Sparse time sequence Bayesian network construction method and system based on power distribution network topology constraint

Abstract

The invention discloses a sparse time sequence Bayesian network construction method and system based on power distribution network topological constraints, and the method specifically comprises the steps: constructing a power distribution network topological graph, and calculating an adjacent matrix between nodes and a k-hop neighborhood matrix Nk; based on the power distribution network topology distance information and the matrix Nk, generating a hard constraint rule, and constructing a topology dependence mask matrix M; calculating an electrical influence range of the fault point on surrounding nodes to obtain an electrical influence matrix E; if the electrical influence coefficient of one node on the other node is smaller than a set value, deleting the corresponding dependent edge; introducing a data source reliability matrix R, and performing hard deletion or soft weakening on edges with reliability lower than a threshold value; combining the matrixes M, E and R to synthesize a sparse structure matrix S; and taking the matrix S as a space skeleton, adding a time dimension autoregression edge, and constructing a complete sparse time sequence Bayesian network. According to the invention, by guiding the rarefaction of the network structure, the number of network edges and the number of parameters are effectively reduced, and the trainability and reasoning efficiency of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology for distribution networks, and in particular to a method and system for constructing sparse temporal Bayesian networks based on distribution network topology constraints. Background Technology

[0002] With the continuous increase in the proportion of renewable energy integration and the comprehensive advancement of new power system construction, the operation of distribution networks exhibits characteristics of high uncertainty, multiple disturbance sources, and complex interactions. Distribution network fault diagnosis and assessment are gradually shifting from traditional rule-based criteria to data-driven and probabilistic reasoning methods. Temporal Bayesian Networks (TBNs), due to their advantages such as modeling multi-source asynchronous information, representing uncertainty, and possessing interpretability, have become an important direction in current research on intelligent diagnosis of distribution networks.

[0003] However, power distribution networks are characterized by a large number of nodes, complex topology, numerous branches, and uneven distribution of measurement points. Traditional Bayesian network construction methods often employ full-structure or weakly constrained structure learning, resulting in a huge number of edges and an explosion of parameters. This leads to the following challenges in practical engineering scenarios:

[0004] (1) The modeling scale is too large: each measurement point, protection action, switch status, etc. are all treated as TBN nodes, and the number of edges of the complete network increases exponentially with the topology scale;

[0005] (2) Training difficulties: The dimension of the conditional probability table (CPT) of TBN expands rapidly with the number of parent nodes, which is particularly prominent in distribution network scenarios where samples are scarce (few real fault samples).

[0006] (3) Inference is time-consuming and cannot be implemented in real time in engineering: Traditional inference algorithms have extremely high time complexity on dense networks, making it difficult to meet the requirements of actual distribution networks for second-level and millisecond-level diagnosis;

[0007] (4) Lack of physical constraints in the structure leads to instability: Some "edges without physical meaning" are learned incorrectly, resulting in poor model interpretability or even inconsistent inference results.

[0008] Therefore, there is an urgent need for a sparse Bayesian network construction method oriented towards the physical topology constraints of distribution networks, which can achieve efficient training, highly reliable inference, and engineering deployability in large-scale networks while ensuring diagnostic accuracy. Summary of the Invention

[0009] The purpose of this invention is to provide a method for constructing sparse temporal Bayesian networks based on distribution network topology constraints. By using multiple engineering priors such as distribution network topology, electrical influence range, and data source reliability, the method guides the sparsification of the network structure, effectively reducing the number of network edges and parameters, and improving the model's trainability and inference efficiency.

[0010] The technical solution to achieve the purpose of this invention is: a method for constructing a sparse temporal Bayesian network based on distribution network topology constraints, comprising the following steps:

[0011] Step 1: Construct the distribution network topology and extract the adjacency matrix: Construct the distribution network topology, calculate the adjacency matrix between each node and the k-hop neighborhood matrix N. k ;

[0012] Step 2, Parent node constraints based on topology radius: Based on distribution network topology distance information and k-hop neighborhood matrix N k Generate hard constraint rules for structure learning and construct the topological dependency mask matrix M;

[0013] Step 3: Dynamic edge reduction based on electrical influence range: Calculate the electrical influence range of the fault point on surrounding nodes to obtain the electrical influence matrix E composed of electrical influence coefficients; In the topology dependency mask matrix M, if the electrical influence coefficient of a node on another node is less than a set value, the corresponding dependency edge is deleted.

[0014] Step 4: Dynamic edge weight adjustment based on data source reliability: Introduce the data source reliability matrix R, and perform hard deletion or soft weakening on edges with reliability below the threshold;

[0015] Step 5: Construct a sparse temporal Bayesian network structure: Combine the topological dependency mask matrix M, the electrical influence matrix E, and the reliability matrix R to synthesize a sparse structure matrix S; use matrix S as the spatial skeleton, add time-dimensional autoregressive edges, and construct a complete sparse temporal Bayesian network.

[0016] Furthermore, the construction of the distribution network topology and extraction of the adjacency matrix in step 1 are as follows:

[0017] Step 1.1: Extract equipment information from the geographic information system (GIS), monitoring and data acquisition system (SCADA), or master station system of the power distribution network. The equipment information includes lines, sections, switches, and protection devices. Construct a power distribution network topology map.

[0018] Step 1.2: Calculate the adjacency matrix A between each node. In the distribution network topology diagram, node i is 1 when there is a direct electrical connection between node i and node j, and 0 when there is no direct electrical connection. A(i,j) is an element in the adjacency matrix A, i.e.:

[0019] A(i,j)=1: Node i is directly connected to node j;

[0020] A(i,j) = 0: Node i and node j are not connected;

[0021] Step 1.3: Calculate the k-hop neighborhood matrix to define the dependency range. The k-hop neighborhood matrix indicates whether there exists a path of length k edges between nodes, i.e., a path passing through k-1 intermediate nodes. The calculation process is as follows:

[0022] First, obtain the adjacency matrix A from the distribution network topology diagram. Then, set the required hop count k. Next, initialize the neighborhood matrix N=A. Perform Boolean matrix multiplication on the initialized neighborhood matrix N to obtain a new matrix. Repeat this Boolean matrix multiplication iteratively to update the new neighborhood matrix k-1 times. Finally, ensure that the diagonal of N is 0 to obtain the k-hop neighborhood matrix N. k .

[0023] Furthermore, the construction of the distribution network topology map described in step 1.1 adopts dynamic topology tracking based on time-series event analysis and switch change-driven methods. Specifically, the topology is regarded as a dynamic graph that evolves over time. The focus is on switch change events that cause topology changes. A time-series log with switch events as the core is established. When a switch change or a group of switch changes is monitored, the system does not perform a full network search again. Instead, it updates the topology locally based on the known topology of the previous moment and applies graph transformation rules.

[0024] Furthermore, the k-hop neighborhood matrix described in step 1.3 only checks the protection and measurement devices within k hops of the fault point when a fault occurs, and each local controller only needs to interact with matrix N. k Instructions to exchange information with neighbors;

[0025] The defined dependency range specifically refers to the topological distance boundary in a distribution network where a state change or event of one node has a significant impact on other nodes. For a given node i and an integer k, the k-hop dependency range is all nodes that satisfy N k The set of nodes j where (i,j)=1.

[0026] Furthermore, the parent node constraint based on the topological radius described in step 2 is as follows:

[0027] Step 2.1: Based on the distribution network topology distance information and the k-hop neighborhood matrix N k The hard constraint rules for generating structure learning stipulate that:

[0028] If the topological distance dist(X, Y) between node X and node Y is greater than k, then Y is prohibited from becoming the parent node of X.

[0029] Step 2.2: Construct the topological dependency mask matrix M:

[0030] If the topological distance dist(X, Y) between node X and node Y is ≤ k, then M(X, Y) = 1;

[0031] Otherwise, M(X, Y) = 0;

[0032] During the structure learning or manual design phase of Temporal Bayesian Networks (TBNs), all edges with M(X,Y)=0 are prohibited from being introduced into the network structure.

[0033] Furthermore, the dynamic edge reduction based on the electrical influence range described in step 3 is as follows:

[0034] Step 3.1: Using electrical theories including line impedance, short-circuit sensitivity, and voltage drop, calculate the electrical influence range of the fault point on surrounding nodes, and obtain the electrical influence matrix E composed of electrical influence coefficients E(i,j).

[0035] Step 3.2: In the topological dependency mask matrix M, if the influence coefficient of a node on another node is less than ε, then delete the corresponding dependency edge, where ε is a judgment threshold determined based on the analysis accuracy requirements, data quality, and industry standard factors.

[0036] Furthermore, the dynamic edge weight adjustment based on data source reliability described in step 4 is as follows:

[0037] A data source reliability matrix R is introduced, including remote signaling reliability RYX, protection action reliability RProt, FA reliability RFA, and current and voltage sampling reliability RMeas;

[0038] If the reliability of a data source is lower than the threshold, the reliability of the specific data source type on the specific evaluation object will be dynamically reduced, or the corresponding parent node edge will be directly deleted.

[0039] Furthermore, the dynamic edge weight adjustment based on data source reliability is implemented through an engineering process and quantitative rules, as follows:

[0040] 1) Engineering quantification of reliability matrix: Based on historical defect elimination records and SOE events, specific reliability values ​​are assigned to multiple data sources, including remote signaling, FA terminals, protection actions, and electrical measurements. Through a combination of long-term statistics and event-level evaluation, the data source reliability matrix R is transformed into engineering parameters that can be actually obtained.

[0041] 2) Specific implementation of dynamic adjustment rules: The reliability threshold is refined into hierarchical operation rules of reduction threshold and weakening threshold. Three processing scenarios are defined: hard reduction, soft reduction and complete retention. Hard reduction is the deletion of edges, and soft reduction is the linear weakening of evidence weights, so as to realize the dynamic sparse adjustment of the network structure.

[0042] Furthermore, the construction of the sparse temporal Bayesian network structure described in step 5 is as follows:

[0043] By combining the topology dependency mask matrix M, the electrical influence matrix E, and the reliability matrix R, a sparse structure matrix S is synthesized:

[0044] S=M∩E∩R

[0045] Temporal Bayesian Networks (TBNs) add autoregressive dependencies in the time dimension to S, becoming a complete model capable of handling dynamic sequences. S is obtained through learning, nodes are copied on the time axis, intra-layer edges are added using S, inter-layer edges are added using autoregressive rules, and finally, conditional probability parameters are learned to obtain the complete sparse temporal Bayesian network.

[0046] A system for constructing sparse temporal Bayesian networks based on distribution network topology constraints is provided. This system implements the aforementioned method for constructing sparse temporal Bayesian networks based on distribution network topology constraints. The system includes:

[0047] Distribution network topology graph construction module: Constructs the distribution network topology graph, calculates the adjacency matrix between each node and the k-hop neighborhood matrix N. k ;

[0048] Parent node constraint module based on topology radius: based on distribution network topology distance information and k-hop neighborhood matrix N k Generate hard constraint rules for structure learning and construct the topological dependency mask matrix M;

[0049] The dynamic edge reduction module based on electrical influence range calculates the electrical influence range of the fault point on surrounding nodes, and obtains the electrical influence matrix E composed of electrical influence coefficients; in the topology dependency mask matrix M, if the electrical influence coefficient of a node on another node is less than a set value, the corresponding dependency edge is deleted.

[0050] The dynamic edge weight adjustment module based on data source reliability introduces a data source reliability matrix R and performs hard deletion or soft weakening on edges with reliability below a threshold.

[0051] Sparse temporal Bayesian network structure construction module: Combine the topological dependency mask matrix M, the electrical influence matrix E, and the reliability matrix R to synthesize the sparse structure matrix S; use matrix S as the spatial skeleton, add time-dimensional autoregressive edges, and construct a complete sparse temporal Bayesian network.

[0052] The significant advantages of this invention compared to existing technologies are:

[0053] (1) The number of network edges is significantly reduced: Topological constraints directly limit the dependency range to a local region, reducing the structural complexity of TBN from O(N) to O(N). 2 The value drops to O(N).

[0054] (2) Improved training efficiency and stability: The sparse structure reduces the CPT parameters, which solves the problem of training failure or overfitting when samples are scarce;

[0055] (3) Significantly improved inference speed: The inference overhead of sparse TBN is reduced, which can support engineering applications at the second level or even the sub-second level;

[0056] (4) Enhance the interpretability and engineering significance of the model: All edges are subject to topological, physical, and electrical logic constraints, avoiding "inference relationships without physical meaning";

[0057] (5) Adapt to real distribution network environments with multiple noises and anomalies: Dynamically reduces the edges of unreliable data sources, improving the robustness and reliability of fault diagnosis;

[0058] (6) Supports large-scale distribution network modeling: It can build usable TBNs at the scale of hundreds or thousands of nodes, which is beneficial for deployment in control centers or distribution automation platforms. Attached Figure Description

[0059] Figure 1 This is a flowchart of the sparse temporal Bayesian network construction method based on distribution network topology constraints of the present invention. Detailed Implementation

[0060] Combination Figure 1 This invention provides a method for constructing sparse temporal Bayesian networks based on distribution network topology constraints, comprising the following steps:

[0061] Step 1: Construct the distribution network topology and extract the adjacency matrix: Construct the distribution network topology, calculate the adjacency matrix between each node and the k-hop neighborhood matrix N. k ;

[0062] Step 2, Parent node constraints based on topology radius: Based on distribution network topology distance information and k-hop neighborhood matrix N k Generate hard constraint rules for structure learning and construct the topological dependency mask matrix M;

[0063] Step 3: Dynamic edge reduction based on electrical influence range: Calculate the electrical influence range of the fault point on surrounding nodes to obtain the electrical influence matrix E composed of electrical influence coefficients; In the topology dependency mask matrix M, if the electrical influence coefficient of a node on another node is less than a set value, the corresponding dependency edge is deleted.

[0064] Step 4: Dynamic edge weight adjustment based on data source reliability: Introduce the data source reliability matrix R, and perform hard deletion or soft weakening on edges with reliability below the threshold;

[0065] Step 5: Construct a sparse temporal Bayesian network structure: Combine the topological dependency mask matrix M, the electrical influence matrix E, and the reliability matrix R to synthesize a sparse structure matrix S; use matrix S as the spatial skeleton, add time-dimensional autoregressive edges, and construct a complete sparse temporal Bayesian network.

[0066] As a specific example, step 1, which involves constructing the distribution network topology and extracting the adjacency matrix, is as follows:

[0067] Step 1.1: Extract equipment information such as lines, sections, switches, and protection devices from the distribution network GIS (Geographic Information System), SCADA (Supervisory and Data Acquisition System), or master station system to construct a distribution network topology map;

[0068] The distribution network topology map constructed here employs dynamic topology tracking based on time-series event analysis and switch change-driven methods. The specific approach is as follows: the topology is viewed as a dynamic graph evolving over time, with a focus on switch change events that cause topology changes. A time-series log centered on these switch events is established. When a switch change or a group of switch changes is detected, the system does not perform a full network search again. Instead, based on the known topology from the previous moment, it applies graph transformation rules to quickly and locally update the topology. For example, if a typical "fault isolation-load transfer" operation sequence is identified, the affected load is automatically switched from the original feeder to a new feeder. This method offers fast response and low computational overhead, making it highly suitable for applications with high real-time requirements.

[0069] Step 1.2: Calculate the adjacency matrix A between nodes. In the distribution network topology diagram, node i is 1 when there is a direct electrical connection between node i and node j, and 0 when there is no direct electrical connection. For example, A(i,j) is an element in the adjacency matrix A, i.e.:

[0070] A(i,j)=1: Node i is directly connected to node j;

[0071] A(i,j) = 0: Node i and node j are not connected;

[0072] Step 1.3: Calculate the k-hop neighborhood matrix to define the dependency range. The k-hop neighborhood matrix indicates whether there exists a path between nodes with a length of exactly k edges (passing through k-1 intermediate nodes). The calculation process is as follows:

[0073] First, obtain the adjacency matrix A from the distribution network topology. Then, set the required hop count k. Next, initialize the neighborhood matrix N=A. Perform Boolean matrix multiplication on the initialized neighborhood matrix N to obtain a new matrix. Repeat this Boolean matrix multiplication iteratively k-1 times. Finally, ensure that the diagonal of N is 0 to obtain the k-hop neighborhood matrix N. k .

[0074] k-hop neighborhood matrix N kIts function is to: when a fault occurs, only check the protection and measuring devices within k jumps of the fault point to narrow down the fault search range; at the same time, each local controller only needs to communicate with matrix N. k By instructing neighbors to exchange information, the communication burden is reduced, and local coordination and global optimization are achieved, thus determining the communication needs between controllers.

[0075] The defined dependency range specifically refers to the topological distance boundary in a distribution network where a state change or event of one node has a significant impact on other nodes. For example, given node i and integer k, its k-hop dependency range is all nodes that satisfy N k The set of nodes j where (i,j)=1.

[0076] As a specific example, the parent node constraint based on topological radius (topological sparsification) described in step 2 is as follows:

[0077] Step 2.1: Based on the distribution network topology distance information and the k-hop neighborhood matrix N k The hard constraint rules for generating structure learning stipulate that:

[0078] If the topological distance dist(X, Y) between node X and node Y is greater than k (in engineering, k=2 or 3 is usually sufficient), then Y is prohibited from becoming the parent node of X.

[0079] Step 2.2: Construct the topological dependency mask matrix M:

[0080] If the topological distance dist(X, Y) between node X and node Y is ≤ k, then M(X, Y) = 1;

[0081] Otherwise, M(X, Y) = 0;

[0082] In the structure learning or manual design phase of TBN (Temporal Bayesian Network), all edges with M(X,Y)=0 are prohibited from being introduced into the network structure. This method achieves automatic sparsification of the global structure.

[0083] As a specific example, the dynamic edge reduction based on electrical influence range described in step 3, based on the sparse candidate edge set defined by the topological dependency mask matrix M output in step 2, further introduces electrical physical characteristics for refined screening, as follows:

[0084] Step 3.1: Using electrical theories including line impedance, short-circuit sensitivity, and voltage sag, calculate the electrical influence range of the fault point on surrounding nodes, obtaining the electrical influence matrix E composed of electrical influence coefficients E(i,j). For example, precise calculations can be performed based on the node impedance matrix. First, construct and solve the node voltage under normal operating conditions, calculate the additional fault conditions, and then calculate the node voltage after the fault. In this way, the influence index E(i,j) can be defined and calculated, and finally, the influence matrix E can be formed.

[0085] Step 3.2: In the topological dependency mask matrix M, if the influence coefficient of a node on another node is less than ε (e.g., current change <5%), then delete the corresponding dependency edge, where ε is a judgment threshold determined based on analysis accuracy requirements, data quality, and industry standard factors.

[0086] The electrical influence matrix E can be used to further reduce the number of edges in BN / TBN (Bayesian Network / Temporal Bayesian Network), making the network strictly modeled around the electrical physical "locality".

[0087] As a specific example, the dynamic edge weight adjustment based on data source reliability described in step 4, following the structural constraints completed in steps 2 and 3, introduces a data quality dimension for final optimization and robustness enhancement of the candidate network structure determined after dual screening by topology and electrical properties, as detailed below:

[0088] A data source reliability matrix R is introduced, including remote signaling reliability RYX, protection action reliability RProt, FA reliability RFA, and current and voltage sampling reliability RMeas;

[0089] If the reliability (long-term reliability or current event reliability) of a certain data source is lower than the threshold, where reliability refers to the reliability evaluation value of a specific data source, which may be any relevant element in the data source reliability matrix R, then the reliability of the specific data source type (such as RYX) on the specific evaluation object (such as a certain switch event) will be dynamically reduced, or the corresponding parent node edge will be directly deleted.

[0090] This allows TBN to operate reliably even in scenarios of "remote communication anomalies / unstable communication".

[0091] The dynamic edge weight adjustment based on data source reliability proposed in this step is implemented through a complete engineering process and quantitative rules, with the specific correspondence as follows:

[0092] 1) Engineering quantification of reliability matrix: Based on historical defect elimination records and SOE events, specific reliability values ​​(such as RYX=0.92, RFA=0.75) are assigned to various data sources such as remote signaling, FA terminals, protection actions, and electrical measurements. By defining a calculation process that combines long-term statistics and event-level evaluation, the abstract reliability matrix R is transformed into practically obtainable engineering parameters.

[0093] 2) Specific implementation of dynamic adjustment rules: The principle of "adjusting edge weights according to reliability threshold" is refined into hierarchical operation rules with explicit thresholds (such as reduction threshold of 0.6 and weakening threshold of 0.8). Specifically, three processing scenarios, namely "hard reduction" (deleting edges), "soft reduction" (linearly weakening evidence weights) and "complete retention", and their mathematical expressions are defined, thus fully realizing the dynamic and fine-grained sparse adjustment of the network structure.

[0094] 3) Quantitative verification of sparsity effect: The number of network edges was further reduced from about 6,000 to about 3,500. This result directly verifies that this step can significantly improve network sparsity based on topological and electrical constraints, and enhance the diagnostic robustness of the model in data anomaly scenarios, thus completing the closed-loop construction from physical structure to reliable engineering model.

[0095] As a specific example, the construction of the sparse temporal Bayesian network structure in step 5 is as follows:

[0096] By combining the topology dependency mask matrix M, the electrical influence matrix E, and the reliability matrix R, a sparse structure matrix S is synthesized:

[0097] S=M∩E∩R

[0098] The final structure of the Temporal Bayesian Network (TBN) is determined by matrix S, ensuring: sparse network structure, controlled CPT size, significantly reduced inference complexity, clear physical meaning, and strong interpretability. The sparse structure matrix S is the mathematical embodiment of the "spatial components" or "intra-layer structure" of the TBN. Based on S, the TBN adds autoregressive dependencies in the time dimension, thus becoming a complete model capable of handling dynamic sequences. By learning S, replicating nodes on the time axis, adding intra-layer edges using S, adding inter-layer edges using autoregressive rules, and finally learning the conditional probability parameters, the complete sparse temporal Bayesian Network can be obtained.

[0099] This invention also provides a sparse temporal Bayesian network construction system based on distribution network topology constraints. This system is used to implement the aforementioned sparse temporal Bayesian network construction method based on distribution network topology constraints. The system includes:

[0100] Distribution network topology graph construction module: Constructs the distribution network topology graph, calculates the adjacency matrix between each node and the k-hop neighborhood matrix N. k ;

[0101] Parent node constraint module based on topology radius: based on distribution network topology distance information and k-hop neighborhood matrix N k Generate hard constraint rules for structure learning and construct the topological dependency mask matrix M;

[0102] The dynamic edge reduction module based on electrical influence range calculates the electrical influence range of the fault point on surrounding nodes, and obtains the electrical influence matrix E composed of electrical influence coefficients; in the topology dependency mask matrix M, if the electrical influence coefficient of a node on another node is less than a set value, the corresponding dependency edge is deleted.

[0103] The dynamic edge weight adjustment module based on data source reliability introduces a data source reliability matrix R and performs hard deletion or soft weakening on edges with reliability below a threshold.

[0104] Sparse temporal Bayesian network structure construction module: Combine the topological dependency mask matrix M, the electrical influence matrix E, and the reliability matrix R to synthesize the sparse structure matrix S; use matrix S as the spatial skeleton, add time-dimensional autoregressive edges, and construct a complete sparse temporal Bayesian network.

[0105] The present invention will be further described in detail below with reference to specific embodiments.

[0106] Example

[0107] Taking a 10kV distribution network in a certain city as an example, this area includes 28 feeders, 760 segment nodes, 540 switches (including FTU / TTU), 220 FA terminals, and a total of over 1500 nodes. We will now verify the construction of a sparse TBN on three typical feeders.

[0108] (1) Results of topological constraint construction

[0109] Based on the GIS and main station topology, an adjacency matrix A (1500×1500) is generated, and topological distances are calculated. Setting k=2 (considering only two-hop neighborhoods), the resulting subgraph size is approximately 10~15. The number of dependent edges is reduced from approximately 1500²≈2.25 million original possible edges to approximately 1500×(10~15)=15,000~23,000, with an average of approximately 18,000 edges, increasing the sparsity to 99.2%.

[0110] (2) Sparsification of electrical effects

[0111] There are usually only 2 to 5 nodes that are directly electrically related to each node in the two-hop neighborhood. By deleting the topological adjacent edges with low impact coefficients (which account for about 70% of the edges retained in step 1), the number of edges is reduced to about 6,000.

[0112] (3) Further reduce TBN connections by leveraging data source reliability.

[0113] Using one year's worth of historical defect elimination records and SOE, the reliability of each data source type was calculated, assuming: remote signaling reliability RYX = 0.92, FA reliability RFA = 0.75 (some FTUs frequently switch), protection action reliability RProt = 0.99, and current and voltage sampling reliability RMeas = 0.95. Edges with reliability below 0.8 were automatically reduced, resulting in approximately 3500 edges in the sparse TBN. Detailed calculation process:

[0114] 3.1) Obtaining data source reliability

[0115] For each type of data source in the distribution network, its historical performance is statistically analyzed according to its channel (such as the remote signaling of the open / close position of the j-th switch, the status of the m-th FA terminal, and the current measurement value of the l-th measuring point) to obtain the long-term reliability. .

[0116] I. Extracting from runtime history Each event log (including actual faults, normal operations, planned maintenance, daily status changes, etc.) is recorded for a specific channel. statistics:

[0117] This channel should have a correct number of times it performs.

[0118] The number of times the data matches the data from the main geological survey station / reliable source during the same period;

[0119] The number of false alarms, false card flips, and false triggers;

[0120] Number of times that should have been submitted but were not.

[0121] II. Define long-term reliability based on statistical results:

[0122]

[0123] in The penalty weight can be 1 or other empirical values.

[0124] III. In the current diagnostic event, calculate the event-level reliability by combining the results of multi-source cross-verification, including voltage-switch consistency, protection-topology consistency, FA flip-flop consistency, and short-circuit current distribution consistency. :

[0125]

[0126] IV. Integrating long-term statistics with event performance yields effective reliability:

[0127]

[0128] in This indicates the dominant role of long-term statistics in reliability assessment.

[0129] 3.2) Reliability-based edge reduction rules

[0130] Set two thresholds:

[0131] Cut edge threshold (e.g., 0.6)

[0132] weaken threshold (e.g., 0.8)

[0133] Perform the following operation on all dependent edges in the temporal Bayesian network that involve the data source node (X^{k}):

[0134] I. Hard cut (deletion)

[0135] like Delete all Dependency edges that are the start or end points, for example:

[0136]

[0137] Once this edge is deleted, subsequent inference will no longer depend on the information from that channel.

[0138] II. Soft Reduction (Weakening of Weights)

[0139] like : Do not delete edges, but weaken their influence in the conditional probability table, for example:

[0140]

[0141] That is, the strength of evidence decreases linearly with reliability, thus achieving "soft sparsity".

[0142] III. Complete Preservation

[0143] like The Bayesian connection relationship for this node remains unchanged.

[0144] The above steps are performed independently for all data source channels.

[0145] 3.3) Edge reduction quantification and result statistics

[0146] After topological constraints (two-hop neighborhood) and electrical sensitivity screening (e.g., a short-circuit current influence threshold of 0.05), the system retains the following number of candidate dependency edges: (Approximately 6000 under the actual measurement conditions of this project).

[0147] After performing reliability edge reduction, statistics are collected for each node. Number of edges deleted:

[0148]

[0149] in This represents the degree of the node in the current sparse structure.

[0150] Finally, we obtain the total number of edges after edge reduction:

[0151]

[0152] The reduction ratio is:

[0153]

[0154] In this embodiment, the FA signal is related to the remote signaling of some switches. Within the range of 0.4 to 0.6, hard pruning is triggered; the average degree of the relevant nodes is relatively high, so a large number of edges are deleted. Ultimately, after reliability-based edge pruning, the number of edges is reduced from approximately 6,000 to approximately 3,500, a reduction ratio of approximately 41.7%, further enhancing the sparsity and robustness of the network model.

[0155] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing sparse temporal Bayesian networks based on distribution network topology constraints, characterized in that, Includes the following steps: Step 1: Construct the distribution network topology and extract the adjacency matrix: Construct the distribution network topology, calculate the adjacency matrix between each node and the k-hop neighborhood matrix N. k ; Step 2, Parent node constraints based on topology radius: Based on distribution network topology distance information and k-hop neighborhood matrix N k Generate hard constraint rules for structure learning and construct the topological dependency mask matrix M; Step 3: Dynamic edge reduction based on electrical influence range: Calculate the electrical influence range of the fault point on surrounding nodes to obtain the electrical influence matrix E composed of electrical influence coefficients; In the topology dependency mask matrix M, if the electrical influence coefficient of a node on another node is less than a set value, the corresponding dependency edge is deleted. Step 4: Dynamic edge weight adjustment based on data source reliability: Introduce the data source reliability matrix R, and perform hard deletion or soft weakening on edges with reliability below the threshold; Step 5: Construct a sparse temporal Bayesian network structure: Combine the topological dependency mask matrix M, the electrical influence matrix E, and the reliability matrix R to synthesize a sparse structure matrix S; use matrix S as the spatial skeleton, add time-dimensional autoregressive edges, and construct a complete sparse temporal Bayesian network.

2. The method for constructing sparse temporal Bayesian networks based on distribution network topology constraints according to claim 1, characterized in that, Step 1, which involves constructing the distribution network topology and extracting the adjacency matrix, is detailed below: Step 1.1: Extract equipment information from the geographic information system (GIS), monitoring and data acquisition system (SCADA), or master station system of the power distribution network. The equipment information includes lines, sections, switches, and protection devices. Construct a power distribution network topology map. Step 1.2: Calculate the adjacency matrix A between each node. In the distribution network topology diagram, node i is 1 when there is a direct electrical connection between node i and node j, and 0 when there is no direct electrical connection. A(i,j) is an element in the adjacency matrix A, i.e.: A(i,j)=1: Node i is directly connected to node j; A(i,j) = 0: Node i and node j are not connected; Step 1.3: Calculate the k-hop neighborhood matrix to define the dependency range. The k-hop neighborhood matrix indicates whether there exists a path of length k edges between nodes, i.e., a path passing through k-1 intermediate nodes. The calculation process is as follows: First, obtain the adjacency matrix A from the distribution network topology diagram. Then, set the required hop count k. Next, initialize the neighborhood matrix N=A. Perform Boolean matrix multiplication on the initialized neighborhood matrix N to obtain a new matrix. Repeat this Boolean matrix multiplication iteratively to update the new neighborhood matrix k-1 times. Finally, ensure that the diagonal of N is 0 to obtain the k-hop neighborhood matrix N. k .

3. The method for constructing sparse temporal Bayesian networks based on distribution network topology constraints according to claim 2, characterized in that, The construction of the distribution network topology map described in step 1.1 adopts dynamic topology tracking based on time-series event analysis and switch change-driven approach. Specifically, the topology is regarded as a dynamic graph that evolves over time. The focus is on switch change events that cause topology changes. A time-series log with switch events as the core is established. When a switch change or a group of switch changes is monitored, the system does not perform a full network search again. Instead, it updates the topology locally based on the known topology of the previous moment and applies graph transformation rules.

4. The method for constructing sparse temporal Bayesian networks based on distribution network topology constraints according to claim 2, characterized in that, The k-hop neighborhood matrix described in step 1.3 only checks the protection and measurement devices within k hops of the fault point when a fault occurs. Simultaneously, each local controller only needs to interact with matrix N. k Instructions to exchange information with neighbors; The defined dependency range specifically refers to the topological distance boundary in a distribution network where a state change or event of one node has a significant impact on other nodes. For a given node i and an integer k, the k-hop dependency range is all nodes that satisfy N k The set of nodes j where (i,j)=1.

5. The method for constructing sparse temporal Bayesian networks based on distribution network topology constraints according to claim 1, characterized in that, The parent node constraint based on the topological radius described in step 2 is as follows: Step 2.1: Based on the distribution network topology distance information and the k-hop neighborhood matrix N k The hard constraint rules for generating structure learning stipulate that: If the topological distance dist(X, Y) between node X and node Y is greater than k, then Y is prohibited from becoming the parent node of X. Step 2.2: Construct the topological dependency mask matrix M: If the topological distance dist(X, Y) between node X and node Y is ≤ k, then M(X, Y) = 1; Otherwise, M(X, Y) = 0; During the structure learning or manual design phase of Temporal Bayesian Networks (TBNs), all edges with M(X,Y)=0 are prohibited from being introduced into the network structure.

6. The method for constructing sparse temporal Bayesian networks based on distribution network topology constraints according to claim 1, characterized in that, The dynamic edge reduction based on the electrical influence range described in step 3 is as follows: Step 3.1: Using electrical theories including line impedance, short-circuit sensitivity, and voltage drop, calculate the electrical influence range of the fault point on surrounding nodes, and obtain the electrical influence matrix E composed of electrical influence coefficients E(i,j). Step 3.2: In the topological dependency mask matrix M, if the influence coefficient of a node on another node is less than ε, then delete the corresponding dependency edge, where ε is a judgment threshold determined based on the analysis accuracy requirements, data quality, and industry standard factors.

7. The method for constructing sparse temporal Bayesian networks based on distribution network topology constraints according to claim 1, characterized in that, The dynamic edge weight adjustment based on data source reliability described in step 4 is as follows: A data source reliability matrix R is introduced, including remote signaling reliability RYX, protection action reliability RProt, FA reliability RFA, and current and voltage sampling reliability RMeas; If the reliability of a data source is lower than the threshold, the reliability of the specific data source type on the specific evaluation object will be dynamically reduced, or the corresponding parent node edge will be directly deleted.

8. The method for constructing sparse temporal Bayesian networks based on distribution network topology constraints according to claim 7, characterized in that, The dynamic edge weight adjustment based on data source reliability is implemented through an engineering process and quantitative rules, as follows: 1) Engineering quantification of reliability matrix: Based on historical defect elimination records and SOE events, specific reliability values ​​are assigned to multiple data sources, including remote signaling, FA terminals, protection actions, and electrical measurements. Through a combination of long-term statistics and event-level evaluation, the data source reliability matrix R is transformed into engineering parameters that can be actually obtained. 2) Specific implementation of dynamic adjustment rules: The reliability threshold is refined into hierarchical operation rules of reduction threshold and weakening threshold. Three processing scenarios are defined: hard reduction, soft reduction and complete retention. Hard reduction is the deletion of edges, and soft reduction is the linear weakening of evidence weights, so as to realize the dynamic sparse adjustment of the network structure.

9. The method for constructing a sparse temporal Bayesian network based on distribution network topology constraints according to claim 1, characterized in that, Step 5 describes the construction of a sparse temporal Bayesian network structure, as follows: By combining the topology dependency mask matrix M, the electrical influence matrix E, and the reliability matrix R, a sparse structure matrix S is synthesized: S=M∩E∩R Temporal Bayesian Networks (TBNs) add autoregressive dependencies in the time dimension to S, becoming a complete model capable of handling dynamic sequences. S is obtained through learning, nodes are copied on the time axis, intra-layer edges are added using S, inter-layer edges are added using autoregressive rules, and finally, conditional probability parameters are learned to obtain the complete sparse temporal Bayesian network.

10. A sparse temporal Bayesian network construction system based on distribution network topology constraints, characterized in that, This system is used to implement the sparse temporal Bayesian network construction method based on distribution network topology constraints as described in any one of claims 1 to 9, the system comprising: Distribution network topology graph construction module: Constructs the distribution network topology graph, calculates the adjacency matrix between each node and the k-hop neighborhood matrix N. k ; Parent node constraint module based on topology radius: based on distribution network topology distance information and k-hop neighborhood matrix N k Generate hard constraint rules for structure learning and construct the topological dependency mask matrix M; The dynamic edge reduction module based on electrical influence range calculates the electrical influence range of the fault point on surrounding nodes, and obtains the electrical influence matrix E composed of electrical influence coefficients; in the topology dependency mask matrix M, if the electrical influence coefficient of a node on another node is less than a set value, the corresponding dependency edge is deleted. The dynamic edge weight adjustment module based on data source reliability introduces a data source reliability matrix R and performs hard deletion or soft weakening on edges with reliability below a threshold. Sparse temporal Bayesian network structure construction module: Combine the topological dependency mask matrix M, the electrical influence matrix E, and the reliability matrix R to synthesize the sparse structure matrix S; use matrix S as the spatial skeleton, add time-dimensional autoregressive edges, and construct a complete sparse temporal Bayesian network.

Images