TOPFARM Lowcost wake in DTU Engineering datadriven models

TOPFARM Low-cost wake (in DTU): Engineering & data-driven models catalogue Tuhfe Göçmen

Low-Cost Wake : What & Why & How • Low-Cost ≈ Low Computational Cost • Why fast(er) estimation? o Wind Farm layout optimization o Operation Monitoring o Ancillary Services (system/market integration : possible power + forecasting) o Wind Farm control § Uncertainty Quantification • Faster by (further) simplification low(er)-fidelity o In terms of governing equations o In terms of ‘sub-components’ of wake modelling • 2 Simple ≠ Easy … DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake : Contents • Governing Equations – Brief Introduction • Subcomponents of Wake Modelling – Inflow Generation – Wake Summation • Low-cost wake models (in DTU) – The Jensen Model Ø with Gaussian deficit – The Larsen Model – FUGA – Dynamic Wake Meandering Model • Better performance & better validation practices – Re-calibration for model improvement – Machine Learning for wakes 3 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake : Governing Equations (briefly) • Unsteady Convection terms Pressure forces Viscous forces External forces Inertial forces 4 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake : Continues… • Governing Equations – Brief Introduction • Subcomponents of Wake Modelling – Inflow Generation – Wake Summation • Low-cost wake models (in DTU) – The Jensen Model Ø with Gaussian deficit – The Larsen Model – FUGA – Dynamic Wake Meandering Model • Better performance & better validation practices – Re-calibration for model improvement – Machine Learning for wakes 5 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake : Subcomponents Inflow Generation if l l a m r s o y t all res area e R t fo x no mple co 6 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake : Subcomponents Wake Summation • 7 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake : Continues… • Governing Equations – Brief Introduction • Subcomponents of Wake Modelling – Inflow Generation – Wake Summation • Low-cost wake models – The Jensen Model Ø with Gaussian deficit – The Larsen Model – FUGA – Dynamic Wake Meandering Model • Better performance & better validation practices – Re-calibration for model improvement – Machine Learning for wakes 8 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake Models : Jensen • Jensen (1983) & Katic (1986) – Simple, practical, robust – The wake is ü Axisymmetric ü Non-turbulent ü Conical shape Jensen N. O. , 1983, A note on wind generator interaction, Technical Report Risø-M-2411, Risø National Laboratory, Roskilde, Denmark Katic I. , Højstrup J. , Jensen N. O. , 1986, A simple model for cluster efficiency, In Proceedings of the European Wind Energy Association Conference, pp. 407 -410, Rome, Italy 9 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake Models : Jensen & stability • if l l a m r s o y t all res area e R t fo x no mple co Alfredo Peña, Pierre‐Elouan Réthoré, M. Paul van der Laan, 2015. On the application of the Jensen wake model using a turbulence ‐dependent wake decay coefficient: the Sexbierum case, Wind Energy DTU Wind Energy, Technical University of Denmark

Low-Cost Wake Models : Jensen with Gaussian Deficit • Bastankhah (2014) & Abkar (2015) – Modified idea of Jensen ü One-parameter model ü Linear wake expansion – Inflow with shear – Axisymmetric Gaussian deficit ü k* defines the ‘width’ & ‘depth’ Majid Bastankhah and Fernando Port e-Agel. A new analytical model for wind-turbine wakes. Renewable Energy, 70: 116– 123, 2014 Mahdi Abkar and Fernando Porte-Agel. Influence of atmospheric stability on wind-turbine wakes: A large-eddy simulation study. Physics of Fluids, 27(3): 1– 20, 2015 11 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake Models : Larsen • Single Wake – Larsen(1988) WF Approach – Larsen(2009) Larsen GC. A simple wake calculation procedure; 1988 Larsen G. A simple stationary semi-analytical wake model. Technical report. Technical Report Risø; 2009 12 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low(er)-Cost Wake Models : Fuga • Ott et al. (2011) Ott S. , Berg J. , Nielsen M. , 2011, Linearised CFD Models for Wakes, Technical Report Risø-R-1772(EN). Risø National Laboratory, Roskilde, Denmark 13 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low(er)-Cost Wake Models : DWM • Larsen et al. (2007, 2008) & HAWC 2 implementation – Madsen et al. (2010) • 3 elements to describe the flow characteristics • Medium-fidelity Larsen G. C. , Madsen H. Aa. , Bingöl F. , Mann J. , Ott S. , Sørensen J. N. , Okulov V. , Troldborg N. , Nielsen M. , Thomsen K. , Larsen T. J. , Mikkelsen R. , 2007, Dynamic wake meandering modeling, Technical Report Risø-R-1607(EN). Risø Larsen G. C. , Madsen H. A. , Thomsen K. , Larsen T. J. , 2008, Wake meandering: a pragmatic approach, Wind Energy, vol. 11, pp. 377– 395 Madsen H. Aa. , Larsen G. C. , Larsen T. J. , Troldborg N. , Mikkelsen R. , 2010, Calibration and validation of dynamic wake meandering model for implementation in an aeroelastic code, Journal of Solar Energy Engineering, vol. 132, no: 4 14 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake : Performance? • How to assess the performance of a wake model? • Which model(s) to choose? – Which application you’d like to use the model for? Ø Long-term & statistical wake properties? ü AEP assessment (e. g. Layout optimization) ü Lifetime assessment / extension (e. g. Extreme winds, certification) Ø Seasonal analysis? ü Focus on certain interval(s) within the year (e. g. Turbine component(s) field test) Ø (Quasi) dynamic performance? ü Online monitoring, short-term power estimation – What type of measurements/data are available? Ø Met-mast, mesoscale, lidars, radars, SCADA, high(er) fidelity simulations, etc. – How “accurate” you’d need the results to be? Ø Risk assessment, grid codes, high sensitivity/dependency on wake behaviour… 15 DTU Wind Energy, Technical University of Denmark 10/31/2020

Model Performance test case : WF efficiency − Øresund, 6 -8 km offshore − 48 SWT 2. 3 MW turbines 110 MW − Close spacing (especially for offshore) − 2 “missing” turbines in the middle − 5 -years SCADA data − Produced Power / ‘Gross’ Production Ø Output uncertainty Ø WD dependency 16 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake Benchmark : WF efficiency 17 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake : Continues… • Governing Equations – Brief Introduction • Subcomponents of Wake Modelling – Inflow Generation – Wake Summation • Low-cost wake models – The Jensen Model Ø with Gaussian deficit – The Larsen Model – FUGA – Dynamic Wake Meandering Model • Better performance & better validation practices – Re-calibration for model improvement – Machine Learning for wakes 18 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake, Re-calibration : Jensen & TI • Duc, T, Giebel, G, Göçmen, T, Korpås, M & Coupiac, O 2017, Optimization of wind farm power production using innovative control strategies. DTU Wind Energy Master Thesis M, vol. 0161 DTU Wind Energy, Technical University of Denmark

Low-Cost Wake, Re-calibration : Gaussian Deficit • 1 Hz SCADA & high performance expectations for short-term wakes – can we actually capture (some) dynamics? 20 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake, Re-calibration : Gaussian Deficit • 1 Hz SCADA & high performance expectations for short-term wakes – can we actually capture (some) dynamics? Göçmen, T & Giebel, G 2018, 'Data-driven Wake Modelling for Reduced Uncertainties in short-term Possible Power Estimation: Paper' Journal of Physics: Conference Series, vol 1037, no. 7, 072002. 21 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake, Re-calibration : Gaussian Deficit • 1 Hz SCADA & high performance expectations for short-term wakes – can we actually capture (some) dynamics? 22 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake, Re-calibration : Larsen 23 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake, Re-calibration : Larsen • Re-calibrated Larsen WF (and time) specific parameters 24 DTU Wind Energy, Technical University of Denmark 10/31/2020

Low-Cost Wake Modelling : Machine Learning • Machine Learning Platform – Tensor. Flow – With Keras wrapper in Python – Fast & easy to apply • The deep learning algorithm – LSTM – Long Short-term Memory – Special building unit for RNN – Shown to perform faster & better for highly fluctuating time series • The inputs from the upstream turbines – Defined at every minute (WD dependent) – WD, Ueff, std(Ueff), ct – Data fed for the previous 1 -hour • Time window of 1 -hour shifted forward at every minute http: //colah. github. io/posts/2015 -08 -Understanding-LSTMs/ Input interval 25 DTU Wind Energy, Technical University of Denmark Output interval 10/31/2020

Low-Cost Wake Modelling : Machine Learning • Machine Learning Platform – Tensor. Flow – With Keras wrapper in Python – Fast & easy to apply • The deep learning algorithm – LSTM – Long Short-term Memory – Special building unit for RNN – Shown to perform faster & better for highly fluctuating time series • The inputs from the upstream turbines – Defined at every minute (WD dependent) – WD, Ueff, std(Ueff), ct – Data fed for the previous 1 -hour • Time window of 1 -hour shifted forward at every minute http: //colah. github. io/posts/2015 -08 -Understanding-LSTMs/ Input interval 26 DTU Wind Energy, Technical University of Denmark Output interval 10/31/2020

Low-Cost Wake Modelling : Machine Learning • Machine Learning Platform – Tensor. Flow • The inputs from the upstream turbines – Defined at every minute (WD dependent) – WD, Ueff, std(Ueff), ct – Data fed for the previous 1 -hour • Time window of 1 -hour shifted forward at every minute • The output – Ueff at the downstream turbine • New network (or model) per WF per turbine per minute – Still feasible real time! • 20 epochs • Batch size = 64 • Single hidden layer with 18 neurons 27 DTU Wind Energy, Technical University of Denmark

Low-Cost Wake : Wrap-up & Summary • Low-Cost ≈ Low Computational Cost • Relevant Applications o Wind Farm layout optimization – priority in TOPFARM o Operation Monitoring o Ancillary Services (system/market integration : possible power + forecasting) o Wind Farm control – highly relevant for TOPFARM § Uncertainty Quantification is crucial for all the applications above! • A number of models available o Application, available measurements, required accuracy o ‘Probabilistic’ as opposed to ‘deterministic’ distribution vs. mean o Learning algorithms : low(er) fidelity no fidelity • 28 Low-cost ≠ Cheap… DTU Wind Energy, Technical University of Denmark 10/31/2020