Modelling wind power in unit commitment models Grid integration session IEA Wind Task 25 Methodologies to estimate wind power impacts to power systems Juha Kiviluoma, Hannele Holttinen, VTT Technical Research Centre of Finland Mark O Malley, Electricity Research Centre, UCD Aidan Tuohy, Electric Power Research Institute Peter Meibom, Risø DTU Erik Ela, Michael Milligan, NREL Madeleine Gibescu, TU Delft Bernhard Lange, Fraunhofer IWES Rüdiger Barth, IER OUTLINE Unit commitment (UC) without wind Variability and wind power forecasts Securing enough available capacity Changes in UC models Illustrative results
UNIT COMMITMENT (UC) Selection of online power plants Enough online capacity to operate system with high level of security Least cost In a market based system takes place through spot market bids Important due to long start-up times, high start-up costs and restrictions on plant cycling UC & SYSTEM OPERATION WITHOUT WIND 14 Forecast 99% Load 12 Forecast 1% UC plans the operation for the next day Reserves handle load forecast errors and plant outages Load (GW) 10 8 6 4 seconds to minutes 2 Sec. freq. control 0 tens of minutes to hours Ter. freq. control day UC and dispatch 0 4 8 12 16 20 24 28 32 36 Time (hours)
UC & SYSTEM OPERATION WITH WIND Expected residual load (load minus wind power production) Variability and uncertainty increases More flexibility required Generation Responsive load Possible curtailment of wind power Net load (GW) 14 12 10 8 6 4 2 0 Original load Forecast 99% Net load Forecast 1% Wind 0 4 8 12 16 20 24 28 32 36 Time (hours) WIND POWER FORECASTS Wind power forecasts are the schedules of wind power plants But: errors are unavoidable and depend on: Forecast horizon NWP model(s) and online measurements (power, wind) available Installed capacity and size of area Forecast system 12 GW installed wind capacity Gibescu et al 2009, Case Study for the Integration of 12 GW Wind Power in the Dutch Power System by 2020. CIGRE/IEEE PES Joint Symposium on the Integration of Wide-Scale Renewable Resources into the Power Delivery System, Calgary, Canada.
WIND INDUCED NEED FOR AVAILABLE CAPACITY Possibility of large forecast errors has to be considered Wind forecast errors are not normally distributed, large errors occur much more frequently This needs to be considered in unit commitment models Probability density 0,1 0,01 1E-3 Actual forecast errors Fitted Gaussian distribution Lange et al (2006), Strategies for Balancing Wind Power in Germany, German Wind Energy Conference DEWEK 2006. 1E-4-8000 -6000-4000 -2000 0 2000 4000 6000 8000 Wind power forecast error [MW] WIND POWER FORECASTS How often new forecasts should/could be available? How to present the uncertainty in the forecasts in UC models? 100 100 80 Probability Steepness and distribution for size production of ramps 80 Timing of ramp events 60 60 40 40 20 20 12 13 14 15 16 17 18 19 20 21 22 23 Hour of the day 12 13 14 15 16 17 18 19 20 21 22 23 Hour of the day
WAYS TO HANDLE THE INCREASED VARIABILITY AND UNCERTAINTY FROM THE UC PERSPECTIVE Incorporate several wind power forecasts in: UC process (e.g. pre-processing stochastic information, stochastic scenarios) Dynamic reserve requirements Use updated wind power forecasts: Intra-day rescheduling of UC realised by liquid intra-day markets Allow for curtailment of wind power when it is the most cost effective solution Perfect forecast Perfect information about wind power production and electricity demand in the day-ahead time scale. Can be used as a base case for estimating the cost of uncertainty. Single forecast with a static reserve Single forecast with a dynamic reserve A rule based adder for reserves is used to account for uncertainty. The UC model is deterministic with one forecast, but the data about wind and demand uncertainty has been preprocessed to create a reserve requirement that is based on probabilistics. UConly in day-ahead With significant wind power penetration a large tertiary reserve would be required. Intra-day UC Decreases the required reserve adder and the amount of wind curtailment. UC only in day-ahead With significant wind power penetration a large tertiary reserve would be required. Intra-day UC Statistical rules Pre-processingis based on statistically derived rules. (LOLP) approaches Convolutionof various probabilities to arrive at the desired LOLP. 10 Stochastic scenarios Optimization of UC decisions over several scenarios for possible outcomes of wind and demand. Scenarios cover part or all of the tertiary reserve requirement.
Advanced Unit Commitment in the Eastern Interconnect Project Team NREL Project Management, data and modeling for the US Risoe DTU developer of the Scheduling Model Univ Stuttgart IER developer of the Scenario Tree Tool ECAR Assistance on analysis and model runs Objectives Understand the impacts of forecast error on the eastern interconnect Understand the impacts and benefits of stochastic planning on the eastern interconnection Understand the impacts and benefits of rolling UC updates on the eastern Interconnection Scope Uses the WILMAR tool Uses consistent data with Eastern Wind Integration and Transmission Study Three data years, 4 unit commitment scenarios, sensitivity requiring coal as must run Study System EWITS Scenario 2: Hybrid with Offshore Region Onshore (MW) Offshore (MW) Total (MW) Annual Energy (TWh) ISO-NE 8,837 5,000 13,837 46 MISO+MAPP 69,444 0 69,444 288 NYISO 13,887 2,620 16,507 48 PJM 28,192 5,000 33,192 97 SERC 1,009 4,000 5,009 16 SPP 86,666 0 86,666 245 TVA 1,247 0 1,247 4 Total 209,282 16,620 225,902 745
The WILMAR tool Improve decision making by using information contained in wind power production and load forecasts Information: Expected wind power production and load, but also precision of forecast, i.e. the distribution of the wind power production forecast errors Information accuracy improves as you get closer to real-time Stochastic Planning: Planning with representation of stochastic variables to provide for a more robust system Rolling Planning: Using updated information to adjust unit commitment decisions more frequently. The WILMAR tool How: Build system-wide stochastic optimisation model with the wind power production and load as a stochastic input parameter Covering both day-ahead scheduling and rescheduling due to updated wind power and load forecasts Consequence: Model makes unit commitment and dispatch decisions being robust towards wind power production and load forecast errors
Scenario Tree Tool Generation of scenario trees containing stochastic input parameter for Scheduling Model: Wind power forecasts Load forecasts Demand for replacement reserves Generation of Semi-Markov processes describing availability / unavailability of power plants Implemented in Matlab Replacement reserve (positive reserves for different forecast horizons): Demand time and scenario dependant (determined for forecast horizons from 1 hour to 36 hours ahead) Demand dependant on wind power and load forecasts Offline units can provide this type of reserve if they have start-up times of less than an hour Replacement Reserves Replacement reserves [MW] 8000 7000 6000 5000 4000 3000 2000 Reserve increases with as uncertainty increases ISO_NE MAPP MISO_Central MISO_East MISO_West NY_ISO PJM SERC SPP_Central SPP_North TVA 1000 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 Forecast hour [h]
Example Wind Forecasts 30000.0 Wind power production [MW] 25000.0 20000.0 15000.0 10000.0 5000.0 0.0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 Forecast 31 33hour 35[h] Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Expected Value Realized Value Example day-ahead scenario tree for PJM net load Residual load (Load minus Wind) [MW] 120000 100000 80000 60000 40000 20000 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 Forecast 31 33hour 35[h] Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Expected Value Realised Value 18
Scheduling model First three hours in scenario tree deterministic: Realized wind power production Realized load Demand replacement reserve (taking uncertainty in wind power production forecasts and load forecasts for forecast horizons 1 to 3 hours ahead into account) Optimization over all outcomes represented by the scenario tree taking both demands for electricity and demand for spinning and replacement reserves into account Minimization of expected costs. Expectation taken over branches in scenario tree Unit restrictions: minimum up time, minimum down time, start-up time, minimum stable operation level, piece-wise linear fuel consumption curve, restriction on ability to provide spinning reserve, ramp up rates, ramp down rates, etc. Implemented in GAMS using CPLEX solver 19 Representation of transmission grid Model area divided into EWITS regions Regions connected by transmission lines No grid representation within a region (only average transmission and distribution loss) Energy exchange between regions. Grid restrictions expressed by usage of NTCs (net transfer capacities) Monthly NTCs obtained from Promod load flow calculations 20
Results Model runs: STO: Stochastic planning using scenario trees with six branches (six forecasts), unit commitment updated every 3 hours UCDay: Stochastic planning, unit commitment for units with start times greater than 1 hour, updated once per day in the day-ahead market OTS: Deterministic planning with forecast error (only one forecast), unit commitment updated every three hours PFC: Deterministic planning with perfect foresight i.e. wind power and load forecasts corresponds to realized wind power and load. Comparisons of these model runs gives insight into the differences of forecast error, stochastic planning, and rolling UC updates 21 Operational costs (mainly fuel costs) Value intra-day rescheduling: 778 M$ (1.3%) (UCDay minus STO) Value stochastic UC: 207 M$ (0.36%) (OTS minus STO) Value perfect foresight: 426 M$ (0.7%) (STO minus PFC) It appears rolling unit commitment gives more value than stochastic planning ) $ (M 59,400 59,200 59,000 58,800 sts 58,600 o C 58,400 n ctio 58,200 u d ro P 58,000 57,800 57,600 57,400 OTS PFC STO UCDay