Applicability of wind power forecasting models in Japan

Transcription

1 Applicability of wind power forecasting models in Japan Atsushi Yamaguchi Takeshi Ishihara Torben Skov Nielsen Technical University of Denmark

2 The outline of the project A national R&D project, Development of Wind Power Prediction Models Based on Numerical Weather Prediction has been carried out since October 25 until March 28. The object of the project is to develop a practical wind power prediction system for power system operation. The Project is funded by New Energy and Industrial Technology Development Organization (NEDO) and cooperated by five organizations. (UT) Itochu Techno-Solutions (CTC) Central Research Institute for Electric Power Industry (CRIEPI) Japan Weather Association (JWA) E & E Solutions (EES)

3 The concept of the project To meet any type of requirement in the future, two types of model will be developed. 数値気象予報データ NWP 単機出力 / ナセル風速データ SCADA a c b d ファーム Ⅲ WF 総出力データ SCADA ファームⅢ ファームⅠ ファームⅣ ファームⅡ 数値気象予報データ NWP 単機出力予測値 a b c d WF 総出力予測値ファームⅠ ファームⅡ ファームⅢ ファームⅣ ファーム Ⅲ WF 総出力予測値 (a+b+c+d) エリア総出力予測値 (Ⅰ+Ⅱ+Ⅲ+Ⅳ) WF model Area model Based on detailed SCADA information. Based on limited SCADA information For wind farm owners. For grid operators.

4 The contents of the project Measurement (both on-line and off-line) EES Development of wind power prediction model CRIEPI + JWA (WF model), UT + CTC (Area model) Verification and error analysis of the model CRIEPI (WF model), UT (Area model) On line implementation of the model JWA (WF model), CTC (Area model) Development of the platform for wind power prediction model CTC, UT Guideline for wind power prediction CRIEPI, UT

5 The topic of this presentation As a part of Development of wind power prediction model, benchmark testing for existing area models were carried out by the University of Tokyo and CTC. The result of the benchmark testing carried out by University of Tokyo will be presented.

6 Wind Farms WF AA, サイト B B サイト and C サイト C WF E E サイト WF H サイト H WF I I サイト The benchmark test was carried out for the sum of the output from six large wind farms (total capacity 167MW) in Tohoku area, northern part of Japan.

7 Day ahead forecast and hourly forecast Day Day ahead ahead forecast Forecast time Forecast horizon : 6: 12: : 12: : Day ahead forecast is carried out once a day at 6: a.m. for the next day Main application: unit commitment Hourly forecast Forecast time Forecast horizon : : 12: : 12: Hourly forecast is carried out every 3minutes for next 24 hours. Main application: load following

8 The conditions of the benchmark test Test models WF Period Geographic information Input data Evaluation data Training data Forecast value Forecast time Forecast horizon University of Tokyo Model, WPPT 6 wind farms in Tohoku area (Total capacity 167 MW) Feb., Apr., Jul. and Oct. 25. (Four months) location of the WF, hub height, elevation, surface roughness NWP(GPV-RSM, 2km resolution.jma), SCADA data from each WF Sum of the SCADA data from six WFs All the data in 24 except for Feb., Apr., Jul and Oct. 3 minutes averaged power from each wind farm and the area 6: a.m. (day ahead forecast) Every 3 minutes (hourly forecast) : to 24: next day 24 hours from the

9 University of Tokyo Model NWP data Geographic imformations SCADA data Mesoscale model RAMS is is used to to downscale the the 2km resolution NWP to to 1km 1km Mesoscale model RAMS WF power curve model P pc t k f u pred t k Power prediction model P a k P b k P pred meas pc t k t t k Predicted power

10 WPPT NWP data SCADA data WF WF power curve and and the the parameters for for power prediction model is is estimated using adoptive method WF power curve model Power prediction model Predicted power P f u,, k pc pred pred t k t kt t kt,, P a k P b k P pp pred meas pred pc t k t kt t t kt t k 2 ht k 2 h c t kt, k cos c t kt, k sin c pred s pred t k WF WF power curve and and the the parameters are are the the function of of the the wind wind direction

11 Summery of the models Spatial resolution of NWP data Adoptive approach for the parameter identification Training period UT model 1km Not used All the data in 25 except for Feb., Apr., Jul. and Oct. WPPT 2km Used All the data in 25 before the forecast time.

12 Day ahead forecast Meas UT WPPT #VALUE! July 25 Power / Rated Power (%) /1 7/2 7/3 7/4 7/5 7/6 7/7 7/8 7/9 7/1 7/11 7/12 7/13 7/14 7/15 Time UT model seems to simulate extreme phenomena better This is because of the resolution

13 Error Analysis In order to show the performance of the models, following measures were used. ME n i 1 P pred i n P meas i Mean Error (bias) RMSE n pred meas P 2 i Pi i 1 n Root Mean Square Error ME and RMSE is calculated as a non dimensional value (%) relative to the total nominal power.

14 RMSE for day ahead forecast 3 3 UT WPPT #VALUE! RMSE(%) 2 RMSE (%) WPPT UT Time time of the day UT model gave better results for all the forecast horizon of the day ahead forecast

15 Summery of the results 2 15 RMSE UT WPPT 1 5 ME UT WPPT Feb. Apr. Jul. Oct. Month -1 Feb. Apr. Jul. Oct. Month For most of the months, UT model shows better performance. For UT model, forecast bias (ME) is not small. * Forecast for February by WPPT may not be appropriate for benchmark testing because of the short time for training.

16 Hourly forecast (6 hour ahead) 1 9 Meas UT WPPT July 25 8 Power / Rated Power (%) /1 7/2 7/3 7/4 7/5 7/6 7/7 7/8 7/9 7/1 7/11 7/12 7/13 7/14 7/15 Time No significant difference can be found. WPPT seems to show slightly better result especially near the sharp peak.

17 RMSE for hourly forecast 3 3 UT WPPT #VALUE! RMSE(%) 2 RMSE (%) WPPT UT Lead Time (h) Forecast horizon (Hours) For the time horizon shorter than 6 hours, WPPT model shows better performance because of the dynamical statistical model

18 Summery (6 hours ahead) 2 15 RMSE UT WPPT 1 5 ME UT WPPT Feb. Apr. Jul. Oct. Month -1 Feb. Apr. Jul. Oct. Month For 6 hour ahead forecast, WPPT shows better results especially the mean error (bias). * Forecast for February by WPPT may not be appropriate for benchmark testing because of the short time for training.

19 High resolution NWP data with WPPT 3 UT WPPT WPPT( 高解像度 ) #VALUE! RMSE (%) 2 WPPT 1 WPPT with high resolution NWP UT Lead Time (h) When high resolution NWP data (1km resolution) was used as an input to the WPPT, it gave similar results upto 15 hours

20 Conclusions For day ahead forecast, prediction error of UT model is smaller than that of WPPT. This is due to the difference in the spatial resolution of the NWP data used. For hourly forecast, prediction error of WPPT is smaller than that of UT model. This is because WPPT uses adoptive approach for the estimation of parameter in the statistical model.