Branimir Novoselnik, Yutao Chen, Yuning Jiang, Tommaso Sartor
University of Freiburg
Tuesday, October 17, 2017, 14:00 - 16:00
Room 02-012, Georges-Köhler Allee 102, Freiburg 79110, Germany
Efficient Algorithm for Model Predictive Control of Systems of Systems via Parametric Optimization
A large-scale complex system comprising many, often spatially distributed, dynamical subsystems with partial autonomy and complex interactions is called system of systems. In this talk we describe an efficient algorithm for model predictive control of a class of system of systems for which the overall objective function is the sum of convex quadratic cost functions of (locally) constrained discrete-time linear subsystems that are coupled through a set of (global) linear constraints on the subsystems coordination parameters.
The proposed control algorithm is based on parametrization and splitting of the underlying optimization problem into one global coordination problem and a set of local optimization problems pertaining to individual subsystems. The local optimization problems are solved off-line, via multi-parametric optimization, while the coordination problem is solved on-line, i.e., at every discrete time step. The properties of the local parametric solutions are utilized to formulate specific structure of the coordination problem that can be solved very efficiently. In particular, it is shown that, for a fixed number of coupling constraints, the coordination problem can be solved with a linear-time algorithm if all subsystems have one-dimensional coordination parameters. The effectiveness of the developed control algorithm is demonstrated on a microgrid case study, where coordinated optimal control is achieved for a cluster of distributed energy sources, storages, and loads.
Partial Updating Schemes for fast NMPC
A bridge between linear and nonlinear MPC is built by using partial sensitivity update for multiple-shooting based SQP methods. A Curvature-like measure of nonlinearity (CMoN) for dynamic systems has been introduced so that only sensitivities that are sufficiently nonlinear are updated. Different updating logic have been developed resulting in different numerical and control performance. Partial condensing and matrix factorization algorithms are proposed to exploit partial sensitivity update. The proposed algorithms are used for a nine degree of freedom dynamic driving simulator and for multi-sensory motion cueing with an active seat. Simulation results have shown the feasibility of such schemes.
Distributed Optimization and Control with ALADIN
Structured nonlinear optimization problems arise in a variety of control applications ranging from nonlinear model predictive control via robust control for uncertain processes to distributed nonlinear control of hybrid systems. The Augmented Lagrangian based Alternating Direction Inexact Newton (ALADIN) method has been proposed to solve non-convex distributed optimization problems to local optimality. My current research focuses on algorithms development based on ALADIN and its applications. In the first part, we have already developed a real time variant for distributed (explicit) model predictive control and a derivative free variant for convex optimization. In the second part, ALADIN or its variants has already been applied for coordinating autonomous vehicles at traffic intersections, AC optimal power flow problem and stochastic robust control in chemical process.
Optimal control for non-prehensile manipulation
Nonprehensile manipulation is the process of manipulating a part without a form- or force-closure grasp. Open loop trajectory been tested in simulation for a 2D non-prehensile system composed by a capsule and a plane. For a simpler case, 1D ball bouncing system, an experimental apparatus have been build and a MPC controller have been tested in real world.