Tools & Applications 1. Introduction 2. Design of Matrix Turbines

Transcription

1 NATIONAL TECHNICAL UNIVERSITY of ATHENS Lab. Thermal Turbomachines Parallel CFD & Optimization Unit Design Optimization Tools & Applications 1. Introduction 2. Design of Matrix Turbines Kyriakos C. Giannakoglou Professor NTUA

2 Introduction to Optimization Problems/Methods From the Analysis to the Design-Optimization Single Objective Optimization, SOO Multi Objective Optimization, MOO Multi Disciplinary Optimization, MDO TERMINOLOGY: Understand d the difference: Optimal Design of a Car Design of Optimal Car Optimal Design of an Optimal Car Optimal Design of an Optimal Car Shape K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 2

3 What is Optimal (Car)? Objective Function The one with max. speed The one with min. fuel consumption The most comfortable one The less expensive The one with min. emissions Optimality : an Objective function (min ή max F) must be carefully defined!!! Objective Function F Cost Function (min) Fitness Function (max) But what if more than one objectives? K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 3

4 Transforming a MOO problem to a SOO one Example: Design of Optimal Wing max C L min -C L min 1/C L max C L min C D min C D min C D max -C D min C D + w/c L min C D + 1/C L min C D + 10/C L min C D subject to: C L =1.2 min C D subject to: C L >1.2 K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 4

5 Objective Function Objective: Minimum DRAG Objective Function: DRAG Coefficient min F=C D K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 5

6 Design (or Optimization) Variables Shape Parametrization Β Α θ y B y A x A x B L N=6 degrees of freedom (dofs) K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 6

7 Evaluation Evaluation Tool: Code for the numerical solution of the Navier-Stokes eqs. Viscous Stresses C D = carcontour forces K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 7

8 Evaluation L x A y A x B y B θ b = b1= b2= b3= b4= b5= b6= F=C D = K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 8

9 Constraints Constraints: Equality & Inequality Constraints!!! Feasible & Infeasible Solutions to the problem K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 9

10 Classification of Optimization Methods Gradient-Based Method vs. Stochastic Methods F(x) () Individual-based Methods vs. Population-based Methods x K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 10

11 Deterministic (Gradient-Based) Optimization F(x) () (if minimization) How to compute the gradient of F: Finite-Differencesit Complex Variable methods Automatic Differentiation Adjoint method x K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 11

12 Deterministic (Gradient-Based) Optimization Στόχος διατριβής Drag sensitivity map (a by-product of the adjoint method) Losses sensitivity map (a by-product of the adjoint method) (bridging the gap between modern design tools and an old-fashioned designer) K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 12

13 Evolutionary Algorithms A Gradient-Free Population-based Algorithm: K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 13

14 Evolutionary Algorithms Evaluation of the Population: F=0.29 F=0.33 F=0.36 F=0.50 F=0.45 F=0.42 K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 14

15 Evolutionary Algorithms Parent Selection: F=0.29 F=0.33 F=0.36 F=0.29 F=0.45 F=0.42 K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 15

16 Evolutionary Algorithms (Two Parents) Crossover (Two Offspring) K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 16

17 Evolutionary Algorithms Mutation K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 17

18 Evolutionary Algorithms The New Offspring Population K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 18

19 The Generalized (μ,λ) Evolutionary Algorithms Generation κ (λ offspring) λ evaluations Parent Selection (μ parents) Generation κ+1 (λ offspring) generated by applying evolution operators (Crossovermutation-elitism) K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 19

20 Multiobjective Optimization The Pareto Front Price Bugatti Veyron km/h VW Golf V km/h Simca km/h (Top Speed) 1 K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 20

21 Multiobjective Optimization The Pareto Front F 2, S ge Front 0 (Pareto) F 1 Front 1 Two-objective Example: min F 1, min F 2 Front 2 K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 21

22 Example: Design of Optimal Power Plants T P T T r P T P Three Objectives: (a) Max. efficiency, (b) Max. power output, (c) minimum investment Capital Cost (MEuro) T T Efficiency G2 Power Output (MW) Research funded by K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 22

23 Unfortunately: Conventional EA/ACO/PSO/BFO etc are computationally expensive, even on parallel platforms! This is where research is focusing during the last decade!!!!!!!!! K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 23

24 Metamodel-Assisted Evolutionary Algorithms (MAEAs) Evolution Evaluation Candidate solution Problem-Specific Evaluation Tool (Exact/Costly Model) Theroleormetamodels or during the evolution is, practically, to interpolate previously evaluated individuals (generated during the EA) for saving CPU cost. EA Fitness or Cost Value Metamodel or Surrogate Evaluation Tool (Approximate/Low-Cost) (F) rformance ( Pe Design Variable (b) K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 24

25 MAEAs with On-Line Trained Metamodels The concept of Inexact Pre-Evaluation within the evolutionary algorithm (EA-IPE or MAEA) K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 25

26 MAEAs with On-Line Trained Metamodels Generation 1 Generation 2 More generations must be performed by the EA; however, the majority of them relies on a limited number of calls to the problem-specific tool (CFD), so the CPU cost per generation is noticeably lower. Generation 3 Generatio n6 Generation 5 Generation 4 K.C. GIANNAKOGLOU: Design of Optimal Aerodynamic Shapes using Stochastic Optimization Methods and Computational Intelligence, Int. Review Journal Progress in Aerospace Sciences, Vol. 38, pp , K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 26

27 Distributed MAEAs (DMAEAs) Basic issues: Number of demes or islands Communication topology Communication frequency Migration algorithm EA set-up per deme M.K. KARAKASIS, A.P. GIOTIS and K.C. GIANNAKOGLOU: Inexact Information Aided, Low-cost, Distributed Genetic Algorithms for Aerodynamic Shape Optimization, Int. J. for Numerical Methods in Fluids, Vol. 43, pp , K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 27

28 Expected Gain in CPU Cost (EA/MAEA/DEA/DMAEA) Airfoil Shape Optimization (min. C D, fixed C L ) Flow Conditions M =0.75, α =2.734 o, Re= C L,ref =0.749, C D,ref =0.0235, β=2 Optimal Solution (SOO): C L,opt =0.744, C D,opt = EA-IPE = MAEA Evaluations K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 28

29 Hierachical EAs or MAEAs Hierarchical Evaluation Hierarchical Search Hierarchical Parameterization Navier-Stokes Turbulence Model Fine Grid, etc. GBM Finer Design Many Control Points Integral Method or Coarse Grid NS, etc EA or MAEA Knot Refinement Knot Removal Rough Design -Few Control Points K.C. GIANNAKOGLOU and I.C. KAMPOLIS, Multilevel Optimization Algorithms based on Metamodel- and Fitness Inheritance-Assisted Evolutionary Algorithms, in Computational Intelligence in Expensive Optimization Problems, Editors: Y. Tenne, C.-K. Goh, Springer-Verlag Series in Evolutionary Learning and Optimization, K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 29

30 Design-Optimization of Matrix Hydraulic Turbines Hydromatrix : a number of small, axial flow turbine generator units assembled in a grid or matrix. Advantages compared to conventional designs (lower cost to power ratio): 1. Minimization of the required civil construction works. 2. Minimum time for project schedules, construction and installation. 3. Small geological and hydrological risks. 4. Minimum environmental inflict. K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 30

31 Design-Optimization of Matrix Hydraulic Turbines Parametrization: Bezier curves are used to parameterize the spanwise distribution ib ti of: mean camber surface angles at LE & TE. circumferential position of the blade LE & TE. mean camber surface curvature. Blade thickness distribution. Total: 52 to 74 design variables K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 31

32 Design-Optimization of Matrix Hydraulic Turbines Objective 1 (G 1 ) : Minimization of the weighted sum of the deviations of the outlet swirl and axial velocity distributions from target curves Objective 2 (G 2 ): the standard deviation of the pressure distribution along the chordwise direction, at eleven equidistant spanwise locations Objective 3 (G 3 ): cavitation index K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 32

33 Design-Optimization of Matrix Hydraulic Turbines Full Load Part Load Best Effic H [m] Q[m 3 /sec] Part Load Best Efficiency Full Load (3 objectives) x (3 operating points) = 9 objectives in total K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 33

34 Design-Optimization of Matrix Hydraulic Turbines F 2 = f(g 3 ) at the three operating points Solution 1 Solution 2 Solution 3 F 1 = f(g 1,G 2 ) at the three operating points K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 34

35 Design-Optimization of Matrix Hydraulic Turbines (1) (2) (3) K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 35

36 Design-Optimization of Matrix Hydraulic Turbines NTUA The Evolutionary Algorithm SYstem K.C. Giannakoglou, Parallel CFD & Optimization Unit, NTUA, Greece 36