Time Series Clustering of Energy Meter Data

Time-dependent electrical measurement data are of utmost benefit to network operators for a multitude of purposes. As a result of network decoupling and decentralization, the data obtained is often incomplete and get hybridized due to the presence of numerous, sparsely placed distributed generation sources and measurement points in the network. This makes the data inexpedient for utilization in smart grids, forecasting, price predictions etc. This work presents two computational techniques for handling such data with the help of unsupervised learning-based clustering algorithms to determine groups of similar time series from large scale data sets of meter data. The approach presented is applied to real electric meter data and then compared based on performance and quality of output. Furthermore, a mechanism to determine features that dominate each cluster and the complete data set is provided that could help provide more meaningful insight and cluster labelling. The proposed technique was able to form clusters from non-linear, noisy and limited data and a deeper insight into influencing features for the clustering was obtained.