1 Powerpoint Templates Day ahead and intra
Transcription
1 Powerpoint Templates Day ahead and intra
CM2E – 4-9 May 2014 Day ahead and intra-day solar forecasting applied to an insular site ! ! M. DAVID, P. LAURET, M.H. DIAGNE, P.J. TROMBE ! PIMENT, Université de La Réunion Pour plus de modèles Colloque : Modèles Powerpoint PPT gratuits Martinique Energie Environnement 4-9 May 2014 Powerpoint Templates 1 Introduction CM2E – 4-9 May 2014 Reunion Island: oSmall tropical Island situated in the Indian Ocean oNon interconnected grid oHigh penetration rate of PV (>25% of total installed power in 2013) oPV output power = 30% of produced power at least 1 time a month (legal limit) oHigh solar fluctuations can affect the grid balance oNeeds in solar forecasting Pour plus de modèles : Modèles Powerpoint PPT gratuits 2 Framework Day ahead forecasts: oScheduling of resources and commitment of units of production oHourly granularity oNWP (ECMWF) and post-processing techniques CM2E – 4-9 May 2014 ! Intra day forecasts: oTo monitor the production and to adjust the scheduling oHourly granularity oTimes series models and machine learning methods ! Measured Data oHourly Global Solar Irradiance (GHI) o1 site -> Saint-Pierre (coastal part with high solar potential) o2 years without missing data: ▪ Year 2012 = training ▪ Year 2013 = validation oFiltering of data: solar zenith angle < 85° Pour plus de modèles : Modèles Powerpoint PPT gratuits Saint-Pierre 3 Day ahead forecasts Initial ECMWF GHI forecast oGenerated at 12h00 UTC (16h00 in Reunion) oSpatial resolution: 0.125° x 0.125° (≈ 14 x 14 km) oTemporal resolution: 1 hour oNo post-processing CM2E – 4-9 May 2014 ! ECMWF + MOS post-processing oApplication of MOS proposed by Lorenz et al. oDesigned in 3 steps : • Spatial aggregation and temporal interpolation • Low Total Cloud Cover = Clear Sky • Bias removal with a polynomial function of fourth order in the clear sky index (kt*) and the cosine of the solar zenith angle (θz) oCalibrated for local conditions ! ECMWF + Neural Network post-processing oNN used to model the bias of the NWP forecast o Inputs: forecasted clear sky index (kt*) and cosine of the solar zenith angle (θz) o Powerpoint complexity controlled with Bayesian techniques Pour plus de modèles : Modèles NN PPT gratuits 4 Day ahead forecasts Focus on the MOS post-processing CM2E – 4-9 May 2014 oNo needs in temporal interpolation oSpatial aggregation does not improve the forecast accuracy oLonger is the period used for the calibration, lower is the error ! ! ! Forecast error over spatial averaging of Day-ahead forecast error over the size of the ECMWFPour day-ahead forecast around Saintsliding window for the calibration of the plus de modèles : Modèles Powerpoint PPT gratuits Pierre polynomial function of fourth order 5 Day ahead forecasts Relative errors of day-ahead forecast CM2E – 4-9 May 2014 (mean GHI = 498.2 W.m-2) Pour plus de modèles : Modèles Powerpoint PPT gratuits 6 Intra day forecasts 1. Single input: measured kt* ! Reference models o Persistence o Climatology (average kt* of year 2012) CM2E – 4-9 May 2014 ! Time series model o ARMA recursive least square (ARMArls) ! ! ! Machine learning techniques o Neural Network (NN) with Bayesian optimization o Support Vector Machine (SVM) o Gaussian Process (GP) 2. Hybrid inputs: measured and forecasted (ECMWF) kt* ! Time series model 1. Kalman filter (Kalman+ECMWF) ! ! Machine learning technique o Neural Network (NN+ECMWF) Pour plus de modèles : Modèles Powerpoint PPT gratuits 7 CM2E – 4-9 May 2014 Intra day forecasts Pour plus de modèles : Modèles Powerpoint PPT gratuits 8 CM2E – 4-9 May 2014 Intra day forecasts Pour plus de modèles : Modèles Powerpoint PPT gratuits 9 CM2E – 4-9 May 2014 Intra day forecasts Pour plus de modèles : Modèles Powerpoint PPT gratuits 10 CM2E – 4-9 May 2014 Intra day forecasts ECMWF Pour plus de modèles : Modèles Powerpoint PPT gratuits 11 Perspectives 1. Test of the models for other sites in Reunion and for other islands (Corsica, Guadeloupe, Martinique ?, Hawaii) ! CM2E – 4-9 May 2014 2. Intra day forecasts: 10 minutes granularity ! 3. Implementation of ECMWF forecasts in an ARMAX Recursive Least Square model ! 4. Use of exogenous data (cloud height, speed and direction) as inputs of the learning machine techniques ! 5. Envelope of uncertainties Acknowledgment The authors would like to thank the European Centre for Medium-Range Weather Forecasts (ECMWF) for providing forecast data. Pour plus de modèles : Modèles Powerpoint PPT gratuits A part of the work was accomplished in the preliminary phase of the SOLFIN project funded by ADEME. 12 CM2E – 4-9 May 2014 Thank you ! Pour plus de modèles : Modèles Powerpoint PPT gratuits ECMWF grid 13