Jayson C. Jueco, Wynn Gaudette P. Navarro, Lester Mel B. Libre, Zyril Mark D. Undalok, Clint Kyle K. Dingcong, Ferdinand F. Batayola, Marvin A. Radaza, Wilen Melsedec O. Narvios, Xavier Julian P. Fernandez, Ike A. Baguio
Accurately clustering and labeling anomalous time series load data from large databases remains challenging due to close similarity and subtle deviations between each observation. This paper addressed the problem with a novel clustering model that cleaves into encoding the load data as images using a modified Gramian Angular Field (GAF) and clustering the anomalous images using K-means. Encoding load data as images enables visual interpretability, which model can distinctively cluster the load data from the temporal and visual features. To evaluate the model, we performed internal and external measures between different values of k to determine the appropriate number of clusters. The average silhouette score of the clustered encoded images is between 0.20 and 0.55. Also, the method consistently yields a minimum Davies-Bouldin Index (DBI) and maximum Calinski-Harabasz Index(CHI) across different datasets for $k=2$, a competitive value against the time series clusters. This suggests the method is suitable for accurately labeling anomalous data from the set. Experimental results are verified to match ground truth historical events correctly. © 2023 IEEE.
Cebu Technological University, Department of Electrical Engineering, Cebu, Philippines; School of Engineering, University of San Carlos, Cebu, Philippines; Cebu Technological University, Department of Industrial Engineering, Cebu, Philippines
