gans enabled super-resolution reconstruction of wind field · 2020. 1. 22. · gans enabled...

Duy Tan H. Tran 1 , Haakon Robinson 1 , Adil Rasheed 1,2 , Omer San 3 , Mandar Tabib 2 , Trond Kvamsdal 2,4 1 Department of Engineering Cybernetics, NTNU | 2 Mathematics and Cybernetics, SINTEF Digital | 3 Mechanical and Aerospace Engineering, OUS | 4 Department of Mathematical Sciences, NTNU Email: [email protected], [email protected], [email protected], [email protected], [email protected] GANs enabled super-resolution reconstruction of wind field HIRLAM (2.5km X 2.5km) SIMRA (200mX200m) G D HIGH RESOLUTION (200mX200m) COARSE RESOLUTION (800mX800m) GENERATED HIGH-RESOLUTION (200mX200m) FAKE REAL 17 th Deep Sea Offshore Wind R&D Conference, EERA DeepWind’ 2020, 15-17 January 2020, Trondheim, Norway Solution: Learn the flow characteristics using Generative Adversarial Networks (GANs) offline for efficient online predictions RESULTS: Comparison between Bicubic Interpolation, GANS and Ground truth Research challenge: Generating high resolution wind field in complex terrain without conducting computationally expensive high resolution numerical simulations Generator competes with discriminator to generate more realistic wind field Discriminator competes with generator to find its fault 3 years hourly data from HARMONIE-SIMRA used to train the GANs ACKNOWLEDGEMENT: The authors acknowledge the financial support from the Norwegian Research Council and the industrial partners of OPWIND: Operational Control for Wind Power Plants (Grant No.: 268044/E20). HIGHLIGHTS OF THE WORK ¾ A novel approach of combing physics based computational fluid dynamics simulator with the power of advanced machine learning algorithm like GANs is demonstrated for generating high resolution wind field in complex terrain. ¾ The GANs network learns the main characteristics of the flow in complex terrain and comfortably outperforms commonly used interpolation technique ¾ A significant improvement in computational efficiency (X100000) L2-norm Unseen image numbers GANs reconstructed super-resolution wind field reconstruction from coarse numerical simulation Offline GANs training with high resolution archived data from numerical simulations ZOOMED-IN VIEW BICUBIC GANs Truth

Upload: others

Post on 28-Jan-2021

3 views

Category:

Documents

0 download

Report

Download

Embed Size (px):

TRANSCRIPT

Duy Tan H. Tran1, Haakon Robinson1, Adil Rasheed1,2, Omer San3, Mandar Tabib2, Trond Kvamsdal2,41Department of Engineering Cybernetics, NTNU | 2Mathematics and Cybernetics, SINTEF Digital | 3Mechanical and Aerospace Engineering, OUS | 4Department of Mathematical Sciences, NTNU

Email: [email protected], [email protected], [email protected], [email protected], [email protected]

GANs enabled super-resolution reconstruction of wind field

HIRLAM (2.5km X 2.5km)SIMRA (200mX200m)

G

DHIGH RESOLUTION

(200mX200m)

COARSE RESOLUTION(800mX800m)

GENERATED HIGH-RESOLUTION(200mX200m)

FAKE

REAL

17th Deep Sea Offshore Wind R&D Conference, EERA DeepWind’ 2020, 15-17 January 2020, Trondheim, Norway

Solution: Learn the flow characteristics using Generative Adversarial Networks (GANs) offline for efficient online predictions

RESULTS: Comparison between Bicubic Interpolation, GANS and Ground truth

Research challenge: Generating high resolution wind field in complex terrain without conducting computationallyexpensive high resolution numerical simulations

Generator competes withdiscriminator to generate more

realistic wind field

Discriminator competes withgenerator to find its fault

3 years hourly data from HARMONIE-SIMRA used to train the GANs

ACKNOWLEDGEMENT: The authors acknowledge the financial supportfrom the Norwegian Research Council and the industrial partners ofOPWIND: Operational Control for Wind Power Plants (Grant No.:268044/E20).

HIGHLIGHTS OF THE WORK

A novel approach of combing physics based computational fluiddynamics simulator with the power of advanced machine learningalgorithm like GANs is demonstrated for generating high resolutionwind field in complex terrain.The GANs network learns the main characteristics of the flow incomplex terrain and comfortably outperforms commonly usedinterpolation techniqueA significant improvement in computational efficiency (X100000)

L2-n

orm

Unseen image numbers

GANs reconstructedsuper-resolution wind

field reconstructionfrom coarse numerical

simulation

Offline GANs training with high resolutionarchived data from

numerical simulations

ZOO

MED

-IN V

IEW

BICUBIC GANs Truth