Energy Economics 117 (2023) 106421

Contents lists available at ScienceDirect

Energy Economics
journal homepage: www.elsevier.com/locate/eneeco

A stochastic time-series model for solar irradiation

Karl Larsson a ,∗, Rikard Green b , Fred Espen Benth c
a
School of Business, Örebro University, Fakultetsgatan 1, S-701 82 Örebro, Sweden
b
Energy Quant Solutions Sweden AB, Ekelundsvägen 25, S-22472 Lund, Sweden
c
Department of Mathematics, University of Oslo, P.O. Box 1053 Blindern, N-0316 Oslo, Norway

ARTICLE INFO ABSTRACT

JEL classification: We propose a novel stochastic time series model able to explain the stylized features of daily irradiation level
C5 data in 5 cities in Germany. The model is suitable for applications to risk management of photovoltaic power
Q4 production in renewable energy markets. The suggested dynamics is a low-order autoregressive time series
Keywords: with seasonal level given by an atmospheric clear-sky model. Moreover, we detect a skewness property in
Solar the residuals which we explain by a winter–summer regime switch. The stochastic variance is modeled by
Time-series a seasonally varying GARCH-dynamics. The winter and summer standardized residuals are proposed to be a
Photovoltaic Gaussian mixture model to capture the bimodal distributions. We estimate the model on the observed data, and
Renewable
perform a validation study. An application to energy markets studying the production at risk for a PV-producer
Energy
is presented.
Risk

1. Introduction using data on production volumes from transmission system opera-

tor (TSO) areas in Germany are (Benth and Ibrahim, 2017; Lingohr
The rapid growth of solar energy production from photovoltaic and Müller, 2019). Aggregated production for large areas have very
technology brings new types of risk to the energy market. With sig- different dynamics compared to single locations. Production at single
nificant volumes of electricity coming from solar power many energy locations or smaller areas are however of critical interest in many
systems are now exposed to the stochastic changes in irradiation levels applications. In addition, production data is only available for large
primarily driven by fluctuating cloud cover. The situation is similar areas. In order to assess the time series dynamics of solar energy pro-
with other renewable energy sources most importantly power pro- duction at specific coordinates a different approach is needed. A way
duction from wind. Stochastic factors such as irradiation levels and forward is to instead model the irradiation level at a given coordinate
wind speeds are intermittent, hard to predict and highly variable, directly and map it to energy output using a production function.
and they bring weather related risk to the forefront of financial risk This approach is also advocated by many market professionals, see
management in many energy markets. New financial instruments have e.g. De Jong (2020), and it is the one we pursue. Data on various
been introduced to allow market participants to manage the related measures of solar irradiation is available for arbitrary coordinates from
risks more efficiently. The growing share of renewables in the energy several different data providers. We use data from CAMS (Copernicus
mix has also changed the way markets operate. The intraday market
Atmosphere Monitoring Service) which is based on satellite imagery
is much more active today with participants trading very short term
and can be downloaded for any coordinate included in the spatial
contracts, for delivery within the hour, in order to manage sudden large
coverage of the service.1 Europe is e.g. fully covered and we choose
changes in the volumes produced from renewable sources.
to work with data for locations in Germany. Production functions
There is an emerging academic literature devoted to the chal-
for power production using different photovoltaic technologies have
lenges brought about by the increased dependency on renewable en-
been extensively studied in the associated engineering literature, see
ergy sources. Of particular interest for risk management purposes are
e.g. Huld et al. (2011) and Kaldellis et al. (2014).
time series models that are able to accurately represent the underlying
processes as well as being useful for calculating prices and risks. The The main contribution of this paper is a stochastic time series model
literature on solar energy modeling from this perspective is still rather for solar irradiation. Having a stochastic, and dynamically consistent,
scarce. Two recent studies that model solar energy production directly time series model is paramount for many applications and we therefore

∗ Corresponding author.
E-mail addresses: karl.larsson@oru.se (K. Larsson), rikard.green73@gmail.com (R. Green), fredb@math.uio.no (F.E. Benth).
1
https://atmosphere.copernicus.eu/data.

https://doi.org/10.1016/j.eneco.2022.106421
Received 30 August 2021; Received in revised form 9 November 2022; Accepted 18 November 2022
Available online 25 November 2022
0140-9883/© 2022 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
K. Larsson et al. Energy Economics 117 (2023) 106421

focus our attention on developing such a model. The model we propose Table 1
Latitude and longitude for the locations.
is demonstrated to capture the most salient features of the time series
dynamics of solar irradiation. Among these features are seasonal and Location Latitude Longitude

auto-regressive effects in both the level and variance of irradiation. Hamburg 53.4361 9.6311
Berlin 52.6306 13.6263
There are also seasonal dependencies in the distributional properties
Nürnberg 49.4073 10.9265
of data. Residuals are shown to follow distinctly different bimodal Stuttgart 48.9296 9.2896
distributions in summer and winter. These distributions are fitted very München 48.4437 11.3660
closely by bimodal Gaussian mixture distributions. Our model is consis-
tent with all these characteristics using only standard time series tools.
The simplicity of the model makes it easy to estimate and evaluate
which is appealing in both academic and applied work. to other data sources, see e.g. Marchand et al. (2020) and Yang and
The paper (Casula et al., 2020) is similar to ours in some respects. Bright (2020). In Marchand et al. (2020) CAMS data is compared to
The authors propose a stochastic time series model for solar irradia- the HelioClim-3 database, which is also based on satellite imagery, for a
tion and also considers financial applications. However, their model number of locations in Germany. Both data sources are found to be able
does not acknowledge the very strong seasonality in the conditional to accurately represent the temporal and spatial variation in irradiation
volatility and disregards the bimodality and seasonal dependency of data. The authors point out that satellite based irradiation data has been
the distribution for the residuals. We find that these features constitute shown to have a higher accuracy compared to data based on reanalysis
important aspects of irradiation data. Accurate description of volatility (as e.g. the ERA5 and MERRA-2 databases) and that CAMS is a widely
and distributional properties are particularly important for applications used and reliable data source.3
in finance and should not be ignored. While Germany has many weather stations that measure solar ir-
We end the paper with an example application to the management radiation, only a few of them can provide high quality data, see
of volume risk for a solar energy producer. There are many other e.g. Marchand et al. (2020) for a discussion of this point. In most
applications that could be addressed with our model. In Cuppari et al. practical cases for energy production the irradiation must be evaluated
(2021) the authors bring attention to the prospect for landowners to at coordinates without available high quality measurements. Hence we
reduce financial risk by co-locating agriculture and solar power pro- prefer to model satellite based data directly since it will be the most
duction technology. Their analysis is built on an interesting application relevant data scenario for applications envisioned for our model. Our
of diversification and takes a stochastic irradiation level as one model view is that working with this data provides the most information for
input. The articles (Benth and Ibrahim, 2017; Casula et al., 2020) potential users.
both employ their models to the valuation of different types of options In this study we choose to model global horizontal irradiance (GHI)
on the revenue stream of a solar park. The authors in Lingohr and which is the radiation received on a horizontal plane from all direc-
Müller (2019) use their model to analyze a futures contract written on tions. It is the most relevant measure for photovoltaic energy produc-
solar energy production. They envision a contract similar in spirit to tion and it is measured in Wh/m2 . We use data on GHI collected at
the wind power production futures contracts traded on the European 11:00 (UCT), around the peak time for solar intensity, for different
Energy Exchange (EEX). Another important application is investment coordinates in Germany. The locations of our selected coordinates are
decisions for the construction, and location, of new solar parks. Our situated near Hamburg, Berlin, Stuttgart, Nürnberg and München. Our
model is perfectly suited for all these applications. It would also be choice of coordinates gives a reasonable coverage of different parts of
of interest to develop multivariate time series models e.g. for wind Germany and makes it possible to address differences between locations
speed, temperature, solar irradiation and electricity prices considered that are both close and far apart.
jointly. Such models should ideally build on, and be consistent with, Our final dataset consists of 11 years of daily observations sampled
the univariate time series dynamics of each component. Our model at 11:00 pm (UCT) for each location during the period 2010-01-01 to
provides a natural benchmark for the irradiation part of such models. 2020-12-31. Data from the 10 year period 2010-01-01 to 2019-12-31 is
Detailed investigations of these interesting topics are left for future used as in-sample for estimation and analysis, and the last year (2020-
research. 01-01 to 2020-12-31) is used in an out-of-sample model prediction
exercise as part of the model validation. Leap year days in 2012, 2016
2. The stochastic model and irradiation data and 2020 were removed. There are very few missing data points with
a maximum of 13 out of 4015 observations for Hamburg and Stuttgart.
2.1. Description of data For the out-of-sample period there are no missing observations. We
replace missing values using linear interpolation between the nearest
We use data obtained from CAMS (Copernicus Atmosphere Moni- observed hours.
toring Service) which is a part of the European Union’s Earth obser- The coordinates for the selected locations are given in Table 1
vation programme and implemented by ECMWF (European Centre for and descriptive statistics are given in Table 2. The mean radiation
Medium-Range Forecasts).2 The data from CAMS is based on satellite increases with decreasing latitude which not unexpectedly tells that
imagery and model input to account for the impact of clouds and other the more southern locations display higher average radiation levels.
atmospheric conditions. We refer to the CAMS user guide (CAMS, 2019; We also observe that the standard deviation is higher for the southern-
Gschwind et al., 2019; Lefevre et al., 2017; Qu et al., 2017) for detailed most locations. Additionally, we notice that the locations situated more
accounts of the methods used. The data is free of charge and can be to the west has a slightly lower standard deviation than the eastern
downloaded for arbitrary coordinates given the spatial coverage of the locations. The maximum and minimum values for the irradiation data
service. One of the stated goals of the service is to provide radiation are confirming higher irradiation for lower latitude. The skewness is
data accurate enough for scientific and commercial use in applications positive for all locations but with the southern locations displaying
to solar energy production. The data is quality controlled and validated lower values.
on a quarterly basis, see CAMS (2019). Data from CAMS has also been
evaluated in several academic studies and fares well in comparison
3
The ERA5 and MERRA-2 databases are operated by ECMWF and NASA
respectively and are both based on reanalysis. We refer to their respective web-
2
Data from the CAMS service is provided by their CAMS data store found sites at https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era5
at https://atmosphere.copernicus.eu/data. and https://gmao.gsfc.nasa.gov/reanalysis/MERRA-2/ for more information.

2
K. Larsson et al. Energy Economics 117 (2023) 106421

Fig. 1. Time series of irradiation 𝐺(𝑡).

Table 2 a stochastic process 𝑍(𝑡) as

Descriptive statistics for the locations.
Location Mean Std.dev. Max Min Skewness 𝐺(𝑡) = 𝑆(𝑡) + 𝑍(𝑡). (1)
Hamburg 382.3726 244.3257 897.8311 36.8759 0.3111 It is natural to base the seasonal component 𝑆(𝑡) on the extraterrestrial
Berlin 393.1779 247.1781 896.1819 38.8503 0.2673
radiation which is the amount of radiation reaching the outer boundary
Nürnberg 450.5042 262.8204 951.5443 47.8573 0.1819
Stuttgart 434.7473 257.1569 940.1480 51.1439 0.2072
of the earth’s atmosphere. The extraterrestrial radiation is determined
München 456.2307 264.8277 956.9816 63.8053 0.1753 by the earth’s movement around the sun and can for all practical
purposes be regarded as a deterministic function of time. At a given
coordinate and time 𝑡 the extraterrestrial radiation is given by
( ( ))
360𝑡
Solar irradiation is influenced by astronomical and atmospherical 𝛬(𝑡) = 𝐾 1 + 0.033 cos cos(𝜃(𝑡)) (2)
365
conditions. Extraterrestrial radiation is the amount of irradiation before
it reaches the atmosphere. It can be regarded as a deterministic function In Eq. (2), 𝐾 is the solar constant which we set to 𝐾 = 1367 (W/m2 )
and is determined from the relative positions of the earth and the sun. and cos(𝜃(𝑡)) is the zenith angle at the selected coordinate at time 𝑡.4
For all our locations the extraterrestrial radiation peaks on June 20 and The function 𝜃(𝑡) is implicitly determined from the solar zenith angle
takes its lowest value on December 20. Extraterrestrial radiation is re- given by
duced through absorption and scattering by particles in the atmosphere. cos(𝜃(𝑡)) = cos(𝑙𝑎𝑡) cos(𝑤(𝑡)) cos(𝛿(𝑡)) + sin(𝑙𝑎𝑡) sin(𝛿(𝑡)) (3)
The irradiation levels reaching the ground is therefore substantially
lower even under clear sky conditions. Due to random interference of where 𝑤(𝑡) and 𝛿(𝑡) are the solar hour angle and declination, resp.,
clouds and particles irradiation levels at the ground are also stochastic. and 𝑙𝑎𝑡 denotes the latitude of the selected coordinate. We refer to
The cloud cover is a very important stochastic component of irradiation Appendix A on how to determine the solar hour angle and declination.
data and it may change drastically even over short periods of time. Such The seasonality in GHI is closely linked to that of the extraterrestrial
rapid changes in the cloud cover occur more frequently in the summer radiation. Actual GHI may also be affected by seasonal components in
causing highly variable irradiation levels. The variations in summer are the atmosphere and most importantly in the cloud cover. Since the ex-
amplified by the seasonally higher levels; the appearance of sudden traterrestrial radiation is not affected by clouds and other atmospheric
clouds may cause the irradiation to drop sharply from a very high level conditions it attains much larger values than actual GHI received at
thus creating large swings. During the winter on the other hand, the the ground level. This means that extraterrestrial radiation must be
variation is smaller due to periods of more persistent overcast skies and calibrated to GHI data 𝐺(𝑡) in order to achieve a proper seasonal
lower overall irradiation levels. adjustment. We propose to use the following seasonality function based
Fig. 1 plots the time series of irradiation values for our 10 year on 𝛬(𝑡),
in-sample period and for all locations. The graphs show the strong sea- ( )
𝑆(𝑡) = 𝛬(𝑡) 𝑎0 + 𝑎1 𝑒−𝑎2 ∕ cos(𝜃(𝑡)) (4)
sonal pattern present in solar irradiation data. Much of the seasonality
is determined from the deterministic extraterrestrial radiation but may where 𝑎𝑖 , 𝑖 = 0, 1, 2 are constants and cos (𝜃(𝑡)) is given in (3). The
also be influenced by seasonal factors in the atmosphere. It can also function (4) can be motivated e.g. by the simple parametric clear sky
be noted from Fig. 1 that the summer periods display larger variations
compared to the winter periods.
4
The solar constant actually varies to some degree, e.g. due to variations
2.2. Time series dynamics in the earth–sun distance, sun spot activity, etc. Different proposed values of
the solar constant can be found in the literature. The differences are small and
In a given location, let 𝐺(𝑡) be the GHI at times 𝑡 = 0, 1, …. We the specific value does not matter for our purposes since we will calibrate the
model 𝐺(𝑡) as the sum of a deterministic seasonal component 𝑆(𝑡) and seasonality function to data.

3
K. Larsson et al. Energy Economics 117 (2023) 106421

model presented in Hottel (1976). A parametric clear sky model is The final model component left to specify is the driving noise
a deterministic model for radiation under the assumption of cloud process 𝑒(𝑡). We find that the irradiation time series display seasonal
free conditions but taking other atmospheric conditions into account, behavior also in the third moment. The skewness of irradiation data
and should give an upper bound for the actual GHI. The clear sky has a clear seasonal pattern turning negative during the summer period
model in Hottel (1976) has additional equations for determining the from being positive in winter. Our empirical results show that this
parameters 𝑎0 , 𝑎1 , 𝑎2 but can be written on the simplified form (4). effect is important for both the understanding of irradiation data and
A clear sky model is calibrated to radiation data under clear sky for model fit. The actual definitions of the summer and winter spans
conditions. However, our aim is not to recreate a clear sky model, we are determined from empirical considerations. All our locations display
are instead using the functional form for 𝑆(𝑡) in order to achieve a similar patterns and it is a feature that we have not seen documented
viable seasonal adjustment of 𝐺(𝑡). This means that we can settle for in other studies. It is common to find bimodality in the distribution
the simpler specification in (4). of high frequency radiation data. The bimodal property is commonly
We model the seasonally adjusted series 𝑍(𝑡) = 𝐺(𝑡) − 𝑆(𝑡) as an explained as representing states of either cloudy or clear sky conditions.
AR(p) process according to To account for both the effects of seasonal skewness and bimodality we
∑
𝑝 model the process 𝑒(𝑡) as
𝑍(𝑡) = 𝛽𝑖 𝑍(𝑡 − 𝑖) + 𝑢(𝑡) (5)
𝑖=1 𝑒(𝑡) = 𝑑(𝑡)𝜖𝑆 (𝑡) + (1 − 𝑑(𝑡))𝜖𝑊 (𝑡) (10)
with constant coefficients 𝛽𝑖 , 𝑖 = 1, … , 𝑝. The modeling approach in
where 𝑑(𝑡) is a simple dummy-process that takes the value 1 in summer,
Eqs. (1) and (5) is strongly supported by the theoretical and empirical
and the value 0 in winter, and the processes 𝜖𝑘 , 𝑘 ∈ {𝑆, 𝑊 }, are as-
analysis in Benth and Šaltytė Benth (2012). With this specification 𝑆(𝑡)
sumed to be independent iid sequences of Gaussian mixture distributed
becomes the seasonal mean function that the irradiation level 𝐺(𝑡) mean
random variables. The probability densities for 𝜖𝑘 , 𝑘 ∈ {𝑆, 𝑊 } are given
reverts to which is an appealing property. The residual process 𝑢(𝑡) in
by,
(5) is specified as

𝑢(𝑡) = 𝜎(𝑡)𝑒(𝑡) (6) 𝑓𝑘 (𝑥) = 𝑞𝑘 𝑓1,𝑘 (𝑥) + (1 − 𝑞𝑘 )𝑓2,𝑘 (𝑥), 𝑥 ∈ R, 𝑘 ∈ {𝑆, 𝑊 } (11)

where 𝜎(𝑡) is a time varying conditional volatility. We find that the where 𝑓𝑚,𝑘 (𝑥) are the densities of Gaussian random variables with
means 𝜇𝑚,𝑘 and variances 𝜈𝑚,𝑘 2 for 𝑚 = 1, 2 and 𝑘 ∈ {𝑆, 𝑊 }. We use the
conditional variance of 𝑢(𝑡) is heteroscedastic, containing both seasonal
and ARCH effects. There are physical reasons for expecting volatility notation GM(𝜇1 , 𝜇2 , 𝜈12 , 𝜈22 , 𝑞) for such a Gaussian mixture distributions
clustering in irradiation data. Scattered and rapidly moving clouds with two components. The parameter 𝑞 is the prior probability of being
will cause more variation in solar irradiation and can persist for some in the first state 𝑚 = 1. With this notation we thus have
days. A thick cloud cover on the other hand will lead to less variation 2 2
𝜖𝑘 (𝑡) ∼ GM(𝜇1,𝑘 , 𝜇2,𝑘 , 𝜈1,𝑘 , 𝜈2,𝑘 , 𝑞𝑘 ), 𝑘 ∈ {𝑆, 𝑊 } (12)
that can persist for a few days. These effects could be sources of
volatility clustering. We account for seasonality and volatility clustering The moments of Gaussian mixture distributions for 𝜖𝑆 (𝑡) and 𝜖𝑊 (𝑡)
by specifying the conditional variance as are easily calculated given the probability densities (11). We intro-
𝜎 2 (𝑡) = 𝜎𝑆2 (𝑡)𝜎𝐺
2
(𝑡) (7) duce the following notation for the expected value, variance and the
centralized third moment of 𝜖𝑘 (𝑡), 𝑘 ∈ {𝑆, 𝑊 },
with a deterministic seasonal component 𝜎𝑆2 (𝑡)
multiplied by a GARCH- [ ]
type component 𝜎𝐺 2 (𝑡). We are using a multiplicative structure similar 𝜇𝑘 = E 𝜖𝑘 (𝑡)
( )
to Benth and Šaltytė Benth (2012). The seasonal component 𝜎𝑆2 (𝑡) is 𝛴𝑘2 = Var 𝜖𝑘 (𝑡)
given by a truncated Fourier series containing three terms, [( [ ])3 ]
( ) ( ) 𝛺𝑘 = E 𝜖𝑘 (𝑡) − E 𝜖𝑘 (𝑡)
2𝜋𝑡 2𝜋𝑡
𝜎𝑆2 (𝑡) = 𝑐0 + 𝑐1 cos + 𝑐2 sin (8)
365 365 and state, without proof, the following explicit expressions,
This formulation is widely used for seasonal volatility modeling, see
𝜇𝑘 = 𝑞𝑘 𝜇1,𝑘 + (1 − 𝑞𝑘 )𝜇2,𝑘
e.g. Campbell and Diebold (2005), Härdle and Lopez Cabrera (2012) ( ) ( )
and Benth and Šaltytė Benth (2012, 2013) for applications to tempera- 𝛴𝑘2 = 𝑞𝑘 𝜈1,𝑘
2
+ 𝜆21,𝑘 + (1 − 𝑞𝑘 ) 𝜈2,𝑘
2
+ 𝜆22,𝑘
ture and wind speed modeling, and Benth et al. (2008) for applications ( ) ( )
√ 2
𝛺𝑘 = 𝑞𝑘 3𝜆1,𝑘 𝜈1,𝑘 + 𝜆31,𝑘 + (1 − 𝑞𝑘 ) 3𝜆2,𝑘 𝜈2,𝑘
2
+ 𝜆32,𝑘
to energy prices. The parameter restriction 𝑐0 > 𝑐12 + 𝑐22 ensures that
the seasonal variance 𝜎𝑆2 (𝑡)
is positive. See Appendix B for details. It is where 𝜆𝑚,𝑘 = 𝜇𝑚,𝑘 − 𝜇𝑘 , for 𝑚 = 1, 2. The skewness of 𝜖𝑘 (𝑡) is given by
possible, and straightforward, to include more trigonometric terms in 𝛺𝑘 ∕𝛴𝑘3 .
(8) but we find that three terms are sufficient. Extending the Fourier Given the independence between 𝜖𝑆 (𝑡) and 𝜖𝑊 (𝑡) it follows that the
series beyond three terms did not result in a substantially different corresponding moments of 𝑒(𝑡) are given by
seasonality curve for the variance. In order to keep the number of
parameters low we settled for the three term model presented here. The E [𝑒(𝑡)] = 𝑑(𝑡)𝜇𝑆 + (1 − 𝑑(𝑡))𝜇𝑊
2 (𝑡) is assumed to take the form
auto-regressive GARCH component 𝜎𝐺 Var (𝑒(𝑡)) = 𝑑(𝑡)𝛴𝑆2 + (1 − 𝑑(𝑡))𝛴𝑊
2

of a standard GARCH(1,1)-process based on the residual series 𝑣(𝑡) = [ 3

]
E (𝑒(𝑡) − E [𝑒(𝑡)]) = 𝑑(𝑡)𝛺𝑆 + (1 − 𝑑(𝑡))𝛺𝑊
𝑢(𝑡)∕𝜎𝑆 (𝑡) standardized by the seasonal component. The specification
becomes Note that with 𝜇𝑆 = 𝜇𝑊 = 0 and 𝛴𝑆2 = 𝛴𝑊 2 = 1 we also have E [𝑒(𝑡)] = 0

2 2 and Var (𝑒(𝑡)) = 1 for all 𝑡. With Var (𝑒(𝑡)) = 1 the skewness of 𝑒(𝑡) is
𝜎𝐺 (𝑡) = 𝜔0 + 𝜔1 𝜎𝐺 (𝑡 − 1) + 𝜔2 𝑣2 (𝑡 − 1) (9)
2
= 1 − 𝜔1 − 𝜔2 + 𝜔1 𝜎𝐺 (𝑡 − 1) + 𝜔2 𝑣2 (𝑡 − 1) Skew (𝑒(𝑡)) = 𝑑(𝑡)𝛺𝑆 + (1 − 𝑑(𝑡))𝛺𝑊 (13)

The unconditional variance of the GARCH part is given by 𝜔0 (1 − which is time-varying and provides a natural channel for explaining
𝜔1 − 𝜔2 )−1 and should be equal to one since it operates on the partly the skewness patterns observed over the year. This completes our
standardized residuals 𝑣(𝑡). We therefore employ the parametrization description of the model. In the next section we demonstrate the models
𝜔0 = 1 − 𝜔 − 𝜔2 which ensures this property and in addition avoids the ability to account for the many particular features of irradiation time
issue of identifying two separate intercepts in 𝜎𝑆2 (𝑡) and 𝜎𝐺
2 (𝑡). series data.

4
K. Larsson et al. Energy Economics 117 (2023) 106421

Fig. 2. Irradiation 𝐺(𝑡) (blue) and fitted seasonal function 𝑆(𝑡) (red). (For interpretation of the references to color in this figure legend, the reader is referred to the web version
of this article.)

Fig. 3. Autocorrelation function (ACF) (left panel) and partial autocorrelation function (PACF) (right panel), for the seasonal adjusted series 𝑍(𝑡) = 𝐺(𝑡) − 𝑆(𝑡).

3. Estimation and validation However, checking crucial model properties at each step shows that
the estimated model performs well in capturing stylized features of the
3.1. Estimation data. For ease of exposition we illustrate the estimation using data only
for the Hamburg location. All parameter estimates for Hamburg are
We propose a stepwise estimation strategy that allows for verifi- presented in Table 3. The corresponding results for the other locations
cation of crucial model properties at each stage. Stepwise estimation
are presented in Table 4 in Section 3.2. The features observed at the
strategies have been used by many other authors in similar models,
Hamburg location are similar to the other locations.
see e.g. Alaton et al. (2002), Härdle and Lopez Cabrera (2012), Camp-
bell and Diebold (2005) and Benth and Šaltytė Benth (2012, 2013). The first step is to calibrate the seasonal function 𝑆(𝑡) given in (4)
Using stepwise estimation has many advantages, being fast and easy to data 𝐺(𝑡). From (4) we have
to perform. Since standard errors are not obtained from a single objec- 𝑆(𝑡) ( )
tive function confidence intervals must be treated with some caution. = 𝑎0 + 𝑎1 exp −𝑎2 ∕ cos(𝜃(𝑡))
𝛬(𝑡)

5
K. Larsson et al. Energy Economics 117 (2023) 106421

Table 3 not sufficient to remove the seasonality from our data which is on
Parameter estimates and standard errors for all model parameters for the Hamburg
a high frequency and not in aggregated form. The irradiation time
location. The parameters 𝑐0 , 𝑐1 , 𝑐2 , and their standard errors, have been divided by
100.
series 𝐺(𝑡) together with the fitted seasonal curve 𝑆(𝑡) are plotted in
Parameter Estimate Std.Err. Parameter Estimate Std.Err.
Fig. 2. The function 𝑆(𝑡) is capturing the seasonality in the level of
𝐺(𝑡) very well. Estimates of 𝑎0 , 𝑎1 , 𝑎2 are presented in Table 3. The
𝑎0 0.2003 0.0374 𝜇1,𝑆 −1.0407 0.0350
𝑎1 0.5993 0.0172 𝜇2,𝑆 0.6688 0.0192
mean of the irradiation 𝐺(𝑡) is 382.37 and the mean of the calibrated
𝑎2 0.4270 0.0779 2
𝜈1,𝑆 0.3703 0.0279 function 𝑆(𝑡) is 382.44. Fig. 3 shows the autocorrelation function (ACF)
𝛽1 0.2259 0.0143 2
𝜈2,𝑆 0.2606 0.0132 and partial autocorrelation function (PACF) for the seasonally adjusted
𝛽2 0.0605 0.0141 𝑞𝑆 0.3912 0.0149 series 𝑍(𝑡) = 𝐺(𝑡) − 𝑆(𝑡). There are no signs of seasonal dependency in
𝑐0 191.0228 4.4450 𝜇1,𝑊 −0.7188 0.0115 the ACF for 𝑍(𝑡) showing that the seasonal level adjustment performs
𝑐1 −173.1130 4.6303 𝜇2,𝑊 0.9479 0.0370
𝑐2 36.5643 2.6998 2
𝜈1,𝑊 0.1341 0.0062
well.
𝜔1 0.6165 0.1277 2
𝜈2,𝑊 0.5601 0.0376 The ACF for 𝑍(𝑡) suggests that a low order AR-structure is a suitable
𝜔2 0.0798 0.0164 𝑞𝑊 0.5687 0.0135 model structure. This is confirmed by the PACF which indicates that
an order of 𝑝 = 2 is sufficient to describe the autocorrelation structure.
Before estimating the AR-parameters in (5) we subtract the mean of
Table 4
the seasonally adjusted series 𝑍(𝑡). Any remaining mean is very small
Parameter estimates and standard errors for all locations except Hamburg. The
corresponding results for Hamburg are given in Table 3. In this table the parameters but may vary between locations depending on the seasonal fit. With an
𝑐0 , 𝑐1 and 𝑐2 have been divided by 100. order of 𝑝 = 2 for the zero mean AR-model for 𝑍(𝑡) we need to estimate
Parameter Berlin Stuttgart Nürnberg Nürnberg the parameters 𝛽1 and 𝛽2 . This is done using a standard least squares
𝑎0 0.0927 −0.2316 −0.2878 0.0019 fit. The estimates for these parameters can be found in Table 3 together
(0.0771) (0.4741) (0.4525) (0.2100) with their estimated standard errors. The AR-parameters 𝛽1 and 𝛽2 are
𝑎1 0.6851 1.0084 1.0565 0.8149 significant and positive.
(0.0492) (0.4383) (0.4195) (0.1678)
From the estimated AR-model part 𝑍(𝑡) we can now obtain observa-
𝑎2 0.3168 0.1777 0.1725 0.2809
(0.0771) (0.1216) (0.1069) (0.1287)
tions of the process 𝑢(𝑡) from (5). The ACFs for 𝑢(𝑡) and 𝑢2 (𝑡) are plotted
𝛽1 0.2686 0.2467 0.3234 0.2728 in Fig. 4. There are no seasonal and serial dependencies in the ACF
(0.0140) (0.0143) (0.0143) (0.0140) for 𝑢(𝑡). The residuals 𝑢(𝑡) also passes the Ljung–Box test with a 𝑝-value
𝛽2 0.0532 0.0641 – – of 0.4001. The ACF for 𝑢2 (𝑡) on the other hand shows strong seasonal
(0.0139) (0.0144) – –
variation. There are also low order ARCH effects but these are hard to
𝑐0 190.1884 261.4136 231.7968 273.2423
(4.8957) (6.3013) (5.9295) (7.3385) glance from Fig. 4. Both these effects are accounted for by the volatility
𝑐1 −172.1055 −210.5509 −192.4294 −217.5883 structure specified by the conditional variance in (7). We estimate the
(5.1640) (7.1180) (6.6297) (8.0930) parameters 𝑐𝑖 , 𝑖 = 0, 1, 2 of the seasonal part 𝜎𝑆2 (𝑡), and the parameters
𝑐2 32.7133 37.2395 35.7742 50.6053 𝜔1 and 𝜔2 of the GARCH part 𝜎𝐺 2 (𝑡) jointly using quasi-maximum-
(2.9577) (4.6673) (4.2530) (5.2746)
𝜔1 0.6829 0.6103 0.5823 0.5103
likelihood based on a Gaussian underlying distribution for 𝑢(𝑡). Then
(0.1188) (0.0838) (0.1522) (0.0697) we fit the Gaussian mixture distributions to the fully standardized
𝜔2 0.0794 0.0963 0.0921 0.1477 residuals 𝑒(𝑡) = 𝑢(𝑡)∕𝜎(𝑡). In a final step we re-estimate the variance
(0.0183) (0.0155) (0.0200) (0.0173) parameters from maximum likelihood based on the fitted Gaussian
𝜇1,𝑆 −0.9814 −1.1085 −1.0022 −1.0530
mixture distributions as recommended e.g. by Engle and Gonzlaez-
(0.0443) (0.0363) (0.0499) (0.0442)
𝜇2,𝑆 0.6785 0.6423 0.6576 0.6465 Rivera (1991). This step does not change the parameter estimates for
(0.0222) (0.0125) (0.0199) (0.0140) the dominating seasonal variance much and overall model performance
2
𝜈1,𝑆 0.4399 0.4863 0.4980 0.5428 is unaffected. Before discussing the estimation of the distributional part
(0.0359) (0.0353) (0.0423) (0.0421) of the model in more detail we first examine the observed residuals 𝑒(𝑡)
2
𝜈2,𝑆 0.2603 0.1724 0.2372 0.1814
and the variance parameters.
(0.0142) (0.0074) (0.0124) (0.0083)
𝑞𝑆 0.4087 0.3669 0.3962 0.3804
From the estimated conditional volatility 𝜎(𝑡) we can now study the
(0.0187) (0.0124) (0.0188) (0.0149) observed residual series 𝑒(𝑡). Fig. 4 shows the ACFs of 𝑒(𝑡) and 𝑒2 (𝑡). As
𝜇1,𝑊 −0.6868 −0.6945 −0.6190 −0.7008 can be seen, both the seasonality and ARCH-effects have been removed.
(0.0138) (0.0128) (0.0163) (0.0125) Judging from these graphs our model adequately describes seasonal
𝜇2,𝑊 0.8987 1.0679 1.1079 1.0210
and auto-regressive components present both in the conditional mean
(0.0522) (0.0267) (0.0432) (0.0319)
2
𝜈1,𝑊 0.1549 0.1830 0.2242 0.1615 and variance of radiation data. The estimates of 𝑐𝑖 , 𝑖 = 0, 1, 2, are
(0.0082) (0.0080) (0.0104) (0.0071) presented in Table 3. The estimate of 𝑐0 corresponds well to the
2
√
𝜈2,𝑊 0.6790 0.3722 0.4731 0.4616 empirical standard deviation of 𝑍(𝑡); 𝑐0 is estimated equal to 138.21
(0.0521) (0.0236) (0.0384) (0.0306) which can be compared to a standard deviation of 138.9 for 𝑍(𝑡).
𝑞𝑊 0.5668 0.6059 0.6416 0.5930
The estimates of 𝑐1 and 𝑐2 multiplying the trigonometric terms in (8)
(0.0184) (0.0111) (0.0152) (0.0123)
are of different signs and magnitude producing a slight shift away
from a symmetric seasonal pattern for the year. The seasonal variance
attains its maximum in late June and its minimum in late December.
where 𝛬(𝑡) is the extraterrestrial radiation in (2). We use this form and The parameters 𝜔1 and 𝜔2 are both significant and fulfill the standard
calibrate the parameters 𝑎0 , 𝑎1 , 𝑎2 by minimizing the function conditions for positivity and stationarity for a GARCH(1,1) process. We
∑𝑇 ( )2 remark that the impact of the GARCH-part is low compared to the
𝐺(𝑡) 𝑆(𝑡) seasonal variance. As mentioned there are physical reasons to expect
𝑄(𝑎0 , 𝑎1 , 𝑎2 ) = − .
𝑡=1
𝛬(𝑡) 𝛬(𝑡) clustered variance, and we do detect some large autocorrelations in
with non-linear least squares. Fitting the parameters to 𝐺(𝑡)∕𝛬(𝑡) in- squared residuals and receive significant parameter estimates, which
stead of directly to 𝐺(𝑡) helps produce more precise estimates. The motivates inclusion of the GARCH-part.
quantity 𝐺(𝑡)∕𝛬(𝑡) is sometimes referred to as the clearness index and To motivate our model for 𝑒(𝑡) in (10) we have plotted the observed
in some studies it is used as the only seasonal adjustment which may residual series 𝑒(𝑡) in Fig. 5. The left panel shows the actual series for
be sufficient for very short periods of time or when observations are 𝑒(𝑡) and the right panel shows a simulated series from our estimated
aggregated and studied on lower time frequencies. This approach is model. The time series plot of 𝑒(𝑡) in the left panel indicate that there

6
K. Larsson et al. Energy Economics 117 (2023) 106421

Fig. 4. Autocorrelation functions (ACF) for 𝑢(𝑡) (upper left), 𝑢2 (𝑡) (upper right), 𝑒(𝑡) (lower left), and 𝑒2 (𝑡) (lower right).

Fig. 5. Time series plots of observed 𝑒(𝑡) (left panel) and simulated 𝑒(𝑡) (right panel).

are still seasonal effects left, despite having successfully adjusted for The remaining seasonality in 𝑒(𝑡), demonstrated in Figs. 5 and
seasonality both in the level and conditional variance of data. The right 6, is related to time variation in the skewness of irradiation data.
panel, with simulated data, shows that our model is consistent with this As explained below the seasonal variation in skewness is due to the
behavior. fact that the distributions for the residuals are distinctly different in
In Fig. 6 we have plotted the ACF for 𝑒3 (𝑡) for both the actual series the summer and winter periods. Fig. 7 plots the monthly empirical
(left panel), and for simulated data from the estimated model (right estimates of the skewness for 𝐺(𝑡) and 𝑒(𝑡) (left panel) and the same
panel). The left panel indicates a weak but clear seasonal pattern in estimate for 𝑒(𝑡) together with the skewness function from the estimated
the ACF over the year. Again, the right panel, based on simulated data, model (right panel). The monthly skewness estimates are calculated
demonstrates our models ability to capture also this feature. from grouping data according to month for all observations during

7
K. Larsson et al. Energy Economics 117 (2023) 106421

Fig. 6. Autocorrelation functions (ACF) for 𝑒3 (𝑡) based on observed (left panel) and simulated (right panel) data.

our 10 year sample. We included skewness estimated for both the regard our simple skewness model to reach a satisfactory level of fit.
original radiation time series 𝐺(𝑡) and the residuals 𝑒(𝑡) in the left panel The empirical skewness for summer is −0.3946 and our model gives
to show that the time variation in skewness is not an artifact of the a skewness of −0.3929. For the winter period the empirical estimate
model steps preceding the construction of 𝑒(𝑡). The small differences is 0.6785 and the model gives 0.6704. Moreover, the comparisons
in skewness estimates between 𝐺(𝑡) and 𝑒(𝑡) are due to the seasonal between the observed and simulated 𝑒(𝑡) in Figs. 5 and 6 show that
adjustment function 𝑆(𝑡). The estimates based on 𝑒(𝑡) show that the our simple model structure is in close alignment with data.
skewness changes in a periodical pattern over the year. It is positive The time variation in skewness is explained by the fact that the
in the months November–February and turns negative in the months distributions for irradiation are very different in the winter and summer
March–October. This seasonal time variation in skewness is a distinct periods. A typical feature of high frequency irradiation data is that
feature of irradiation data. It is observed at all our locations and we the distribution is bimodal. In Fig. 8 we plot the histogram of the
have not seen it documented in the related literature before. The reason residual series 𝑒(𝑡) where we have superimposed a standard kernel den-
for these differences are likely to be related to different cloud conditions sity estimate using a Gaussian kernel with correspondingly optimized
during the year with more sustained periods of overcast skies in winter. bandwidth. The bimodal feature is clearly displayed. All our data series
This explanation is very much in line with the results we find below show similar behavior. The shape of the histogram of 𝑒(𝑡) reveals an
when we proceed to fit different Gaussian mixture distributions to the important characteristics of the data and it is one we would like to be
summer and winter residuals. able to recreate with our model. However, as observed in our previous
Since the months with positive/negative skewness are the same discussion the residual series 𝑒(𝑡) is not an iid sequence and hence
across locations we label them as summer (March-October) and winter the histogram is not a valid representation of the actual distribution
(November-February). The skewness estimates displayed in Fig. 7 show of 𝑒(𝑡). Fitting a Gaussian mixture distribution directly to 𝑒(𝑡) would
that the shift from positive to negative, and back again, occur quite provide an accurate recreation of the histogram, however, due to the
rapidly during the spring and autumn seasons. In between the shifts temporal dependence in the data series, it would not be consistent. By
the skewness varies much less. This behavior motivates our simple instead separating 𝑒(𝑡) into summer and winter residuals our model can
definition of the summer and winter without sacrificing any crucial accomplish this is in a fully consistent way.
information. We have tested basing the summer definition instead on We assume Gaussian mixture distributions with two components
the period where the extraterrestrial radiation exceeds its mean level. for both the summer and winter residuals. The parameters of these
Since extraterrestrial radiation is deterministic this gives the same distributions are estimated using the EM-algorithm where the 𝑘-means
summer period for any year and does not depend on data. This led algorithm is employed to initialize a grouping of data into the different
to a summer definition that again is the same across locations (up to a components. This is a standard method of estimating Gaussian mixture
maximum difference of one day) and gave similar empirical results. We distributions, see e.g. McLachlan and Peel (2000), and we have used the
prefer to use our empirically motivated definition since it is the same Matlab routine fitgmdist. The estimates are presented in Table 3. In our
for all locations, transparent, and yielding a slightly better model fit. model setup we fit different Gaussian mixture distributions for the sum-
In the right panel of Fig. 7 we show the skewness function (13) im- mer and winter residuals. At the model level these distributions should
plied by our model compared to the skewness estimates from residual have zero expected value and unit variance to avoid introducing a small
data 𝑒(𝑡). Our model generates a skewness function that is constant in time-variation in the mean and variance. In empirical studies there is
the summer and winter periods, and explains rather well the overall nearly always some mismatch between the empirical moments of a
skewness pattern of the data. Comparing the model implied skewness residual series and the corresponding moments implied by the fitted
with the empirical estimates for the summer and winter periods we distribution. Such deviations usually have no empirical consequences

8
K. Larsson et al. Energy Economics 117 (2023) 106421

Fig. 7. Empirical skewness per month plotted against the day of year. Left panel plots skewness estimated for radiation data 𝐺(𝑡) (green) and for residuals 𝑒(𝑡) (blue). Right panel
show estimated skewness for 𝑒(𝑡) (blue) and the skewness function obtained from the estimated model (orange). (For interpretation of the references to color in this figure legend,
the reader is referred to the web version of this article.)

Fig. 8. Histogram (normalized) for residuals 𝑒(𝑡) and corresponding kernel density estimate (red curve).

since they typically are very small. We nevertheless employ a simple procedure have negligible effects on data since the mean and variance
practical procedure to ensure that the estimated distributions have are already close to zero and one, and it does not affect the skewness.
expected value of zero and unit variance. The EM-algorithm generates We have compared estimation results obtained with and without this
parameter estimates that closely matches the first three moments of extra standardization and the differences in parameter estimates are
the residuals. In order to cancel out any small deviations from zero very small and does not affect the distributional fit to any significant
mean and unit variance in the data, we estimate the Gaussian mixtures degree. We still keep it since it gives consistency with regard to the
on residuals modified by an extra standardization using the empirical model and its components. The procedure ensures that the estimated
mean and standard deviation. By virtue of the EM-algorithm this minor distributions for 𝜖𝑆 (𝑡) and 𝜖𝑊 (𝑡) have expected values of zero and
adjustment gives us parameter estimates that imply theoretical zero unit variances, and hence by implication that also E[𝑒(𝑡)] = 0 and
expected values and unit variances to very close approximation. This Var (𝑒(𝑡)) = 1.

9
K. Larsson et al. Energy Economics 117 (2023) 106421

Fig. 9. Histogram (normalized) for winter (left panel) and summer (right panel) residuals 𝜖𝑊 (𝑡) and 𝜖𝑆 (𝑡) respectively.

The random variables 𝜖𝑆 (𝑡) and 𝜖𝑊 (𝑡) follow Gaussian mixture simulated 10-year sample of 𝑒(𝑡) generated by the estimated model.
distributions with two components and thus have the stochastic rep- From Fig. 9 we know that our model is highly consistent with the
resentations, empirical distributions of summer and winter residuals, and Fig. 10
shows that the model is also consistent with the histogram of 𝑒(𝑡). The
𝜖𝑘 = 𝐵𝑘 𝑌1,𝑘 + (1 − 𝐵𝑘 )𝑌2,𝑘 , 𝑘 ∈ {𝑆, 𝑊 }
fit in Fig. 10 is rather remarkable given that the model is fitted to
where 𝐵𝑘 is a Bernoulli random variable with parameter 𝑞𝑘 (i.e. P(𝐵𝑘 = summer and winter residuals separately and not on the observed 𝑒(𝑡),
1) = 𝑞𝑘 ), and the components 𝑌1,𝑘 are Gaussian random variables. It and the fact that the kernel estimate representing the model is obtained
is natural to interpret the Gaussian components 𝑌1 and 𝑌2 as states from a single run of simulated data. Different simulations will produce
of either cloudy or clear sky conditions respectively. We consistently different levels of agreement but we find the fit to be quite stable across
estimate one Gaussian component with a negative mean and one with a simulations.
positive mean. The interpretation is that the component with negative As a final remark on the estimation results we emphasize that in
mean corresponds to the state of cloudy conditions and we refer this addition to providing a convincing representation of the irradiation
state to component 1 in both summer and winter. Component 2 is then time series our model is highly tractable. The model structure is both
representing the state for clear sky conditions and the parameter 𝑞𝑘 is simple and transparent. It is based on standard time series model
the a priori probability of being in the cloudy state. components and it is very easy to estimate and simulate.
Fig. 9 shows the normalized histograms of the summer and winter
residuals together with the Gaussian mixture densities implied by the 3.2. Validation
parameter estimates. The fit is excellent in both cases and the graphs
display interesting differences between the summer and winter periods. We validate our proposed model along two different dimensions.
Both densities are bimodal but with shoulder-like appearances having First, we investigate the model’s ability to represent irradiation data
one peak significantly higher than the other. For the winter period at other locations. This part is done by estimating the model on the
the mixture distribution has a positive skewness and a significant left remaining four locations in our dataset. We present and discuss the
peak indicating that the state of cloudy conditions is dominant and overall model performance at these locations. With residuals obtained
carries more probability mass. The opposite situation is seen for the for the summer and winter periods at each location we also calculate
summer period which has a negative skewness and a significant right empirical correlations between the different sites. Second, we perform
peak with the probability mass shifted towards the clear sky state. The a standard validation exercise where we study day-ahead predictions
estimate of the a priori probability for the cloudy state is 𝑞𝑆 = 0.3912 and associated confidence intervals.
for the summer, which corresponds to a 0.6088 probability for the clear Parameter estimates for the four remaining locations are presented
sky state. In the winter the probability for the cloudy state is higher in Table 4. We use the PACF of 𝑍(𝑡) = 𝐺(𝑡)−𝑆(𝑡) to determine the order
and estimated to 𝑞𝑊 = 0.5687. We note that the underlying Gaussian 𝑝 in the AR-part (5) of the model. The AR-order is 𝑝 = 2 for Hamburg,
distributions for the cloudy state has a higher variance in summer while Berlin and Stuttgart but for Nürnberg and München an order of 𝑝 = 1
the opposite is true in the winter. This is also a natural result consistent is found to be sufficient. Overall, the model fit is similar across all
with irradiation levels being more dispersed and variable in cloudy locations and with rather small differences in parameter estimates. We
conditions during summer. observe the same basic features as presented for the Hamburg location
In Fig. 10 we show that our model is capable of reproducing the only with slight variations.
histogram of the observed series 𝑒(𝑡). In Fig. 10 we have again plotted Our definition of the summer and winter periods is based on the
the normalized histogram for 𝑒(𝑡) together with the kernel density months where we observe negative empirical skewness. These periods
estimate, but also added the kernel density estimate obtained from a are the same for all locations. Fig. 11 plots the empirical monthly

10
K. Larsson et al. Energy Economics 117 (2023) 106421

Fig. 10. Histogram (normalized) for residuals 𝑒(𝑡) and corresponding kernel density estimate (red curve). The green curve is the kernel density estimate obtained from simulated
𝑒(𝑡) from the estimated model. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

skewness together with the skewness function implied by the parameter Table 5
Correlation estimates calculated from summer residuals.
estimates for all locations. The skewness patterns are very similar across
locations. Hamburg Berlin Stuttgart Nürnberg Nürnberg

In Fig. 12 we plot the histogram of the residual series 𝑒(𝑡) together Hamburg 1.0000 0.4165 0.1626 0.1786 0.0809
Berlin 1.0000 0.1794 0.2833 0.1473
with superimposed kernel density estimates both from data and simu-
Stuttgart 1.0000 0.6581 0.5488
lations constructed from the estimated models. The bimodal mixture Nürnberg 1.0000 0.6276
distributions provide an excellent fit in all cases. There are some München 1.0000
notable differences in the histograms. Hamburg and Berlin display two
peaks of comparable height while the southern locations Nürnberg and
Table 6
München exhibit a significantly lower left peak. Correlation estimates calculated from winter residuals.
The estimation results from the different locations show that the Hamburg Berlin Stuttgart Nürnberg Nürnberg
model works well for all the selected coordinates thus providing further
Hamburg 1.0000 0.3458 0.1167 0.1299 0.0195
validation. Indeed, irradiation data at all locations display the same fea- Berlin 1.0000 0.1977 0.2463 0.1697
tures and are in fact very similar with regard to underlying stochastic Stuttgart 1.0000 0.5579 0.5262
and seasonal properties. The small differences in the AR-order, ranging Nürnberg 1.0000 0.5874
between 𝑝 = 1 and 𝑝 = 2, is also a sign of the model specification München 1.0000

being stable. We conclude that the model structure we have proposed is

stable across locations and provides a satisfactory fit at each coordinate. Table 7
These results indicate that the model structure should be relevant Distances (in km) between locations.
for Germany and most likely for large parts of Europe that share a Hamburg Berlin Stuttgart Nürnberg Nürnberg
similar climate. However, other climate types may certainly display Hamburg 0.0 281.7 501.7 456.9 568.2
quite different properties of solar irradiation and require modifications Berlin 0.0 512.0 405.1 492.1
to the model structure. Stuttgart 0.0 130.3 161.7
Nürnberg 0.0 111.9
The estimated models allow us to identify the summer and winter
München 0.0
residuals 𝜖𝑆 (𝑡) and 𝜖𝑊 (𝑡) for each coordinate. Since these are assumed
iid we can calculate empirical (sample) correlations for the summer and
winter periods between the different locations. These correlations are
reported in Tables 5 and 6. As can be expected the correlations decrease [ 𝑡 at time
based on the information ] 𝑡, is given by the conditional
with physical distance. The distances (in km) are reported in Table 7. ̂ + 1) = E 𝐺(𝑡 + 1)|𝑡 . From (1) and (5) we have
expectation 𝐺(𝑡
Taking Hamburg as a reference point we see that e.g. the summer corre-
∑
𝑝
lations are steadily decreasing from a correlation of 0.4165 with Berlin 𝐺(𝑡 + 1) = 𝑆(𝑡 + 1) + 𝛽𝑖 𝑍(𝑡 + 1 − 𝑖) + 𝜎(𝑡 + 1)𝑒(𝑡 + 1)
to 0.0809 with München. The southern locations Stuttgart, Nürnberg 𝑖=1
and München are more closely situated and display higher correlations. from which it follows that
We find the highest correlation in summer to be 0.6581, found between ∑
𝑝
Nürnberg and Stuttgart, while the highest winter correlation is 0.5874 ̂ + 1) = 𝑆(𝑡 + 1) +
𝐺(𝑡 𝛽𝑖 𝑍(𝑡 + 1 − 𝑖)
found between Nürnberg and München. 𝑖=1
[ ]
We also study one-day-ahead predictions of the irradiation level using that E 𝜎(𝑡 + 1)𝑒(𝑡 + 1)|𝑡 = 0 since 𝜎(𝑡+1) is known at 𝑡 and 𝑒(𝑡+1)
using as out-of-sample the year 2020. The purpose of our prediction is independent of 𝑡 . From (5) we have that 𝑍(𝑡+1−𝑖) is known at time
study is to confirm that the model produces sensible predictions and 𝑡 for 𝑖 > 0 and can be expressed in terms of 𝑆(𝑡 + 1 − 𝑖) and 𝑍(𝑡 + 1 − 𝑖).
associated confidence intervals. The predicted irradiation at time 𝑡 + 1 The one-day predictions are thus easily calculated. Confidence intervals

11
K. Larsson et al. Energy Economics 117 (2023) 106421

Fig. 11. Empirical monthly skewness estimates (blue) and model fitted skewness functions (red) for all locations. (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)

Fig. 12. Histogram (normalized) for residuals 𝑒(𝑡) and corresponding kernel density estimate (red) for all locations. The green curves are the kernel density estimates obtained
from simulated 𝑒(𝑡) from the estimated models. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

̂ + 1) can be calculated from simulations. To simulate from the

for 𝐺(𝑡 We illustrate the predictions for the location near Hamburg. In
conditional distribution of 𝐺(𝑡 + 1) at 𝑡 we only need to draw random Fig. 13 we have plotted data and one-day predictions for each day in
numbers from the distribution of 𝑒(𝑡 + 1). This amounts to simulating 2020 together with the estimated 95% confidence bounds, calculated
from a Gaussian mixture distribution, either from 𝜖𝑆 (𝑡 + 1) or 𝜖𝑊 (𝑡 + 1) from using 𝑀 = 50,000 simulations each day. The prediction errors
depending on the value of 𝑑(𝑡 + 1), and can be done very efficiently and the length of the confidence intervals varies with the seasonal
using standard methods. From a simulated sample of 𝑀 observations conditional volatility. In the Hamburg case the confidence bounds are
from 𝑒(𝑡+1) we calculate the empirical quantiles 𝜆𝛼∕2 and 𝜆1−𝛼∕2 to form exceeded 3.84% of the 365 days consistent with the assumed confi-
̂ + 1).
100(1 − 𝛼)% confidence intervals for 𝐺(𝑡 dence level of 5%. Predicting irradiation levels from one day to the next

12
K. Larsson et al. Energy Economics 117 (2023) 106421

Fig. 13. Time series of irradiation data 𝐺(𝑡) (blue), and day-ahead predictions 𝐺(𝑡)̂ (green) with 95% confidence bounds (red), for the out-of-sample period of 2020. Hamburg
location. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

is a challenging task given the large day-to-day variations present in of 20%, a reference temperature of 𝑇𝑟𝑒𝑓 = 25 degrees Celsius, and a
data. Given our statistical model setup, based only on irradiation data temperature coefficient of 0.5%. This corresponds to 𝜅1 = 0.2 and 𝜅2 =
itself, the predictions are reasonably accurate. The mean prediction 0.005. Under the given assumptions, and a solar irradiation of 1000 W
error in summer is −12.70 and in winter it is 8.06. These numbers per square meter, a 50 000 m2 solar park produces 10 MWh of electricity
are small compared to the mean irradiation levels of 517.53 and per hour for a constant operating temperature at 25 degrees Celsius.
131.88 and correspond to −2.45% and 6.11% in summer and winter This situation is a slight variation of the example studied in De Jong
respectively. (2020).
At time 𝑡 the produced volume (in MWh) from this solar park is
4. Application to PV risk management in power market
𝑉 (𝑡) = 0.05 × ℎ(𝐺(𝑡), 𝑇 (𝑡))
Production volumes from renewable generation assets cannot be We treat the temperature as a deterministic external input. The impact
perfectly forecasted due to the intermittent and stochastic fluctuations of temperature on produced volumes is small compared to irradiation
in solar irradiation. The deviation between the forecasted day-ahead and its contribution to the variance of production even smaller. Con-
production and the actual production typically gives rise to an imbal- ditional on time 𝑡 information the distribution function of 𝑉 (𝑡 + 1)
ance cost component, which impacts the producer in the imbalance is
market, see De Jong and Kovaleva (2021) for a detailed discussion. For
reasons of operational risk management and for investment decisions, it 𝐹𝑡 (𝑥) = P𝑡 (𝑉 (𝑡 + 1) ≤ 𝑥) = P𝑡 (𝐺(𝑡 + 1)𝐻(𝑡 + 1) ≤ 𝑥)
is therefore of practical interest to properly understand the distribution [ ( )]
where we denote 𝐻(𝑡) = 0.05𝜅1 1 − 𝜅2 𝑇 (𝑡) − 𝑇𝑟𝑒𝑓 . It is straightfor-
of day-ahead production volume as this is the major determinant of
ward to show that in our proposed model
future income streams. In analogy with the commonly used Value-at-
( ) ( )
Risk measure used in financial markets we define the Production-at- 𝐹𝑡 (𝑥) = 𝑞𝑘 𝐹1,𝑘 𝑦𝑡 (𝑥) + (1 − 𝑞𝑘 )𝐹2,𝑘 𝑦𝑡 (𝑥) (15)
Risk (P@R) as a quantile of the distribution for day-ahead production
volume. where 𝑘 = 𝑆 if 𝑑(𝑡+1) = 1 and 𝑘 = 𝑊 otherwise, 𝑦𝑡 (𝑥) is the transformed
We apply the proposed model to estimate the P@R for a hypotheti- point
( )
cal PV solar park by simulating the future solar irradiation distribution 1 𝑥 ∑𝑝

and transforming it to production volumes by using a production 𝑦𝑡 (𝑥) = − 𝑆(𝑡 + 1) − 𝛽𝑗 𝑍(𝑡 + 1 − 𝑗) (16)
𝜎(𝑡 + 1) 𝐻(𝑡 + 1) 𝑗=1
function. Several mathematical models of PV production functions have
been developed and tested for different solar cell technologies, see and 𝐹𝑚,𝑘 (𝑥) are the corresponding Gaussian distribution functions for
e.g. Huld et al. (2011) and Kaldellis et al. (2014). We employ a simple components 𝑚 = 1 and 𝑚 = 2. Note that at time 𝑡 all the terms
linear model which is also used in De Jong (2020) and in Kaldellis et al. in (15) are known; they are either deterministic or can be computed
(2014). The PV power production per square meter, ℎ(𝐺, 𝑇 ), is defined from historic values in our model. The conditional distribution function
as a function of the solar irradiation 𝐺 and the operating temperature in (15) is itself useful for risk management allowing probabilities for
𝑇 according to the expression different scenarios to be easily calculated.
[ ( )] For a given value of 𝛼 ∈ (0, 1) the day-ahead P@R at time 𝑡, denoted
ℎ(𝐺, 𝑇 ) = 𝜅1 𝐺 1 − 𝜅2 𝑇 − 𝑇𝑟𝑒𝑓 (14)
𝜆(𝑡, 𝛼), fulfills the relation
where the parameter 𝜅1 is the efficiency of transforming solar energy
𝐹𝑡 (𝜆(𝑡, 𝛼)) = P𝑡 (𝑉 (𝑡 + 1) ≤ 𝜆(𝑡, 𝛼)) = 𝛼 (17)
into electrical power at the reference temperature 𝑇𝑟𝑒𝑓 , and 𝜅2 is the
temperature coefficient, which is determined by the cell technology and The P@R value 𝜆(𝑡, 𝛼) must be solved for numerically from (17). This
the reference temperature. In this application we assume an efficiency can be done using simulation or a standard root-finding algorithm

13
K. Larsson et al. Energy Economics 117 (2023) 106421

Fig. 14. Production-at-Risk at 10% (blue) and 5% (red) for a 50 000 m2 solar park at the Hamburg location for year 2020. Included are also predicted day-ahead production
volume (orange) and seasonal volume (green). All numbers expressed in MWh. (For interpretation of the references to color in this figure legend, the reader is referred to the
web version of this article.)

where (15) and (16) are used to express 𝐹𝑡 (𝜆(𝑡, 𝛼)) in (17). In our exam- calculated for May 28 were 3.59, at 10%, and 3.05 at 5%. In contrast,
ple we focus on the left tail of the distribution but a corresponding P@R the lowest production in 2020 is 0.45 MWh recorded on Dec 23 with
for the right tail can be defined similarly with obvious modifications. P@R values calculated to 0.41, at 10%, and 0.35, at 5%. The difference
The P@R has a clear dependence on the volatility and shape of the in P@R for the 10% and 5% levels on the day of maximum production
distribution of solar irradiation data. As demonstrated in this paper is 0.55 which is substantial in terms of MWH. On the day of minimum
the solar irradiation distribution exhibits a complicated time structure, production the difference is significantly smaller at 0.07 MWh. These
which makes it a non-trivial task to dynamically estimate the P@R. differences are in fact rather typical for the summer and winter periods
We illustrate this by calculating P@R for the hypothetical solar park and are caused by the different stochastic and seasonal properties of
described above on each day in our out-of-sample data for the year solar irradiation over the year.
2020. The solar park is assumed to be located at the Hamburg coordi-
nate and we use the parameter estimates in Table 3 in the calculations. 5. Concluding remarks
The impact of the temperature on the variation of produced volume
is minor and for simplicity we cancel it from our calculations by We have proposed a stochastic time series model for the daily
assuming 𝑇 = 𝑇𝑟𝑒𝑓 = 25. This is obviously unrealistic, especially for (around noon) irradiance in 5 different cities in Germany. Our model
the winter period, but the point of this exercise is to illustrate the is based on a careful analysis of the stylized facts of irradiance levels
temporal behavior of the P@R on which the temperature can only observed in a long time series of data recordings. Based on astronomical
have a small influence compared to irradiation. Fig. 14 show the knowledge of the sun’s position to the earth, we have explained the
P@R calculated at the 5% and 10% levels for each day in 2020. The mean variation in irradiance by a clear sky model, while the first
predicted and seasonal volumes are also plotted for reference. These are order stochastic effects are captured by an autoregressive time series
obtained from plugging the predicted, 𝐺(𝑡 ̂ + 1), and the seasonal 𝑆(𝑡 + 1) of order 2. The residuals show seasonality in both their variance and
irradiation, into the production function. The P@R-levels in Fig. 14 skewness, where the latter effect has to the best of our knowledge
display significant seasonal variations both in the level and variance. never been observed nor modeled effectively. We suggest a summer
The P@R values naturally follows the seasonality in the irradiation and winter parameter changing between two different regimes. The
level. It is also clearly visible that the P@R is more variable in the variance is modeled as the product of a seasonally varying function
summer. It can also be noted that the 5% and 10% P@R values are and a GARCH(1,1)-model. The standardized residuals turn out to be
much closer during winter compared to summer. This is an effect of uncorrelated Gaussian mixture models, where the bimodality of the
the seasonal dependencies present in the distribution of irradiation. distributions in winter and summer can be attributed to the clouds’
The P@R values plotted in Fig. 14 were calculated using simulation. interference with the sun.
The actual production volumes exceeded the 10% and 5% P@R during Our proposed model is estimated to daily data, and validated in a
2020 at 36 and 14 instances respectively. This corresponds to 9.86% prediction study. It is demonstrated that the model captures well the
and 3.84% of the days which is in close agreement with the selected stylized features of irradiance in the studied locations. We presented an
probability levels. application to energy markets, where we show how PV-producers may
To give some numerical insight we compare the P@R on three use the model to assess their risk based on the measure Production at
different dates. On March 16 the actual produced volume is 4.44 MWh Risk.
and the 10% and 5% P@R are 2.22 and 1.84. This means that the Our model and analysis can be extended in several directions. A
probabilities that the next day production falls below 1.84 and 2.22 natural first step is to move towards a higher time resolution and
MWh are 5% and 10% respectively. The 10% P@R corresponds to a consider an hourly stochastic time series model. Such a model will be
50% drop in production from the 16th to the 17th. The maximum interesting for applications in the intraday power market, where PV
production in 2020 occurs on May 28 at 8.68 MWh. The P@R values producers say, can hedge their exposure in the day ahead market. We

14
K. Larsson et al. Energy Economics 117 (2023) 106421

expect an hourly model to lead to further challenges in capturing a Declaration of competing interest
daily seasonality profile, in particular around the hours of sunrise and
sunset. Most likely, the memory effect in the time series model will The authors declare that they have no known competing finan-
show a higher order of autoregression. cial interests or personal relationships that could have appeared to
In our study we have considered the correlation across 5 locations influence the work reported in this paper.
in Germany. Several interesting applications call for a spatial stochastic
model, which is another interesting extension of our proposed irra- Acknowledgments
diation dynamics. Our analysis indicates a declining correlation of
irradiance residuals with distance. A spatial model must detect the We thank two anonymous referees, Pascal Heider, Kamil Kladivko
spatial correlation structure in terms of distance to define a random and seminar participants at EnBW Germany for helpful comments
field model (see Cressie and Wikle (2011)). Moreover, one must also be on the paper. F. E. Benth acknowledges financial support from the
able to describe the other parameters of the model as functions of their thematic research group SPATUS funded by UiO:Energy, University of
geographic coordinates. Such a model will be very useful in studies of Oslo.
finding the optimal location of a PV installation, or in combining PV
sites in a risk hedging study analogous to Benth et al. (2021). Appendix A
The irradiation in a location is sensitive to variations in the cloud
cover. Our model does not explicitly take into account the effect of the In this appendix we describe how the solar hour angle 𝑤 and decli-
shifting cloud cover over a location. We remark that including cloud nation 𝛿 are determined. We give only a brief description and we refer
cover data for specific coordinates would likely have to use different to Duffie and Beckmann (2013) where more details and explanations
data sources and that it then may prove difficult to reach adequate can be found. The solar hour angle is given by
consistency between the series. Furthermore, variation in cloud cover
is closely connected with the wind field at cloud level, which also 𝑤 = 15◦ (𝐿𝑆𝑇 − 12)
correlates with temperature. The power curve of a PV-panel depends where 𝐿𝑆𝑇 is the local solar time. The local solar time 𝐿𝑆𝑇 is obtained
on ground temperature. It is a challenge to pin down realistic stochastic from the local time 𝐿𝑇 and a time correction 𝑇 𝐶 as
models combining all these factors.
Intraday solar production depends on the movements of the solar 𝐿𝑆𝑇 = 𝐿𝑇 + 𝑇 𝐶∕60.
irradiation over the course of the day, reaching its peak around midday.
The time correction 𝑇 𝐶 is calculated from
Unfortunately, this pattern does not coincide well with the demand
profile for electricity, which typically peaks in the morning/evening 𝑇 𝐶 = 4(𝐿𝑆𝑇 𝑀 − 𝑙𝑜𝑛) + 𝐸
due to the increased human activities at these times of the day. This
where 𝐿𝑆𝑇 𝑀 = 15◦ 𝑈 where 𝑈 is the offset to UTC time (U = +1 in
supply/demand discrepancy leads to a situation where the net demand
Central European Time (CET)), 𝑙𝑜𝑛 is the longitude, and 𝐸 denotes the
(demand minus renewable production) falls into a slump during the
so called Equation of Time given by
midday hours, with relatively low demand and high solar generation,
and in the morning/evening it sharply peaks due to high demand and 𝐸 = 229.2 (0.000075 + 0.001868 cos(𝐵) − 0.032077 sin(𝐵)
low solar generation. The intraday net demand visually resembles the − 0.014615 cos(2𝐵) − 0.04089 sin(2𝐵))
shape of a duck and is commonly known as the ‘‘duck curve’’. The
duck curve points to a problematic situation since it originates from a with 𝐵 = (𝑛−1)360∕365 where 𝑛 is the yearday which may be expressed
supply/demand mismatch, which needs to be compensated by conven- as a continuous variable.
tional production technologies, such as gas or coal, and their flexibility The declination 𝛿 is directly given by
comes at a high economic cost. With an increasing share of renewable ( )
360(284 + 𝑛)
production in the power system the duck curve is, ceteris paribus, likely 𝛿 = 23.45◦ sin
365
to be even more pronounced in the future. This makes a challenge
There are more accurate, and complicated, expressions available to
for the renewable expansion and several solutions to the problem are
determine the Equation of Time 𝐸 and the declination 𝛿, see Duffie
currently being explored. One promising alternative, which is presently
and Beckmann (2013). The formulas presented here are sufficient for
investigated and implemented, is to store the oversupply from the
our purposes since they are only used to construct the seasonal function
midday solar production in short-term batteries in order to flatten
that we calibrate to data.
the duck curve. In fact, the stochastic control problem for battery
optimization requires a stochastic solar irradiation model as input data,
Appendix B
which is an important motivation for the model proposed in this paper.
We leave these interesting and challenging questions of potential
In this appendix we provide details on the parameter restriction for
model extensions and applications to future studies.
the seasonal variance. The seasonal variance part can be written

CRediT authorship contribution statement 𝜎𝑆2 (𝑡) = 𝑐0 + 𝑐1 cos(𝑘𝑡) + 𝑐2 sin(𝑘𝑡)

where 𝑘 = 2𝜋∕365. Using the trigonometric relation

Karl Larsson: Conceptualization, Data curation, Formal analysis,
Investigation, Methodology, Project administration, Resources, Soft- sin(𝑘𝑡 + 𝑧) = cos(𝑘𝑡) sin(𝑧) + sin(𝑘𝑡) cos(𝑧)
ware, Supervision, Validation, Visualization, Writing – original draft, we can write
Writing – review & editing. Rikard Green: Conceptualization, Investi-
gation, Resources, Methodology, Supervision, Validation, Visualization, 𝜎𝑆2 (𝑡) = 𝑐0 + 𝑐 sin(𝑘𝑡 + 𝑧)
Writing – original draft, Writing – review & editing. Fred Espen Benth: √
Conceptualization, Formal analysis, Investigation, Resources, Method- where 𝑐 = 𝑐12 + 𝑐22 and 𝑧 is the number such that tan(𝑧) = 𝑐1 ∕𝑐2 , 𝑐2 ≠ 0.
The smallest possible value of 𝜎𝑆2 (𝑡) is therefore 𝑐0 −𝑐 and the restriction
ology, Supervision, Validation, Visualization, Writing – original draft, √
Writing – review & editing. 𝑐0 > 𝑐12 + 𝑐22 thus ensures positivity.

15
K. Larsson et al. Energy Economics 117 (2023) 106421

Appendix C. Supplementary data Duffie, J.A., Beckmann, W.A., 2013. Solar Engineering of Thermal Processes, fourth ed.
John Wiley & Sons..
Engle, R.F., Gonzlaez-Rivera, G., 1991. Semiparametric ARCH models. J. Bus. Econom.
Supplementary material related to this article can be found online
Statist. 9, 345–359.
at https://doi.org/10.1016/j.eneco.2022.106421. Gschwind, B., Wald, L., Blanc, P., Lefevre, M., Schroedter-Homscheidt, M., Arola, A.,
2019. Improving the McClear model estimating the downwelling solar radiation at
References ground level in cloud free conditions – McClear-v3. Meteorol. Z. 28 (2), 147–163.
http://dx.doi.org/10.1127/metz/2019/0946.
Alaton, P., Djehiche, B., Stillberger, D., 2002. On modelling and pricing weather Härdle, W., Lopez Cabrera, B., 2012. The implied market price of weather risk. Appl.
derivatives. Appl. Math. Finance 9 (1), 1–20. Math. Finance 19 (1), 59–95.
Benth, F.E., Christensen, T.S., Rohde, V., 2021. Multivariate continuous-time modeling Hottel, H.C., 1976. A simple model for estimating the transmittance of direct solar
of wind indexes and hedging of wind risk. Quant. Finance 21 (1), 165–183. radiation through clear atmospheres. Sol. Energy 18 (2), 129–134. http://dx.doi.
Benth, F.E., Ibrahim, N.A., 2017. Stochastic modeling of photovoltaic power generation org/10.1016/0038-092X(76)90045-1.
and electricity prices. J. Energy Mark. 10 (3), 1–33. http://dx.doi.org/10.1080/ Huld, T., Friesen, G., Skoczek, A., Kenny, R.P., Sample, T., Field, M., Dunlop, E.D.,
17442508.2016.1177057. 2011. A power-rating model for crystalline silicon PV modules. Sol. Energy Mater.
Benth, F.E., Šaltytė Benth, J., 2012. A critical view on temperature modelling for Sol. Cells 95 (2011), 3359–3369.
application in weather derivatives markets. Energy Econ. 34, 592–602. Kaldellis, J.K., Kapsali, M., Kavadias, K.A., 2014. Temperature and wind speed
Benth, F.E., Šaltytė Benth, J., 2013. Modeling and Pricing in Financial Markets for impact on the efficiency of PV installations. Experience obtained from outdoor
Weather Derivatives. World Scientific. measurements in Greece. Renew. Energy 66, 612–624, 2014.
Benth, F.E., Šaltytė Benth, J., Koekebakker, S., 2008. Stochastic Modelling of Electricity Lefevre, M., Oumbe, A., Blanc, P., Espinar, B., 2017. Mcclear: a new model estimating
and Related Markets. World Scientific. downwelling solar radiation at ground level in clear-sky conditions. Atmos Meas
Campbell, S.D., Diebold, F.X., 2005. Weather forecasting for weather derivatives. J. Tech 6, 2403–2418. http://dx.doi.org/10.5194/amt-6-2403-2013.
Amer. Statist. Assoc. 100 (469), 6–16. Lingohr, D., Müller, G., 2019. Stochastic modelling of intraday photovoltaic power
CAMS, 2019. User Guide To the CAMS Radiation Service. ECMWF Copernicus report, generation. Energy Econ. 81, 175–186.
available at https://atmosphere.copernicus.eu, last accessed 2021-04-19. Marchand, M., Saint-Drenan, Y.-M., Saboret, L., Wey, E., Wald, L., 2020. Performance
Casula, L., D’Amico, G., Masala, G., Petroni, F., 2020. Performance estimation of of CAMS radiation service and HelioClim-3 databases of solar radiation at surface:
photovoltaic energy production. Lett. Spatial Resour. Sci. 13, 267–285. evaluating the spatial variation in Germany. Adv. Sci. Res. 17, 143–152. http:
Cressie, N., Wikle, C.K., 2011. Statistics for Spatio-Temporal Data. John Wiley & Sons. //dx.doi.org/10.5194/asr-17-143-2020.
Cuppari, R.I., Higgins, C.E., Characklis, G.W., 2021. Agrivoltaics and weather risk: a McLachlan, G., Peel, D., 2000. Finite Mixture Models. Wiley.
diversification strategy for landowners. Appl. Energy 291, 1–16. http://dx.doi.org/ Qu, Z., Oumbe, A., Blanc, P., Espinar, B., Gesell, G., Gschwind, B., Klüser, L.,
10.1016/j.apenergy.2021.116809. Lefevre, M., Saboret, L., Schroedter-Homscheidt, M., Wald, L., 2017. Fast radiative
De Jong, C., 2020. Production Patterns of Wind and Solar. Kyos Energy Consulting, transfer parameterisation for assessing the surface solar irradiance: The Heliosat-4
available at https://www.kyos.com/wp-content/uploads/2020/05/Production- method. Meteorol. Z. 26, 33–57. http://dx.doi.org/10.1127/metz/2016/0781.
patterns-of-wind-and-solar-the-financials-of-renewable-power-and-PPA- Yang, D., Bright, J.M., 2020. Worldwide validation of 8 satellite-derived and reanalysis
contracts.pdf. solar radiation products: a preliminary evaluation and overall metrics for hourly
De Jong, C., Kovaleva, S., 2021. Short Term Forecasting and Imbalance Costs. Kyos data over 27 years. Sol. Energy 210, 3–19. http://dx.doi.org/10.1016/j.solener.
Energy Consulting, available at https://www.kyos.com/wp-content/uploads/2021/ 2020.04.016.
02/Imbalance-the-financials-of-renewable-power-and-PPA-contracts.pdf.

16