Urban medium-low voltage distribution network power dispatching method

Abstract

The invention belongs to the technical field of power dispatching and control, and particularly discloses a city middle and low voltage power distribution network power dispatching method, which comprises the following steps of: acquiring current power dispatching procedure text data, analyzing violation operation description and boundary condition definition by adopting a natural language processing algorithm, extracting a keyword group and a logical relationship from the text data, and calculating the power dispatching procedure text data according to the keyword group and the logical relationship; obtaining a standard violation feature set; according to the standard violation feature set, knowledge graph representation is constructed, nodes correspond to violation types and operation modes, edges represent condition dependency relations, and a dynamic knowledge structure is determined; obtaining an operation instruction sequence from massive historical scheduling logs, performing matching query for the dynamic knowledge structure, and marking potential violation points if connection switches in the sequence are closed and do not conform to breaking conditions; the objective of the invention is to solve the problem that the existing power dispatching checking method is difficult to dynamically adapt to standard updating, cannot effectively recognize complex violation scenes, and is easy to cause missing detection and false alarm, so that the recognition accuracy and authority of a system are insufficient.

Description

Technical Field

[0001] This invention relates to the technical field of power dispatching and control, specifically to a power dispatching method for urban low-voltage distribution networks. Background Technology

[0002] Urban low- and medium-voltage distribution networks are a critical link in the power system facing end users. The accuracy and standardization of their dispatch operations directly affect power supply reliability and grid security. Therefore, rigorous verification of dispatch instructions and operational processes is essential. Currently, verification methods relying on manual or simple rule matching are insufficient to address the complex, dynamic, and hidden violation patterns in dispatch instructions. Their limitations lie in insufficient depth and adaptability of verification, failing to continuously uncover potential risk points that meet the latest professional standards from massive historical and real-time data.

[0003] This raises a core technical challenge in the field: the dynamic coordination between dispatch verification standards and identification capabilities. Power dispatching procedures and operational specifications are updated and revised as technology advances, requiring the criteria for error pattern identification to remain up-to-date. However, standards are dynamically changing textual knowledge, while the rule base upon which the identification system relies is often static and fixed, creating a gap in update lag and coverage. This contradiction directly leads to another closely related technical challenge: the lack of a continuously reliable benchmark for the verification and calibration of the identification system. Due to the lack of a mechanism for real-time benchmarking against current standards and systematically verifying the conformity of identification results, the system may miss violations defined under new standards and falsely report outdated rules, thus challenging the authority of its early warning judgments.

[0004] Specifically, in operational practice, every operational instruction issued by a dispatcher may contain multiple complex violation characteristics, such as mistakenly closing a tie switch that should have been disconnected under a specific operating mode. Ensuring that the identification system can not only detect these explicitly prohibited typical errors, but also effectively identify violations that have been added with the promulgation of new regulations or have more complex boundary conditions, and that its judgment capabilities are always in line with authoritative industry standards, has become a core and critical issue in ensuring the safety and lean management of urban power distribution network dispatch. Summary of the Invention

[0005] This invention provides a power dispatching method for urban medium and low voltage distribution networks, aiming to solve the problems of existing power dispatching verification methods, such as difficulty in dynamically adapting to standard updates, inability to effectively identify complex violation scenarios, and susceptibility to missed detections and false alarms, resulting in insufficient system identification accuracy and authority.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A power dispatching method for urban low-voltage distribution networks includes: collecting textual data of current power dispatching procedures, parsing the descriptions of violations and boundary condition definitions using natural language processing algorithms, extracting keyword phrases and logical relationships to obtain a standard violation feature set; constructing a knowledge graph representation based on the standard violation feature set, where nodes correspond to violation types and operating modes, and edges represent conditional dependencies, thus determining a dynamic knowledge structure; obtaining operation instruction sequences from massive historical dispatching logs, performing matching queries against the dynamic knowledge structure, and marking potential violation points if the closing of tie switches in the sequence does not meet the disconnection conditions, thereby obtaining a preliminary risk list; and employing knowledge... The graph traversal algorithm scans the surrounding nodes of each point in the initial risk list, integrates newly added violation attributes to expand the query scope, and determines whether it covers complex scenarios at the boundaries of the new standard, thus obtaining an expanded risk set. By comparing the expanded risk set with the real-time operation command input, a similarity score is calculated. If the score exceeds a preset threshold, an alarm mechanism is triggered to confirm the violation result. Instances of missed detections and false alarms are extracted from the violation confirmation result and fed back into the knowledge graph update process to adjust node weights and edge strengths, thereby obtaining an optimized knowledge structure. Based on the optimized knowledge structure, the real-time operation command is re-parsed. If composite violation features are detected, an authoritative identification report is output to determine the overall consistency of the verification.

[0008] In one aspect of the invention, the method involves collecting textual data of current power dispatching procedures, using natural language processing algorithms to parse the descriptions of violations and boundary condition definitions, extracting keyword phrases and logical relationships, and obtaining a standard violation feature set, including:

[0009] Obtain the text set of power dispatching regulations;

[0010] Natural language processing algorithms are used to parse the descriptions of violations and the definitions of boundary conditions in the text set;

[0011] Extract the keyword groups corresponding to the descriptions of the violations from the analysis results to obtain a set of violation predicate groups;

[0012] Extract the logical relationships corresponding to the boundary conditions from the parsing results to obtain the set of conditional relationship relationships;

[0013] By integrating the set of violating predicate phrases and the set of conditional semantic relations, a standard set of violation features is formed;

[0014] Based on the standard violation feature set, a violation operation identification model is constructed.

[0015] In one aspect of the invention, the step of constructing a knowledge graph representation based on a standard violation feature set, wherein nodes correspond to violation types and operating modes, edges represent conditional dependencies, and determining a dynamic knowledge structure includes:

[0016] Obtain the standard violation feature set;

[0017] Knowledge graphs are stored using graph database technology;

[0018] Generate a graph node set based on the violation type and operation mode definition in the feature set;

[0019] Generate a graph edge set based on the conditional dependencies defined in the feature set;

[0020] If there is a logical conflict between the node set and the edge set, the consistency check rule is invoked to correct it;

[0021] The verified node set and edge set are imported into the graph database to obtain a dynamic knowledge graph.

[0022] In one aspect of the invention, the step of obtaining the operation instruction sequence from massive historical scheduling logs, performing a matching query against a dynamic knowledge structure, and marking potential violations if the closing of a communication switch in the sequence does not meet the disconnection conditions, thereby obtaining a preliminary risk list, includes:

[0023] Obtain historical scheduling log data source;

[0024] Log data is processed using text parsing technology to extract structured operation instruction sequences;

[0025] Associate and map the switch operation events in the instruction sequence with nodes in the dynamic knowledge graph;

[0026] For each associated switch operation event, obtain its final state in the instruction sequence;

[0027] Query the preset disconnection conditions of the switch node under the current operating mode from the dynamic knowledge graph;

[0028] If the switch status in the instruction sequence is closed, and the preset disconnection condition in the diagram is true, then a condition conflict is determined to exist.

[0029] Switch operation events with conflicting conditions and their context information are recorded as risk entries, and a risk list is generated by summarizing them.

[0030] In one aspect of the invention, the step of employing a knowledge graph traversal algorithm to scan the surrounding nodes of each point in the initial risk list, integrating newly added violation attributes to expand the query scope, determining whether it covers complex scenarios at the boundary of the new standard, and obtaining an expanded risk set includes:

[0031] Obtain each risk entry from the preliminary risk list;

[0032] For each risk item, obtain its corresponding switch node from the knowledge graph;

[0033] Starting from the switch node, a breadth-first search algorithm is used to traverse the graph and obtain the set of topological nodes directly connected to it.

[0034] According to the preset violation judgment rules, the status of each node in the topology node set is checked;

[0035] If a node's status meets the violation determination rules, then the node is marked as a derivative risk node;

[0036] Summarize all derived risk nodes corresponding to risk entries, remove duplicate records, and generate an extended risk set.

[0037] In one aspect of the invention, the step of calculating a similarity score by comparing an extended risk set with real-time operation command input, and triggering an alarm mechanism if the score exceeds a preset threshold to determine the violation confirmation result, includes:

[0038] Obtain the real-time operation command stream and extended risk set;

[0039] A text vectorization model is used to process each risk description in the operation instruction stream and risk set;

[0040] Calculate the cosine similarity between the operation instruction vector and each risk description vector;

[0041] The maximum value among all similarity calculation results is selected as the final similarity score;

[0042] If the final similarity score is greater than the preset threshold, an alarm will be triggered.

[0043] After the alarm is triggered, the acknowledgment device outputs the result confirming the violation.

[0044] In one aspect of the invention, the step of extracting missed and false alarm instances from the violation confirmation results and feeding them back into the knowledge graph update process to adjust node weights and edge strengths to obtain an optimized knowledge structure includes:

[0045] Obtain the set of missed detection instances and the set of false alarm instances from the violation confirmation results;

[0046] Merge the sets of missed detection instances and false alarm instances to generate a set of feedback instances;

[0047] Locate the graph node associated with each instance within the set of feedback instances from the knowledge graph;

[0048] If an instance belongs to the set of missed instances, increase the weight value of the node associated with that instance.

[0049] If an instance belongs to the set of false positive instances, then the connection strength of the edges between the associated nodes of that instance and its neighboring nodes is reduced.

[0050] A graph embedding model was used to process and update the knowledge graph with weights and edge strengths, generating new node vector representations.

[0051] Based on the new node vector representation, the similarity between nodes is recalculated to form an optimized knowledge structure.

[0052] In one aspect of the invention, the step of re-analyzing real-time operation instructions based on optimized knowledge structure, and outputting an authoritative identification report if composite violation characteristics are detected, to determine overall verification consistency, includes:

[0053] Relevant data on real-time instructions are obtained from the knowledge structure, and the instruction content is initially broken down to obtain the results of operation analysis.

[0054] Based on the results of the operation analysis, the violation features in the knowledge structure are matched. If a pattern that matches the composite features is detected, it is determined that there is a potential violation.

[0055] For identified potential violations, a pre-established feature matching mechanism is used to analyze the specific manifestations of composite features and obtain detailed records of feature matching.

[0056] By using detailed records of feature matching, a corresponding recognition report is generated, and the credibility level of the report is determined by combining authoritative evaluation standards.

[0057] Based on the credibility level of the identification report, analyze the relevant indicators of consistency verification to obtain the basis data for the overall assessment;

[0058] Based on the overall assessment data, a logistic regression model is used to process the various indicators that are consistent with the verification, and the final compliance conclusion is determined.

[0059] Based on the compliance conclusions, organize the final content of the identification report and generate a structured output document.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. Achieved dynamic adaptive updating of verification basis: Automatically parses the latest power dispatching regulations text through natural language processing technology, extracts standard violation features and constructs a dynamic knowledge graph, fundamentally solving the technical problem that traditional systems based on static rule bases cannot adapt to the dynamic revision of standard texts, and ensuring the timeliness and authority of the identification benchmark.

[0062] 2. Improved ability to identify complex and hidden violation scenarios: By using knowledge graphs to represent violation types, operating methods and conditional dependencies in a networked manner, and combining them with graph traversal algorithms such as breadth-first search, it can systematically discover complex boundary violation scenarios and composite violation features composed of multiple devices and multiple conditions coupled together, overcoming the problem of high false negative rate of simple rule matching.

[0063] 3. A closed-loop optimization mechanism has been formed to continuously improve the accuracy of identification: By matching real-time operation commands with the extended risk set for similarity and issuing alarms, and feeding back confirmed missed detections and false alarms to the knowledge graph, the node weights and edge strengths are dynamically adjusted, enabling the system to have online learning and optimization capabilities, effectively reducing the false alarm rate and missed detection rate.

[0064] 4. The verification process has been automated and the report has been structured: The entire process from instruction parsing, risk retrieval, intelligent alarm to the generation of authoritative identification reports has been automated, and the structured report containing risk details, confidence level and basis has been output, which greatly improves the efficiency, standardization and traceability of the scheduling and verification work. Attached Figure Description

[0065] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0066] Figure 1 This is one of the flowcharts for a power dispatching method for urban low-voltage distribution networks according to the present invention.

[0067] Figure 2 This is the second flowchart of a power dispatching method for urban low-voltage distribution networks according to the present invention.

[0068] Figure 3 This is the third flowchart of a power dispatching method for urban low-voltage distribution networks according to the present invention. Detailed Implementation

[0069] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0070] Please see Figures 1-3 As shown in the figure, this embodiment discloses a power dispatching method for urban low-voltage distribution networks, which may specifically include:

[0071] S101. By collecting text data of current power dispatching regulations, natural language processing algorithms are used to parse the descriptions of violations and the definitions of boundary conditions, extracting key phrases and logical relationships to obtain a standard violation feature set.

[0072] Obtain the text set of power dispatching regulations. Use a BERT-based pre-trained model for text semantic parsing, leveraging its multi-head attention mechanism to extract keyword phrases indicating violations, obtaining a set of violation predicates. Extract the logical relationships corresponding to boundary condition definitions from the parsing results, obtaining a set of conditional semantic relations. Merge the set of violation predicates and the set of conditional semantic relations to form a standard violation feature set. Construct a violation operation identification model based on the standard violation feature set.

[0073] In this step, the BERT (Bidirectional Encoder Representations from Transformers) pre-trained model based on the Transformer architecture is chosen for semantic parsing. The main reason is that violation descriptions in power dispatching regulations often contain complex conditional clauses, passive voice, and negative forms of technical terms. The BERT model, through its bidirectional attention mechanism, can deeply integrate the contextual semantic information of words, achieving higher accuracy in understanding structured semantics such as "under what conditions is what operation prohibited" compared to traditional bag-of-words models or recurrent neural networks (RNNs). By fine-tuning the domain text, the model can more accurately identify entities involved in violations and entities involved in boundary conditions.

[0074] In the data collection phase of current power dispatching regulations, 120 regulations documents published within the past five years were downloaded in batches from the power company's internal document management system or retrieved via API interface using information technology. The documents were in PDF and DOCX formats, with a total data volume of approximately 850MB. Subsequently, OCR technology was used to perform text recognition on the scanned PDFs, ensuring a text extraction accuracy of no less than 99.5%. All text was then uniformly converted to UTF-8 encoded plain text format and stored in a non-relational database to establish the raw corpus for subsequent processing. Moving to the natural language processing (NLP) algorithm parsing phase, the corpus was first preprocessed, including word segmentation using the jieba word segmentation tool with enhanced professional dictionaries. A custom dictionary was added, including 350 professional terms such as "overload," "island operation," and "asynchronous closing," and stop words were removed. Next, the BERT-based pre-trained model "bert-base-chinese" was used to embed sentences, utilizing its 12-layer Transformer architecture to capture contextual semantics. To analyze the descriptions of violations and the definitions of boundary conditions, a sequence labeling model was fine-tuned using 5000 manually annotated sentences. The BIO labeling system was used to identify entities. For example, in the sentence "Line load exceeding the stability limit by 120% for 10 minutes is considered a violation," the model accurately labeled the violation operation "line load exceeding the stability limit" and the boundary conditions "120%" and "10 minutes." In the stage of extracting keyword phrases and logical relationships, dependency parsing was performed on the identified entities. The Stanford CoreNLP tool was used to parse the syntax tree, extracting the dependency relationships between core verb phrases and modifiers. For example, "exceeding" is the core verb, "load" is its subject, "limit" is its object, and "120%" is the adverbial of "exceeding." Through rule templates (such as "If [condition], then [operation] is considered a violation") and a relational statistical model based on graph neural networks, structured tuples such as "operation-value-unit-time" were extracted from the analysis results. Finally, in the stage of obtaining the standard violation feature set, the extracted tuples are merged and standardized, for example, "120%" and "1.2 times" are unified into "1.2", and a feature vector containing fields such as violation type, key parameter threshold, logical operators (such as >, ≥, &&), and duration is constructed. The feature vector is clustered using a clustering algorithm (such as DBSCAN, with neighborhood radius eps=0.5 and minimum number of samples min_samples=5), similar features are merged, and finally a standard violation feature set containing 15 major categories and a total of 287 specific features is generated. Each feature is stored in machine-readable JSON format, for example {"Violation type": "Overload", "Equipment": "Line", "Parameter": "Load rate", "Operator": ">", "Threshold": 1.2, "Duration": 600, "Unit": "Seconds"}.

[0075] S102. Based on the standard violation feature set, construct a knowledge graph representation, where nodes correspond to violation types and operation modes, edges represent conditional dependencies, and determine the dynamic knowledge structure.

[0076] To further support the identification of complex violation scenarios, a "complex violation feature" node is introduced into the knowledge graph. This node is connected to multiple basic "violation type" nodes through logical edges (such as "AND" and "OR"), indicating that these basic violations must be satisfied simultaneously or at least one of them to constitute a higher-level violation. For example, when the feature set contains a rule "Prohibit load transfer when the bus is overloaded and the reserve capacity is insufficient," the system will create a "complex violation feature: invalid transfer under overload" node and establish "AND" relationship edges from this node to the "bus overload" node and the "insufficient reserve capacity" node, respectively.

[0077] Obtain a standard violation feature set. Store the knowledge graph using graph database technology. Generate a graph node set based on the violation types and operation methods defined in the feature set. Generate a graph edge set based on the conditional dependencies defined in the feature set. If logical conflicts exist between the node set and the edge set, apply consistency verification rules for correction. Import the verified node set and edge set into the graph database to obtain a dynamic knowledge graph.

[0078] Based on the established standard violation feature set containing 287 JSON-formatted features, constructing a knowledge graph representation first requires defining the node and edge patterns of the graph. Nodes are designed in two categories: one is the "violation type" node, whose attributes are mapped from the "violation type" and "device" fields of the feature set, such as "line overload" and "transformer temperature rise exceeding limits"; the other is the "operation mode" node, whose attributes are derived from the combination of power grid topology and switch states, such as "single busbar segmented operation" and "double circuit parallel operation". Edges are designed to represent conditional dependencies, meaning that a specific operation mode node is a necessary precondition for triggering a certain type of violation.

[0079] For example, if the feature set contains the rule "Under single-circuit power supply mode, if the standby power supply automatic transfer device is deactivated, the load transfer operation must be completed within 15 minutes," a "prone to triggering" relationship edge can be established between the "single-circuit power supply" operation mode node and the "load transfer timeout" violation type node through parsing. To achieve automated construction from the feature set to the knowledge graph, a conversion engine is developed. This engine reads each feature JSON and uses rule templates (e.g., if the "applicable scenario" field in the feature contains "island," then create or associate it with the "island operation" operation mode node) and a TF-IDF-based keyword matching algorithm (setting a similarity threshold of 0.7) to determine and create nodes and edges. For determining the dynamic knowledge structure, the key is to enable the knowledge graph to update the "activation" status of its nodes according to the real-time status of the power grid. This is achieved by associating the knowledge graph with the telemetry and communication data streams of the Supervisory Control and Data Acquisition (SCADA) system. A state mapping module is designed to subscribe to SCADA switch change and topology analysis results messages. When a change to "single bus segmented operation" is detected in the power grid topology, the "current state" attribute of the corresponding "operation mode" node in the knowledge graph is marked as "true". Simultaneously, a graph traversal algorithm (such as depth-first search) is periodically executed. Starting with all operation mode nodes with a "true" state, the algorithm traverses along "prone to triggering" edges to find all downstream "violation type" nodes and dynamically updates the "risk level" attribute of these violation nodes to "high," thus forming a hotspot view that reflects the potential violation risks under the current power grid operation mode in real time. To optimize the performance of this dynamic query, the graph data is stored and indexed using the Neo4j graph database, and its Cypher query language is used to implement complex real-time queries such as "finding all violation types that may be triggered by the current active operation modes and whose threshold parameter is greater than 1.5".

[0080] S103. Obtain the sequence of operation instructions from the massive historical scheduling logs, perform matching queries against the dynamic knowledge structure, and mark potential violations if the closing of the communication switch in the sequence does not meet the disconnection conditions, thereby obtaining a preliminary risk list.

[0081] Obtain historical scheduling log data sources. Process the log data using text parsing technology to extract structured operation instruction sequences. Associate and map the switch operation events in the instruction sequences with nodes in the dynamic knowledge graph. For each associated switch operation event, obtain its final state in the instruction sequence. Query the preset disconnection conditions for that switch node under the current operating mode from the dynamic knowledge graph. If the switch state in the instruction sequence is closed, and the preset disconnection condition in the graph is true, a condition conflict is determined. Record the switch operation events with condition conflicts and their context information as risk entries, and summarize them to generate a risk list.

[0082] First, the sequence of operation instructions needs to be obtained from massive historical scheduling logs. This requires designing a log parser based on regular expression patterns.

[0083] For example, matching operation verbs such as "close," "open," "engage," and "exit," along with equipment numbers like "2216 switch," extracts structured operation events from unstructured log text. Each event record includes a timestamp, operation object, operation type, and pre-operation state. For instance, from the log entry "14:30:05 Closed the 10kV bus tie 2216 switch at Chengdong Substation," the event {"time": "14:30:05", "object": "2216 switch", "action": "closed", "pre_state": "opened"} is extracted. Next, a matching query is performed on a dynamic knowledge structure. Here, the knowledge structure specifically refers to a real-time updated power grid topology knowledge graph, where nodes represent equipment such as switches and busbars, edges represent electrical connections, and their states (such as "closed" and "opened") are synchronized in real-time by the SCADA system. The matching query process is implemented through a sequence alignment algorithm, which compares the parsed sequence of operation instructions (such as "turn on switch 2216", "turn on switch 2217") with the real-time status of the device recorded in the knowledge graph.

[0084] Specifically, for each "close" operation instruction, the algorithm will query the knowledge graph for the current state of the operation object and the associated topological constraint rules.

[0085] For example, when the command "Close switch 2216" is detected, the algorithm queries the topology to find that switch 2216 connects bus section I and bus section II. It further queries to find that bus section I is currently powered by main transformer T1 with a load rate of 85%, while bus section II is in a "hot standby" state powered by backup power supply B1. Simultaneously, it matches a constraint from the rule base: "When the load rate of any bus exceeds 80%, closing the bus tie switch for parallel operation is prohibited to prevent circulating current impact." If closing the tie switch in the sequence does not meet this disconnection condition, a potential violation point is marked. The judgment logic here is a rule-based inference engine that comprehensively calculates the operation command, real-time topology status, and predefined safety constraint rules (stored in "IF-THEN" format, such as "IF Device Type = 'Bus Tie Switch' AND Operation = 'Close' AND (Bus A. Load Rate > 0.8 OR Bus B. Load Rate > 0.8) THEN Violation Type = 'Circulating Current Risk'"). When all conditions are met, a violation mark is generated. Finally, the system aggregates all marked violations within a time window (such as the past 24 hours) and associates them with their corresponding original log entries, triggered rule IDs, and the power grid snapshot data at that time to form a preliminary risk list. This list can be sorted according to dimensions such as violation frequency and importance of associated equipment for subsequent in-depth analysis.

[0086] S104. Use a knowledge graph traversal algorithm to scan the surrounding nodes of each point in the preliminary risk list, integrate newly added violation attributes to expand the query range, determine whether it covers complex scenarios at the boundary of the new standard, and obtain an expanded risk set. Specifically, the knowledge graph is stored in the Neo4j graph database, and graph traversal is implemented using the Cypher query language.

[0087] In this step, a breadth-first search (BFS) algorithm is used to traverse the graph starting from the risk point, primarily based on the topological characteristics of power grid risk propagation. The BFS algorithm expands the traversal range hierarchically, prioritizing the discovery of equipment nodes that are electrically closest to and directly connected to the source risk node. This aligns with the characteristic of faults or violations in the power grid spreading instantly along physical connection paths. Compared to depth-first search (DFS), BFS is more comprehensive and efficient in exploring the local impact range of risks, avoiding premature exploration of single branches and missing key related nodes at the same level, thus ensuring coverage of "complex boundary scenarios."

[0088] Among them, a knowledge graph traversal algorithm is used to scan the surrounding nodes of each point in the preliminary risk list. Specifically, the breadth-first search (BFS) algorithm is used to traverse the graph, because it can efficiently cover the local topology and is suitable for discovering related violation nodes, thereby improving the recognition rate of complex scenes.

[0089] Among them, covering complex scenarios under the new standard means identifying complex violation scenarios with multiple constraints as defined in the new standard.

[0090] Retrieve each risk entry from the initial risk list. For each risk entry, obtain its corresponding switch node from the knowledge graph. Starting from this switch node, traverse the graph using a breadth-first search algorithm to obtain the set of directly connected topology nodes. Verify the state of each node in the topology node set according to preset violation judgment rules. If a node's state meets the violation judgment rules, mark it as a derived risk node. Summarize all derived risk nodes corresponding to risk entries, remove duplicate records, and generate an extended risk set.

[0091] After obtaining the initial risk list, the system will launch a knowledge graph traversal algorithm based on breadth-first search for each marked violation point in the list, starting from the violation device node and exploring its associated topology and operating environment.

[0092] For example, for a risk point marked as "bus load rate exceeds limit when bus tie switch is closed," the algorithm starts from the bus tie switch node, sets the traversal depth to 3, first obtains the nodes of bus section I and bus section II directly connected to it, then obtains the power supply nodes supplying power to these two bus sections (such as main transformer T1 and backup power supply B1), and further obtains the upstream incoming line switches of these power supplies and the nodes of the substation on the opposite side. During the traversal, the system will integrate newly added violation attributes synchronized from the SCADA system that exceed the original rule base in real time for expanded query.

[0093] For example, in addition to load factor, the algorithm also queries attributes such as "voltage deviation," "harmonic distortion rate," and "protection setting group status" for each node along the traversal path. Suppose that when traversing to the main transformer T1 node, it finds that its current harmonic distortion rate (THD) is 4.8%, while according to the newly promulgated "Power Quality Technical Standard for Power Grids," the THD limit for the 110kV main transformer power supply bus is 4.0%. This attribute constitutes a new violation condition. The algorithm then determines whether this scenario covers the boundary complex scenario under the new standard, namely, "parallel operation is performed under high bus load (>80%), and the power quality (THD>4.0%) of one of the power sources exceeds the standard." The system uses a scenario matching engine to match the multi-dimensional states (load rate = 85%, THD = 4.8%, operation = bus tie) obtained during the current traversal with composite rules in the new standard. These rules might be expressed as "IF bus tie operation AND (Bus A. load rate > 0.8 OR Bus B. load rate > 0.8) AND (Power source A. THD > 0.04 OR Power source B. THD > 0.04) THEN Risk level upgraded to 'High'". Because both load and power quality violations are triggered simultaneously, the original risk point is judged to cover a more complex risk scenario and is thus added to the extended risk set, along with a detailed topology impact path (bus tie switch -> bus section I -> main transformer T1) and superimposed violation attribute values, providing a more comprehensive network-based perspective for risk assessment.

[0094] S105. By comparing the extended risk set with the real-time operation command input, a similarity score is calculated. If the score exceeds the preset threshold, an alarm mechanism is triggered to confirm the violation result.

[0095] The threshold was determined based on a historical data-driven optimization process: 20,000 data samples containing confirmed violations and safe operations were randomly selected from historical scheduling logs to form the training and test sets. Feature vectors of operation instructions and risk scenarios were extracted using this method, and similarity was calculated. Five-fold cross-validation was employed, with the F1 score (harmonic mean of precision and recall) as the core evaluation metric, and a grid search was performed within a threshold range of 0.5 to 0.9. Experimental results show that when the threshold is set to 0.75, the system achieves the highest overall F1 score (0.91) on the test set, achieving the optimal balance between precision and recall; therefore, this threshold was set as the system's default alarm threshold. This threshold can also be calibrated through the system management interface according to actual operational needs.

[0096] The system acquires the real-time operation command stream and an expanded risk set. Each risk description in the operation command stream and risk set is processed using a text vectorization model. The cosine similarity between the operation command vector and each risk description vector is calculated. The maximum similarity score among all similarity calculations is selected as the final similarity score, with a threshold of 0.75. This value, validated through data training, achieves an optimal balance between accuracy and false positive rate. If the final similarity score exceeds the preset threshold, an alarm is triggered. After the alarm is triggered, the confirmer outputs a confirmation result of the violation.

[0097] After generating the extended risk set, the system compares and analyzes it with the sequence of pending operation instructions obtained in real time from the operation ticket system. In practice, the system first transforms each risk scenario in the extended risk set—for example, a risk identified due to "the automatic transfer switch of the backup power supply at the opposite substation not being properly locked during line maintenance"—into a feature vector. This vector includes dimensions such as risk type code, core equipment ID (e.g., the automatic transfer switch AAT of line L1 and the opposite substation S2), and risk attribute value (e.g., the lockout status is "not locked," and the maintenance status is "yes"). Simultaneously, the system also parses a pending instruction, such as "open switch 101 of line L1," into an instruction feature vector containing dimensions such as the operation object (switch 101 of line L1), operation type (opening), and associated equipment (opposite substation S2). Next, the system uses an improved cosine similarity algorithm to calculate the matching score between the two. This algorithm not only calculates the cosine of the angle between the vectors but also assigns a higher weight factor to the "device association" dimension. For example, the weight of matching associated devices is set to 0.6, while the weight of matching general attributes is 0.4. The calculation process is as follows: First, the risk vector R and the instruction vector I are normalized; then, the similarity component of each dimension is calculated. For discrete values ​​such as device ID, a perfect match results in a component of 1, otherwise it is 0; for associated devices, if the operation object (switch 101) in the instruction is located on the topological path (line L1) affected by the risk scenario, the component is recorded as 1. Finally, the weighted sum is used to obtain the final similarity score S=Σ(wi*sim_i). The system's preset alarm trigger threshold is 0.75. Assuming that the calculated similarity between the above risk scenario and the instruction "open switch 101 on line L1" reaches 0.82, the key dimension score is high because the operation object is directly located on the risk line and the associated device (opposite station S2) is a perfect match. Since 0.82 > 0.75, the system immediately triggers a real-time alarm and marks the operation instruction as "high-risk confirmed violation". At the same time, it automatically associates and pushes the details of its corresponding extended risk set item, including the risk path (line L1 -> opposite substation S2.AAT device) and core risk attributes (standby automatic transfer switch not locked), thereby completing automated risk interception and confirmation before the instruction is executed.

[0098] S106. Extract missed and false alarm instances from the violation confirmation results, feed them back to the knowledge graph update process, adjust node weights and edge strengths, and obtain an optimized knowledge structure.

[0099] The graph embedding model used here is the Node2Vec algorithm. This algorithm simulates a biased random walk in the graph, mapping each node in the knowledge graph (representing a device, violation type, etc.) to a low-dimensional dense vector. After updating the node weights and edge strengths, rerunning the Node2Vec algorithm can bring semantically similar nodes (e.g., those with close electrical connections or frequently triggering violations simultaneously) closer together in the vector space. Subsequently, by calculating the cosine similarity between node vectors, the strength of the association between nodes can be more accurately quantified, forming an optimized knowledge structure and providing a more accurate semantic similarity foundation for subsequent risk matching.

[0100] Obtain the sets of missed detection instances and false positive instances from the violation confirmation results. Merge the sets of missed detection instances and false positive instances to generate a feedback instance set. Locate the graph nodes associated with each instance in the feedback instance set from the knowledge graph. If an instance belongs to the missed detection instance set, increase the weight value of the node associated with that instance. If an instance belongs to the false positive instance set, decrease the connection strength of the edges between the node associated with that instance and its neighboring nodes. Use a graph embedding model to update the knowledge graph with weights and edge strengths, generating new node vector representations. Based on the new node vector representations, recalculate the similarity between nodes to form an optimized knowledge structure.

[0101] After completing real-time alarms and violation confirmations, the system continuously collects two types of key feedback data: one type is missed detection instances where a violation is confirmed but the system does not generate an alarm. For example, an operation command targeting the risk of "accidental disconnection of the secondary circuit breaker of the busbar PT" was manually reviewed and found to be risky, but the system's similarity score was only 0.68, below the threshold of 0.75. The other type is false alarm instances where the system triggers an alarm but is manually confirmed as a safe operation. For example, the system gave a score of 0.80 and issued an alarm for the command "open the isolating switch of line L2" because the weight of the associated equipment was too high, but in fact there was no related risky equipment on the other side of the line. For missed detection instances, the system will parse the feature vector of the operation command and the corresponding risk scenario, and use the gradient descent method to fine-tune the initial weights of the relevant nodes in the knowledge graph.

[0102] Specifically, for the device node "Bus PT Secondary Circuit Breaker", its initial influence factor weight is 0.5. The system calculates the gradient using the loss function L = (target score 0.85 - actual score 0.68)^2 and updates the node's weight to 0.53 with a learning rate of 0.01. Simultaneously, it strengthens the edge strength between this node and the "Voltage Transformer Secondary Circuit Anomaly" risk type node, increasing the association strength value from 0.7 to 0.73. For false alarm instances, the system performs the opposite attenuation process. Analysis reveals that false alarms mainly originate from an erroneous association between "Line L2" and an unrelated "Opposite Substation S3" node in the topology map formed by an outdated maintenance record. This edge strength is 0.65. After identifying this abnormal association using the false alarm analysis algorithm, the system applies the exponential decay formula to update the edge strength value to 0.65 * 0.9 = 0.585. If this edge is not activated by a valid instance within the next three consecutive cycles, the strength value will further decay to below 0.5 and be automatically hidden by the system. These dynamic adjustment processes are integrated and calculated as a batch every 24 hours. The graph neural network algorithm is used to perform a forward propagation and weight rebalancing of the entire knowledge graph, so that the relationship between "equipment-risk-operation" in the graph can continuously evolve based on the latest practical feedback data, thereby optimizing the accuracy of subsequent risk reasoning.

[0103] S107. Based on the optimized knowledge structure, re-parse the real-time operation instructions. If composite violation characteristics are detected, output an authoritative identification report to determine the overall consistency of the verification.

[0104] Relevant data on real-time instructions are extracted from the knowledge structure. The instruction content is initially broken down to obtain operation analysis results. Based on the operation analysis results, violation features in the knowledge structure are matched. If a pattern matching the composite features is detected, a potential violation is identified. For the identified potential violations, a pre-established feature matching mechanism is used to analyze the specific manifestations of the composite features, obtaining detailed records of feature matching. A corresponding identification report is generated based on these detailed records, and its credibility level is determined using authoritative evaluation standards. Based on the credibility level of the identification report, relevant consistency indicators are analyzed to obtain data for overall evaluation. Using the overall evaluation data, a logistic regression model is used to process the consistency indicators to determine the final compliance conclusion. Based on the compliance conclusion, the final content of the identification report is compiled to generate a structured output document.

[0105] On the basis of optimizing the knowledge structure, the system first deeply analyzes the real-time operation instructions, converts the instruction text into a structured feature vector through natural language processing technology. For example, for the instruction of "switching the operation mode of main transformer M1", the system extracts key fields such as the main transformer device identifier, operation type, and context timestamp, and generates a 128-dimensional feature vector. Subsequently, it uses a pre-trained embedding model to calculate the cosine similarity between this vector and the known risk scenarios in the knowledge graph. Suppose the calculation result is 0.82, exceeding the preset risk threshold of 0.78, and the system marks it as a potential risk instruction. Then, the system conducts multi-dimensional analysis on the detected compound violation features, invokes a classification algorithm based on decision trees, inputs the feature vector into the model, combines the association probability of 0.67 between "main transformer switching" and "overload risk" in historical data, and outputs the confidence level that this instruction may cause compound violations as 0.85. At the same time, it cross-verifies the association strength of 0.72 with the "abnormal cooling system status" node, and comprehensively judges the possibility of multiple risk superposition. To further generate an authoritative identification report, the system automatically integrates the analysis results, and based on the confidence level of 0.85 and the association strength of 0.72, generates a detailed report including risk types, impact ranges, and triggering conditions. For example, the report points out that "the switching operation of main transformer M1 may result in an overload probability of 0.67 and a cooling abnormality probability of 0.45", and attaches the visualization data of the relevant paths in the knowledge graph. Finally, the system verifies the overall analysis process through a consistency verification algorithm, adopts an entropy-based evaluation method to calculate the distribution uniformity of the output results of each module. Suppose the entropy value is 0.92, which is higher than the preset threshold of 0.88, and it is determined that the verification consistency is qualified. If the consistency does not meet the standard, the system automatically triggers a backup analysis model to recalculate the feature vector until the entropy value meets the requirements to ensure the reliability of the results. The entire process is automatically processed by the system, relying on algorithms and data analysis to form a closed-loop logic, ensuring the rigor and accuracy from instruction parsing to report output.

[0106] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A power dispatching method for urban low-voltage distribution networks, characterized in that, include: By collecting textual data of current power dispatching regulations, natural language processing algorithms are used to analyze the descriptions of violations and the definitions of boundary conditions, extracting key phrases and logical relationships to obtain a standard violation feature set. Based on the standard violation feature set, a knowledge graph representation is constructed, where nodes correspond to violation types and operation modes, edges represent conditional dependencies, and a dynamic knowledge structure is determined. The dynamic knowledge structure refers to a knowledge representation form based on a graph database that can adjust the relationships between nodes and edges in real time as the standard is updated. The operation instruction sequence is obtained from the massive historical scheduling logs, and a matching query is performed based on the dynamic knowledge structure. If the contact switch in the sequence is closed and does not meet the disconnection condition, the potential violation point is marked and a preliminary risk list is obtained. A knowledge graph traversal algorithm is used to scan the surrounding nodes of each point in the initial risk list, and the newly added violation attributes are integrated to expand the query scope. It is then determined whether the extended risk set is covered to see if it covers complex scenarios at the boundary of the new standard. By comparing the extended risk set with the real-time operation command input, a similarity score is calculated. If the score exceeds a preset threshold, an alarm mechanism is triggered to confirm the violation. Extract missed and false alarm instances from the violation confirmation results, feed them back into the knowledge graph update process, adjust node weights and edge strengths, and obtain an optimized knowledge structure; Based on the optimized knowledge structure, real-time operation instructions are re-analyzed. If composite violation characteristics are detected, an authoritative identification report is output to determine the overall consistency of the verification. Composite violation characteristics refer to situations where multiple procedural clauses are violated simultaneously in the same operation instruction.

2. The power dispatching method for urban low-voltage distribution networks according to claim 1, characterized in that, The process involves collecting textual data from current power dispatching regulations, using natural language processing algorithms to parse the descriptions of violations and boundary condition definitions, extracting keyword phrases and logical relationships, and obtaining a standard violation feature set, including: Obtain the text set of power dispatching regulations; Natural language processing algorithms are used to parse the descriptions of violations and the definitions of boundary conditions in the text set; Extract the keyword groups corresponding to the descriptions of the violations from the analysis results to obtain a set of violation predicate groups; Extract the logical relationships corresponding to the boundary conditions from the parsing results to obtain the set of conditional relationship relationships; By integrating the set of violating predicate phrases and the set of conditional semantic relations, a standard set of violation features is formed; Based on the standard violation feature set, a violation operation identification model is constructed.

3. The power dispatching method for urban low-voltage distribution networks according to claim 1, characterized in that, The process involves constructing a knowledge graph representation based on a standard violation feature set, where nodes correspond to violation types and operational modes, edges represent conditional dependencies, and a dynamic knowledge structure is determined, including: Obtain the standard violation feature set; Knowledge graphs are stored using graph database technology; Generate a graph node set based on the violation type and operation mode definition in the feature set; Generate a graph edge set based on the conditional dependencies defined in the feature set; If there is a logical conflict between the node set and the edge set, the consistency check rule is invoked to correct it; The verified node set and edge set are imported into the graph database to obtain a dynamic knowledge graph.

4. The power dispatching method for urban low-voltage distribution networks according to claim 1, characterized in that, The process involves retrieving operation instruction sequences from massive historical scheduling logs, performing matching queries against a dynamic knowledge structure, and marking potential violations if a communication switch in the sequence is closed, thus obtaining a preliminary risk list, including: Obtain historical scheduling log data source; Log data is processed using text parsing technology to extract structured operation instruction sequences; Associate and map the switch operation events in the instruction sequence with nodes in the dynamic knowledge graph; For each associated switch operation event, obtain its final state in the instruction sequence; Query the preset disconnection conditions of the switch node under the current operating mode from the dynamic knowledge graph; If the switch status in the instruction sequence is closed, and the preset disconnection condition in the diagram is true, then a condition conflict is determined to exist. Switch operation events with conflicting conditions and their context information are recorded as risk entries, and a risk list is generated by summarizing them.

5. A power dispatching method for urban low-voltage distribution networks according to claim 1, characterized in that, The algorithm employs a knowledge graph traversal to scan the surrounding nodes of each point in the initial risk list, integrates newly added violation attributes to expand the query scope, and determines whether it covers complex scenarios at the boundary of the new standard, thus obtaining an expanded risk set, including: Obtain each risk entry from the preliminary risk list; For each risk item, obtain its corresponding switch node from the knowledge graph; Starting from the switch node, a breadth-first search algorithm is used to traverse the graph and obtain the set of topological nodes directly connected to it. According to the preset violation judgment rules, the status of each node in the topology node set is checked; If a node's status meets the violation determination rules, then the node is marked as a derivative risk node; Summarize all derived risk nodes corresponding to risk entries, remove duplicate records, and generate an extended risk set.

6. The power dispatching method for urban low-voltage distribution networks according to claim 1, characterized in that, The process involves comparing an expanded risk set with real-time operation command inputs to calculate a similarity score. If the score exceeds a preset threshold, an alarm mechanism is triggered to confirm the violation, including: Obtain the real-time operation command stream and extended risk set; A text vectorization model is used to process each risk description in the operation instruction stream and risk set; Calculate the cosine similarity between the operation instruction vector and each risk description vector; The maximum value among all similarity calculation results is selected as the final similarity score; If the final similarity score is greater than the preset threshold, an alarm will be triggered. After the alarm is triggered, the acknowledgment device outputs the result confirming the violation.

7. A power dispatching method for urban low-voltage distribution networks according to claim 1, characterized in that, The process of extracting missed and false alarm instances from violation confirmation results and feeding them back into the knowledge graph update process to adjust node weights and edge strengths to obtain an optimized knowledge structure includes: Obtain the set of missed detection instances and the set of false alarm instances from the violation confirmation results; Merge the sets of missed detection instances and false alarm instances to generate a set of feedback instances; Locate the graph node associated with each instance within the set of feedback instances from the knowledge graph; If an instance belongs to the set of missed instances, increase the weight value of the node associated with that instance. If an instance belongs to the set of false positive instances, then the connection strength of the edges between the associated nodes of that instance and its neighboring nodes is reduced. A graph embedding model was used to process and update the knowledge graph with weights and edge strengths, generating new node vector representations. Based on the new node vector representation, the similarity between nodes is recalculated to form an optimized knowledge structure.

8. A power dispatching method for urban low-voltage distribution networks according to claim 1, characterized in that, The process involves re-analyzing real-time operation instructions based on an optimized knowledge structure. If composite violation characteristics are detected, an authoritative identification report is output to determine overall consistency, including: Relevant data on real-time instructions are obtained from the knowledge structure, and the instruction content is initially broken down to obtain the results of operation analysis. Based on the results of the operation analysis, the violation features in the knowledge structure are matched. If a pattern that matches the composite features is detected, it is determined that there is a potential violation. For identified potential violations, a pre-established feature matching mechanism is used to analyze the specific manifestations of composite features and obtain detailed records of feature matching. By using detailed records of feature matching, a corresponding recognition report is generated, and the credibility level of the report is determined by combining authoritative evaluation standards. Based on the credibility level of the identification report, analyze the relevant indicators of consistency verification to obtain the basis data for the overall assessment; Based on the overall assessment data, a logistic regression model is used to process the various indicators that are consistent with the verification, and the final compliance conclusion is determined. Based on the compliance conclusions, organize the final content of the identification report and generate a structured output document.

Images