Unsupervised learning method for clustering dynamic behavior in the context of power systems. Issue 2 (2020)

• Record Type: Journal Article
• Authors: Mitrentsis, Georgios; Lens, Hendrik
• Is Part Of: IFAC-PapersOnLine. Volume 53:Issue 2(2020)
• Page Start: 13024
• Page End: 13029
• Publication Date: 2020
• Subjects: active distribution networks -- unsupervised learning -- clustering -- dynamic models -- parameter identification -- measurement-based approach -- multisignal analysis; Automatic control -- Periodicals; 629.805
• DOI: 10.1016/j.ifacol.2020.12.2170
• Languages: English
• ISSNs: 2405-8963

Abstract

Abstract: Aggregated dynamic equivalent models of active distribution networks (ADNs) are commonly derived using the measurement-based approach. This method exploits acquired data in order to estimate the model parameters using system identification techniques. However, most of the approaches assume that the system maintains the same dynamics for different operating conditions, even though the load mix and the distributed generation (DG) composition are constantly changing. To this end, this paper presents a novel method, which can be used as the first step of the system identification procedure, in order to take into account different system dynamics in ADN modeling. To do so, three unsupervised learning methods for clustering the various dynamic behaviors are introduced, yielding groups of measurements that represent different dynamics. In this context, the proposed methods leverage four clustering algorithms of different notion and complexity, namely k -means++, k -medoids, fuzzy c -means (FCM) and hierarchical clustering. To assess the validity of the proposed approach, real measurements acquired within a year in six real substations in Southern Germany are processed. The results highlight the remarkable difference in system dynamics justifying the necessity of an initial cluster analysis. Finally, the ratio of "Within Cluster sum of squares" to "Between Cluster Variation" (WCBCR) is deployed to compare the effectiveness of the clustering algorithms.