A remote meter reading method for concentrators

Abstract

The present application relates to the field of data processing, and more particularly to a remote meter reading method for a concentrator, comprising: collecting historical sample data, each historical sample data comprising metering feature data, state feature data and communication feature data; obtaining a plurality of clustering clusters by using K-means clustering according to the calculated optimal K value, and assigning a label to each clustering cluster, the label comprising normal or abnormal; obtaining real-time sample data after preprocessing, calculating the Euclidean distance between the sample data and each clustering cluster, obtaining the label of the sample data according to the Euclidean distance, and determining that the collected sample data is incorrect if the label of the sample data is abnormal. The present application dynamically adjusts the K value and evaluates the stability of the cluster, so that the model can better adapt to new data. When the data distribution changes, the model can automatically adjust the K value to adapt to the new data structure, improving the accuracy of anomaly detection.

Description

Technical Field

[0001] This invention relates to the field of data processing. In particular, it relates to a remote meter reading method for a concentrator. Background Technology

[0002] In power systems, remote meter reading technology using concentrators is a crucial link in achieving automatic collection and management of electricity consumption information. By remotely reading data from smart meters or data acquisition terminals through concentrators, electricity data collection and analysis can be efficiently completed, providing important data for power system operation management, user billing, power dispatching, and fault diagnosis. However, in actual remote meter reading processes, anomalies may occur in the meter reading data. Abnormal meter reading data may mask the actual operating status of power equipment, leading to untimely detection and handling of equipment faults, or errors in user electricity bill calculations. Therefore, detecting abnormal meter reading data is particularly important.

[0003] Chinese patent application CN112819299A discloses a center-optimized differential K-means load clustering method, comprising: collecting load data of user electricity consumption; obtaining initial cluster centers based on the local density of load data curves; calculating the differential Euclidean distance of the load data, and combining the differential Euclidean distance and the Euclidean distance of the load data as a similarity measure of the load data; randomly selecting K initial cluster centers as cluster centers for the load data, performing clustering based on the similarity measure and updating the cluster centers, and outputting the K divided clusters and the cluster label of each cluster after the update iteration is completed; calculating the mean of the load data curves in each category of clusters based on the cluster labels to obtain the typical load curve of each category of user, and determining the electricity consumption type of each type of residential electricity user.

[0004] However, the actual data distribution changes with time, environment, or business needs. A fixed K value and initial center cannot adapt to the new data distribution, causing the clustering results to be out of sync with the actual data distribution, affecting the accuracy of anomaly detection. In addition, if the K value is mismatched, normal data may be misjudged as anomalies (false alarms) or real anomaly data may be missed (false negatives), further reducing the reliability of anomaly detection. Summary of the Invention

[0005] To address the technical problem of using a fixed K value to cluster meter reading data, which cannot adapt to data changes and results in low anomaly detection accuracy, this invention provides a remote meter reading method for concentrators.

[0006] A remote meter reading method for a concentrator includes:

[0007] Collect historical sample data, each of which includes measurement characteristic data, status characteristic data, and communication characteristic data;

[0008] Based on the calculated optimal K value, K-means clustering is used to obtain multiple clusters, and each cluster is assigned a label, which includes normal or abnormal.

[0009] Obtain preprocessed real-time sample data, calculate the Euclidean distance between the sample data and each cluster, obtain the label of the sample data based on the Euclidean distance, and determine that the collected sample data is incorrect if the label of the sample data is abnormal.

[0010] The calculation process for the optimal K value includes:

[0011] Based on a given initial K value, historical sample data are clustered to obtain multiple initial clusters;

[0012] Calculate the stability score of the initial clusters. When the stability score is greater than or equal to a preset threshold, the corresponding initial cluster is marked as a stable cluster. When the stability score is less than the preset threshold, the corresponding initial cluster is marked as an unstable cluster. When an unstable cluster meets the first condition, it is split to obtain a new cluster. When an unstable cluster meets the second condition, it is merged with other clusters to obtain a new cluster.

[0013] The stability scores of the clusters generated after splitting or merging operations, as well as the original stable clusters, are counted. The stability scores of these clusters are recalculated until no further splitting or merging operations are performed. The number of clusters at this point is the optimal K value.

[0014] Preferably, the process of obtaining the first condition is as follows:

[0015] Calculate the intra-cluster dispersion of all initial clusters, determine the splitting threshold based on the percentage method, and take the intra-cluster dispersion of unstable clusters that is greater than the splitting threshold as the first condition.

[0016] If the intra-cluster dispersion of an unstable cluster is greater than the splitting threshold, it indicates that the intra-cluster differences of the unstable cluster are large, and it needs to be split into smaller clusters to improve the stability and accuracy of the final clustering results.

[0017] Preferably, the process of obtaining the second condition is as follows:

[0018] Calculate the inter-cluster distances of all initial clusters, determine the merging threshold based on the percentage method, and use the condition that the inter-cluster distance between an unstable cluster and a certain cluster is less than or equal to the merging threshold as a second condition.

[0019] If the distance between an unstable cluster and other clusters is less than or equal to the merging threshold, it indicates that the two clusters are highly similar and can be merged to form a new cluster, thereby improving the stability and accuracy of the final clustering results.

[0020] Preferably, the calculation process for the optimal K value also includes:

[0021] When an unstable cluster does not meet either the first or the second condition, it is determined whether the unstable cluster meets the third condition. If it does, the unstable cluster is merged into the cluster with the smallest inter-cluster distance to obtain a new cluster; otherwise, it is treated as a stable cluster and participates in subsequent steps.

[0022] Preferably, the process for obtaining the third condition is as follows:

[0023] Calculate the ratio of the number of historical sample data contained in the unstable cluster to the total number of historical sample data, and use a ratio that is less than a preset value as the third condition.

[0024] Preferably, the stability score is calculated as follows:

[0025] For each initial cluster, several perturbation clusters are generated. The Jaccard index values ​​between the initial cluster and each of its corresponding perturbation clusters are calculated, and the mean of all corresponding Jaccard index values ​​is taken as the stability score of the initial cluster.

[0026] Preferably, the process of generating several perturbation clusters is as follows:

[0027] Each initial cluster contains multiple historical sample data. For each initial cluster, the Bootstrap sampling method is used to extract historical sample data. This process is repeated to obtain several perturbation subsets. For each perturbation subset, a preset proportion of Gaussian noise is added, and clustering is performed to obtain perturbation clusters.

[0028] By adding tiny Gaussian noise to the data, we can simulate the uncertainty of data in the real world. This perturbation can test the robustness of clustering results in the face of small random fluctuations.

[0029] Preferably, the historical sample data includes multiple feature dimensions, and the process of adding Gaussian noise is as follows:

[0030] Noise is added to the feature data of each feature dimension of the historical sample data in the perturbation subset, where the noise is 5% of the standard deviation of the corresponding feature dimension in the corresponding initial cluster.

[0031] In the real world, data always contains noise and uncertainty. By adding Gaussian noise with a standard deviation of 5%, natural fluctuations can be simulated, making the clustering results more realistic.

[0032] Preferably, the intra-cluster dispersion is the average Euclidean distance from all historical sample data within the initial cluster to the cluster center.

[0033] The present invention has the following effects:

[0034] 1. This invention classifies clusters by their stability scores, splits or merges unstable clusters, dynamically adjusts the K value, and automatically finds the optimal number of clusters based on the inherent structure of the data, ensuring that the final clustering results have high stability and accuracy. This method avoids the deviation of clustering results caused by manually preset K values, making the clustering results more stable and reliable.

[0035] 2. This invention enables the model to better adapt to new data by dynamically adjusting the K value and evaluating the stability of the cluster. When the data distribution changes, the model can automatically adjust the K value to adapt to the new data structure, thereby improving the accuracy of anomaly detection.

[0036] 3. By detecting anomalies in the remote meter reading data of the concentrator in real time, the accuracy and reliability of meter reading data can be significantly improved, power equipment failures can be prevented and diagnosed, power resource allocation can be optimized, the safety and stability of the power system can be improved, the operation and management of the smart grid can be supported, and the economic efficiency of the power system can be enhanced. Attached Figure Description

[0037] Figure 1 This is a flowchart of steps S1-S3 in a remote meter reading method for a concentrator according to an embodiment of the present invention.

[0038] Figure 2 This is a flowchart of steps S20-S25 in a remote meter reading method for a concentrator according to an embodiment of the present invention. Detailed Implementation

[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0040] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0041] K-means clustering is a classic unsupervised learning method used to divide n d-dimensional samples into K clusters, minimizing intra-cluster scatter. Traditional K-means clustering requires pre-setting a value for K and randomly initializing K initial centers; different K values ​​and initial centers can lead to significantly different clustering results. This invention addresses the issue that a fixed, pre-set K value cannot adapt to data variations, causing the clustering results to deviate from the actual data distribution and affecting anomaly detection accuracy. Therefore, it proposes a remote meter reading method that uses the stable value of the data itself to split or merge clusters, dynamically adjusting the K value to adapt to changes in data distribution.

[0042] Reference Figure 1 A remote meter reading method for a concentrator includes steps S1-S3, as detailed below:

[0043] S1. Collect historical sample data. Each historical sample data includes measurement characteristic data, status characteristic data, and communication characteristic data.

[0044] Historical (e.g., the previous 3 months) electricity consumption data is collected from smart meters or data acquisition terminals connected to the concentrator, including active power (unit: kW, reflecting electricity load) and daily electricity consumption (unit: kWh, reflecting total electricity consumption). Historical operation logs of the remote meter reading system are collected from the concentrator's stored records, including user type data (residential / commercial / industrial) and power supply voltage (unit: V / A, reflecting power supply stability). Historical communication transmission quality data is collected from the communication module between the concentrator and the master station / terminal, including signal strength (unit: dBm, reflecting wireless signal quality) and transmission latency (unit: ms, the time it takes for data to travel from the terminal to the master station). Missing and duplicate values ​​are removed from all data.

[0045] Data with different dimensions and value ranges are normalized: Numerical data are normalized using Min-Max standardization, such as active power, daily electricity consumption, and power supply voltage; categorical data are encoded using one-hot encoding, converting each category into a binary feature, such as user type data. All collected data are standardized to the range [0, 1].

[0046] Key features for each sample are extracted based on the normalized data, including three categories: metering features, status features, and communication features. Each category has its corresponding normalized sub-features. Metering features correspond to active power and daily electricity consumption; status features correspond to user type and power supply voltage; and communication features correspond to signal strength and transmission delay. The three categories of features for all samples are concatenated sequentially to construct their corresponding feature vectors, ensuring consistent feature vector structures, thus obtaining all historical sample data. In this embodiment, the historical sample data package contains seven feature dimensions: active power, daily electricity consumption, residential power supply voltage, commercial power supply voltage, industrial power supply voltage, signal strength, and transmission delay.

[0047] S2. Based on the calculated optimal K value, use K-means clustering to obtain multiple clusters, and assign a label to each cluster, including normal or abnormal.

[0048] Traditional K-means clustering requires pre-setting the K value, but the cluster structure of actual data is often unknown. Dynamically adjusting the K value can automatically find the optimal number of clusters based on the inherent structure of the data, avoiding bias in clustering results caused by inappropriate K value selection. Clustering based on the optimal K value ensures that data points within each cluster have high similarity, while data points between different clusters have significant differences, making the clustering results clearer and improving the accuracy of anomaly detection. By assigning "normal" or "abnormal" labels to each cluster, abnormal behavior in the data can be quickly identified.

[0049] Among them, such as Figure 2 As shown, the calculation process for the optimal K value includes steps S20-S25, as detailed below:

[0050] S20. Based on the given initial K value, cluster the historical sample data to obtain multiple initial clusters.

[0051] In this embodiment, the initial K value is set to 5. Five different sample data are selected from all historical sample data as initial cluster centers. The Euclidean distance from each historical sample data to the five initial cluster centers is calculated. Each historical sample data is assigned to the cluster corresponding to the initial cluster center with the closest Euclidean distance. The average value of all historical sample data in each cluster is recalculated and used as the new cluster center. The steps of assigning historical sample data and updating cluster centers are repeated until the cluster centers no longer change. The initial clustering is then completed, resulting in five initial clusters.

[0052] S21. Calculate the stability score of the initial cluster. When the stability score is greater than or equal to the preset threshold, mark the corresponding initial cluster as a stable cluster; when the stability score is less than the preset threshold, mark the corresponding initial cluster as an unstable cluster.

[0053] The stability score of the initial cluster is calculated as follows: For each initial cluster, the Bootstrap sampling method is used to extract 80% of the historical sample data from the historical sample dataset of the initial cluster with replacement, generating a perturbation subset. The above sampling steps are repeated until 10 perturbation subsets are obtained.

[0054] Calculate the standard deviation of each feature dimension in the initial cluster. Add noise to the feature data of each feature dimension of the historical sample data in the perturbation subset. The noise is 5% of the standard deviation of the corresponding feature dimension in the initial cluster, thus obtaining 10 perturbation datasets. The calculation formula for adding noise is as follows:

[0055]

[0056] In the formula, This represents the first historical sample data of the initial cluster. Standard deviation of each feature dimension; This indicates that the mean is 0 and the variance is 0. Gaussian distribution; Indicates the first subset of the perturbation The first historical sample data 3D eigenvalues; This indicates the first element in the perturbation subset after adding noise. The first historical sample data 3D eigenvalues. This allows noise to be added to each historical sample data point in the perturbation subset, more realistically simulating random perturbations during data acquisition;

[0057] Perform K-means clustering with a K value of 1 on each perturbation dataset to generate 10 perturbation clusters. Calculate the Jaccard index value between the initial cluster and each of its corresponding perturbation clusters, and take the mean of all the corresponding Jaccard index values ​​as the stability score of the initial cluster.

[0058] The calculation of the Jaccard index value follows the formula:

[0059]

[0060] In the formula, Indicates the initial cluster; Represents the first cluster corresponding to the initial cluster. A cluster of disturbances; This represents the historical sample dataset in the initial cluster; Indicates the first A sample dataset of perturbation clusters; This represents the number of samples that belong to both the initial cluster and the perturbation cluster; This represents the total number of samples in the initial cluster and the perturbation cluster; Indicates the first The initial cluster and its corresponding perturbation cluster in the i-th initial cluster are the i-th... The Jaccard index value of each perturbation cluster is in the range of [0,1]. The larger the index value, the higher the similarity between the initial cluster and the perturbation cluster, and the less affected the initial cluster is by the perturbation, that is, the more stable the initial cluster is.

[0061] In this embodiment, the preset threshold is taken as an empirical value of 0.7. In other embodiments, it can also be adjusted according to the business background. When the stability score of the initial cluster is greater than or equal to 0.7, it indicates that the initial cluster is not sensitive to disturbances and the sample composition is consistent. The initial cluster can be determined to be a stable cluster and no adjustment is needed. When the stability score is less than 0.7, it indicates that the initial cluster is relatively sensitive to disturbances and the initial cluster is marked as an unstable cluster.

[0062] The K-means algorithm is prone to getting trapped in local optima, especially when the initial cluster centers are poorly chosen. Data perturbation can generate multiple different perturbed datasets, increasing the chances of finding the global optimum. Furthermore, in practical applications, data is often noisy; adding small amounts of Gaussian noise can simulate the uncertainty of real-world data, and this perturbation can test the robustness of the clustering results to small random fluctuations.

[0063] S22. When an unstable cluster satisfies the first condition, the unstable cluster is split to obtain a new cluster.

[0064] If an initial cluster is marked as unstable, it is necessary to first determine whether the initial cluster meets the first condition. The process of obtaining the first condition is as follows: Calculate the intra-cluster dispersion of all initial clusters. The intra-cluster dispersion of each initial cluster is the mean of the Euclidean distances from all historical sample data within the cluster to the cluster center. Arrange all initial clusters in ascending order of intra-cluster dispersion to obtain the intra-cluster dispersion distribution of all initial clusters. Determine the splitting threshold using the percentage method. Select the distance value corresponding to a high percentile (e.g., 95%) as the splitting threshold. That is, find a value among all intra-cluster dispersions such that 95% of the intra-cluster dispersions are less than or equal to this value. This value is the distance value corresponding to the 95th percentile. The intra-cluster dispersion of unstable clusters being greater than the splitting threshold is taken as the first condition.

[0065] The formula for calculating the intra-cluster dispersion of the initial cluster is as follows:

[0066]

[0067] In the formula, Indicates the first The intra-cluster dispersion of the initial cluster, i.e., the first... The mean Euclidean distance from all historical sample data within an initial cluster to the cluster center; Indicates the first The number of all historical samples in each initial cluster; Represents the initial cluster A historical sample of data within; Indicates the first The cluster centers of the initial clusters.

[0068] If the intra-cluster dispersion of an unstable cluster is greater than the splitting threshold, it indicates that the intra-cluster differences of the unstable cluster are large, and it needs to be split into smaller clusters. The unstable cluster is then split into two new clusters by performing a K-means clustering algorithm with a K value of 2.

[0069] S23. When an unstable cluster satisfies the second condition, the unstable cluster is merged with other clusters to obtain a new cluster.

[0070] If an unstable cluster does not meet the first condition, there is no need to split it. Then, determine whether the unstable cluster meets the second condition. The process for obtaining the second condition is as follows: Calculate the inter-cluster distance of all initial clusters pairwise, i.e., calculate the Euclidean distance between the centers of every two initial clusters. Arrange the initial clusters pairwise according to their inter-cluster distance from smallest to largest to obtain the distribution of inter-cluster distances for all initial clusters. Determine the merging threshold using the percentage method, selecting the distance value corresponding to a low percentile (e.g., 5%) as the merging threshold. That is, find a value among all inter-cluster distances such that 5% of the inter-cluster distances are less than or equal to this value; this value is the distance value corresponding to the 5th percentile. The condition that the inter-cluster distance between the unstable cluster and other clusters is less than or equal to the merging threshold is used as the second condition.

[0071] The formula for calculating the inter-cluster distance between two initial clusters is as follows:

[0072]

[0073] In the formula, Represents the initial cluster To the initial cluster The inter-cluster distance is the Euclidean distance between the centers of the two initial clusters. Represents the initial cluster The first cluster center sample 3D eigenvalues ​​( ); Represents the initial cluster The first cluster center sample Dimensional eigenvalues.

[0074] If the distance between an unstable cluster and other clusters is less than or equal to the merging threshold, it indicates that the two clusters are highly similar and can be merged to form a new cluster.

[0075] S24. When an unstable cluster does not meet the first condition or the second condition, determine whether the unstable cluster meets the third condition. If it does, merge the unstable cluster into the cluster with the smallest inter-cluster distance to obtain a new cluster. If it does not meet the third condition, it will be treated as a stable cluster and participate in subsequent steps.

[0076] When an unstable cluster does not meet either the first or second condition, it is determined whether the unstable cluster meets the third condition. The process for obtaining the third condition is as follows: Calculate the ratio of the number of historical sample data contained in the unstable cluster to the total number of historical sample data. If this ratio is less than a preset value, it is taken as the third condition. In this embodiment, the preset value is an empirical value of 3%. If the ratio is less than 3%, it indicates that the sample size of the unstable cluster is too small, and it needs to be merged into the cluster with the smallest inter-cluster distance to form a new cluster. If the ratio is greater than or equal to 3%, it is labeled as a stable cluster.

[0077] S25. Count the clusters generated after splitting or merging operations and the original stable clusters. Recalculate the stability score of these clusters until no more splitting or merging operations are performed on the clusters. The number of clusters at this point is the optimal K value.

[0078] In the above steps, unstable clusters are split or merged to form new clusters. The new clusters and all stable clusters are counted, and the stability scores of these clusters are recalculated in the same way as the initial clusters. If the stability score of all clusters is greater than or equal to 0.7, then no further splitting or merging is needed, and the number of clusters at this point is the optimal K value. If any cluster has a stability score less than 0.7, it is marked as an unstable cluster, and steps S22 to S25 are repeated. It can be understood that when repeating these steps, the initial clusters mentioned in the steps refer to the clusters generated after splitting or merging operations and all stable clusters counted after each iteration.

[0079] For ease of understanding, this embodiment provides a simplified description of the above iterative process: After forming 5 initial clusters and calculating the stability scores of the 5 initial clusters, it is assumed that there is 1 unstable cluster and 4 stable clusters.

[0080] First, determine whether the unstable cluster meets the first condition, that is, whether the intra-cluster dispersion of the unstable cluster is greater than the splitting threshold. If it is greater than the splitting threshold, split the unstable cluster into two new clusters. The two new clusters and the original four stable clusters form a new round of clusters. Count the number of clusters in this round and calculate the stability score of the clusters in this round. If the stability score of all clusters in this round is greater than or equal to 0.7, then take the number of all clusters in this round as the optimal K value, that is, K value is 6.

[0081] If an unstable cluster does not meet the first condition, then it is determined whether the unstable cluster meets the second condition, that is, whether the distance between the unstable cluster and other clusters is less than or equal to the merging threshold. If it is less than or equal to the merging threshold, then the unstable cluster is merged with other clusters to form a new cluster. This new cluster, together with the original three stable clusters (one of which was used to merge with an unstable cluster, so the merged cluster needs to be removed), forms a new round of clusters. The number of clusters in this round is counted and the stability score of the clusters in this round is calculated. If the stability score of all clusters in this round is greater than or equal to 0.7, then the number of all clusters in this round is taken as the optimal K value, that is, K value is 4.

[0082] If an unstable cluster satisfies neither the first nor the second condition, then it is determined whether the unstable cluster satisfies the third condition: whether the ratio of the number of historical samples within the unstable cluster to the total number of historical samples is less than 3%. If it is less than 3%, it indicates that the sample size is too small. This unstable cluster is then merged into the cluster with the closest inter-cluster distance, forming a new cluster. This new cluster, along with its original three stable clusters, forms a new round of clusters. The number of clusters in this round is counted, and the stability score of all clusters in this round is recalculated. If the stability score of all clusters in this round is greater than or equal to 0.7, then the number of clusters in this round is taken as the optimal K value, i.e., K = 4. If the ratio is greater than 3%, then the unstable cluster is treated as a stable cluster and marked. At this point, all five initial clusters are stable clusters, and the optimal K value is 5. It is understandable that since the above description is based on the assumption that there is one unstable cluster, the optimal K value is directly obtained after this unstable cluster is treated as a stable cluster. If there are two unstable clusters among the five initial clusters, after one of the unstable clusters is treated as a stable cluster because it does not meet the first and second conditions and its sample ratio is greater than 3%, this unstable cluster will participate in all calculations normally as a stable cluster.

[0083] If, during the process described above, after this unstable cluster splits or merges, there are still clusters with a stability score less than 0.7 in the new round of clusters formed with the original stable cluster, then the clusters with a stability score less than 0.7 are judged to be unstable clusters. The above process is repeated until all the new round of clusters are stable clusters. The number of all clusters in this round is the optimal K value.

[0084] K-means clustering is performed on the collected historical sample data based on the optimal K value: K different sample data are randomly selected from all historical sample data as initial cluster centers. The Euclidean distance from each historical sample data to the K initial cluster centers is calculated. Each historical sample data is assigned to the cluster corresponding to the initial cluster center with the closest Euclidean distance. The average value of all historical sample data in each cluster is recalculated and used as the new cluster center. The steps of assigning historical sample data and updating cluster centers are repeated until the cluster centers no longer change. The clustering is then completed, and K clusters are obtained.

[0085] After clustering, K clusters are obtained with pre-defined business rules. The validity of these clusters is then verified to match actual business models. Before setting the rules, the business scenarios and key objects involved in the remote meter reading data used by the concentrator are clearly defined, resulting in a business scenario list. Based on this list, industry standards, and equipment parameters, the normal range for each feature is determined, and business rules are formulated. For all clusters obtained after clustering using the optimal K-value, their compliance with the business rules is verified one by one. Based on the verification results, business labels are added to the clusters. Clusters that comply with the business rules are labeled as normal, while those that do not are labeled as abnormal. This process yields all clusters with known business labels.

[0086] S3. Obtain the preprocessed real-time sample data, calculate the Euclidean distance between the sample data and each cluster, obtain the label of the sample data based on the Euclidean distance, and determine that the collected sample data is incorrect if the label of the sample data is abnormal.

[0087] Real-time data undergoes preprocessing and feature alignment. The preprocessed samples are transformed into feature vectors with the same feature structure as historical sample data, forming real-time sample data. The Euclidean distance between the real-time sample data and the cluster center of each cluster is calculated. The cluster with the smallest Euclidean distance is selected as the cluster to which the sample data belongs. The business label of the belonging cluster is used to determine whether the real-time sample data is normal. If the business label of the belonging cluster is abnormal, the sample data is determined to be abnormal.

[0088] By detecting anomalies in remote meter reading data from concentrators in real time, the accuracy and reliability of meter reading data can be significantly improved, power equipment failures can be prevented and diagnosed, power resource allocation can be optimized, the safety and stability of the power system can be improved, the operation and management of smart grids can be supported, and the economic efficiency of the power system can be enhanced.

[0089] In each iteration, this invention splits or merges unstable clusters, dynamically adjusts the K value, and automatically finds the optimal number of clusters based on the inherent structure of the data. This ensures that the final clustering results have high stability and accuracy. This method avoids the deviation of clustering results caused by manually preset K values, making the clustering results more stable and reliable. Moreover, by dynamically adjusting the K value and evaluating the stability of clusters, the model can better adapt to new data. When the data distribution changes, the model can automatically adjust the K value to adapt to the new data structure.

[0090] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for remote meter reading for concentrators, characterized by, The method comprises the following steps: Collect historical sample data, each historical sample data comprising metering feature data, state feature data and communication feature data; According to the calculated optimal K value, use K-means clustering to obtain a plurality of clustering clusters, and assign a label to each clustering cluster, the label comprising normal or abnormal; Obtain real-time sample data after preprocessing, calculate the Euclidean distance between the sample data and each clustering cluster, and obtain the label of the sample data according to the Euclidean distance; if the label of the sample data is abnormal, it is determined that the collected sample data is incorrect; The calculation process of the optimal K value comprises: Based on the given initial K value, the historical sample data is clustered to obtain a plurality of initial clustering clusters; Calculate the stability score of the initial clustering cluster; when the stability score is greater than or equal to the preset threshold, mark the corresponding initial clustering cluster as a stable cluster; when the stability score is less than the preset threshold, mark the corresponding initial clustering cluster as an unstable cluster; when the unstable cluster meets the first condition, split the unstable cluster to obtain a new clustering cluster; when the unstable cluster meets the second condition, merge the unstable cluster with other clusters to obtain a new clustering cluster; Statistically analyze the clustering clusters generated after the splitting or merging operation and the original stable clusters, recalculate the stability score of the clustering clusters, and stop the splitting or merging operation until the clustering clusters no longer meet the conditions; at this time, the number of clustering clusters is the optimal K value; The first condition is obtained as follows: calculate the intra-cluster dispersion of all initial clustering clusters, determine the splitting threshold based on the percentage method, and take the intra-cluster dispersion of the unstable cluster greater than the splitting threshold as the first condition; The second condition is obtained as follows: calculate the inter-cluster distance of all initial clustering clusters, determine the merging threshold based on the percentage method, and take the inter-cluster distance between the unstable cluster and a certain clustering cluster less than or equal to the merging threshold as the second condition.

2. A method for remote meter reading for concentrator as claimed in claim 1, wherein, The calculation process of the optimal K value further comprises: When the unstable cluster does not meet the first condition and the second condition, determine whether the unstable cluster meets the third condition; if it meets the third condition, merge the unstable cluster into the clustering cluster with the smallest inter-cluster distance to obtain a new clustering cluster; if it does not meet the third condition, it is taken as a stable cluster to participate in the subsequent steps.

3. A method for remote meter reading for concentrator as claimed in claim 2 wherein, The third condition is obtained as follows: Calculate the ratio of the number of historical sample data contained in the unstable cluster to the total number of historical sample data, and take the ratio less than a preset value as the third condition.

4. The method of claim 1, wherein the concentrator is a power line concentrator. The calculation process of the stability score is as follows: For each initial clustering cluster, generate a plurality of perturbation clusters, calculate the Jaccard index value between the initial clustering cluster and each corresponding perturbation cluster, and take the average of all corresponding Jaccard index values as the stability score of the initial clustering cluster.

5. A method for remote meter reading for concentrator as claimed in claim 4 wherein, The process of generating a plurality of perturbation clusters is as follows: Each initial clustering cluster contains a plurality of historical sample data; for each initial clustering cluster, use the Bootstrap sampling method to extract historical sample data, repeat the process to obtain a plurality of perturbation subsets; for each perturbation subset, add a preset proportion of Gaussian noise and perform clustering to obtain a perturbation cluster.

6. A method for remote meter reading for concentrator as claimed in claim 5 wherein, The historical sample data comprises a plurality of feature dimensions, and the process of adding Gaussian noise is as follows: Adding noise to each feature dimension of each historical sample data in the disturbance subset, wherein the noise is 5% of the standard deviation of the corresponding feature dimension in the initial clustering cluster.

7. The method of claim 1, wherein the concentrator is a power line concentrator. The intra-cluster dispersion is the average Euclidean distance of all historical sample data in the initial clustering cluster to the cluster center. The intra-cluster dispersion is the average Euclidean distance of all historical sample data in the initial clustering cluster to the cluster center.

Images