| dc.description.abstract | In this dissertation we develop multiple algorithms for efficient parallel solution of structured nonlinear programming problems by decomposition of the linear augmented system solved at each iteration of a nonlinear interior-point approach. In particular, we address large-scale, block-structured problems with a significant number of complicating, or coupling variables. This structure arises in many important problem classes including multi-scenario optimization, parameter estimation, two-stage stochastic programming, optimal control and power network problems. The structure of these problems induces a block-angular structure in the augmented system, and parallel solution is possible using a Schur-complement decomposition. Three major variants are implemented: a serial, full-space interior-point method, serial and parallel versions of an explicit Schur-complement decomposition, and serial and parallel versions of an implicit PCG-based Schur-complement decomposition. All of these algorithms have been implemented in C++ in an extensible software framework for nonlinear optimization.
The explicit Schur-complement decomposition is typically effective for problems with a few hundred coupling variables. We demonstrate the performance of our implementation on an important problem in optimal power grid operation, the contingency-constrained AC optimal power ow problem. In this dissertation, we present a rectangular IV formulation for the contingency-constrained ACOPF problem and demonstrate that the explicit Schur-complement decomposition can dramatically reduce solution times for a problem with a large number of contingency scenarios. Moreover, a comparison of the explicit Schur-complement decomposition implementation and the Progressive Hedging approach provided by Pyomo is provided, showing that the internal decomposition approach is computationally favorable to the external approach. However, the explicit Schur-complement decomposition approach is not appropriate for problems with a large number of coupling variables because of the high computational cost associated with forming and solving the dense Schur-complement.
We show that this bottleneck can be overcome by solving the Schur-complement equations implicitly using a quasi-Newton preconditioned conjugate gradient method.
This new algorithm avoids explicit formation and factorization of the Schur-complement.
The computational efficiency of the serial and parallel versions of this algorithm are compared with the serial full-space approach, and the serial and parallel explicit
Schur-complement approach on a set of quadratic parameter estimation problems and nonlinear optimization problems. These results show that the PCG implicit Schur-complement approach dramatically reduces the computational expense for problems with many coupling variables. | en |
