基于HDBSCAN动态跟踪客户用电行为模式

Dynamic Tracking of Customer's Load Behavior Pattern by HDBSCAN

为了自动辨识出优质电力大客户并快速感知其负荷行为变化模式,文章通过HDBSCAN算法(hierarchical density-based spatial clustering of applications with noise),对大工业客户1个月内分钟级的负荷行为数据进行自动分类。依据聚类结果筛选出潜在优质的用电客户,对其负荷行为模式进行动态跟踪分析(这里所说的"动态"是指相邻时间段内不同负荷状态的转换模式,综合考虑状态特征和时域特征的变化),以找出用电行为异常、或存在负荷结构变化的客户,增强对电网系统的动态感知能力,降低潜在风险。该算法最大程度地避免了人为主观性经验的参与调整参数,采用这种无监督机器学习技术能极大程度地提高整体分析效率;属于自下而上的数据驱动感知用户侧精细行为模式,将能大面积快速感知到诸多潜在风险模式和异常行为模式。

To automatically identify high quality electricity customers whose changed(or abnormal)load behavior pattern can be quickly perceived,this article uses the HDBSCAN algorithm(Hierarchical Density-Based Spatial Clustering of Applications with Noise)to automatically cluster the customer’s load behavior data in minute level within a month.According to the clustering results,potential customers with high quality are filtered through dynamic tracking and analysis of their load behaviors(here,the“dynamic“refers to transitions between different load levels in adjacent periods,and the change of state feature and time domain feature is considered comprehensively).Detecting customer’s abnormal load behavior or changes in load structure,can enhance the dynamic perception of power grid system and reduce the potential risks.This algorithm which is a bottom-up data-driven method in quickly perceiving customer’s granularly behaviors where maybe exist potential risk patterns or abnormal patterns in a large area,avoids the subjective participation in adjusting parameters to the greatest extent.