Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

PrDbabilistic MDdeling Df SDlar PDwer Systems

Fayssal M. Safie, Ph.D.; Reliability Engineer; NASA

Key WDrds: SDlar phDtDvDltaic pDwer systems, stochastic processes, MarkDv Chain TheDry.

Abstract most suited for use for detailed analysis of PV systems.


However, the variation in SOLCEL data of up to !20 per­
Performance evaluation Df stand-alDne phDtovoltaic cent is creating the primary errors in the predicted
(PV) power systems has, in general, been dDne using a system performance." In fact, due to their deterministic
dynamic simulation approach. This paper presents a nature most models have relatively low efficiency and
probabilistic approach based on MarkDv Chain Theory to their reliability is limited to the reliability of
model stand-alone PV power systems and predict their hourly, daily, or monthly data used.
long-term service perfDrmance. The majDr advantage of
this approach is that it allows designers and developers This paper presents a probabilistic approach to
Df these systems to analyze the system perfDrmance as analyze the service performance of a stand-alone PV
well as the battery subsystem performance in the IDng system. Figure 1 shows a simplified diagram of a stand­
run, and determine the system design requirements that alone PV power system. The method is based on Markov
meet a specified service performance level. The Chain Theory. A Markov chain model for solar insolation
methodology presented is illustrated by using data for is first developed. The transition probab�ities of the
a radio repeater system for the Boston location. matrix are estimated from macro data usL.,..., _ least­
squares estimation technique and a quadraclc p rogramming
IntroductiDn formulation. A set of energy flow equations for a
constant load demand and a constant voltage ampere-hour
Society continues to seek alternative energy sources storage battery are developed to track flDw of energy
that are efficient and more reliable. One alternative through the system. The insolation model, the constant
is solar photovoltaic energy whereby solar energy is load demand model, and the flow of energy equations are
converted directly to electricity through the use of then used with other system data tD construct a Markov
solar cells. This methodology has been known since one step transition probability matrix to monitor the
1839 and the first practical solar cell was developed in performance of the system. From the performance matrix,
1954 (Ref. 1). The relatively high cost and low effi­ the steady-state probabilities are computed and used to
ciency of these early cells, tDgether with the consumer determine the long-term system service performance.
skepticism which is a typical reactiDn to many new
products, prevented their wide spread use.

During the 1960's, PV power systems were first used


in space vehicles. Until the energy crisis created by
the oil embargD in the early 1970's, terrestrial PV
applicatiDns were limited to small, special purpose
installations where the advantage of a totally isolated
power supply made it possible to justify the high cost
of the solar PV power system. Since then, a great deal
Df effDrt has been directed toward the use of solar PV
power as an alternative energy source for terrestrial
applications.
ELECTRICAL
DEMAND
With increasing cost of conventional energy, sDlar
PV power systems can compete with other energy systems, ENERGy lOST ___-'
prDvided improvements can be made in the methods and
techniques for designing and manufacturing PV devices. Figure 1. A stand-alone photovoltaic power system.
This required special attention in the design and per­
formance analysis of the PV system. Basic Approach to Model Development

Modeling Efforts In a stand-alone PV power system, the battery sub­


system acts as a buffer that provides stability and
Since the early 1970's, several dynamic simulation reliability for running the loads when the amount of
models have been used for designing and analyzing stand­ solar insolation available is not enough tD produce the
alone PV power systems. The NASA Lewis Research Center energy required to satisfy the demand. In general, the
has developed three simulation programs: SEEDRA, insolation and the load demand are both random and
SEBIFANO, and SHEMESH (Ref. 2). Sandia Laboratories accordingly the battery will behave in some stochastic
have developed SOLCEL and SOLCEL-II (Refs. 3,4) which fashion. In fact, the PV system shown in Figure 1 can
provided more detailed analysis than those of NASA Lewis be considered as a special type of inventory system that
Research Center. Other commonly used design programs has received attention in the industrial engineering and
have been developed by corporations that produce solar operations research literature. Using analogy, the
PV devices such as SOLAVOLT International (Ref. 5). energy generated by the array is like items coming into
a central distribution center from some number of manu­
While the models developed to date have apparently facturing facilities where they are transferred to con­
been successfully used, the probabilistic behavior of sumers (loads) or else stored in a warehouse (battery
the system has not been considered. Gupta (Ref. 6) storage) for later use. For certain types of perishable
reports that "of the analysis models currently available. items, if distribution is not made by a certsin time
the Sandia developed Code, SOLCEL. represents the code then the value of the items is altered or diminished.

0149 144X/89/0000-0425SOl.00 © 1989 IEEE


1989 PROCEEDINGS Annual RELIABILITY AND MAINTAINABILITY Symposium 425
This is similar to a PV system with energy generated
from the array in excess of battery capacity and load
requirements. Using this analogy. Markov chains which
have been successfully applied to inventory and other
storage systems. have been used as a first attempt to
model PV power systems.

Consider a stochastic process [X : t � 0] with a


t
state space W and time parameter T � O. We characterize
the state space W by the th ree- tupl e (S,I,L) where S is
the battery state of charge in ampere-hours (amp-hrs).

I is the insolation in KWH/m2-day. and L corresponds to


the load demand in amp-hrs. Hence. we have:

S S S
l' m

I I
11' n

L L1• L
Y
=
W [w S L ) i=lt ... ,ill, j=l, . . . ,TI,
( i' Ir k :
k=l .. . . . y].

Here we shall consider a discrete time parameter, that


is the parameter space T consists of a discrete set of
points (t , t2,
1
t )' Thus. we have for ou r general
z
• . • •

model [X : te: T] and at fixed time vectors tEo T. the


t
corresponding X is a random variable of appropriate
t
dimensions. We shall assume that (X : t E.. T) possesses
t
Markov dependence in that the pro bab ility of the system
being in a certain state at time t given the entire past
history of the system actually depends only on the state
of c-he--system at the inunediately preceding time instant.
This assumption is common and is of practical signifi­
cance. However, its statistical validity must be Figure 2. System performance model.
examined in practice.
specific water levels. In the insolation case, since
Figure 2 shows a flow diagram of the approach used there is no a priori procedure to segment the continuous
in developing the system performance model. The system process. an arbitrary number of intervals (insolation
performance Markov chain matrix is a three-tuple chain. levels) is chosen, and a sensitivity analysis is carried
where each state consists of a battery state, insolation out to adequately determine the number of states and the
state, and load demand state. size of the intervals.

Insolation Model In order to estimate the transition probabilities


of the Markov chain matrix, different methods are used
In the last few years, stochastic behavior of solar depending on the type of data available. If micro data
insolation has been studied by many researchers. is available then the transition probabilities can be
Mustacchi (Ref. 7) indicated that stochastic simulation es timated using the "micro" maximum likelihood estima­
of solar insolation carried out with Auto-Regressive tion technique. If macro data on state occupation
Moving Average and Factor Analysis techniques did not frequency is available. the least-squares technique can
adequately describe the statistical features of the be used (Ref. 9). In this p a per, the least-squares
insolation sequence. A Markov transition probability estimator of the Markov chain matrix has been deter­
matrix approach provided a simple yet effective simula­ mined from an aggregate frequency insolation data using
tion deVice. Brinkworth (Ref. 8) found that the auto­ a quadratic programming formulat ion . The reason for
correlation for daily insolation sequence is not sig­ Using aggregate frequency data is mainly to reduce the
nificantl y greater than zero except for time lags of a model data requirements.
few days away. This lower order correlation suggests
the possibility of Markov dependence. Load Demand Model and Battery Characteristics

In- thIs- paper a--firSt�order-Markov model for solar The performance model presented in this paper has
insolation is assumed. That is, each daily insolation been developed for a single 24-hour constant load demand
value depends on its own immediate past value. This is and a simple constant voltage amp-hrs battery subsystem.
logically consistent with the fact that when the sky is Accordingly, the flow o f energy through the PV system
heavily overcast. a random disturbance will tend to can be described as follows: The amp-hrs output of the
increase its clarity. whereas if the sky is very clear PV array goes to the loads. If the load demand is less
it will be more likely to decrease its clarity the next than the array output. the excess energy goes to the
day. In order to get the state space for the Markov battery. If the battery is fully charged. the excess
chain. the continuous process is segmented into inter­ energy is lost. If the load demand exceeds the array
vals. Approximating a continuous process by a discrete output, the difference is withdrawn from the battery.
one has been routinely done for many applications. One This flow of energy can be done on an hourly. daily,
well known example is the modeling of water content of weekly, or monthly basis. In this study, we have
a dam by a Markov chain; where the states of the matrix accounted for the state-of-charge on a daily basis. In
are defined by intervals which describe a set of such a case, the 24-hour load demand must be split into

426 1989 PROCEEDINGS Annual RELIABIUTY AND MAlNTAINABIUTY Symposium


nighttime and daytime load demands to account for the b. AO
< Ld
zero insolation (array output) during the evening hours.
During these hours, the demand is totally supplied by f (L - AO) /L (8)
the battery, whereas, during the daytime hours the
demand is supplied either by the array or by the battery b 1 + f * (r-l) (9)
or both depending on the amount of insolation available.
S
2
=
31 + (AO - L)/b (10)
The av�rage daytime and nighttime load demands are
calculated as follows: If 3
2
< 3 then set S S (11)
min' 2 min

Daytime load L L * H /24 (1) If S then set S S (1 2)


d d 32 > max' 2 max
=

Nighttime load L L * H /24 (2)


n n Case 2: S S
I = min
where
a. A L L (13)
O d
average number of nighttime hours/day
Hd L /L (14)
n
H average number of daytime hours/day
n b 1 + f * (r-l) (15)
L total daily load demand .
S A - L (L /b) (16)
=

2 = 81 +
-
O d n
Note that the calculation of H and H require the deter­
d n If S then S S (17)
32 < min' 2 min
=

mination of sunset and sunrise times for the specific


location. b. AO
< Ld

Energy Flow Equations S S (18)


82 I min
=

In order to track the flow of energy through the


Case 1 indicates that when the array output equals
PV system shown in Figure 1, a set of equations has been
or exceeds the load demand during the daytime (A L L ),
developed according to the flow of energy procedure O d
described in the previous section. Given an initial the fraction of the load met by the battery, f, is equal
battery state of charge (SOC), an insolation level and to the ratio of the nighttime load to the total daily
a daily load demand, the final SOC is determined accord­ load demand. This is expressed in equation (3). Equa­
ing to the following procedure: tion (4) calculates the battery efficiency, b, for a
given battery round trip efficiency r. Equation (5)
Case 1: 3 S < 3 calculates the final battery SOC, S ' which is then
min < I max 2
checked against the battery lower and upper cut-off
a. AO L L
d points (S and S ) as expressed in equations (6) and
min max
f L /L (3) (7). The remaining equations can be explained similarly.
n

b 1 + f * (r-l) (4) Note that the flow of energy equations represents


the interrelationship between the main system variables.

52 31
+ A
O
-
Ld
- L /b
n
(5)
System Performance Matrix
=

If 3 S
2 � min'
then set S
2
S
min
(6)
Using the flow of energy equations, the Markov chain
If S S then set S S matrix for insolation and the load demand model, a Markov
2
(7)
2 2. max' max
one-step transition probability matrix is generated.
where Figure 3 shows the general form of the matrix obtained.
This matrix, called the performance matrix, contains the
AO is the array output in amp-hrs for the given transitions from one combination of battery state, inso­
insolation level, lation state, and load demand state to another combina­
tion of the same variables. The number of states of the
S is the lower cut-off point for the battery matrix is equal to the product of the number of states
min
SOC, of the insolation, load demand, and battery SOC. Since
the construction of the performance matrix involves
S is the upper cut-off point for the battery discretizing a continuous process, a sensitivity analysis
max
SOC, to determine the appropriate number of states is per­
formed.
L is the daytime load demand,
d
On the other hand, the transition probabilities of
L is the nighttime load demand, the matrix are generated in a straight forward manner.
n
With the battery SOC, the insolation, and the load demand
L is the total load demand, denoted by S, I, L, respectively, for each combination of
S, I, and L, a new battery SOC is determined using the
S is the initial battery SOC, flow of energy equations. Unfortunately, these equations
I
do not indicate what will be the values of I and L for
S is the final battery SOC, the next time period. However, a joint distribution for
2
I and L is computed from the insolation and load demand
b is the battery efficiency, models. Assuming the load demand and the insolation to
be independent, their joint probability for any combina­
r is the battery round trip efficiency, tion of the two variables is the product of the indi­
Vidual probabilities. For the special case where the
f is the fraction of the l oad met by the battery. load is constant, the joint distribution is merely the

1989 PROCEEDINGS Annual REUABILITY AND MAINTAINABIUTY Symposium 427


..... ..... .....
��
..J

-�
_o. -�

'" '" <i' ","


-E
'" '"
L
P12 • • • • Pin PI, n+1 P1,2n P1. T.n+1 PT,T

(S,.I n-1 ,L)

TRANSITION PROBABILITIES OF THE ARST ROW OF THE


PERFORMANCE MATRIX

S 1 ,11 ,L 000•• 000 A11 A'2 ••••••••••• A1n 000•••000

FT, T-n+1 Figure 4. Calculation of the transition


probabilities of the performance matrix.

The procedure developed to identify the cases of


5, . BATTERYStATEOFCHARGELEVELiO�1,2, • ,m)
• •
unsatisfied demand and to calculate F can be described
I,
� INSOLATIONLEVELi(i.1,2, • • • ,n)
T. (nJ· (m)
• TOTAL NUMBER OF STATES OF THE ERFORMANCE MATRIX
P
as follows:

Figure 3. Performance matrix. A_ Cases of Unsatisfied Demand and the Associated


Long Term Probabilities
insolation distribution; that is, the transition prob­
abilities of the row of the insolation matrix cor­ Case 1
responding to the present level of insolation.
T
To illustrate the calculation of the transition
E P . (5,I,L) (20)
J
j�1
probabilities of the performance matrix, consider the
first row of the matrix shown in Figure 3. Assume that
S.T,
the present state of the system is (S ' I ' L). Using
I I
the energy flow equations, the battery SOC for the next S = S
min
(21)
time period is determined. If the next battery SOC
happens to be S ' then the transition from 8
r 1
to S
r
can (22)
be described as shown in Figure 4_ In Figure 4, A 'S
ij
Case 2
are obtained from the joint distribution of the load and
insolation models. The transition probabilities for the
T
other rows can be generated in a similar manner.
E PJ. (5,I,L) (23)
j=1
Evaluation of Sy�tt!JlI. Performance

S_T.
Service performance of stand-alone PV systems has
been of major concern to system designers and engineers.
In addition to component failure and downtime due to
S
min
< S <
Smin + LIb (24)
maintenance, lack of adequate solar isolation can greatly
A
O
- L � 0
decrease the reliability of these systems. In fact, d
solar insolation is' the major factor in determining the
S + AO - L - {L /[1 + (L /L) (r - 1)1} < Smin
service performance level (5PL) in stand-alone PV power d n n
systems.
(25)
In our model, the SPL is determined using the sys­
Case 3
tem performance matrix. The model computes the steady­
stste probability vector of the performance matrix and
T
then identifies the states which indicate unsatisfied
load demand . The sum of the steady-state probabilities
P
u3
a E P
j
(S,I,L) (26)
j=1
corresponding to the cases of unsatisfied demand con­
stitutes the portion of time the system fails to satisfy
S.T.
the load. If we denote this portion of time byF. then
the service performance level is: S
min < S < Smin + LIb (27)
SPL �l-F . (19)
A - L < 0 (28)
O d

428 1989 PROCEEDINGS Annual REUABIUTY AND MAlNTAINABIUTY Symposium


5 + {(A - L)/[I + «L - A )/L) (r - I)l) < 5 Table 1. PV System Design for the
O O min
Boston Location
(29)
PV Array
B. Calculation of F
Number of Modules Per Series String ; 2
(30) Number of Parallel Strings of Modules 47
Total Number of Modu le s ; 94
where P ' P ' and P are the sum of the steady-state Array Tilt Angle ; 65·
uI u2 u3
probabilities associated with Cases I, 2, and 3, respec­ Module Area ; 0.436 m2
tively. Module Maximum Power at 28·C ; 40 Watts
Module Packing Efficiency ; 70.6%
In addition to the 5PL, the long-term battery per­ Cell Efficiency at 28·C ; 13%
formance can also be determined from the performance
matrix steady-state probabilities. In fact, the per­ Battery Subsystem
formance matrix and its steady-state probability vector
contains all the information needed to analyze the per­ Number of Batteries Per Series String ; 12
formance of the PV system. Number of Parallel Strings of Batteries = 7
Total N umbe r of Batteries = 84
Model Validation Capacity Per Battery = 50 Amp-Hrs
Maximum Depth of Discharge = 50%
In o rde r to demonstrate and validate the Markov Round Trip Efficiency ; 75%
chain approach, the model developed was applied to a
radio repeater system for the Boston location. For the System Voltage ; 24 Volts
given PV system design shown in Table 1 and for an
average daily load demand of 240 amp-hrs, the model
yielded a 98.3 percent service performance level. That
Table 2. Battery Long-Term Performance
is_ for sllC"h a gy�t�m design� the load demand is
expected to be satisfied 98.3 percent of the time. On Battery Cumulative
the other hand, the l ong-term battery performance sum­ State of Charge (%) Probabilities
marized in Table 2 indicates that 13.32 percent of the
time the battery is expected to be fully charged. To 100 0.1332
check for the validity of the model, the same example 99 0.1269
application was used. The performance result of a 98 0.1906
system design of 94 modules and 96 batteries were com­ 97 0.02152
pared to results obtained by SOLAVOLT International (51) 96 0.2376
dynamic simulation program (Ref. 5). Both results 95 0.2892
showed that for such a system design, the load demand is 94 0.3367
expected to be satisfied 100 percent of the time. For 93 0.3805
further check, results obtained by the Markov chain 92 0.4208
model were also compared to results obtained by SEOSCAR 91 0.4579
programs (Ref. 1). The results of numerous c omput er 90 0.5253
runs similar to those of Table 3 showed a goo d comparison
with 5EOSCAR resu lts. Finally, the results of the model
were checked for consist ency. Computer runs for differ­
ent battery and array sizes were performed and found to
yield consistent results. For instance, when the model Table 3. Comparison of Markov Chain (MC) Model
was run for 84 batteries and 94 modules, the SPL was
Results to SEOSCAR Res ult s
98.3 percent while for 96 batteries and 94 modules the
SPL increased to 100 percent. This shows that for a System Size SPL (%)
larger battery subsystem and a constan t array siz e , the
chance that the system will meet the load demand is Modules Ba t t eries Me Mod el SEOSCAR
high er. On the other hand, the model results showed
that for smaller array size and constant battery subsys­ 82 102 89.3 90
tem, the SPL is lower. 82 96 89.3 90
88 60 89.0 90
Conclusions 90 48 90.5 90
86 120 97.9 99
The Markov Chain model presented in this paper has 88 96 98.2 99
proven to yield results consistent and comparable to 94 84 98.3 99
other commonly used models. The major advantage of th e
model is that it provides a probabilistic approach for
mod eling stand-alone PV power systems. An approach that
system. Such an expansion requires the modification of
allows designers and developers of these systems to
th e state space of the p erformanc e matrix and the flow
analyze the system as well as the battery performance
of energy equations.
in the long run and determine the s ystem design that
satisfies a required service performance level. Also,
Ref er ences
since the model incorporates Markov chains into the
dynamics of the system, the model is expected to reduce
1. Rosenblum, L. (University of Michigan): "Prac­
the large CPU time required by standard dynamic simula­
tical Aspects of Photovoltaic Technology, Applications,
tion models when used for design optimization.
and Cost." Prepared for NASA-Lewis Research Center,
NASA CR-168025, December 1982.
It is wo rthy to note that although the model
2. NASA-Lewis Research Center, Photovoltaics
d escr ib ed was developed for a simple stand-alone PV
Application Section, Cleveland, Ohio. Private communica­
po wer system with a single constant load, the model can
tion with J. Martz and A. Ratajczak.
be expanded to include multiple loads and a backup

1989 PROCEEDINGS Annual RELIABILITY AND MAINTAINABILITY Symposium 429


3. Lin, J. K.: "Optimization of Terrestrial Fayssal M. Safie, Ph.D.
Photovoltaic Systems," IEEE, Photovoltaic Specialists NASA/Marshall Space Flight Center
Conferences, June 1978, � 1166-1170. Mail Code CT13
4. Hoover, E., (Sandia Laboratories, Alburquerque, Huntsville, Alabama 35812
New Mexico): "SOLCEL-Il: An Improved Photovoltaic System
Analysis Program." SAN 79-1785, February 1980. Dr. F. Safie is a Reliability Engineer at NASA/Marshall
5. Kaszeta, W. J., (Solarvolt International): Space Flight C enter (MSFC). Before joining MSFC in
"Qualification Testing of Solar Photovoltaic Powered 1986, he served as a visiting Assistant Professor at
Refrigerator Freezers for Medical Use in Remote Geo­ Cleveland State University. Since joining MSFC, Dr.
graphic Locations." Prepared for NASA-Lewis Research Safie has been involved in several reliability studies
Center under Contract No. DEN 3-240. related to the various elements of the Space Shuttle
6. Gupta, Y. P.: "A Solar Photovoltaic Flat Plate System. Dr. Safie received his Ph.D. from Cleveland
Applications Experiment at the Oklahoma C enter for State University. He is a member of ASQC and a senior
Science and Arts." Science Applications, Inc., June member of AIlE.
1979.
7. Mustacchi, C., et al.: "Stochastic Simulation
of Hourly Global Radiation Sequences," Solar Energy,
Vol. 23, 1979, pp. 4 7-51.
8. Brinkworth, B. J.: "Autocorrelation and
Stochastic Modeling of Insolation Sequences." Solar
Energy, Vol. 19, 1977, pp. 343-347.
9. Kelton, C. M. L.: "Estimation of Time­
Independent Markov Process with Aggregate Data: A Com­
parison of Techniques." Econometrica, Vol. 49, No. 2,
March 1981.

-- - - -- 430--- ---- 1989 PROCFFniNGS Annual REUABIUTY AND MAINTAINABIUTY Symposium

You might also like