1

Modeling Solar Irradiance for Short-term

Dynamic Analysis of Power Systems
Guðrún Margrét Jónsdóttir, Student Member, IEEE and Federico Milano, Fellow, IEEE
School of Electrical & Electronic Engineering, University College Dublin, Ireland
[email protected], [email protected]

Abstract—This paper presents and compares four solar ir- of the sun. This trajectory can be accurately predicted based
radiance models for short-term power system analysis. Three on location, time of year and day using clear-sky irradiance
of these models can be found in the literature but are based models [5], [6]. The stochastic variability is dependent on
on approximations and shortcomings, which are duly discussed.
The fourth model is novel and is formulated through stochastic the cloud coverage and can be expressed using the clear-sky
differential equations with jumps. The case study illustrates index (the ratio between the global solar radiation and the
the ability of the proposed solar irradiance model to generate corresponding clear-sky radiation).
synthetic processes that reproduce the stochastic properties of In the dynamic analysis of power systems with PV gen-
flickering events taken from measurement data. eration, the solar irradiance is either assumed to be constant
Index Terms—Solar irradiance, clear-sky index, Poisson pro- [7], [8] or to vary with random steps [9]–[11]. These models
cess, stochastic differential equations. do not capture the actual intermittency of the solar irradiance.
Measurement data has to be utilized to build more accurate
I. I NTRODUCTION models.
A. Motivation In [12], the solar irradiance variability is modeled by com-
bining a Poisson jump process and an Autoregressive Moving
With an installed capacity that has doubled in the last three
Average (ARMA) model. Stochastic Differential Equations
years, solar Photovoltaic (PV) generation is the fastest growing
(SDEs) with jumps are defined in [4] for modeling the clear-
energy source in power systems worldwide [1]. Solar genera-
sky index. Both these methods define model parameters based
tion is renewable and eco-friendly but also highly volatile due
on measurements. However, these models do not adequately
to the position of the sun and clouds changing [2]. The impact
capture the volatility of the solar irradiance in the seconds to
of solar generation fluctuations on the dynamic behavior
minutes time scale.
of power systems has not been thoroughly investigated and
remains a relevant research question. Accurate models are
required to represent the solar irradiance fluctuations in power C. Contributions
system simulations. This paper addresses this modeling need. The contributions of this paper are twofold.
• To describe three models of solar irradiance volatility
B. Literature Review
previously presented in [10], [12], and [4], respectively,
The output of PV solar power plants naturally changes and discuss the shortcomings of such models.
throughout the day because of the daily path the sun follows • To propose a novel model of the clear-sky index for short-
across the sky. During sunrise and sunset the output of the PV term dynamic analysis.
plant will change by about 10% in just 15 min. The daily sun
path can be easily and accurately predicted. On the other hand, In the proposed model, clear-sky conditions are represented
PV power plants are also a significant source of intermittency through a SDE and the jumps caused by cloud movements
due to cloud coverage. The change in solar irradiance caused are simulated with two jump diffusion processes. The paper
by cloud movement can be over 60% of the peak irradiance also shows how the proposed approach overcomes the issues
within a few seconds [3]. These variations can be smoothed of the other models.
and their transient effects minimized if considering a large PV
power plant or an aggregated model of several geographically D. Organization
distributed plants [2]. However, if a single PV power plant The reminder of this work is organized as follows. Sec-
covers a relatively small area, e.g., in microgrid applications, tion II describes the solar irradiance measurement data utilized
its power output fluctuations have to be properly modelled [4]. throughout the paper and outlines the modeling of solar
Based on the discussion above, solar irradiance variations irradiance. Section III presents four solar irradiance models
can be divided into a deterministic component and a stochas- and Section IV shows how the proposed model can be used
tic one. The deterministic component is the variations at a to generate synthetic solar irradiance processes that accurately
minutely or hourly scale due to the daily apparent movement capture the intermittent behaviour of real-world data. Finally,
Guðrún Margrét Jónsdóttir and Federico Milano are supported by Science Section V draws conclusions and outlines relevant areas for
Foundation Ireland (SFI) under Grant No. SFI/15/IA/3074. future work.

978-1-7281-1981-6/19/$31.00 ©2019 IEEE

 2

II. M ODELING OF S OLAR I RRADIANCE

This section presents the procedure to identify the de-
terministic and stochastic part of the solar irradiance from 1.4 09.01.2011
10.01.2011
measurement data. With this aim and for further analysis in 1.2 11.01.2011
the paper, the solar irradiance data collected by the National
Renewable Energy Laboratory (NREL), gathered in Kalaeloa, 1.0

Clear sky index

Hawaii, are used [13]. This data set consists of one year of 0.8
measurements gathered with a 1 Hz sampling frequency from
April 2010 to March 2011. Each day consists of measurements 0.6
from 5am to 8pm.
0.4
Irradiance is a measure of the power of sunlight (W/m2 ).
The power output of a PV panel is proportional to the 0.2
solar irradiance that hits the panel. Figure 1 shows the
0.0
solar irradiance measurements over three whole days, with 8 9 10 11 12 13 14 15 16 17
different clouding conditions. The effective Global Horizontal Time [h]

Irradiance (GHI) on the solar panel can be modeled in two

Fig. 2. Clear-sky index found for the measured solar irradiance data shown
parts. The deterministic part, which is the estimated clear-sky in Fig. 1. In some cases, the clear-sky index becomes bigger than zero during
irradiance based on the position of the sun and the stochastic flickering conditions. This is due to cloud enhancement, i.e., sunlight being
part, dependent on the cloud movement. reflected by surrounding clouds.

where z is the zenith angle, which is estimated based on

1200 the location and time of day. The parameters a and b are
09.01.2011 determined by fitting equation (2) to the measured data for
10.01.2011
1000 clear-sky days [6]. These coefficients change day by day and
11.01.2011
are thus found for each clear day of the data set. For the
Solar irradiance [W/m2 ]

Clear sky irradiance

800 remaining days, a and b coefficients are estimated based on
the parameters for the clear days.
600 The data sets of clear-sky indexes for one day are limited
by the sunrise and sunset, that is when GC (t) > 0. The low
400 values of GHI occurring just after sunrise and sunset result
in higher uncertainties in the clear-sky index [14]. Because of
200 this, only solar irradiance data from 8am to 5pm are used.

0
6 8 10 12 14 16 18 20
III. M ODELS OF THE C LEAR -S KY I NDEX
Time [h] This section presents four models to model the fluctuations
of the clear-sky index. The first three models have been pre-
Fig. 1. Measured solar irradiance [13] and estimated clear-sky global solar
irradiance using (2). sented in the literature to represent solar irradiance fluctuations
for the short-term (seconds-to-minutes time scale) analysis of
The temporal variability of solar irradiance, due to cloud power systems. The fourth model, which is proposed in this
coverage, can be modeled through the clear-sky index, k. This paper, addresses the shortcomings of the available models and
is defined as the ratio between the measured GHI, G, and the proposes a novel way to reproduce the jumps in the clear-sky
estimated global horizontal clear-sky irradiance, GC : index.
G
k= . (1) A. Model I
GC
This model is a simple way to represent the clear-sky
The clear-sky index for the three days in Fig. 1 is shown in index variations in simulation [10], [11]. Such variations are
Fig. 2. represented by a random signal between 1 and 0.4, with 5
The clear-sky global solar irradiance is the maximum ir- second steps. Figure 3 shows an example of a simulated clear-
radiance arriving at earth’s surface at a specific location and sky index obtained with Model I. The range of the jump size
time, i.e., when no clouds are present. The clear-sky irradiance and the waiting time between jumps can be changed to fit
depends on the site, the solar elevation angle and various different locations. However, it is not possible to vary the
atmospheric conditions [6]. waiting time between consecutive jumps or to consider the
A number of models of varying complexity have been correlation of jump amplitudes.
suggested in the literature to model the clear-sky irradiance. In
this paper, a clear-sky model of the same form as the Adnot- B. Model II
Bourges-Campana-Gicquel model is used [5]:
This model, presented in [12], as well as the following two
GC = a · cos(z)b , (2) models, split the representation of the clear-sky index into two
 3

1.0 1.0
Clear sky index

Clear sky index

0.8 0.8

0.6
0.6

0.4
0.4
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Time [s] Time [s]

Fig. 3. Clear-sky index generated with Model I. Fig. 4. Clear-sky index generated with Model II.

parts: C. Model III

1. The baseline of the clear-sky index varying around 1, This model was presented in [4] and uses jump diffusion
which models the clear-sky condition. processes, i.e., SDEs with jumps, to represent the clear-sky
2. The jumps of the clear-sky index due to cloud coverage. index. A general jump diffusion process is defined as:
The baseline is modeled using an Autoregressive Moving
Average (ARMA) model. ARMA models are discrete, with dY (t) = a(Y (t), t) + b(Y (t), t)dW (t) + ξdJ(t) , (6)
a fixed time step that must match the sampling interval where a(Y (t), t) and b(Y (t), t) are the drift and diffusion
of the data. Measured solar irradiance data sampled every terms, respectively; ξ is a jump size that is assumed to be
minute is used to define the parameters of the model. In [12], a normally distributed random number, ξ ∼ N (µξ , σξ ); and
interpolation is used to convert the model from a time-step of J(t) is a Poisson process with jump rate λ, as defined in (3).
one minute to one second. In the following, for comparison, it is assumed that the first
The number of cloud events E (jumps) are modeled using a two terms on the right-hand side of (6) represent an Ornstein-
Poisson random variable, with the mean λ and the probability Uhlenbeck process, as in (5).
density function: In [4], three scenarios are modeled: cloudy, flickering and
λx sunny. For the cloudy and sunny scenarios, no jumps are
f (x, λ) = exp(−λ) , (3)
x! considered (ξ = 0). Then, a non-parametric estimation method
where x = 0, 1, 2.... The inter-event waiting time, i.e., the is used for estimating the parameters of the model. Figure 5
time between cloud events, is itself a random variable with an shows an example of a simulated clear-sky index in the flick-
exponential distribution with mean µW : ering state obtained with Model III. The parameters defined
1 in [4] for the flickering state are used, namely, λ = 0.01,
fW (t) = exp(−t/µW ) , (4) σξ = 0.028 and µξ = 0.7. The number of times the clear-sky
µW
index data crosses its mean value (∼ 0.7) is used for defining
where t ≥ 0. Finally, the duration TD of a cloud event is the jump rate λ. This assumption clearly prevents modeling
assumed to be exponentially distributed with mean µD . A any jump that is smaller than the threshold defined by the
detailed description of how the parameters for the cloud events mean value.
are derived can be found in [12].
In the following, the ARMA model is substituted with an
Ornstein-Uhlenbeck Stochastic Differential Equation (SDE) to
1.0
illustrate the properties of this model. This substitution has no
side effect as the variations of the baseline are minimal in the 0.8
Clear sky index

considered time scale. The SDE is defined as: 0.6

dX(t) = θ(µ − X(t))dt + σdW (t) , (5) 0.4

where µ, θ > 0 and σ > 0 are parameters and W (t) is a 0.2

Wiener process. The process described by (5) is a continuous- 0.0
time equivalent of an ARMA(1,0) process. Models III and IV, 0 50 100 150 200 250 300 350 400
Time [s]
which are discussed below, also utilize the process in (5) to
represent the stochastic clear-sky conditions.
Fig. 5. Clear-sky index generated with Model III.
Figure 4 shows an example of a simulated clear-sky index,
generated with Model II and with the parameters of the jump
process that represent the spring data set, i.e., λ = 7.4178,
µW = 46.5186 and µD = 54.0616 [12]. In this model, the D. Model IV (Proposed Model)
jump amplitude is the same for all cloud events, which is not If the clear-sky indexes generated using Models I to III
realistic. above are compared with measurement data, two limitations
 4

become apparent. Firstly, they are based on the whole data set, TABLE I
not on the flickering cloudy conditions solely. For this reason, R ANGE OF PARAMETERS FOR THE JUMP MODELS OF M ODEL IV DEFINED
BASED ON THE ANALYSIS OF 60 CLOUDING EVENTS
these models cannot capture the dynamics of fast variations in
the time scale of seconds to minutes. Secondly, small jumps Parameters Jump model 1 Jump model 2
of the clear-sky index are neglected. λ 0.005 − 0.05 0.05 − 0.1
µm 0.6 − 0.8 0
The model proposed in this section is aimed at capturing σm2 0.0005 − 0.005 0.01 − 0.1
clear-sky index variations for flickering clouding events over σδ2 10 − 50 1−5
the time scale of seconds to minutes. The proposed method
utilizes the Ornstein-Uhlenbeck SDE in (5) to represent the TABLE II
clear-sky stochastic variations in the same way as is done in PARAMETERS OF M ODEL IV FOR THE CLOUDING EVENT SHOWN IN F IG . 6
Method II and III. Since the jumps do not depend on the Parameters Jump model 1 Jump model 2
stochastic variable X(t) and are additive noise (see (6)), they λ 0.007 0.05
are added directly to the X(t), thus simplifying the numerical µm 0.7 0
σδ2 30 3
integration. The interested reader can find the detailed proce- σm2 0.05 0.001
dure to integrate jump diffusion processes in [15].
The jumps are modeled as:
IV. C ASE S TUDY
H(t) = mP (t) , (7)
A cloud event from the data set presented in Section II is
where m is the jump amplitude assumed to be a normally considered in this section to illustrate Model IV and compare
distributed random number, namely, m ∼ N (µm , σm 2
). P (t) its output with Models I to III discussed above. The event
is a step function that can get only 0/1 values, where the occurred on the 1st of December 2010 and its duration was
number of transitions per period are determined with Poisson 450 s. The analysis of the behavior of the clear-sky index
distribution with parameter λ as in (3). The duration of each during this event allows determining the parameters for the
jump is determined with a normal distribution δ ∼ N (0, σδ2 ). clear-sky index through Model IV.
In turn, each time P (t) switches from 0 to 1 (or to 1 to 0), it
remains constant for a time δ.
Visual inspection of the measured clear-sky index data
1.0
allows identifying two types of jumps of the clear-sky index:
Clear sky index

• Jump model 1 (JM1): Big clouds passing over the PV 0.8

that block most of the solar irradiance.
0.6
• Jump model 2 (JM2): Small clouds that typically pass
by more frequently and only partially reduce the solar 0.4
irradiance.
0.2
The resulting proposed model of the clear-sky index is: 0 100 200 300 400 500 600
Time [s]

k(t) = X(t) + u(t) G(t) , (8)

Fig. 6. Clear-sky index obtained from measurement data [13] and clear-sky
model (2).
where X(t) is defined by (5) and u(t) is a function that defines
the duration of a clouding event: The first step is to identify the parameters for JM1. Three
( periods of low clear-sky indexes can be clearly seen from
1 if ustart ≤ t ≤ ustop Fig. 6. From this values, one obtains that the number of jumps
u(t) = (9)
0 otherwise , in the unit of time (450 s), is 3/450 = 0.007. This value is
within the range given in Table I and, just for the sake of
where ustart and ustop are the starting and ending times of example, the Poisson distribution parameter is set λ = 0.007.
the clouding event and With similar assumptions, the average and variance of the
2
( jumps amplitudes are set to µm = 0.7 and σm = 0.001,
−H1 (t) + H2 (t) if H1 (t) > 0 respectively, whereas the variance that defines the distribution
G(t) = (10)
−H2 (t) otherwise , of the duration of big clouding events is set to σδ2 = 30.
The second step is to identify the parameters for JM2. Only
where H1 (t) and H2 (t) are JM1 and JM2, respectively. jumps in the measured clear-sky index that exceed 0.1 are
The data set presented in Section II is used for evaluating considered. Such a number is found to be 23 and, hence, λ =
the parameters of the jump models. Five clouding events for 23/450 ≈ 0.05 is used. The remaining parameters for JM2
each month, for a total of 60 events, are analyzed. This analysis are identified in the same way as the parameters for JM1. All
leads to the parameters shown in Table I. These parameters are parameters of Model IV for the clouding event in Fig. 6 are
utilised in the remainder of the paper to represent the jumps of shown in Table II.
the clear-sky index for Model IV, as discussed in the following Using Model IV and the parameters of Table II, one can
section. generate synthetic scenarios that are comparable with the
 5

measurement data clouding event shown in Fig. 6. Two sample V. C ONCLUSIONS

processes are shown in Fig. 7. Visual inspection reveals that This paper deals with the modeling of solar irradiance for
the proposed model is able to reproduce the main features short-term analysis of PV solar generation. Three models pre-
of the clouding event of the measurement data in the time- viously proposed in the literature are discussed. It is concluded
domain. that these models do not fully capture the behavior of big
jumps of the solar irradiance and fail to model small jumps.
To cope with these issues, a novel model is proposed in
the paper based on measurement data. The model considers
individual cloudy flickering events. Jumps are grouped con-
1.0 sidering two thresholds. In this way, the proposed model is
0.8
able to capture both big jumps, caused by full cloud coverage,
and small jumps, due to partial blockage of the sun. Finally,
0.6 it is demonstrated how synthetic data can be generated to
replicate the actual behavior of an individual clouding event,
0.4
as observed in real-world measurements.
Clear sky index

Model IV − a
Future work will focus on studying the impact of the
flickering of the solar irradiance in power system models. This
is done by including the proposed solar irradiance model in
1.0
power system models with PV solar power plants.
0.8
R EFERENCES
0.6
[1] BP, Statisical Review of World Energy (June 2018). [Online]. Available:
https://www.bp.com/en/global/corporate/energy-economics/statistical-
0.4
Model IV − b
review-of-world-energy/renewable-energy/solar-energy.html
[2] A. Mills, M. Ahlstrom, M. Brower, A. Ellis, R. George, T. Hoff,
0 100 200 300 400 500 600 B. Kroposki, C. Lenox, N. Miller, M. Milligan, et al., “Dark shadows,”
Time [s] Power and Energy Magazine, vol. 9, no. 3, pp. 33–41, 2011.
[3] A. Mills, M. Ahlstrom, M. Brower, A. Ellis, R. George, T. Hoff,
B. Kroposki, C. Lenox, N. Miller, J. Stein, et al., “Understanding
Fig. 7. Two sample synthetic clear-sky index processes, generated with Model variability and uncertainty of photovoltaics for integration with the
IV and based on the cloud event shown in Fig. 6. electric power system,” The Electricity Journal, 2009.
[4] M. Anvari, B. Werther, G. Lohmann, M. Wächter, J. Peinke, and H.-
P. Beck, “Suppressing power output fluctuations of photovoltaic power
Figure 8 compares the probability distributions of the mea- plants,” Solar Energy, vol. 157, pp. 735–743, 2017.
sured and simulated clear-sky indexes. Model I does not [5] Y. Dazhi, P. Jirutitijaroen, and W. M. Walsh, “The estimation of clear sky
global horizontal irradiance at the equator,” Energy Procedia, vol. 25,
capture the two peaks in the probability distribution, while pp. 141–148, 2012.
Models II and III capture the peaks but not the distribution [6] M. J. Reno, C. W. Hansen, and J. S. Stein, “Global horizontal irra-
between the peaks. Model IV is able to better reproduce diance clear sky models: Implementation and analysis,” Sandia report
(SAND2012-2389), 2012.
the clear-sky indexes probability distribution and time-domain [7] R. Elliott, A. Ellis, P. Pourbeik, J. Sanchez-Gasca, J. Senthil, and
behavior when compared to Models I-III. J. Weber, “Generic photovoltaic system models for WECC-a status
report,” in Power & Energy Society General Meeting, IEEE, 2015.
[8] M. Hossain, T. Saha, N. Mithulananthan, and H. Pota, “Robust control
strategy for PV system integration in distribution systems,” Applied
Energy, vol. 99, pp. 355–362, 2012.
[9] T. Alquthami, H. Ravindra, M. Faruque, M. Steurer, and T. Baldwin,
Data Model I “Study of photovoltaic integration impact on system stability using
5 custom model of PV arrays integrated with PSS/E,” in North American
Power Symposium (NAPS). IEEE, 2010, pp. 1–8.
[10] R. Ayyanar and A. Nagarajan, “Distribution system analysis tools for
0 studying high penetration of PV with grid support features,” Arizona
Model II Model III State University, Tech. Rep, 2013.
Probability

5 [11] A. Nagarajan and R. Ayyanar, “Methods for dynamic analysis of distri-

bution feeders with high penetration of PV generators,” in Photovoltaic
Specialists Conference (PVSC), IEEE, 2016, pp. 1848–1852.
0 [12] J. Sexauer and S. Mohagheghi, “Hybrid stochastic short-term models for
Model IV − a Model IV − b wind and solar energy trajectories,” in Green Technologies Conference
5 (GreenTech), 2015 Seventh Annual IEEE, 2015, pp. 191–198.
[13] M. Sengupta, A. Andreas (2010). Oahu Solar Measurement
Grid (1-Year Archive): 1-Second Solar Irradiance; Oahu, Hawaii
0 (Data); NREL Report No. DA-5500-56506. [Online]. Available:
0.0 0.5 1.0 0.0 0.5 1.0 http://dx.doi.org/10.5439/1052451
Clear sky index [14] M. Lave, J. Kleissl, and E. Arias-Castro, “High-frequency irradiance
fluctuations and geographic smoothing,” Solar Energy, vol. 86, no. 8,
pp. 2190–2199, 2012.
Fig. 8. Histograms of the clear-sky index during the clouding event shown [15] E. Platen and N. Bruti-Liberati, Numerical Solution of Stochastic Differ-
in Fig. 6 and the generated clear-sky indexes obtained with Models I-IV. ential Equations with Jumps in Finance. Berlin Heidelberg: Springer
- Verlag, 2010.