Power distribution network fault identification detection method based on graph computation model

Abstract

The application relates to a power distribution network fault identification detection method based on a graph calculation model, which comprises the following steps: constructing a power distribution network multi-source heterogeneous data graph calculation model; loading the graph calculation model in a storage-computation integrated graph calculation acceleration engine, correcting and processing the graph calculation model by using a power grid parameter identification deep learning model, and outputting an optimized graph model; performing topological analysis and fault analysis and judgment processing on the optimized graph model, and outputting a fault identification result. Compared with the prior art, the application can realize high-precision and high-real-time fault identification detection of a super-large-scale power distribution network by constructing a novel power distribution network data graph calculation model, building a storage-computation integrated graph calculation acceleration engine, and fusing topological analysis, parameter identification and deep learning technology.

Description

Technical Field

[0001] This invention relates to the field of power system fault detection technology, and in particular to a method for identifying and detecting distribution network faults based on a graph computation model. Background Technology

[0002] With the in-depth advancement of the construction of new power systems, distribution networks are showing typical characteristics such as deep interaction between power sources, grids, loads and storage, large-scale growth of distributed power sources, and hybrid operation of AC and DC. The distribution networks of megacities have formed ultra-large-scale complex networks covering all voltage levels from 1000 to 0.4kV, with hundreds of millions of equipment nodes and tens of thousands of electrical nodes, which puts forward extremely high requirements for the real-time performance, accuracy and fault tolerance of distribution network fault identification and detection.

[0003] Traditional methods for identifying and detecting faults in power distribution networks have the following core shortcomings, making them unable to meet the development needs of modern power systems.

[0004] I. Insufficient data model adaptability: Traditional distribution network data modeling often adopts structured models oriented towards single business, which cannot adapt to diverse and heterogeneous power grid resources, assets, topology, measurement, and business event information. It is difficult to support unified fault analysis in scenarios of source-grid-load-storage interaction and AC / DC hybrid interconnection. The fragmented data foundation leads to the lack of dimensions for fault judgment and cannot cover all elements of fault scenarios.

[0005] Second, the computational efficiency and real-time performance cannot meet the requirements: Traditional centralized computing architecture cannot adapt to the high-dimensional data processing needs of hundreds of millions of equipment nodes in ultra-large-scale power grids. The computational latency of topology analysis and fault judgment is high, which cannot meet the requirements of real-time optimization and online fault detection of tens of thousands of electrical nodes, and it is difficult to support rapid fault handling and coordinated load transfer between the main and distribution networks.

[0006] 3. Insufficient accuracy in topology and parameter identification: Traditional fault detection methods rely on offline topology and fixed parameter models. When faced with dynamic changes in topology and noise in measurement data brought about by the access of distributed power sources, topology errors and parameter identification deviations are prone to occur, resulting in inaccurate fault location, misjudgment and missed judgment. The algorithm has poor fault tolerance and cannot adapt to the uncertainty of scenarios with a high proportion of new energy.

[0007] IV. Weak global situation awareness: It is impossible to achieve real-time awareness of the global situation of the distribution network, and fault handling is mostly a reactive response after the fact. It is impossible to achieve large-scale real-time dynamic optimization and it is difficult to provide early warning of fault risks. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a distribution network fault identification and detection method based on a graph computing model, which can achieve high-precision and high-real-time fault identification and detection in ultra-large-scale distribution networks.

[0009] The objective of this invention can be achieved through the following technical solution: a method for fault identification and detection in a distribution network based on a graph computation model, comprising the following steps: S1. Construct a calculation model for multi-source heterogeneous data graphs of the power distribution network; S2. Load the graph computing model into the in-memory graph computing acceleration engine, use the grid parameter identification deep learning model to correct the graph computing model, and output the optimized graph model. S3. For the optimized graph model, perform topology analysis and fault assessment, and output the fault identification results.

[0010] Furthermore, S1 specifically involves abstracting the power sources, loads, energy storage, power distribution equipment, measuring devices, assets, and business events in the distribution network as graph nodes, and abstracting electrical connections, power transmission, business associations, and asset ownership as graph edges, thereby constructing a hierarchical temporal graph model.

[0011] Furthermore, the graph nodes include electrical nodes, asset nodes, business nodes, and measurement nodes, and the graph edges include electrical edges, asset edges, business edges, and measurement edges. The electrical nodes include power sources (centralized power sources, distributed photovoltaics, wind power, virtual power plants, energy storage), loads (industrial and commercial, residential, charging piles), and electrical equipment (transformers, lines, switches, converter stations, busbars). The asset nodes include equipment assets, maintenance personnel, maintenance plans, and spare parts; The business nodes include scheduling instructions, market transactions, carbon flow quotas, and user services; The measurement nodes include SCADA (Supervisory Control And Data Acquisition) measurements, distribution network automation measurements, distributed power generation measurements, and user-side measurements. The electrical edge includes line connections, power flow, voltage / current transmission, and electrical attributes; The asset edge includes asset ownership, operation and maintenance association, and repair association; The business edge includes scheduling association, market transaction association, carbon flow association, and user service association; The measurement edge includes the association between the measurement device and the equipment, and the association for the transmission of measurement data.

[0012] Furthermore, S1 includes the following process: Acquire all heterogeneous data from the distribution network, including structured, semi-structured, and unstructured data; Based on graph data transformation rules, all heterogeneous data of the distribution network are mapped to the node attributes or edge attributes of the graph model; Add timestamp attributes to all nodes and edges, and build a layered decoupled architecture according to the physical topology layer, business association layer, and situational awareness layer to obtain the graph computing model.

[0013] Furthermore, the physical topology layer is used to store electrical topology, device parameters, and real-time measurement data information; The business association layer is used to store asset, scheduling, market, and carbon flow business data information; The situation awareness layer is used to store power grid operating status, fault risks, and early warning information.

[0014] Furthermore, the in-memory computing graph computing acceleration engine in S2 adopts a collaborative architecture of in-memory computing and persistent storage, accelerates graph computing through a distributed parallel framework, and provides standardized APIs / SDKs for integrating deep learning models for power grid parameter identification, topology analysis, and fault judgment and processing algorithms. It also supports unified scheduling of heterogeneous computing power from CPU, GPU, NPU, and in-memory computing chips, as well as automatic elastic scaling up and down based on computing load.

[0015] Furthermore, the in-memory computing and persistent storage collaborative architecture specifically stores hot data (real-time measurements, current topology) in memory to achieve microsecond-level access; stores cold data (historical topology, fault records) in distributed persistent storage to reduce costs; and uses adjacency list compressed storage and attribute block storage to reduce memory access overhead and improve graph traversal efficiency. The distributed parallel framework specifically partitions the graph according to the main distribution network, voltage level, and geographical region to improve computational locality and reduce cross-node communication; and adopts an incremental computation optimization strategy: for topology changes and measurement updates, only the changed nodes / edges are updated to avoid full recalculation and reduce topology analysis latency.

[0016] Furthermore, the deep learning model for power grid parameter identification in S2 adopts a hybrid driving architecture of mathematical equations and neural networks to identify line parameters, equipment parameters and AC / DC hybrid parameters online. Combined with a power-specific artificial intelligence pre-trained model, it completes topology error identification, state estimation and bad data detection.

[0017] Furthermore, the topology analysis in S3 specifically employs an incremental graph update algorithm to update the topology in real time in response to changes. It then integrates measurement data and state estimation, and identifies topology errors based on graph connectivity analysis and power flow verification. The topology changes include switching operations and distributed power supply access / exit, and the topology errors include false alarms in switch status and wiring errors.

[0018] Furthermore, the fault assessment and processing in S3 includes fault location, fault type identification, and fault impact analysis. Specifically, the fault location is based on graph traversal and graph neural network, which integrates real-time measurements to locate the fault section and fault point. The fault type identification specifically involves extracting the electrical characteristics (current, voltage, power, harmonics) of graph nodes / edges, and then automatically identifying short circuit, grounding, open circuit, and arcing faults based on deep learning. The fault impact analysis is specifically based on graph connectivity analysis and power flow calculation to generate a fault impact list, including users experiencing power outages and affected power sources.

[0019] Compared with the prior art, the present invention has the following advantages: This invention first constructs a multi-source heterogeneous data graph calculation model for a distribution network. Then, it loads the graph calculation model into an in-memory computing-integrated graph calculation acceleration engine. A deep learning model for power grid parameter identification is used to refine the graph calculation model, resulting in an optimized graph model. Finally, topology analysis and fault assessment are performed on the optimized graph model, outputting fault identification results. Thus, by constructing a novel distribution network data graph calculation model and building an in-memory computing-integrated graph calculation acceleration engine, integrating topology analysis, parameter identification, and deep learning technologies, high-precision, high-real-time fault identification and detection can be achieved in ultra-large-scale distribution networks.

[0020] This invention constructs a novel multi-element heterogeneous data graph calculation model for power distribution networks. After acquiring all heterogeneous data of the power distribution network, based on graph data conversion rules, it abstracts power sources, loads, energy storage, power distribution equipment, measuring devices, assets, and business events in the power distribution network into graph nodes, and abstracts electrical connections, power transmission, business associations, and asset ownership into graph edges. It constructs a hierarchical temporal graph model, realizing unified data modeling for all elements of the power source, grid, load, and storage, all voltage levels, and AC / DC hybrid scenarios. This solves the problem of fragmentation in traditional data models, provides a complete data foundation for fault identification and detection, and perfectly adapts to the complex operating scenarios of the new power system power distribution network.

[0021] This invention designs an in-memory computing graph acceleration engine, employing a collaborative architecture of in-memory computing and persistent storage. It accelerates graph computing through a distributed parallel framework and provides standardized APIs / SDKs for integrating deep learning models for power grid parameter identification, topology analysis, and fault assessment algorithms. It also supports unified scheduling of heterogeneous computing power from CPUs, GPUs, NPUs, and in-memory computing chips, as well as automatic elastic scaling based on computing load. This enables parallel computing acceleration for hundreds of millions of device nodes and tens of thousands of electrical nodes, effectively reducing computational latency for topology analysis and fault assessment, and meeting the needs of online real-time fault detection in distribution networks.

[0022] In this invention, the deep learning model for power grid parameter identification adopts a hybrid driving architecture of mathematical equations and neural networks. By identifying line parameters, equipment parameters, and AC / DC hybrid parameters online, and combining them with a power-specific artificial intelligence pre-trained model, it completes topology error identification, state estimation, and bad data detection. In addition, it identifies topology errors through topology analysis and performs fault location, fault type identification, and fault impact analysis through fault judgment processing. Thus, it integrates topology analysis, parameter identification, and deep learning technologies to achieve automatic topology error identification and high-precision online parameter identification. Combined with iterative optimization of the model based on real fault feedback, it significantly improves the accuracy of fault location and type identification, solving the problems of misjudgment and omission and poor fault tolerance of traditional methods. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 A schematic diagram of the application process in Example 2. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0025] Example 1 like Figure 1 As shown, a method for fault identification and detection in a distribution network based on a graph computation model includes the following steps: S1. Construct a calculation model for multi-source heterogeneous data graphs of the power distribution network; S2. Load the graph computing model into the in-memory graph computing acceleration engine, use the grid parameter identification deep learning model to correct the graph computing model, and output the optimized graph model. S3. For the optimized graph model, perform topology analysis and fault assessment, and output the fault identification results.

[0026] In step S1, when constructing the graph calculation model, the entire elements of the distribution network (source, grid, load, and storage) are taken as objects. The power sources, loads, energy storage, power distribution equipment, measuring devices, assets, and business events in the distribution network are abstracted as graph nodes, and electrical connections, power transmission, business associations, and asset ownership are abstracted as graph edges. A hierarchical temporal graph model is constructed to achieve unified integration and real-time updates of diverse and heterogeneous data, thereby forming a unified data foundation that is compatible with hundreds of millions of equipment nodes and tens of thousands of electrical nodes.

[0027] In step S2, this solution constructs an in-memory computing graph acceleration engine to meet the computational needs of hundreds of millions of equipment nodes and tens of thousands of electrical nodes in the power distribution network: It enables distributed storage and parallel computing acceleration of graph data, supports multi-temporal topology analysis and real-time graph model optimization calculation, and effectively reduces the computational latency of topology analysis and fault assessment. It meets the real-time fault detection requirements of tens of thousands of electrical nodes, provides high-performance computing support for fault identification and detection in distribution networks, and adapts to the requirements of integrated online analysis of main and distribution systems in ultra-large-scale power grids.

[0028] This in-memory computing graph acceleration engine adopts an in-memory computing and persistent storage collaborative architecture, designs a distributed parallel graph computing acceleration framework, optimizes graph partitioning and incremental computing strategies, supports multi-temporal graph computing and heterogeneous computing power scheduling, and realizes low-latency computing for topology analysis of hundreds of millions of device nodes and real-time optimization of tens of thousands of electrical nodes, providing high-performance computing support for fault identification and detection.

[0029] In step S2, this scheme constructs a deep learning model for power grid parameter identification: Using mathematical equations such as power flow equations of distribution networks as physical constraints, and combining them with deep learning models, we can achieve high-precision online identification of line parameters and equipment parameters of distribution networks, thereby correcting the parameter deviations of the graph calculation model. By combining pre-trained AI models specifically designed for the power industry, the training of downstream models for distribution network situational awareness can be supported, thereby improving the accuracy of subsequent fault identification.

[0030] In step S3, on the one hand, topology analysis is used to realize real-time updates of the topology structure and automatic identification of topology errors, solving the problem of dynamic topology changes caused by distributed power supply access and providing an accurate topology basis for fault location. On the other hand, fault judgment and processing are carried out by integrating graph traversal algorithms and graph neural network (GNN) technology to realize fault location, fault type identification and fault impact analysis.

[0031] When performing actual distribution network fault identification, the graph computing model is updated by accessing multi-source data of the distribution network in real time. The in-memory computing graph computing acceleration engine drives the algorithm and deep learning to run in tandem to output fault identification information.

[0032] Example 2 This solution applies the approach described in Implementation Example 1. First, it connects to the existing business system to collect multi-source heterogeneous data: structured data (SCADA measurement data, PMS equipment parameter data, marketing load data), semi-structured data (maintenance records, fault history data), and unstructured data (inspection images / videos). Data preprocessing is performed, including data cleaning (removing outliers and missing values), data standardization (unifying coding, units, and formats), and data anonymization (protecting user privacy and sensitive power grid information), in order to build a data resource pool and ensure that data quality meets standards.

[0033] Next, select suitable hardware and software equipment, including distributed servers, in-memory computing chips, and GPU computing clusters, to ensure that the storage and computing needs of hundreds of millions of device nodes are met; set up the development environment: select suitable graph computing frameworks (improved Pregel / GraphX), deep learning frameworks (TensorFlow / PyTorch), and databases (distributed graph databases), and complete environment deployment and debugging.

[0034] After completing the above-mentioned preliminary accurate work, such as Figure 2 As shown, this embodiment establishes four core components (a multi-source heterogeneous data graph calculation model for the distribution network, an in-memory computing graph calculation acceleration engine, a distribution network topology analysis and fault judgment algorithm, and a deep learning model for power grid parameter identification) and coordinates them. Specifically: I. Constructing a new multi-element heterogeneous data graph calculation model (data foundation) for power distribution networks 1. Abstraction and standardized definition of all elements in a graphical model Taking the 1000-0.4kV full voltage level distribution network as the object, a standardized diagram abstraction of all elements of source, grid, load and storage is performed. Among them, the node types are divided into the following categories: Electrical nodes: power sources (centralized power sources, distributed photovoltaic / wind power, virtual power plants, energy storage), loads (industrial and commercial, residential, charging piles), and electrical equipment (transformers, lines, switches, converter stations, busbars). Asset nodes: equipment assets, maintenance personnel, maintenance plans, spare parts; Business nodes: scheduling instructions, market transactions, carbon flow quotas, and user services; Measurement nodes: SCADA measurement, distribution network automation measurement, distributed power generation measurement, user-side measurement.

[0035] Edge types are categorized as follows: Electrical side: Line connections, power flow, voltage / current transmission, and labeling electrical attributes such as impedance, capacity, and voltage level; Asset-related aspects: asset ownership, operation and maintenance association, and repair association; Business-side: scheduling association, market transaction association, carbon flow association, user service association; Measurement edge: The relationship between the measurement device and the equipment, and the relationship of measurement data transmission.

[0036] This approach follows the power industry standard to unify the coding, attributes, and semantics of nodes / edges, enabling semantic interoperability between the main distribution network, power generation, load and storage systems, and AC / DC systems, thus solving the semantic fragmentation problem of traditional models.

[0037] 2. Fusion mapping and temporal graph design of multi-source heterogeneous data Multi-source data integration and intelligent conversion: Access all heterogeneous data from the distribution network (structured: SCADA / PMS / marketing system; semi-structured: maintenance / fault records; unstructured: inspection images / videos), and map the heterogeneous data to the node / edge attributes of the graph model based on graph data conversion rules; Temporal graph model design: Add timestamp attributes to all nodes / edges to support the storage, query and calculation of multi-temporal topologies, realize historical topology backtracking, real-time topology update and future topology prediction, and adapt to multi-timescale requirements such as fault detection (second level), unit combination (minute level) and carbon flow tracking (day level); Data consistency verification: A real-time verification mechanism is adopted to automatically compare the graphical model with the source system data, correct data deviations, and ensure the consistency between the graphical model and the actual power grid.

[0038] 3. Layered graphical model architecture design A layered decoupling architecture is adopted to decouple data from business logic, thereby improving maintainability. This layered decoupling architecture includes: Physical topology layer (core layer): Stores electrical topology, equipment parameters, and real-time measurements, providing a foundation for fault identification, topology analysis, and power flow calculation; Business-related layer: Stores business data such as assets, scheduling, market, and carbon flow, supporting scenarios such as spot market clearing, carbon flow tracking, and equipment management; Situation awareness layer: Stores power grid operating status, fault risks, and early warning information to support situation awareness and decision optimization.

[0039] II. Building an In-Store Computing Graph Acceleration Engine (Computing Power Support) 1. In-memory computing collaborative architecture design To address the sparsity and locality of power grid graph data, a collaborative architecture of in-memory computing and persistent storage is designed. Hot data (real-time measurements, current topology) is stored in memory for microsecond-level access, while cold data (historical topology, fault records) is stored in distributed persistent storage to reduce costs. Furthermore, the storage structure is optimized by using adjacency list compression storage and attribute block storage to reduce memory access overhead and improve graph traversal efficiency.

[0040] 2. Distributed Parallel Graph Computing Acceleration Framework On the one hand, the map is divided according to the main distribution network, voltage level, and geographical region to improve computational locality and reduce cross-node communication; on the other hand, for topology changes and measurement updates, only the changed nodes / edges are updated to avoid full recalculation and reduce topology analysis latency.

[0041] 3. Heterogeneous computing power scheduling It unifies the scheduling of computing power from CPU, GPU, NPU, and in-memory computing chips, automatically allocates the optimal computing power to different algorithms (CPU for topology analysis, GPU for deep learning), and supports hardware acceleration of core operators by FPGA / in-memory computing chips. Elastic scaling: Automatically scales up and down based on computing load to improve resource utilization.

[0042] 4. Standardized interfaces and system integration It provides standardized APIs / SDKs to support rapid integration with upper-layer algorithm libraries and deep learning engines, enabling seamless integration with existing power grid business systems (intelligent dispatching, spot market, equipment management) and achieving data sharing and reuse.

[0043] III. Constructing Algorithms for Distribution Network Topology Analysis and Fault Assessment 1. Fast topology analysis and tracing Real-time incremental topology update: For topology changes such as switching operations, distributed power supply access / exit, and faults, an incremental graph update algorithm is used to achieve millisecond-level topology updates; Topology error identification and verification: By integrating measurement data and state estimation, and based on graph connectivity analysis and power flow verification, topology errors such as false alarms of switch status and wiring errors are automatically identified; Integrated topology analysis of main and distribution networks: spanning all voltage levels from 1000 to 0.4kV, it enables unified analysis of the main and distribution network topologies, solving the problem of fragmented main and distribution topologies.

[0044] 2. Online comprehensive fault analysis Rapid fault location: Based on graph traversal (DFS / BFS) and graph neural network (GNN), and combined with real-time measurement, it realizes the location of fault section / fault point, and is suitable for AC / DC hybrid and distributed power supply high proportion scenarios. Fault type identification: Extract electrical features (current, voltage, power, harmonics) of graph nodes / edges, and automatically identify faults such as short circuits, grounding, open circuits, and arcing based on deep learning; Fault impact assessment: Based on graph connectivity analysis and power flow calculation, a fault impact list (power outage users, affected power sources, etc.) is automatically generated, which is beneficial for supporting subsequent scheduling decisions.

[0045] Furthermore, in practical applications, the algorithm can be automatically iterated and optimized by combining real-world fault feedback with comparative calculations and deductions, thereby improving the algorithm's fault tolerance and forming a closed loop of "data-algorithm-feedback-optimization".

[0046] IV. Constructing a Deep Learning Model for Power Grid Parameter Identification 1. Hybrid Mathematical-Neural Network Driven Core Architecture To address the shortcomings of purely physics-driven (poor noise resistance) and purely data-driven (poor generalization) approaches, a hybrid architecture integrating the physical and data layers is constructed: Physical constraint layer: using power flow equations and Kirchhoff's laws as hard constraints to ensure that the model conforms to the physical laws of the power system; Data-driven layer: Based on neural networks such as GNN and LSTM, it learns the nonlinear characteristics and parameter deviations of the power grid; Fusion Optimization Layer: Combines physical constraints and data-driven approaches, taking into account both physical rationality and data accuracy, to solve the problem of inaccurate parameter identification.

[0047] 2. Online identification of all parameters of the distribution network Line parameter identification: In response to parameter changes caused by line aging and distributed power supply access, online identification of line resistance, reactance, and capacitance is performed. Equipment parameter identification: Online identification of equipment parameters for transformers (turn ratio, loss), converter stations, energy storage, and distributed power sources, and correction of parameter deviations in graphical models; AC / DC hybrid parameter identification: A hybrid drive model is adopted to achieve unified identification of AC / DC system parameters, adapting to AC / DC hybrid scenarios.

[0048] 3. Topology error identification and state estimation Topology error identification: By integrating parameter identification results and measurement data, topology errors can be automatically identified, solving the problems of misjudgment and omission in traditional methods; State estimation: The state estimation problem is transformed into a graph optimization problem. Based on a hybrid driving model, the accuracy of state estimation is improved, and it is adapted to high-dimensional and high-noise distribution network measurement data. Bad data detection: Based on graph anomaly detection algorithm, it automatically identifies bad measurement data, improves data quality, and ensures the accuracy of fault diagnosis.

[0049] 4. Power-specific map pre-training large model and situational awareness Pre-trained large model construction: Based on massive power grid map data (topology, measurement, fault, meteorology, market), train a pre-trained large model of power grid special map to learn the physical laws, fault modes and business logic of the power grid; Downstream task adaptation: Through fine-tuning, it can quickly adapt to scenarios such as fault identification, parameter identification, carbon flow tracking, and spot market clearing, reducing the need for labeled data; Global situational awareness: Based on pre-trained models and graph computing engines, the system can perceive the operating status and fault risks of the distribution network in real time, enabling proactive early warning and facilitating the upgrade from passive response to proactive prevention and control. Similarly, in practical applications, by combining on-site fault feedback, errors in basic data can be automatically identified, the model can be iteratively optimized, and the robustness of the algorithm can be continuously improved.

[0050] In summary, this solution first constructs a multi-dimensional heterogeneous data graph computation model for the distribution network, forming a unified data foundation; then, it loads this graph model into an in-memory computing acceleration engine to achieve large-scale, efficient in-memory computing; finally, it uses parameter identification deep learning to correct errors and refine parameters in the graph model; and finally, using the corrected, high-quality graph model as input, it runs topology analysis and fault judgment algorithms on the acceleration engine to quickly and accurately output fault identification results. This solution can meet the operational needs of new power systems, including source-grid-load-storage interaction, AC / DC hybrid interconnection, and mega-city power grids, achieving high-precision, high-real-time fault identification and detection for ultra-large-scale distribution networks.

Claims

1. A method for fault identification and detection in distribution networks based on a graph computation model, characterized in that, Includes the following steps: S1. Construct a calculation model for multi-source heterogeneous data graphs of the power distribution network; S2. Load the graph computing model into the in-memory graph computing acceleration engine, use the grid parameter identification deep learning model to correct the graph computing model, and output the optimized graph model. S3. For the optimized graph model, perform topology analysis and fault assessment, and output the fault identification results.

2. The method for distribution network fault identification and detection based on a graph computation model according to claim 1, characterized in that, Specifically, S1 involves abstracting power sources, loads, energy storage, power distribution equipment, measuring devices, assets, and business events in the distribution network as graph nodes, and abstracting electrical connections, power transmission, business associations, and asset ownership as graph edges to construct a hierarchical temporal graph model.

3. The method for distribution network fault identification and detection based on a graph computation model according to claim 2, characterized in that, The graph nodes include electrical nodes, asset nodes, business nodes, and measurement nodes; the graph edges include electrical edges, asset edges, business edges, and measurement edges; and the electrical nodes include power sources, loads, and electrical equipment. The asset nodes include equipment assets, maintenance personnel, maintenance plans, and spare parts; The business nodes include scheduling instructions, market transactions, carbon flow quotas, and user services; The measurement nodes include SCADA measurements, distribution network automation measurements, distributed power generation measurements, and user-side measurements; The electrical edge includes line connections, power flow, voltage / current transmission, and electrical attributes; The asset edge includes asset ownership, operation and maintenance association, and repair association; The business edge includes scheduling association, market transaction association, carbon flow association, and user service association; The measurement edge includes the association between the measurement device and the equipment, and the association for the transmission of measurement data.

4. The method for distribution network fault identification and detection based on a graph computation model according to claim 3, characterized in that, S1 includes the following process: Acquire all heterogeneous data from the distribution network, including structured, semi-structured, and unstructured data; Based on graph data transformation rules, all heterogeneous data of the distribution network are mapped to the node attributes or edge attributes of the graph model; Add timestamp attributes to all nodes and edges, and build a layered decoupled architecture according to the physical topology layer, business association layer, and situational awareness layer to obtain the graph computing model.

5. The method for distribution network fault identification and detection based on a graph computation model according to claim 4, characterized in that, The physical topology layer is used to store electrical topology, equipment parameters, and real-time measurement data. The business association layer is used to store asset, scheduling, market, and carbon flow business data information; The situation awareness layer is used to store power grid operating status, fault risks, and early warning information.

6. The method for fault identification and detection in a distribution network based on a graph computation model according to claim 1, characterized in that, The in-memory computing and persistent storage-integrated graph computing acceleration engine in S2 adopts a collaborative architecture of in-memory computing and persistent storage, accelerates graph computing through a distributed parallel framework, and provides standardized APIs / SDKs for integrating deep learning models for power grid parameter identification, topology analysis, and fault judgment and processing algorithms. It also supports unified scheduling of heterogeneous computing power from CPU, GPU, NPU, and in-memory computing chips, as well as automatic elastic scaling up and down based on computing load.

7. The method for distribution network fault identification and detection based on a graph computation model according to claim 6, characterized in that, The in-memory computing and persistent storage collaborative architecture specifically stores hot data in memory for microsecond-level access; stores cold data in distributed persistent storage to reduce costs; and uses adjacency list compressed storage and attribute block storage to reduce memory access overhead and improve graph traversal efficiency. The distributed parallel framework specifically partitions the graph according to the main distribution network, voltage level, and geographical region to improve computational locality and reduce cross-node communication; and adopts an incremental computation optimization strategy: for topology changes and measurement updates, only the changed nodes / edges are updated to avoid full recalculation and reduce topology analysis latency.

8. The method for fault identification and detection in a distribution network based on a graph computation model according to claim 1, characterized in that, The deep learning model for power grid parameter identification in S2 adopts a hybrid driving architecture of mathematical equations and neural networks to identify line parameters, equipment parameters and AC / DC hybrid parameters online. Combined with a power-specific artificial intelligence pre-trained model, it completes topology error identification, state estimation and bad data detection.

9. The method for fault identification and detection in a distribution network based on a graph computation model according to claim 1, characterized in that, The topology analysis in S3 specifically employs an incremental graph update algorithm to update the topology in real time in response to changes. It then integrates measurement data and state estimation, and identifies topology errors based on graph connectivity analysis and power flow verification. The topology changes include switching operations and distributed power supply access / exit, and the topology errors include false alarms in switch status and wiring errors.

10. The method for fault identification and detection in a distribution network based on a graph computation model according to claim 1, characterized in that, The fault assessment and processing in S3 includes fault location, fault type identification and fault impact analysis. The fault location is specifically based on graph traversal and graph neural network, which integrates real-time measurements to locate the fault section and fault point. The fault type identification specifically involves extracting the electrical features of graph nodes / edges, and then automatically identifying short circuit, grounding, open circuit, and arcing faults based on deep learning; The fault impact analysis is specifically based on graph connectivity analysis and power flow calculation to generate a fault impact list, including users experiencing power outages and affected power sources.

Images