wind turbine simulations using cpu/gpu heterogeneous...

1
Wind Turbine Simulations Using CPU/GPU Heterogeneous Computing Yong Su Jung Graduate Research Assistant University of Maryland, College Park James Baeder Professor University of Maryland, College Park Contact info. Yong Su Jung (Dept. AE) [email protected] 14th Symposium on Overset Composite Grids and Solution Technology A heterogeneous solution framework using both CPUs and GPUs has been used to numerically simulate flow over the NREL Phase VI horizontal-axis wind turbine. An in-house line-based unstructured flow solver on CPU (HAMSTR) is coupled to an in-house structured solver on GPUs (GARFIELD) through a lightweight Python-based framework within an overset mesh system. The full wind turbine simulations including the nacelle and tower are used to understand blade-tower interference on both upwind and downwind configurations and the predictions have been compared against experimental data in terms of blade airloads. The effects of laminar-turbulent transition are also investigated on the blade using transition model, whose inclusion resulted in a more accurate torque prediction and better resolved the unsteady airloads arising from blade-tower interference. A normal wind profile model is used to simulate free-stream wind shear for wind turbines operating in an atmospheric boundary layer. Couple unstructured grid based CPU solver and structured GPU accelerated solver Exploit advantages of both multiple mesh paradigms and line-based method Apply to full wind turbine simulation for investigating the effect of flow unsteadiness Apply laminar-turbulent transition model on wind turbine simulation to study the effect on turbine performance Abstract Objectives NRTEL Phase VI Wind Turbine Overset Grids Two-bladed horizontal-axis-wind-turbine model Rotor diameter of 10.058 m and hub height of 12.19 m Tower clearance of 1.401 m Operates on upwind/downwind configuration at 72 RPM Test wind speed : 7 m/s, 10 m/s, and 20 m/s Condition Rigid/Teetered Coning angle Blade tip pitch Upwind Rigid 0.0° 3.0° Downwind Teetered (Avg. 3.0°) 3.4° 3.0° *Dimensions are based on NASA Ames wind tunnel section Near body (CPU) Cylindrical nest (GPU) Background (GPU) 5.1 millions cell 5.8 millions cell 4.1 millions cell Coupled CPU-GPU Flow Solver HAMSTR: Line Method on Unstructured Grids for turbine blade, nacelle, and tower components GARFIELD : GPU Accelerated RANS Solver for background domain Communication between HAMSTR and GARFIELD is accomplished with overset grid assembler (TIOGA). Data pointers are shared in an Python framework which allow for the flexibility of a familiar scripting language without sacrificing performance. Laminar-Turbulent Transition Model transition model: Local correlation-based transition model for Spalart-Allmaras one equation turbulence model V (m/s) Torque [N-m] Experiment [N-m] Thrust [N] Experiment [N] 7 766 (696) 805 1,185 (1,126) 1,154 10 1,196 (1,172) 1,340 1,673 (1,665) 1,675 20 1,134 (1,084) 1,110 3,278 (3,261) 3,005 The use of transition model better estimates the torque. Turbulent transition onset along with laminar separation bubble was predicted at mid-chord along the span at wind speed of 7 m/s. Isolated Rotor Computation = 10 m/s = 20 m/s =7 m/s At operating condition of 7m/s, flow is fully attached on the blade. Separated flow presented on suction side at higher wind speeds. Explored CPU/GPU heterogeneous CFD framework for wind turbine simulation Observed improvements in turbine performance prediction using transition model, especially in attached flow condition Investigated blade-tower interference for both upwind and downwind configurations at wind speed of 7 m/s Studied effect of free-stream wind shear on the rotor performance and turbine wakes As a future work, study on optimal load balancing between CPU solver and GPU solver can be performed for a wind farm. Conclusions References 1. M. Hand et al, “Unsteady Aerodynamics Experiment Phase VI: Wind Tunnel Test Configurations and Available Data Campaigns,” NREL/TP-500-29955, National Renewable Energy Laboratory, 2001. 2. D. Simms et al, “NREL Unsteady Aerodynamics Experiment in the NASA-Ames Wind Tunnel: A Comparison of Predictions to Measurements,” NREL/TP-500-29494, National Renewable Energy Laboratory, 2001. 3. Y. Jung, B. Govindarajan, J. Baeder, “Turbulent and Unsteady Flows on Unstructured Line-Based Hamiltonian Paths and Strand Grids,” AIAA Journal, Vol. 55, pp.1986-2001, 2017. 4. D. Jude and J. Baeder, “Extending a Three-Dimensional GPU RANS Solver for Unsteady Grid Motion and Free-Wake Coupling,” 54 th AIAA SciTech Forum, San Diego, CA, 2016 5. S. Medida, “Correlation-based Transition Modeling for External Aerodynamic Flows,” University of Maryland PhD Dissertation, 2014. Atmospheric Boundary Layer Free-stream wind shear was described by normal wind profile using power law. Circular form low momentum wake zone lifted upward in wind shear. Large airload variations of sectional force occur near blade tip. Variation of blade airloads according to azimuth angle can induce highly unsteady hub moment. Opposite effects between two blades canceled each other in terms of rotor performance. Low momentum wake zone Tower shedding = ( ) at 3R at 6R Full Configuration Computation (Left: upwind, Right: downwind) decelerated Slower axial velocity due to tower blockage effect reduces blade aerodynamic force. The region affected by tower extends to half rotor revolution. Inner section experience more interaction due to lower tangential velocity The rotor was modeled without coning and teetered hub based on the blade being approximately parallel to the tower axis during tower passage. Blade shed wake interaction (BWI) was captured in downwind configuration. More severe blade-tower interference than upwind configuration, indicated by sharp drop in blade airloads. Over 30% reduction in single blade thrust and 50% reduction in torque at = 180° Transition Fully turbulence Transition Fully turbulence Leeward surface Windward surface Quadrangulated triangles Strand volume grid Quad-dominant grid 10 % Region of effect Unstructured Grid CPU Structured Grid GPU 1.2 R 2.5 R Cylindrical nested mesh Chord-wise pressure distribution (Top: r/R=0.3, Bottom: r/R=0.8) Iso-surface of vorticity magnitude contour Inflow distribution ( ) at sectional plane of 0.63R Inflow distribution ( ) at sectional plane of 0.63R 60 % 25 % 15 % Normal wind profile (=0.2) [m/s] Sectional force distribution along radial direction Flowchart showing function calls through Python

Upload: others

Post on 17-Apr-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Wind Turbine Simulations Using CPU/GPU Heterogeneous …2018.oversetgridsymposium.org/assets/posters/Jung_Poster_OGS20… · numerically simulate flow over the NREL Phase VI horizontal-axis

Wind Turbine Simulations Using CPU/GPU Heterogeneous ComputingYong Su Jung

Graduate Research AssistantUniversity of Maryland, College Park

James BaederProfessor

University of Maryland, College Park

Contact info.Yong Su Jung (Dept. AE)[email protected]

14th Symposium on Overset Composite Grids and Solution Technology

A heterogeneous solution framework using both CPUs and GPUs has been used tonumerically simulate flow over the NREL Phase VI horizontal-axis wind turbine. An in-houseline-based unstructured flow solver on CPU (HAMSTR) is coupled to an in-house structuredsolver on GPUs (GARFIELD) through a lightweight Python-based framework within anoverset mesh system. The full wind turbine simulations including the nacelle and tower areused to understand blade-tower interference on both upwind and downwind configurationsand the predictions have been compared against experimental data in terms of bladeairloads. The effects of laminar-turbulent transition are also investigated on the blade using𝛾𝛾 − 𝑅𝑅𝑅𝑅𝜃𝜃𝜃𝜃 − 𝑆𝑆𝑆𝑆 transition model, whose inclusion resulted in a more accurate torqueprediction and better resolved the unsteady airloads arising from blade-tower interference. Anormal wind profile model is used to simulate free-stream wind shear for wind turbinesoperating in an atmospheric boundary layer.

• Couple unstructured grid based CPU solver and structured GPU accelerated solver

• Exploit advantages of both multiple mesh paradigms and line-based method

• Apply to full wind turbine simulation for investigating the effect of flow unsteadiness

• Apply laminar-turbulent transition model on wind turbine simulation to study the effect on turbine performance

Abstract

Objectives

NRTEL Phase VI Wind Turbine Overset Grids• Two-bladed horizontal-axis-wind-turbine model

• Rotor diameter of 10.058 m and hub height of 12.19 m

• Tower clearance of 1.401 m

• Operates on upwind/downwind configuration at 72 RPM

• Test wind speed : 7 m/s, 10 m/s, and 20 m/s

Condition Rigid/Teetered Coning angle Blade tip pitchUpwind Rigid 0.0° 3.0°

Downwind Teetered (Avg. 3.0°)

3.4° 3.0°

*Dimensions are based on NASA Ames wind tunnel section

Near body (CPU)

Cylindrical nest (GPU)

Background (GPU)

5.1 millions cell 5.8 millions cell 4.1 millions cell

Coupled CPU-GPU Flow Solver

• HAMSTR: Line Method on Unstructured Grids for turbine blade, nacelle, and tower components

• GARFIELD : GPU Accelerated RANS Solver for background domain

• Communication between HAMSTR and GARFIELD is accomplished with overset grid assembler (TIOGA).

• Data pointers are shared in an Python framework which allow for the flexibility of a familiar scripting language without sacrificing performance.

Laminar-Turbulent Transition Model• 𝜸𝜸 − 𝑹𝑹𝑹𝑹𝜽𝜽𝜽𝜽 − 𝑺𝑺𝑺𝑺 transition model: Local correlation-based transition model for Spalart-Allmaras one equation

turbulence model

V∞(m/s)

Torque[N-m]

Experiment[N-m]

Thrust[N]

Experiment[N]

7 766 (696) 805 1,185 (1,126) 1,15410 1,196 (1,172) 1,340 1,673 (1,665) 1,67520 1,134 (1,084) 1,110 3,278 (3,261) 3,005

• The use of transition model better estimates the torque.

• Turbulent transition onset along with laminar separation bubble was predicted at mid-chord along the span at wind speed of 7 m/s.

Isolated Rotor Computation𝑽𝑽∞= 10 m/s 𝑽𝑽∞= 20 m/s𝑽𝑽∞=7 m/s

• At operating condition of 7m/s, flow is fully attached on the blade.• Separated flow presented on suction side at higher wind speeds.

• Explored CPU/GPU heterogeneous CFD framework for wind turbine simulation

• Observed improvements in turbine performance prediction using 𝛾𝛾 − 𝑅𝑅𝑅𝑅𝜃𝜃𝜃𝜃 − 𝑆𝑆𝑆𝑆 transition model, especially in attached flow condition

• Investigated blade-tower interference for both upwind and downwind configurations at wind speed of 7 m/s

• Studied effect of free-stream wind shear on the rotor performance and turbine wakes

• As a future work, study on optimal load balancing between CPU solver and GPU solver can be performed for a wind farm.

Conclusions

References1. M. Hand et al, “Unsteady Aerodynamics Experiment Phase VI: Wind Tunnel Test Configurations and

Available Data Campaigns,” NREL/TP-500-29955, National Renewable Energy Laboratory, 2001.2. D. Simms et al, “NREL Unsteady Aerodynamics Experiment in the NASA-Ames Wind Tunnel: A Comparison

of Predictions to Measurements,” NREL/TP-500-29494, National Renewable Energy Laboratory, 2001.3. Y. Jung, B. Govindarajan, J. Baeder, “Turbulent and Unsteady Flows on Unstructured Line-Based

Hamiltonian Paths and Strand Grids,” AIAA Journal, Vol. 55, pp.1986-2001, 2017.4. D. Jude and J. Baeder, “Extending a Three-Dimensional GPU RANS Solver for Unsteady Grid Motion and

Free-Wake Coupling,” 54th AIAA SciTech Forum, San Diego, CA, 20165. S. Medida, “Correlation-based Transition Modeling for External Aerodynamic Flows,” University of Maryland

PhD Dissertation, 2014.

Atmospheric Boundary Layer• Free-stream wind shear was described by normal wind profile using power law.

• Circular form low momentum wake zone lifted upward in wind shear.

• Large airload variations of sectional force occur near blade tip.

• Variation of blade airloads according to azimuth angle can induce highly unsteady hub moment.

• Opposite effects between two blades canceled each other in terms of rotor performance.

Low momentum wake zone

Tower shedding

𝑉𝑉 𝑧𝑧 = 𝑉𝑉ℎ𝑢𝑢𝑢𝑢( ⁄𝑧𝑧 𝑧𝑧ℎ𝑢𝑢𝑢𝑢)𝛼𝛼

at 3R at 6R

Full Configuration Computation (Left: upwind, Right: downwind)

decelerated

• Slower axial velocity due to tower blockage effect reduces blade aerodynamic force.

• The region affected by tower extends to half rotor revolution.

• Inner section experience more interaction due to lower tangential velocity

• The rotor was modeled without coning and teetered hub based on the blade being approximately parallel to the tower axis during tower passage.

• Blade shed wake interaction (BWI) was captured in downwind configuration.

• More severe blade-tower interference than upwind configuration, indicated by sharp drop in blade airloads.

• Over 30% reduction in single blade thrust and 50% reduction in torque at 𝜓𝜓 = 180°

Transition Fully turbulence Transition Fully turbulence

Leeward surface Windward surface

Quadrangulated triangles

Strand volume grid

Quad-dominant grid

10 %

Region of effect

Unstructured Grid CPU

Structured Grid GPU

1.2 R2.5 R

Cylindrical nested mesh

Chord-wise pressure distribution (Top: r/R=0.3, Bottom: r/R=0.8)

Iso-surface of vorticity magnitude contourInflow distribution ( ⁄𝑤𝑤 𝑉𝑉∞) at sectional plane of 0.63R

Inflow distribution ( ⁄𝑤𝑤 𝑉𝑉∞) at sectional plane of 0.63R

60 %25 %

15 %

Normal wind profile (𝛼𝛼=0.2)

𝑉𝑉 [m/s]

Sectional force distribution along radial direction

Flowchart showing function calls through Python