ISBN 9783843957663

978-3-8439-5766-3, Reihe Regelungstechnik

Gösta Stomberg
Real-Time Optimization for Distributed Nonlinear Model Predictive Control

184 Seiten, Dissertation Technische Universität Hamburg (2026), Softcover, A5

Zusammenfassung / Abstract

The widespread availability of computing and communication technology promotes the development of Cyber Physical Systems of Systems (CPSoS) which couple subsystems in the physical and cyber domains. CPSoS enable cooperation between subsystems and facilitate novel applications in diverse domains such as energy systems, robotic teams, and beyond. Distributed Model Predictive Control (DMPC) is a framework for control of CPSoS as it builds upon centralized Model Predictive Control (MPC) while also addressing the needs of CPSoS. One promising avenue relies on the use of decentralized optimization to solve a cooperative Optimal Control Problem (OCP) online without a coordinator. However, the iterative nature of decentralized optimization induces the need for repeated communication between coupled subsystems, which may jeopardize real-time feasibility. For centralized nonlinear MPC, real-time feasibility is achieved through Real-Time Iterations (RTIs) that enable fast control sampling. This thesis presents a novel decentralized RTI (dRTI) scheme for DMPC. The framework builds upon a novel decentralized Sequential Quadratic Programming (dSQP) method which parallelizes expensive computations among subsystems via the alternating direction method of multipliers. We first prove the local convergence of dSQP to regular OCP solutions in the presence of non-convex constraints. Subsequently, we show the local exponential stability of the closed-loop system-optimizer dynamics and we provide quantifiable bounds on the number of optimizer iterations that guarantee stability. We further demonstrate the efficacy of dRTI in hardware experiments for the formation control of mobile robots and holonomic hovercraft, studying scenarios of increasing difficulty ranging up to collision avoidance in dynamic environments. In addition to the above theoretical and practical contributions, we showcase the computational efficacy of dRTI through simulation case studies.