Distributed FA cooperative control method for edge computing nodes of distribution network terminal

Abstract

The invention provides a distributed FA cooperative control method for edge computing nodes of a distribution network terminal, and the method comprises the steps: synchronously collecting high-frequency electrical and partial discharge signals through an edge computing terminal, carrying out the multi-dimensional processing, and forming a fault feature vector; each edge node monitors a fault in real time by using a lightweight algorithm, the fault is broadcasted to an adjacent node after being subjected to Hilbert-Huang transform feature enhancement, accurate positioning is realized by adopting multi-node cross validation, and edge cloud collaboration is linked during complex topology. And after the fault is positioned, starting dual verification, sending a tripping instruction through an encrypted GOOSE protocol, searching an optimal recovery path based on a mixed integer linear programming model, and preferentially utilizing a distributed power supply to reversely supply power. A control strategy is dynamically optimized through federated learning and an MADDPG algorithm, and a digital twinborn model is trained to improve adaptability. According to the invention, the problems of slow response, low positioning precision and the like of the traditional centralized FA are effectively solved, and the fault processing efficiency, reliability and intelligent level of the distribution network are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power distribution network control technology, specifically to a distributed FA collaborative control method for edge computing nodes in power distribution networks. Background Technology

[0002] With the deepening of smart grid construction, the distribution network, as a key link connecting the power system and users, faces higher requirements for operational reliability and intelligence. Traditional distribution network feeder automation (FA) systems mainly adopt a centralized control mode, uploading field data to a central control station for unified processing before issuing control commands to handle faults. However, this mode has revealed many technical bottlenecks in practical applications: First, centralized control suffers from significant response delays. Traditional power distribution (FA) systems require transmitting electrical quantity data (such as current and voltage traveling waves) distributed across a wide-area distribution network to a central node via a communication network. After data processing and fault analysis, control commands are then issued. This process is limited by communication bandwidth and the computing load of the central node, resulting in fault handling times typically ranging from hundreds of milliseconds to seconds, which is insufficient to meet the reliability requirements of modern distribution networks. For example, in densely populated urban distribution areas, a single power outage can affect thousands of users. Delayed power restoration increases the average outage time for users, severely impacting the quality of power service.

[0003] Secondly, the fault location accuracy is insufficient and blind spots exist. Traditional single-end traveling wave location methods rely on the assumption of wave velocity and the precise measurement of the arrival time of the fault traveling wave. However, in actual distribution networks, factors such as the dispersion of line parameters and the reflection / refraction of traveling waves caused by distributed power sources can amplify the location error. For long overhead lines, the location error is even greater, requiring manual segment-by-segment inspection, which greatly increases the time required for fault handling. In addition, in complex topologies such as multi-branch and ring networks, single-end location cannot distinguish faulty sections, creating location blind spots.

[0004] Furthermore, the reliability of the communication network restricts system performance. Centralized FA (Feature Provider) has extremely high requirements for the real-time performance and stability of the communication network. When a communication link fails (such as a broken fiber optic cable or wireless signal interference), the central node cannot obtain field data, causing the FA function to fail. In severe weather conditions such as thunderstorms and freezing, the probability of distribution network communication interruption increases.

[0005] Furthermore, the integration of distributed generation (DG) presents challenges for coordinated control. With the large-scale integration of distributed generation (DG) sources such as photovoltaics and wind power, the distribution network topology is shifting from a traditional radial structure to a multi-source grid-connected structure. Traditional power distribution (FA) systems lack the ability to perceive and coordinate the output characteristics of DG. During faults, reverse power supply from DG may lead to maloperation of protection systems, or DG may not be effectively utilized for load restoration after a fault.

[0006] Finally, there is insufficient capacity for processing massive amounts of data. High-frequency traveling wave signals (sampling rate ≥ 1MHz) and partial discharge signals generated during distribution network faults contain rich fault characteristics. However, traditional centralized systems, limited by transmission bandwidth and storage capacity, typically only collect power frequency data (50Hz), leading to the loss of fault characteristics. For example, for high-resistance grounding faults, their transient characteristics are mainly concentrated in the 10-100kHz frequency band. Traditional systems, due to insufficient sampling rate, cannot effectively identify these faults, causing them to be misjudged as normal operating conditions.

[0007] To address the aforementioned issues, the introduction of edge computing technology offers a new solution for power distribution network (FA). The local computing capabilities of edge nodes enable real-time data processing, reducing reliance on central nodes. However, building efficient collaboration mechanisms among edge nodes, accurately extracting fault characteristics, and optimizing distributed control strategies remain key technical challenges that urgently need to be overcome. Summary of the Invention

[0008] To address the problems of high response latency, insufficient positioning accuracy, strong communication dependence, weak distributed power supply coordination capabilities, and low efficiency in processing massive amounts of data in traditional centralized power distribution (FA) systems, this invention aims to provide a distributed FA collaborative control method based on edge computing nodes. By constructing a distributed collaborative architecture of edge computing nodes and combining intelligent algorithms and digital twin technology, this method enables rapid identification, accurate location, reliable isolation, and efficient recovery of distribution network faults, thereby improving the operational reliability and intelligence level of distribution networks in complex topologies and scenarios with a high proportion of distributed power supply access.

[0009] To address the aforementioned technical problems, embodiments of the present invention provide the following technical solutions: A distributed FA collaborative control method for edge computing nodes in a distribution network terminal includes the following steps: Edge computing terminals with integrated AI chips are deployed at key nodes of the power distribution network to simultaneously collect raw data on high-frequency electrical signals and partial discharge signals. The raw data is then denoised, enhanced, and standardized to form a multi-dimensional fault feature vector. Each edge node runs a lightweight self-learning algorithm to monitor the line status in real time. After identifying a fault, it extracts the transient waveform, enhances the fault characteristics through Hilbert-Huang transform, compresses and encodes the fault information, and broadcasts it to neighboring nodes. After receiving data from neighboring nodes, the fault location is calculated in real time through topology calculation, blind spots are eliminated through multi-node cross-verification, and edge-cloud collaboration is introduced for complex topologies. After fault location, the two nodes near the fault point initiate dual verification of local policy matching and confirmation by adjacent nodes, send trip command through encrypted GOOSE protocol, embed anti-replay mechanism to ensure command reliability, update topology after action and broadcast isolation result; Based on the updated topology, the optimal recovery path is searched using a mixed-integer linear programming model. Distributed power sources are used to provide reverse power first. Closing commands are sent in sequence and monitored in real time. In case of an anomaly, a backup plan is activated or support is requested. Each node periodically exchanges data and policy parameters, builds a global knowledge base through federated learning, dynamically optimizes control parameters using the MADDPG algorithm, and trains a digital twin model to simulate fault scenarios.

[0010] Optionally, the denoising, feature enhancement, and standardization processing of the original data to form a multi-dimensional fault feature vector specifically involves: A dual-redundant time synchronization mechanism is constructed by using the BeiDou satellite time synchronization module and a local temperature-controlled crystal oscillator. When the difference between the primary and backup clocks exceeds the threshold for three consecutive sampling cycles, a switching is triggered and a timestamp is added to the signal. The initial wave velocity was calculated based on the IEC 60641 standard, and the actual wave velocity was fitted using the least squares method with fault data. For high-frequency traveling waves, db6 wavelet decomposition and adaptive threshold denoising are used, and power frequency quantities are processed by moving average filtering combined with Kalman filtering. Eigenmode functions are generated through empirical mode decomposition. After extracting components and reconstructing the waveform, parameters such as peak fault current, harmonic content, and wavelet packet energy entropy are extracted from the time domain, frequency domain, and joint time-frequency domain. Feature vectors are generated through Z-Score standardization and principal component analysis for dimensionality reduction. These vectors are then compared with typical patterns in the local knowledge base using cosine similarity. Features with a similarity ≥ 0.9 are considered valid. These vectors are stored in a lightweight database with timestamp indexes and fault type labels.

[0011] Optionally, after identifying the fault, the transient waveform is extracted, the fault characteristics are enhanced through Hilbert-Huang transform, and after compression encoding, node information is added and broadcast to neighboring nodes, including: Lightweight neural networks are deployed on edge nodes to analyze standardized fault feature vectors in real time. When the cosine similarity between the fault feature vector and the typical pattern in the local knowledge base is ≥0.9 and lasts for 2 sampling periods, it is determined to be a line fault. After identifying the fault, the transient waveform of the fault is extracted, and the waveform is decomposed into time and frequency using Hilbert-Huang transform to enhance the transient characteristic components of the fault. Wavelet packet compression algorithm is used to compress and encode the waveform, and node ID, timestamp and fault type label are added, and broadcast to neighboring nodes through 5G or fiber optic communication.

[0012] Optionally, the step of calculating the fault location through real-time topology calculation, eliminating blind spots through multi-node cross-verification, and introducing edge-cloud collaboration for complex topologies specifically involves: When an edge node receives fault traveling wave data from a neighboring node, it extracts the polarity characteristics of the traveling wave arriving at each node. If the same fault traveling wave has the same polarity in neighboring nodes, it is confirmed as a valid fault traveling wave. If the polarities are opposite, it is determined to be electromagnetic interference or a non-fault traveling wave and is filtered out. When three or more edge nodes are involved in fault location, a set of location equations is constructed based on the arrival time and polarity characteristics of the traveling waves collected by each node. The optimal fault location is solved by the least squares method to control the location error and eliminate single-end or double-end location blind spots. If the distribution network topology is a complex form that is not limited to multiple branch lines, ring network structure or distributed power supply access, the edge node automatically triggers the edge cloud collaboration mechanism to upload real-time topology data and traveling wave characteristics to the edge cloud, and uses the cloud global topology model and optimization algorithm to perform fault location calculation.

[0013] Optionally, the two nodes near the fault point initiate a dual verification process involving local policy matching and confirmation from adjacent nodes, specifically including: Once the fault location is determined, the two edge nodes near the fault point perform similarity matching between the current fault feature vector and typical strategies in the pre-stored fault handling strategy library. When the matching degree is ≥0.9, the local strategy execution process is triggered. At the same time, a fault confirmation request is sent to the adjacent node. After receiving the request, the adjacent node performs real-time analysis of the electrical signals it has collected. If no fault features are detected, it sends back a no-fault signal. When both nodes receive no-fault confirmation from the adjacent node and the local strategy matching is successful, a trip command is sent through the encrypted GOOSE protocol.

[0014] Optionally, the step of searching for the optimal recovery path using a mixed-integer linear programming model specifically includes: A mixed-integer linear programming model is constructed with the objective functions of maximizing the restored load capacity and minimizing the restoration path loss. The constraints include keeping the node voltage deviation within the rated value, the line current not exceeding the rated current, the number of switching operations not exceeding 3, and the output of distributed power sources not exceeding 90% of their rated capacity. When solving the model, the distributed power source closest to the fault point is selected first for reverse power supply, and the switching action sequence is generated according to the closing timing optimization algorithm. After sending the closing command, the voltage and current parameters of each node are monitored in real time. When a voltage over-limit or overcurrent abnormality occurs, the backup recovery plan is immediately activated or support is requested from the edge cloud.

[0015] Optionally, the nodes periodically exchange data and policy parameters to build a global knowledge base through federated learning, specifically including: Each edge node exchanges historical fault data feature vectors and local policy parameters through an encrypted communication link at a preset period. During the exchange, only the model parameter gradients are uploaded, not the original data. The edge cloud uses the FedAvg algorithm to aggregate parameters of each node and update the global knowledge base model. The training data covers fault types such as normal state, short circuit, grounding, and open circuit, and includes operational data under different seasons and load scenarios. The updated global model is distributed to each edge node to replace the local model parameters. At the same time, a digital twin model is trained based on historical fault data to simulate the response characteristics under different fault scenarios and optimize the fault handling strategy library.

[0016] Optionally, the step of dynamically optimizing control parameters using the MADDPG algorithm and training a digital twin model to simulate fault scenarios specifically includes: Each edge node is defined as an independent intelligent agent. The state space includes local fault feature vectors, state parameters of adjacent nodes and global topology matrix. The action space includes trip decision threshold adjustment, closing timing optimization and traveling wave positioning parameter correction. The reward function is constructed into a multi-objective optimization system based on fault handling time, positioning accuracy, and power restoration rate, and the policy gradient is generated by offline training of fault simulation scenarios. The edge cloud uses the MADDPG algorithm to aggregate experience data from various agents, dynamically update the control parameter matrix, and distribute it to the edge nodes. A digital twin model is trained based on historical fault data and real-time operating parameters to simulate transient response characteristics under fault scenarios, including but not limited to three-phase short circuits and single-phase grounding. The fault handling strategy library is optimized to achieve adaptive evolution of control parameters.

[0017] Optionally, the encrypted GOOSE protocol uses the AES-256 algorithm to encrypt and protect the tripping command, and combines HMAC-SHA256 to generate a message authentication code. The anti-replay sequence embedded in the command consists of a nanosecond-level timestamp and a random number. The receiving node verifies the timestamp deviation and that the random number does not appear in the command.

[0018] Optionally, the digital twin model is trained using an LSTM neural network architecture, with inputs including time-domain / frequency-domain features of historical fault data, real-time operating parameters, and a topology matrix, to simulate transient response characteristics of faults, including but not limited to three-phase short circuits, single-phase grounding, and open-circuit faults.

[0019] The beneficial effects of the above-described technical solution of the present invention are as follows: 1. This invention achieves local real-time data processing and distributed collaborative control by deploying edge computing terminals with integrated AI chips at key nodes of the power distribution network, avoiding the data upload delays and central node processing bottlenecks of traditional centralized systems. Each edge node can independently complete fault feature extraction, identification, and transient waveform broadcasting. Combined with lightweight algorithms and distributed collaborative mechanisms, the fault handling cycle is significantly shortened, enabling rapid response throughout the entire process from fault occurrence to isolation and recovery, effectively reducing power outage time due to power distribution network faults and improving power supply continuity.

[0020] 2. This invention employs multi-node polarity feature cross-validation to eliminate blind spots in single-end positioning. Combined with collaborative calculations between the edge and cloud in complex topology scenarios, it achieves accurate fault location determination. This mechanism, through multi-node data interaction and solving the positioning equation set, effectively overcomes the impact of factors such as the dispersion of line parameters and distributed power supply access on positioning accuracy, making fault location results more reliable, providing accurate evidence for rapid fault isolation, and reducing the workload of manual troubleshooting.

[0021] 3. This invention employs a dual verification mechanism—local policy matching at the nearest fault node and confirmation from adjacent nodes—combined with encrypted GOOSE protocol and anti-replay sequence design, to ensure reliable execution and secure transmission of tripping commands. This mechanism effectively avoids malfunctions caused by misjudgments from a single node. Through distributed collaborative verification and encrypted communication protection, it enhances system stability under communication anomalies or electromagnetic interference environments, ensuring the correct operation of distribution network switchgear and reducing the risk of unplanned power outages.

[0022] 4. This invention uses a mixed-integer linear programming model to search for the optimal recovery path, prioritizing the use of distributed power sources for reverse power supply to achieve efficient load recovery after a fault. This strategy fully considers changes in the distribution network topology and the output characteristics of distributed power sources, reducing network losses and improving the economic efficiency of distribution network operation while restoring load capacity. Real-time monitoring and backup scheme switching mechanisms ensure the stability of the recovery process, adapting to the operational needs of modern distribution networks with a high proportion of distributed power sources.

[0023] 5. This invention constructs a global knowledge base through federated learning and dynamically optimizes control parameters using the MADDPG algorithm, enabling the system to possess self-learning and policy evolution capabilities. Each edge node periodically exchanges parameter gradients, and the model is updated by aggregation between the edge and cloud. Combined with digital twin technology, it simulates different fault scenarios, achieving adaptive optimization for new fault characteristics and complex operating conditions. This mechanism allows the system to continuously improve fault handling efficiency and policy accuracy during long-term operation, reducing manual intervention and adapting to the needs of intelligent distribution network development. Attached Figure Description

[0024] Figure 1 This is a flowchart of the distributed FA collaborative control method for the edge computing node of the distribution network terminal according to the present invention.

[0025] Figure 2 This is a flowchart illustrating the method for broadcasting additional node information to adjacent nodes after a fault in the distributed FA collaborative control method for edge computing nodes of distribution network terminals according to the present invention.

[0026] Figure 3 The flowchart of the edge-cloud collaborative implementation method is introduced for the distributed FA collaborative control method of the distribution network terminal edge computing node of the present invention.

[0027] Figure 4 The flowchart illustrates the method for training a digital twin model to simulate fault scenarios in the distributed FA collaborative control method for edge computing nodes of the distribution network terminal according to the present invention. Detailed Implementation

[0028] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0029] Example 1

[0030] like Figure 1 As shown, this invention proposes a distributed FA collaborative control method for edge computing nodes in a distribution network terminal, comprising the following steps: S1. Deploy edge computing terminals with integrated AI chips at key nodes of the power distribution network to simultaneously collect raw data on high-frequency electrical signals and partial discharge signals, and perform noise reduction, feature enhancement and standardization on the raw data to form a multi-dimensional fault feature vector. S2. Each edge node runs a lightweight self-learning algorithm to monitor the line status in real time. After identifying the fault, it extracts the transient waveform, enhances the fault characteristics through Hilbert-Huang transform, and broadcasts the compressed and encoded node information to adjacent nodes. S3. After receiving data from adjacent nodes, the fault location is calculated in real time through topology calculation, and blind spots are eliminated through multi-node cross-verification. Edge-cloud collaboration is introduced for complex topologies. S4. After fault location, the two nodes near the fault point initiate dual verification of local policy matching and confirmation by adjacent nodes, send trip command through encrypted GOOSE protocol, embed anti-replay mechanism to ensure command reliability, update topology and broadcast isolation result after action; S5. Based on the updated topology, search for the optimal recovery path through a mixed integer linear programming model, prioritize the use of distributed power sources for reverse power supply, send closing commands in sequence and monitor in real time, and activate backup schemes or request support in case of abnormality. S6. Each node periodically exchanges data and strategy parameters, builds a global knowledge base through federated learning, dynamically optimizes control parameters using the MADDPG algorithm, and trains a digital twin model to simulate fault scenarios.

[0031] In this embodiment, step S1 involves denoising, feature enhancement, and standardization of the original data to form a multi-dimensional fault feature vector, specifically as follows: A dual-redundant time synchronization mechanism is constructed by using the BeiDou satellite time synchronization module and a local temperature-controlled crystal oscillator. When the difference between the primary and backup clocks exceeds the threshold for three consecutive sampling cycles, a switching is triggered and a timestamp is added to the signal. The initial wave velocity was calculated based on the IEC 60641 standard, and the actual wave velocity was fitted using the least squares method with fault data. For high-frequency traveling waves, db6 wavelet decomposition and adaptive threshold denoising are used, and power frequency quantities are processed by moving average filtering combined with Kalman filtering. Eigenmode functions are generated through empirical mode decomposition. After extracting components and reconstructing the waveform, parameters such as peak fault current, harmonic content, and wavelet packet energy entropy are extracted from the time domain, frequency domain, and joint time-frequency domain. Feature vectors are generated through Z-Score standardization and principal component analysis for dimensionality reduction. These vectors are then compared with typical patterns in the local knowledge base using cosine similarity. Features with a similarity ≥ 0.9 are considered valid. These vectors are stored in a lightweight database with timestamp indexes and fault type labels.

[0032] In this embodiment, as Figure 2 As shown, in step S2, after identifying the fault, the transient waveform is extracted, the fault characteristics are enhanced through Hilbert-Huang transform, and after compression encoding, node information is added and broadcast to neighboring nodes, including: Lightweight neural networks are deployed on edge nodes to analyze standardized fault feature vectors in real time. When the cosine similarity between the fault feature vector and the typical pattern in the local knowledge base is ≥0.9 and lasts for 2 sampling periods, it is determined to be a line fault. After identifying the fault, the transient waveform of the fault is extracted, and the waveform is decomposed into time and frequency using Hilbert-Huang transform to enhance the transient characteristic components of the fault. Wavelet packet compression algorithm is used to compress and encode the waveform, and node ID, timestamp and fault type label are added, and broadcast to neighboring nodes through 5G or fiber optic communication.

[0033] In this embodiment, as Figure 3 As shown, step S3 calculates the fault location through real-time topology calculation, eliminates blind spots through multi-node cross-verification, and introduces edge-cloud collaboration for complex topologies, specifically as follows: When an edge node receives fault traveling wave data from a neighboring node, it extracts the polarity characteristics of the traveling wave arriving at each node. If the same fault traveling wave has the same polarity in neighboring nodes, it is confirmed as a valid fault traveling wave. If the polarities are opposite, it is determined to be electromagnetic interference or a non-fault traveling wave and is filtered out. When three or more edge nodes are involved in fault location, a set of location equations is constructed based on the arrival time and polarity characteristics of the traveling waves collected by each node. The optimal fault location is solved by the least squares method to control the location error and eliminate single-end or double-end location blind spots. If the distribution network topology is a complex form that is not limited to multiple branch lines, ring network structure or distributed power supply access, the edge node automatically triggers the edge cloud collaboration mechanism to upload real-time topology data and traveling wave characteristics to the edge cloud, and uses the cloud global topology model and optimization algorithm to perform fault location calculation.

[0034] In this embodiment, step S4, which involves two nodes near the fault point initiating dual verification of local policy matching and neighboring node confirmation, specifically includes: Once the fault location is determined, the two edge nodes near the fault point perform similarity matching between the current fault feature vector and typical strategies in the pre-stored fault handling strategy library. When the matching degree is ≥0.9, the local strategy execution process is triggered. At the same time, a fault confirmation request is sent to the adjacent node. After receiving the request, the adjacent node performs real-time analysis of the electrical signals it has collected. If no fault features are detected, it sends back a no-fault signal. When both nodes receive no-fault confirmation from the adjacent node and the local strategy matching is successful, a trip command is sent through the encrypted GOOSE protocol.

[0035] In this embodiment, step S5 searches for the optimal recovery path using a mixed-integer linear programming model, specifically including: A mixed-integer linear programming model is constructed with the objective functions of maximizing the restored load capacity and minimizing the restoration path loss. The constraints include keeping the node voltage deviation within the rated value, the line current not exceeding the rated current, the number of switching operations not exceeding 3, and the output of distributed power sources not exceeding 90% of their rated capacity. When solving the model, the distributed power source closest to the fault point is selected first for reverse power supply, and the switching action sequence is generated according to the closing timing optimization algorithm. After sending the closing command, the voltage and current parameters of each node are monitored in real time. When a voltage over-limit or overcurrent abnormality occurs, the backup recovery plan is immediately activated or support is requested from the edge cloud.

[0036] In this embodiment, as Figure 4 As shown, in step S6, each node periodically exchanges data and policy parameters, constructs a global knowledge base through federated learning, dynamically optimizes control parameters using the MADDPG algorithm, and trains a digital twin model to simulate fault scenarios. Specifically, this includes: Each edge node exchanges historical fault data feature vectors and local policy parameters through an encrypted communication link at a preset period. During the exchange, only the model parameter gradients are uploaded, not the original data. The edge cloud uses the FedAvg algorithm to aggregate parameters of each node and update the global knowledge base model. The training data covers fault types such as normal state, short circuit, grounding, and open circuit, and includes operational data under different seasons and load scenarios. The updated global model is distributed to each edge node to replace the local model parameters. At the same time, a digital twin model is trained based on historical fault data to simulate the response characteristics under different fault scenarios and optimize the fault handling strategy library. Each edge node is defined as an independent intelligent agent. The state space includes local fault feature vectors, state parameters of adjacent nodes and global topology matrix. The action space includes trip decision threshold adjustment, closing timing optimization and traveling wave positioning parameter correction. The reward function is constructed into a multi-objective optimization system based on fault handling time, positioning accuracy, and power restoration rate, and the policy gradient is generated by offline training of fault simulation scenarios. The edge cloud uses the MADDPG algorithm to aggregate experience data from various agents, dynamically update the control parameter matrix, and distribute it to the edge nodes. A digital twin model is trained based on historical fault data and real-time operating parameters to simulate transient response characteristics under fault scenarios, including but not limited to three-phase short circuits and single-phase grounding. The fault handling strategy library is optimized to achieve adaptive evolution of control parameters.

[0037] In this embodiment, the encrypted GOOSE protocol uses the AES-256 algorithm to encrypt and protect the tripping command, and combines it with HMAC-SHA256 to generate a message authentication code. The anti-replay sequence embedded in the command consists of a nanosecond-level timestamp and a random number. The receiving node verifies the timestamp deviation and that the random number does not appear in the command.

[0038] In this embodiment, the digital twin model is trained using an LSTM neural network architecture. The input includes time-domain / frequency-domain features of historical fault data, real-time operating parameters, and a topology matrix, simulating transient response characteristics of faults, including but not limited to three-phase short circuits, single-phase grounding, and open circuits.

[0039] Example 2

[0040] I. Data Acquisition and Feature Processing Examples

[0041] Edge computing terminals are deployed at key nodes such as the outgoing terminals of 10kV distribution transformers and cable branch boxes to synchronously collect three-phase current, voltage traveling wave signals and partial discharge signals.

[0042] Dual redundancy timing is implemented: The BeiDou satellite timing module provides the global positioning system clock. When the difference between the primary and backup clocks exceeds 500ns for three consecutive sampling periods, a clock switching mechanism is triggered. A nanosecond-level timestamp is added to the signal via a field-programmable gate array (FPGA). The timestamp format is as follows: Year-Month-Day T Hour:Minute:Second.Millisecond+Nanosecond (e.g. 2025-06-29T15:30:22.123456789Z).

[0043] Wave velocity fitting algorithm: The initial wave velocity is taken as 2.85 × 10^8 m / s (0.95 times the speed of light in a vacuum) according to IEC 60641 standard. The actual wave velocity is fitted using the least squares method based on historical fault data. Taking two-node positioning as an example, the arrival times of the fault traveling wave at node 1 and node 2 are t1 and t2 respectively, and the node spacing is L. Then the actual wave velocity v satisfies:

[0044] Where Δt is the clock skew correction value.

[0045] Signal denoising and feature extraction: High-frequency traveling waves are decomposed into 5 levels using the db6 wavelet, with an adaptive threshold set to 3 times the noise standard deviation, and denoising is achieved through a soft thresholding function; power frequency quantities are denoised using a 51-point moving average filter combined with a Kalman filter, with the state equation as follows:

[0046] in Here is the state transition matrix. The process noise matrix is... For the observation matrix, , The noise is Gaussian white noise. Empirical Mode Decomposition (EMD) is used to generate Eigenmode Functions (IMFs). The first three IMF components are extracted to reconstruct the waveform. Feature parameters are extracted from the time domain (fault current peak, rise time, half-peak width), frequency domain (3rd / 5th / 7th harmonic content, calculated using FFT), and time-frequency joint domain. Dimensionality reduction is achieved through Z-score normalization and principal component analysis. Cosine similarity is calculated with typical fault modes in the local knowledge base. Features with a similarity ≥ 0.9 are considered valid and stored in the SQLite database. Timestamp indexes and fault type labels (such as A-phase grounding, BC-phase short circuit, etc.) are set.

[0047] II. Fault Identification and Feature Broadcasting Examples

[0048] A lightweight MobileNet-v3 neural network is deployed at the edge nodes, taking a feature vector as input and outputting a fault probability value (0-1). The model uses the ReLU activation function, employs the Adam optimizer during training, and uses binary cross-entropy as the loss function. When the cosine similarity corresponding to the output probability is ≥0.9 and persists for 2 sampling periods, a line fault is identified, triggering the transient waveform extraction module.

[0049] Hilbert-Huang transform enhancement features: Extract transient waveforms from 20ms before the fault to 50ms after the fault, decompose the waveforms into 5 IMF components through empirical mode decomposition, perform Hilbert transform on each IMF to generate a Hilbert spectrum, retain the time-frequency components with the first 80% of energy, and enhance the transient features of the fault (such as the arrival time of the traveling wave and the point of amplitude change).

[0050] Compression Coding and Broadcasting: A wavelet packet compression algorithm is used to decompose the enhanced waveform into three layers, and the sub-bands with concentrated energy are selected for encoding. An additional node ID (a 64-bit unique code, such as ED-001-2025), timestamp, and fault type label are added and broadcast to neighboring nodes via 5G or fiber optic communication. The broadcast period is 10ms, ensuring that neighboring nodes receive fault information within 50ms.

[0051] III. Fault Location and Cloud Collaboration Examples

[0052] After receiving broadcast data from neighboring nodes, edge nodes extract the arrival polarity characteristics of the traveling wave (+1 for the rising edge and -1 for the falling edge). Taking nodes A and B as an example, if the same fault traveling wave is +1 at both nodes A and B, it is confirmed as a valid traveling wave; if it is +1 at A and -1 at B, it is determined to be electromagnetic interference or a non-fault traveling wave and is filtered out.

[0053] Construction of multi-node localization equations: When three edge nodes (A, B, C) participate in localization, let the fault location be (x, y), the arrival times of the traveling wave be tA, tB, and tC, the wave velocity be v, and the node coordinates be (xA, yA), (xB, yB), and (xC, yC), respectively. Then the equation system is as follows:

[0054] in, At the moment the fault occurs, solve for x, y, and y using the least squares method. The objective function is:

[0055] When solving iteratively, the convergence threshold is set to 1m, the maximum number of iterations is 50, and the positioning error is controlled to be ≤50m.

[0056] Complex Topology Cloud Collaboration: When the distribution network topology consists of three or more branches, a ring network structure, or includes distributed power supply access, the edge nodes automatically trigger the cloud collaboration mechanism. Real-time topology data (node ​​connection relationships, switch status, DG access location and capacity) and traveling wave characteristics (polarity, arrival time, waveform feature vector) are uploaded to the edge cloud via a VPN encrypted channel.

[0057] IV. Dual Verification and Trip Execution Examples

[0058] After fault location, the system uses Euclidean distance calculation to select the two edge nodes closest to the fault point (such as node X and node Y) and initiates a dual verification mechanism. Based on a pre-stored fault handling strategy library (containing 200+ typical fault strategies, each strategy corresponding to a feature vector template and a switch action sequence), nodes X and Y perform cosine similarity matching between their current 8-dimensional feature vector and the strategies in the library. When the matching degree is ≥0.9, the local strategy execution process is triggered.

[0059] Neighbor node confirmation process: Node X sends a fault confirmation request to neighboring node M, and node Y sends a request to neighboring node N. After receiving the request, neighboring nodes M / N perform a 100ms window analysis on the electrical signals they have collected, extract feature vectors and compare them with fault feature templates. If the similarity is < 0.5, they will return a no-fault signal.

[0060] Encrypted GOOSE command transmission: When both nodes X and Y receive fault-free confirmations from their neighboring nodes and their local policies match successfully, a trip command is sent via the encrypted GOOSE protocol. The command is encrypted using the AES-256 algorithm, with the key updated every 24 hours. A 64-byte message authentication code is generated using HMAC-SHA256. The command format is as follows: [Frame header (8B)]+[Node ID (8B)]+[Timestamp (8B)]+[Random number (16B)]+[Switch number (4B)]+[Action type (1B)]+[Encrypted data (32B)]+[HMAC (64B)]; The random number is generated using 128 bits. The receiving node verifies that the time difference between the timestamp and the local time is ≤5ms, and the random number does not appear in the most recent 100 instructions to ensure that the instruction has not been replayed. After the trip instruction is sent, the node monitors the switch action feedback signal in real time. If no confirmation signal is received within 100ms, the instruction is resent, up to a maximum of 3 times, to ensure the reliability of instruction execution.

[0061] V. Optimal Recovery Path Search Example

[0062] Based on the updated topology (switch status, fault isolation section), a mixed-integer linear programming model is constructed with the optimization objectives of maximizing the restored load capacity and minimizing the restoration path loss:

[0063] in, Let i be the capacity (kVA) of load node i. This indicates the load recovery status (1 for recovered, 0 for not recovered). Let be the resistance (Ω) of line j. Line current (A), Power supply time (h), , For node voltage and rated voltage (kV), The rated current of the line (A) This indicates the switch's operating state (1 for active, 0 for inactive). , The output and rated capacity (kW) of the distributed power source.

[0064] Distributed power source priority strategy: During model solving, the Euclidean distance of each distributed power source from the fault point is calculated, and the closest DG (distance weight coefficient 0.7) is selected for reverse power supply. A genetic algorithm is used to generate the closing sequence, with a population size of 50, a crossover probability of 0.7, and a mutation probability of 0.03. The objective function is to minimize the time of the action sequence and the number of switching actions. The generated action sequence is as follows: "Close DG outlet switch S2 (t=0ms) → Close tie switch S3 (t=500ms) → Close load switch S4 (t=1000ms)".

[0065] Real-time monitoring and anomaly handling: After sending the closing command, the voltage and current parameters of each node are monitored at a sampling rate of 100Hz. When the voltage deviation exceeds ±10% of the rated value or the current exceeds 120% of the rated value, the backup recovery plan is immediately activated (such as switching to the nearest DG power supply). If the backup plan still cannot resolve the anomaly, support is requested from the edge cloud. The cloud uses a global optimization algorithm to regenerate the recovery path and sends it to the edge node for execution.

[0066] VI. Examples of Federated Learning and Policy Optimization

[0067] Each edge node exchanges historical fault data feature vectors and local policy parameters via a VPN encrypted link every 72 hours. Only model parameter gradients are uploaded, not the raw data, with a communication data volume of approximately 10MB per cycle. The edge cloud uses the FedAvg algorithm to aggregate the parameters of each node; the aggregation formula is as follows:

[0068] in, represents the global parameters after aggregation in round t, and N represents the number of participating nodes (≥5). The weight of node i is based on empirical values ​​for handling node faults; for example, the more faults handled, the higher the weight, ranging from 0.1 to 0.3. These are the local parameters for node i.

[0069] Training data coverage: The training data includes eight fault types, such as normal state, three-phase short circuit, two-phase short circuit, single-phase grounding, two-phase grounding, open circuit, and open circuit grounding, as well as load scenarios in spring (load rate 30%-50%), summer (50%-80%), autumn (30%-50%), and winter (60%-90%). The total number of samples is 5000+, and each sample contains an 8-dimensional feature vector, fault type label, and processing strategy.

[0070] Digital twin model training: The updated global model is distributed to each edge node, and the digital twin model is trained simultaneously based on historical fault data and real-time operating parameters (voltage, current, load). An LSTM neural network architecture is used, with 24 neurons in the input layer (8-dimensional features + 8-dimensional operating parameters + 8-dimensional topology matrix), 2 hidden layers (128 and 64 neurons respectively), and 10 neurons in the output layer (simulating transient response features of 10 fault scenarios). During training, the mean squared error loss function, Adam optimizer, learning rate 0.0001, batch size 16, and 500 training cycles are used. Incremental updates are performed quarterly based on newly added fault data (≥500 records) to improve the model's ability to simulate new faults.

[0071] VII. Dynamic Optimization and Digital Li Generation Implementation Examples

[0072] Each edge node is defined as an independent intelligent agent. The state space S contains a local 8-dimensional feature vector, 3-dimensional voltage parameters of adjacent nodes, 3-dimensional current parameters, and a 10x10 global topology matrix, for a total of 21 dimensions. The action space A includes trip decision threshold adjustment, closing timing optimization, and traveling wave positioning parameter correction, for a total of 3 discrete actions.

[0073] Reward function design:

[0074] Where t is the fault handling time (ms, upper limit 100ms), and e is the positioning error (m, upper limit 50m). , To restore the load and total load (kVA). When t < 50ms, the bonus factor is increased to 0.5; when e < 10m, the bonus factor is increased to 0.4.

[0075] MADDPG Algorithm Implementation: Experience data (state-action-reward-next state) from each agent is collected at the edge cloud, aggregated every 24 hours, and the policy network and value network are updated using an Actor-Critic architecture. The Actor network takes a 21-dimensional state as input and outputs a 3-dimensional action probability distribution; its network structure is a 3-layer fully connected network (128, 64, 3 neurons). The Critic network takes a 21-dimensional state and a 3-dimensional action as input and outputs a value function; its network structure is a 3-layer fully connected network (256, 128, 1 neuron). During training, experience replay (buffer size 10^5) and a target network (update cycle 1000 steps) are used to improve training stability. The control parameter matrix is ​​dynamically updated and distributed to edge nodes; the parameter update cycle is 7 days.

[0076] Digital twin simulation scenarios: Based on a trained model, the transient response characteristics of 10 fault scenarios are simulated, including three-phase short circuit, single-phase grounding, two-phase short circuit, open circuit, high-resistance grounding, arc grounding, distributed power grid connection fault, ring network fault, multi-branch fault, and compound fault (such as short circuit + open circuit). Parameters include voltage drop depth, current mutation rate, and traveling wave propagation time. The simulation scenario weights are adjusted every six months based on field fault statistics, and the fault handling strategy library is optimized to achieve adaptive evolution of control parameters. This improves the fault handling efficiency of the system by more than 20% after one year of operation.

[0077] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A distributed FA collaborative control method for edge computing nodes in a distribution network terminal, characterized in that, Includes the following steps: Edge computing terminals with integrated AI chips are deployed at key nodes of the power distribution network to simultaneously collect raw data on high-frequency electrical signals and partial discharge signals. The raw data is then denoised, enhanced, and standardized to form a multi-dimensional fault feature vector. Each edge node runs a lightweight self-learning algorithm to monitor the line status in real time. After identifying a fault, it extracts the transient waveform, enhances the fault characteristics through Hilbert-Huang transform, compresses and encodes the fault information, and broadcasts it to neighboring nodes. After receiving data from neighboring nodes, the fault location is calculated in real time through topology calculation, blind spots are eliminated through multi-node cross-verification, and edge-cloud collaboration is introduced for complex topologies. After fault location, the two nodes near the fault point initiate dual verification of local policy matching and confirmation by adjacent nodes, send trip command through encrypted GOOSE protocol, embed anti-replay mechanism to ensure command reliability, update topology after action and broadcast isolation result; Based on the updated topology, the optimal recovery path is searched using a mixed-integer linear programming model. Distributed power sources are used to provide reverse power first. Closing commands are sent in sequence and monitored in real time. In case of an anomaly, a backup plan is activated or support is requested. Each node periodically exchanges data and policy parameters, builds a global knowledge base through federated learning, dynamically optimizes control parameters using the MADDPG algorithm, and trains a digital twin model to simulate fault scenarios.

2. The distributed FA collaborative control method for edge computing nodes of distribution network terminals according to claim 1, characterized in that, The denoising, feature enhancement, and standardization processes performed on the original data to form a multi-dimensional fault feature vector are as follows: A dual-redundant time synchronization mechanism is constructed by using the BeiDou satellite time synchronization module and a local temperature-controlled crystal oscillator. When the difference between the primary and backup clocks exceeds the threshold for three consecutive sampling cycles, a switching is triggered and a timestamp is added to the signal. The initial wave velocity was calculated based on the IEC 60641 standard, and the actual wave velocity was fitted using the least squares method with fault data. For high-frequency traveling waves, db6 wavelet decomposition and adaptive threshold denoising are used, and power frequency quantities are processed by moving average filtering combined with Kalman filtering. Eigenmode functions are generated through empirical mode decomposition. After extracting components and reconstructing the waveform, parameters such as peak fault current, harmonic content, and wavelet packet energy entropy are extracted from the time domain, frequency domain, and joint time-frequency domain. Feature vectors are generated through Z-Score standardization and principal component analysis for dimensionality reduction. These vectors are then compared with typical patterns in the local knowledge base using cosine similarity. Features with a similarity ≥ 0.9 are considered valid. These features are stored in a lightweight database with timestamp indexes and fault type labels.

3. The distributed FA collaborative control method for edge computing nodes of distribution network terminals according to claim 1, characterized in that, After identifying the fault, the transient waveform is extracted, the fault characteristics are enhanced by Hilbert-Huang transform, and after compression encoding, node information is added and broadcast to neighboring nodes, including: Lightweight neural networks are deployed on edge nodes to analyze standardized fault feature vectors in real time. When the cosine similarity between the fault feature vector and the typical pattern in the local knowledge base is ≥0.9 and lasts for 2 sampling periods, it is determined to be a line fault. After identifying the fault, the transient waveform of the fault is extracted, and the waveform is decomposed into time and frequency using Hilbert-Huang transform to enhance the transient characteristic components of the fault. Wavelet packet compression algorithm is used to compress and encode the waveform, and node ID, timestamp and fault type label are added, and broadcast to neighboring nodes through 5G or fiber optic communication.

4. The distributed FA collaborative control method for edge computing nodes of distribution network terminals according to claim 1, characterized in that, The method involves calculating the fault location through real-time topology calculation, eliminating blind spots through multi-node cross-verification, and introducing edge-cloud collaboration for complex topologies. Specifically: When an edge node receives fault traveling wave data from a neighboring node, it extracts the polarity characteristics of the traveling wave arriving at each node. If the same fault traveling wave has the same polarity in neighboring nodes, it is confirmed as a valid fault traveling wave. If the polarities are opposite, it is determined to be electromagnetic interference or a non-fault traveling wave and is filtered out. When three or more edge nodes are involved in fault location, a set of location equations is constructed based on the arrival time and polarity characteristics of the traveling waves collected by each node. The optimal fault location is solved by the least squares method to control the location error and eliminate single-end or double-end location blind spots. If the distribution network topology is a complex form that is not limited to multiple branch lines, ring network structure or distributed power supply access, the edge node automatically triggers the edge cloud collaboration mechanism to upload real-time topology data and traveling wave characteristics to the edge cloud, and uses the cloud global topology model and optimization algorithm to perform fault location calculation.

5. The distributed FA collaborative control method for edge computing nodes of distribution network terminals according to claim 1, characterized in that, The dual verification process, which involves local policy matching and confirmation from adjacent nodes, is initiated between the two nodes near the fault point. Specifically, it includes: Once the fault location is determined, the two edge nodes near the fault point perform similarity matching between the current fault feature vector and typical strategies in the pre-stored fault handling strategy library. When the matching degree is ≥0.9, the local strategy execution process is triggered. At the same time, a fault confirmation request is sent to the adjacent node. After receiving the request, the adjacent node performs real-time analysis of the electrical signals it has collected. If no fault features are detected, it sends back a no-fault signal. When both nodes receive no-fault confirmation from the adjacent node and the local strategy matching is successful, a trip command is sent through the encrypted GOOSE protocol.

6. The distributed FA collaborative control method for edge computing nodes of distribution network terminals according to claim 1, characterized in that, The search for the optimal recovery path using a mixed-integer linear programming model specifically includes: A mixed-integer linear programming model is constructed with the objective functions of maximizing the restored load capacity and minimizing the restoration path loss. The constraints include keeping the node voltage deviation within the rated value, the line current not exceeding the rated current, the number of switching operations not exceeding 3, and the output of distributed power sources not exceeding 90% of their rated capacity. When solving the model, the distributed power source closest to the fault point is selected first for reverse power supply, and the switching action sequence is generated according to the closing timing optimization algorithm. After sending the closing command, the voltage and current parameters of each node are monitored in real time. When a voltage over-limit or overcurrent abnormality occurs, the backup recovery plan is immediately activated or support is requested from the edge cloud.

7. The distributed FA collaborative control method for edge computing nodes of distribution network terminals according to claim 1, characterized in that, The nodes periodically exchange data and policy parameters, and construct a global knowledge base through federated learning, specifically including: Each edge node exchanges historical fault data feature vectors and local policy parameters through an encrypted communication link at a preset period. During the exchange, only the model parameter gradients are uploaded, not the original data. The edge cloud uses the FedAvg algorithm to aggregate parameters of each node and update the global knowledge base model. The training data covers fault types such as normal state, short circuit, grounding, and open circuit, and includes operational data under different seasons and load scenarios. The updated global model is distributed to each edge node to replace the local model parameters. At the same time, a digital twin model is trained based on historical fault data to simulate the response characteristics under different fault scenarios and optimize the fault handling strategy library.

8. The distributed FA collaborative control method for edge computing nodes of distribution network terminals according to claim 1, characterized in that, The method of dynamically optimizing control parameters using the MADDPG algorithm and training a digital twin model to simulate fault scenarios specifically includes: Each edge node is defined as an independent intelligent agent. The state space includes local fault feature vectors, state parameters of adjacent nodes and global topology matrix. The action space includes trip decision threshold adjustment, closing timing optimization and traveling wave positioning parameter correction. The reward function constructs a multi-objective optimization system based on fault handling time, positioning accuracy, and power restoration rate, and generates policy gradients through offline training of fault simulation scenarios. The edge cloud uses the MADDPG algorithm to aggregate experience data from various agents, dynamically update the control parameter matrix, and distribute it to the edge nodes. A digital twin model is trained based on historical fault data and real-time operating parameters to simulate transient response characteristics under fault scenarios, including but not limited to three-phase short circuits and single-phase grounding. The fault handling strategy library is optimized to achieve adaptive evolution of control parameters.

9. The distributed FA collaborative control method for edge computing nodes of distribution network terminals according to claim 5, characterized in that, The encrypted GOOSE protocol uses the AES-256 algorithm to encrypt and protect the trip command, and combines it with HMAC-SHA256 to generate a message authentication code. The anti-replay sequence embedded in the command consists of a nanosecond-level timestamp and a random number. The receiving node verifies the timestamp deviation and that the random number does not appear in the command.

10. The distributed FA collaborative control method for edge computing nodes of distribution network terminals according to claim 8, characterized in that, The digital twin model is trained using an LSTM neural network architecture. The input includes time-domain / frequency-domain features of historical fault data, real-time operating parameters, and topology matrix. It simulates transient response characteristics of faults, including but not limited to three-phase short circuits, single-phase grounding, and open circuits.

Images