prevention safety mechanism for identification margin shown in the figure, after the traversal comparison is completely finished, the scheduling control module will accurately extract the optimal candidate topology with the maximum SSI score from the global SSI score vector (denoted as T opt , with the score being SSI max ), as well as the sub-optimal candidate topology with the second largest SSI score (denoted as T sub , with the score being SSI sub ). Subsequently, the system calculates the absolute score difference between the two in real time in the memory (i.e., ΔSSI = SSI max - SSI sub ), and defines it as the "identification margin" of the current topology identification conclusion.

[0103] Safety defense line interception verification: The system performs forced comparison and interception between the calculated identification margin ΔSSI and the preset safety margin threshold of the system (e, which is calibrated to 0.05 according to the sensitivity of the 69-node grid in this embodiment).

[0104] Misjudgment-prevention safety mechanism triggering and suspected set degradation output,

[0105] Identification risk and automatic blocking: When the safety verification module determines that the current identification margin satisfies ΔSSI<e, the system identifies from the underlying logic that there exists an interference topology with extremely similar electrical characteristics in the current power grid, or the data structure of the current real-time observation matrix has been damaged by the extremely harsh external environment to the extent that it cannot support a unique deterministic conclusion. At this time, the system immediately executes fusing interception, pauses and blocks the conventional automatic output path of a single determined topology conclusion.

[0106] Conservative alarm and manual collaboration: The system instead triggers the misjudgment-prevention fault-tolerant degradation mechanism, and sets T opt , T subEven the top K most similar configurations with similar scores are packaged together to generate a "suspected topology set" with a "low confidence" flag and a red alarm. The system simultaneously pushes this suspected set, the score difference panel, and the corresponding underlying physical ledger vouchers to the power grid master station dispatch console, forcibly prompting manual intervention. The dispatcher can then use this reduced suspected set, combined with on-site switch feedback information or inspection records, to cross-check and collaboratively determine the solution, thereby fundamentally preventing serious power misoperation accidents that might be caused by the algorithm's "blind guessing" under extreme data degradation conditions.

[0107] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technology of the present invention without departing from the scope of the present invention are within the protection scope of the present invention.

Claims

1. A method for topology identification of distribution networks based on structural similarity index, characterized in that: The method includes the following steps: S1. Obtain real-time load data and input it into the candidate topology set to perform power flow simulation. Calculate the full network voltage and current distribution of each topology in the topology library. Based on the physical distribution of sensors, extract the measurement data of the corresponding n monitoring locations from the simulation results and generate the reference state vector of each candidate topology. S2. Using high-resolution on-site measurement equipment, real-time electrical parameters of w time frames are acquired at a uniform frequency within a continuous observation time window. The measurement data of n monitoring locations are arranged according to the spatiotemporal dimension to construct a w×n-dimensional original observation matrix. S3. Copy and tile the reference state vector w times along the time axis to construct a w×n dimensional simulation reference matrix that matches the dimension of the original observation matrix. Complete the missing data in the original observation matrix. Perform normalization preprocessing on the original observation matrix and the simulation reference matrix to map the heterogeneous measurement data to a unified numerical range. S4. Calculate the structural similarity between the simulated reference matrix and the preprocessed original observation matrix of each candidate topology using the structural similarity index. The structural similarity index is quantitatively evaluated from three dimensions: mean deviation, fluctuation variance, and spatial cocorrelation. Select the candidate topology with the highest structural similarity index score as the actual operating topology of the current distribution network.

2. The distribution network topology identification method based on structural similarity index according to claim 1, characterized in that: The step of extracting measurement data corresponding to n monitoring locations from the simulation results based on the physical distribution of sensors, and generating the baseline state vector for each candidate topology, includes: By utilizing the mapping relationship between the distribution network topology map and the equipment ledger in the topology library, the node voltage vector or branch current vector in the distribution network power flow calculation result corresponding to the real-time load distribution data is spatially filtered according to the physical node index and branch index of the actual deployed high-resolution measurement equipment, and combined to form a 1×n-dimensional static electrical feature vector, which serves as the initial physical reference benchmark for the candidate topology.

3. The distribution network topology identification method based on structural similarity index according to claim 1, characterized in that: In step S3, the missing data completion in the original observation matrix includes: When it is detected that the measurement data of a certain monitoring position in the w×n-dimensional original observation matrix is ​​lost or invalid in a specific time frame, the measurement values ​​of the effective time frames before and after the monitoring position on the time axis are extracted, and the arithmetic mean is calculated for interpolation and filling. If multiple consecutive frames of data are missing on the timeline, the system first attempts to interpolate and fill the missing data by combining the measurement values ​​of the adjacent valid time frames before and after the monitoring location on the timeline. If the filling fails, the column data corresponding to the measurement point is removed, and the corresponding column data in the simulation reference matrix is ​​also removed.

4. The distribution network topology identification method based on structural similarity index according to claim 1, characterized in that: The normalization preprocessing of the original observation matrix and the simulation reference matrix includes: A per-unit scaling algorithm is used to independently scale the voltage observation column and the current observation column in the w×n-dimensional original observation matrix and the simulation reference matrix, respectively, to eliminate the differences in the inherent magnitudes of different physical quantities.

5. The distribution network topology identification method based on structural similarity index according to claim 1, characterized in that: The calculation of the structural similarity between the simulated reference matrix and the preprocessed original observation matrix of each candidate topology using the structural similarity index includes: The mean comparison function, variance comparison function, and covariance comparison function between the original observation matrix and the simulated reference matrix are calculated respectively to quantify the features of the three dimensions of mean deviation, fluctuation variance, and spatial cocorrelation. By adaptively adjusting the weighting factors of the electrical characteristics of the distribution network, the mean comparison function, variance comparison function, and covariance comparison function are multiplied and fused to obtain the comprehensive score of the structural similarity index.

6. The distribution network topology identification method based on structural similarity index according to claim 1, characterized in that: After selecting the candidate topology with the highest structural similarity index score as the actual operating topology of the current distribution network, a security verification mechanism based on discrimination margin is also included, specifically including: Extract the optimal candidate topology that obtains the maximum structural similarity index score and the second-best candidate topology that obtains the second-largest structural similarity index score, and calculate the score difference between the two as the discrimination margin. When the discrimination margin is less than the preset safety margin threshold, the output of a single, definitive automated identification conclusion is paused, and instead a set of suspected topologies with low confidence alarms is output to prompt manual intervention for verification.

7. The distribution network topology identification method based on structural similarity index according to claim 1, characterized in that: The candidate topology set is constructed in the following ways: Obtain the physical infrastructure ledger and connectivity parameters of the distribution network at the bottom layer, and abstract the actual distribution network into a mathematical graph theory model G=(V, E), where the node set V represents the bus, transformer and load connection point, the edge set E represents the physical feeder, and the edge set is further subdivided into uncontrollable static line segments and controllable switching equipment. Driven by the graph theory spanning tree algorithm and with the radial operation constraints of the distribution network as a prerequisite, all combinations of switch states that satisfy the characteristics of connectivity and no loops are extracted in the graph model space to generate an initial radial topology library. The initial radial topology library is pruned by applying multiple operational boundary constraints, including system global connectivity and island avoidance constraints, power supply capacity range constraints of upstream power sources, node voltage over-limit constraints, and branch thermal capacity constraints, ultimately generating a candidate topology set.