Distributed K-means Power User Classification via Adjacency Matrix
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The lack of shared power consumption data among load aggregators leads to inconsistent user categorization, making it difficult for grid operators and multiple load aggregators to participate in the power market and formulate demand response mechanisms effectively.
Innovation Solution
A method of power user classification based on distributed K-means, where load aggregators normalize time series load data, form an adjacency matrix to represent information sharing relationships, and perform K-means clustering with iterative calculations to achieve globally consistent centroids without sharing actual power consumption data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If load aggregators do not share power consumption data to protect user privacy and maintain competition, then user privacy and market competition are protected, but user category classification becomes inconsistent across different aggregators
Solution Approach 1:
The patent segments the user classification task by allowing each load aggregator to perform local K-means clustering independently on their own data, while sharing only the clustering results (centroids and cluster assignments) rather than raw data. This segmentation enables privacy preservation while achieving consistent classification through distributed computation.
Solution Approach 2:
The patent introduces an intermediary mechanism where load aggregators share clustering results through an adjacency matrix that represents their information sharing relationships. This intermediary allows the system to propagate classification information across aggregators without direct data sharing, resolving the contradiction between privacy and consistency.
2Device complexity
If load aggregators perform independent K-means clustering without data sharing, then computation complexity is reduced and privacy is protected, but the resulting user categories differ across aggregators
Solution Approach 1:
The patent merges the independent clustering results from multiple load aggregators by sharing centroids and cluster assignments through an adjacency matrix. This combining mechanism allows each aggregator to maintain simple local computation while achieving global consistency through the aggregation of results.
Solution Approach 2:
The patent implements feedback mechanisms where clustering results from one iteration are used as initial values for the next iteration. This feedback loop allows the system to converge toward consistent user categories across all aggregators while maintaining relatively simple local computations at each step.
3Adaptability or versatility
If multiple load aggregators participate in demand response mechanisms with different user categories, then service customization is enabled, but coordination and grid operator management becomes inconvenient
Solution Approach 1:
The patent changes the parameters of user classification by using normalized load data and standardized K-means clustering parameters across all aggregators. This parameter standardization enables service customization through different cluster results while maintaining consistent classification frameworks that ease grid operator coordination.
Data Source
AI summary
A method of a power user classification based on distributed K-means, a storage medium and a classification device are provided. The method includes: obtaining, by N load aggregators, power consumption data of power users managed by respective load aggregators; performing, by each load aggregator, a normalization operation on time series load data of the power users managed by the load aggregator; forming a N×N dimensional adjacency matrix A; performing K-means clustering on normalized time series load data, to obtain the respective centroids and user groups characterized by the respective centroids; sharing, by the respective load aggregators, the centroids and the number of users under the respective centroids based on the adjacency matrix A, and obtaining consistent centroids by multiple load aggregators; after an overall iteration ends, obtaining, by the respective load aggregators, the consistent centroids consistent with the K-means centroid based on global data, to realize user classification.