Upload
sandia-national-laboratories-energy-climate-renewables
View
383
Download
0
Embed Size (px)
Citation preview
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany 1 1
Local and regional PV power forecasting based on
PV measurements, satellite data and numerical weather predictions
Elke Lorenz, Jan Kühnert, Björn Wolff,
Annette Hammer, Detlev Heinemann
1)Energy Meteorology Group, Institute of Physics, University of Oldenburg
1
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany 2
Outline
grid integration of PV power in Germany
overview on PV power prediction system
evaluation:
different data and models for different forecast horizons
combination of different models
summary and outlook
2
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Grid integration of PV in Germany
installed Power: 38.5GWpeak (end of 2014)
3
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Grid integration of PV in Germany
installed Power: 38.5GWpeak (end of 2014)
marketing of PV power at the European Energy Exchange
3
control areas
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Grid integration of PV in Germany
installed Power: 38.5GWpeak (end of 2014)
marketing of PV power at the European Energy Exchange
by transmission system operators regional forecasts
3
control areas
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Grid integration of PV in Germany
installed Power: 38.5GWpeak (end of 2014)
marketing of PV power at the European Energy Exchange
by transmission system operators regional forecasts
direct marketing local forecasts
3
control areas
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Grid integration of PV in Germany
installed Power: 38.5GWpeak (end of 2014)
marketing of PV power at the European Energy Exchange
by transmission system operators regional forecasts
direct marketing local forecasts
day ahead: for the next day
3
control areas
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Grid integration of PV in Germany
installed Power: 38.5GWpeak (end of 2014)
marketing of PV power at the European Energy Exchange
by transmission system operators regional forecasts
direct marketing local forecasts
day ahead: for the next day
intraday: until 45 minutes before delivery
3
control areas
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Grid integration of PV in Germany
installed Power: 38.5GWpeak (end of 2014)
marketing of PV power at the European Energy Exchange
by transmission system operators regional forecasts
direct marketing local forecasts
day ahead: for the next day
intraday: until 45 minutes before delivery
here: 15 min to 5 hours ahead
3
control areas
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
10
PV power measurement
hours forecast horizon
Overview of forecasting scheme
4
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
11
PV power measurement
satellite cloud motion forecast CMV
hours forecast horizon
Overview of forecasting scheme
4
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
12
PV power measurement
satellite cloud motion forecast CMV
days hours
NWP: numerical weather prediction
forecast horizon
Overview of forecasting scheme
4
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
13
PV power predictions
PV power measurement
satellite cloud motion forecast CMV
days hours
PV power forecasting: PV simulation and statistical models
NWP: numerical weather prediction
forecast horizon
Overview of forecasting scheme
4
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Numerical weather predictions
global model forecast (IFS) of the European Centre for Medium-Range Weather Forecasts (ECWMF)
regional model forecasts (COSMO EU) of the German Meteorological service (DWD)
COSMO EU, dir. irradiance 2104-05-02, 12:00
5
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Numerical weather predictions
global model forecast (IFS) of the European Centre for Medium-Range Weather Forecasts (ECWMF)
regional model forecasts (COSMO EU) of the German Meteorological service (DWD)
Post processing:
bias correction and combination with linear regression
COSMO EU, dir. irradiance 2104-05-02, 12:00
5
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Meteosat Second Generation (high resolution visible range)
cloud index from Meteosat images with Heliosat method* resolution in Germany
1.2km x 2.2 km 15 minutes
Irradiance prediction based on satellite data
6
*Hammer et al 2003
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
cloud index from Meteosat images with Heliosat method
cloud motion vectors by identification of matching cloud structures in consecutive images
extrapolation of cloud motion to predict future cloud index
Irradiance prediction based on satellite data
Meteosat Second Generation (high resolution visible range)
6
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
satellite derived irradiance maps
cloud index from Meteosat images with Heliosat method
cloud motion vectors by identification of matching cloud structures in consecutive images
extrapolation of cloud motion to predict future cloud index
irradiance from predicted cloud index images with Heliosat method
Irradiance prediction based on satellite data
200W/m2 900W/m2
6
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Measurement data
March- November 2013
15 minute values
921 PV systems1) in Germany
information on PV system tilt and orientation
19
1)Monitoring data base of Meteocontrol GmbH
7
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
20
PV power predictions
PV power measurement
satellite cloud motion forecast CMV
days hours
PV power forecasting: PV simulation and statistical models
NWP: numerical weather prediction
forecast horizon
Overview of forecasting scheme
8
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
21
PV power predictions
PV power measurement
satellite cloud motion forecast CMV
days hours
NWP: numerical weather prediction
forecast horizon
Overview of forecasting scheme
8
Pmeas(t-Dt)
Pclear(t-Dt)
persistence: constant ratio of measured PV power Pmeas to clear sky PV power Pclear
Ppers (t)= Pclear(t)
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Different input data and models
22
PV power predictions
PV power measurement
satellite cloud motion forecast CMV
days hours
NWP: numerical weather prediction
forecast horizon
persistence
8
PV simulation:
Diffuse fraction model: Skartveith et al, 1998
Tilt model: Klucher 1970
Parametric model for MPP efficiency: Beyer et al 2004
Linear regression
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
PV power predictions
PV power measurement
satellite cloud motion forecast CMV
days hours
PV simulation
NWP: numerical weather prediction
forecast horizon
Different input data and models
PV simulation
persistence
8
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Regional forecasts: persistence, CMV and NWP based forecasts
9
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany 9
Regional forecasts: persistence, CMV and NWP based forecasts
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany 9
Regional forecasts: persistence, CMV and NWP based forecasts
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany 9
Regional forecasts: persistence, CMV and NWP based forecasts
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany 9
Regional forecasts: persistence, CMV and NWP based forecasts
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany 9
Regional forecasts: persistence, CMV and NWP based forecasts
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Rmse in dependence of forecast horizon
15 minute values
normalization to installed power Pinst only daylight values, calculation time of CMV: sunel > 10° only hours with all models available included in dependence of forecast horizon
N
iinst
pred
inst
meas
P
P
P
P
Nrmse
1
2
1
10
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Rmse in dependence of forecast horizon
forecasts for German average
CMV forecasts better than NWP based forecast up to 4 hours ahead
10
N
iinst
pred
inst
meas
P
P
P
P
Nrmse
1
2
1
15 minute values normalization to installed power Pinst only daylight values, calculation time of CMV: sunel > 10° only hours with all models available included in dependence of forecast horizon
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Rmse in dependence of forecast horizon
forecasts for German average
CMV forecasts better than NWP based forecast up to 4 hours ahead
persistence better than CMV forecasts up to 1.5 hour ahead
10
N
iinst
pred
inst
meas
P
P
P
P
Nrmse
1
2
1
15 minute values normalization to installed power Pinst only daylight values, calculation time of CMV: sunel > 10° only hours with all models available included in dependence of forecast horizon
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Rmse in dependence of forecast horizon
11
comparison of German average and single site forecasts:
rmse for German about 1/3 of single sites rmse for NWP forecasts
German average single sites
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Rmse in dependence of forecast horizon
11
comparison of German average and single site forecasts:
rmse for German about 1/3 of single sites rmse for NWP forecasts
improvements with persistence and CMV larger for regional forecasts
German average single sites
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
35
PV power predictions
PV power measurement
satellite cloud motion forecast CMV
days hours
PV simulation*
NWP: numerical weather prediction
forecast horizon
Different input data and models
PV simulation*
persistence
12
*)PV simulation with bias correction
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
36
PV power predictions
PV power measurement
satellite cloud motion forecast CMV
days hours
PV simulation
NWP: numerical weather prediction
forecast horizon
Combination of different models
PV simulation
persistence
combination
12
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
13
Combination of forecasting methods
combination of forecast models with linear regression:
Pcombi=aNWPPNWP + aCMVPCMV + apersistPpersist + a0
coefficients aNWP, aCMV, apersist, a0 are fitted to measured data
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
13
Combination of forecasting methods
combination of forecast models with linear regression:
Pcombi=aNWPPNWP + aCMVPCMV + apersistPpersist + a0
coefficients aNWP, aCMV, apersist, a0 are fitted to measured data
in dependence of
forecast horizon
hour of the day
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
13
Combination of forecasting methods
combination of forecast models with linear regression:
Pcombi=aNWPPNWP + aCMVPCMV + apersistPpersist + a0
coefficients aNWP, aCMV, apersist, a0 are fitted to measured data
in dependence of
forecast horizon
hour of the day
training data:
for single site forecasts: each PV system separately
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
13
Combination of forecasting methods
combination of forecast models with linear regression:
Pcombi=aNWPPNWP + aCMVPCMV + apersistPpersist + a0
coefficients aNWP, aCMV, apersist, a0 are fitted to measured data
in dependence of
forecast horizon
hour of the day
training data:
for single site forecasts: each PV system separately
for regional forecasts: average of sites
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
14
Combination of forecasting methods
How many days to train forecast combination?
Improvement score:
with respect to best single model
𝑟𝑚𝑠𝑒𝑟𝑒𝑓 − 𝑟𝑚𝑠𝑒𝑐𝑜𝑚𝑏𝑖
𝑟𝑚𝑠𝑒𝑟𝑒𝑓
all sites average, May to November, 2012
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
14
Combination of forecasting methods
How many days to train forecast combination?
Improvement score:
with respect to best single model
𝑟𝑚𝑠𝑒𝑟𝑒𝑓 − 𝑟𝑚𝑠𝑒𝑐𝑜𝑚𝑏𝑖
𝑟𝑚𝑠𝑒𝑟𝑒𝑓
all sites average, May to November, 2012 independent test year for model configuration
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
14
Combination of forecasting methods
How many days to train forecast combination?
Improvement score:
with respect to best single model
last 30 days
𝑟𝑚𝑠𝑒𝑟𝑒𝑓 − 𝑟𝑚𝑠𝑒𝑐𝑜𝑚𝑏𝑖
𝑟𝑚𝑠𝑒𝑟𝑒𝑓
all sites average, May to November, 2012
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Regional forecasts
regression coefficients in dependence of forecast horizon
15
bars:
standard deviation for all hours
regression coefficients (weights) reflect horizon dependent forecast performance of different models
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Regional forecasts
Rmse in dependence of forecast horizons
15
Considerable improvement with combined model over single model forecasts
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
16
Single site forecasts
regression coefficients in dependence of forecast horizon
bars:
standard deviation
for all hours & sites
regression coefficients (weights) reflect horizon dependent forecast performance of different models for single sites
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Regression coefficients
16
bars:
standard deviation
for all hours & sites
horizon dependent regression coefficients different for regional and single site forecasts
German average single sites
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Rmse in dependence of forecast horizon
forecast combination outperforms single model forecasts for all horizons
improvements with combination larger for regional forecasts
17
German average single sites
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
49
Summary
PV power prediction contributes to successful grid integration of more than 38 GWpeak PV power in Germany
18
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
50
Summary
PV power prediction contributes to successful grid integration of more than 38 GWpeak PV power in Germany
PV power forecasts based on satellite data (CMV) significantly better than NWP based forecasts up to 4 hours ahead
18
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
51
Summary
PV power prediction contributes to successful grid integration of more than 38 GWpeak PV power in Germany
PV power forecasts based on satellite data (CMV) significantly better than NWP based forecasts up to 4 hours ahead
significant improvement by combining different forecast models with PV power measurements, in particular for regional forecasts
18
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
52
Outlook
combination of machine learning with PV simulation for integration of additional data sources:
additional meteorological parameters
additional NWP systems
uncertainty information
19
October 22nd 2015, 4th PV Modelling Workshop, Köln, Germany
Thank you for your attention!
EU-FP7 grant agreement no: 308991
20