Decomposition Strategies for Nonconvex Problems, a Parametric Approach

This thesis deals with the development of numerical methods for solving nonconvex optimisation problems by means of decomposition and continuation techniques. We first introduce a novel decomposition algorithm based on alternating gradient projections and augmented Lagrangian relaxations. A proof of local convergence is given under standard assumptions. The effect of different stopping criteria on the convergence of the augmented Lagrangian loop is investigated. As a second step, a trust region algorithm for distributed nonlinear programs, named TRAP, is introduced. Its salient ingredient is an alternating gradient projection for computing a set of active constraints in a distributed manner, which is a novelty for trust region techniques. Global convergence as well as local almost superlinear convergence are proven. The numerical performance of the algorithm is assessed on nonconvex optimal power flow problems. The core of this thesis is the development and analysis of an augmented Lagrangian algorithm for tracking parameter-dependent optima. Despite their interesting features for large-scale and distributed optimisation, augmented Lagrangian methods have not been designed and fully analysed in a parametric setting. Therefore, we propose a novel optimality-tracking scheme that consists of fixed number of descent steps computed on an augmented Lagrangian subproblem and one dual update per parameter change. It is shown that the descent steps can be performed by means of first-order as well as trust region methods. Using the Kurdyka-Lojasiewicz property, an analysis of the local convergence rate of a class of trust region Newton methods is provided without relying on the finite detection of an optimal active set. This allows us to establish a contraction inequality for the parametric augmented Lagrangian algorithm. Hence, stability of the continuation scheme can be proven under mild assumptions. The effect of the number of primal iterations and the penalty is analysed by means of numerical examples. Finally, the efficacy of the augmented Lagrangian continuation scheme is successfully demonstrated on three examples in the field of optimal control. The first two examples consists of a real-time NMPC algorithm based on a multiple-shooting discretisation. In particular, it is shown that our C++ software package is competitive with the state-of-the-art codes on NMPC problems with long prediction horizons, and can address a more general class of real-time NMPC problems. The third case study is the distributed computation of solutions to multi-stage nonconvex optimal power flow problems in a real-time setting.

Splitting Methods for Distributed Optimization and Control

Ye Pu

This thesis contributes towards the design and analysis of fast and distributed optimization algorithms based on splitting techniques, such as proximal gradient methods or alternation minimization algorithms, with the application of solving model predictive control (MPC) problems. The first part of the thesis focuses on developing an efficient algorithm based on the fast alternating minimization algorithm to solve MPC problems with polytopic and second-order cone constraints. Due to the requirement of bounding the online computation time in the context of real-time MPC, complexity bounds on the number of iterations to achieve a certain accuracy are derived. In addition, a discussion of the computation of the complexity bounds is provided. To further improve the convergence speed of the proposed algorithm, an o-line pre-conditioning method is presented for MPC problems with polyhedral and ellipsoidal constraints. The inexact alternating minimization algorithm, as well as its accelerated variant, is proposed in the second part of the thesis. Different from standard algorithms, inexact methods allow for errors in the update at each iteration. Complexity upper-bounds on the number of iterations in the presence of errors are derived. By employing the complexity bounds, sufficient conditions on the errors, which guarantee the convergence of the algorithms, are presented. The proposed algorithms are applied for solving distributed optimization problems in the presence of local computation and communication errors, with an emphasis on distributed MPC applications. The convergence properties of the algorithms for this special case are analysed. Motivated by the complexity upper-bounds of the inexact proximal gradient method, two distributed optimization algorithms with an iteratively refining quantization design are proposed for solving distributed optimization problems with a limited communication data-rate. We show that if the parameters of the quantizers satisfy certain conditions, then the quantization error decreases linearly, while at each iteration only a fixed number of bits is transmitted, and the convergence of the distributed algorithms is guaranteed. The proposed methods are further extended to distributed optimization problems with time-varying parameters.

EPFL2016

Two-Layered Real-Time Optimization of a Solid Oxide Fuel Cell Stack

Dominique Bonvin, Gene Bunin, Grégory François

The optimal operation of a solid oxide fuel cell stack is addressed in this paper. Real-time optimization, performed at a slow time scale via constraint adaptation, is used to account for uncertainty and degradation eﬀects, while model-predictive control is performed at a faster time scale to reject process disturbances and to safely adapt the system to the specified output constraints following changes in cell power demand. To ensure that these constraints are strictly honored, a novel adaptation algorithm that uses the built-in constraint handling of quadratic programming is implemented within the model-predictive controller. An additional feature of this algorithm - its ability to adapt the feasibility region in view of uncertainty - is shown as well. Simulation results illustrate the efficacy of this approach in the solid oxide fuel cell system.

2010

Low-Complexity Optimization-Based Control: Design, Methods and Applications

Ivan Pejcic

Optimization-based controllers are advanced control systems whose mechanism of determining control inputs requires the solution of a mathematical optimization problem. In this thesis, several contributions related to the computational effort required for optimization-based controller execution are provided. The content of the thesis is divided into three parts: The first part provides methods capable of performing automatic controller tuning for constrained control of nonlinear systems. Given a specified controller structure, the presented methods are able to perform an offline tuning of the controller parameters such that some user-specified performance metric is optimized while imposing stability guarantees on the obtained closed-loop system. The methods are characterized by a broad flexibility that allows their application to many control schemes that are widely popular in practice, but also to novel user-specified control schemes that are convenient from a computational or some other point of view. The controller tuning is formulated as an optimization problem that can be tackled by black-box optimization techniques such as Bayesian optimization. The methods are demonstrated by application examples involving speed control of a permanent magnet synchronous machine and position control of a mechanical gyroscopic system. The second part provides an accelerated version of the alternating direction method of multipliers (ADMM) optimization algorithm derived by using a recently proposed accelerated Douglas-Rachford (DR) splitting. The obtained method is an accelerated ADMM version that replaces the internal proximal point convergence mechanism of the classical ADMM by the accelerated gradient method applied on a specially constructed scaled DR envelope function. The form of the accelerated ADMM is derived and conditions are provided under which the underlying accelerated DR splitting is validly addressing the Fenchel dual problem. The third part describes a model predictive control scheme for power electronics control which involves a combination of the integral of squared predicted tracking error as the controller's cost function together with offline computed optimal steady-state voltage signals. These offline computed optimal steady-state signals are in the power electronics community referred to as Optimized Pulse Patterns (OPPs). The method is presented by considering an industrial case study involving a grid-tied converter with LC filter. After introducing an optimal control problem based on OPPs, low computational complexity approximate versions are provided. The resulting approximate controller versions are addressed by using memory storage of the dynamic behavior of the system, leading to controller forms whose execution can be performed on embedded hardware.