Copyright
C “IIE”
A distributed generation system that integrates wind and solar energy has emerged as a new paradigm for power supply. The goal of
distributed generation planning is to determine the generation capacity, placement, and maintenance such that the system performance
is optimized. In prior studies, the decision is often made assuming a deterministic operating condition. Power intermittency and
equipment cost are the main challenges in deploying a wind- and solar-based energy solution. This article proposes a multi-
criteria generation planning model to maximize the renewable energy throughput while minimizing its cost. The system is designed
under stringent reliability and power quality criteria stipulated as loss-of-load-probability and voltage variation, respectively. A
variance propagation model is derived to decouple the voltage correlations between upstream and downstream nodes. A two-stage,
metaheuristic algorithm is developed to search for the non-dominant solution set. Our approach differs from existing planning
methods in that we implement a statistical quality control mechanism to reduce the voltage drops. A 13-node distribution network is
used to demonstrate the performance of the proposed method.
Keywords: Distributed energy resources, onsite generation, multi-criteria optimization, power quality, variation propagation
0740-817X
C 2015 “IIE”
Integrating distributed wind and solar generation 1123
methods can be classified into two categories: deterministic PV cells as a new source into the DG portfolio along with
modeling and probabilistic modeling. In the former, both WTs.
the load and the power are treated as constant parameters in The contributions of this article are twofold. First, in-
the optimization process (Rau and Wan, 1994; El-Khattam spired by the multistage assembly manufacturing design
et al., 2004; Celli et al., 2005). Deterministic planning mod- principle (Shi and Zhou, 2009), we decouple the corre-
els are applicable to generating units with a relatively stable lation between nodal voltages and propagate its varia-
output power, such as fuel cells, micro-turbines, and diesel tion along the distribution line based on the mean and
engines. the variance data. We show that WTs and PV cells are
In probabilistic planning models, the load and the power cost-competitive solutions to improving power quality and
are treated as random variables. Wang and Nehrir (2004) energy reliability, despite their power intermittency. Sec-
make use of time-series data to characterize the non- ond, we propose a two-stage decision process, in which
stationary load profile in their optimization model. Haghi- the first stage allocates DER capacity and placement and
fam et al. (2008) characterize the load pattern and the the second stage optimizes their maintenance times. A
system power as a time-varying process considering de- two-stage optimization allows the system planner to dis-
mand variability and wind speed volatility. In these models, card inferior design options at an early stage, hence sav-
both the load and the power are divided into several levels, ing on computational efforts. In application, this article
each being assigned a discrete probability value. Recently, enriches the reliability engineering and quality manage-
continuous probabilistic modeling is increasingly used, due ment literature by extending it from multistage assembly
to advanced computing and data analytics. Based on the manufacturing systems to distributed energy production
normality assumption, Liu et al. (2011) use a chance con- systems.
straint to mitigate the thermal stresses of distribution lines. This article is organized as follows. Sections 2 and 3 de-
Novoa and Jin (2011) and Jin et al. (2013) characterize fine the energy reliability and power quality criteria for
the loss-of-load probability and the thermal constraint us- DG systems, respectively. The voltage variation propaga-
ing the statistical moments of nodal demand and DER tion model is also derived in Section 3. Section 4 estimates
power. the energy cost and the carbon savings under power in-
Studies (Wiser and Bolinger, 2008; Koutroulis and Blaab- termittency. Section 5 presents a multi-criteria DG plan-
jerg, 2013) show that maintenance and repair of WT and ning model along with the two-stage optimization algo-
PV equipment may account for 20–25% of their life cycle rithm. In Section 6, we demonstrate the proposed model
cost. In most DG planning models, preventive maintenance on a 13-node distribution network. Section 7 concludes the
is rarely treated as a critical lever to lower the system cost. article.
Jin et al. (2013) take an early step to jointly coordinate
the maintenance policy with the sizing and siting of WTs
to minimize the energy cost as well as to maximize equip-
ment reliability. Power quality is often manifested as voltage 2. Energy reliability in DG systems
variation along the distribution lines. In prior studies, the
control scheme for the nodal voltage is often solely limited Figure 1(a) shows a typical centralized distribution sys-
to the mean value. The increase of the types and number of tem consisting of one substation and five nodes (or buses)
variable generation units causes more uncertainties in nodal with D1 to D5 being the power demands. Connected to large
voltage. Hence, the fluctuation of nodal voltage needs to be power plants through the main grid, the substation supplies
precisely quantified and appropriately incorporated in the all the electrical energy to these nodes. In a DG system (see
DG planning. Fig. 1(b)), WTs and solar PV units can be placed on any
This article proposes a multi-criteria DG planning model nodes to co-supply the power along with the substation.
to facilitate the integration of wind and solar energy into Although wind and solar energy can reduce the carbon
a distribution network with dual objectives: first, minimiz- footprint and improve the nodal voltage, power intermit-
ing the energy production cost and, second, maximizing tency and equipment cost are major concerns. Hence, the
the renewable energy throughput. We extend and general- goal for planning a DG system is to minimize the system
ize the framework of Jin et al. (2013) in three aspects. First, cost and to maximize the power output with ensured energy
we develop a variance propagation model that is used in a reliability and power quality.
statistical power quality control scheme to ensure voltage We adopt the moment method of Novoa and Jin (2011)
stability. Second, we propose a two-stage optimization al- to characterize the power intermittency and load variabil-
gorithm to sequentially optimize the sizing and the siting ity. Let m be the number of DER types that can be poten-
of DER units and their maintenance time. Third, instead tially placed in the distribution network. For instance, if
of maximizing equipment reliability, this article maximizes the available DER units are a 1 MW WT, 2 MW WT, and
the annual carbon savings as one of the objectives. Hence, 0.5 MW PV cell, then m = 3. Let n be the number of nodes
we expand the generation portfolio by incorporating solar where the DER units can be placed. For instance, n = 5 in
Fig. 1(b). Let xij be the number of DERs of type i to be
1124 Jin et al.
placed on node j. Then x = [x11 , x12 , . . ., xmn ] is the decision distribution network. Based on the Central Limit Theo-
vector corresponding to the sizing and siting of DER units. rem, P(x) can be approximated as a normal distribution
The aggregate system power and load, denoted as P(x) and for large n. A discussion of compact distribution networks
D, are given as in suburban areas, where nodal generations are correlated,
is available in Jin et al. (2013).
m
n
P(x) = xi j Pi j , (1) Loss-Of-Load-Probability (LOLP) is a fundamental
i =1 j =1
metric used by utility companies to measure the energy
supply reliability. It is defined as the probability that the
power is less than the load at any point in time. Let α 1 be
n
the maximum acceptable LOLP value. The energy reliabil-
D= Dj , (2)
ity criterion for a DG system can be expressed as
j =1
Pr {P(x) < D} ≤ α1 . (7)
where Pij is the power output of a single unit of DER type
i at node j and Dj is the load of node j. Both Pij and Dj are If the distribution functions of P(x) and D are known, the
treated as random variables. The mean and the variance of relation between x and α 1 can be explicitly characterized.
D and P(x) can be estimated as follows: This will be further discussed in Section 5 where Equation
(7) is translated into its deterministic counterpart under the
n
μ D = E[D] = E[D j ], (3) normality condition.
j =1
3.1. Model of the propagation of the variation of the voltage Here Pk (x) is the aggregate renewable power on node k,
Without loss of generality, we use a four-node distribu- and Pk(x) = im xi k Pi k for k = j, j + 1, ...., J. A distinction
tion line, as shown in in Fig. 2, to derive a generic model must be made between J and n as the latter represents the
of the propagation of the nodal voltage under variable total number of nodes in a distribution network that may
power integration. Both Ohm’s law and Kirchhoff’s law consist of multiple feeders (see Fig. 4). Although Equation
are used during the derivation. Ohm’s law states that the (14) establishes the quantitative relation between Vk and Ij ,
voltage drop is always equal to the product of the line resis- it still does not allow us to calculate Vk , due to the unknown
tance and the current. Kirchhoff’s law states that the sum value of Ij . For a distribution network operating in a stable
of the currents entering a node is the same as the sum of condition, the voltage variation between the substation and
the currents exiting that node. its downstream nodes is relatively small compared with
Let Vj denote the voltage at node j for j = 1, 2, 3, and VDG . Hence, the following approximation generally holds:
4 by realizing that j = 1 is the substation node, and VDG
is the nominal operating voltage. Note that Dj and Pj (x), VDG ∼
= V1 ∼
= ··· ∼
= VJ . (15)
respectively, represent the nodal demand and the aggregate
renewable power on node j, and Rj is the line resistance Using this approximation allows us to estimate the line
between nodes j and j + 1. We assume that all renew- current without the necessity of solving the system of com-
able power is absorbed by the local node. This assumption plex simultaneous power flow equations. By substituting
is reasonable due to the fact that today’s distribution net- Equation (15) into Equation (14), the current leaving node
works are generally not capable of handling reverse current, j can be approximated as
which may damage the upstream transformers and circuits.
1
J
Technically, power flow is similar to a multistage assembly
Ij ∼
= (Dk − Pk(x)), for j = 1, 2, . . . , J − 1.
manufacturing system where materials usually move from VDG
k= j +1
upstream to downstream in a unidirectional mode (Shi and
(16)
Zhou, 2009). Starting from node 4, according to Kirch-
Equation (16) represents an important result as it allows
hoff’s law, we have
us to estimate the line current solely based on Dk and Pk (x).
I3 = ID4 − IP4 , (9) By applying Ohm’s law between nodes 1 and 2, the voltage
where I 3 is the current from node 3 to node 4, IP4 is the at node 2 can be expressed as
current supplied by DER 4, and ID4 is the current entering
R1
J
D4 . By applying Ohm’s law, I 3 can be estimated as V2 ∼
= V1 − R1 I1 = V1 − (Dk − Pk(x)), for j = 2.
D4 − P4 (x) VDG
k=2
I3 = , (10) (17)
V4
Similarly, by using the chain rule, the voltage of all the
where P4 (x) is the aggregate DER power at node 4, and V 4 downstream nodes from the substation to node J can be
is the voltage of node 4. Similarly, by applying Kirchhoff’s estimated accordingly, and the result is given as follows:
law to node 3, we have
Rj −1
J
I2 = ID3 + I3 − IP3 . (11)
Vj ∼
= Vj −1 − (Dk − Pk(x)), for j = 3, 4, . . . ,J.
By substituting Equation (10) into Equation (11) along VDG k= j
with the use of Ohm’s law on node 3, we have (18)
D3 − P3 (x) D4 − P4 (x) Note that Vj-1 is the voltage of the adjacent upstream
I2 = + , (12) node. When j = 1, we have V 1 = VDG .
V3 V4
1126 Jin et al.
3.2. Mean and variance of Vj and Ij independent for given k. Now the mean and the variance
of Vj can be obtained as follows:
Obviously Vj and Ij in Equations (16) and (18) are random
variables due to the stochastic nature of Pj (x) and Dj . The 1
J
objective of this section is to estimate the mean and the vari- μVj (x) ∼
= VDG − rk (E[Dk] − E[Pk(x)]),
ance of Vj and Ij based on the first two moments of Pj (x) VDG
k=2
or Dj . Equation (18) shows that Vj at node j depends on the for j = 2, 3, . . . ,J, (26)
upstream voltage and the downstream demand and power.
A close analogy to Vj is the multi-operational machining
1 2
process where the overall parts variation at a current stage J
is influenced by the variations of the preceding stage and σV2j (x) ∼
= 2
rk (Var(Dk) + Var(Pk(x))),
VDG
the errors from the current stage (Huang and Shi, 2004). k=2
It is usually difficult to directly estimate the mean and the for j = 2, 3, . . . ,J. (27)
variance of Vj due to the voltage correlations among the
To ensure the power quality, the voltage variation on
upper and downstream nodes. Here we present a variance
node j is controlled according to Equation (8) by placing
propagation model, in which Vj (x) is decomposed as a su-
WTs and PV units across the distribution network. Based
perposition of downstream power and load only. Let us
on Equation (16), we can also estimate the mean and the
rewrite Equations (17) and (18) as follows:
variance of the line current as follows:
R1
J
1
J
for j = 2 : ∼
V2 = VDG − (Dk − Pk(x)),
VDG
(19) E[I j ] ∼
= (E[Dk] − E[Pk(x)]),
k=2 VDG
k= j +1
for j = 1, 2, . . . , J − 1, (28)
R2
J
for j = 3 : V3 ∼
= V2 − (Dk − Pk(x)), (20)
VDG
1
k=3 J
Var(I j ) ∼
= 2
(Var(Dk) + Var(Pk(x))),
VDG
Rj −1
J k= j +1
for any j : Vj ∼
= Vj −1 − (Dk − Pk(x)), (21) for j = 1, 2, . . . , J − 1, (29)
VDG k= j
where Ij is the current flowing from node j to j + 1 via link
j, and there is a total of J − 1 links in the interconnected
RJ−1
for j =J VJ ∼
= VJ−1 − (DJ − PJ (x)). (22) distribution line.
VDG We use the system in Fig. 1(b) to illustrate how DER
Summation of Equations (19) to (22) results in the fol- units improve the voltage profile across a distribution line.
lowing voltage estimate: Let VDG = 33 kV, Dj = 5 MW for j = 1, 2, . . ., 5, and Rj =
1 for the line resistance between two adjacent nodes. For
j −1
1
J
this example, the resistance between substation and node 1
Vj ∼
= VDG − Ri (Dk − Pk(x)), is R0 = 1 as well. It is assumed that a 2 MW PV system
VDG
i =1 k=i +1 is installed at node 5, and a WT is to be placed at node 3.
for j = 2, 3, . . . ,J. (23) Based on Equation (24), we compute the nodal voltage by
In fact, Equation (23) can be expressed more concisely varying the wind power between 0 and 5 MW. As shown in
as Fig. 3, the voltage of all nodes increases with the amount of
wind energy. For instance, given P3 = 5 MW, the voltage of
1
J
nodes 3 becomes 31.82 kV as opposed to 31.18 kV prior to
Vj ∼
= VDG − rk(Dk − Pk(x)), for j = 2, 3, . . . ,J, integration. An interesting observation is that a node with-
VDG
k=2
out a DER unit also benefits from its neighboring nodes
(24)
where DER units are placed. For instance, the voltage of
with
⎧ k−1 nodes 2 and 4 increases as more wind power is injected
⎪ into node 3. These observations further show that voltage
⎪
⎨ Ri , for 2 ≤ k ≤ j − 1,
i =1
variation is mutually correlated long the same line.
rk = (25)
⎪ j
⎪
−1
⎩ Ri , for j ≤ k ≤ J.
i =1
4. System cost and carbon savings
Pk (x) depends on the local wind speed and solar irradi-
ance, whereas the power demand Dk is associated with the Two quantitative methods are commonly used to predict
users’ activities on node k. Thus, Pk (x) and Dk are mutually the output of WT and PV generating units. One is to directly
Integrating distributed wind and solar generation 1127
where Pm is the rated power or the turbine capacity, and γ
= Pm /vr 3. The WT power curve consists of four operating
phases: the standby phase when 0 < y < vc ; the nonlinear
production phase when vc ≤ y < vr ; the constant output
phase when vr ≤ y ≤ vs ; and the cut-off phase when y > vs .
Let Pw (Y ) be the random output power at wind speed Y .
The mean and the variance of Pw (Y ) can be estimated by
vr
E[Pw (Y)] = γ y3 fw (y)dy + Pm (Fw (vs ) − Fw (vr )) ,
vc
(31)
vr
E[Pw2 (Y)] = γ 2 y6 fw (y)dy + Pm2 (Fw (vs ) − Fw (vr )) ,
Fig. 3. Impact of wind and solar power on nodal voltage. vc
(32)
use the historical data to estimate the energy throughput Var(Pw (Y)) = E[Pw2 (Y)] − (E[Pw (Y)])2 , (33)
using time-series models (Billinton et al., 1996). Another
approach, also called the moment method, combines the where Fw (y) and fw (y) are the cumulative distribution and
wind speed (or solar irradiance) distribution with the DER probability density functions of Y , respectively. As shown
power curve to estimate the mean and variance of the out- by Karki et al. (2006) and Weekes and Tomlin (2014), Y
put power (Novoa and Jin, 2011). The moment method can be fitted with normal or Weibull distributions depend-
is preferred when historical generation data are scarce or ing on the meteorological data. For a Weibull wind speed
not available for new DG planning. We briefly review the distribution with scale parameter cw and shape parameters
moment-based prediction method and further discuss the dw , its probability density function is
DG life cycle cost and the potential carbon savings.
dw −1
dw y
e−(y/cw ) , for y ≥ 0. (34)
dw
fw (y) =
cw cw
4.1. Modeling wind power generation
Let vc be the cut-in speed, vr be the rated speed, and vs be
the cut-off speed for a WT. Assume Y is the random wind 4.2. Modeling solar energy generation from a PV system
speed, and y is its realization. The following cubic power The output power of a PV system depends on multiple
curve is commonly used to characterize the output power factors, including the panel size and orientation, the PV
of a WT system (Jangamshetti and Rau, 2001): efficiency, the tilt angle, the daily hour, the latitude, and the
⎧ weather condition. Let S be the random solar irradiance
⎨0 0 < y < vc , or y > vs incident on the PV panel and s is its realization. A generic
Pw (y) = γ y3 vc ≤ y < vr , (30) PV output model incorporating all these factors is given by
⎩ P vr ≤ y ≤ vs
m Lave and Kleissl (2011) as follows:
Ps (S) = ηs As S (1 − 0.005(To − 25)) , (35)
where Ps (S) is the output of a PV cell in W/m2, ηs is the
efficiency of the cell and it is between 10 and 20% for current
commercial products, As is the panel size (m2), and To is the
operating temperature of the cell (◦ C). According to Atwa
et al. (2009), S for a specific region during the course of a
year can be modeled by a beta distribution as follows:
as −1
(as + bs ) s s bs −1
fs (s) = 1− ,
sm (as )(bs ) sm sm
for 0 ≤ s ≤ sm , (36)
where sm in W/m2 is the maximum solar irradiance on the
PV surface in a year. Both as and bs are the shape parame-
Fig. 4. A 13-node distribution network with multiple feeders. ters of the beta distribution. The mean and the variance of
1128 Jin et al.
Ps (S) are now given as follows: type i at node j. Although there are no fuel consumptions
as sm ηs As (1 − 0.005(To − 25)) for WT and PV units, a major operating cost is associated
E[Ps (S)] = , (37) with leasing the land for placing the DER units (U.S. De-
as + bs partment of Energy, 2014). In Equation (42), cij is a negative
as bs sm2 ηs2 A2s (1 − 0.005(To − 25))2 coefficient representing the carbon credits received by the
Var(Ps (S)) = . (38)
(as + bs )2 (as + bs + 1) DG investors. ti a is the annual operating hours, and ti a =
8740 hours for a WT and 4380 hours for a PV cell (solar
is only available in the daytime). Pij is the random output
4.3. Annualized cost of a DG system power of a DER of type i at node j and varies between zero
and Pij c.
The cost of a DG system typically consists of installa-
Equation (43) captures the system maintenance cost with
tion cost Ce (x), operation cost Co (x), maintenance cost
ui (τ i ) being the cost per unit time for DER type i. Note that
Cm (x, τ ), and carbon credits Cc (x). The last item is the
Ri (t) is the reliability function, and Fi (t) = 1 − Ri (t). Costs cfi
income to the renewable DG investor. We first revisit the
and cpi are associated with performing a corrective replace-
mean cost model by Jin et al. (2013) and further derive
ment (i.e., failed) and a planned replacement, respectively.
the cost variance estimate. For maintenance strategy, we
The times required to perform a corrective or a planned
adopt time-based Preventative Maintenance (PM) policy
replacement are tfi and tpi , respectively. The definition of
for DER units as it is widely used by the power industry
ui (τ i ) is slightly different from existing PM models in that
(Bertling et al., 2005). Nevertheless, the proposed planning
the equipment downtime tfi F(τ i )+tpi R(τ i ) is also included
model can be extended to DG systems under condition-
in the denominator in Equation (43). If the downtime is
based maintenance (Liao et al., 2006; Byon et al., 2010;
small, it is usually ignored for computational simplicity
Ye et al., 2014). Let x = [x11 , x12 , . . ., xmn ] and τ =
(Chen and Jin, 2003; Elsayed, 2012).
[τ 1 , τ 2 ,. . . τ m ] be the decision vectors, corresponding to
Valenzuela and Marzumdar (2000) argue that the cost
DER sizing and siting and maintenance times, respectively.
of producing electricity is a random variable, as it depends
The annualized DG system cost is given as follows:
on the availability of generating units and the uncertain
CDG (x, τ ) = Ce (x) + Co (x) + Cc (x) + Cm (x, τ ), (39) demand. This is echoed by Equation (39) where CDG (x, τ )
is a random variable due to the stochastic nature of Dj and
with Pij . Particularly, the cost variation of a DG system can be
m
n
r (1 + r )h ascribed to the variable output of WT and PV units. The
Ce (x) = xi j ai j Picj mean and the variance of CDG (x, τ ) can be estimated as
(1 + r )h − 1
i =1 j =1
m
n
m
n
m
n
=φ xi j ai j Picj , (40) E[CDG (x, τ )] = φ xi j ai j Picj + xi j bi j Picj
i =1 j =1 i =1 j =1 i =1 j =1
m n
m
n + tia xi j ci j E[Pi j ]
Co (x) = xi j bi j Picj , (41) i =1 j =1
i =1 j =1
m
n
+ tia xi j u i (τi ),
m
n i =1 j =1
Cc (x) = tia xi j ci j Pi j , (42) (44)
i =1 j =1
m
n
Var(CDG (x, τ )) = Var(CDG (x)) = tia xi2j ci2j Var(Pi j ).
⎛
⎞
i =1 j =1
m
ti c f i Fi (τi ) + c pi Ri (τi )
a
n
Cm (x, τ ) = ⎝ τ xi j ⎠ (45)
i =1
Ri (t)dt + t f i Fi (τi ) + tpi Ri (τi )
0
i
j =1
⎛ ⎞
m
n
= ⎝tia u i (τi ) xi j ⎠. (43) 4.4. Carbon savings
i =1 j =1
A major motivation behind the use of renewable energy is
In Equation (40), φ = [r (1 + r )h ]/[(1 + r )h − 1] is the to control and reduce carbon emissions. Let q be the carbon
capital recovery factor where h is the years to pay off the intensity; i.e., the amount of CO2 emitted when producing
loan and r is the annual interest rate. Note that aij is the 1 MWh of electricity. Depending on the type of fossil fuel
capacity cost per MW when installing DER type i at node (i.e., gas, oil, or coal), it is generally known that q = 0.6–0.9
j, and Pij c is the allocated capacity of DER type i at node j. ton/MWh. For a renewable DG system, the amount of
In Equation (41), bij is the operating cost per MW of DER carbon emissions avoided (i.e., carbon savings) in a year,
Integrating distributed wind and solar generation 1129
denoted as Q(x, τ ), can be estimated as in node j should not exceed 100λ percent of the mean load.
Here λ is the renewables penetration rate and λ ∈ [0, 1]. This
m
n
Q(x, τ ) = q tia Ai (τi )xi j Pi j , (46) requirement is necessary to prevent the upstream circuits
i =1 j =1
from being damaged by a reverse current. With the emer-
gence of smart grid systems, two-way energy flow becomes
with feasible and λ could be larger than one.
τi
Ri (t)dt
Ai (τi ) = τ i 0
,
0 Ri (t)dt + t f i F i (τi ) + tpi Ri (τi ) 5.2. Risk-averse multi-criteria optimization
for i = 1, 2, . . . ,m, (47)
Although the formulation in Problem P1 accommodates
where Ai (τ i ) is the operational availability of DER type i. power intermittency and load variation, it is difficult, if not
If the lifetime of a DER unit follows the Weibull distribu- impossible, to find the optimal solution given the random
tion with scale parameter ηi and shape parameter β i , its nature of Pij and Dj . To make the problem tractable, we
reliability function is transform Problem P1 to a deterministic counterpart as
βi follows.
− t
Ri (t) = e ηi
, for t ≥ 0, and i = 1, 2, . . . , m. (48)
Problem P2:
Min : f1 (x, τ ) = E[CDG (x, τ )] + Z1−θ Var(CDG (x, τ )),
5. Multi-criteria planning model (55)
We design a renewable DG system to achieve two objectives: Max : f2 (x, τ ) = q tia xi j Ai (τi )E[Pi j ], (56)
minimizing the energy production cost and maximizing the i =1 j =1
Problem P3:
Stage 1:
6. Numerical experiments
j=1 SS SS SS SS SS SS SS SS
2 5 4 4 4 5 3 1 3
3 4 5 5 5 4 5 5 5
4 5 5 5 5 5 5 5 5
5 0 2 2 2 2 1 2 2
6 0 2 2 3 1 3 3 3
7 5 1 1 2 2 3 3 3
8 2 2 2 2 3 3 3 3
9 1 0 0 2 0 1 1 2
10 2 1 3 1 3 2 3 2
11 1 2 0 1 2 1 3 3
12 4 5 5 1 2 1 1 3
13 1 0 2 1 2 2 2 2
τ 1 (hour) 2860 2793 2876 2943 2936 3019 3105 n/a
τ 2 (hour) 2820 2734 2808 2892 2889 3024 3118 2734
τ 3 (hour) n/a n/a 2869 2925 2957 3041 3195 2773
τ 4 (hour) 10 732 10 615 10 743 10 887 10 860 n/a n/a n/a
τ 5 (hour) 7644 7419 7635 7861 7886 8150 8490 7419
1134 Jin et al.
Table 7. LOLP and lower voltage limit (unit: kV) at α 2 = 90%
Solution 1 2 3 4 5 6 7 8
LOLP
(×10−4) 9.2 9.7 9.2 9.9 9.9 9.5 8.7 6.3
LOLP is suppressed below α 1 = 0.001 in all eight candidate approach that employs local search methods in mathemati-
designs. We also estimate the lower voltage limits of indi- cal programing problems. As a key feature of TS, short-term
vidual nodes at the α 2 = 90% confidence level. As shown memory is used to avoid re-evaluation of solutions already
in Table 7, the lower voltage limit of all the nodes is no less assessed, thus achieving diversification and intensification
than the minimum threshold V min = 31.35 kV, as opposed of the search. TS is quite effective in tackling non-linear
to three nodes that fail to meet this criterion in the original mixed-integer programing problems arising from DG plan-
system (see Fig. 5). Finally, we check the thermal condition ning (Novoa and Jin, 2011; Mohammadi and Nafar, 2013;
of individual links based on the Pareto solutions, and the Arefifar and Mohamed, 2014). In our setting, TS is em-
results are listed in Table 8. The upper current limits of all ployed to search for the sizing and siting of DER units in
12 links are smaller than their maximum thresholds at α 3 Stage 1, and in Stage 2 the gradient method further identi-
= 90% confidence level. In the original network, links 1, 4, fies the maintenance times.
and 10 violate the thermal constraints due to the excessive The TS algorithm is implemented in Matlab 2013 and
current flow (see Fig. 6). run on a Dell PC featuring an Intel(R) Core 2 Duo CPU
@2.93GHz and 8 GB memory. Figure 8 plots the Pareto op-
timality in Stage 2 generated from the TS-based algorithm
6.6. Comparing with tabu search
in conjunction with those from the GA-based algorithm
For comparison purposes, problem P3 is also solved us- (see Table 6). TS can be easily implemented using a high-
ing the Tabu Search (TS) algorithm. Originally created by level computer language such as Matlab and also possesses
Glover (Glover, 1986), TS is a metaheuristic optimization computational advantages over GA in terms of time. On
j=1 SS SS SS SS SS SS SS SS
2 5 1 5 5 2 5 2 3
3 4 5 4 4 5 4 5 5
4 5 5 5 5 5 5 5 5
5 2 4 2 2 2 4 5 5
6 5 2 5 5 4 5 2 2
7 0 5 0 1 2 3 2 2
8 4 3 3 3 2 2 5 5
9 2 4 2 2 3 3 3 3
10 3 0 3 3 1 5 3 3
11 4 2 4 4 2 3 2 2
12 5 4 5 5 4 3 3 3
13 0 1 0 0 2 2 5 5
τ 1 (hour) n/a 2874 n/a 2805 2834 n/a n/a n/a
τ 2 (hour) 2708 2760 2759 2717 2785 2779 2791 2692
τ 3 (hour) 2763 2787 2728 2845 2765 2808 2806 2755
τ 4 (hour) 10 657 10 545 10 675 10 646 10 699 10 579 n/a n/a
τ 5 (hour) 7480 7 438 7410 7410 7452 7422 7378 7428
7. Conclusions
A2. The derivative of A(τ ) where w1 + w2 = 1. By taking the derivative with respect to
The operational availability of a DER unit is given by τ i (for i = 1, 2, . . ., m), and setting it to zero, we have
Equation (47), and we rewrite A(τ ) by suppressing the index ∂ fc (τ ; x)
as follows: = w1 ϕi (c f i − c pi ) − w2 δi (tpi − t f i )
τ ∂τi
τi
0 R(t)dt
A(τ ) = τ . (A5) × Ri (τi ) + fi (τi )
2
Ri (t)dt
0 R(t)dt + (tp − t f )R(τ ) + t f 0
1138 Jin et al.
−Ri (τi ) w1 ϕi c f i + w2 δi t f i M.S. degrees from the Nanjing University of Aeronautics and Astronau-
tics University and a Ph.D. in Systems Engineering from Southeast Uni-
+w1 ϕi fi (τi )(tpi c f i − t f i c pi ) = 0, (A14) versity, China. She worked as a software engineer at Zhongxing Telecom-
with fi (τi ) = −∂ Ri (τi )/∂τi . The optimal maintenance time munication Equipment Corporation in 2001–2002. During 2008–2009,
she was a visiting scholar at Texas A&M International University. Her
τ i is found by solving Equation (A14). To generate a set research is funded by the National Natural Science Funds of China. Her
of non-dominant solutions, we solve Equation (A14) by current research interests include reliability modeling and system opti-
varying w1 and w2 . mization, project scheduling, power market simulation and optimization,
and bidding mechanism design.