07.02.2013 Views

Optimization and Computational Fluid Dynamics - Department of ...

Optimization and Computational Fluid Dynamics - Department of ...

Optimization and Computational Fluid Dynamics - Department of ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

3 Mathematical Aspects <strong>of</strong> CFD-based <strong>Optimization</strong> 75<br />

ing technique in general <strong>and</strong> especially in the context <strong>of</strong> Nonlinear Model<br />

Predictive Control (NMPC), the reader is referred to [13, 14].<br />

Within classical multiple shooting, these blockwise matrices, also called<br />

Wronskians, are built up in a sensitivity approach, which means that the<br />

Jacobian matrix <strong>of</strong> the constraints are explicitly constructed <strong>and</strong> completely<br />

stored. Matrix factorization techniques dovetailed to the multiple shooting<br />

QP are applied in order to solve the quadratic programs. Note, however, that<br />

these Jacobians need only be approximated in a coarse manner, when using<br />

formulation (3.16, 3.17) such as by secant update techniques. In case <strong>of</strong> CFD<br />

optimization problems they should only capture unstable <strong>and</strong> slowly decaying<br />

modes. In certain cases, the frequent generation <strong>of</strong> the constraint Jacobian<br />

can also be avoided, so that only the very first Jacobian is constructed <strong>and</strong><br />

factorized by the use <strong>of</strong> formulation (3.16, 3.17), where the matrix A denotes<br />

the (multiple shooting) Jacobian at the data <strong>of</strong> the first optimization<br />

iteration.<br />

3.3.2 Parallel Multiple Shooting<br />

Parallelization <strong>of</strong> multiple shooting is trivially based on the idea that the<br />

initial value solvers on each multiple shooting interval can work completely<br />

independently, provided the initial values si are available. Parallel realizations<br />

<strong>of</strong> multiple shooting have existed for a long time, cf. [15, 24, 36], <strong>and</strong> more<br />

recently as parareal in [27]. Ulbrich has described the use <strong>of</strong> this approach<br />

within an optimal control context [35] in a formulation similar to Eqs. (3.16)-<br />

(3.17). There the initial values are provided by a coarse discretization scheme<br />

(<strong>of</strong>ten implicit Euler) which operates only on the multiple shooting nodes.<br />

It is obvious that this approach works only well with CFD problems, which<br />

are dissipative enough so that even coarse time discretizations give some<br />

consistent information.<br />

3.3.3 Real-time <strong>Optimization</strong> <strong>and</strong> Nonlinear Model<br />

Predictive Control<br />

Real-time optimization is not just fast optimization, rather it means a shift<br />

in philosophy: in order to minimize the delay time such as for a control<br />

response to a perturbation <strong>of</strong> the process, one tries to compute as much<br />

information as possible before data about the values <strong>of</strong> states <strong>and</strong> parameters<br />

or <strong>of</strong> scenarios for the real process become available in real-time. In the CFD<br />

context this means in particular that matrix factorizations, which are <strong>of</strong>ten<br />

avoided for computational speed, or the pre-computation <strong>of</strong> sophisticated<br />

pre-conditioners, receive an increased attention, ins<strong>of</strong>ar as they can <strong>of</strong>ten be

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!