Time-series modelling of aggregate wind power output Alexander Sturt, Goran Strbac 17 March 2011 Introduction Eastern Wind Integration and Transmission Study (EWITS) (2010) • Wind datasets prepared by AWS Truewind over 9 month period • Created by simulation using mesoscale Numerical Weather Prediction (NWP) model • 3 years of synthetic data, 1326 sites (freely available online) • Hardware used: 80 x dual CPU quad core penguin workstations (640 cores) • Run time per year of simulation: 21 days (in theory...) What if this level of detail isn’t needed? What if we need a model of aggregated wind output? What if we need to understand the statistical properties? Modelling strategy • • Univariate model for aggregate wind power, not wind speed Autoregressive driver: AR(p), hourly (or half-hourly) timesteps X k 1 X k 1 2 X k 2 ... k • iid N(0,1) Include diurnal variation with periodic additive term: X k X k k mod n n = number of data points per day • Fit to long-term distribution with transformation function: • Use different models for the different seasons Pk W X k Model calibration 1. Choose these to satisfy long-term distribution and diurnal variation, assuming X~N(0,1) 1.4 1.2 1 X 0.8 0.6 0.4 0.8 0.2 0.7 1 0 0.6 W -0.2 -0.4 0.5 Σ 0.4 0.5 0.4 0.3 0.2 0.1 0 -0.1 -0.2 -0.3 -0.4 -0.5 0.3 μ 0.2 0 -6 -4 -2 0 2 4 6 0.1 0 P Model calibration 1.4 1.2 1 2. Choose parameters of AR model to fit shortterm transitional properties and N(0,1) asymptotic distribution 0.8 X 0.8 0.6 0.4 0.2 0.7 1 0 0.6 W -0.2 -0.4 0.5 Σ 0.4 0.5 0.4 0.3 0.2 0.1 0 -0.1 -0.2 -0.3 -0.4 -0.5 0.3 μ 0.2 0 -6 -4 -2 0 2 4 6 0.1 0 P Case study: GB2030 model • • • • 6 years of hourly wind speed data taken from MIDAS dataset by Olmos (2009) 116 sites (onshore only) 10m anemometer data extrapolated to hub-height and converted to wind power using turbine curve Regional weightings chosen to match core 2030 buildout scenario used by Poyry (2009); offshore capacity mapped to nearest onshore regions Olmos Poyry GB2030: modelling strategy • • • • Weighted regional power output aggregated to produce a univariate time series Split into four seasons For each season, calibrate model to reproduce asymptotic distribution, diurnal variation and short-term volatility, using AR(2) model Tweak to approximate effect of offshore component GB2030 (untweaked): distribution and volatility 1800 1200 1000 Sim Lower Hist2001-2 Hist2003-4 Hist2005-6 Hist Av 800 600 400 0.20 RMS change (p.u.) 1400 Sim Upper Sim Mean Hist2002-3 Hist2004-5 Hist2006-7 0.15 Sim Upper Sim Mean Hist2002-3 Hist2004-5 Hist2006-7 0.10 0.05 200 0 Sim Lower Hist2001-2 Hist2003-4 Hist2005-6 HistAv 0.00 0-0.05 0.05-0.1 0.1-0.15 0.15-0.2 0.2-0.25 0.25-0.3 0.3-0.35 0.35-0.4 0.4-0.45 0.45-0.5 0.5-0.55 0.55-0.6 0.6-0.65 0.65-0.7 0.7-0.75 0.75-0.8 0.8-0.85 0.85-0.9 0.9-0.95 0.95-1 Occurrences per year 1600 0.25 0 5 10 15 Time horizon (hr) Power output bucket (p.u.) Power output distribution Volatility curve 20 GB2030 (untweaked): distribution of absolute power output changes 10000 10000 1000 1 hr Occurrences per year 100 10 1 Sim Upper Sim Mean Hist2002-3 Hist2004-5 Hist2006-7 0.1 Sim Lower Hist2001-2 Hist2003-4 Hist2005-6 HistAv 10 Sim Upper Sim Mean Hist2002-3 Hist2004-5 Hist2006-7 1 0.1 0.01 Sim Lower Hist2001-2 Hist2003-4 Hist2005-6 HistAv 0.6-0.65 0.55-0.6 0.5-0.55 0.45-0.5 0.4-0.45 0.35-0.4 0.3-0.35 0.25-0.3 0.2-0.25 0.15-0.2 0.1-0.15 0.05-0.1 0-0.05 0.25-0.3 0.2-0.25 0.15-0.2 0-0.05 0.05-0.1 0.1-0.15 0.01 Power output change bucket (p.u.) Power output change bucket (p.u.) 10000 10000 8 hr 100 10 Sim Upper Sim Mean Hist2002-3 Hist2004-5 Hist2006-7 1 0.1 24 hr 1000 Occurrences per year 1000 Sim Lower Hist2001-2 Hist2003-4 Hist2005-6 HistAv 100 10 1 0.1 0.01 Sim Upper Sim Mean Hist2002-3 Hist2004-5 Hist2006-7 Sim Lower Hist2001-2 Hist2003-4 Hist2005-6 HistAv Power output change bucket (p.u.) 0.75-0.8 0.7-0.75 0.65-0.7 0.6-0.65 0.55-0.6 0.5-0.55 0.45-0.5 0.4-0.45 0.35-0.4 0.3-0.35 0.25-0.3 0.2-0.25 0.15-0.2 0.1-0.15 0.05-0.1 0.01 0-0.05 Occurrences per year 4 hr 100 0-0.05 0.05-0.1 0.1-0.15 0.15-0.2 0.2-0.25 0.25-0.3 0.3-0.35 0.35-0.4 0.4-0.45 0.45-0.5 0.5-0.55 0.55-0.6 0.6-0.65 0.65-0.7 0.7-0.75 0.75-0.8 0.8-0.85 0.85-0.9 0.9-0.95 Occurrences per year 1000 Power output change bucket (p.u.) GB2030: variation of 4hr volatility with power level 0.18 0.16 0.14 1 0.12 W(x) 0.1 0.08 Power output bucket (p.u.) x 0 0.9-1.0 0.8-0.9 -6 0.7-0.8 0.2-0.3 0.1-0.2 0 0.6-0.7 0.02 Sim Lower Hist2001-2 Hist2003-4 Hist2005-6 HistAv 0.4-0.5 0.04 0.3-0.4 Sim Upper Sim Mean Hist2002-3 Hist2004-5 Hist2006-7 0.5-0.6 0.06 0-0.1 Mean absolute change (p.u.) 0.2 -4 -2 0 2 4 6 What about turbine cutout? Denmark, distribution of 4-hour changes (non-rolling window) Sim Upper Sim Mean Hist2004-5 Hist2006-7 Hist2008-9 1000 Sim Lower Hist2003-4 Hist2005-6 Hist2007-8 100 8 Jan 2005 10 1 Power output change bucket (p.u.) 0.65-0.7 0.6-0.65 0.55-0.6 0.5-0.55 0.45-0.5 0.4-0.45 0.35-0.4 0.3-0.35 0.25-0.3 0.2-0.25 0.15-0.2 0.1-0.15 0.05-0.1 0.1 0-0.05 Occurrences per year 10000 GB2030: tweaking strategy (1) Diurnal variation is too great • Lunchtime wind speed peak at hub height is less pronounced than at anemometer height (insolation reduces stability) Mean power ouptut (p.u.) 0.45 0.4 0.35 0.3 0.25 Olmos Summer 0.2 Olmos Winter 0.15 NG Summer 0.1 NG Winter 0.05 0 0 4 8 12 16 20 24 Hour (GMT) • Offshore component has no diurnality => Reduce μ values by a factor of 4 GB2030: tweaking strategy (2) Offshore component increases mean capacity factor (28% -> 33%) => Stretch W function so as to match duration curves shown in Poyry (2009). Use same AR parameters as untweaked model 40000 Year 1 35000 Year 2 Year 3 Wind output (MW) 30000 Year 4 25000 Year 5 Year 6 20000 Year 7 Year 8 15000 10000 5000 0 100% Poyry 2030 data (43GW capacity) 80% 60% 40% 20% Synthetic data from tweaked GB2030 model 0% 800 600 1200 1000 Untweaked Tweaked 400 200 0 Power output bucket (p.u.) Power output distribution RMS power output change (p.u.) 0-0.05 0.05-0.1 0.1-0.15 0.15-0.2 0.2-0.25 0.25-0.3 0.3-0.35 0.35-0.4 0.4-0.45 0.45-0.5 0.5-0.55 0.55-0.6 0.6-0.65 0.65-0.7 0.7-0.75 0.75-0.8 0.8-0.85 0.85-0.9 0.9-0.95 0.95-1 Occurrences / year GB2030: Effect of tweak 1400 0.25 0.2 0.15 Untweaked 0.1 Tweaked 0.05 0 0 4 8 12 Volatility curve 16 Time horizon (hr) 20 24 Wind output (GW) GB2030: Time history sample (“Turing test”) Poyry data 45 Wind output (GW) 40 35 30 25 20 15 10 5 0 01-Dec-08 11-Dec-08 21-Dec-08 31-Dec-08 10-Jan-09 Tweaked GB2030 synthetic winter data 20-Jan-09 30-Jan-09 Conclusions • Non-Gaussian wind power time series can be transformed to a Gaussian (X) domain and modelled with a Gaussian time series model • Synthetic time series reproduce the important long-term and transitional properties (for power system simulation) • Simplicity of model makes it possible to write down formulae for any desired statistic • Transformation to Gaussian domain simplifies modelling of correlated RVs: • Forecast errors (anti-correlated with wind realisation to prevent forecast biasing) • Multi-bus models • Combined demand / wind model References • Sturt, A. and Strbac, G. “Time series modelling of power output for large-scale wind fleets”, Wind Energy, 2011 (to be published) • Enernex Corporation “Eastern Wind Integration and Transmission Study”, 2010 http://www.nrel.gov/wind/systemsintegration/ewits.html • Olmos, P. “Probability distribution of wind power during peak demand”, MSc dissertation, University of Edinburgh, 2009 • Olmos, P.E., Dent, C., Harrison, G.P. and Bialek, J.W. “Realistic calculation of wind generation capacity credits”, CIGRE/IEEE Symposium on integration of wide-scale renewable resources into the power delivery system, Calgary, 2009 • Poyry Energy Consulting, “Impact of intermittency: how wind variability could change the shape of the British and Irish electricity markets: summary report”, 2009 http://www.poyry.com • Sturt, A. and Strbac, G. “A time series model for the aggregate GB wind output circa 2030”, 2011 http://www.ee.ic.ac.uk/%20alexander.sturt07/GB2030SOM.pdf