A Data-Driven Customer Segmentation Strategy Based on Contribution to System Peak Demand
A Data-Driven Customer Segmentation Strategy Based on Contribution to System Peak Demand
复制标题
DOI：
10.1109/tpwrs.2020.2979943
复制
发表时间：
2018-10
影响因子：
6.6
通讯作者：
Yuxuan Yuan;K. Dehghanpour;Fankun Bu;Zhaoyu Wang
中科院分区：
文献类型：
--
作者：
Yuxuan Yuan;K. Dehghanpour;Fankun Bu;Zhaoyu Wang
高级计量基础设施（AMI）使公用事业公司能够获得精细的能耗数据，这为根据客户对配电网中各种运营指标的影响设计客户细分策略提供了独特的机会。然而，执行公用事业规模的细分为不可观察的客户，只有每月的计费信息，仍然是一个具有挑战性的问题。为了应对这一挑战，我们提出了一个新的指标，一致的月度峰值贡献（CMPC），量化的贡献，个别客户的系统峰值需求。此外，一种新的多状态机器学习为基础的分割方法，估计CMPC的客户没有智能电表（SM）：首先，聚类技术被用来建立一个数据库，包含典型的日常负荷模式在不同的季节使用SM数据的可观察的客户。接下来，为了将不可观察的客户与所发现的典型负荷分布相关联，利用分类方法来计算不同不可观察的家庭的日常消费模式的可能性。在第三阶段中，加权聚类回归（WCR）模型被用来估计CMPC的不可观察的客户使用他们的每月账单数据和分类模块的结果。所提出的分割方法已被测试和验证使用真实的公用事业数据。
Advanced metering infrastructure (AMI) enables utilities to obtain granular energy consumption data, which offers a unique opportunity to design customer segmentation strategies based on their impact on various operational metrics in distribution grids. However, performing utility-scale segmentation for unobservable customers with only monthly billing information, remains a challenging problem. To address this challenge, we propose a new metric, the coincident monthly peak contribution (CMPC), that quantifies the contribution of individual customers to system peak demand. Furthermore, a novel multi-state machine learning-based segmentation method is developed that estimates CMPC for customers without smart meters (SMs): first, a clustering technique is used to build a databank containing typical daily load patterns in different seasons using the SM data of observable customers. Next, to associate unobservable customers with the discovered typical load profiles, a classification approach is leveraged to compute the likelihood of daily consumption patterns for different unobservable households. In the third stage, a weighted clusterwise regression (WCR) model is utilized to estimate the CMPC of unobservable customers using their monthly billing data and the outcomes of the classification module. The proposed segmentation methodology has been tested and verified using real utility data.