Method and system for generating large model of power monitoring system
Through the combination of t-SNE dimensionality reduction and ranking federated learning, the problems of data heterogeneity and communication overhead in the power grid system are solved, efficient training and security of the power monitoring system are achieved, and data privacy protection is ensured.
Patent Information
- Application Number
- CN202510463361.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-14
- Publication Date
- 2025-08-15
AI Technical Summary
There are problems in power grid systems with high data heterogeneity and large communication overhead, which affect the accuracy and efficiency of the model. The existing federated learning methods have shortcomings in data privacy protection and communication optimization.
The t-SNE dimensionality reduction technology is used to reduce the dimensionality of high-dimensional heterogeneous data. Combined with the ranking federated learning mechanism, through the collaboration between the power grid headquarters and sub-nodes, the local score is initialized using random seeds, and the local ranking is calculated while the pop-up algorithm is used to calculate the local ranking, and the global ranking is updated through the number of votes, and finally a power monitoring model with minimized average losses is generated.
It effectively solves the problems of high data heterogeneity and large communication overhead in power grid systems, ensures data privacy, improves the efficiency and security of model training, reduces communication overhead, and ensures the smoothness of large-scale grid model training.
Smart Images
Figure CN120493686A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of large-scale model generation, and in particular to a method and system for generating a large-scale model of an electric power monitoring system. Background Art
[0002] With the surge in power grid system data and the diversification of its features, power grid system data has become high-dimensional and dynamically changing, which poses scalability and real-time challenges for traditional machine learning models. In this context, large models have attracted attention due to their powerful data processing capabilities and flexibility. They can better adapt to the complexity and dynamic changes of power grid data, thus playing a key role in power monitoring systems.
[0003] As the scale of power systems continues to expand, uploading data from distributed nodes to a central node poses not only the risk of data leakage but also incurs significant communication overhead due to the surge in data volume. Furthermore, the increasing heterogeneity of power grid system data significantly impacts model accuracy. To address these challenges, adopting a federated learning architecture for large-scale model training has become an effective solution. Federated learning allows data to remain locally, minimizing the risk of information leakage. Furthermore, it only requires the transmission of local model parameters to a central server, effectively reducing communication overhead.
[0004] However, federated learning requires a large number of uploaded model parameters, meaning there's room for optimization in communication overhead. Furthermore, a variety of feature extraction techniques exist, and the appropriate one must be selected based on the unique scenarios of power grid systems. Therefore, how to efficiently train large models while protecting data privacy, thereby improving the effectiveness and security of power monitoring systems, remains a pressing challenge. Summary of the Invention
[0005] The purpose of this application is to overcome the above-mentioned defects in the prior art and provide a method and system for generating a large model of an electric power monitoring system.
[0006] This application provides a method for generating a large model of a power monitoring system, which is applied to a federated learning process between a power grid headquarters and power grid sub-nodes. The method includes:
[0007] The power grid sub-node obtains high-dimensional heterogeneous data of the power grid sub-node, and generates low-dimensional feature data by t-SNE dimension reduction, wherein the low-dimensional feature data includes a local structure;
[0008] The power grid headquarters distributes random seeds to the power grid sub-nodes;
[0009] The power grid sub-node initializes a local score under the local structure constraint based on the random seed;
[0010] The power grid headquarters initializes global model parameters, generates a global ranking based on the local scores, and broadcasts the global ranking to all the power grid nodes;
[0011] The power grid sub-node calculates a local ranking based on the low-dimensional feature data and the global ranking using an edge pop-up algorithm, generates a voting number based on the local ranking and uploads it to the power grid headquarters;
[0012] The power grid headquarters counts the total number of votes based on the number of votes, updates the global ranking based on the total number of votes, and generates a power monitoring model that minimizes average loss using the updated global ranking.
[0013] Optionally, generate low-dimensional feature data through t-SNE dimensionality reduction, including:
[0014] Calculating a high-dimensional similarity matrix of the high-dimensional heterogeneous data, wherein the similarity of each pair of data points is represented by a Gaussian distribution conditional probability;
[0015] Initialize the low-dimensional embedding space data and calculate the low-dimensional similarity matrix using the t distribution with 1 degree of freedom;
[0016] By minimizing the KL divergence between the high-dimensional similarity matrix and the low-dimensional similarity matrix, the low-dimensional embedding space data is iteratively updated until convergence.
[0017] Optionally, local ranking is calculated using an edge pop-up algorithm, including:
[0018] Generating a binary mask according to the local score, and removing edges in the binary mask with scores lower than a preset score to form a subnetwork;
[0019] Based on the subnetwork, updating the local score through a straight-through gradient estimator, wherein the discontinuous operation of the binary mask is an identity function in the back propagation of the straight-through gradient estimator;
[0020] The above steps are iterated until a local ranking matching the low-dimensional feature data is generated.
[0021] Optionally, counting the total number of votes includes:
[0022] Converting the local ranking of each of the power grid sub-nodes into a voting number, wherein the parameter ranking order of the power grid sub-nodes is positively correlated with the voting number;
[0023] The votes for the same parameter are weighted and summed to obtain the total votes.
[0024] Optionally, generating a power monitoring model that minimizes average loss using the updated global ranking includes:
[0025] Calculating the decrease rate of the local loss function value of all the power grid sub-nodes under the model corresponding to the global ranking;
[0026] If the descent rate is less than the preset threshold, the model is determined to have converged and the training is terminated. If the descent rate is not less than the preset threshold, the previous step is repeated until the maximum number of iterations is reached.
[0027] Optionally, a high-dimensional similarity matrix is calculated for high-dimensional heterogeneous data, including:
[0028] Apply the Gaussian kernel function to the Euclidean distance of each pair of data points to calculate the conditional probability and generate a similarity matrix;
[0029] Normalize the conditional probability of each row so that the sum of the probabilities of each row is 1;
[0030] A symmetric processing is performed on the similarity matrix.
[0031] Optionally, initialize the low-dimensional embedding space data, including:
[0032] Use a normal distribution with a mean of 0 and a variance of 1 to randomly generate low-dimensional coordinates;
[0033] Normalizing the low-dimensional coordinates to limit their numerical range to a preset interval;
[0034] The low-dimensional coordinates are updated by mini-batch gradient descent to initialize the low-dimensional embedding space data, where some data points are sampled in each iteration.
[0035] Optionally, generating a binary mask includes:
[0036] Generate a binary mask based on the preset subnetwork ratio.
[0037] The present application also provides a large-scale model generation system for a power monitoring system, including a power grid sub-node module and a power grid headquarters module connected via a wireless or wired network:
[0038] The power grid sub-node module obtains high-dimensional heterogeneous data of the power grid sub-nodes and generates low-dimensional feature data through t-SNE dimensionality reduction, wherein the low-dimensional feature data includes local structure;
[0039] The grid headquarters module distributes random seeds to the grid sub-node modules;
[0040] The grid sub-node module initializes a local score based on the random seed;
[0041] The power grid headquarters module initializes global model parameters, generates a global ranking based on the local scores, and broadcasts the global ranking to all the power grid node modules;
[0042] The power grid sub-node module calculates the local ranking based on the low-dimensional feature data and the global ranking through an edge pop-up algorithm, generates a voting number according to the local ranking and uploads it to the power grid headquarters module;
[0043] The power grid headquarters module counts the total number of votes based on the number of votes, updates the global ranking based on the total number of votes, and generates a power monitoring model that minimizes average loss based on the updated global ranking.
[0044] Optionally, the power grid node module generates low-dimensional feature data through t-SNE dimensionality reduction, including:
[0045] Calculating a high-dimensional similarity matrix of the high-dimensional heterogeneous data, wherein the similarity of each pair of data points is represented by a Gaussian distribution conditional probability;
[0046] Initialize the low-dimensional embedding space data and calculate the low-dimensional similarity matrix using the t distribution with 1 degree of freedom;
[0047] By minimizing the KL divergence between the high-dimensional similarity matrix and the low-dimensional similarity matrix, the low-dimensional embedding space data is iteratively updated until convergence.
[0048] Optionally, the power grid sub-node module calculates the local ranking by an edge pop-up algorithm, including:
[0049] Generating a binary mask according to the local score, and removing edges in the binary mask with scores lower than a preset score to form a subnetwork;
[0050] Based on the subnetwork, updating the local score through a straight-through gradient estimator, wherein the discontinuous operation of the binary mask is an identity function in the back propagation of the straight-through gradient estimator;
[0051] The above steps are iterated until a local ranking matching the low-dimensional feature data is generated.
[0052] Optionally, the grid headquarters module counting the total number of votes includes:
[0053] Converting the local ranking of each of the power grid sub-nodes into a voting number, wherein the parameter ranking order of the power grid sub-nodes is positively correlated with the voting number;
[0054] The votes for the same parameter are weighted and summed to obtain the total votes.
[0055] Optionally, the power grid headquarters module generates a power monitoring model that minimizes average loss based on the updated global ranking, including:
[0056] Calculating the decrease rate of the local loss function value of all the power grid sub-nodes under the model corresponding to the global ranking;
[0057] If the descent rate is less than the preset threshold, the model is determined to have converged and the training is terminated. If the descent rate is not less than the preset threshold, the previous step is repeated until the maximum number of iterations is reached.
[0058] Optionally, a high-dimensional similarity matrix is calculated for high-dimensional heterogeneous data, including:
[0059] Apply the Gaussian kernel function to the Euclidean distance of each pair of data points to calculate the conditional probability and generate a similarity matrix;
[0060] Normalize the conditional probability of each row so that the sum of the probabilities of each row is 1;
[0061] A symmetric processing is performed on the similarity matrix.
[0062] Optionally, initialize the low-dimensional embedding space data, including:
[0063] Use a normal distribution with a mean of 0 and a variance of 1 to randomly generate low-dimensional coordinates;
[0064] Normalizing the low-dimensional coordinates to limit their numerical range to a preset interval;
[0065] The low-dimensional coordinates are updated by mini-batch gradient descent to initialize the low-dimensional embedding space data, where some data points are sampled in each iteration.
[0066] Optionally, generating a binary mask includes:
[0067] Generate a binary mask based on the preset subnetwork ratio.
[0068] The beneficial effects of this application are:
[0069] The present application provides a method for generating a large model of an electric power monitoring system, which is applied to the federated learning process between the power grid headquarters and the power grid sub-nodes. The method includes: the power grid sub-node obtains the high-dimensional heterogeneous data of the power grid sub-node, generates low-dimensional feature data through t-SNE dimensionality reduction, and the low-dimensional feature data includes local structure; the power grid headquarters distributes random seeds to the power grid sub-nodes; the power grid sub-node initializes local scores based on the random seeds under the local structure constraints; the power grid headquarters initializes global model parameters, generates global rankings based on the local scores, and broadcasts the global rankings to all the power grid sub-nodes; the power grid sub-node calculates local rankings based on the low-dimensional feature data and the global rankings through an edge pop-up algorithm, generates votes based on the local rankings and uploads them to the power grid headquarters; the power grid headquarters counts the total number of votes based on the votes, updates the global ranking based on the total number of votes, and generates an electric power monitoring model with minimized average loss based on the updated global ranking. The present application effectively solves the problems of high data heterogeneity and high communication overhead in the power grid system by combining t-SNE dimensionality reduction technology and ranking federated learning mechanism. BRIEF DESCRIPTION OF THE DRAWINGS
[0070] Figure 1 This is a schematic diagram of the process of generating a large model of the power monitoring system in this application;
[0071] Figure 2 This is a schematic diagram of the online stage of generating a large model of the power monitoring system in this application;
[0072] Figure 3 This is a schematic diagram of the offline stage of generating a large model of the power monitoring system in this application. DETAILED DESCRIPTION
[0073] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that various forms of implementation of the present disclosure are not limited to the embodiments set forth herein. Rather, the embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.
[0074] The large-scale model generation for the power monitoring system utilizes a federated learning framework, comprised of a grid headquarters and multiple grid nodes. These two collaborate to generate the large-scale model. The core goal is to efficiently train an anomaly detection model applicable to global grid monitoring while protecting data privacy.
[0075] The power grid headquarters is responsible for aggregating, updating, and coordinating global model parameters, and does not directly access the raw data of sub-nodes. Grid sub-nodes are distributed across different regions and are responsible for local data collection, pre-processing, local model training, and ranking calculations, with all data retained locally throughout the process.
[0076] It consists of two stages:
[0077] In the offline phase, this phase is independently executed by the sub-nodes: the sub-nodes use t-SNE technology to reduce the dimensionality of high-dimensional heterogeneous data, retain local structural information, and provide adaptive input for subsequent federated learning.
[0078] The online phase is coordinated by the headquarters and branch nodes, including:
[0079] Initialization: The headquarters distributes random seeds, and the nodes initialize local model parameters.
[0080] Global ranking drive: The headquarters broadcasts the global ranking, and the nodes optimize the local model based on the ranking using the edge pop-up algorithm to generate local votes for upload.
[0081] Iterative convergence: The headquarters aggregates votes to update the global ranking until the average loss of the model is minimized.
[0082] The large-scale model generation method for the power monitoring system based on t-SNE and ranking technology proposed in this application mainly ensures that the original data can smoothly participate in the calculation of the online stage after being pre-processed in the offline stage by combining the offline stage. At the same time, the t-SNE dimension reduction method and the ranking mechanism significantly reduce the communication overhead in the model construction process and improve the construction efficiency. In this application, on the one hand, the data is retained locally to ensure data privacy; on the other hand, it can flexibly respond to the increase in data volume and data dimension in the power grid system, ensuring the smoothness of large-scale power grid model training.
[0083] Specifically, during the offline phase, each grid node uses t-SNE technology to preprocess data to ensure that the data can effectively participate in the online calculation phase. During the online phase, this solution reduces communication overhead by transmitting parameter ranking information instead of traditional gradient information. By combining t-SNE technology with the ranking mechanism, the overall grid model construction process is more efficient and smooth.
[0084] Please refer to Figures 1 to 3 As shown, the present application provides a method for generating a large model of a power monitoring system, which is applied to a federated learning process between a power grid headquarters and power grid sub-nodes. The method includes:
[0085] S101, the power grid sub-node obtains high-dimensional heterogeneous data of the power grid sub-node, and generates low-dimensional feature data through t-SNE dimensionality reduction, wherein the low-dimensional feature data includes local structure;
[0086] Each power grid node collects a high-dimensional heterogeneous data set X i ={x1,x2,…,x N,…}, such as dynamic monitoring data such as power consumption and power generation. Among them, X i , represents the i-th grid node, x N Indicates the Nth data.
[0087] Specifically, data on electricity consumption and power generation in a region is collected by various nodes. Each node uses t-SNE technology to extract features from the collected heterogeneous data and perform dimensionality reduction while preserving the data's inherent structure.
[0088] Similarity matrix calculation: Calculate the Euclidean distance d(x i ,x j ).
[0089] Calculate the conditional probability P using Gaussian distribution ij , satisfying the symmetry P ij =P ji , and satisfy the normalization condition for each row: the sum of the probabilities of each row is 1.
[0090] Low-dimensional space initialization: Randomly generate low-dimensional embedding data Y with mean 0 and variance 1.
[0091] Iterative optimization:
[0092] Calculate the low-dimensional similarity matrix Q between low-dimensional data points based on t distribution ij , and minimize the KL divergence of the high / low dimensional similarity difference:
[0093]
[0094] Among them, Q represents the similarity matrix of low-dimensional data Y, Q ij Represents the similarity between data points, and KL(P||Q) is used to represent the "information loss" from P to Q.
[0095] Update the low-dimensional coordinate Y by gradient descent until convergence, including the objective function change approaching zero.
[0096] Output the low-dimensional feature data Y after dimensionality reduction, retaining the local structural information of the original data for online model training.
[0097] By retaining structural information, the data can be fully clustered during model construction in the online stage, providing highly adaptable low-dimensional input for subsequent federated learning, avoiding the degradation of model performance caused by information loss in traditional dimensionality reduction methods.
[0098] S102, the power grid headquarters distributes random seeds to the power grid sub-nodes;
[0099] The power grid headquarters selects the finite field random seed SEED←Z q, and distribute it to all grid nodes. This ensures the consistency of the initialization of the model parameters and scores of each node, and achieves the reproducibility of the distributed system by sharing the seed. q represents a finite field.
[0100] S103, the grid sub-node initializes a local score under the local structure constraint based on the random seed;
[0101] After receiving the random seed, each node performs the following operations:
[0102] Initialize local model weights θ ω ←θ,SEED.
[0103] Initialize the model parameter score θ s ←θ,SEED.
[0104] Among them, θ ω represents the weight of the grid node, θ s represents the fraction of the grid model parameter θ.
[0105] The initialization process needs to ensure the independence of parameter distribution between nodes to avoid privacy leakage.
[0106] S104: The power grid headquarters initializes global model parameters, generates a global ranking based on the local scores, and broadcasts the global ranking to all the power grid nodes;
[0107] The first-round global ranking Rg1 is determined by the initialization score θs, and the calculation formula is:
[0108]
[0109] in, Represents the global ranking of the dth round.
[0110] In subsequent rounds, the ranking is updated based on the total number of votes. The steps are as follows:
[0111] Initial score reordering:
[0112]
[0113] Returns the normalized θ s ;
[0114] Updated scores:
[0115]
[0116] Returns the updated local score
[0117] Calculate local ranking:
[0118]
[0119] Where EP represents the edge pop-up algorithm, E represents the number of local rounds, and Ω m represents the local data of the grid node m, η represents the learning rate, Represents the local ranking of grid node m in round d.
[0120] The power grid headquarters will rank the current round globally Broadcast to all sub-nodes to drive the optimization direction of the local model.
[0121] S105: The power grid sub-node calculates a local ranking based on the low-dimensional feature data and the global ranking using an edge pop-up algorithm, generates a voting number based on the local ranking, and uploads it to the power grid headquarters;
[0122] The local scores of the parameters are re-ranked according to the global ranking of the current round of power grid model parameters received.
[0123] Edge Popup Algorithm (EP) performs:
[0124] Forward propagation: global ranking Re-rank local scores θ S , get the normalized score. Generate binary mask m=sort(θ S ), keep the high-scoring edges (mask is 1) and remove the low-scoring edges (mask is 0) in proportion:
[0125] t=int((1-k)*len(m))
[0126] Sub-network weight calculation: θ ω ⊙m, that is, the weight is multiplied by the mask element by element.
[0127] Use the masked sub-network for forward calculation and output the prediction result θω(b), where b is the input data.
[0128] The mask function h(·) (discontinuous) is regarded as the identity function (through) in back propagation, and the gradient is directly transferred.
[0129] Backpropagation: Calculating loss gradients in is the loss function, I v is the input of the neuron, and the edge score s is updated through the direct gradient estimator uv :
[0130]
[0131] Among them, W uv is a fixed weight, Z uis the output of the neurons in the previous layer.
[0132] Updated scores:
[0133]
[0134] Repeat the forward-backward step for E rounds and update the local score sm.
[0135] Calculating local ranking Convert rankings to votes And upload to headquarters.
[0136] S106. The power grid headquarters calculates the total number of votes based on the number of votes, updates the global ranking based on the total number of votes, and generates a power monitoring model that minimizes average loss based on the updated global ranking;
[0137] The power grid headquarters aggregates the votes of all sub-nodes and calculates the total number of votes:
[0138] v total =∑v m
[0139] Generate a new global ranking based on the total number of votes:
[0140]
[0141] Repeat S104-S106 until the global ranking is stable (average loss is minimized), and output the final power grid model.
[0142] This application uses a ranking federation mechanism to reduce high-dimensional parameter transmission and combines it with t-SNE dimensionality reduction preprocessing to reduce communication overhead.
[0143] t-SNE dimensionality reduction preserves the local structure of heterogeneous data, allowing low-dimensional features to be directly used for ranking calculations, avoiding ranking redundancy caused by high-dimensional data.
[0144] If only t-SNE is used, traditional gradient transmission still has high communication overhead; if only ranking federation is used, high-dimensional data requires more interaction rounds to converge. Combining these two methods can achieve dual optimization of communication and computational costs.
[0145] Furthermore, the mask mechanism in the edge pop-up algorithm needs to be aligned with the low-dimensional features after t-SNE dimensionality reduction to ensure that voting focuses on key structural information.
[0146] The present application also provides a large-scale model generation system for a power monitoring system, including a power grid sub-node module and a power grid headquarters module connected via a wireless or wired network:
[0147] The power grid sub-node module obtains high-dimensional heterogeneous data of the power grid sub-nodes and generates low-dimensional feature data through t-SNE dimensionality reduction, wherein the low-dimensional feature data includes local structure;
[0148] The grid headquarters module distributes random seeds to the grid sub-node modules;
[0149] The grid sub-node module initializes a local score based on the random seed;
[0150] The power grid headquarters module initializes global model parameters, generates a global ranking based on the local scores, and broadcasts the global ranking to all the power grid node modules;
[0151] The power grid sub-node module calculates the local ranking based on the low-dimensional feature data and the global ranking through an edge pop-up algorithm, generates a voting number according to the local ranking and uploads it to the power grid headquarters module;
[0152] The power grid headquarters module counts the total number of votes based on the number of votes, updates the global ranking based on the total number of votes, and generates a power monitoring model that minimizes average loss based on the updated global ranking.
[0153] Furthermore, the power grid node module generates low-dimensional feature data through t-SNE dimensionality reduction, including:
[0154] Calculating a high-dimensional similarity matrix of the high-dimensional heterogeneous data, wherein the similarity of each pair of data points is represented by a Gaussian distribution conditional probability;
[0155] Initialize the low-dimensional embedding space data and calculate the low-dimensional similarity matrix using the t distribution with 1 degree of freedom;
[0156] By minimizing the KL divergence between the high-dimensional similarity matrix and the low-dimensional similarity matrix, the low-dimensional embedding space data is iteratively updated until convergence.
[0157] Furthermore, the grid node module calculates the local ranking by using an edge pop-up algorithm, including:
[0158] Generating a binary mask according to the local score, and removing edges in the binary mask with scores lower than a preset score to form a subnetwork;
[0159] Based on the subnetwork, updating the local score through a straight-through gradient estimator, wherein the discontinuous operation of the binary mask is an identity function in the back propagation of the straight-through gradient estimator;
[0160] The above steps are iterated until a local ranking matching the low-dimensional feature data is generated.
[0161] Furthermore, the grid headquarters module calculates the total number of votes including:
[0162] Converting the local ranking of each of the power grid sub-nodes into a voting number, wherein the parameter ranking order of the power grid sub-nodes is positively correlated with the voting number;
[0163] The votes for the same parameter are weighted and summed to obtain the total votes.
[0164] Furthermore, the power grid headquarters module generates an average loss-minimizing power monitoring model based on the updated global ranking, including:
[0165] Calculating the decrease rate of the local loss function value of all the power grid sub-nodes under the model corresponding to the global ranking;
[0166] If the descent rate is less than the preset threshold, the model is determined to have converged and the training is terminated. If the descent rate is not less than the preset threshold, the previous step is repeated until the maximum number of iterations is reached.
[0167] Furthermore, the high-dimensional similarity matrix of high-dimensional heterogeneous data is calculated, including:
[0168] Apply the Gaussian kernel function to the Euclidean distance of each pair of data points to calculate the conditional probability and generate a similarity matrix;
[0169] Normalize the conditional probability of each row so that the sum of the probabilities of each row is 1;
[0170] A symmetric processing is performed on the similarity matrix.
[0171] Furthermore, initialize the low-dimensional embedding space data, including:
[0172] Use a normal distribution with a mean of 0 and a variance of 1 to randomly generate low-dimensional coordinates;
[0173] Normalizing the low-dimensional coordinates to limit their numerical range to a preset interval;
[0174] The low-dimensional coordinates are updated by mini-batch gradient descent to initialize the low-dimensional embedding space data, where some data points are sampled in each iteration.
[0175] Furthermore, generating a binary mask includes:
[0176] Generate a binary mask based on the preset subnetwork ratio.
[0177] The above description of the embodiments is intended to facilitate understanding and application of the present invention by those skilled in the art. It will be readily apparent to those skilled in the art that various modifications to the above embodiments can be made, and the general principles described herein can be applied to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the present disclosure are intended to fall within the scope of protection of the present invention.
Claims
1. A method for generating a large model of an electric power monitoring system, characterized in that: Applied to the federated learning process between the power grid headquarters and the grid nodes, the method includes: The power grid sub-node obtains high-dimensional heterogeneous data of the power grid sub-node, and generates low-dimensional feature data by t-SNE dimension reduction, wherein the low-dimensional feature data includes a local structure; The power grid headquarters distributes random seeds to the power grid sub-nodes; The power grid sub-node initializes a local score under the local structure constraint based on the random seed; The power grid headquarters initializes global model parameters, generates a global ranking based on the local scores, and broadcasts the global ranking to all the power grid nodes; The power grid sub-node calculates a local ranking based on the low-dimensional feature data and the global ranking using an edge pop-up algorithm, generates a voting number based on the local ranking and uploads it to the power grid headquarters; The power grid headquarters counts the total number of votes based on the number of votes, updates the global ranking based on the total number of votes, and generates a power monitoring model that minimizes average loss using the updated global ranking.
2. The method for generating a large model of a power monitoring system according to claim 1, characterized in that: Generate low-dimensional feature data through t-SNE dimensionality reduction, including: Calculating a high-dimensional similarity matrix of the high-dimensional heterogeneous data, wherein the similarity of each pair of data points is represented by a Gaussian distribution conditional probability; Initialize the low-dimensional embedding space data and calculate the low-dimensional similarity matrix using the t distribution with 1 degree of freedom; By minimizing the KL divergence between the high-dimensional similarity matrix and the low-dimensional similarity matrix, the low-dimensional embedding space data is iteratively updated until convergence.
3. The method for generating a large model of a power monitoring system according to claim 1, characterized in that: The local ranking is calculated using the edge pop-up algorithm, including: Generating a binary mask according to the local score, and removing edges in the binary mask with scores lower than a preset score to form a subnetwork; Based on the subnetwork, updating the local score through a straight-through gradient estimator, wherein the discontinuous operation of the binary mask is an identity function in the back propagation of the straight-through gradient estimator; The above steps are iterated until a local ranking matching the low-dimensional feature data is generated.
4. The method for generating a large model of a power monitoring system according to claim 1, characterized in that: The total number of votes counted includes: Converting the local ranking of each of the power grid sub-nodes into a voting number, wherein the parameter ranking order of the power grid sub-nodes is positively correlated with the voting number; The votes for the same parameter are weighted and summed to obtain the total votes.
5. The method for generating a large model of a power monitoring system according to claim 1, wherein: Generating an average loss-minimizing power monitoring model using the updated global ranking, comprising: Calculating the decrease rate of the local loss function value of all the power grid sub-nodes under the model corresponding to the global ranking; If the descent rate is less than the preset threshold, the model is determined to have converged and the training is terminated. If the descent rate is not less than the preset threshold, the previous step is repeated until the maximum number of iterations is reached.
6. The method for generating a large model of a power monitoring system according to claim 2, characterized in that: Calculate high-dimensional similarity matrices for high-dimensional heterogeneous data, including: Apply the Gaussian kernel function to the Euclidean distance of each pair of data points to calculate the conditional probability and generate a similarity matrix; Normalize the conditional probability of each row so that the sum of the probabilities of each row is 1; A symmetric processing is performed on the similarity matrix.
7. The method for generating a large model of a power monitoring system according to claim 2, characterized in that: Initialize low-dimensional embedding space data, including: Use a normal distribution with a mean of 0 and a variance of 1 to randomly generate low-dimensional coordinates; Normalizing the low-dimensional coordinates to limit their numerical range to a preset interval; The low-dimensional coordinates are updated by mini-batch gradient descent to initialize the low-dimensional embedding space data, where some data points are sampled in each iteration.
8. The method for generating a large model of a power monitoring system according to claim 3, characterized in that: Generating a binary mask comprises: Generate a binary mask based on the preset subnetwork ratio.
9. A large-scale model generation system for an electric power monitoring system, characterized in that: It includes power grid sub-node modules and power grid headquarters modules connected via wireless or wired networks: The power grid sub-node module obtains high-dimensional heterogeneous data of the power grid sub-nodes and generates low-dimensional feature data through t-SNE dimensionality reduction, wherein the low-dimensional feature data includes local structure; The grid headquarters module distributes random seeds to the grid sub-node modules; The grid sub-node module initializes a local score based on the random seed; The power grid headquarters module initializes global model parameters, generates a global ranking based on the local scores, and broadcasts the global ranking to all the power grid node modules; The power grid sub-node module calculates the local ranking based on the low-dimensional feature data and the global ranking through an edge pop-up algorithm, generates a voting number according to the local ranking and uploads it to the power grid headquarters module; The power grid headquarters module counts the total number of votes based on the number of votes, updates the global ranking based on the total number of votes, and generates a power monitoring model that minimizes average loss based on the updated global ranking.
10. A large-scale model generation system for a power monitoring system according to claim 9, characterized in that: The power grid node module generates low-dimensional feature data through t-SNE dimensionality reduction, including: Calculating a high-dimensional similarity matrix of the high-dimensional heterogeneous data, wherein the similarity of each pair of data points is represented by a Gaussian distribution conditional probability; Initialize the low-dimensional embedding space data and calculate the low-dimensional similarity matrix using the t distribution with 1 degree of freedom; By minimizing the KL divergence between the high-dimensional similarity matrix and the low-dimensional similarity matrix, the low-dimensional embedding space data is iteratively updated until convergence.
11. The large-scale model generation system for a power monitoring system according to claim 9, characterized in that: The grid node module calculates the local ranking using an edge pop-up algorithm, including: Generating a binary mask according to the local score, and removing edges in the binary mask with scores lower than a preset score to form a subnetwork; Based on the subnetwork, updating the local score through a straight-through gradient estimator, wherein the discontinuous operation of the binary mask is an identity function in the back propagation of the straight-through gradient estimator; The above steps are iterated until a local ranking matching the low-dimensional feature data is generated.
12. The large-scale model generation system for a power monitoring system according to claim 9, characterized in that: The grid headquarters module counts the total number of votes including: Converting the local ranking of each of the power grid sub-nodes into a voting number, wherein the parameter ranking order of the power grid sub-nodes is positively correlated with the voting number; The votes for the same parameter are weighted and summed to obtain the total votes.
13. The large-scale model generation system for a power monitoring system according to claim 9, characterized in that: The power grid headquarters module generates a power monitoring model that minimizes average loss based on the updated global ranking, including: Calculating the decrease rate of the local loss function value of all the power grid sub-nodes under the model corresponding to the global ranking; If the descent rate is less than the preset threshold, the model is determined to have converged and the training is terminated. If the descent rate is not less than the preset threshold, the previous step is repeated until the maximum number of iterations is reached.
14. The large-scale model generation system for a power monitoring system according to claim 10, characterized in that: Calculate high-dimensional similarity matrices for high-dimensional heterogeneous data, including: Apply the Gaussian kernel function to the Euclidean distance of each pair of data points to calculate the conditional probability and generate a similarity matrix; Normalize the conditional probability of each row so that the sum of the probabilities of each row is 1; A symmetric processing is performed on the similarity matrix.
15. The large-scale model generation system for a power monitoring system according to claim 10, characterized in that: Initialize low-dimensional embedding space data, including: Use a normal distribution with a mean of 0 and a variance of 1 to randomly generate low-dimensional coordinates; Normalizing the low-dimensional coordinates to limit their numerical range to a preset interval; The low-dimensional coordinates are updated by mini-batch gradient descent to initialize the low-dimensional embedding space data, where some data points are sampled in each iteration.
16. A large-scale model generation system for a power monitoring system according to claim 11, characterized in that: Generating a binary mask comprises: Generate a binary mask based on the preset subnetwork ratio.