Time Scale Separation and Hierarchical Control with the Koopman Operator

TL;DR

This paper develops Koopman Operator-based methods to handle hierarchical control and time scale separation in nonlinear systems, enabling better stability analysis and control policy computation across different time scales.

Contribution

It introduces domain-aware Koopman formulations tailored for systems with hierarchical control and time scale separation, advancing their analysis and control capabilities.

Findings

Koopman formulations quantify cross-scale interaction impacts on stability.

They enable optimal control policy calculation at multiple time scales.

The approach integrates domain knowledge into Koopman-based modeling.

Abstract

The Koopman Operator (KO) is a mathematical construct that maps nonlinear (state space) dynamics to corresponding linear dynamics in an infinite-dimensional functional space. For practical applications, finite-dimensional approximations can be constructed with machine learning. The linearity of the KO facilitates the use of linear tools and theories on nonlinear dynamical systems without relying on local approximations. Additionally, the KO has the ability to incorporate forms of domain knowledge such as stability and known variable interactions. It therefore constitutes part of the broader interest in domain-aware or physics-informed machine learning. Hierarchical control and time scale separation are two common properties in engineered systems -- often found together -- that can pose both computational and practical challenges. These properties provide opportunities for further KO-based domain knowledge incorporation, but this has seldom been taken advantage of in the KO literature. This paper focuses on developing and using domain-aware Koopman formulations for systems with hierarchical control, systems with time scale separation, and systems with both. We show how those formulations can be leveraged to quantify the impact of cross-scale interactions on overall stability and to calculate optimal control policies at both fast and slow time scales.