2nd Int'l Conference on Large-Scale Grid Integration of Renewable Energy in India| New Delhi, India | 4-6 Sep 2019

Improving Grid Integration of Renewable Energies

A success story based on combinational PV power forecasts

Dr. Christian Kurz

Head of Prognoses & Data Analytics
meteocontrol GmbH – Energy & Weather Services
Augsburg, Germany
c.kurz@meteocontrol.com

In this paper we present an integrated combinational solar of renewable energies. The associated energy flow can lead
power forecast based on machine learning algorithms. The to a high load on individual power lines. To prevent this, the
forecasting system is used to optimize the grid and market network operators carry out redispatch measures. This can
integration of Renewable Energies. mean that in case of strong wind or solar power production
renewable energy plants have to be reduced in output in
Solar Power Forecasts; Machine Learning; Grid Integration certain grid areas. At the same time, conventional plants
I. INTRODUCTION have to be ramped up. Figure 2 shows the strong increase in
redispatch measures (in terms of redispatch energy).
Renewable energies represent an increasingly important Improved short-term forecasts can be an effective means of
contribution to our energy supply system. In Germany reducing the further increase in this energy and the
alone, installed photovoltaic (PV) capacity is around 45 GW associated further increase in costs.
(as of July 2019). However, a strong increase in capacity can
also lead to challenges in terms of secure grid integration.
One way of making it easier to plan the production and feed-
in of renewable energies is to draw up and optimize solar
power forecasts. This paper presents a combinational
approach to improve solar power production forecasts,
especially in the short-term area (intraday). The results are
discussed as examples for the German market, but can be
generalized to other countries.
II. NEED FOR SOLAR POWER FORECASTS
A. Energy Trading
Until a few years ago, the market integration of solar
power in Germany took place exclusively in the day ahead
segment. Due to the uncertainties in the weather models,
however, this is associated with forecasting errors that have
to be compensated during the actual day. As the share of
renewable energies increased, the amount of unbalanced Figure 1. Intraday trading volume in Germany from 2009 to 2018.. Data
energy due to these forecasting errors for the following day source: Epex Spot – European Power Exchange.
became increasingly large. This was associated with steadily
rising costs due to the procurement of balancing energy. For
this reason, intraday trading was also opened for renewable III. COMBINATIONAL FORECASTS
energies. This market has developed rapidly since its
introduction. Figure 1 shows that the intraday trading State of the art forecasting methods combine different
volume has increased tenfold within only 10 years to around information or input forecasts. Especially for short-term
50 TWh (for the year 2018). It can be assumed that this applications, numerical weather prediction models, satellite
process will develop even faster in new, faster growing images and current PV production values from monitoring
markets. New, optimized short-term forecasts are therefore systems are usually used. These 3 products are described in
essential. more detail in the following sections. The combination
factors are usually learned from machine learning methods
B. Grid Stability and Redispatch based on historic measured values.
Another important area of application for short-term
forecasts is grid stability. The spatial distribution of energy
generation and load often differs in grids with a high share
 2nd Int'l Conference on Large-Scale Grid Integration of Renewable Energy in India| New Delhi, India | 4-6 Sep 2019

A. Cloud Motion Vector Satellite Forecast horizon, the position of the sun, and other parameters.
For the short-term range of some hours a 'cloud motion Machine learning methods can be used for the combination.
vector' prediction based on current satellite images is used In the present case, a simple linear regression approach was
[1,2]. The displacement of cloud structures is determined implemented. Production data in 15 minute resolution from
from two consecutive satellite images by pattern 786 solar plants in Germany for one year were used for
recognition. The method is applied to each pixel of the validation. A new simulation was calculated every 15
image. This results in a vector field that describes the speed minutes. The forecast was adjusted separately for each
and direction of the atmospheric flow. These vectors are quarter of an hour of the day to the measured values.
used to extrapolate the current cloud structures into the near
future. In the last step irradiance at the ground is calculated
from the predicted cloudiness [3]. The latest development is
the use of satellite images in the thermal (infrared) spectral
range [4]. This makes it possible to calculate forecasts for
the early morning hours, even if no satellite information in
the visible spectral range is available at the time of
calculation. The 'cloud motion vector' forecasts are produced
in cooperation with the University of Oldenburg.

Figure 3. Weighting factors for the 3 input forecasts in dependence of the

day in the year. Forecast horizon: 15 minutes.

Figure 3 shows an example of the combination factors

determined for locations in Germany for the year 2017 for a
forecast horizons of 15 minutes depending on the day in the
year. The variation of the weighting factors depending on
the season is clearly visible. For this short forecast horizon
of 15 minutes the persistence forecast dominates in winter.
Summer, on the other hand, was characterized by unstable
weather conditions with frequently changing clouds.
Figure 2. Redispatch energy in Germany from 2010 to 2017. Data source: Accordingly, the cloud motion vector satellite forecast was
Bundesnetzagentur. increased in weighting by the regression models. The
numerical weather models receive only a low weight
B. Numerical Weather Prediction throughout the year.
Numerical weather prediction models have been state of
the art in calculating renewable energy production forecasts
for many years [5]. The models of several international and
national weather services are combined in order to generate
the best possible combinational forecast depending on the
weather situation. The main focus for the use of numerical
weather forecasts is the day ahead application. For a long
time, these models were also used for intraday short-term
forecasting. Due to the complexity and the associated long
computing time, however, the accuracy of the numerical
models is particularly limited in the short-term range.
C. PV Monitoring Data Figure 4. As figure 3, but for a forecast horizon of 2 hours.
Further important information is provided by measured
values of solar power production from monitoring systems. A changed picture emerges in the short-term range of 2
In contrast to the two methods described so far, in which the hours, see figure 4. Even in winter, the dominance of the
PV output is only calculated, monitoring systems provide persistence forecast is no longer very pronounced. In
real PV power measurements. The data can therefore be summer it even receives the lowest weight of all 3 input
used very well for the calibration of satellite and numerical forecasts. Here, as already for the 15 minute horizon, the
forecasts. At the same time, the actual production value can satellite forecast dominates. The consistently high weight of
be used to create an estimate for the next few minutes by the numerical forecast is conspicuous, especially in
means of a persistence forecast. Naturally, the accuracy of a combination with the satellite forecast. This can be
forecast based on a persistence approach, i.e. the assumption explained by the cloud motion vector prediction method.
of constant cloud structures, may be very limited depending This is based on a shift of existing cloud patterns into the
on the weather condition. future. Meteorological effects such as cloud formation or
dissipation are not represented in this method. This
D. Combination by Machine Learning information was taken from the numerical weather models,
The decisive step in the forecasting process is the whose weighting was chosen accordingly high by the
optimal combination of the three input forecasts mentioned regression algorithm.
above. The ideal model combination depends on the
geographical location of the PV system, the forecast
 2nd Int'l Conference on Large-Scale Grid Integration of Renewable Energy in India| New Delhi, India | 4-6 Sep 2019

IV. RESULTS grid integration and successful market integration of

In summary, figure 5 shows the accuracy achieved with renewable energies.
the combined forecast system as a function of the forecast ACKNOWLEDGMENT (HEADING 5)
horizon. Results are normalized to the root mean square
error of the numerical weather forecast, which was used as This study was supported by the German Federal
standard for forecasts also in the short-term range for a long Ministry for Environment, Nature Conservation and Nuclear
time. The graph also summarizes the respective advantages Safety on the basis of a decision by the German Bundestag
and disadvantages of the individual forecast approaches. The under contract number 0325583B.
persistence forecast derived from the measured values has
advantages especially for the current time step and the short
term range, since it concerns real power measurements. REFERENCES
However, the accuracy decreases rapidly as the forecast [1] E. Lorenz, J. Kühnert, and D. Heinemann. Short Term Forecasting of
horizon increases. The cloud motion vector forecast derived Solar Irradiance by Combining Satellite Data and Numerical
from the satellite data is, as described above, subject to Weather Predictions. In: Proceedings of 27th EUPVSEC. Frankfurt,
certain uncertainties due to the conversion of a cloud image Germany (2012).
to irradiance. Nevertheless, it is the most precise source for [2] J. Kühnert, E. Lorenz, and D. Heinemann. Satellite-Based Irradiance
the forecast period between about 2 and 4 hours. Decisive and Power Forecasting for the German Energy Market. In: Elsevier
Book on Solar Energy Forecasting and Resource Assessment.
for the approach chosen here is the result that the calculated Boston: Academic Press (2013), 267–297.
combination forecast always has a higher accuracy than [3] A. Hammer, D. Heinemann, C. Hoyer, R. Kuhlemann, E. Lorenz, R.
each of the 3 individual input forecasts. This applies to all Müller, and H. Beyer. Solar energy assessment using remote sensing
forecast horizons. technologies. In: Remote Sensing of Environment 86.3 (2003), 423–
432. doi : 10.1016/S0034-4257(03)00083-X.
[4] A. Hammer, J. Kühnert, K. Weinreich, and E. Lorenz. Short-Term
Forecasting of Surface Solar Irradiance Based on Meteosat-SEVIRI
Data Using a Night-time Cloud Index. In: Remote Sensing 7.7
(2015), 9070–9090. doi : 10.3390/rs70709070.
[5] E. Lorenz, T. Scheidsteger, J. Hurka, D. Heinemann, and C. Kurz.
Regional PV power prediction for improved grid integration. In:
Progress in Photovoltaics: Research and Applications 19 (2011),
757–771. doi : 10.1002/pip.

BIOGRAPHICAL INFORMATION
Figure 5. Normalized root mean square error of the 3 input forecasts, and
the combinational forecast in dependence of the forecast horizon. Dr. Christian Kurz is a senior expert in energy meteorology for
almost 20 years. After his studies of Atmospheric Physics at Munich
V. CONCLUSION University he worked at the German Aerospace Center (DLR) for
many years. He mainly worked on the parameterization of clouds
Current forecast solutions for the short-term sector and cloud-irradiation interactions in a Global Atmospheric
generally consist of a combination of different forecast Circulation model. His Ph.D. thesis was about the development of a
coupled model for atmospheric chemistry and circulation.
approaches. This paper presented a combination of
Since 2007 he is Head of Prognoses & Data Analytics at the
numerical weather prediction, a cloud motion vector forecast international energy service provider meteocontrol in Augsburg,
derived from satellite data, and measured production data Germany. He is responsible for the development of Renewable
from a monitoring system. The weighting factors were Energies power forecasts for grid operators and trading companies.
learned by linear regression models depending on the Together with his team he set up the first operational system for
season, the weather situation, and the forecast horizon. It online feed-in estimations and forecasts of PV power for the German
Transmission System Operators. A further focus of his work is Data
was shown that by this combination the forecast quality in Analytics of monitoring data by using machine learning methods.
the short-term range could be clearly increased. These
improved forecasts are an important prerequisite for secure