Power distribution network topology identification method

Abstract

The invention belongs to the technical field of power electronics, and relates to a power distribution network topology identification method, which comprises the following steps of: 1, constructing a topology priority vector TPV to carry out heuristic search; 2, constructing and training a terminal node connectivity prediction network TNCPN; 3, executing iterative topology search, confirmation and aggregation; 4, calculating equivalent injection and executing dynamic model correction; 5, determining iteration termination and generating a final topological structure; the method completely gets rid of dependence on physical parameters such as accurate line impedance and susceptance, a decision model only needs to utilize node voltage amplitude and power injection measurement data which are commonly available in a power distribution network, and the characteristic fundamentally solves the technical problem that identification fails due to inaccurate parameters of a traditional model-based driving method.

Description

Technical Field

[0001] This invention belongs to the field of power electronics technology, and specifically relates to a method for identifying the topology of a power distribution network. Background Technology

[0002] Distribution network topology identification methods refer to the technical means of determining the physical connection relationships (i.e., topology) of various devices in the distribution network by analyzing their measurement data or operating status information. Based on different implementation principles and data sources, they can be mainly divided into two categories: The first category is model-driven methods. The core of these methods lies in using an accurate physical model of the power grid to infer the topology. Methods represented by Mixed Integer Linear Programming (MILP) set the on / off states of power grid switches as integer variables in the mathematical model, and then use optimization algorithms to find the switch combinations that minimize the error between the model calculation results and the actual measurement data, thereby determining the network topology. While the physical meaning of these methods is clear, their practicality is greatly limited because they are highly dependent on accurate and difficult-to-obtain power grid parameters such as line impedance. Once these parameters are inaccurate, the identification accuracy cannot be guaranteed. The second category is data-driven methods. These methods aim to avoid dependence on accurate parameters and directly learn patterns from measurement data. Early solutions, such as voltage correlation analysis, utilized the statistical characteristics of highly synchronized voltage changes at connected nodes. Current mainstream technology uses deep learning to treat topology identification as a classification problem. The common approach is to train a deep neural network (DNN) using simulation data, allowing it to learn the complex mapping from input features such as node voltage and power to the output result of "connected" or "unconnected". However, a key drawback of this method is that its model is static: once trained, the model is fixed. In actual iterative identification processes, as local topologies are identified and aggregated, the equivalent structure of the network continuously evolves. This static model cannot adapt to these changes, leading to a significant performance degradation in the later stages of identification and the accumulation of errors, ultimately affecting the overall accuracy.

[0003] Therefore, existing methods for identifying distribution network topologies have the following drawbacks: Strong dependence on parameters: Model-based methods require highly accurate line parameters, but in actual distribution networks, these parameters are often unknown, inaccurate, or change over time, leading to a serious decrease in identification accuracy.

[0004] Insufficient physical interpretability: Some data-driven methods, especially complex deep learning models, are like a "black box," with opaque decision-making processes and difficult-to-understand physical basis, making it difficult for power grid dispatchers to fully trust their identification results.

[0005] High cost of measurement equipment: Some advanced data-driven methods rely on high-precision, phase-angle-inclusive measurement data provided by synchronous phasor measurement units (PMUs), but the deployment of PMUs in distribution networks is costly and has not yet become widespread.

[0006] Limitations and error accumulation of static models: This is a core flaw in existing data-driven methods. Topology identification is essentially an iterative process; as the local topology is confirmed, the network's analysis objects (nodes) aggregate and change. Existing static models remain fixed after training. When faced with more complex "aggregated nodes" with electrical characteristics differing from the initial training data distribution, their identification ability significantly decreases, leading to incorrect judgments. These errors accumulate in subsequent iterations, ultimately causing overall identification failure.

[0007] Therefore, a method or apparatus with strong interpretability and high identification accuracy is needed to solve the above-mentioned technical problems. Summary of the Invention

[0008] This invention provides the following technical solution: a method for identifying the topology of a power distribution network, comprising the following steps: Step 1: Construct a Topology Priority Vector (TPV) for heuristic search. Using historical voltage measurement snapshot data collected under different operating conditions of the distribution network, construct a Topology Priority Vector (TPV) that reflects the electrical hierarchy of nodes.

[0009] Step 2: Construct and train the Terminal Node Connectivity Prediction Network (TNCPN). Construct and complete the initial training of the TNCPN. The TNCPN is a deep learning model. The TNCPN is used as the core decision-making tool in the subsequent topology identification process to determine the possibility of a direct physical connection between any two nodes.

[0010] Step 3: Perform iterative topology search, confirmation, and aggregation. This is achieved through an iterative execution process. In each iteration, the system performs bottom-up topology reconstruction based on the electrical hierarchy established by the topology priority vector TPV.

[0011] Step 4: Calculate equivalent injection and perform dynamic model correction. After the node aggregation operation in step 3 is completed, perform equivalent injection calculation and dynamic model correction to ensure the physical consistency and identification accuracy of the algorithm in subsequent iterations.

[0012] Step 5: Determine if the iteration terminates and generate the final topology. The iterative identification process is executed periodically. After each iteration, a status evaluation is performed to determine whether the preset system-level termination conditions used to terminate the entire identification process are met.

[0013] The overall technical architecture of this invention is a closed-loop system that tightly integrates "bottom-up iterative aggregation" with "online dynamic model correction". This framework is not a one-time static judgment, but a continuously evolving identification process, in which the output of each iteration (the newly generated aggregation node) is used as the input for the next iteration and the basis for model optimization, thus forming a complete and self-improving identification process.

[0014] The dynamic model correction mechanism of this invention is the core technology. Specifically, during the iterative identification process, new training samples are automatically and online generated based on the electrical characteristics of the newly aggregated, more complex virtual nodes. These samples are then used to fine-tune or retrain the machine learning model (TNCPN) used for connection relationship judgment. The purpose of this mechanism is to enable the identification ability of the decision model to evolve in sync with the complexity of the identification task, thereby fundamentally solving the performance degradation problem of static models caused by data distribution shifts in the later stages of iteration.

[0015] The calculation results of this invention not only include the algebraic sum of the net injected power of all original nodes within the virtual node, but also further include a compensation value for the total power loss of the internal connecting lines calculated based on the Line Loss Estimation Network (LEN). This method ensures that the aggregated virtual node is physically complete in terms of external electrical characteristics, providing a crucial data foundation for the effectiveness of subsequent dynamic model correction.

[0016] Preferably, step 1 specifically includes: Step 1-1: Perform pairwise comparisons on all nodes in the network. For any two nodes, count the number of times in all data snapshots that the voltage of one node is higher than that of the other node.

[0017] Step 1-2: Calculate a quantified “explicit score” for each node based on the statistical results.

[0018] Steps 1-3: All nodes are sorted from high to low according to the "explicit score" to form the final topology priority vector (TPV); this TPV is used as a guide for the search order in the subsequent identification process.

[0019] Steps 1-4: Start with the lowest-ranked node in the TPV and perform topology identification in a bottom-up order; when searching for the parent node of any child node, the range of candidate parent nodes is limited to all nodes in the TPV that are ranked higher than the child node.

[0020] Preferably, step 2 specifically includes: Step 2-1: Training the model involves a dataset containing positive and negative labeled samples. Positive samples consist of known, directly connected parent-child node pairs, while negative samples consist of non-directly connected node pairs. The label for positive samples is 1, and the label for negative samples is 0.

[0021] Step 2-2: Use a hard negative sampling strategy to select node pairs that are electrically close but have no direct physical connection in order to enhance the model's ability to distinguish in fuzzy scenes.

[0022] Steps 2-3: The feature vector of any node pair to be evaluated input to TNCPN is a four-dimensional array, which includes: the voltage magnitude of the candidate parent node, the voltage magnitude of the child node, the net active power injection and the net reactive power injection of the child node.

[0023] Steps 2-4: After model processing, the output is a single numerical value between 0 and 1, which represents the probability that there is a direct physical connection between two nodes. This initially trained TNCPN will be used to perform the first round of topology search and confirmation.

[0024] Preferably, step 3 specifically includes: Step 3-1: Perform an upward search operation. Select the node with the lowest ranking in TPV and whose upstream connection relationship has not yet been determined as the current target node to be identified. Then, form a set of candidate parent nodes for all nodes in TPV with a higher ranking than the target node.

[0025] Step 3-2: Perform connection confirmation operation. For each candidate parent node in the candidate parent node set, construct a feature vector of a preset dimension with the electrical state information of the target node and input it into the current TNCPN for processing to obtain a quantized connection probability value. After completing the probability calculation for all candidate parent nodes, determine the candidate parent node with the highest connection probability value as the unique parent node of the target node and establish a determined topology connection.

[0026] Step 3-3: Perform node aggregation operation. After the iteration is completed, aggregate all identified nodes that are determined to be connected to the same parent node to form a new equivalent virtual node. In subsequent iterations, this equivalent virtual node will participate in the subsequent topology identification process as an independent electrical entity with equivalent electrical characteristics. The iteration process is executed cyclically until the connection relationship of all nodes in the network is determined.

[0027] Preferably, step 4 specifically includes: Step 4-1: Calculate the equivalent power injection; for each newly generated equivalent virtual node, calculate the equivalent power injection value presented to the outside world.

[0028] Step 4-2: Perform dynamic model calibration. Using newly generated virtual nodes and physically consistent equivalent power injection, construct a new set of training samples for more complex aggregation structures. These new samples are used to perform online fine-tuning of the current Terminal Node Connectivity Prediction Network (TNCPN). This dynamic calibration process allows the TNCPN's decision boundary to adaptively optimize based on the more structurally complex nodes that emerge during iterations, thereby improving its model generalization ability and maintaining high accuracy in subsequent iterations when making connectivity judgments for large-scale aggregation nodes.

[0029] More preferably, the calculation method of step 4-1 includes two parts: the first part is to perform an algebraic summation of the net injected power of all original nodes contained in the virtual node; the second part is to add a compensation value for the total power loss of all connection lines inside the virtual node on this basis; the total power loss is estimated by a pre-trained deep learning-based line loss estimation network.

[0030] Preferably, in step 5, the termination condition includes: from the perspective of the algorithm, all nodes representing the original load or branch in the distribution network have been merged into a single equivalent entity that represents the entire network and is directly connected to the system root node through hierarchical aggregation.

[0031] If the system determines that the termination condition has not been met after the state assessment, it will proceed to the next iteration and continue to identify the remaining, more complex nodes using the TNCPN whose performance has been enhanced after the previous correction. If the system determines that the termination condition has been met after the state assessment, the iteration process will terminate immediately.

[0032] More preferably, after the iteration process terminates, a topology finalization step is initiated, systematically integrating and parsing all parent-child node connection relationships recorded and stored in all historical iteration rounds; performing "reverse parsing" or "de-virtualization" on connection relationships involving virtual nodes; parsing all connection relationships, converting all indirect, virtual node-based connections into direct, original physical node-based connections; finally, the fully parsed connection relationships are compiled into a definitive final topology data structure that can completely describe all physical connections of the distribution network and output as the final identification result.

[0033] More preferably, the "reverse resolution" or "de-virtualization" means that when the parent node of a node is recorded as a virtual node, the system traces the aggregation history of the virtual node to determine the original node contained within it that serves as the actual connection point.

[0034] The beneficial effects of this invention are: 1. This invention achieves complete decoupling of power grid line parameters, possessing broad engineering applicability. It completely eliminates the reliance on precise physical parameters such as line impedance and susceptance; its decision model only needs to utilize readily available node voltage amplitude and power injection measurement data in distribution networks. This characteristic fundamentally solves the technical problem of identification failure caused by inaccurate parameters in traditional model-driven methods, greatly reducing the application threshold and deployment cost of this technology, enabling it to operate effectively even in real-world power grid environments with unknown or frequently changing parameters.

[0035] 2. This invention significantly improves identification accuracy and algorithm robustness by introducing a dynamic model correction mechanism. The core innovation of this invention lies in the fact that its identification model is not statically fixed but can be adaptively fine-tuned during iteration. This dynamic correction mechanism successfully overcomes the inherent defect of existing static data-driven models, which suffer performance degradation in the later stages of iteration when facing complex aggregation nodes. By continuously learning the newly emerging and more complex electrical characteristics of nodes, this invention can effectively suppress the accumulation and propagation of early small errors, ensuring that the algorithm maintains a high level of judgment ability throughout the entire "bottom-up" identification process, ultimately achieving a topology identification accuracy close to 100%.

[0036] 3. This invention enhances the physical interpretability of the identification process and the credibility of the results. The core decision-making mechanism of this invention, the Terminal Node Connectivity Prediction Network (TNCPN), is based on well-defined power grid physical laws (voltage-power correlation patterns), and is not an unanalyzable "black box." The algorithm's behavior during execution—making high-confidence judgments only after virtual nodes have undergone complete aggregation and their electrical characteristics have regained physical consistency—further verifies the high degree of self-consistency between its internal logic and physical reality. This interpretability enhances the trust of technical personnel in the identification results, facilitating the analysis and tracing of abnormal results.

[0037] 4. This invention improves identification efficiency and ensures the clarity of decision-making. By constructing a Topology Priority Vector (TPV) in the initial stage, this invention provides efficient heuristic guidance for the subsequent search process, intelligently pruning the large candidate parent node search space and avoiding global brute-force search, thereby significantly improving the computational efficiency of the algorithm. Simultaneously, the TNCPN, after sufficient training and dynamic correction, exhibits a clear polarization in its output connection probabilities: for true connections, the probability value is extremely close to 1.0; for non-connections, it is extremely close to 0.0. This high-confidence, low-ambiguity decision-making ensures the reliability of topology confirmation at each step. Attached Figure Description

[0038] Figure 1This is a comparison diagram of the voltage difference-load correlation patterns between directly connected nodes and non-adjacent nodes in a power distribution network topology identification method according to the present invention. Figure 2 This is a structural diagram of the adaptive iterative topology aggregation distribution network topology identification method of the present invention; Figure 3 This is a flowchart of the steps of the present invention. Detailed Implementation

[0039] The related technologies of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0040] like Figures 1-3 As shown, the distribution network topology identification method in this embodiment includes: Step 1: Construct a Topology Priority Vector (TPV) for heuristic search. This step utilizes historical voltage measurement snapshots collected from the distribution network under different operating conditions to construct a Topology Priority Vector (TPV) reflecting the electrical hierarchy of nodes. Specifically, the algorithm performs pairwise comparisons of all nodes in the network. For any two nodes, the system counts the number of times one node's voltage is higher than the other's across all data snapshots. Based on this statistical result, a quantified "Dominance Score" is calculated for each node. Subsequently, all nodes are sorted from highest to lowest according to their "Dominance Score," forming the final Topology Priority Vector (TPV). In this vector, nodes with higher scores (typically corresponding to upstream nodes with higher electrical hierarchy) are listed first, and nodes with lower scores (typically corresponding to downstream end nodes with lower electrical hierarchy) are listed last. This TPV is used as a guide for the search order in subsequent identification processes. The topology identification process starts with the node with the lowest ranking in the TPV and proceeds in a bottom-up order. When searching for the parent node of any child node, the range of candidate parent nodes is limited to all nodes with a higher ranking in the TPV than that child node.

[0041] Step 2: Construct and train the Terminal Node Connectivity Prediction Network (TNCPN) The core task of this step is to construct and complete the initial training of a Terminal Node Connectivity Prediction Network (TNCPN). This TNCPN is a deep learning model that serves as the core decision-making tool in the subsequent topology identification process, determining the likelihood of a direct physical connection between any two nodes. As shown in Figure 1, the model is built upon a physical principle: in a power grid, the voltage difference between two directly connected nodes is strongly correlated with the power flow on the lines between them; however, for indirectly connected nodes, this relationship becomes weaker and more complex due to the combined influence of multiple power flow paths. The TNCPN is trained to learn and distinguish between these two different voltage-power correlation patterns.

[0042] Training this model requires a dataset containing both positive and negative labeled samples, typically generated through power grid simulations. Positive samples (labeled 1) consist of known, directly connected parent-child node pairs. Negative samples (labeled 0) consist of non-directly connected node pairs. Specifically, a hard negative sampling strategy is employed to select node pairs that are electrically close (e.g., neighbors at voltage levels) but without a direct physical connection, enhancing the model's ability to distinguish between nodes in ambiguous scenarios. For any node pair to be evaluated (including a child node i to be identified and its candidate parent node j), the feature vector input to the TNCPN is a four-dimensional array, specifically including: the voltage magnitude of the candidate parent node, the voltage magnitude of the child node, the net active power injection of the child node, and the net reactive power injection. After processing, the model outputs a single value between 0 and 1, representing the probability that a direct physical connection exists between the two nodes. This initially trained TNCPN will be used to perform the first round of topology search and confirmation.

[0043] Step 3: Perform iterative topology search, confirmation, and aggregation. like Figure 2 As shown, the topology identification method of the present invention is implemented through an iterative execution process. In each iteration, the system performs a bottom-up topology reconstruction based on the electrical hierarchy established by the Topology Priority Vector (TPV). This process specifically includes the following operations: First, an upward search operation is performed. The system selects the node with the lowest ranking in the TPV whose upstream connection relationship has not yet been determined as the current target node to be identified. Then, all nodes in the TPV with a higher ranking than the target node are used to form a set of candidate parent nodes.

[0044] Next, a connection confirmation operation is performed. For each candidate parent node in the candidate parent node set, a feature vector of a preset dimension is constructed by combining it with the electrical state information of the target node. This vector is then input into the current Terminal Node Connectivity Prediction Network (TNCPN) for processing to obtain a quantified connection probability value. After the system completes the probability calculation for all candidate parent nodes, the candidate parent node corresponding to the highest connection probability value is determined as the unique parent node of the target node, thereby establishing a defined topology connection.

[0045] Finally, node aggregation is performed. After one iteration, all identified nodes connected to the same parent node are aggregated to form a new equivalent virtual node. This equivalent virtual node will participate in subsequent topology identification processes as an independent electrical entity with equivalent electrical characteristics in subsequent iterations. This iterative process is repeated until the connection relationships of all nodes in the network are determined.

[0046] Step 4: Calculate equivalent injection and perform dynamic model correction. After the node aggregation operation described in step 3 is completed, the system will perform two core steps: equivalent injection calculation and dynamic model correction, to ensure the physical consistency and identification accuracy of the algorithm in subsequent iterations.

[0047] First, the equivalent power injection is calculated. For each newly generated equivalent virtual node, the system calculates its external equivalent power injection value. This calculation involves two parts: the first part is the algebraic summation of the net injected power of all original nodes contained in the virtual node; the second part is to add a compensation value for the total power loss of all internal connection lines of the virtual node. The total power loss of the internal lines is estimated using a pre-trained, deep learning-based Line Loss Estimation Network (LEN).

[0048] Secondly, dynamic model calibration is performed. The system automatically constructs a new set of training samples for more complex aggregation structures using newly generated virtual nodes and their physically consistent equivalent power injections. These new samples are used to perform online fine-tuning of the current Terminal Node Connectivity Prediction Network (TNCPN). This dynamic calibration process allows the TNCPN's decision boundary to adaptively optimize based on the more structurally complex nodes that emerge during iterations, thereby improving its model generalization ability and maintaining high accuracy in subsequent iterations when making connectivity judgments for large-scale aggregation nodes.

[0049] Step 5: Determine if the iteration terminates and generate the final topology. The iterative identification process described in this invention, which includes a series of operations such as topology search, confirmation, aggregation, and dynamic model correction, will be executed periodically. After each iteration, the system will perform a state evaluation to determine whether a preset system-level termination condition for terminating the entire identification process is met.

[0050] The termination condition is strictly defined as follows: From the algorithm's perspective, all nodes representing the original loads or branches in the distribution network have been aggregated into a single equivalent entity that represents the entire network and is directly connected to the system root node through a hierarchical aggregation process. This state signifies that all scattered nodes in the network have been successfully assigned a unique affiliation in the upstream topology, and the network has been reconstructed into a loop-free, coherent tree structure. If, after state evaluation, the system determines that the termination condition has not yet been met, i.e., there are still two or more independent nodes or virtual nodes in the network that have not been aggregated into a single entity, the system will automatically enter the next iteration and use the Terminal Node Connectivity Prediction Network (TNCPN), whose performance has been enhanced after the previous correction, to continue identifying the remaining, more structurally complex nodes.

[0051] Once the system determines that the termination condition has been met after a certain iteration, the iteration process terminates immediately. Subsequently, the system initiates a topology finalization step. The core task of this step is to systematically integrate and parse all parent-child node connections recorded and stored in all historical iterations. A key operation in this process is the "reverse parsing" or "de-virtualization" of connections involving virtual nodes. That is, when a node's parent node is recorded as a virtual node, the system needs to trace the aggregation history of that virtual node to accurately determine the original node contained within it that serves as the actual connection point. By performing this parsing on all connections, the system converts all indirect, virtual node-based connections into direct, original physical node-based connections. Finally, these fully parsed connections are compiled into a definitive final topology data structure that can completely describe all physical connections of the distribution network, such as an adjacency list or connection matrix, and output as the final identification result of this invention.

[0052] In summary, this invention can completely decouple power grid line parameters and has broad engineering applicability. This invention completely eliminates the dependence on precise physical parameters such as line impedance and susceptance. Its decision model only needs to utilize the node voltage amplitude and power injection measurement data that are commonly available in the distribution network. This feature fundamentally solves the technical problem of identification failure caused by inaccurate parameters in traditional model-driven methods, greatly reducing the application threshold and deployment cost of this technology, and enabling it to operate effectively in actual power grid environments with unknown or frequently changing parameters.

[0053] It should be emphasized that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for identifying the topology of a power distribution network, characterized in that, Includes the following steps: Step 1: Construct a Topology Priority Vector (TPV) for heuristic search. Using historical voltage measurement snapshot data collected under different operating conditions of the distribution network, construct a Topology Priority Vector (TPV) that reflects the electrical hierarchy of nodes. Step 2: Construct and train the Terminal Node Connectivity Prediction Network (TNCPN). Construct and complete the initial training of the TNCPN. The TNCPN is a deep learning model. The TNCPN is used as the core decision-making tool in the subsequent topology identification process to determine the possibility of a direct physical connection between any two nodes. Step 3: Perform iterative topology search, confirmation and aggregation. This is achieved through an iterative execution process. In each iteration, the system performs bottom-up topology reconstruction based on the electrical hierarchy established by the topology priority vector TPV. Step 4: Calculate equivalent injection and perform dynamic model correction. After the node aggregation operation in Step 3 is completed, perform equivalent injection calculation and dynamic model correction to ensure the physical consistency and identification accuracy of the algorithm in subsequent iterations. Step 5: Determine if the iteration terminates and generate the final topology. The iterative identification process is executed periodically. After each iteration, a status evaluation is performed to determine whether the preset system-level termination conditions used to terminate the entire identification process are met.

2. The distribution network topology identification method according to claim 1, characterized in that, Step 1 specifically includes: Step 1-1: Perform pairwise comparisons on all nodes in the network. For any two nodes, count the number of times in all data snapshots that the voltage of one node is higher than that of the other node. Step 1-2: Calculate a quantified "explicit score" for each node based on the statistical results; Steps 1-3: All nodes are sorted from high to low according to the "explicit score" to form the final topology priority vector (TPV); Steps 1-4: Start with the lowest-ranked node in the TPV and perform topology identification in a bottom-up order; when searching for the parent node of any child node, the range of candidate parent nodes is limited to all nodes in the TPV that are ranked higher than the child node.

3. The distribution network topology identification method according to claim 1, characterized in that, Step 2 specifically includes: Step 2-1: Training the model involves a dataset containing positive and negative labeled samples. Positive samples consist of known, directly connected parent-child node pairs, while negative samples consist of non-directly connected node pairs. The label for positive samples is 1, and the label for negative samples is 0. Step 2-2: Use a hard negative sampling strategy to select node pairs that are electrically close but have no direct physical connection in order to enhance the model's ability to distinguish in fuzzy scenes; Steps 2-3: The feature vector of any node pair to be evaluated input to TNCPN is a four-dimensional array, which includes: the voltage magnitude of the candidate parent node, the voltage magnitude of the child node, the net active power injection and the net reactive power injection of the child node. Steps 2-4: After model processing, the output is a single value between 0 and 1, which represents the probability that there is a direct physical connection between the two nodes.

4. The distribution network topology identification method according to claim 1, characterized in that, Step 3 specifically includes: Step 3-1: Perform an upward search operation. Select the node with the lowest ranking in TPV and whose upstream connection relationship has not yet been determined as the current target node to be identified. Then, form a set of candidate parent nodes for all nodes in TPV with a higher ranking than the target node. Step 3-2: Perform connection confirmation operation. For each candidate parent node in the candidate parent node set, construct a feature vector of a preset dimension with the electrical state information of the target node and input it into the current TNCPN for processing to obtain a quantized connection probability value. After completing the probability calculation for all candidate parent nodes, determine the candidate parent node with the highest connection probability value as the unique parent node of the target node and establish a determined topology connection. Step 3-3: Perform node aggregation operation. After the iteration is completed, aggregate all identified nodes that are determined to be connected to the same parent node to form a new equivalent virtual node. The iteration process is executed repeatedly until the connection relationship of all nodes in the network is determined.

5. The distribution network topology identification method according to claim 1, characterized in that, Step 4 specifically includes: Step 4-1: Calculate the equivalent power injection; for each newly generated equivalent virtual node, calculate the equivalent power injection value presented to the outside world. Step 4-2: Perform dynamic model correction, using newly generated virtual nodes and physically consistent equivalent power injection to construct a new set of training samples for more complex aggregation structures.

6. The distribution network topology identification method according to claim 5, characterized in that, The calculation method in step 4-1 includes two parts: the first part is to perform an algebraic summation of the net injected power of all original nodes contained in the virtual node; the second part is to add a compensation value for the total power loss of all connection lines inside the virtual node on this basis; the total power loss is estimated by a pre-trained deep learning-based line loss estimation network.

7. The distribution network topology identification method according to claim 1, characterized in that, In step 5, the termination condition includes: from the perspective of the algorithm, all nodes representing the original load or branch in the distribution network have been merged into a single equivalent entity that represents the entire network and is directly connected to the root node of the system through hierarchical aggregation. If the system determines that the termination condition has not been met after the state assessment, it will proceed to the next iteration and continue to identify the remaining, more complex nodes using the TNCPN whose performance has been enhanced after the previous correction. If the system determines that the termination condition has been met after the state assessment, the iteration process will terminate immediately.

8. The distribution network topology identification method according to claim 7, characterized in that, After the iteration process terminates, the topology finalization step is initiated, systematically integrating and parsing all parent-child node connection relationships recorded and stored in all historical iteration rounds; performing "reverse parsing" or "de-virtualization" on connection relationships involving virtual nodes; parsing all connection relationships, converting all indirect, virtual node-based connections into direct, original physical node-based connections; finally, the fully parsed connection relationships are compiled into a definitive final topology data structure that can completely describe all physical connections of the distribution network and output as the final identification result.

9. The distribution network topology identification method according to claim 8, characterized in that, The "reverse resolution" or "de-virtualization" refers to the following: when the parent node of a node is recorded as a virtual node, the system traces the aggregation history of the virtual node to determine the original node contained within it that serves as the actual connection point.

Images