Current best practice of uncertainty forecast for wind energy Dr. Matthias Lange Stochastic Methods for Management and Valuation of Energy Storage in the Future German Energy System 17 March 2016
Overview Company profile Wind power forecasting Facts on forecasting errors Approaches to estimate uncertainty
Company profile Integration of renewables into grids and markets Service provider for energy meteorology since 2004 Wind and solar power predictions forecasting service worldwide for traders and grid operators Virtual power plants energy trading, remote control of decentralized units, reserve power Software as a Service National and international research projects 60 people
Areas with operational forecasting experience 85 GW of installed wind power currently predicted Previento wind power forecasting for grid operators for traders
How does a wind power prediction look like? Schedule containing 24 or 96 values per day Single plants or aggregate Uncertainty forecast included Delivered prediction Measurement Confidence bands
Previento wind power forecasting system Previento Physical Model: Spatial refinement Thermal stratification Site-specific power curve Forecast uncertainty energy & meteo systems
Input: Wind speed forecast from numerical weather models
Combination: Do not trust in one weather model only!
Types of forecasting error prediction measurement phase error: fronts not in time NWP amplitude error: wind speed at hub height wrongly predicted NWP or wind profile
Types of forecasting error prediction measurement shifted diurnal cycle: thermal stratification wrongly predicted wind profile machine dependent errors: cut-off, curtailment, availability machine data
How accurate are wind power predictions in general? Different metrics Mean absolute error (MAE) Root mean square error (RMSE) Bias Correlation Skill scores Contingency tables CSI (ramps) Different normalizations Installed power Average output Actual forecast value Visusal / subjective Scatter plots Time series plots (ramps)
RMSE forecasting error for real sites (day-ahead) Long-term results from wind farms and aggregations over regions in Europe, North America and Australia
Increase of forecasting error with prediction horizon Previento prediction of total aggregate of German wind farms 2014
Forecast errors in scatter plot underestimated Scatter over 1 year overestimated
Distribution of wind power prediction error (dayahead) Large region installed power: 2250 MW
Distribution of wind power prediction error (dayahead) Small region installed power: 37 MW
Transformation of wind speed errors to power errors power curve
Artificial forecasting errors Wind farm availability below 100 % Wind farm shut down due to grid congestion
Quantile forecasts as uncertainty estimate Statistical training based on historical errors Quantile forecast P85, i.e. with probability of 85% observed value is lower combination prediction Measurement Quantile forecast P15, i.e. with probability of 15% observed value is lower
Now for something completely different: Weather Exploit that atmosphere is chaotic.
Sensitive dependence on initial conditions Lorenz System Very small differences in starting point of simulation.
Sensitive dependence on initial conditions Lorenz System
Solution jumps chaotically between two states Lorenz System
Ensemble uncertainty tries to exploit chaotic behaviour COSMO-DE-EPS Multimodel-Ensemble Not enough spread Spread covers error range
Different NWP models represent different future scenarios combination of NWP
Different kinds of uncertainty estimates Spread of ensemble Quantiles with fixed confidence level Prediction of January 24, 2016, 8 UTC
Different kinds of uncertainty estimates Spread of ensemble Quantiles with fixed confidence level Prediction of January 15, 2016, 8 UTC
Spread of ensemble updates 19/07/2015 to 22/07/2015 Prediction of 20/07/2015, 08 UTC
Spread of ensemble updates 19/07/2015 to 22/07/2015 Spreads are more weather specific! Prediction of 21/07/2015, 08 UTC
Approaches to estimate uncertainty Quantile prediction Based on historical timeseries of forecasting error, i.e. really observed deviations Conditional error distributions can be considered, e.g. foreast horizon or production level Advantage: well-defined confidence levels Disadvantage: only weak dependence on current weather situation Spread of ensemble Based on deviation between several predicted scenarios for the upcoming weather situation Ensemble members represent realistic weather situations with equal probability Better results for intraday and day-ahead if poor mans ensemble is used instead of ensemble model Advantage: clear relation to current weather situations Disadvantage: so far no well-defined confidence levels
Thanks for your attention! www.energymeteo.com Contact: Dr. Matthias Lange matthias.lange@energymeteo.com