An edge collaboration system for intelligent fusion terminals in power substations

Abstract

The present invention relates to an edge collaboration system of an intelligent fusion terminal in an electric power substation, characterized in that it includes several intelligent fusion terminals and a cloud. Specifically: a data hierarchical processing mechanism is introduced in the intelligent fusion terminal, and real-time processing is realized by dividing sensor data into high priority and low priority; and based on a noise reduction algorithm for time series anomaly detection, the sensor data is subjected to noise reduction processing; a dynamic collaboration network is established among several intelligent fusion terminals in the substation according to the distribution priority of the computing amount, and a hierarchical task unloading collaboration mechanism is adopted to dynamically allocate computing tasks to other intelligent fusion terminals; the cloud is based on an LSTM model, and the load change of the substation is predicted in real time according to the processed sensor data; a distribution diagnosis model based on a graph neural network is introduced in the cloud; and fault self-healing is realized through a dynamic collaboration network of intelligent fusion terminals. The present invention significantly improves the intelligence level of the substation and the operation efficiency of the power grid.

Description

Technical Field

[0001] The present invention relates to the field of power edge collaboration, and in particular to an edge collaboration system for intelligent fusion terminals in power substations. Background Art

[0002] As the power system develops towards intelligence and distribution, traditional distribution substation management faces multiple challenges, including the increasing complexity of equipment access, the uncertainty of load changes, the need for coordinated optimization of multi-energy systems, and the response requirements for fault diagnosis and rapid recovery. In this context, intelligent fusion terminals and edge computing technologies have gradually become important means to improve the operating efficiency and reliability of substations. Intelligent fusion terminals can monitor the real-time operating status of substations based on sensor data and support data analysis and decision-making on the edge side, thereby significantly reducing latency while reducing the need for data transmission to the cloud.

[0003] On the other hand, with the expansion of the scale of substations and the widespread access to distributed energy (such as photovoltaics, electric vehicle charging, etc.), the dynamics and complexity of power grid operation continue to increase. The traditional centralized computing model cannot efficiently cope with the dynamic task allocation and real-time control requirements of the distribution network. Summary of the invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide an edge collaboration system of intelligent fusion terminals in power substations, which further realizes self-healing of faults in the substations through the dynamic collaboration network of intelligent fusion terminals, and significantly improves the operating efficiency and resilience of the power system.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] An edge collaboration system of intelligent fusion terminals in power substations includes several intelligent fusion terminals and a cloud. Specifically:

[0007] Introducing a data hierarchical processing mechanism in the intelligent fusion terminal, real-time processing is achieved by dividing sensor data into high priority and low priority. Based on the noise reduction algorithm of time series anomaly detection, the sensor data is subjected to noise reduction processing to obtain processed sensor data.

[0008] Establish a dynamic collaborative network among several intelligent fusion terminals in the area according to the distribution priority of computing volume, directly form a dynamic collaborative relationship on the edge side, and adopt a hierarchical task offloading coordination mechanism to dynamically allocate computing tasks to other intelligent fusion terminals;

[0009] Based on the LSTM model, the cloud predicts the load changes in the substation area in real time according to the processed sensor data, and uses federated learning to conduct collaborative training between nodes in the dynamic collaborative network, so that each intelligent fusion terminal can jointly improve the accuracy of the prediction model without exchanging original data;

[0010] The cloud introduces a distribution diagnosis model based on graph neural network to realize topological fault diagnosis based on processed sensor data and the connection relationship between substation equipment.

[0011] Based on the load forecast results of the substation and the topology-level fault diagnosis results, the cloud realizes self-healing of faults through the dynamic collaborative network of intelligent fusion terminals.

[0012] Furthermore, the intelligent converged terminal adopts a heterogeneous computing architecture and adopts a CPU+GPU+NPU collaborative processing method, in which the CPU is responsible for control and management, the GPU handles high-concurrency tasks, the NPU is used for deep learning reasoning, and is equipped with an FPGA module, so that the terminal supports real-time algorithm acceleration and flexible hardware upgrades.

[0013] Furthermore, a data hierarchical processing mechanism is introduced in the intelligent fusion terminal, and real-time processing is achieved by dividing the sensor data into high priority and low priority. Based on the noise reduction algorithm of time series anomaly detection, the sensor data is subjected to noise reduction processing to obtain the processed sensor data; the details are as follows;

[0014] The collected sensor data, including current, voltage, and temperature, is divided into high-priority data and low-priority data in real time through rule threshold algorithms;

[0015] Use the terminal's memory queue to cache high-priority data, perform preliminary analysis in the intelligent fusion terminal, traverse the currently cached high-priority data in the memory queue one by one, and compare it with the predefined normal range value to identify whether there is an anomaly; calculate the data change rate to determine whether there is a rapid fluctuation anomaly; analyze continuous high-priority data to determine whether the anomaly persists. If an anomaly is detected, extract and record relevant information, including the anomaly type, timestamp, specific value, and duration, and upload the anomaly report to the cloud;

[0016] The low-priority data is sent to the noise reduction module, and after filtering out the noise, it is stored in the time series database of the edge terminal for subsequent performance analysis and trend model training.

[0017] Furthermore, several intelligent fusion terminals in the substation area are connected to establish a dynamic collaborative network based on the distribution priority of the computing volume, and a dynamic collaborative relationship is directly formed on the edge side. A hierarchical task offloading collaborative mechanism is adopted to dynamically allocate computing tasks, as follows;

[0018] According to the computing performance of the intelligent fusion terminal, the intelligent fusion terminals in the substation area are divided into core nodes and auxiliary nodes; each core node and its N most adjacent auxiliary nodes form a local group. Within the group, full-connection communication is adopted. Each auxiliary node can communicate directly with the core node, and the auxiliary nodes can also directly cooperate with each other; local groups realize cross-group communication through core nodes to avoid network congestion during global broadcasting, and build a dynamic collaborative network with local full connection;

[0019] Each local group auxiliary node and core node broadcasts its own load information and available resources in a time period;

[0020] Core nodes prioritize task allocation within groups. In each local group, core nodes C k Decompose the tasks and balance the load according to the load of each auxiliary node:

[0021] ;

[0022] Among them, M i Assigned to secondary node A k,i The total amount of tasks; is the remaining computing capacity of the auxiliary node; M is the total amount of assigned tasks; A k,i represents the i-th auxiliary node of the k-th group; N is the number of nodes in the group; j is the index of N;

[0023] When the resources in a group are insufficient, that is, when the core node C of a group k The combined load exceeds the threshold L threshold , the core node selects candidate target nodes C with strong computing power and low communication overhead globally p :

[0024] ;

[0025] in, For node C p The remaining computing power; C k With C p Communication distance; is the communication bandwidth between nodes; is the final target node;

[0026] The auxiliary nodes receive tasks within the group and feedback the results to the core nodes after completion; cross-group tasks are reallocated by the target core node to the auxiliary nodes in its group for processing.

[0027] Furthermore, the cloud uses the LSTM model to predict the load changes in the substation area in real time based on the processed sensor data, and uses federated learning to conduct collaborative training between nodes in the dynamic collaborative network, as follows:

[0028] Each terminal uses local data to train the load prediction model and uploads the model parameters to the cloud. In each local group, the core node coordinates the auxiliary nodes to complete local training and summarize the model parameters in the group:

[0029]

[0030] Among them, W group,k is the aggregation model parameter of the kth group; N is the number of nodes in the group; Auxiliary node A k,j Local model parameters of ; Auxiliary node A k,j The training sample size;

[0031] The core node uploads the model parameters after aggregation within the group to the cloud, and the cloud completes the global model aggregation:

[0032] ;

[0033] Among them, W global is the global model parameter; is the training sample size of the kth group; K is the number of groups;

[0034] The cloud aggregates the model parameters of all terminals, updates the global model and sends it to each terminal; the terminal nodes use the updated model to make real-time predictions of the substation load, and the cloud performs load trend analysis based on the global model. Based on the predicted load trend, it dynamically adjusts the task allocation of the collaborative network: if the predicted load increases, more tasks are allocated to high-performance terminals in advance; if the predicted load decreases, the task allocation is reduced.

[0035] Furthermore, the load forecasting model is built based on the LSTM model, as follows:

[0036] The input features are multivariate time series data, including: voltage, current, power, environmental data and time characteristics;

[0037] The model structure includes an input layer, a hidden layer and an output layer; the input layer inputs multivariate time series data (of shape T×F, where T is the time step and F is the number of features); the hidden layer uses multi-layer LSTM units to capture time series features; the output layer predicts the load value of the next H steps;

[0038] The state of the LSTM unit is updated:

[0039] Forget Gate: ;

[0040] Input Gate: ;

[0041] Candidate values: ;

[0042] Cell status update: ;

[0043] Output Gate: ;

[0044] Hide status updates: ;

[0045] in, is the input feature vector at time step t; is the hidden state at time step t; For the Gate of Oblivion; is the input gate; is the candidate value; that is, the candidate memory cell at the current moment; is the cell state; it indicates the memory cell state at the current moment; is the output gate; is the hidden state at the current moment; , , , They are the weight matrices of the forget gate, input gate, candidate value, and output gate respectively; , , , are the bias vectors of the forget gate, input gate, candidate value, and output gate, respectively. is the activation function.

[0046] Furthermore, the distribution diagnosis model based on graph neural network is as follows:

[0047] A dynamic directed weighted graph is used to model the substation equipment and their connection relationships. The graph network is defined as G t :

[0048] G t =(V,E,X t ,A t );

[0049] Among them, V is the node set; node feature x i ∈Xt represents the sensor state of the device; E is a set of directed edges, representing the electrical connection between devices; A t For a dynamic adjacency matrix, the connection weights of the edges are defined, and the edge characteristics depend on the electrical flow;

[0050] Dynamic adjacency matrix between nodes i and j The calculation formula is:

[0051] ;

[0052] Among them, f w Based on node features and edge features The dynamic weight function of , which expresses the transmission strength of the edge;

[0053] The node features are passed through GNN for multi-layer message passing, and the spatial topological features are obtained by dynamic graph convolution. The node embedding update formula is:

[0054]

[0055] in, is the hidden state of node i in layer l; For edge e ij The feature weights of is the trainable parameter matrix; σ is the nonlinear activation function; N(i) is the set of adjacent nodes of node i; c ij is the normalization factor of information transfer from neighbor node j to node i; is the bias vector;

[0056] The update formula for edge features is:

[0057] ;

[0058] in, is the edge feature weight matrix of the lth layer; is a nonlinear activation function;

[0059] The characteristics of the time dimension obtained by the load forecasting model And the spatial topological features obtained by GNN Fusion, the total embedding becomes h i final :

[0060] ;

[0061] Through the node classification task, each node is predicted as a faulty or non-faulty state, and the fault probability p is output. i :

[0062] ;

[0063] Among them, w o ,b o are the weight and bias of the output layer respectively;

[0064] Fault propagation path analysis: Calculate the possibility of the fault propagation path according to the edge characteristics and node weights:

[0065] ;

[0066] where v ij is the fault weight propagated from node i to j; The fault propagation analysis is performed by determining , is the propagation threshold.

[0067] Furthermore, based on the台区 load prediction results and the topological-level fault diagnosis results, the cloud realizes fault self-healing through the dynamic cooperation network of intelligent fusion terminals, specifically as follows:

[0068] Based on the input 台区 load prediction and fault diagnosis results, the self-healing strategy includes:

[0069] a. Fault isolation: It includes disconnecting the faulty device, isolating the faulty area and its influence range, and each isolation operation is based on the influence propagation weight ;

[0070] b. According to the predicted load calculation results and the power supply capacity, preferentially transfer the load to the standby path; Allocate the optimized load power to other normal nodes:

[0071] ;

[0072] where β ij is the load transfer allocation ratio, is the total power demand of node j after transferring or adjusting the load; P j is the original load power of node j; is the allocation ratio of the load transferred from node i to node j;

[0073] c. If the transferred load R i <re, where re is a preset threshold, is insufficient to meet the power supply demand, then enable the standby power supply to supplement the power supply:

[0074] ;

[0075] where, is the standby power supply power of node i; is the total load demand of node i; is the total load increment allocated to node i through load transfer;

[0076] It should be noted that the term "台区" in the original text may need to be further clarified according to the specific context to ensure more accurate translation. Here it is directly translated as "台区" for the time being.d. Dynamic collaborative network coordinates intelligent fusion terminals to complete distributed control. Node operations include: circuit breaker switch control, reducing or inhibiting power flow to faulty nodes; power supply priority management, adjusting transmission power according to path importance;

[0077] The core node collects status information within the group, determines the feasibility of the self-healing action, and the cloud sends action instructions to the core node. The core node quickly transmits the information through the internal collaborative network of the group and simultaneously updates the status of the entire network.

[0078] Each intelligent fusion terminal executes actions based on real-time status feedback. If some terminals fail to execute, the core node updates the strategy to adjust the tasks of other terminals.

[0079] The present invention has the following beneficial effects:

[0080] 1. The present invention further realizes fault self-healing in the substation through the dynamic collaborative network of intelligent fusion terminals, significantly improving the operating efficiency and resilience of the power system;

[0081] 2. The present invention constructs a hybrid multi-layer dynamic collaborative network, enhances the dynamic unloading and task sharing capabilities of multiple intelligent fusion terminals in the substation area, effectively addresses the problem of unbalanced terminal loads in the substation area, reduces communication delays by adopting local grouping and core node collaboration, and introduces load prediction to avoid preemptive task unloading caused by short-term fluctuations. Combined with dynamic priority setting and performance-sensitive scheduling, the task execution efficiency is improved;

[0082] 3. The present invention integrates intelligent load forecasting, topology information and collaborative networks to construct an efficient and robust substation fault self-healing technology framework, effectively improving the reliability and intelligence level of the distributed distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 This is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0084] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0085] refer to Figure 1 In this embodiment, an edge collaboration system of an intelligent fusion terminal in a power substation is provided, including several intelligent fusion terminals and a cloud. Specifically:

[0086] Introducing a data hierarchical processing mechanism in the intelligent fusion terminal, real-time processing is achieved by dividing sensor data into high priority (such as fault information) and low priority (such as operation trend data); high-priority data is cached in real time using the edge terminal memory, and preliminary analysis is completed at the millisecond level; and based on the noise reduction algorithm of time series anomaly detection (such as Fast Fourier Transform (FFT) combined with time series analysis), the sensor data is subjected to noise reduction processing to obtain processed sensor data;

[0087] A dynamic collaborative network is established among several intelligent fusion terminals in the substation according to the distribution priority of the computing volume, and a dynamic collaborative relationship is directly formed on the edge side to share the task load through a distributed computing architecture. When the load of a certain intelligent fusion terminal is too high (for example, a sharp increase in computing pressure due to fault analysis), a hierarchical task offloading coordination mechanism is adopted to dynamically allocate computing tasks to other intelligent fusion terminals.

[0088] Based on the LSTM model, the cloud predicts the load changes of the substation in real time according to the processed sensor data, and uses Federated Learning to conduct collaborative training between nodes in the dynamic collaborative network, so that each intelligent fusion terminal can jointly improve the accuracy of the prediction model without exchanging original data;

[0089] The cloud introduces a distribution diagnosis model based on graph neural network to realize topological fault diagnosis based on processed sensor data and the connection relationship between substation equipment.

[0090] Based on the load forecast results of the substation and the topology-level fault diagnosis results, the cloud realizes self-healing of faults through the dynamic collaborative network of intelligent fusion terminals.

[0091] In this embodiment, the intelligent converged terminal adopts a heterogeneous computing architecture and adopts a CPU+GPU+NPU collaborative processing method, in which the CPU is responsible for control and management, the GPU processes high-concurrency tasks, the NPU is used for deep learning reasoning, and is equipped with an FPGA module to enable the terminal to support real-time algorithm acceleration and flexible hardware upgrades.

[0092] In this embodiment, a data hierarchical processing mechanism is introduced in the intelligent fusion terminal, and real-time processing is achieved by dividing the sensor data into high priority and low priority. The sensor data is subjected to noise reduction processing based on a noise reduction algorithm for time series anomaly detection to obtain processed sensor data.

[0093] The collected sensor data, including current, voltage, and temperature, is divided into high-priority data and low-priority data in real time through a rule threshold algorithm; (the goal of the priority division is to quickly identify critical fault information (high-priority data) and process it in real time, while reducing noise and storing the operating trend data (low-priority data);

[0094] Use the terminal's memory queue to cache high-priority data, perform preliminary analysis in the intelligent fusion terminal, and traverse the currently cached high-priority data in the memory queue one by one, and compare it with the predefined normal range value to identify whether there is an anomaly; calculate the data change rate (such as the unit time change of voltage / current) to determine whether there is a rapid fluctuation anomaly; analyze continuous high-priority data to determine whether the anomaly persists. If an anomaly is detected, extract and record relevant information, including the anomaly type, timestamp, specific value, and duration, and upload the anomaly report to the cloud;

[0095] The low-priority data is sent to the noise reduction module, and after filtering out the noise, it is stored in the time series database of the edge terminal for subsequent performance analysis and trend model training.

[0096] In this embodiment, a number of intelligent fusion terminals in the station area are used to establish a dynamic collaboration network based on the distribution priority of the computing amount, and a dynamic collaboration relationship is directly formed on the edge side. A hierarchical task offloading collaboration mechanism is adopted to dynamically allocate computing tasks, as follows;

[0097] According to the computing performance of the intelligent fusion terminal, the intelligent fusion terminals in the substation area are divided into core nodes and auxiliary nodes; each core node and its N most adjacent auxiliary nodes form a local group. Within the group, full-connection communication is adopted. Each auxiliary node can communicate directly with the core node, and the auxiliary nodes can also directly cooperate with each other; local groups realize cross-group communication through core nodes to avoid network congestion during global broadcasting, and build a dynamic collaborative network with local full connection;

[0098] Each local group auxiliary node and core node broadcasts its own load information and available resources in a time period;

[0099] Core nodes prioritize task allocation within groups. In each local group, core nodes C k Decompose the tasks and balance the load according to the load of each auxiliary node:

[0100] ;

[0101] Among them, M i Assigned to auxiliary node A k,i The total amount of tasks; is the remaining computing capacity of the auxiliary node; M is the total amount of assigned tasks; A k,i represents the i-th auxiliary node of the k-th group; N is the number of auxiliary nodes in the group; j is the index of N;

[0102] When the resources in a group are insufficient, that is, when the core node C of a group k The combined load exceeds the threshold Lthreshold , the core node selects candidate target nodes C with strong computing power and low communication overhead globally p :

[0103] ;

[0104] in, For node C p The remaining computing power; C k With C p Communication distance; is the communication bandwidth between nodes; is the final target node;

[0105] The auxiliary nodes receive tasks within the group and feedback the results to the core nodes after completion; cross-group tasks are reallocated by the target core node to the auxiliary nodes in its group for processing.

[0106] In this embodiment, the cloud uses the LSTM model to predict the load change of the substation in real time according to the processed sensor data, and uses federated learning to perform collaborative training between nodes in the dynamic collaborative network, as follows:

[0107] Each terminal uses local data to train the load prediction model and uploads the model parameters to the cloud. In each local group, the core node coordinates the auxiliary nodes to complete local training and summarize the model parameters in the group:

[0108]

[0109] Among them, W group,k is the aggregation model parameter of the kth group; N is the number of nodes in the group; Auxiliary node A k,j Local model parameters of ; Auxiliary node A k,j The training sample size;

[0110] The core node uploads the model parameters after aggregation within the group to the cloud, and the cloud completes the global model aggregation:

[0111] ;

[0112] Among them, W global is the global model parameter; is the training sample size of the kth group; K is the number of groups;

[0113] The cloud aggregates the model parameters of all terminals, updates the global model and sends it to each terminal; the terminal nodes use the updated model to make real-time predictions of the substation load, and the cloud performs load trend analysis based on the global model. Based on the predicted load trend, it dynamically adjusts the task allocation of the collaborative network: if the predicted load increases, more tasks are allocated to high-performance terminals in advance; if the predicted load decreases, the task allocation is reduced.

[0114] In this embodiment, the load forecasting model is constructed based on the LSTM model, as follows:

[0115] The input features are multivariate time series data, including voltage, current, power, environmental data (such as temperature and humidity), and time features (such as hour, week, and holiday identifiers).

[0116] The model structure includes an input layer, a hidden layer and an output layer; the input layer inputs multivariate time series data (of shape T×F, where T is the time step and F is the number of features); the hidden layer uses multi-layer LSTM units to capture time series features; the output layer predicts the load value of the next H steps;

[0117] The state of the LSTM unit is updated:

[0118] Forget Gate: ;

[0119] Input Gate: ;

[0120] Candidate values: ;

[0121] Cell status update: ;

[0122] Output Gate: ;

[0123] Hide status updates: ;

[0124] in, is the input feature vector at time step t; is the hidden state at time step t; For the Gate of Oblivion; is the input gate; is the candidate value; that is, the candidate memory cell at the current moment; is the cell state; it indicates the memory cell state at the current moment; is the output gate; is the hidden state at the current moment; , , , They are the weight matrices of the forget gate, input gate, candidate value, and output gate respectively; , , , are the bias vectors of the forget gate, input gate, candidate value, and output gate, respectively. is the activation function.

[0125] In this embodiment, the power distribution diagnosis model based on the graph neural network is as follows:

[0126] A dynamic directed weighted graph is used to model the substation equipment and their connection relationships. The graph network is defined as G t :

[0127] G t =(V,E,X t ,A t );

[0128] Where V is a node set, including equipment (such as transformers, switches, circuit breakers, etc.); node feature x i ∈Xt represents the sensor state of the device; E is a set of directed edges, representing the electrical connection between devices; A t It is a dynamic adjacency matrix that defines the connection weights of the edges, and the edge characteristics depend on the electrical flow (such as active / reactive power, impedance, etc.);

[0129] Dynamic adjacency matrix between nodes i and j The calculation formula is:

[0130] ;

[0131] Among them, f w Based on node features and edge features The dynamic weight function of , which expresses the transmission strength of the edge;

[0132] The node features are passed through GNN for multi-layer message passing, and the spatial topological features are obtained by dynamic graph convolution. The node embedding update formula is:

[0133]

[0134] in, is the hidden state of node i in layer l; For edge e ij The feature weights of is the trainable parameter matrix; σ is the nonlinear activation function; N(i) is the set of adjacent nodes of node i; c ij is the normalization factor of information transfer from neighbor node j to node i; is the bias vector;

[0135] The update formula for edge features is:

[0136] ;

[0137] in, is the edge feature weight matrix of the lth layer; is a nonlinear activation function;

[0138] The characteristics of the time dimension obtained by the load forecasting model And the spatial topological features obtained by GNN Fusion, the total embedding becomes h i final :

[0139] ;

[0140] Through the node classification task, each node is predicted as a faulty or non-faulty state, and the fault probability p is output. i :

[0141] ;

[0142] Among them, w o ,b o are the weight and bias of the output layer respectively;

[0143] Fault propagation path analysis: Calculate the possibility of fault propagation path based on edge characteristics and node weights:

[0144] ;

[0145] Among them, v ij is the fault weight propagated from node i to node j; fault propagation analysis is performed by determining , is the propagation threshold.

[0146] In this embodiment, the cloud implements fault self-healing through a dynamic collaborative network of intelligent converged terminals based on the load forecast results of the substation area and the topology-level fault diagnosis results, as follows:

[0147] Based on the input area load forecast and fault diagnosis results, the self-healing strategies include:

[0148] a. Fault isolation: including disconnecting the faulty device, isolating the faulty area and its impact range. Each isolation operation is based on the impact propagation weight. ;

[0149] b. Based on the predicted load calculation results and power supply capacity, the load is transferred to the backup path first; the optimized load power is allocated to other normal nodes:

[0150] ;

[0151] wherein, β ij is the load transfer distribution ratio, is the total power demand of node j after transferring or adjusting the load; P j is the original load power of node j; is the distribution ratio of the load transferred from node i to node j;

[0152] c. If the transferred load R i <re, where re is a preset threshold, is insufficient to meet the power supply demand, then the backup power supply is enabled to supplement the power supply:

[0153] ;

[0154] wherein, is the backup power supply power of node i; is the total load demand of node i; is the total load increment allocated to node i through load transfer;

[0155] d. The dynamic cooperation network coordinates the intelligent fusion terminals to complete distributed regulation. The node operations include: circuit breaker switch control to reduce or inhibit the power flow to the faulty node; power supply priority management to adjust the transmission power according to the path importance;

[0156] The core node collects the status information within the group, judges the feasibility of the self-healing action, the cloud sends the action instruction to the core node, and the core node quickly transmits it through the internal cooperation network of the group and synchronously updates the status of the whole network;

[0157] Each intelligent fusion terminal executes the action according to the real-time status feedback. If the execution of some terminals fails (such as switch failure, communication interruption), the core node updates the strategy to adjust the tasks of other terminals.

[0158] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0159] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0160] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0161] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0162] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. An edge collaboration system for intelligent fusion terminals in power substations, characterized in that: Including several intelligent fusion terminals and cloud, specifically: Introducing a data hierarchical processing mechanism in the intelligent fusion terminal, real-time processing is achieved by dividing sensor data into high priority and low priority. Based on the noise reduction algorithm of time series anomaly detection, the sensor data is subjected to noise reduction processing to obtain processed sensor data. Establish a dynamic collaborative network among several intelligent fusion terminals in the area according to the distribution priority of computing volume, directly form a dynamic collaborative relationship on the edge side, and adopt a hierarchical task offloading coordination mechanism to dynamically allocate computing tasks to other intelligent fusion terminals; Based on the LSTM model, the cloud predicts the load changes in the substation area in real time according to the processed sensor data, and uses federated learning to conduct collaborative training between nodes in the dynamic collaborative network, so that each intelligent fusion terminal can jointly improve the accuracy of the prediction model without exchanging original data; The cloud introduces a distribution diagnosis model based on graph neural network to realize topological fault diagnosis based on processed sensor data and the connection relationship between substation equipment. Based on the load forecast results of the substation and the topology-level fault diagnosis results, the cloud realizes self-healing of faults through the dynamic collaborative network of intelligent fusion terminals.

2. According to claim 1, the edge collaboration system of the intelligent fusion terminal of the power substation is characterized in that: The intelligent fusion terminal adopts a heterogeneous computing architecture and adopts a CPU+GPU+NPU collaborative processing method, in which the CPU is responsible for control and management, the GPU processes high-concurrency tasks, the NPU is used for deep learning reasoning, and is equipped with an FPGA module, so that the terminal supports real-time algorithm acceleration and flexible hardware upgrades.

3. According to claim 1, the edge collaboration system of the intelligent fusion terminal of the power substation is characterized in that: The data hierarchical processing mechanism is introduced into the intelligent fusion terminal, and the sensor data is divided into high priority and low priority to achieve real-time processing, and the sensor data is subjected to noise reduction processing based on the noise reduction algorithm of time series anomaly detection to obtain the processed sensor data; the details are as follows; The collected sensor data, including current, voltage, and temperature, is divided into high-priority data and low-priority data in real time through rule threshold algorithms; Use the terminal's memory queue to cache high-priority data, perform preliminary analysis in the intelligent converged terminal, traverse the currently cached high-priority data in the memory queue one by one, and compare it with the predefined normal range value to identify whether there is an abnormality; Calculate the data change rate to determine whether there is a rapid fluctuation anomaly; analyze continuous high-priority data to determine whether the anomaly persists. If an anomaly is detected, extract and record relevant information, including the anomaly type, timestamp, specific value, duration, and upload the anomaly report to the cloud; The low-priority data is sent to the noise reduction module, and after filtering out the noise, it is stored in the time series database of the edge terminal for subsequent performance analysis and trend model training.

4. According to claim 1, the edge collaboration system of the intelligent fusion terminal of the power substation is characterized in that: The multiple intelligent fusion terminals in the station area are established into a dynamic collaborative network according to the distribution priority of the computing amount, and a dynamic collaborative relationship is directly formed on the edge side, and a hierarchical task offloading collaborative mechanism is adopted to dynamically allocate computing tasks, as follows; According to the computing performance of the intelligent fusion terminal, the intelligent fusion terminals in the substation area are divided into core nodes and auxiliary nodes; each core node and its N most adjacent auxiliary nodes form a local group. Within the group, full-connection communication is adopted. Each auxiliary node can communicate directly with the core node, and the auxiliary nodes can also directly cooperate with each other; local groups realize cross-group communication through core nodes to avoid network congestion during global broadcasting, and build a dynamic collaborative network with local full connection; Each local group auxiliary node and core node broadcasts its own load information and available resources in a time period; Core nodes prioritize task allocation within groups. In each local group, core nodes C k Decompose the tasks and balance the load according to the load of each auxiliary node: Among them, M i Assigned to auxiliary node A k,i The total amount of tasks; is the remaining computing capacity of the auxiliary node; M is the total amount of assigned tasks; A k,i represents the i-th auxiliary node of the k-th group; N is the number of nodes in the group; j is the index of N; When the resources in a group are insufficient, that is, when the core node C of a group k The combined load exceeds the threshold L threshold , the core node selects candidate target nodes C with strong computing power and low communication overhead globally p : in, For node C p The remaining computing power; C k With C p Communication distance; is the communication bandwidth between nodes; is the final target node; The auxiliary nodes receive tasks within the group and feedback the results to the core nodes after completion; cross-group tasks are reallocated by the target core node to the auxiliary nodes in its group for processing.

5. According to claim 1, the edge collaboration system of the intelligent fusion terminal of the power substation is characterized in that: The cloud is based on the LSTM model, predicts the load change of the substation in real time according to the processed sensor data, and uses federated learning to perform collaborative training between nodes in the dynamic collaborative network, as follows: Each terminal uses local data to train the load prediction model and uploads the model parameters to the cloud. In each local group, the core node coordinates the auxiliary nodes to complete local training and summarize the model parameters in the group: Among them, W group,k is the aggregation model parameter of the kth group; N is the number of nodes in the group; Auxiliary node A k,j Local model parameters of ; Auxiliary node A k,j The training sample size; The core node uploads the model parameters after aggregation within the group to the cloud, and the cloud completes the global model aggregation: Among them, W global is the global model parameter; is the training sample size of the kth group; K is the number of groups; The cloud aggregates the model parameters of all terminals, updates the global model and sends it to each terminal; the terminal nodes use the updated model to make real-time predictions of the substation load, and the cloud performs load trend analysis based on the global model. Based on the predicted load trend, it dynamically adjusts the task allocation of the collaborative network: if the predicted load increases, more tasks are allocated to high-performance terminals in advance; if the predicted load decreases, the task allocation is reduced.

6. The edge collaboration system of the intelligent fusion terminal of the power substation according to claim 5 is characterized in that: The load forecasting model is built based on the LSTM model, as follows: The input features are multivariate time series data, including: voltage, current, power, environmental data and time characteristics; The model structure includes an input layer, a hidden layer and an output layer; the input layer inputs multivariate time series data; the hidden layer uses multi-layer LSTM units to capture time series features; the output layer predicts the load value of the next H steps; The state of the LSTM unit is updated: 。 7. The edge collaboration system of the intelligent fusion terminal of the power substation according to claim 1 is characterized in that: The power distribution diagnosis model based on graph neural network is as follows: A dynamic directed weighted graph is used to model the substation equipment and their connection relationships. The graph network is defined as G t : G t =(V,E,X t ,A t ), where V is the node set; node feature x i ∈Xt represents the sensor state of the device; E is a set of directed edges, representing the electrical connection between devices; A t is the dynamic adjacency matrix, which defines the connection weights of the edges, and the edge characteristics depend on the electrical flow; the dynamic adjacency matrix between nodes i and j The calculation formula is: ; Among them, f w Based on node features and edge features The dynamic weight function expresses the transmission strength of the edge; the node features are passed through GNN for multi-layer message transmission, and the spatial topological features are obtained by dynamic graph convolution. The node embedding update formula is: ; The update formula for edge features is: ; Through the node classification task, each node is predicted to be in a faulty or non-faulty state, and the fault probability is output p i : Among them, W o ,b o are the weight and bias of the output layer respectively; Fault propagation path analysis: According to the edge characteristics and node weights, the possibility of fault propagation path is calculated: V ij =e ij ⋅p i ;in, V ij is the fault weight propagated from node i to node j; fault propagation analysis is performed by determining V ij >ϵ, ϵ is the propagation threshold.

8. The edge collaboration system of the intelligent fusion terminal of the power substation according to claim 1 is characterized in that: The cloud realizes fault self-healing through a dynamic collaborative network of intelligent fusion terminals based on the load forecast results of the substation area and the topology-level fault diagnosis results, as follows: Based on the input load prediction and fault diagnosis results of the substation area, the self-healing strategy includes: a. Fault isolation: including disconnecting the faulty equipment, isolating the faulty area and its influence range, and each isolation operation is based on the influence propagation weight v ij > ϵ; b. According to the predicted load calculation results and power supply capacity, preferentially transfer the load to the standby path; allocate the optimized load power to other normal nodes: ; where β ij is the load transfer allocation ratio, is the total power demand of node j after transferring or adjusting the load; P j is the original load power of node j; is the allocation ratio of the load transferred from node i to node j; c. If the transferred load R i < re, where re is a preset threshold, is not enough to meet the power supply demand, then enable the standby power supply to supplement the power supply: where, is the standby power supply power of node i; is the total load demand of node i; is the total load increment allocated to node i through load transfer; d. The dynamic cooperation network coordinates the intelligent fusion terminal to complete distributed regulation. The node operations include: circuit breaker switch control to reduce or suppress the power flow to the faulty node; power supply priority management to adjust the transmission power according to the path importance; the core node collects the status information within the group, judges the feasibility of the self-healing action, the cloud sends the action instruction to the core node, and the core node quickly transmits it through the internal cooperation network of the group and synchronously updates the status of the whole network; Each intelligent fusion terminal executes actions based on real-time status feedback. If some terminals fail to execute, the core node updates the strategy to adjust the tasks of other terminals.

Images