A Power Load Clustering Method Based on Limited DTW Algorithm
Mingyuan Gao, Taorong Gong, Rongheng Lin, Hua Zou · 2019 IEEE 3rd Information Technology, Networking, Electronic and Automation Control Conference (ITNEC) · 2019
With the rapid development of smart grids, power load analysis has become the main method to divide user groups and understand user consumption behaviors. In this paper, a power load clustering method based on limited dynamic time warping algorithm (LimitDTW) is proposed. On account of the disadvantages of original DTW algorithm, including high time complexity and malignant matching, the LimitDTW algorithm limits the search range to areas near the diagonal of distance matrix. While limiting unreasonable curve scaling, the time complexity is greatly reduced from O(N2) to average O(N) level. Furthermore, the method introduces finite curve translation to optimize time alignment within the search range, especially for users with similar patterns but slight time offsets. And finally, the K-Medoids algorithm is applied for clustering based on user similarity. An experiment is conducted on actual load data of users in the United States during one year. The clustering results divide all 800 users into 9 clusters with different consumption behaviors. Referring to 16 standard types of commercial building defined by the U.S. Department of Energy, the accuracy of clustering reaches 74.125% and performs 9 percent higher than original DTW algorithm. At the same time, the runtime cost of LimitDTW algorithm reaches linear complexity which is far superior to the original one, with both Silhouette Coefficient and Calinski-Harabaz Score better than original DTW algorithm, widely-used fastDTW and LCSS algorithms. At last, we also analyze the load consumption behaviors of each cluster respectively so as to extract typical load patterns and look for the reasons of abnormal fluctuations.