Optimal combinaison of CFD modeling and statistical learning for short-term wind power forecasting
2 likes943 views
This document discusses the optimal combination of CFD modeling and statistical learning for short-term wind power forecasting. It emphasizes the importance of integrating physical and statistical models to improve forecast accuracy, particularly in complex terrains, achieving RMSE and MAE reductions. Additionally, it highlights the use of artificial neural networks for error correction based on high-resolution mesoscale and microscale modeling data.
Optimal combinaison of CFD modeling and statistical learning for short-term wind power forecasting
- 1. Optimal combination of CFD modeling and statistical learning for short–term wind power forecasting Stéphane SANQUER & Jérémie JUBAN- meteodyn
- 2. Wind power depends on the volability of the wind Two times scale are relevant : one for the wind turbine controle (up to few sec.), one for the integration of power in the grid (minutes to weeks) Why a forecast ? -Optimise the planning of conventional power plants (3-10h) -Optimise the value of produced electricity in the market (0-48h) -Schedule the maintenance of the farm and the transmission lines (day to week)
- 3. The model can be physical, statistical or twice Very short term statistical approaches use scada as input (look ahead time <6h) A forecast system for horizons >6h always includes a Numerical Weather Prediction system (NWP) and sometimes a Model Output Statistic system to optimize the forecast (MOS) •Source : Anemos Project ” The State-Of-The-Art in Short-Term Prediction of Wind Power”
- 4. Average NMAE for 12 hours forecast horizon vs RIX Source: Best Practice in Short-Term Forecasting. A Users Guide Gregor Giebel(Risø National Laboratory, DTU), George Kariniotakis( Ecole des Mines de Paris) 12 models tested on various terrains to consider the local effects Errors increase with the terrain complexity. Terrain modeling can be introduced to improve the performance of the forecast system.
- 5. To define an optimal combination of both physical and statistical modeling in order to reach the highest forecast performance To use a learning model (« black box ») based on a data set of couples measurement/prediction. Here we use a Artificial Neural Network (ANN) To minimize the prediction error by introducing automatic error corrections while keeping the advantage of the full physical modeling
- 7. Global model 0.125 or 0.5 degrees Resolution GFS, ECM WF Mesoscale model . 1 km to 15 km resolution WRF. Microscale model 25 m resolution Meteodyn WT Statistical modelling DATA
- 8. Mesoscale models compute the wind above the ground with a resolution from 1 km to 5 km. Mesoscale models consider the thermal effects on the boundary layer behaviors. The NWP data defines the stability class at each time step. Mesoscale models can not compute well enough the effects of complex terrains and should be mixed with microscale models. Microscale computations are carried out for various stability classes
- 9. The mesoscale points are transfered to each wind turbine thanks to the « speed coefficients » obtained by the CFD model Local effects taken into account : Orography, Land-use The windspeed coefficients allow the statistical correction of NWP data and power curves correction, by using met mast measurements. Calibration takes into account seasonal variations (snow, foliage density, …)
- 10. Global model 0.125 or 0.5 degrees Resolution GFS, ECM WF Mesoscale model . 1 km to 15 km resolution WRF. Microscale model 25 m resolution Meteodyn WT Statistical modelling DATA
- 11. Global model 0.125 or 0.5 degrees Resolution GFS, ECM WF Mesoscale model . 1 km to 15 km resolution WRF. Microscale model 25 m resolution Meteodyn WT Statistical modelling DATA
- 12. How to define Neural Network Architecture? (number of layers, number of neurons) Increasing complexity Map several inputs to an output Input: Forecast power, NWP variables and production data Output: wind power or wind speed The supervised mapping function is learnt from data
- 13. Define three sets A Testing set choose architecture (testing error) A Training set training the network (training error) Finally, a validation set computes true error. Training Error Testing error Expected minimum error
- 14. Wind Farm in China with a complex terrain and weather regimes Learning period : 06/2010 to 02/2012 Testing period : 03/2012 to 11/2012 Forecast horizons : +6h to 46h Forecast steps : 15 min. Runs :4/day Input variables NWP : V, Dir, S, T, r,Patm Park production
- 15. Mesoscale modeling is used to compute the wind above the site Model GFS/WRF Resolution 5 km
- 16. Wind speed and production are computed by considering all the relevant parameters Orography and roughness of terrains Density of air Power curves Wake effects CFD
- 17. After learning of the ANN model, the production is forecast and compared to the real production Production is globally well forecasted Some time lags are observed
- 18. ANN model reduce forecasting errors of pure physical approach Improvements on MAE and RMSE are respectively 5% and 16% RMSE reduced to 16% bound MAE reduced to 10.5% bound
- 19. RMSE is roughly constant and depends slightly on the look ahead time ANN model benefits are the same in the ranges 6h-30h and 22h-46h 18 19 15 16 0 2 4 6 8 10 12 14 16 18 20 6h-30h 22h-46h RMSE WT WT+ANN
- 20. An optimal combination of statistical and physical modeling is central to high performance forecasting Even for complex terrains, as soon as micro-CFD modeling is performed, Even with weather regime, by coupling NWP with Statistical learning for short term wind power forecasting, RMSE about 15% can be achieved for horizons in the range 6h-48h. MAE reach 10% bound as for flat terrains. Introducing advanced statistical learning leads to significant improvement over a pure (even advanced) physical approach