The State of the Art in Short-term Prediction of Wind Power

Transcription

1 The State of the Art in Short-term Prediction of Wind Power Dr. Gregor Giebel DTU Wind Energy And numerous co-authors, not all knowingly SANEDI, Sandton, South Africa, 19 Sept 2017

2 Outline Why predictions? and for whom? Predictions: The general information flow Some actual implementations Some knowledge from research Best Practice

3 State-of-the-Art for Wind Power

4 State-of-the-Art in Short-term Prediction

5 State-of-the-Art in Short-term Prediction 111 pages >380 references >700 citations DOI: /DTU:

6 Source: WindEurope Daily Wind newsletter, valid for 13 Sept September

7 Record in DK: 140% wind power! Low demand and high wind production in a summer night. Here given for all of Denmark. Source: September

8 Smoothing

9 All Europe is connected The image is of 1997, but it s not much changed since then However, local or cross-border transmission can be a bottle-neck

10 Data from 60 meteorological stations One year of three-hourly wind speed data from (mostly) 10 m Also many inland sites Then spatially averaged over all time series at every time step => Total generation profile is less variable!

11 Cross-correlation versus distance Calculate cross-correlation Correlation coefficient (lag=0 hours) Distance [km] coefficient between every pair of stations Result: Cross-correlation decreases with distance Exponential fit has shape parameter of ca. 700 km

12 Users

13 Users of forecasts Who needs forecasts: Transmission companies in areas with high wind penetration (eg Energinet.dk, Tennet, 50Hertz, Red Electrica de España, CaISO, AEMO, ESKOM ) Wind power owners/operators with own market access (Iberdrola, DONG, Vattenfall, RWE Innogy, Vindkraft, DTEK, ) Electrical utilities (eg DONG Energy, Vattenfall, Acciona, Iberdrola, E.On, NUON, RWE, EnBW, ESKOM ) Everyone trading on markets with sizeable shares of wind power

14 Thomas Ackermann about TSOs Source: 1 st Workshop on Short-term Forecasting, Uni NSW, Sydney (AUS), December 2005

15 Level of users 2016 AU.SA Slide source: Waldl: Operational wind & solar power forecasting - The Perfect Wind Power Prediction. Talk on the 1 st Workshop on Large-scale Grid Integration of Renewables in India, Dehli, 6-8 Sept 2017

16 Timescales for wind forecasts Min/sec Source: The Future of Wind. White paper on Wind Power Monthly Expert reports, MetOffice, 2013

17 Cost functions Different users have different cost functions Even within their organisations Might not even be obvious for them Uses: trading wind power definition of reserve requirements unit commitment and economic dispatch operation of combined wind-hydro operation of wind associated with storage design of optimal trading strategies electricity market design etc.

18 Strategic bidding vs TSO responsibility REE showed first that aggregating the forecasts coming from the market participants was worse than their in-house state-of-the-art tool Due to better information available at the TSO Forecasts had variable quality (usually coming from the lowest bidder, not necessarily the best) Potential for strategic bidding of the market participants: E.g. allowing for a +- 20% corridor means that bidding the full volume increases the risk of error safe bid is 80% max. Requires online SCADA data at TSO, including active power, curtailment, maintenance information, Often mandated in grid codes September

19 Predictions HowTo

20 Short-Term Prediction Overview Orography Roughness Wind farm layout Online data GRID End user TRADING Numerical Weather Prediction Prediction model Image sources: DWD, WAsP, Joensen/Nielsen/Madsen EWEC 97, Pittsburgh Post-Gazette, Red Electrica de España.

21 Short-Term Prediction Overview Orography Roughness Wind farm layout Online data GRID End user TRADING Numerical Weather Prediction Prediction model Image sources: DWD, WAsP, Joensen/Nielsen/Madsen EWEC 97, Pittsburgh Post-Gazette, Red Electrica de España.

22 Stakeholders End user Numerical Weather Prediction Prediction model Image sources: DWD, WAsP, Joensen/Nielsen/Madsen EWEC 97, Red Electrica de España.

23 Statistical power curve estimation Establish best connection between NWP wind speed and measured power Often non-parametric and not a function Often recursively adapted with new online data

24 Data with turbine availability and curtailment Power [%] 50 Unconstrained wind farm production Limited turbine availability Grid curtailment 15 4:30 5:30 6:30 7:30 8:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30 16:30 17:30 18:30 19:30 20:30 21:30 Time [hours] Slide source: Waldl: Operational wind & solar power forecasting - The Perfect Wind Power Prediction. Talk on the 1 st Workshop on Large-scale Grid Integration of Renewables in India, Dehli, 6-8 Sept 2017

25 Performance

26 Source: Advanced Short Range Wind Energy Forecasting Technologies-- Challenges, Solutions, and Validation Kristin Larson and Tillman Gneiting, Global WINDPOWER 2004, March 31, 2004

27 Time and space scale of atmospheric motion Typical sizes globalscale 2000 km synopticscale 2000 km microscale mesoscale 20 km Thunderstorms tornadoes waterspouts 2 m small turbulent eddies Land-sea breeze Mountainvalley breeze Typhones Tropical Storms Mid latitudes Hs & Ls fronts Long waves secs to mins mins to hours hours to days days to a week or more Source: Jesper Nissen Typical life span

28 Synoptic scale meteorology High pressure system Low pressure system Cold front Warm front ( sure.html#view) Source: Jesper Nissen

29 Mesoscale Meteorology Thunderstorm Coastal low level jets Picture from Sea breeze circulation L H H SEA L LAND Source: Jesper Nissen

30 Micro scales There are always unresolved processes that cannot be represen by a numerical model. These features are approximated through Parametrization! Turbulent strees seen on the sea surface scales m Dust devils scale m Source: Jesper Nissen

31 NWP: DMI-HIRLAM

32 Level vs. Phase errors There are two error categories from the NWP: level errors and phase errors Level error means not to predict the intensity of a storm Phase errors are misjudging the timing of the storm Phase errors are more frequent these days (but how to report them?) Ca 80% of the error comes from the NWP!

33 Phase and Level errors Errors can be phase (timing) or level errors Actual Production Short Term Forecast MW /01 00:00 09/01 04:48 09/01 09:36 09/01 14:24 09/01 19:12 10/01 00:00 Time 10/01 04:48 10/01 09:36 10/01 14:24 10/01 19:12 11/01 00:00

34 Common evaluation criteria Since none were available, the ANEMOS project codified common criteria for performance measurements of short-term forecasting systems: Mean Error Mean Absolute Error Root Mean Square Error R 2 (coefficient of determination) Histogram of errors Also, use separate training and validation datasets Present the errors normalised with the installed capacity Madsen, H., P. Pinson, G. Kariniotakis, H.Aa. Nielsen, T.S. Nielsen: Standardizing the Performance Evaluation of Short-term Wind Power Prediction Models. Wind Engineering 29(6), pp , 2005

35 Typical results (1996 now more like 10%) RMS Error [% of nameplate capacity] Nøjsomhedsodde wind farm Peak capacity=5030 kw Persistence HWP Mean RMS Error [kw] Forecast length [h]

36 Typical results (1996 now more like 10%) RMS Error [% of nameplate capacity] Nøjsomhedsodde wind farm Peak capacity=5030 kw Persistence HWP NewRef Mean RMS Error [kw] Forecast length [h]

37 Typical results (1996 now more like 10%) RMS Error [% of nameplate capacity] Nøjsomhedsodde wind farm Peak capacity=5030 kw Typical forecast accuracy of a modern system Persistence HWP NewRef HWP / MOS RMS Error [kw] Forecast length [h]

38 Wind Speed Dependency Wind speed errors from the NWP are quite similar across the whole range of wind speeds Folded through the power curve gives large errors for rising parts, small errors for flat part Source: M. Lange, D. Heinemann: Accuracy of Short Term Wind Power Predictions Depending on Meteorological Conditions. Proceedings of the Global Wind Power Conference, Paris 2002 Central plot from Lange, M., and U. Focken: Physical Approach to Short-Term Wind Power Prediction. Berlin: Springer-Verlag, 2005

39 Forecast accuracy, historical (eg ISET) Forecasting got better during the last years Some of it piggybacks on improvements in meteorology Some is due to better interface to meteorological models (e.g., using 100m wind speed) Some is using multi-model approach Graph shows error in E.On control zone over the years, with references from the paper B. Lange et.al.: Wind Power Prediction in Germany - Recent advances and future challenges. Paper on the EWEC 2006 in Athens.

40 Smoothing of forecast errors Focken et al looked into the spatial smoothing of forecast errors left is actual, right is derived model Therefore, predictions for a region always are better than predictions for a single wind farm Source: Lange, M., and U. Focken: Physical Approach to Short-Term Wind Power Prediction. Berlin: Springer-Verlag, 2005

41 History

42 Ed McCarthy Predicted for the large wind farms in California (Altamont, San Gorgognio etc) Was run in the summers of On a HP 41CX programmable calculator Using meteorological observations and local upper air observations The program was built around a climatological study of the site and had a forecast horizon of 24 hours. It forecast daily average wind speeds with better skill than either persistence or climatology alone.

43 Prediktor Applied in Eastern Denmark between 1993 and 1999 Similar: Previento

44 Previento Similar to Prediktor, but uses more stringent physical downscaling (incl stability) and specialised upscaling Operational at EWE, E.On, RWE, Vattenfall, EnBW University of Oldenburg / energy & meteo systems GmbH Power Prediction ω ω *

45 Wind Power Prediction Tool Developed at IMM/DTU Operational in Western DK 1994 Operational for all of DK 1999 Statistical non-parametric adaptive models for prediction of representative farms Upscaling statistically to installed capacity Employs data cleaning Similar: Sipreólico, WPMS, MORE-CARE

46 Fraunhofer IWES WPMS Wind Power Management System = Nowcasting + Forecasting In use at E.On Netz since 2001, RWE since 6/2003, Vattenfall Europe 2004 E.On case: 50 representative wind farms (soon more) from WMEP -> ANN upscaling = Nowcast DWD Lokalmodell and others provide for forecast Accuracy: after 7 hours purely NWP dominated (5% RMS for E.On Netz total area)

47 New for WE: 100m winds 100m wind forecasts and analysis (from ECMWF) publically available from August Featured in the Weather Eye column of The Times (19th of August 2010) Evaluation of 100m deterministic and ensemble forecasts Focus on sites with available observations (e.g. Fino) Potential benefit from getting more model-level wind forecasts

48 1990 Evolution of the state of the art Deterministic approaches Hybrid approaches Next generation of tools Focus on extremes Diversify predicted information Spatio-temp Ramp forecasting Cut-off forecasting Alarming, risk indices Link to meteorology etc SafeWind 2002 Anemos ANEMOS./plus Functions considered: Power system scheduling Reserves estimation Probabilistic forecasting Probabilistic approaches 1st Benchmarking Towards standardisation Combined forecasts (multi model/nwps) Congestion management Wind/ storage coordination Optimal trading ANEMOS.plus demos End-users : TSOs DSOs Island system operators Utilities Traders Link forecasts to the application Power system management/trading Stochastic optimisation Demonstration

49 Research results

50 Doubling the number of NWP Used DMI and DWD for six test cases in Denmark Result: the combination of inputs is better Klim Tunø Knob Middelgrunden Fjaldene Syltholm Hagesholm G. Giebel, A. Boone: A Comparison of DMI-Hirlam and DWD-Lokalmodell for Short-Term Forecasting. Poster on the EWEC, London, Nov 2004

51 Benefit of multiple NWPs Combining two NWPs improves results Alaiz data, Hirlam and MM5 (CENER) as NWP H.Aa. Nielsen: Slides on project meeting, PSO project Intelligent Prognosis Systems, January 2006 at Risø

52 Spatio-temporal improvement of forecasts Work done by DTU-IMM (now DTU-Compute) and ENFOR in the SafeWind project. J. Tastu, P. Pinson, E. Kotwa, H.Aa. Nielsen, H. Madsen (2011). Spatio-temporal analysis and DTU modeling Wind Energy, of wind Technical power University forecast of errors. Denmark Wind Energy 14(1), pp

53 Anemos

54 Overview 3 EU sponsored projects (20 M total, 15 M from EU) Started 2002 Most important European institutes, many important end users, some meteorological providers Common shell for models developed and installed at utilities / TSOs Comparison of existing models Advancing the state-of-the-art in statistical, physical and offshore predictions Also commercial installations: AEMO, SONI Many slides in this section come from Georges Kariniotakis, Anemos project leader at Ecole des Mines / Armines.

55 The Consortium of Anemos IASA

56 Evolution Accuracy! Uncertainty! 1990ies to 2005 Meteorology Wind power forecasting technology Graphics: George Kariniotakis Operational decision making

57 Evolution Feedback to meteorology, dedicated wind power meteorological forecasts Meteorology Wind power forecasting technology Graphics: George Kariniotakis Operational decision making

58 Evolution Specialised forecast products for power system issues: Trading Ramps / Variability Medium-range forecasts Congestion / Storage management Meteorology Wind power forecasting technology Graphics: George Kariniotakis Operational decision making

59 Evolution Specialised forecast products for power system issues: Trading Ramps / Variability Medium-range forecasts Congestion / Storage management Meteorology Wind power forecasting technology Graphics: George Kariniotakis Operational decision making

60 Evolution Increased interaction between wind power forecasting and meteorological community Meteorology Wind power forecasting technology Graphics: George Kariniotakis Operational decision making

61 Evolution Increased interaction between wind power forecasting and meteorological community Meteorology Wind power forecasting technology Graphics: George Kariniotakis Operational decision making

62 The forecasting system Generic architecture An operational prediction platform is developed covering a wide range of end-user requirements: single wind farm forecasting up to regional/national forecasting.

63 Portfolio of models Short-term (0-6 hours) statistical Medium-term (0-48/72 hours) statistical & physical from the leading model providers (choose some or all) Combined approaches Regional / National prediction On-line uncertainty estimation Probabilistic forecasts Risk assessment Numerical Weather Predictions by alternative models as input

64 Operational experience France (utility) Germany (utility) Greece (utility) Ireland (TSO) Spain (TSO, developer) Real-time operation & demonstration in 9 countries by 10 end-users. Portugal (TSO) UK (Northern Ireland TSO) Denmark (utility) Australia (TSO, commercial installation)

65 IEA Wind Task 36

66 IEA Wind Task 36 Forecasting International Energy Agency (IEA) has several Energy Technology Networks, e.g. Wind Power. Forecasting Task (36) runs 1/ /2018 (and probably new phase after that). Some 200 people from weather services, operational forecasters, academia and end users (TSOs, operators, traders, ). Participation is open for IEA Wind member countries, but news are distributed via a mailing list (send mail to grgi@dtu.dk). See details at September

67 Technical Results Mainly: published 5 lists, useful for peers Tall masts for NWP verification, and how to access their data Field experiments in wind power meteorology Openly available benchmarks for power forecasts Research projects in the field Future research issues

68 Use of probabilistic forecasting Open Access journal paper 48 pages on the use of uncertainty forecasts in the power industry Definition Methods Communication of Uncertainty End User Cases Pitfalls - Recommendations Source: September

69 Best Practice

70 Best Practice Get a model (24/7) Get another model (NWP and / or short-term forecasting model) Use online power data for first hours and power curve calibration Work together with service provider / academia / weather service to continuously improve model accuracy Reduce error by predicting for a larger area (smoothing) Balance all errors together, not just wind Use the uncertainty / pdf Do forecasting on TSO level, not necessarily on wind farm / developer level Use intraday trading Use longer forecasts for maintenance planning Meteorological training for the operators Meteorological hotline for special cases Also in report on powwow.risoe.dk (Giebel and Kariniotakis: Best Practice in Short-term Forecasting. A User s Guide. Project report for the POW WOW project, 6 pages, 2009)