Smart Meter Data Analytics Based on Modified Streaming k-Means
Wendong Zhu, Weiqing Yu, Bowen Kan, Guangyi Liu · 2017
k-Means clustering is one of the central problems in data analysis and remains as one of the most popular data processing algorithms. There are streaming variant of k-Means clustering algorithms to address the continuous and evolving data. Stream Computing allows users to capture and analyze all events and data, all the time and just in time. Streaming models have advantages over batch models due to the real time, dynamic and incremental features. However, streaming k-Means is sensitive to initial assignment of cluster centers, and the results are not accurate in the case that cluster centers drift significantly. This paper presents a Modified Streaming k-Means algorithm which is dynamic and incremental, as well as being able to take into consideration of long term patterns of data. Compared with streaming k-means clustering, the Modified Streaming k-Means clustering has better convergence ability and more stable results. We explore the application of our Modified Streaming k-Means algorithm in consumer segmentation problem, and demonstrate the effectiveness of our method using real-world smart meter data.