An Optimal Wavelet Detailed-Coefficient Determination Using Time-Series Clustering

Abstract

Time-series clustering is used to find out similar patterns which occur in a time-series data. Time-series sequences are characterized by many features. The trend and seasonality, for instance, are the two important features which reveal the behavior of a time-series sequence and can be used to group them. One of the common dimensionality reduction method for time-series data is the usage of wavelets. In this paper, we analyze the importance of wavelet-based time-series clustering which helps to reduce the length of time-series dimensions, in-turn reducing the time needed for the clustering task. Determining an optimal set of coefficients for wavelets is always challenging. While it plays an important role in reducing the time taken for clustering, to test/verify our claims, we analyze a renewable energy dataset which contains the periodical energy load of various generators located in Europe. The clustering of this time-series data helps to identify similar electrical nodes. We compared eight different clustering algorithms with and without using wavelet decomposition on the renewable energy data. Minibatch k-means and Brich clustering show better performance on using wavelet-based decomposition.