Synthetic Generation of Hourly Solar Irradiance

Using a Multi-State Markov Model

Vinay Chamola and Biplab Sikdar
Department of Electrical and Computer Engineering
National University of Singapore, Singapore
AbstractSolar irradiance data is a key factor in dimensioning
photovoltaic (PV) panel size and energy storage for solar powered
systems. Unfortunately, long term solar irradiance data is not
available for many locations, thereby affecting the accuracy
and effectiveness of dimensioning efforts. To address this issue,
this paper presents a model for generating accurate, synthetic
traces of arbitrary length for hourly solar irradiance. The
proposed model is based on an approach that combines daily
correlations in weather conditions affecting the solar irradiance
with finer-grained, hourly transitions in solar irradiance levels.
The proposed model has been evaluated by comparing various
statistical parameters of hourly solar irradiance against real data.

I. I NTRODUCTION
With the increase in global carbon emissions and the
increasing awareness of government agencies and industries,
green energy solutions have emerged as a viable solution and
active area of research in many domains such as transportation,
communications, electrification etc. The primary objectives
of solar powered systems are to either provide electricity
based services in areas without grid power or to reduce the
dependence on traditional energy sources and their associated
environmental impact. Solar energy based systems are further
motivated by the increasing prices of conventional energy
sources and sharp decreases in the prices of PV panels
for harvesting solar energy. The greater popularity of solar
energy is also because it is more readily available than other
renewable resources, is more reliable and designing systems
based on it is simpler as compared to others. For example,
90% of the cellular base stations powered by green energy in
the world are powered by solar energy [1].
A solar powered system typically consists of solar panels (for harvesting solar energy), batteries (to store energy
for nights and weather periods), and an inverter (to covert
AC/DC). The problem of dimensioning such systems involves
estimating the required solar panel size and the number of
batteries in order to satisfy the systems operational requirements. Inaccuracies in the estimated dimensions may lead to
degraded or unacceptable performance as well as higher costs
[2]. Also, in the specific case of off-grid applications, any
errors in the estimation of the dimensions may cause very
unreliable performance and much higher outages than what
the system was designed for. A critical input that determines
the accuracy of the dimensioning results is the solar irradiance data. While historical long-term solar irradiance data is
available for certain places, such data is usually not available
for much of the developing world where the lack of reliable

grid power makes solar powered systems more important. To

address this issue, this paper develops a simple model for
generating synthetic, long-term traces of location-specific solar
irradiance.
Existing approaches to model the daily/hourly solar irradiance include Markov models [3], [4], autoregressive moving
average (ARMA) models [5] - [7], and models based on artificial neural network [8], [9]. The Markovian approach assumes
the solar irradiance at a given time (either day or hour) to
be dependent only on the solar irradiance at the last time
unit. Further, daily or hourly transition statistics are derived
to characterize the transition from one solar irradiance level
to other. The ARMA approach uses autocorrelation between
consecutive days to capture the nature of the solar radiation.
The artificial neural network method for characterizing solar
irradiance has been developed in the recent past [9]. However,
their applicability is for the prediction of solar irradiance based
on past observations, and they require long term solar data
for training. It has been shown that Markov models perform
better than ARMA models [10], [11]. While most Markov
models focus on daily irradiance levels, some models exist
for modeling the hourly solar irradiance [10]. However, they
lack accuracy since they do not consider the day to day
correlations in the solar irradiance. To address this issue, this
paper proposes a Markov model which captures the solar
irradiance characteristics both on an hourly as well as on a
daily scale.
This paper proposes a model to generate synthetic traces of
hourly solar irradiance values for any given site. The proposed
model is based on combining two Markov models. The first
model captures the day to day correlations in the weather that
affects the solar irradiance levels over a day. The second model
captures the hourly variations in the solar irradiance levels. The
proposed model only requires a years data for training and its
effectiveness is compared against real data for two different
locations. Our results show that the proposed model generates
solar irradiance traces that match real data in terms of multiple
statistical parameters.
It is to be noted that this paper uses the Global Horizontal
Irradiance (GHI) as the solar irradiance. GHI is the total
amount of solar irradiance received from above by a surface
which is horizontal to the ground. It includes both Direct
Normal Irradiance (DNI) and Diffuse Horizontal Irradiance
(DHI). DNI is the solar radiation that comes in a straight line
from the direction of the sun at its current position in the

V. Chamola and B. Sikdar, Synthetic Generation of Hourly Solar Irradiance Using a Multi-State Markov Model, Proc. IEIE/IEEE ICEIC, Singapore, Jan
2015. Presented at International Conference on Electronics, Information and Communication (ICEIC 2015), Singapore

pgb

irradiance level. The overall state of the process representing

the solar irradiance at any hour is denoted by
St : St {Bx,y , Gx,y },

pgg
Good

Bad

x {1, 2, , 24}, y {1, 2, 3, 4}

(2)
where B and G correspond to bad and good weather days
respectively, x denotes the hour of the day and y denotes
the state of the solar irradiance in that hour. We denote the
expected value of the solar irradiance in each state by ESt and
it can be obtained by taking the average of the solar irradiance
in that particular state (using the empirical data). Examples of
the four states for each hour of good and bad weather days are
shown in Figures 2 and 3 respectively. The methodology for
determining the values of ESt and the other parameters used in
our model is described in Section II-A. It should be noted that
though one can use more than four states for characterizing
the solar irradiance in any hour of the day, our experimental
results show that there is only a marginal improvement in the
accuracy by including additional states. Thus the addition of
extra states at the cost of increased complexity is not justified.

pbb

pbg
Fig. 1. State diagram for transitions between good and bad days.

sky whereas DHI is the solar radiation that does not arrive
on a direct path from the sun, but has been scattered by the
molecules and particles in atmosphere and comes equally from
all directions [12].
The rest of the paper is organized as follows. Section II
presents the proposed model for generating solar irradiance
data. Section III presents the validation results while Section
IV concludes the paper.

Based on the Markovian assumption for the hourly solar

irradiance, the state of solar irradiance at a given hour depends
only on the state of solar irradiance in the previous hour and
the transition probability of going from the previous state to
the current state. This may be expressed as

II. P ROPOSED M ODEL

In this section we describe the proposed framework to
model the location-specific solar irradiance on an hourly basis.
The proposed model combines a Markov model for capturing
the day to day correlations in the weather patterns with
another model that captures the hourly variations in the solar
irradiance.
In existing literature, it has been shown that the solar irradiance levels for a given location are well modeled by a Markov
process [10]. In this paper, we refine the existing models to
propose a multi-state Markov model for characterizing the
solar irradiance. In the proposed model, we classify any given
day as either a bad or good weather day based on the
daily solar irradiance of that particular day. Intuitively one
can think of the sunny days to be good days and those with
clouds for a major portion of the day as bad days. From the
complete set of days, we then select % of the days as bad
days. Given a bad or a good day, the next day may be either
bad or good day. In order to capture occurrence of consecutive
bad or good weather days, we model this transition as a twostate Markov process. The transition probability matrix of this
Markov process is given by


pbb pbg
T=
(1)
pgb pgg

P [St |St1 , St2 , ...S0 ] = P [St |St1 ].

(3)

Given that the solar irradiance is currently in a given state, the

state may transition to any of the four states in the next hour
(for both good and bad weather days). For a bad weather day,
the transition probability matrix can be given as

b(1,1)(1,1) b(1,1)(24,4)

..
..
..
(4)
B=

.
.
.
b(24,4)(1,1)

b(24,4)(24,4)

1
min/max
cutoffs
average

0.9

Solar irradiance (kW/m2)

State 4

where pbb (resp. pgg ) is the probability of transition from a bad

day (good day) to a bad day (good day), and pbg = 1 pbb
(resp. pgb = 1pgg ) is the probability of transition from a bad
day (good day) to a good day (bad day). The state transitions
between the good and bad days are shown in Figure 1.
Now, in order to capture the hourly variations of the solar
irradiance, each hour is categorized into one of the four
possible categories. The four categories for bad and good days
are different and each category is characterized by its solar

0.8
0.7

State 3

0.6
0.5

State 2

0.4
0.3

State 1

0.2
0.1
0
0

10

12

14

16

18

20

22

Hour of the day

Fig. 2. An example of the demarcation of states and their expected solar

irradiance for good days.

V. Chamola and B. Sikdar, Synthetic Generation of Hourly Solar Irradiance Using a Multi-State Markov Model, Proc. IEIE/IEEE ICEIC, Singapore, Jan
2015. Presented at International Conference on Electronics, Information and Communication (ICEIC 2015), Singapore

A. Parameter Estimation

1
0.9

min/max
cutoffs
average

Solar irradiance (kW/m2)

In this section we describe the methodology for obtaining

the parameters described in the previous section. For the
0.7
parameter estimation we use historical data of solar irradiance,
0.6
State 4
like that provided by the National Renewable Energy Labora0.5
State 3
tory (NREL), USA [13]. For any given location, we consider
0.4
the solar irradiance data for one year and use the monthly
State 2
0.3
data to extract the model parameters for each month. For each
State 1
0.2
day of the month, we first calculate the total solar irradiance
0.1
received over the day. The days are then sorted based on the
0
0
2
4
6
8
10
12
14
16
18
20
22
received solar irradiance, and a fraction of the days with
Hour of the day
the lowest solar irradiance are marked as bad weather days
while the rest of the days are marked as good weather days.
Fig. 3. An example of the demarcation of states and their expected solar
irradiance for bad days.
This data is then used to calculate the transition probability of
going from a bad day to bad day, and good day to good day
so as to obtain the transition probabilities pgg , pgb , pbb and
pbg for Eqn. (1).
with
Further, in order to obtain the average solar irradiance values
(
r(i,j)(k,l) k = (i+1) mod 24; j, l {1, 2, 3, 4} in each of the four possible states in each hour, we first analyze
b(i,j)(k,l) =
the hourly solar irradiance data to record the hourly maximum
0
otherwise
(5) and minimum values of solar irradiance received for both bad
where b(i,j)(k,l) is the probability of transition from the j-th and good weather days. The region between the maximum and
state in the i-th hour to the l-th state in k-th hour on a bad minimum solar irradiance values for each hour is divided by
weather day. Note that in a given hour, from a particular state, defining cutoffs which uniformly partition the region into four
the solar irradiance can only go to one of the states in the sub-regions. Each of the four sub-region specifies the state
next hour. The numerical value of the transition probability is of solar irradiance at that hour. Using the empirical data, we
obtain the average hourly irradiance values for each of those
denoted by r(i,j)(k,l) and it satisfies
states.
4
X
r(i,j)(k,l) = 1,
k = (i + 1) mod 24, j {1, 2, 3, 4}.
l=1
III. R ESULTS
(6)
0.8

We can similarly define the the transition probability matrix

for a good weather day. The hourly state transitions for a bad
day are shown in Figure 4.

In this section we evaluate the proposed model for generating the solar irradiance data by comparing its results with
empirical data. To validate the methodology presented in this
paper, we consider two locations: Kolkata (India) and San
Diego (USA) and for both locations, our model used = 0.2.
The solar irradiance data for these locations was obtained from
the NREL database.

Synthetic
1.2

1.1

1.1

0.9

0.9

Solar Irradiance (kW/m2)

Solar Irradiance (kW/m2)

Empirical
1.2

0.8
0.7
0.6
0.5
0.4

0.7
0.6
0.5
0.4

0.3

0.3

0.2

0.2

0.1
0

0.1
2000

4000

Hour

Fig. 4. Markov state diagram showing transition from a state in an hour to

state in next hour

0.8

6000

8000

2000

4000

6000

8000

Hour

Fig. 5. Empirical vs synthetic solar irradiance data for Kolkata (2009).

V. Chamola and B. Sikdar, Synthetic Generation of Hourly Solar Irradiance Using a Multi-State Markov Model, Proc. IEIE/IEEE ICEIC, Singapore, Jan
2015. Presented at International Conference on Electronics, Information and Communication (ICEIC 2015), Singapore

Synthetic
10

1.1

1.1

0.9

0.9
2

Solar irradiance (kW/m2)

1.2

Solar Irradiance (kW/m )

Solar Irradiance(kW/m2)

Empirical
1.2

0.8
0.7
0.6
0.5
0.4

0.8
0.7
0.6
0.5
0.4

7
6
5
4
3

0.3

0.3

0.2

0.2

0.1

0.1

2000

4000
6000
Hour

8000

0
Jan
2000

4000
6000
Hour

Empirical: Daily mean

Model: Daily mean
Empirical: Daily variance
Model: Daily variance

Feb

Mar

Apr

May

Jun

8000

July Aug Sept

Oct

Nov

Dec

Month

Fig. 8. Comparison of the mean and variance of the daily solar irradiance
values, empirical vs model generated data: San Diego (2009)

Fig. 6. Empirical vs synthetic solar irradiance data for San Diego (2009).
7
Empirical: Daily mean

1
Empirical: Hourly mean
Model: Hourly mean
Empirical: Hourly variance
Model: Hourly variance

Model: Daily mean

0.9

Empirical: Daily variance

0.8
2

Solar irradiance (kW/m )

Solar irradiance (kW/m2)

Model: Daily variance

0.7
0.6
0.5
0.4
0.3
0.2

0.1
0
Jan

Feb

Mar

Apr

May

Jun

July Aug Sept

Oct

Nov

Dec

0
1

Month

11

13

15

17

19

21

23

Hour

Fig. 7. Comparison of the mean and variance of the daily solar irradiance
values, empirical vs model generated data: Kolkata (2009)

Fig. 9. Comparison of the mean and variance of the hourly solar irradiance
values, empirical vs model generated data: Kolkata (June 2009)

A. Yearly Solar Irradiance

that these are the months in the respective locations which

exhibited maximum variability in terms of solar irradiance,
for the year 2009. Thus we focus on these months to further
validate the performance of the proposed model.
To get a finer look at the performance of the proposed
model, next we compare the mean and variance of the solar
irradiance values in each hour of the day for these two
locations for the month with maximum variability. Figure
9 and 10 show these results for Kolkata and San Diego,
respectively. It can be seen that the hourly mean and variance
of the synthetic trace generated by our model matches closely
with that of empirical data.

To ascertain the accuracy of the proposed model, we first

consider the solar irradiance trace for an entire year, as
generated by our model and compare it against real data.
Figures 5 and 6 show the hourly trace of solar irradiance
for the two locations for the year 2009. We note the close
match between the overall trends and values of the data for
the synthetic and empirical traces.
B. Mean and Variance of Irradiance Levels
To get a more detailed look at the performance of the
proposed model, next we consider the mean and variance of
the daily solar irradiance values. It is to be noted that for
our model, long duration synthetic data was generated for
comparing the mean and variance values. Figures 7 and 8 show
the mean and variance of the daily solar irradiance values,
calculated on a monthly basis for the year 2009. Again we
see that the results from the proposed model match well with
those from empirical data.
From Figure 7 and 8, we observe that the month of June for
Kolkata and the month of May for San Diego have the highest
variance in the daily solar irradiance values. This indicates

C. Irradiance State Statistics

While the hourly mean and variance results provide an
indication of the overall accuracy of the proposed model, a
more finer look at the match with empirical data is obtained
by considering the statistics related to the hourly irradiance
values. For this comparison we generate a month long synthetic trace for both these months. The solar irradiance from
0 to 1000 W/m2 is divided into 10 bins. For the months of
June for Kolkata and May for San Diego in the year 2009, for

V. Chamola and B. Sikdar, Synthetic Generation of Hourly Solar Irradiance Using a Multi-State Markov Model, Proc. IEIE/IEEE ICEIC, Singapore, Jan
2015. Presented at International Conference on Electronics, Information and Communication (ICEIC 2015), Singapore

1.1
Empirical: Hourly mean
Model: Hourly mean
Empirical: Hourly variance
Model: Hourly variance

Solar irradiance (kW/m2)

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0

11
13
Hour

15

17

19

21

23

Fig. 10. Comparison of the mean and variance of the hourly solar irradiance
values, empirical vs model generated data: San Diego (May 2009)

Fig. 12. Comparison of number of hours in a particular irradiance slot,

empirical vs model generated data: San Diego (May 2009)

R EFERENCES
[1] http://www.gsma.com/mobilefordevelopment/programmes/green-powerfor-mobile/tracker, Last accessed: August 23, 2014.
[2] Zhou, Wei, et al. Current status of research on optimum sizing of standalone hybrid solarwind power generation systems, Applied Energy 87.2
(2010), pp. 380-389.
[3] Knight, K. M., S. A. Klein, and J. A. Duffie. A methodology for the
synthesis of hourly weather data, Solar Energy 46.2 (1991), pp. 109120.
[4] A. Maafi and A. Adane, Analysis of the performances of the first-order
two-state Markov model using solar radiation properties, Renewable
Energy, vol. 13, no. 2, pp. 175-193, 1998.
[5] Festa, R., S. Jain, and C. F. Ratto. Stochastic modelling of daily global
irradiation, Renewable energy 2.1 (1992), pp. 23-34.
[6] Hokoi, S., M. Matsumoto, and M. Kagawa. Stochastic models of
solar radiation and outdoor temperature, ASHRAE Transactions (American Society of Heating, Refrigerating and Air-Conditioning Engineers);(United States) (1990), pp. 245-252
[7] Mora-Lopez, L. L., and M. Sidrach-de-Cardona. Multiplicative ARMA
models to generate hourly series of global irradiation, Solar Energy
63.5 (1998), pp. 283-291.
[8] Dorvlo, Atsu SS, Joseph A. Jervase, and Ali Al-Lawati. Solar radiation
estimation using artificial neural networks, Applied Energy 71.4 (2002),
pp. 307-319.
[9] Rehman, Shafiqur, and Mohamed Mohandes. Artificial neural network
estimation of global solar radiation using air temperature and relative
humidity, Energy Policy 36.2 (2008), pp. 571-576.
[10] Kamal, Lalarukh, and Yasmin Zahra Jafri.Stochastic modeling and
generation of synthetic sequences of hourly global solar irradiation at
Quetta, Pakistan, Renewable energy 18.4 (1999), pp. 565-572.
[11] Aguiar, R. J., M. Collares-Pereira, and J. P. Conde. Simple procedure
for generating sequences of daily radiation values using a library of
Markov transition matrices, Solar Energy 40.3 (1988), pp. 269-279.
[12] http://www.3tier.com/en/support/solar-prospecting-tools/, Last accessed:
August 23, 2014.
[13] http://www.nrel.gov/rredc/solar_data.html, Last accessed: August 23, 2014.

Fig. 11. Comparison of number of hours in a particular irradiance slot,

empirical vs model generated data: Kolkata (June 2009)

both the empirical and synthetic data, we count the number of

occurrences of the hourly solar irradiance being in a particular
irradiance level. The comparison is shown in Figures 11 and
12. It can again be observed that there is a close match between
the number of hours in a given irradiance level for the model
generated synthetic trace and that obtained from empirical
data.
IV. C ONCLUSION
This paper presented a framework for generating synthetic
traces of hourly solar irradiance for the purposes of dimensioning solar-powered systems. The proposed framework is based
on combining Markov models for the daily changes in weather
patterns with one for an hourly variation in the solar irradiance
levels. The proposed model has been validated by comparing
it against empirical data from geographically diverse regions.