Country scale solar irradiance forecasting for PV power trading The benefits of the nighttime satellite-based forecast Sylvain Cros, Laurent Huet, Etienne Buessler, Mathieu Turpin
European power exchange (EPEX) Power organised SPOT market Day-Ahead auctions: Electricity traded for delivery the following day at 24-hour time step The daily auction takes place at 12:00 pm, 7 days a week Intra-day trading: Electricity traded for delivery on the same or the following day at 15 min time step Trading is continuous 7 days a week and 24 hours a day (up to 30 min. before physical delivery
Intraday trading Solar energy is a variable source of electricity Intraday power trading leads to fill the gaps at the best price
Unbalances adjustements PV power forecast helps to : Optimize trading prices Avoid costly adjustment
Intraday forecast for EPEX market: a German use-case Every morning at 6 am CET, the customer wants: 25000 Forecast of total Germany PV power production eligible to EPEX 15 min time step until 12am CET German PV production 2016-08-07 (MW) PV capacity map 20000 15000 Actual production 10000 5000 0 02:00 04:00 06:00 08:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00
How to get an intraday country scale PV power forecast? German PV production 2016-08-07 (MW) 30000 25000 Actual production 20000 15000 10000 NWP 5000 0 02:00 04:00 06:00 08:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00
How to get an intraday country scale PV power forecast? German PV production 2016-08-07 (MW) 30000 Actual production 25000 20000 NWP 15000 10000 Satellite 5000 0 02:00 04:00 06:00 08:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 Forecast delivery time
How to get an intraday country scale PV power forecast? German PV production 2016-08-07 (MW) 30000 Actual production 25000 20000 NWP 15000 Satellite 10000 5000 Night Satellite 0 02:00 04:00 06:00 08:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 Forecast delivery time
How to get an intraday country scale PV power forecast? 30000 German PV production 2016-08-07 (MW) Actual production 25000 NWP 20000 Satellite 15000 10000 5000 Night Satellite Machine-learning combination 0 02:00 04:00 06:00 08:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 Objective: quantifiying the benefit of night satellite-based forecast for trading applications
Country forecast using satellite data Extrapolation of cloud index pattern (Cros et al., 2014) from Meteosat-10 satellite Cloud motion vector field on T0 HRV map Forecasted Cloud index maps Up to 6 hours (step 15 min)
Forecasting before the sunrise If the sun is below the horizon : No cloud data available with VIS channel: => forecast is not possible Delivering forecast at 6:00 CET only possible few days in summer
A night cloud index Using Meteosat-10 IR channels (10.8 and 3.6 µm) to synthetize a «Heliosat compatible cloud index» before the sunrise Approach of Hammer et al. (2015): Cloud maintain their distinctive features during few hours between night and day time Statistical relationships between HRV cloud index and brightness temperature related value are established 2016-05-26 00:00 to 10:00 UTC Cloud index 2016-05-26 7th Solar Integration Workshop on integration of Solar Power into Power Systems 24-25 October, Berlin, 00:00 Germany to 10:00 UTC
A night cloud index Statistical relationships between HRV channel cloud index and brightness temperature related value are established From Hammer et al. (2015) Fog and stratus Very cold clouds Obtaining cloud index - not only a cloud mask - without NWP profiles
HRV reflectance correction Our modifications for night-day transition 85 < SZA < 89.8 Cloud index is overestimate when the sun is low An airmass correction is needed for pixel reflectance X = 1/(cos θ + 0.025*exp(-11* cos θ)) When 85 < θ < 89.8 Rozenberg (1966) θ: solar zenithal angle
HRV reflectance correction Cloud index 2016-07-16 4:15 UTC Cloud index 2016-07-16 4:15 UTC Airmass correction from Hammer et al., (2015)
HRV reflectance correction Empirically decreasing highest reflectance values Cloud index 2016-07-16 4:15 UTC New airmass correction more appropriate for reflectance computed from calibrated radiance on Heliosat-2 (Rigollier et al., 2004) based method Cloud index 2016-07-16 4:15 UTC ᑭ = ln (1+ ᑭ/a)*a a = 15/( θ-75 +000.1 ) * 2
Irradiance night satellite forecast assessment Full year 2016 over 23 DWD stations (hourly irradiation, free access)
Relative RMSE (%) Satellite-based forecast assessment Night: θ > 89.8 ; Day θ < 89.8 70 Relative RMSE (%) 60 50 40 30 20 All Night Day 10 0 1 2 3 4 5 6 Time horizon (h)
Satellite-based forecast assessment Night: θ > 89.8 ; Day: θ < 89.8 Night Day All Time horizon: 3h
Country forecast Statistical approach NWP (ECMWF, ICON, GFS) Satellite (Hourcast Reuniwatt) Night Satellite (Hourcast Reuniwatt) Random forest Algorithm Training Country power forecasts PV power capacity map Seasonnality Historical PV production Measurements (EPEX SPOT archives)
Country Forecasting Statistical Approach Typical successful case: night satellite forecast «warns» NWP that Germany is more cloudy than expected this early morning Random forest algorithm behaves according to its training and to the symmetry of the daily production profile 30000 25000 20000 15000 NWP Night Satellite Machine-learning combination 10000 5000 0 02:00 04:00 06:00 08:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00
Impact evaluation of satellite forecasts Producing day and night satellite forecast for years 2015 and 2016 Training random forest algorithm over 2015 and testing over 2016 without satellite data with daytime satellite data with day and night satellite data Comparison between forecasts and PV power on a daily basis (including night value). Installed capacity: 39.5 GW Tests RMSE (MW) rrmse/ Pinst (%) MAE (MW) rmae Pinst (%) No sat 1997 5.0 1975 3.5 Day sat 1540 3.9 1264 3.2 D&N Sat 1066 2,7 987 2.5
Impact evaluation of satellite forecasts MAE in function of time horizon 8 rmae / Pinst (%) 7 6 5 4 3 2 1 0 NWP NWP + Day Sat NWP + D&N Sat 0 5 10 15 20 25 Time horizon (h)
Conclusion A country PV power forecast method has been presented and evaluated A satellite-based algorithm delivering forecast before sunrise has been presented and assessed The case-study clearly showed the benefits of satellite imagery for PV power spot market trading Further progress can be undertaken: Improving the cloud index mapping during night/day transition Using surface temperature from NWP for a better detection of low cloud Progress margin exists in random forest algorithm setting
References Cros S., Sébastien N., Liandrat O., Schmutz N., Cloud pattern prediction from geostationary meteorological satellite images for solar energy forecasting, SPIE Remote Sensing Conference, September 22-25 2014, Amsterdam, The Netherlands. Proc. SPIE 9242, Remote Sensing of Clouds and the Atmosphere XIX; and Optics in Atmospheric Propagation and Adaptive Systems XVII, 924202 (21 October 2014) Hammer, A., Kühnert, J., Weinreich, K., & Lorenz, E. (2015). Short-term forecasting of surface solar irradiance based on Meteosat-SEVIRI data using a nighttime cloud index. Remote Sensing, 7(7), 9070-9090. Rigollier, C., Lefèvre, M., & Wald, L. (2004). The method Heliosat-2 for deriving shortwave solar radiation from satellite images. Solar Energy, 77(2), 159-169. Rozenberg, G.V. Twilight: A Study in Atmospheric Optics; Plenum Press: New York, NY, USA,1966
Find more information on our website www.reuniwatt.com Thank you! sylvain.cros@reuniwatt.com