Cluster Survival Model of Concept Drift in Load Profile Data

Abstract

An accurate scenario of customer’s power consumption patterns is a worthwhile asset for electricity provider. This paper proposes a cluster survival model of concept drift in load profile data. The cluster survival model of concept drift retrieves the dynamic behaviors of the clusters over time. We formulate a new data stream clustering algorithm, I-niceStream, which identifies the number of clusters and initial cluster centers automatically for producing the clustering results. We derive a modified Kullback–Leibler divergence for computing the concept drift scores from the clustering results. The concept drift scores are used to estimate the related clusters and the clustering patterns. The survival model categorizes the clustering patterns into sustaining, fading, and emerging types. Experiments were conducted on both synthetic datasets and real-world load profile dataset collected from different factories at Guangdong province in China. Experimental results show that the cluster survival model is able to identify the clustering patterns effectively from load profile data stream. The I-niceStream algorithm significantly outperformed three state-of-the-art algorithms in clustering accuracy on synthetic stream datasets.