Design and Voltage Supply of High-Speed Induction - Aaltodoc
Design and Voltage Supply of High-Speed Induction - Aaltodoc
Design and Voltage Supply of High-Speed Induction - Aaltodoc
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
whole solution space determined by the limits for the variables <strong>and</strong> other constraints. Thus, it is<br />
more likely that the global optimum is found.<br />
• When designing new kinds <strong>of</strong> motors like high-speed motors, the optimal designs may be quite<br />
different from the conventional designs. The GA is not sensitive to the initial point <strong>of</strong> the<br />
design, which is <strong>of</strong>ten the case with the deterministic algorithms. The GA is more like an<br />
intelligent stochastic algorithm, combining the ‘freedom <strong>of</strong> prejudice’ <strong>of</strong> a stochastic algorithm<br />
with the ability <strong>of</strong> an deterministic algorithm to ‘concentrate on the relevant designs’.<br />
The downside <strong>of</strong> using a stochastic algorithm is that it takes a large number <strong>of</strong> function evaluations<br />
to get the result. Using a heavy model like FEM means that each function evaluation takes several<br />
or several tens <strong>of</strong> minutes to calculate. In order to decrease the optimization times <strong>and</strong> make more<br />
use <strong>of</strong> the time spent, a practical improvement not found in the literature review is suggested. This<br />
improvement is applicable to any algorithm but it is most efficient when used with a stochastic<br />
optimization algorithm running a heavy model.<br />
9.1.1 Combination <strong>of</strong> discrete search space <strong>and</strong> solution history<br />
Solving heavy models like FEM can take most <strong>of</strong> the time used in optimization. In this work, the<br />
FEM model discussed in previous sections <strong>and</strong> Appendix A uses roughly 95 % <strong>of</strong> the computation<br />
time. Hence, the lower number <strong>of</strong> designs calculated the better. Unfortunately, this is contradictory<br />
to the nature <strong>of</strong> the stochastic optimization algorithms.<br />
The idea <strong>of</strong> the discretization <strong>of</strong> the search space is to avoid evaluation <strong>of</strong> designs that are<br />
practically the same. For example, an induction motor with an air gap <strong>of</strong> 2.0 mm is no different<br />
from a motor with an air gap <strong>of</strong> 2.001 mm. The discretization should be selected so that the smallest<br />
difference between two designs is noticeable <strong>and</strong> realizable. In the above example, the machine<br />
produced would probably have the air gap somewhere near 2.0 ± 0.05 mm. <strong>High</strong>er tolerances would<br />
be possible but not necessarily cost-effective.<br />
For optimization algorithms using heavy models, the discretization is not a problem since the<br />
gradient <strong>and</strong> Hessian information <strong>of</strong> the objective function is unavailable anyway. Finite difference<br />
approximations could still be used. A proper discretization could actually improve the use <strong>of</strong> finite<br />
differences because <strong>of</strong> noise inherent in the models. However, the relation between discretization or<br />
relative perturbations <strong>and</strong> the accuracy <strong>of</strong> finite differences is a complicated issue (Palko 1996) <strong>and</strong><br />
avoided by using stochastic algorithms.<br />
107