A method for identifying household change relationships based on single-step active annotation
By using a single-step active annotation method and optimizing the annotation budget with a feature extraction network and classifier, the problem of high cost and low accuracy in identifying the topological relationship between users and distribution transformers in low-voltage power distribution systems is solved, achieving efficient and low-cost identification results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID HUBEI ELECTRIC POWER RES INST
- Filing Date
- 2026-02-13
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies for identifying the topological relationship between users and distribution transformers in low-voltage power distribution systems are costly, time-consuming, and susceptible to noise and communication interference, especially in areas with uneven user numbers where the accuracy of identification decreases.
We adopt a single-step active annotation method, which uses a feature extraction network to obtain user data features, build a classifier and train it with manual annotation, optimize the annotation budget using uncertainty and diversity criteria, directly process noisy data, and reduce manual costs.
It enables direct processing of noisy data without installing additional equipment, achieving accuracy close to manual screening. The process is simple, the annotation cost is low, it is suitable for scenarios with uneven user numbers, and the recognition accuracy is high.
Smart Images
Figure CN122132943A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power grid branch topology identification technology, and relates to a method for identifying household-transformer relationships based on single-step active annotation. Background Technology
[0002] In low-voltage power distribution systems, accurately identifying the topological relationship between users (meters) and distribution transformers (substations) is crucial for power distribution network operation monitoring, fault location, and intelligent management. However, existing technologies primarily rely on two methods to obtain this relationship: manual on-site inspection or signal injection, and algorithms based on unsupervised clustering or graph models. Manual on-site inspection or signal injection requires manual verification or the deployment of signal injection / receiving equipment on the lines, resulting in high costs, long processing times, and susceptibility to noise and communication interference. Algorithms based on unsupervised clustering or graph models typically assume noise-free data and a balanced number of users in each distribution area, generating topological relationships through similarity analysis, clustering, or graph models. However, in real-world scenarios, user voltage / current time-series data suffers from various noise issues, missing data, and clock asynchrony problems, leading to decreased identification accuracy. Furthermore, unsupervised methods cannot incorporate prior knowledge or focus on key samples, resulting in blurred decision boundaries, particularly in areas with unevenly distributed user numbers.
[0003] Therefore, a new household change relationship identification method based on single-step active annotation is needed to solve the above problems. Summary of the Invention
[0004] The purpose of this invention is to provide a method for identifying household change relationships based on single-step active annotation, so as to solve the problems of the prior art.
[0005] The technical solution of the present invention is as follows: 1) Obtain all user data All user data Input a feature extraction network and output c-dimensional features. ; 2) Establish a classifier, which is represented by the following formula: , In the formula, For each user, the normalized features are, where, S is the optimization objective, dist() is the distance metric function, N is the total number of users, and B is the labeling budget. This is an unlabeled user dataset, including voltage data from N users in k transformer substations. Label for users, For user data, User data after feature extraction; 3) c-dimensional features obtained from step 1) Select B users for manual annotation to obtain a subset of data. Using data subsets The classifier is trained, and the remaining unlabeled dataset is processed. Input the classifier to obtain the final household change relationship identification result.
[0006] Furthermore, the loss function of the classifier described in step 2) is: , In the formula, For the uncertainty criterion formula, The formula for diversity labeling criteria, This is the hyperparameter for regularization strength; , In the formula, KL represents the KL divergence. For the i-th user parameter vector Mixed Gaussian kernel density estimation, Let E be the Gaussian kernel density estimate of the data distribution of the i-th user with noise, and E be the expected value. For unlabeled feature sets, The normalized features of the i-th user; , In the formula, For the j-th parameter vector, For the k-th parameter vector, The temperature parameter controls the smoothness of the distribution, and B is the labeling budget. This formula is the objective function for maximum likelihood estimation, serving as the optimization objective during training to narrow the distance between features and class centers, enabling the model to classify correctly. The parameter can be adjusted according to different user data distributions, and is generally set to 1.
[0007] Furthermore, the parameter vector The mixture Gaussian kernel density estimate is expressed by the following equation: , In the formula, For Gaussian mixture kernel density estimation of parameter vectors, For normalization term, Temperature parameters used to control the smoothness of the distribution, The distribution of noisy user data obtained by clustering algorithm. For the normalized features of the i-th user, Let B be a parameter vector containing B objectives to be optimized, where B is the labeled budget. For the j-th parameter vector, This represents the partition function, which is the sum of the exponential scores for all possible categories.
[0008] Furthermore, the parameter vector The following equation is obtained through iteration: , In the formula, The distribution of noisy user data obtained by clustering algorithm. For the normalized features of the i-th user, For normalization term, For the j-th parameter vector, Temperature parameters used to control the smoothness of the distribution, This represents the probability predicted by the model.
[0009] Furthermore, the feature extraction network described in step 1) is a multilayer perceptron (MLP).
[0010] Furthermore, the feature extraction network described in step 1) includes three fully connected layers with dimensions of 64, 128, and c, respectively.
[0011] Furthermore, in step 1), c is 256.
[0012] Beneficial effects: The household change relationship identification method based on single-step active annotation of this invention can directly process noisy data without installing additional equipment, and its accuracy is close to that of manual screening. The process is simple and easy to deploy. Single-step active annotation does not require multiple rounds of interaction, reducing the complexity of annotation management. Only a small amount of manual annotation cost is required to achieve the same recognition accuracy as manual screening. Attached Figure Description
[0013] Figure 1 This is a flowchart of a household change relationship identification method based on single-step active annotation. Detailed Implementation
[0014] Please see Figure 1 As shown, the household change relationship identification method based on single-step active annotation of the present invention includes the following steps:
[0015] Step 1: Unsupervised feature extraction and clustering: Noisy unlabeled user dataset Total of Each TV station area There are [number] users, and the voltage data sequence length for each user is [length]. All user data is input into a feature extraction network consisting of a multilayer perceptron composed of fully connected layers. The data and sampling time are encoded, and c-dimensional features are output. Clustering algorithms are used in high-dimensional space to obtain a low-dimensional manifold representation of user data in high-dimensional space, which is then used for subsequent single-step active annotation analysis.
[0016] Step 2: Single-step active annotation sample selection:
[0017] Given annotation budget The goal of the algorithm is to In the case of, from Select a subset from The algorithm then performs manual annotation and uses the labeled data for supervised training to achieve near-fully supervised performance. Specifically, the algorithm will extract data from a high-dimensional feature space. Select Each user manually labeled the data, then supervised training was performed using labeled data, and the results were analyzed through... Add a category module to output the categories each user belongs to. Confidence level of each station area.
[0018] Step 2.1 Feature normalization and optimization target construction:
[0019] Let the normalized features of each user be denoted as . and satisfy ,in If the feature extractor is composed of multilayer perceptrons, then the problem of "representing the distribution of user data in the entire distribution area with as little user data as possible" can be modeled as an optimization problem: , in, To optimize the objective, For distance measurement. The method will Constructing to include The parameter vector of the target to be optimized Aligning feature dimensions with high-dimensional features of user data Let represent the expected feature center of high-value users. Then, the Gaussian kernel density estimate formed by the parameter vectors is: , in, For normalization term, Temperature parameters used to control the smoothness of the distribution, This represents the distribution of noisy user data obtained by a clustering algorithm.
[0020] Step 2.2 Optimization of Uncertainty Criteria:
[0021] The KL divergence of the distribution is used to represent the uncertainty of the samples. The optimization objective of the method is to make the feature distribution of the parameter vector as close as possible to the user feature distribution, expressed as:
[0022] Where KL represents the KL divergence, substituting the Gaussian mixture kernel density estimate into the equation transforms the problem of minimizing the distribution distance into finding its maximum value, and the problem of minimizing the distribution distance into a solvable optimization problem.
[0023] By iterating through the above formula, the optimized parameter vector can eventually be obtained. Through calculation and The B users with the most similar feature distributions are selected as the samples to be labeled. The distance metric used is Manhattan distance. The optimized result is... The distribution is close to the distribution of the entire data to be labeled, which greatly improves the labeling efficiency.
[0024] Step 2.3 Optimization of Diversity Criteria
[0025] In practical applications, it was found that when the number of users in a distribution area is uneven, this method tends to allocate the annotation budget to distribution areas with a large number of users in order to fully fit the dataset. This wastes valuable annotation budget in these areas, while distribution areas with fewer users may not be able to fit effectively. To address this, a diversity annotation criterion is added, which aims to increase the feature distance between labeled samples as much as possible, preventing the annotation budget from being concentrated in high-user distribution areas and instead distributing it as widely as possible among the distribution areas. The formula for the diversity annotation criterion is: , The final optimization objective can be expressed as: ,,
[0026] Step 3: Active labeling and household transformer identification
[0027] Manually select a subset of data The dataset was labeled, and a fully connected layer was added after the original feature extractor to act as a classifier for training. Then, the remaining unlabeled dataset was processed. Input the classifier to obtain the final documentation result.
[0028] Existing methods for identifying low-voltage transformer substations require additional measurement equipment, resulting in high installation and maintenance costs, and they may fail in complex scenarios. Unsupervised clustering algorithms require manual noise filtering and screening of the data, consuming significant manpower and resources. This invention eliminates the need for additional equipment and can directly process noisy data, while achieving accuracy close to that of manual screening.
[0029] Compared to fully supervised algorithms, this invention requires only one-quarter of the manual annotation cost to achieve almost the same recognition accuracy. Compared to unsupervised clustering combined with simple uncertainty labeling, this invention improves accuracy by more than 20% when the annotation budget coefficient is 0.2.
[0030] This invention exhibits excellent robustness and generalization ability, enabling it to directly process data containing missing, asynchronous, and anomalous elements without the need for manual data cleaning. The sample selection balances representativeness and diversity, effectively covering different types and sizes of transformer substations. Especially in scenarios with uneven user numbers, its accuracy approaches that of fully supervised algorithms.
[0031] The invention features a simple and easy-to-deploy process, with single-step active annotation eliminating the need for multiple rounds of interaction, thus reducing the complexity of annotation management. It can also be integrated with existing smart power grid platforms and implemented through software, thereby enhancing its engineering application value.
[0032] Example 1
[0033] The test dataset consists of meter current and voltage data from a real power supply area in a province, including 8 transformers and 2913 users. The power supply area is located in the old city area, with a complex user structure, including 2278 residential community users (R), 142 street-front commercial users (C), and 307 self-built households (S). Other users are public infrastructure (P), including communication base stations, public lighting, photovoltaic power stations, police stations, and electric vehicle charging stations. The user information for the dataset is shown in Table 1.
[0034] Table 1. Dataset User Distribution
[0035] This experiment collected voltage data from users at 96 points per day for one month, from March 1st to 30th, 2025. All data were raw data with noise. The known clock deviation range for user data is 5–30 minutes, the meter voltage data error accuracy is ±2%, and approximately 20% of users have missing data, with a missing rate of about 5%, the most severe missing rate being 20%. To facilitate this experiment, with the assistance of the power company, all transformer-customer files within the power supply area were manually checked, organized, and marked.
[0036] First, we tested the experimental results under different labeling budgets. All profile labels were removed from the dataset. We set the labeling budget to B = {0.05, 0.1, 0.15, 0.2, 0.3}N and compared it with the random selection strategy. The random selection strategy simulates the performance of supervised algorithms under insufficient budget conditions. Profiled users were used to train the transformer area classifier, and the remaining unlabeled users were input into the trained classifier for identification. Finally, we also tested the experimental results in a fully supervised scenario, i.e., the dataset was divided into training and test sets in a 4:1 ratio, equivalent to a scenario with a labeling budget of 0.8. The experimental results are shown in the table below:
[0037] Table 2 Experiment with different budget coefficients
[0038] The experimental results show that when the annotation budget coefficient is only 0.05, that is, when only 146 users are manually verified, the model performance significantly exceeds that of the random annotation strategy. When the annotation budget is increased to 0.2, it can achieve an accuracy close to that of full supervision, reaching 97.8%, while the annotation cost is only one-quarter of that of full supervision.
Claims
1. A method for identifying household change relationships based on single-step active annotation, characterized in that, Includes the following steps: 1) Obtain all user data All user data Input a feature extraction network and output c-dimensional features. ; 2) Establish a classifier, which is represented by the following formula: ， In the formula, For each user, the normalized features are, where, S is the optimization objective, dist() is the distance metric function, N is the total number of users, and B is the labeling budget. This is an unlabeled user dataset, including voltage data from N users in k transformer substations. Label for users, For user data, User data after feature extraction; 3) c-dimensional features obtained from step 1) Select B users for manual annotation to obtain a subset of data. Using data subsets The classifier is trained, and the remaining unlabeled dataset is processed. Input the classifier to obtain the final household change relationship identification result.
2. The household change relationship identification method based on single-step active annotation as described in claim 1, characterized in that, The loss function of the classifier in step 2) is: ， In the formula, For the uncertainty criterion formula, The formula for diversity labeling criteria, This is the hyperparameter for regularization strength; ， In the formula, KL represents the KL divergence. For the i-th user parameter vector Mixed Gaussian kernel density estimation, Let E be the Gaussian kernel density estimate of the data distribution of the i-th user with noise, and E be the expected value. For unlabeled feature sets, The normalized features of the i-th user; ， In the formula, For the j-th parameter vector, For the k-th parameter vector, The temperature parameter used to control the smoothness of the distribution is B, which is the standard budget.
3. The household change relationship identification method based on single-step active annotation as described in claim 2, characterized in that, The parameter vector The mixture Gaussian kernel density estimate is expressed by the following equation: ， In the formula, For Gaussian mixture kernel density estimation of parameter vectors, For normalization term, Temperature parameters used to control the smoothness of the distribution, The distribution of noisy user data obtained by clustering algorithm. For the normalized features of the i-th user, Let B be a parameter vector containing B objectives to be optimized, where B is the labeled budget. For the j-th parameter vector, This represents the partition function, which is the sum of the exponential scores for all possible categories.
4. The household change relationship identification method based on single-step active annotation as described in claim 2, characterized in that, The parameter vector The following equation is obtained through iteration: ， In the formula, The distribution of noisy user data obtained by clustering algorithm. For the normalized features of the i-th user, For normalization term, For the j-th parameter vector, Temperature parameters used to control the smoothness of the distribution, This represents the probability predicted by the model.
5. The household change relationship identification method based on single-step active annotation as described in claim 2, characterized in that, The feature extraction network described in step 1) is a multilayer perceptron (MLP).
6. The household change relationship identification method based on single-step active annotation as described in claim 1, characterized in that, The feature extraction network described in step 1) consists of three fully connected layers with dimensions of 64, 128 and c, respectively.
7. The household change relationship identification method based on single-step active annotation as described in claim 1, characterized in that, In step 1), c is 256.