Shapelet based classification of customer consumption patterns

Abstract

Energy consumption time series show similarities among clients which can be further used to classify consumption based on common patterns. The recently introduced shapelet algorithm enables us to perform such classification out of reach for clustering algorithms (e.g., k-means) usually used in smart grid analysis. In this paper we analyze the efficiency of the shapelet algorithm in classifying various weekend consumption patterns extracted from real-life data. We test both the accuracy of the classification and the scalability in comparison with other algorithms such as KNN, DBSCAN, and OPTICS. Results show that while the shapelet algorithm achieves the highest accuracy (« 89%) training time takes about 5.8 hours. However, once the algorithm is trained classification is as fast as KNN. KNN is second best with « 82% accuracy but results depend highly on the value for k and the distance metric. Finally, DBSCAN and OPTICS show poor classification success mainly due to their unsupervised approach. Overall, this analysis demonstrates the efficiency of supervised clustering based on shapelets to identify meaningful information in consumption time series, representing a further step in the area of smart grid data analysis which is currently dominated by unsupervised clustering.