A method and system for early warning of power distribution equipment status based on edge computing

Abstract

This invention relates to the field of power distribution equipment condition monitoring technology, and specifically to a power distribution equipment condition early warning method and system based on edge computing. The method includes: an edge computing unit loading and running a dynamic digital twin model for each downlink connected power distribution equipment. The model uses real-time collected power distribution equipment operating data as driving input, and calculates and outputs a twin state vector representing the ideal operating state of the power distribution equipment. The power distribution equipment operating data and the twin state vector are subtracted in real-time to obtain a twin-residual vector. Causal relationship analysis is then performed on the time series of this vector to construct a residual causal graph. The method determines whether a preset abnormal causal propagation chain appears in the residual causal graph. After the determination, the fault root cause, abnormal causal propagation chain, and early warning information are encapsulated into a data packet and reported as an event. This invention enables rapid and accurate localized early warning of power distribution equipment fault root causes.

Description

Technical Field

[0001] This invention relates to the field of power distribution equipment condition detection technology, and more specifically, to a power distribution equipment condition early warning method and system based on edge computing. Background Technology

[0002] Driven by the construction of new power systems, the digital and intelligent transformation of the power grid is accelerating. As the bridge connecting the power grid and end users, the reliability and health of distribution equipment are crucial to ensuring regional power supply quality. With the integration of numerous distributed renewable energy sources and electric vehicle charging stations, the operating conditions of distribution equipment are becoming increasingly complex, facing more severe electromagnetic, thermal, and mechanical stress shocks. Therefore, there is an urgent need to develop more advanced condition monitoring and intelligent early warning technologies. Through in-depth analysis and mining of equipment operating data, a shift from traditional passive, planned maintenance to proactive, predictive maintenance can be achieved, thereby improving the asset management efficiency of distribution equipment and the overall operational resilience of the power grid. Currently, the condition monitoring of power distribution equipment relies heavily on centralized analysis by the cloud master station. This not only leads to huge network bandwidth pressure and communication delays caused by the massive data transmission, but also significantly reduces the real-time performance of early warning judgments. In addition, existing analysis models are mostly based on static thresholds, making it difficult to deeply explore the hidden dynamic correlations between data. They often confuse normal operating condition fluctuations with early signs of faults, resulting in insufficient accuracy and reliability of early warning results. Furthermore, they cannot effectively guide maintenance personnel to investigate the root causes after an alarm occurs, and lack the ability to gain insight into the fault evolution mechanism. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for early warning of the status of power distribution equipment based on edge computing. By performing twin-residual causal analysis at the edge, it achieves rapid and accurate localized early warning of the root causes of power distribution equipment faults.

[0004] This invention is achieved through the following technical solution:

[0005] A method for early warning of power distribution equipment status based on edge computing, the steps of which include:

[0006] In its internal containerized environment, the edge computing unit loads and runs a dynamic digital twin model for each downlink connected power distribution device. Based on real-time collected power distribution device operation data as driving input, the dynamic digital twin model calculates and outputs a twin state vector representing the ideal operating state of the power distribution device.

[0007] The operating data of the power distribution equipment is subtracted from the twin state vector in real time in local memory to obtain the twin-residual vector. Causal relationship analysis is performed on the time series of the twin-residual vector to construct a residual causal graph that represents the intrinsic driving relationship of the residuals in each dimension.

[0008] Monitor the residual cause-effect graph and determine whether a preset abnormal cause-effect propagation chain appears in the residual cause-effect graph. If so, directly determine the occurrence of the warning event locally and identify the root node of the abnormal cause-effect propagation chain as the root cause of the fault.

[0009] After completing the local determination, the edge computing unit encapsulates the root cause of the fault, the abnormal causal propagation chain, and the early warning information into a data packet, and selectively reports the event to the main station or cloud platform through its uplink communication module.

[0010] Optionally, the dynamic digital twin model includes:

[0011] The physical mechanism module contains a benchmark mathematical model that characterizes the ideal physical properties of the power distribution equipment. The benchmark mathematical model solves for the theoretical basis state vector based on the input power distribution equipment operation data.

[0012] The data-driven compensation module is specifically a pre-trained neural network model. The input is set to the operating data of the power distribution equipment, and the output is set to the compensation vector. It is used to fit and compensate for the difference between the physical mechanism module and the actual operating state of the physical equipment.

[0013] The state fusion module is used to perform vector weighted fusion of the theoretical basis state vector output by the physical mechanism module and the compensation vector output by the data-driven compensation module to generate and output the twin state vector.

[0014] Optionally, the process of solving the twin-residual vector is as follows:

[0015] At the preset calculation time point, the edge computing unit synchronously locks the power distribution equipment operation data to form a reference data vector;

[0016] The edge computing unit inputs the power distribution equipment operation data into the dynamic digital twin model, calculates and outputs the twin state vector corresponding to the power distribution equipment operation data in timestamp;

[0017] Mandatory data structure validation is performed on the power distribution equipment operation data and twin state vectors to ensure that the number of data dimensions, their order of arrangement, and the data types contained in the two are completely consistent.

[0018] Subtract the corresponding data value in the same dimension from the data value in the twin state vector for each dimension of the power distribution equipment operation data, and combine all the differences obtained by subtraction one by one into a new vector in the original dimensional order, namely the twin-residual vector.

[0019] Optionally, the dynamic digital twin model is further provided with adaptive update logic, which is as follows:

[0020] Using the twin-residual vector as an error feedback signal, and through backpropagation or Kalman filtering algorithms, the parameters of the neural network model in the data-driven compensation module are fine-tuned online to achieve adaptive tracking of the aging process of individual power distribution equipment by the digital twin.

[0021] Optionally, the causal relationship analysis of the time series of the twin-residual vectors specifically includes:

[0022] From the time series of the twin-residual vector, extract a data window with a preset time length as an analysis sample;

[0023] For any two time series of different dimensions within the data window, pairwise causal relationship quantification is performed using the transfer entropy algorithm to determine the unidirectional driving strength of one dimension sequence on the other dimension sequence.

[0024] Each dimension of the twin-residual vector is defined as a node in the residual causal graph. If the unidirectional driving strength between two dimensions exceeds a preset threshold, a directional edge is constructed between the corresponding two nodes to represent the driving relationship.

[0025] Iterate through all dimension pairs and repeat the above process until a residual causal graph representing the intrinsic driving relationship of the residuals in each dimension is generated.

[0026] Optionally, the specific logic for the step of monitoring the residual cause-effect graph and determining whether a preset abnormal cause-effect propagation chain has occurred is as follows:

[0027] The edge computing unit has a pre-set fault mode library, which includes templates of abnormal causal propagation chains corresponding to multiple known fault modes.

[0028] The residual cause-effect graph is compared with the templates of each abnormal cause-effect propagation chain in the failure mode library using graph matching.

[0029] If the nodes, edges, and topology of the subgraph in the residual cause-effect graph completely match any abnormal cause-effect propagation chain template in the fault mode library, then it is determined that the preset abnormal cause-effect propagation chain has appeared.

[0030] Optionally, in the successfully matched abnormal causal propagation chain, all nodes in the subgraph are traversed, and the starting node that is not pointed to by any other node in the abnormal causal propagation chain is identified. The residual dimension represented by the starting node is confirmed as the root cause of the fault.

[0031] Optionally, the encapsulation is a data packet, specifically:

[0032] Extract and combine the following data elements: root cause of the fault, successfully matched abnormal causal propagation chain, warning level associated with abnormal causal propagation chain, and timestamp of the event occurrence.

[0033] All data elements are serialized into structured data objects to form an event data packet.

[0034] Optionally, the step of selectively reporting events to the main station or cloud platform via its uplink communication module has the following specific logic:

[0035] Configure reporting strategies in edge computing units to bind different warning levels with different reporting behaviors;

[0036] After generating the event data packet, query the reporting strategy based on the warning level it contains;

[0037] If the policy specifies that the current level is to report immediately, the data packet will be sent immediately through the uplink communication module; if the policy specifies that it is not urgent, the data packet will be temporarily stored locally and awaited for periodic batch reporting.

[0038] The edge computing-based power distribution equipment status early warning system includes:

[0039] Processor and local memory;

[0040] Downlink data interface, used to collect real-time operating data of power distribution equipment connected to the downlink;

[0041] The twin computing module is configured by the processor to: load and run a dynamic digital twin model corresponding to the power distribution equipment in its internal containerized environment, and calculate and output a twin state vector representing the ideal operating state of the power distribution equipment based on the power distribution equipment operating data as the driving input.

[0042] The causal analysis module, configured by the processor, is as follows:

[0043] In the local memory, the operating data of the power distribution equipment is subtracted from the twin state vector in real time to obtain the twin-residual vector. Causal relationship analysis is then performed on the time series of the twin-residual vector to construct a residual causal graph that characterizes the intrinsic driving relationship of the residuals in each dimension.

[0044] The early warning determination module is configured by the processor to: monitor the residual cause-effect graph, and when a preset abnormal cause-effect propagation chain appears in the determination graph, determine that an early warning event has occurred locally, and identify the root node of the abnormal cause-effect propagation chain as the root cause of the fault.

[0045] The uplink communication module is configured by the processor to selectively report data packets containing the root cause of the fault, the abnormal causal propagation chain, and early warning information to the main station or cloud platform after completing local determination.

[0046] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0047] This invention, on the one hand, deeply integrates dynamic digital twin models and causal analysis algorithms into edge computing units, enabling local real-time diagnosis of power distribution equipment status. This significantly reduces reliance on cloud communication and data transmission costs, while ensuring ultra-low latency and high timeliness in early warning judgments. On the other hand, by innovatively performing causal analysis on twin-residual vectors, this invention can penetrate data appearances, accurately identifying the initial root cause of faults and clearly depicting the propagation path of anomalies. This transforms passive threshold alarms into proactive, interpretable causal early warnings, significantly improving the efficiency and accuracy of fault diagnosis. Attached Figure Description

[0048] Figure 1 A schematic diagram illustrating the principle of the power distribution equipment status early warning system based on edge computing provided by the present invention;

[0049] Figure 2 This is a flowchart illustrating the edge computing-based power distribution equipment status early warning method provided by the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0051] like Figure 1 As shown, this invention provides a power distribution equipment status early warning system and method based on edge computing. Unlike traditional models that rely on centralized analysis from a cloud-based master station, this invention decentralizes core intelligent computing capabilities to the power distribution site, closer to the data source. Its core carrier is a high-performance, compact power distribution energy-saving communication and control unit, which is physically deployed directly next to the power distribution substation, ring main unit, or power distribution transformer. Through its highly integrated internal functional modules, it can independently complete the entire intelligent early warning process, from data acquisition, twin modeling, causal analysis, fault determination to selective reporting, thereby achieving rapid, accurate, and interpretable localized diagnosis of early-stage faults in power distribution equipment.

[0052] The core of the power distribution status early warning system in this embodiment is a power distribution conservation communication and control unit. In terms of hardware, this unit integrates an industrial-grade processor with a high-performance multi-core ARM architecture and sufficient local memory, including at least 4GB of DDR4 RAM for high-speed computing and no less than 64GB of eMMC flash memory for storing the operating system, applications, and historical data.

[0053] Its hardware interface configuration is as follows:

[0054] Downlink Data Interface: To connect to power distribution equipment from different manufacturers and eras, the downlink data interface features multi-protocol acquisition capabilities. Its physical layer includes multiple RS-485 / RS-232 serial ports and at least two industrial Ethernet ports. The protocol layer incorporates parsing libraries for Modbus RTU / TCP, IEC 60870-5-101 / 104, CDT, and some proprietary protocols from power distribution equipment manufacturers. This allows the unit to simultaneously acquire operational data from various devices such as distribution transformers, switchgear, DC power supplies, and environmental monitoring instruments.

[0055] Uplink Communication Module: The uplink communication module is responsible for communicating with the main station or cloud platform. It adopts a dual redundancy design of wired + wireless, including a gigabit Ethernet uplink interface and a built-in wireless communication module that supports 4G / 5G full network compatibility. This ensures that the optimal communication link can always be selected for data reporting in different field environments, reflecting its cost-effective and highly reliable design philosophy.

[0056] This embodiment pre-installs a lightweight embedded Linux operating system at the software level. The operating system kernel has been trimmed and hardened to meet the stability and security requirements of industrial environments. On top of the operating system, this embodiment deploys a Docker containerized environment. Containerization technology is one of the key supports for this invention, packaging different applications and their dependencies into independent, portable containers. This embodiment includes the following core containerized applications:

[0057] Twin Computing Module Container: Specifically responsible for loading and running dynamic digital twin models.

[0058] Causal analysis module container: responsible for performing the calculation of twin-residual vectors and subsequent causal graph construction.

[0059] Early warning and judgment module container: responsible for monitoring cause-effect graphs, matching fault modes and identifying root causes.

[0060] The container-based modular architecture enables resource isolation and independent lifecycle management between various functional modules, greatly improving the stability and maintainability of the system. This allows the updating and deployment of algorithm models to be as convenient as plugging and unplugging a USB drive, without affecting other running services.

[0061] like Figure 2 As shown, this embodiment takes the monitoring of dry-type distribution transformers by a communication control unit deployed in a 10kV substation as an example to illustrate the entire execution process of the method in this embodiment.

[0062] Step 1: Construction and Operation of Dynamic Digital Twin Model

[0063] During the system initialization phase, the twin computing module of the communication control unit loads a dynamic digital twin model for the monitored dry-type transformer. In this embodiment, the dynamic digital twin model is not a single algorithm, but a complex model comprising multiple sub-modules, with the specific structure as follows:

[0064] The physical mechanism module contains a benchmark mathematical model that characterizes the ideal physical properties of a dry-type transformer. Understandably, this embodiment is based on the IEC 60076-7 thermal model. The model uses parameters such as transformer load current, winding resistance, and core loss collected in real-time by the downlink interface. By solving thermodynamic differential equations, it calculates the theoretical fundamental state vector of the winding and core under the current load. This theoretical fundamental state vector mainly includes the theoretical temperature value.

[0065] Data-driven compensation module: Since the physical model is idealized, it cannot describe individual differences caused by factors such as manufacturing processes, material aging, and environmental changes. Therefore, this embodiment sets up a pre-trained neural network model as a compensation module. Specifically, the neural network model used in this embodiment is a Long Short-Term Memory (LSTM) network. The input of the LSTM model is also real-time operating data of the power distribution equipment (such as load current, ambient temperature, humidity, vibration signals, etc.), and its output is a compensation vector. The significance of the compensation vector is to fit and compensate for the systematic deviation between the theoretical calculation values ​​of the physical mechanism module and the actual operating state of the equipment.

[0066] State fusion module: It receives the outputs from the two modules mentioned above, namely the theoretical basis state vector and the compensation vector. It performs vector addition or weighted averaging on the two through a weighted fusion algorithm, and finally generates and outputs a twin state vector that highly approximates the most ideal and healthiest state of the power distribution equipment under the current operating conditions.

[0067] The dynamic digital twin model is encapsulated in a separate Docker container and runs in real time, driven by the processor, continuously outputting twin state vectors that correspond one-to-one with the real devices.

[0068] Step 2: Solving for the twin-residual vector

[0069] At each preset calculation time point, the communication control unit first performs a snapshot lock on the raw, unprocessed power distribution equipment operation data collected in real time through the downlink data interface in its local memory, forming a time-fixed reference data vector.

[0070] The locked baseline data vector is used as the driving input and fed into the running dynamic digital twin model. The dynamic digital twin model calculates and outputs the corresponding twin state vector at the same timestamp.

[0071] Before the subtraction, the processor performs a mandatory data structure check, specifically verifying that the number of dimensions, the order of the dimensions, and the data type of each dimension of the two vectors, the base data vector and the twin state vector, are completely consistent.

[0072] After successful verification, vector subtraction is performed in local high-speed memory. The value of each dimension in the baseline data vector is subtracted from the value of the same dimension in the twin state vector. For example, the temperature residual = 85.1 - 82.5 = 2.6°C. All these differences are recombine in their original order to form a new vector, the twin-residual vector. The non-zero values ​​in the twin-residual vector represent the degree to which the operating state of the power distribution equipment deviates from its ideal baseline. The solved twin-residual vector is then stored in a FIFO memory queue, forming a time series of the twin-residual vector.

[0073] Furthermore, the dynamic digital twin model also possesses adaptive update logic. The calculated twin-residual vector is used as a feedback signal, and through the backpropagation algorithm, the network weights of the data-driven compensation module are fine-tuned online and in small batches. This enables the digital twin to automatically track and adapt to the performance aging and characteristic drift of physical equipment due to long-term operation, ensuring the accuracy of the model throughout the entire lifecycle of the equipment.

[0074] Step 3: Causal relationship analysis and residual cause-effect graph construction

[0075] From the generated twin-residual vector time series queue, a fixed-length sliding data window is extracted. For example, 300 twin-residual vector data points from the past 5 minutes are extracted as samples for a causal analysis. Pairwise dimensional analysis is performed on the 300 sample data windows. First, the time series of temperature residuals and vibration residuals are extracted. Then, the transfer entropy algorithm is used for calculation. Transfer entropy quantitatively measures the extent to which past information of one time series reduces the uncertainty of the future of another time series, providing a unidirectional causal strength value from vibration to temperature. Understandably, this embodiment chooses transfer entropy specifically because it can capture nonlinear relationships and has a wider range of applications.

[0076] In memory, create a graph structure. Each dimension of the residual vector becomes a node in the graph. Then, based on the calculation results of the previous step, set up directed edges. The process of setting up directed edges is as follows: if the calculated transfer entropy value from the vibration residual to the temperature residual exceeds a preset engineering threshold, create an edge in the graph pointing from the vibration node to the temperature node, i.e., a directed edge.

[0077] Causal relationship analysis and graph structure creation are repeatedly performed on all ordered dimensions in the data window until the traversal is complete. Ultimately, a complete residual causal graph is formed in memory. The nodes and directed edges in the residual causal graph can display various minor anomalies in the current state of the power distribution equipment.

[0078] Step 4: Monitoring, Judgment, and Root Cause Identification

[0079] This embodiment pre-configures a fault mode library. The fault mode library stores multiple known abnormal causal propagation chain templates corresponding to specific faults, such as winding loosening, insulation aging, and core overheating, in a structured manner. Understandably, the winding loosening fault template in this embodiment is defined as a path with a specific topological structure: vibration residual → noise residual → partial discharge residual.

[0080] The early warning determination module continuously performs graph matching or subgraph isomorphic search on the real-time residual cause-effect graph against all templates in the library. Once a structure in the real-time residual cause-effect graph is found to perfectly match a template in the library, the system directly determines that an early warning event has occurred locally. After determining the early warning, the system immediately searches for the root node in the subgraph of the anomaly causal propagation chain of the matching event. The root node is defined as a node with no in-degree edges in the propagation chain. In the example of the loose winding mentioned above, the vibration residual node is this root node because it drives the subsequent anomaly, while it itself is not driven by other nodes in the chain. Therefore, the system identifies the vibration anomaly as the root cause of this early warning event.

[0081] After completing the judgment and root cause identification locally, the system encapsulates all key information of this warning event into a data packet according to a predefined JSON format. Before sending the data packet, the system queries the locally configured reporting policy. This policy binds different warning levels with reporting behaviors:

[0082] Level 1 Warning (Severe): The policy stipulates that the report should be immediately sent via the 4G / 5G wireless module to ensure that the main station receives it as quickly as possible.

[0083] Level 2 Warning (Important): The policy specifies reporting via wired Ethernet. If Ethernet is interrupted, it will switch to 4G / 5G.

[0084] Level 3 Early Warning (General): The strategy stipulates that data packets are temporarily stored locally and reported in batches along with periodic heartbeat packets at the next hour to save communication bandwidth.

[0085] Based on the above-mentioned early warning method, this embodiment provides a highly interpretable diagnostic report that includes the root cause of the fault and the abnormal propagation path. It can also intelligently select the reporting time and channel according to the urgency of the event, thus perfectly realizing the localization, intelligence and conservation of communication resources of status early warning.

[0086] In a preferred embodiment of the present invention, this embodiment runs in parallel with the original decision logic based on fault mode library matching, aiming to overcome the limitation of the original logic that it can only identify known fault modes. The specific implementation process is as follows:

[0087] The core idea of ​​this embodiment is that in a healthy state, the causal relationship network between the variables within any physical system, i.e., the residual causal graph, will exhibit a relatively stable or regularly changing topological structure. When an early anomaly occurs in the device, even if the anomaly pattern has not yet been entered into the fault database, it will inevitably cause a significant change in the structure of this causal relationship network. This embodiment specifically achieves early warning of unknown faults by quantifying and monitoring this structural change, and the specific implementation includes the following stages:

[0088] Phase 1: This phase is the initial stage after the deployment of the communication control unit. Its goal is to establish a normal reference standard for subsequent anomaly detection.

[0089] The communication control unit continuously executes steps one through three of the above embodiments to dynamically generate time series of residual causal graphs. For each causal graph in the time series, the early warning judgment module additionally calculates a set of core quantitative indicators to describe its network topology, specifically including global indicators and node-level indicators. The global indicators include: graph density, the ratio of the actual number of directed edges in the graph to the theoretically maximum number of edges, reflecting the overall tightness of the correlation between each residual dimension; and the average clustering coefficient, which measures the degree of clustering of nodes in the graph, i.e., the tightness of the connections between a node's neighboring nodes. The node-level indicators include: node out-degree, calculating the number of directed edges emitted by each node (i.e., each residual dimension), representing the driving ability of that dimension on other dimensions; node in-degree, calculating the number of directed edges pointing to each node, representing the degree to which that dimension is influenced by other dimensions; and node betweenness centrality, measuring the frequency with which a node appears on the shortest path between any two other nodes in the graph, representing the importance of that dimension as a bridge in the causal transmission network. Statistical baseline modeling involves collecting time-series data of all the aforementioned topological indicators within a set learning period and establishing a statistically significant normal baseline for each indicator. This baseline can be a Gaussian distribution model, containing the mean and standard deviation of each indicator under healthy conditions. The mean and standard deviation parameters are permanently stored in the local memory of the communication control unit, forming a topological fingerprint of the device's health status.

[0090] Phase Two: After completing self-learning and entering normal operation, the system will execute the following two decision logics in parallel:

[0091] Logic A: That is, the original implementation of step four, performing graph matching based on the fault mode library to diagnose known faults.

[0092] Logic B: In this preferred embodiment, anomaly detection based on topology baseline is performed to discover unknown anomalies.

[0093] Specifically, the process of logic B is as follows:

[0094] For the latest generated residual causal graph, the system calculates in real time a complete set of topological indices that are exactly the same as those in the learning phase.

[0095] Each indicator value calculated in real time is compared with its mean and standard deviation stored in the baseline to calculate the degree of deviation from the normal state. In the optional calculation process of this embodiment, the Z-score is calculated: Z = (current value – mean) / standard deviation. The larger the absolute value of the Z-score, the further the indicator deviates from the normal state.

[0096] The Z-scores of all indicators are combined to form an overall topology anomaly score. When this total score exceeds a preset judgment threshold, the system will determine that an anomaly event of an unknown pattern has occurred, even if logic A does not match any known fault template.

[0097] By introducing this optimized implementation, when logic A is triggered, the system issues a high-confidence, explicit fault warning (e.g., loose winding) and provides the root cause defined by the template. When logic A is not triggered, but logic B triggers an alarm, the system issues an unknown anomaly warning. In this case, although a specific fault name cannot be given, it can analyze which nodes' topology metrics (such as out-degree and centrality) deviate the most, indicating to maintenance personnel which dimension(s) of residuals exhibit the most abnormal causal driving pattern. This also provides valuable clues for manual troubleshooting.

[0098] This optimized embodiment frees the system from absolute reliance on the completeness of expert knowledge. For unprecedented and complex fault chains caused by factors such as equipment aging, sudden changes in operating conditions, and external interference, the system possesses the ability to detect them, achieving a leap from a closed-world diagnostic system to an open-world intelligent system with self-learning and discovery capabilities.

[0099] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early warning of power distribution equipment status based on edge computing, characterized in that, The steps of this method include: In its internal containerized environment, the edge computing unit loads and runs a dynamic digital twin model for each downlink connected power distribution device. Based on real-time collected power distribution device operation data as driving input, the dynamic digital twin model calculates and outputs a twin state vector representing the ideal operating state of the power distribution device. The operating data of the power distribution equipment is subtracted from the twin state vector in real time in local memory to obtain the twin-residual vector. Causal relationship analysis is performed on the time series of the twin-residual vector to construct a residual causal graph that represents the intrinsic driving relationship of the residuals in each dimension. Monitor the residual cause-effect graph and determine whether a preset abnormal cause-effect propagation chain appears in the residual cause-effect graph. If so, directly determine the occurrence of the warning event locally and identify the root node of the abnormal cause-effect propagation chain as the root cause of the fault. After completing the local determination, the edge computing unit encapsulates the fault root cause, abnormal causal propagation chain and early warning information into a data packet, and selectively reports the event to the main station or cloud platform through its uplink communication module. The causal relationship analysis of the time series of twin-residual vectors specifically includes: From the time series of the twin-residual vector, extract a data window with a preset time length as an analysis sample; For any two time series of different dimensions within the data window, pairwise causal relationship quantification is performed using the transfer entropy algorithm to determine the unidirectional driving strength of one dimension sequence on the other dimension sequence. Each dimension of the twin-residual vector is defined as a node in the residual causal graph. If the unidirectional driving strength between two dimensions exceeds a preset threshold, a directional edge is constructed between the corresponding two nodes to represent the driving relationship. Iterate through all dimension pairs and repeat the above process until a residual causal graph representing the intrinsic driving relationship of the residuals in each dimension is generated.

2. The method for early warning of power distribution equipment status based on edge computing according to claim 1, characterized in that, The dynamic digital twin model includes: The physical mechanism module contains a benchmark mathematical model that characterizes the ideal physical properties of the power distribution equipment. The benchmark mathematical model solves for the theoretical basis state vector based on the input power distribution equipment operation data. The data-driven compensation module is specifically a pre-trained neural network model. The input is set to the operating data of the power distribution equipment, and the output is set to the compensation vector. It is used to fit and compensate for the difference between the physical mechanism module and the actual operating state of the physical equipment. The state fusion module is used to perform vector weighted fusion of the theoretical basis state vector output by the physical mechanism module and the compensation vector output by the data-driven compensation module to generate and output the twin state vector.

3. The method for early warning of power distribution equipment status based on edge computing according to claim 2, characterized in that, The process of solving the twin-residual vector is as follows: At the preset calculation time point, the edge computing unit synchronously locks the power distribution equipment operation data to form a reference data vector; The edge computing unit inputs the power distribution equipment operation data into the dynamic digital twin model, calculates and outputs the twin state vector corresponding to the power distribution equipment operation data in timestamp; Mandatory data structure validation is performed on the power distribution equipment operation data and twin state vectors to ensure that the number of data dimensions, their order of arrangement, and the data types contained in the two are completely consistent. Subtract the corresponding data value in the same dimension from the data value in the twin state vector for each dimension of the power distribution equipment operation data, and combine all the differences obtained by subtraction one by one into a new vector in the original dimensional order, namely the twin-residual vector.

4. The method for early warning of power distribution equipment status based on edge computing according to claim 3, characterized in that, The dynamic digital twin model also includes adaptive update logic, specifically: Using the twin-residual vector as an error feedback signal, and through backpropagation or Kalman filtering algorithms, the parameters of the neural network model in the data-driven compensation module are fine-tuned online to achieve adaptive tracking of the aging process of individual power distribution equipment by the digital twin.

5. The method for early warning of power distribution equipment status based on edge computing according to claim 4, characterized in that, The specific logic for monitoring the residual cause-effect graph and determining whether a preset abnormal cause-effect propagation chain has occurred is as follows: The edge computing unit has a pre-set fault mode library, which includes templates of abnormal causal propagation chains corresponding to multiple known fault modes. The residual cause-effect graph is compared with the templates of each abnormal cause-effect propagation chain in the failure mode library using graph matching. If the nodes, edges, and topology of the subgraph in the residual cause-effect graph completely match any abnormal cause-effect propagation chain template in the fault mode library, then it is determined that the preset abnormal cause-effect propagation chain has appeared.

6. The method for early warning of power distribution equipment status based on edge computing according to claim 5, characterized in that, In the abnormal causal propagation chain that is determined to be successfully matched, all nodes in the subgraph are traversed, and the starting node that is not pointed to by any other node in the abnormal causal propagation chain is identified. The residual dimension represented by the starting node is confirmed as the root cause of the fault.

7. The method for early warning of power distribution equipment status based on edge computing according to claim 6, characterized in that, The encapsulation into a data packet specifically refers to: Extract and combine the following data elements: root cause of the fault, successfully matched abnormal causal propagation chain, warning level associated with abnormal causal propagation chain, and timestamp of the event occurrence. All data elements are serialized into structured data objects to form an event data packet.

8. The method for early warning of power distribution equipment status based on edge computing according to claim 7, characterized in that, The specific logic of the step of selectively reporting events to the main station or cloud platform through its uplink communication module is as follows: Configure reporting strategies in edge computing units to bind different warning levels with different reporting behaviors; After generating the event data packet, query the reporting strategy based on the warning level it contains; If the policy specifies that the current level is to report immediately, then the data packet will be sent immediately through the uplink communication module; If the policy specifies that it is not urgent, the data packets will be temporarily stored locally and await periodic batch reporting.

9. A power distribution equipment status early warning system based on edge computing, characterized in that, include: Processor and local memory; Downlink data interface, used to collect real-time operating data of power distribution equipment connected to the downlink; The twin computing module is configured by the processor to: load and run a dynamic digital twin model corresponding to the power distribution equipment in its internal containerized environment, and calculate and output a twin state vector representing the ideal operating state of the power distribution equipment based on the power distribution equipment operating data as the driving input. The causal analysis module, configured by the processor, is as follows: In the local memory, the operating data of the power distribution equipment is subtracted from the twin state vector in real time to obtain the twin-residual vector. Causal relationship analysis is then performed on the time series of the twin-residual vector to construct a residual causal graph that characterizes the intrinsic driving relationship of the residuals in each dimension. The early warning determination module is configured by the processor to: monitor the residual cause-effect graph, and when a preset abnormal cause-effect propagation chain appears in the determination graph, determine that an early warning event has occurred locally, and identify the root node of the abnormal cause-effect propagation chain as the root cause of the fault. The uplink communication module is configured by the processor to selectively report data packets containing the root cause of the fault, the abnormal causal propagation chain, and early warning information to the main station or cloud platform after completing local determination. The causal relationship analysis of the time series of twin-residual vectors specifically includes: From the time series of the twin-residual vector, extract a data window with a preset time length as an analysis sample; For any two time series of different dimensions within the data window, pairwise causal relationship quantification is performed using the transfer entropy algorithm to determine the unidirectional driving strength of one dimension sequence on the other dimension sequence. Each dimension of the twin-residual vector is defined as a node in the residual causal graph. If the unidirectional driving strength between two dimensions exceeds a preset threshold, a directional edge is constructed between the corresponding two nodes to represent the driving relationship. Iterate through all dimension pairs and repeat the above process until a residual causal graph representing the intrinsic driving relationship of the residuals in each dimension is generated.

Images