Data-driven Influence Based Clustering of Dynamical Systems

TL;DR

This paper introduces a novel data-driven clustering method for dynamical systems that accounts for their evolution over time by using a influence-based similarity measure derived from the Koopman operator framework, applicable to various complex systems.

Contribution

It proposes a new influence-based similarity measure for clustering dynamical systems directly from time-series data, leveraging Koopman operator theory to handle nonlinearities.

Findings

Successfully clustered linear, nonlinear, and atmospheric systems.

Demonstrated effectiveness on power grid and environmental data.

Framework applicable to diverse dynamical systems.

Abstract

Community detection is a challenging and relevant problem in various disciplines of science and engineering like power systems, gene-regulatory networks, social networks, financial networks, astronomy etc. Furthermore, in many of these applications the underlying system is dynamical in nature and because of the complexity of the systems involved, deriving a mathematical model which can be used for clustering and community detection, is often impossible. Moreover, while clustering dynamical systems, it is imperative that the dynamical nature of the underlying system is taken into account. In this paper, we propose a novel approach for clustering dynamical systems purely from time-series data which inherently takes into account the dynamical evolution of the underlying system. In particular, we define a \emph{distance/similarity} measure between the states of the system which is a function of the influence that the states have on each other, and use the proposed measure for clustering of the dynamical system. For data-driven computation we leverage the Koopman operator framework which takes into account the nonlinearities (if present) of the underlying system, thus making the proposed framework applicable to a wide range of application areas. We illustrate the efficacy of the proposed approach by clustering three different dynamical systems, namely, a linear system, which acts like a proof of concept, the highly non-linear IEEE 39 bus transmission network and dynamic variables obtained from atmospheric data over the Amazon rain forest.