The streaming data scenario has brought about unique challenges with it, like outliers detection, large dimensionality and the issue of scalability being at primary focus. The temporal locality is a quite important while, processing evolving data stream (EDS). The inherit patterns present in the data evolves, and hence, the past clusters are no longer valid to the future and also the initial centroids. Thus, a method of centroid initialization has an important role to play in the clusters formation. The proposed work, analyzes the effect of different initialization methods is presented. The three methods are studied K-nearest neighborhood (K-NN), random and uniform. The results have been validated on the nine datasets. K-NN methods have shown promising results with the formation of compact clusters still being a bit time-consuming approach. While in the case of other methods the centroid remains concentrated. These methods highly depended on the shape of the clusters present and the distribution.