You are on page 1of 16

Applied Energy 99 (2012) 455–470

Contents lists available at SciVerse ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

An efficient scenario-based stochastic programming framework


for multi-objective optimal micro-grid operation
Taher Niknam ⇑, Rasoul Azizipanah-Abarghooee, Mohammad Rasoul Narimani
Department of Electrical and Electronics Engineering, Shiraz University of Technology, Shiraz, Iran

h i g h l i g h t s

" Proposes a stochastic model for optimal energy management.


" Consider uncertainties related to the forecasted values for load demand.
" Consider uncertainties of forecasted values of output power of wind and photovoltaic units.
" Consider uncertainties of forecasted values of market price.
" Present an improved multi-objective teaching–learning-based optimization.

a r t i c l e i n f o a b s t r a c t

Article history: This paper proposes a stochastic model for optimal energy management with the goal of cost and emis-
Received 27 January 2012 sion minimization. In this model, the uncertainties related to the forecasted values for load demand,
Received in revised form 29 February 2012 available output power of wind and photovoltaic units and market price are modeled by a scenario-based
Accepted 12 April 2012
stochastic programming. In the presented method, scenarios are generated by a roulette wheel mecha-
Available online 2 July 2012
nism based on probability distribution functions of the input random variables. Through this method,
the inherent stochastic nature of the proposed problem is released and the problem is decomposed into
Keywords:
a deterministic problem. An improved multi-objective teaching–learning-based optimization is imple-
Improved teaching–learning-based
algorithm
mented to yield the best expected Pareto optimal front. In the proposed stochastic optimization method,
Micro grid a novel self adaptive probabilistic modification strategy is offered to improve the performance of the pre-
Multi-objective stochastic optimization sented algorithm. Also, a set of non-dominated solutions are stored in a repository during the simulation
Renewable energy management process. Meanwhile, the size of the repository is controlled by usage of a fuzzy-based clustering tech-
Self adaptive probabilistic modification nique. The best expected compromise solution stored in the repository is selected via the niching mech-
strategy anism in a way that solutions are encouraged to seek the lesser explored regions. The proposed
Uncertainty framework is applied in a typical grid-connected micro grid in order to verify its efficiency and feasibility.
Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction and consequently appearance of the more advanced energy storage


facilities, these devices are playing an important role in low voltage
Deregulating in power systems has motivated the electricity power networks. They can save energy at low-price hours and sell it
companies to utilize the Distributed Generators (DGs) near the at high-price hours which this helps the network to work more effi-
energy consumers [1,2]. DGs include different types of power ciently and economically [9]. Meanwhile the operation and control
sources such as diesel engines, Micro Turbines (MTs), Fuel Cells of DG units in accompany with storage facilities such as flywheels,
(FCs), Photo Voltaic (PV), and small Wind Turbines (WTs) [3,4]. energy capacitors and batteries are getting more complicated and
From the consumers’ point of view, using DGs causes the decrease challengeable [10].
in electricity cost, higher service reliability, higher power quality, Above claims and important role of the Micro Grids (MGs) in
increase of energy efficiency, and energy independence [5]. modern power systems have caused many researches to be carried
Furthermore, using Renewable Energy Sources (RESs) like wind or out in order to scrutinize different aspects of the MG problem.
photovoltaic can satisfy the entire environment concerns [6–8]. Chakraborty et al. [11] utilized a linear programming algorithm
Therefore, nowadays with progress in power electronic equipments in order to minimize MG operation cost and to optimize the states
of battery charge. Sortomme and El-Sharkawi [12] used Particle
⇑ Corresponding author. Tel.: +98 7117264121; fax: +98 7117353502. Swarm Optimization (PSO) for decreasing the costs of MGs includ-
E-mail address: niknam@sutech.ac.ir (T. Niknam). ing controllable loads and battery storage by selling stored energy

0306-2619/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.apenergy.2012.04.017
456 T. Niknam et al. / Applied Energy 99 (2012) 455–470

(a) Ordinary MCS scenario-based stochastic programming framework is applied to


1 solve the proposed problem in which at first, the discrete control
0.8 variables (here-and-now) are generated for all generation units
Random Value

and after that the continuous dispatch variables (wait-and-see)


0.6
are produced by scenario based technique [21]. These scenarios
0.4 are generated using the Roulette Wheel Mechanism (RWM) and
0.2 Lattice Monte Carlo Simulations (LMCS) method. Moreover, for
0 avoiding the intractable computation, a scenario reduction tech-
0 20 40 60 80 100 120 140 160 180 nique should be applied.
Sample No. It should be noted that there are four techniques including sce-
nario reduction, fine tuning of the solution algorithm, decomposi-
(b) Rank-1 LMCS tion, and manipulation of the model for reducing the cumbersome
1
computation [21].
0.8 To solve the proposed problem, two aforementioned first tech-
Random Value

0.6 niques are applied in this paper to decline the computation time.
0.4
Also, the Teaching–Learning-Based Optimization (TLBO) algorithm
as one of the fast evolutionary algorithms, is utilized to optimize
0.2
the proposed problem [22]. It should be noted that the proposed
0 Stochastic Multi-objective Optimal MG Operation (SMOMGO) is a
0 20 40 60 80 100 120 140 160 180
mixed integer nonlinear and non-differentiable optimization prob-
Sample No.
lem. The TLBO algorithm is a new efficient optimization algorithm
Fig. 1. Random numbers generated by (a) ordinary MCS and (b) rank-1 LMCS. which has been inspired by a learning mechanism in a class [22].
To avoid trapping in local optima, the proposed optimization algo-
rithm is equipped with a Self Adaptive Probabilistic Modification
at high prices. Hernandez-Aramburo et al. [13] decreased the fuel Strategy (SAPMS) which is called Improved TLBO (ITLBO) algo-
consumption rate of the system while constrained this rate to sat- rithm. Finding a set of solutions in accompany with the best com-
isfy the local electrical and thermal energy demands and provided promise one can help to MG Central Controller (MGCC) to have
a certain minimum reserve power, this study considered operation more flexibility in their choices. In this respect, a fuzzy clustering
management of MGs as a single objective optimization problem. method is used to prune the size of the repository without destroy-
Also, storage facilities and selling or purchasing power to/from ing its characteristics [23]. To incorporate Decision Makers (DMs)
utility were not taken into account. Mohamed and Koivo consid- biased preference through the search process perfectly, a min–
ered the presence of storage systems in [14] but they did not for- max approach is developed to decide on the best candidate solu-
mulate the problem on period with several time intervals and tions for the next optimization part of the proposed algorithm. In
therefore did not consider the issue of charge and discharge cycles order to encourage individuals to seek lesser explored regions with
and other pricing issues in [15]. Logenthiran and Srinivasan [16] higher probability of Pareto-Optimal Front (POF), a niching tech-
developed the classical unit commitment problem using a three nique is employed to select the expected best compromise solution
steps method for minimizing the cost objective function. Morais [23].
et al. [17] proposed the classical unit commitment. Also, all previ- The main contributions of this work can be summarized as fol-
ous researches have considered the operation management of MGs lows: (i) Formulate the SMOMGO by consideration of cost and
as a deterministic problem. Because of uncertainty in forecasted emission objectives simultaneously, (ii) Implement a Pareto-based
values of available output powers of WT and PV units, market approach based on a novel ITLBO for solving the proposed problem,
prices and load demand considering this problem as deterministic (iii) Model and formulate the Spinning Reserve Requirements
one fails to obtain reliable solutions. (SRRs) to overcome the sudden fault in the system, (iv) Decrease
Dukpa et al. [18] proposed a new optimal participation strategy the computation time with backward scenario reduction method
for a wind power generator that employs an energy storage device (v) The performance and potential of the proposed approach is suc-
for participating in a day-ahead unit commitment process consid- cessfully validated with numerical simulations.
ering stochastic power output. The authors have also proposed a
novel smart energy management system which can coordinate
2. Stochastic model description
power forecasting, energy storage and energy exchanging and then
make a proper short-term scheduling to minimize the total opera-
To solve the proposed problem, a two-stage stochastic scenario-
tion cost.
based method is implemented in this paper. At the first stage, all
The important drawback of above study is that it does not con-
discrete control variables (here-and-now) are made for all genera-
sider all uncertainties of the problem. Also, nowadays restraining
tion units before the uncertainty of input random variables is real-
emissions of greenhouse gases becomes one of the most important
ized. The uncertain variables are the load demands, the market
issues of the electricity generation sector. The MG operators need
prices, and powers output of PV and WT units in the study horizon.
powerful energy management tools which can dispatch the power
These uncertainties are realized using a scenario-based technique
between DG sources, substation and storage facilities in order to
in which scenarios are generated by RWM and LMCS. In the next
satisfy economic and environmental goals. In this regard, solving
stage, continuous dispatch variables (wait-and-see) are generated.
a Multi-objective Optimization Problem (MOP) [19,20] over 24 h
The scenario generation and reduction techniques are formulated
considering the associated uncertainties is inevitable. The novelty
as follows.
of the proposed approach resides in considering of the optimal
scheduling of generation units accompany with the inherent
uncertainties which leads to getting close to actual conditions. Fur- 2.1. Scenario generation
thermore, the proposed approach considers Spinning Reserve
Requirements (SSRs) in order to protect system from sudden fault, Random sampling is a corner stone of the Monte Carlo method,
which can provide high reliability for MG systems. The two-stage hence using a most efficient method for scenario generation is
T. Niknam et al. / Applied Energy 99 (2012) 455–470 457

Interval 1

Interval 2

Interval 3

Interval 4

Interval 5

Interval 6
Interval 7
Probability
Density
β 1,t Interval 1

0 1
β 2,t , β 3,t Interval 3 Interval 2
Fig. 3. Accumulated normalized probabilities of the forecast error intervals.

β 4,t , β 5,t Interval 5 Interval 4


error for unit WT, photovoltaic power forecasted error for unit PV,
β 6,t , β 7,t Interval 7 Interval 6
load demand forecasted error for load level ld and price forecasted
-3 σ -2 σ -σ 0 σ 2σ 3σ
error at time t in scenario s, respectively. NWT, NPV, ND, T, and Ns are
Load Forecast Error ( Δ PD ,ld ,t ,s )
the numbers of WT and PV units, load level, time intervals and sce-
narios, correspondingly. Different realizations of the random vari-
Fig. 2. Typical discretization of the PDF of the load level forecast error.
ables and their probabilities are obtained by usage of the
scenario-generation method described in [24]. This scenario-gener-
crucial. In this regard, a random number between [0, 1] is sepa- ation process is based on LMCS and RWM [26] summarized as
rately generated for each input random variable. In the state of follows.
the ordinary MCS, the random numbers are uniformly distributed At first, according to the desired preciseness, the distribution
in the intervals, but in the LMCS, an N-point lattice rule of rank-r function centered on the zero mean (which represents the distri-
in d-dimension is computed as follows [24]: bution mean) is divided into some class intervals (here, seven seg-
! ments). Each class interval determines one standard deviation
Xr
kj XN
error (r) wide [27]. Moreover, each interval l is associated with a
v i mod1; kj ¼ 1; . . . ; N j ¼ 1; . . . ; r ð1Þ
probability denoted by bl,t. Subsequently, according to different
j¼1
nj i¼1
intervals and their probabilities obtained by the PDF, RWM [26]
where v1, . . . , vN are vectors with dimension d obtained by the ordin- are applied to generate scenarios for each hour. In this regard, at
ary MCS. Dimension d indicates the number of random values in first, the probabilities of different intervals are normalized in
each scenario. One LMCS scenario includes a vector with dimension which their summation becomes equal to unity. In addition, each
d of random numbers in the range of [0, 1] which is constructed by interval is associated with an accumulated normalized probability,
means of a set of values {kj, j = 1, . . . , r}. The difference among scat- as shown in Fig. 3. Therefore, each scenario comprises a vector of
tering of the scenarios related to MCS and LMCS in two-dimensional binary parameters identifying the load demand, market price,
space is shown in Fig. 1 which reveals the superiority of the LMCS powers output of PV and WT in each period:
rank-1. n o
All uncertainties are modeled based on their error determined Scenario s ¼ W L1;t;s ;...;W L7;t;s ;W WT WT PV PV Price Price
1;t;s ;...;W 7;t;s ;W 1;t;s ;...;W 7;t;s ;W 1;t;s ;...;W 7;t;s
t¼1;...;T
by the Probability Distribution Function (PDF). As an example, a ð6Þ
typical continuous PDF accompany with its discretization is shown
in Fig. 2 for the load forecast error in time period t [25]. The prob- where W Ll;t;s ; W WT
and uw;t;s ; W PV
pv ;t;s ,
are binary parameters indi- W Price
pri;t;s

ability distribution of a random variable is represented by a finite cating whether the lth load interval, the uwth wind power interval,
set of scenarios. This procedure is common in a stochastic pro- the pvth photovoltaic power output, and the prith market price
gramming. Also, each scenario corresponds with a single realiza- interval which are selected in scenario s (W Ll;t;s ; W WT PV
uw;t;s ; W pv ;t;s , and

tion of the random variables throughout the study horizon. In W pri;t;s W l;s ¼ 1) or not (W l;t;s ; W uw;t;s ; W pv ;t;s ; and W pri;t;s W Ll;s ¼ 0),
Price L L WT PV Price

addition, each scenario has an associated probability of occurrence. respectively. In other word, for each time interval and each random
Thus load demand, market price, powers output of PV and WT lev- variable, a random number is produced between 0 and 1 which fol-
els for each scenario can be presented as follows: lows the LMCS strategy. The first interval with an accumulated nor-
PWT;t;s ¼ Pforecast þ DP WT;t;s ; WT ¼ 1; . . . ; NWT ; t malized probability less than or equal to this random number is
WT;t
selected so that its associated binary parameter becomes equal to
¼ 1; . . . ; NT ; s ¼ 1; . . . ; Ns ð2Þ
1, the parameters of the non-selected intervals equal to 0, reversely.
This trend is continued until the desired number of scenarios is
PPV;t;s ¼ Pforecast þ DPPV;t;s ; PV ¼ 1; . . . ; NPV ; t ¼ 1; . . . ; NT ; s
PV;t generated. Finally, the normalized probability of each scenario is
¼ 1; . . . ; Ns ð3Þ calculated according to the following equation:
QT QND P7  P
7
 P
7
 P
7
 
t¼1 ld¼1 l¼1 W Ll;t;s bl;t uw¼1 W WT
uw;t;s buw;t
PV
pv ¼1 W pv ;t;s bpv ;t
Price
pri¼1 W pri;t;s bpri;t
PD;ld;t;s ¼ P forecast
D;ld;t þ DP D;ld;t;s ; ld ¼ 1; . . . ; ND; t ¼ 1; . . . ; NT ; s ps ¼ PNs QN QND P7  P
7
 P
7
 P
7
  s
T
s¼1 t¼1 ld¼1 l¼1 W Ll;t;s bl;t WT
uw¼1 W uw;t;s buw;t
PV
pv ¼1 W pv ;t;s bpv ;t
Price
pri¼1 W pri;t;s bpri;t

¼ 1; . . . ; Ns ð4Þ ¼ 1;...;N s
ð7Þ
forecast
Priceutility;t;s ¼ Priceutility;t þ DPriceutility;t;s ; t ¼ 1; . . . ; NT ; s where ND is the total number of load levels in each hour. bl,t, buw,t,
¼ 1; . . . ; Ns ð5Þ bpv,t, bpri,t are the probability of the lth load interval, uwth wind
power interval, pvth photovoltaic power interval and prith market
where PWT,t,s, PPV,t,s, PD,ld,t,s, Priceutility,t,s are the wind power output of price interval, respectively.
unit WT (kW), photovoltaic power output of unit PV (kW), ldth load
level (kW), and market price (€ct) at time t in scenario s, respec- 2.2. Scenario reduction
forecast
tively. P forecast forecast forecast
WT;t ; P PV;t PD;ld;t ; Priceutility;t are the forecasted wind
power of unit WT (kW), forecasted photovoltaic power of unit PV A higher number of generated scenarios results provide a better
(kW), the ldth forecasted load level (kW), and forecasted market modeling of uncertainties but with the cost of higher cumbersome
price (€ct) at time t, respectively. DPWT,t,s (kW), DPPV,t,s (kW), computation. Therefore, scenario-reduction techniques are essen-
DPD,ld,t,s (kW), and DPriceutility,t,s (€ct) are wind power forecasted tial to achieve computation tractability while keeping the stochas-
458 T. Niknam et al. / Applied Energy 99 (2012) 455–470

tic information embedded in the original scenario set as much as Priceutility;s ¼ ½Priceutility;1;s ; Priceutility;2;s ; . . . ; Priceutility;T;s 1T
possible. The simultaneous backward method is implemented to
trim down the number of deteriorated scenarios, as least as possi- Putility;s ¼ ½P utility;1;s ; Putility;2;s ; . . . ; Putility;T;s 1T
ble, this procedure depends on the accuracy of the approximation.
Consider ns(s = 1, . . . , Ns) as different Ns scenarios, each with a prob- PD;s ¼ ½PD;1;s ; PD;2;s ; . . . ; PD;T;s ð1ðNDTÞÞ ; PD;t;s
ability of ps and DT s;s0 as distance of scenario pair (s, s0 ). For using
¼ ½P D;1;t;s ; PD;2;t;s ; . . . ; PD;ND;t;s ð1NDÞ
the simultaneous backward method following steps should be ap-
plied [24]:
PWT;s ¼ ½PWT;1;s ; PWT;2;s ; . . . ; PWT;T;s ð1ðNWT TÞÞ ; PWT;t;s
Step1: Consider S as an initial set of scenarios, DS is the scenarios ¼ ½P1;t;s ; P2;t;s ; . . . ; PNWT ;t;s ð1NWT Þ
which should be deleted. The initial DS is null. Compute the dis-
tances of all scenario pairs: DT s;s0 ¼ DTðns ; ns0 Þ; s; s0 ¼ 1; . . . ; N s as PBATT;s ¼ ½PBATT;1;s ; PBATT;2;s ; . . . ; PBATT;T;s ð1ðNBATT TÞÞ ; PBATT;t;s
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pd ffi
s0 2
i¼1 ðv i  v i Þ .
s
DT s;s0 ¼ ¼ ½P1;t;s ; P2;t;s ; . . . ; PNBATT ;t;s ð1NBATT Þ
Step2: For each scenario k, DT k;r ¼ min DT k;s0 ; s0 ; k 2 S and s0 – k,
r is the scenario index which has the minimum distance with PFC;s ¼ ½PFC;1;s ; PFC;2;s ; . . . ; PFC;T;s ð1ðNFC TÞÞ ; PFC;t;s
scenario k.
Step3: Calculate PDk,r = pk  DTk,r, k e S. Select d in which ¼ ½P1;t;s ; P2;t;s ; . . . ; PNFC ;t;s ð1NFC Þ
PDd = min PDk, k e S.
Step4: S = S  {d}, DS = DS + {d}, pr = pr + pd. PPV;s ¼ ½PPV;1;s ; PPV;2;s ; . . . ; PPV;T;s ð1ðNPV TÞÞ ; PPV;t;s
Step5: Repeat steps 2–4 till the number is mitigated meet our
¼ ½P1;t;s ; P 2;t;s ; . . . ; PNPV ;t;s ð1NPV Þ
favorite request.

PMT;s ¼ ½PMT;1;s ; PMT;2;s ; . . . ; PMT;T;s ð1ðNMT TÞÞ ; PMT;t;s


2.3. Stopping rule
¼ ½P1;t;s ; P 2;t;s ; . . . ; PNMT ;t;s ð1NMT Þ
In this paper, a criterion is utilized for choosing a number of
Us ¼ ½UG;s ; US;s 1ððNG þNS ÞTÞ
samples in order to decide during the course of simulation,
whether or not the estimation seems to be accurate enough. This
method divides a simulation results into a number of batches, UG;s ¼ ½UWT;s ;UFC;s ;UPV;s ;UMT;s ;UBATT;s ;Uutility;s 1ððNWT þNFC þNPV þNMT þNBATT ÞTÞ
which each batch includes a specific number of scenarios. With
this manner it is not essential to check a stopping rule after each UWT;s ¼ ½UWT;1;s ; UWT;2;s ; . . . ; UWT;T;s ð1ðNWT TÞÞ ; UWT;t;s
scenario because the stopping rule is checked after each batch
and if the results are not acceptable, another batch should be ¼ ½U 1;t;s ; U 2;t;s ; . . . ; U NWT ;t;s ð1NWT Þ
run. This criterion is called coefficient of variation cvf and is defined
as follows: UPV;s ¼ ½UPV;1;s ; UPV;2;s ; . . . ; UPV;T;s ð1ðNPV TÞÞ ; UPV;t;s

rf ¼ ½U 1;t;s ; U 2;t;s ; . . . ; U NPV ;t;s ð1NPV Þ


cv f ¼ pffiffiffiffiffiffi ð8Þ
lf N s
UFC;s ¼ ½UFC;1;s ; UFC;2;s ; . . . ; UFC;T;s ð1ðNFC TÞÞ ; UFC;t;s
where rf and lf are the standard deviation and mean values of the
¼ ½U 1;t;s ; U 2;t;s ; . . . ; U NFC ;t;s ð1NFC Þ
output random variable f, respectively. If the value of cvf is less than
a pre-specific tolerance, then the outcomes can be considered as
acceptable results and the simulation can be terminated [28,29].
UMT;s ¼ ½UMT;1;s ; UMT;2;s ; . . . ; UMT;T;s ð1ðNMT TÞÞ ; UMT;t;s
¼ ½U 1;t;s ; U 2;t;s ; . . . ; U NMT ;t;s ð1NMT Þ
3. Stochastic multi-objective optimal micro-grid operation
(SMOMGO) UBATT;s ¼ ½UBATT;1;s ; UBATT;2;s ; . . . ; UBATT;T;s ð1ðNBATT TÞÞ

SMOMGO considering hybrid RESs is a nonlinear optimization UBATT;t;s ¼ ½U 1;t;s ; U 2;t;s ; . . . ; U NBATT ;t;s ð1NBATT Þ
problem with random continuous and discrete parameters and
variables. Penetration of load and generation uncertainties into Uutility;s ¼ ½U utility;1;s ; U utility;2;s ; . . . ; U utility;T;s ð1TÞÞ
the problem can affect the power outputs of all other deterministic
DGs and storage devices in the MG systems. In the following, the where X is the set of scenarios. NG, NS are the number of generating
objective functions and constraints are presented. Superscriptsˆ de- units and storage devices, respectively. Ps, Us, and fs are the vector
note the expectation of output random variables. of control variables, state variables and output random variables for
scenario s, respectively. PFC,s, PMT,s, PBATT,s, Putility,s are the FC, MT,
3.1. Decision variables Battery (BATT), and utility production vectors in scenario s, respec-
tively. PWT,s, PPV,s, PD,s, Priceutility,s are the WT production, PV produc-
tion, load demand, and market price vectors in scenario s,
correspondingly. PFC,t,s, PMT,t,s, PBATT,t,s, Putility,t,s are the FC, MT, BATT,
X ¼ ½X1 ; X2 ; . . . ; XNs 1Ns nT
and utility production vectors at time t in scenario s, correspond-
ingly. PWT,t,s, PPV,t,s, PD,t,s, Priceutility,t,s are the WT production, PV pro-
Xs ¼ ½Ps ; Us ; fs 1nT ; n ¼ 2NG þ 2NS þ ND þ 2
duction, load demand, and market price vectors at time t in scenario
s, respectively. UG,s, US,s are the status vectors for generation and
Ps ¼ ½PFC;s ; PMT;s ; PBATT;s ; Putility;s 1ððNFC þNMT þNBATT þ1ÞTÞ
storage units in scenario s, respectively. UWT,s, UFC,s, UPV,s, UMT,s,
UBATT,s, Uutility,s are the WT, FC, PV, MT, BATT, and utility status vec-
fs ¼ ½PWT;s ; PPV;s ; PD;s ; Priceutility;s 1ððNWT þNPV þNDþ1ÞTÞ tors in scenario s, correspondingly. UWT,t,s, UFC,t,s, UPV,t,s, UMT,t,s, and
T. Niknam et al. / Applied Energy 99 (2012) 455–470 459

Fig. 4. Flowchart of the proposed multi-objective ITLBO algorithm.


460 T. Niknam et al. / Applied Energy 99 (2012) 455–470

UBATT,t,s are the WT, FC, PV, MT, and BATT status vectors at time t in X
NG X
NS X
ND

scenario s, correspondingly. UWT,t,s, UFC,t,s, UPV,t,s, UMT,t,s, UBATT,t,s, and Pi;t;s U i;t;s þ Pj;t;s U j;t;s þ Putility;t;s U utility;t;s ¼ PD;ld;t;s ; t
i¼1 j¼1 ld¼1
Uutility,t,s are the states of the WT, FC, PV, MT, BATT, and utility at
time t in scenario s, respectively. ¼ 1; . . . ; T; s ¼ 1; . . . ; Ns ð14Þ
Active power limits of units [5]
3.2. Objective functions
Pmin
i;t 6 Pi;t;s 6 Pmax
i;t ; i ¼ 1; . . . ; NG; t ¼ 1; . . . ; T; s
Minimize [5]
¼ 1; . . . ; Ns ð15Þ
X
Ns X
Ns X
T
^f 1 ðXÞ ¼ ps f1;s ðXs Þ ¼ ps Costs;t
s¼1 s¼1 t¼1 Pmin max
j;t 6 P j;t;s 6 P j;t ; j ¼ 1; . . . ; NS; t ¼ 1; . . . ; T; s
(
X
T XNG
¼ 1; . . . ; Ns ð16Þ
Ns
¼ ums¼1 ps ½U i;t;s Pi;t;s Bi;t þ Start i maxð0; U i;t;s  U i;t1;s Þ
t¼1 i¼1

X
NS Pmin max
utility;t 6 P utility;t;s 6 P utility;t ; t ¼ 1; . . . ; T; s ¼ 1; . . . ; Ns ð17Þ
þShut i maxð0; U i;t1;s  U i;t;s Þ þ ½U j;t;s Pj;t;s Bj;t
j¼1 where P min min min
i;t ; P j;t ; and P utility;t are the minimum active powers of the
þStart j maxð0; U j;t;s  U j;t1;s Þ þ Shut j maxð0; U j;t1;s  U j;t;s Þ ith DG, the jth storage device and the utility at the time t, respec-

þU utility;t;s Putility;t;s Priceutility;t;s ð9Þ tively. Similarly, P max max max
i;t ; P j;t ; and P utility;t are the capacity powers of
corresponding units at time t.
where Bi,t and Bj,t are the bids of the DGs and storage units at time t, Spinning Reserve Requirements (SRRs) [5]
(Starti, Startj) or (Shuti, Shutj) represent the start-up or shut-down
costs for the ith DG and the jth storage unit, respectively. X
NG X
NS
U i;t;s Pmax
i;t þ U j;t;s Pmax max
j;t þ U utility;t;s P utility;t
Minimize [5] i¼1 j¼1

X
Ns X
Ns X
T X
ND
^f 2 ðXÞ ¼ ps f2;s ðXs Þ ¼ ps Emissions;t P PD;ld;t;s þ SRt;s ; t ¼ 1; . . . ; T; s ¼ 1; . . . ; Ns ð18Þ
s¼1 s¼1 t¼1 ld¼1
(
X
Ns X
T XNG X
NS
¼ ps ½U i;t;s Pi;t;s Ei;t  þ ½U j;t;s Pj;t;s Ej;t  where SRt,s is the SRRs at time t in scenario s.
s¼1 t¼1 i¼1 j¼1 Charge and discharge rate limit related to storage device [5]

þU utility;t;s Putility;t;s Eutility;t ð10Þ wj;t;s ¼ wj;t1;s þ gcharge;j Pcharge;j;t;s Dt
where Ei,t, Ej,t and Euility,t are described as the amount of pollutants 1
 Pdischarge;j;t;s Dt; j
emission in kg MWh1 for each generation unit, storage unit and gdischarge;j
utility at time t, respectively. These emission variables can be ex- ¼ 1; . . . ; NS; t ¼ 1; . . . ; T; s ¼ 1; . . . ; Ns ; Dt ¼ 1h ð19Þ
pressed as follows [5]:

Ei;t ¼ CO2;DG;i;t þ SO2:DG;i;t þ NOx;DG;i;t ð11Þ wmin


j 6 wj;t;s 6 wmax
j ; j ¼ 1; . . . ; NS; t ¼ 1; . . . ; T; s
¼ 1; . . . ; Ns ð20Þ
where CO2,DG,i,t, SO2,DG,i,t and NOxDG,i,t are the amounts of CO2, SO2
and NOx emissions related to the ith DG unit at time t,
respectively[5]. Pcharge;j;t;s 6 Pmax
charge;j ; j ¼ 1; . . . ; NS; t ¼ 1; . . . ; T; s
¼ 1; . . . ; Ns ð21Þ
Ej;t ¼ CO2;STOR;j;t þ SO2:STOR;j;t þ NOx;STOR;j;t ð12Þ

where CO2,STOR,j,t, SO2,STOR,j,t and NOx,STOR,j,t are the amounts of CO2, Pdischarge;j;t;s 6 Pmax
discharge;j ; j ¼ 1; . . . ; NS; t ¼ 1; . . . ; T; s
SO2 and NOx emissions related to the jth storage unit at time t,
¼ 1; . . . ; Ns ð22Þ
respectively [5].
where, wj,t,s and wj,t1,s are the amounts of energy storage inside the
Eutility;t ¼ CO2;utility;t þ SO2;utility;t þ NOx;utility;t ð13Þ
battery storage j at time t and t  1 in scenario s, respectively.
Pcharge,j,t,s(Pdischarge,j,t,s) is the jth battery permitted rate of charge (dis-
where CO2,utility,t, SO2,utility,t and NOx,utility,t are the amounts of CO2,
charge) during a definite period of time (Dt = 1 h) at time t in sce-
SO2 and NOx emissions related to utility at time t, respectively.
nario s. gcharge,j(gdischarge,j) is the efficiency of the jth battery during
The multi-objective SMOMGO problem consists of two compet-
the charge(discharge) process. wmin j and wmax
j are the lower and
ing objectives, i.e., expected total operation costs (9) and expected
upper limits on amount of energy storage inside the jth battery.
total operation emissions (10). The total operating cost of the MG
Pmax max
charge;j and P discharge;j are the maximum rate of the jth battery
in €ct (Euro cent) includes the fuel costs of DGs, start-up/shut-
charge(discharge) during each time interval Dt = 1h, respectively.
down costs and the costs of power exchange between the MG
and the utility [5]. The emission of pollutant gases (10) is modeled
in which three of the most important pollutants are involved: CO2 4. Solution methodology
(carbon dioxide), SO2 (sulfur dioxide) and NOx (nitrogen oxides)
[5]. In this section, the ITLBO algorithm is presented followed by the
multi-objective solution methodology including a fuzzy model for
3.3. Constraints [5] objective functions and Pareto optimization method. Thirdly, the
application of the proposed algorithm is presented. Finally, some
The constraints of the proposed problem are listed as follows: applicable tool usages of the SMOMGO problem are provided.
Power balance [5] The flowchart of the whole process is given in Fig. 4.
T. Niknam et al. / Applied Energy 99 (2012) 455–470 461

4.1. Improved teaching–learning-based optimization (ITLBO) xxiter


m;disc;jj;new1  xxmin
xyiter
m;disc;jj;new1 ¼ ; m ¼ 1; . . . ; Nlearner ; jj
xxmax  xxmin
TLBO is a new evolutionary algorithm which uses a population
¼ 1; . . . ; Ndisc ð32Þ
of learners to proceed to the global solution [22]. In this algorithm,
the population is considered as a class of learners. In TLBO, differ- where xxiter iter
m;disc;jj;new1 and xym;disc;jj;new1 are the pseudo-probability and
ent design variables are resembled different subjects offered to the actual probability for the jjth discrete variable of new solution
learners and the learners’ remarks are resembled the ‘fitness func- of the mth learner, respectively. xxmin and xxmax are the predeter-
tion’. The teacher is considered as the best solution obtained so far. mined as the minimum and maximum value of xxiter m;disc;jj;new1 ,
Learners can improve their knowledge through teacher teaching correspondingly. It is notable that the initial value of xx0m;disc;jj
and interaction between themselves. Therefore, the TLBO’s process parameters are equal to zero and xxmin and xxmax are equal to
is divided into two parts include teacher part and learners part. 50 and 50, respectively. Thus, the initial values for xy0m;disc;jj;new1
are 0.5 for m = 1, . . . , Nlearner and jj = 1, . . . , Ndisc.
4.1.1. Teacher part After that, a random number (rand) is generated and compared
A good teacher can bring up the mean of the class to some ex- with the probability xyiter iter
m;disc;jj;new1 . The value of xm;disc;jj;new1 will be
tend based on their talents and abilities [22]. To clarify this process,
determined as follows [30]:
the following parameters are defined: MEiter is a vector including
(
the mean of each decision variable for all learners and Titer is the 1 if rand 6 xyiter
iter m;disc;jj;new1 ;
teacher at the iterth iteration of the evolution procedure. Based xm;disc;jj;new1 ¼ m ¼ 1;...;N learner ; jj ¼ 1;...;N disc
0 else;
on the TLBO concept, Titer tries to move mean MEiter up to its
own grade. The structure of each learner is formed by putting ð33Þ
the discrete and continuous control variables beside each other The final control vector is generated by putting the discrete and
as follows: continues control variables beside each other as follows:
h i h i
Xiter iter iter
m ¼ Xm;disc ; Xm;cont
Xiter iter iter
m;new1 ¼ Xm;disc;new1 ;Xm;cont;new1
h i h i
¼ xiter iter iter iter
m;disc;1 ; . . . ; xm;dis;Ndisc ; xm;cont;1 ; . . . ; xm;cont;Ncont ð23Þ ¼ xiter iter iter iter
m;disc;1;new1 ;...;xm;dis;N disc ;new1 ;xm;cont;1;new1 ;...;xm;cont;N cont ;new1

ð34Þ
where Ndisc and Ncont are the number of discrete and continuous
variables, respectively. The structure of the teacher and the mean After generating all vectors, they are compared with the old one
value of the class are defined as follows: according to the min–max method as follows [31,32]:
8
h i >
> 1 ^f q ðXÞ 6 ^f min ;
ME iter
MEiter iter >
> q
¼ disc ; MEcont < ^max
f q ^f q ðXÞ ^f min < ^f q ðXÞ 6 ^f max ;
h i l^ fq ðXÞ ¼ ^f max ^f min i q q ¼ 1; 2 ð35Þ
¼ meiter iter iter iter >
disc;1 ; . . . ; medisc;Ndisc ; mecont;1 ; . . . ; mecont;Ncont ð24Þ > q q
>
>
:0 ^f q ðXÞ > ^f max ;
q

xiter
1;disc;jj þ xiter
2;disc;jj þ  þ xiter
Nlearner ;disc;jj
meiter
disc;jj ¼ ð25Þ ^f ¼ Maxððl
ref ;1  lf1 Þ; ðlref ;2  lf2 ÞÞ
Nlearner ^ ^ ð36Þ

xiter iter iter


1;cont;ii þ x2;cont;ii þ    þ xNlearner ;cont;ii
where ^f min
q and ^f max
q are the minimum and maximum acceptable ex-
meiter
cont;ii ¼ ð26Þ pected level of the qth objective function, respectively. l
^ fq is the ex-
Nlearner
pected membership value for the qth objective function.
Since the proposed problem has both discrete and continuous con- lref,qq = 1, 2 is the reference membership value for the qth objective
trol variables so the generations of population involves two steps. function. If the new learner has lower value of Eq. (36) then it sub-
The continuous solutions (learners) can be updated according to stitutes the old one. It should be noted that the final results of this
the following equation [22]: part are the inputs for the next part, i.e. learner part.

 
DMiter ¼ randðÞ Titer  T iter iter 4.1.2. Learner part
F M ð27Þ
In this part, interaction between learners can improve their
h i knowledge (situations). Each learner interacts with the other learn-
DMiter ¼ DMiter
disc ; DMcont
iter
ers randomly by the use of group discussions, presentations, formal
h i communications, etc. [22]. Thus, each learner can gain knowledge
iter iter iter iter
¼ dmdisc;1 ; . . . ; dmdisc;Ndisc ; dmcont;1 ; . . . ; dmcont;Ncont ð28Þ of something new if the other ones have more knowledge than
him/her. This procedure is described in the following form [22].
T iter ¼ roundð1 þ randÞ ð29Þ For the mth learner in the class, an individual (n) is selected ran-
F
domly in a way that m – n. Now a new solution is generated as fol-
where T iter
is the teaching factor in the iterth iteration. rand is a ran-
F
lows [22]:
dom function generator in the range [0, 1]. Now, the new solution      
can be produced according to the following equation [22]: Xiter iter iter iter
m;new2 ¼ Xm þ randðÞ Xm  Xn if ^f Xiter
m < ^f Xiter
n
  ð37Þ
Xiter iter iter
m;cont;new1 ¼ Xm;cont þ DMcont ; m ¼ 1; . . . ; Nlearner ð30Þ Xiter iter iter iter
m;new2 ¼ Xm þ randðÞ Xn  Xm otherwise

The discrete variables (learners) can be updated as follows [30]: After carrying out this part, Xiter iter iter
m is replaced by Xm;new2 if Xm;new2 has a

xxiter iter iter better function value according to Eq. (36). Note that like the ‘tea-
m;disc;jj;new1 ¼ xxm;disc;jj þ dmdisc;jj ; m ¼ 1; . . . ; Nlearner ; jj
cher part’, the final results of the learner part are the inputs for
¼ 1; . . . ; Ndisc ð31Þ the next part i.e. improved part.
462 T. Niknam et al. / Applied Energy 99 (2012) 455–470

4.1.3. Improved part logðNlearner  m þ 1Þ


wwm ¼ ; m ¼ 1; . . . ; Nlearner ð40Þ
In comparison with other evolutionary algorithms, TLBO has logð1Þ þ    þ logðNlernear Þ
many major advantages that have boosted its usage in several ap-
proaches such as economic dispatch applications [33]. These van- wwmm
acummethod ¼ acummethod þ ; mm ¼ 1; . . . ; Nmethod ð41Þ
tages are its easy implementation, simple concept, minimal Nmethod
storage requirements and no need for tuning the algorithm param-
where Nmethod is a number of learners which select methodth mod-
eters. It means this algorithm reaches the optimal solution without
ification method and wwmm (mm = 1, . . . , Nmethod) are the weight fac-
any user defined parameter and this feature is a major advantage
tors corresponding to them. The excitation probability is calculated
of this algorithm. Despite these features, TLBO may experience
as:
unsuitable convergence, so this paper devises the following modi-
fication to cope with this problem and avoid trapping in local op- acummethod
probmethod ¼ ð1  aÞ  probmethod þ a  ðmethod
tima. Since the proposed problem has continuous and discrete Iter max
control variables so like the generation of each vector, this section ¼ 1; 2Þ ð42Þ
has different rules for generating new vectors for continuous and
discrete variables. Concerning the continuous variables, two differ- where a is a learning rate to control the learning speed in the ITLBO
ent modified rules are applied which each learner according to a algorithm and it is considered equal to a = 0.142 in this paper. Final-
probability model chooses one of these methods. ly the RWM is applied to choose the methodth modification method
for each learner based on normalized probability values as follows:
4.1.4. Modification method 1 probmethod
This modification can be considered as globally correlated and probmethod ¼ ðmethod ¼ 1; 2Þ ð43Þ
prob1 þ prob2
therefore effectively enhances the TLBO’s global optimization
capabilities as follows [34]: After the above procedure the new solution is generated for each
  learner m as Xiter
m;cont;new3 .
Xiter iter
mm;cont;method1 ¼ Xmm;cont þ rand Titer iter
cont  Xmm;cont ; mm The proposed modification for all discrete variables has simple
mechanism in which for each learner the discrete value is updated
¼ 1; . . . ; N1 ð38Þ as follows [30]:
where Xiter
mm;cont;method1is the new solution which is generated from if rand < 0:08(
modification method1. Titer cont is the continues variables part of tea- 1 if xiter
m;disc;jj ¼ 0 ð44Þ
cher Titer. N1 is the number of learners which select the modification xiter
m;disc;jj;new3 ¼
method 1.
0 if xiter
m;disc;jj ¼ 1

After generating discrete and continuous modification vector, they


4.1.5. Modification method 2 are put beside each other for generating a control variable vector.
According to [34], the trigonometric modification operation of Eq. (36) is computed for each modification vector. ^f is considered
Eq. (39) biases the new trial solution heavily in the direction of as a criterion for substituting the modification vector with the rel-
the best one of three individuals chosen for modification process. evant vector which had been generated by the learner section.
It is worthwhile to note that it is a local search operator. In this
way, the proposed modification operator fulfills a wider search in 4.2. Multi-objective optimization problem (MOP)
the solution space and has significant effect in many generations
along the ITLBO algorithm. This method is formulated as follows 4.2.1. Pareto-based approach
[34]: Multi-objective optimization methods are specific tools to ex-
! plore the search space of optimization problem for optimal solu-
Xiter iter
r1 þ Xr2 þ Xr3
iter
tion/solutions in a reasonable time. The proposed Pareto method
Xiter
mm;cont;method2 ¼ þ ðR1
3 obtains several non-inferior solutions of the optimization problem
    in a single run and preserves the diversity of Pareto optimal solu-
 R2 Þ Xiter
r1  Xr2
iter
þ ðR3  R2 Þ Xiter iter
r2  Xr3 tions besides determines the best one. The Pareto method is based
  on the non-dominated solutions concept. In a minimization prob-
þ ðR1  R3 Þ Xiter
r3  Xr1
iter
; mm lem, a solution X1 dominates X2 if and only if the two following
¼ 1; . . . ; N2 ð39Þ conditions are satisfied simultaneously [23,33,35]:

   8q 2 f1; 2g : ^f q ðX1 Þ 6 ^f q ðX2 Þ


j^f ðXiter
r1 Þj j^f ðXiter Þj j^f ðXiter
r3 Þj   ð45Þ
where, R1 ¼ R0 , R2 ¼ Rr20 , R3 ¼ R0 , R ¼ ^f Xiter
0
r1 þ 9q 2 f1; 2g : ^f q ðX1 Þ < ^f q ðX 2 Þ
     
^ iter  ^ iter  iter
f Xr2  þ f Xr3 . Xmm;cont;method2 is the new solution which is gen- X1 is a non-dominated solution within the set {X1, X2}, if X1 domi-
erated from modification method2. N2 is the number of learners nates the solution X2. The solutions that are non-dominated within
which select the modification method 2. the entire search space are denoted as Pareto-optimal solutions.
In each iteration (iter) and for each solution mm, three vectors, These non-dominated solutions are stored in the repository during
r1, r2, r3, are selected from the existing population such that the algorithm search process.
r1 – r2 – r3 – mm.
The probabilistic solution strategy is defined as follows: at 4.2.2. Fuzzy-based clustering to reduce the repository size
first the probability of both modification methods are considered The POF for most of the problems is enormously large and may
probmethod = 0.5 (method = 1, 2) and a parameter called accumulator even include an infinite number of individuals. The large number
is allocated to each of them in which acummethod = 0 (method = 1, 2). of Pareto-optimal solutions leads to more computation time [23].
In each iteration, a weight factor is allocated to each learner Moreover, there are always memory constraints. Therefore,
after sorting the population according to Eq. (40). It is clear that decreasing the size of the repository without destroying the char-
the best learner gets the larger weight factor. After that the related acteristic of the POF is necessary. Fuzzy clustering method is ap-
accumulator of each method is brought up to date as Eq. (41): plied for fixing the size of repository, it is worthwhile to note
T. Niknam et al. / Applied Energy 99 (2012) 455–470 463

that the applied method dose not destroy the characteristics of the Step 2: Scenario generation: Generate 1000 scenarios and
POF [23]. In this regard, when the number of solutions in the repos- reduce them with the backward scenario reduction to reach
itory reaches its maximum limit, pruning all non-dominated solu- the desired batch including Ns scenarios. This batch should
tions (members of the repository and non-dominated solutions be satisfied Eq. (8). For each scenario, the hourly load
wishing to enter the repository) is done by the proposed fuzzy demand, market price, powers output of PV and WT units
clustering method. In order to generate a compromise value for are deterministic.
each member of the repository, the MGCC is asked to specify the Step 3: Represent and initialize the learners for the population:
desired weight factor for each of the objective functions which rep- A learner is a vector representing the generation levels and
resent the importance of each objective named as wq, q = 1, 2. Then statuses of each unit (i.e. MT, Phosphoric Acid Fuel Cell
the following normalized weighted membership approach can be (PAFC), Battery, and utility) at each hour of the day. At the
used to extract better solutions in the premiere region with itera- beginning, the learners of the population are generated ran-
tively search process which it is expected to find the closest solu- domly in the range of [0, 1] and located between the maxi-
tions with respect to the MGCC’s requirements [23]: mum and minimum operating limits of the unit. For each
P2 learner in the population, total operation costs f1,s
q¼1 wq lfq ðn1 Þ
^
Nl ðn1 Þ ¼ PNrep P2 ð46Þ s = 1, . . . , Ns and total operation emissions f2,s s = 1, . . . , Ns are
q¼1 wq lfq ðn1 Þ
n1 ¼1
^ calculated according to Eqs. (9) and (10), respectively. It
should be noted that all the constraints should be satisfied
In the proposed clustering approach, at first each non-dominated for each scenario contributed to each learner. A penalty term
solution builds a separate cluster. After that the solutions which is added to f1,s and f2,s for each violation of the constraints.
have small distance with this cluster adjoined to this cluster until Aggregated total operation costs and emissions for each lear-
the number of clusters becomes equal to the repository size. At this ner are calculated using Eqs. (9) and (10), respectively. After
time, an individual with higher normalized weighted membership that, compute Eq. (36).
value from each cluster is selected to be stored in the repository. Step 4: Constitute the repository: Save the non-dominated
solutions of the current population in the repository.
4.2.3. Niching technique for diversity preservation Step 5: Reduce the repository size: Use the fuzzy clustering
Niching technique [23], is utilized in this paper for avoiding the mechanism based on Section 4.2.2 to achieve the desired
learners accumulating towards densely populated area and conse- repository size.
quently obtaining a uniformly distributed POF. This technique Step 6: Update: Titer: The niching mechanism is applied on the
implements the fitness sharing which is based on the fundamental set of non-dominated solutions in the repository and the best
idea that the resources in a particular niche have to be shared be- compromise solution is selected as Titer.
tween the present individuals [23]. Therefore, the higher number Step 7: Incorporate teacher part as described in Section 4.1.1.
of the neighboring individuals leads to more inferiority of the fit- Step 8: Incorporate learner part: The learners would try to
ness. In the proposed approach, updating rule of the teacher part improve themselves by the use of knowledge interaction as
is changed by incorporating a niching-based procedure that guides described in Section 4.1.2.
the learners towards sparse regions. In the proposed method, the Step 9: Use SAPMS by incorporating modification part as
position of the teacher is selected from the repository by usage described in Section 4.1.3.
of the niching method. The Pseudocode of this process is shown Step 10: Update the repository: Check for non-domination.
at the end this section. In this regard, sharing distance dn1 n2 be- Update the repository.
tween repository member n1 and n2 should be computed. The Step 11: Check the convergence criteria: Go to step 5 for the
Shðdn1 n Þ which is the corner stone of the niching method computed next iteration. This loop can be terminated after a predeter-
as follow [23]: mined number of iterations and the teacher is selected as
8 d 2 the expected best compromise solution.
< n n
1 r1 2 if dn1 n2 6 rshare
Shðdn1 n2 Þ ¼ share ð47Þ
: 4.4. Tool usage
0 otherwise

where rshare is a niche radius which is set to 2 in this paper. After If the uncertainty of load demands, market prices, and available
computing the Sh(dij), the niche count M n1 can be computed and powers output of PV and WT units are neglected by the MGCC and
after that fitness of each individual which is inversely proportional the deterministic MOMGO is solved instead of the stochastic
to its niche count will be calculated. The global best solution in the MOMGO problem, the MGCC will almost certainly fail to meet
repository is selected by the RWM [36]. The niching approach can the challenges of supplying the load demands and satisfying the
be expressed as follows [23]: problem constraints. By the time that the dispatch results of MOM-
GO problem must be taken place at the beginning of the next day.
Therefore, the sources of the system uncertainties should be con-
For n1 ¼ 1 to N rep For n2 ¼ 1 to N rep Calculate
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  sidered in order to avoid inconsistency between the real conditions
P2 ^f q ðXn Þ^f q ðXn Þ 2
dn1 n2 ¼ q¼1 fq
1
^max ^min
f q
2
and Shðdn1 n2 Þ according to (47). and ante-scheduling patterns of MOMGO problem. Additionally, in
each time period, the forecast should be updated and a new
M n1 ¼ M n1 þ Shðdn1 n2 Þ EndFor (it refers to set n2) fit n1 ¼ M1n EndFor (it refers to
1
SMOMGO should be run while the powers output of all DG sources,
set n1)
Nickel-Metal-Hydride-Battery (NiMH-Battery) storage and pur-
Normalize the fitness function and use RWM to select the Titer. The chase power from utility in the previous hour are taken into ac-
initial values for Mn1 ; n1 ¼ 1; . . . ; N rep are zero. Nrep is the number of count. Consequently, to handle the system uncertainties as well
repository members. fit n1 ; Mn1 are the fitness and niche count of the as the constraints of the proposed framework, implementing the
n1th repository member, respectively. SMOMGO to cope with the stochastic nature of MOMGO problem
is a vital importance.
4.3. Application of the proposed algorithm In the conventional vertically integrated utilities, MGCCs are in
charge of load supply and generation scheduling, simultaneously.
Step 1: Input all required data for the system Therefore, to cope with the load and market price uncertainties be-
464 T. Niknam et al. / Applied Energy 99 (2012) 455–470

Fig. 5. Typical low-voltage MG model.

Table 1 Table 2
Results obtained by different methods for DOMGO1 (S1). Results obtained by different methods for DOMGO1 (S2).

Solution Total operation costs (€ct) CPU time Solution Total operation costs (€ct) CPU time
technique (min) technique (min)
Best Mean Worst Best Mean Worst
value value value value value value
GA [37] 277.7444 290.4321 304.5889 NA GA 298.1642 310.8519 325.0587 0.323
PSO [37] 277.3237 288.8761 303.3791 NA PSO 297.7435 309.2959 323.7989 0.296
FSAPSO [37] 276.7867 280.6844 291.7562 NA TLBO 291.6320 292.8431 293.5299 0.006
CPSO-T [37] 275.0455 277.4045 286.5409 NA TLBO-method1 288.8204 289.4275 290.0067 0.020
CPSO-L [37] 274.7438 276.3327 281.1187 NA TLBO-method2 288.3759 288.7199 289.0355 0.017
AMPSO-T [37] 274.4317 274.5643 274.7318 NA
ITLBO 287.4798 287.4798 287.4798 0.014
AMPSO-L [37] 274.5507 274.9821 275.0905 NA
TLBO 272.7610 273.1998 273.8411 0.005
TLBO-method1 271.1361 271.7890 272.0866 0.015
TLBO-method2 270.9328 271.2512 271.6319 0.013
ITLBO 269.7600 269.7600 269.7600 0.012 Table 3
Results obtained by different methods for DOMGO1 (S3).
NA: not available in the referred literature.
Solution Total operation costs (€ct) CPU time
technique (min)
Best Mean Worst
side PV and WT power generation variations, the proposed SMOM- value value value
GO framework is suitable for the DMs of the system. Also, in the GA 334.3679 336.3074 345.3569 1.294
deregulated power systems, both of MGCCs and upstream net- PSO 327.9226 331.0284 340.2094 1.040
works can use this framework. The MGCCs can use the proposed TLBO 313.8403 315.7245 316.9180 0.021
stochastic framework to handle their uncertainties and also, the TLBO-method 1 306.1463 307.4896 308.5608 0.065
TLBO-method 2 304.1127 305.5970 306.3305 0.061
upstream networks can implement the proposed framework for
bidding strategy to manage their uncertainties regarding load, ITLBO 302.2767 302.2783 302.2812 0.056

market price, PV and WT forecasted errors. Therefore, utilities


can adopt appropriate strategy to participate in power markets
while they consider the probable scenarios. such as MT, PAFC, PV, WT, and a NiMH-Battery as portrayed in
Fig. 5. The system data is extracted from [5], where a complete data
set can be found. The SRRs are set to 5% of the load demand in each
5. Simulation and numerical results hour.
It should be noted that a time period of 1 day with hourly time
In order to verify the effectiveness and superiority of the pro- step is considered as the study horizon. All the DG units produce
posed ITLBO algorithm for the SMOMGO problem, it is applied on active power at unity power factor. It means that these units re-
a typical low voltage MG. This system contains various DG sources quire or produce no reactive power. Also, the thermal load is not
T. Niknam et al. / Applied Energy 99 (2012) 455–470 465

Fig. 6. Generation levels of different units for DOMGO1 (S1) in kW. Fig. 7. Generation levels of different units for DOMGO1 (S2) in kW.

considered in the proposed MG system. Furthermore, there is a


power exchange link between the utility and the MG during the
time step in the study period based on the decisions made by the
MGCC. In this paper, it is assumed that the MGCC buys the
available powers of the PV and WT units at each hour. In order
to better exposition of our proposed framework, three different
strategies are considered. In the first Strategy (S1), all the DG units
are in service state during the time horizon and also the initial
charge of the NiMH-Battery is infinite. In the second Strategy
(S2), the NiMH-Battery starts with the 100 kW initial charge and
all the units are in service. In the third Strategy (S3), the NiMH-Bat-
tery starts with no charge at the beginning of the first hour of the
day and all units are allowed to start up or shut down for flexible
operation of the MG.
For comparison purposes, the single objective versions of
SMOMGO have been solved, named the SOMGO1, referred to cost,
and the SOMGO2, referred to emission. Similarly, the deterministic
version of SMOMGO, referred to as DMOMGO, has also been
solved. The simulations are carried out on a Pentium P4, Core 2
Duo 2.4 GHz personal computer with 1 GB RAM memory. The soft- Fig. 8. Generation levels of different units for DOMGO1 (S3) in kW.

ware is developed using MATLAB 7.8. The parameter setting of the


optimization algorithm is as follows: (i) the number of the popula-
tion is equal to 40. (ii) Maximum number of the iterations is 50. 650
It should be pointed out that all evolutionary methods require 600 Deterministic POF
tuning of different algorithm parameters for their proper search- Stochastic POF
ing. A small change in these parameters may result in a large 550

change in the algorithm performance. TLBO overcomes such diffi- 500


culties, because it does not require any parameter for tuning. It
Emission (Kg)

450
means this algorithm reaches the optimal solution without adjust-
ing any parameter and this feature is a significant advantage of this 400
Expected best compromise value
algorithm.
350
For a better illustration of the performance of the proposed
method in the stochastic optimization, six batches of scenarios 300
are considered. They include 6  30, 5  30, 4  30, 3  30,
250
2  30, and 30 samples of scenarios, respectively. As mentioned
before, a scenario is constructed by samples of the market price, 200
load demand, and available output powers of the PV and WT units
150
for all hours in the time period. Thus, each scenario has 82 parts. 200 400 600 800 1000 1200 1400 1600
Obviously, the computational effort of the scenario-based ap- Cost (Ect)
proach is directly proportional to the number of scenarios in each
Fig. 9. Deterministic and stochastic POF achieved by ITLBO algorithm for S1 (80
batch so the backward scenario reduction method would reduce
non-dominated solutions).
the total number of scenarios considering a tradeoff between
CPU time and solution accuracy. 5.1. Deterministic single objective and multi-objective solutions of the
It should be mentioned that due to alternative nature of the DMOMGO
evolutionary algorithm, 30 independent trial runs are carried out
to extract the statistical information such as best, worst, and mean This case is a base case which the input random variables are
value of the obtained solution. considered in the deterministic quantities. Firstly, the single
466 T. Niknam et al. / Applied Energy 99 (2012) 455–470

750

700 Deterministic POF


Stochastic POF

650

600
Emission (Kg)

550

500

450

400

350
200 400 600 800 1000 1200 1400 1600
Cost (Ect)
Fig. 12. Generation levels of different units for the best compromise solution (S1) in
Fig. 10. Deterministic and stochastic POF achieved by ITLBO algorithm for S2 (80
kW.
non-dominated solutions).

750

700 Deterministic POF


Stochastic POF

650
Emission (Kg)

600

550

500

450

400
200 400 600 800 1000 1200 1400
Cost (Ect)
Fig. 13. Generation levels of different units for the best compromise solution (S2) in
Fig. 11. Deterministic and stochastic achieved by ITLBO algorithm POF for S3 (80
kW.
non-dominated solutions).

objective versions of DMOMGO have been solved, namely DOM-


GO1, referred to cost, and DOMGO2, referred to emission. This pro-
cedure should be done to extract the extreme points of the tradeoff
surface and judge the diversity characteristics of the POF. The qual-
ity of the solutions found by ITLBO and TLBO has been assessed
through the comparison with the results reported in the literature
by available methods [37], which did not model SRRs. Since this
paper considers the uncertainty in load demands, the market
prices, and powers output of PV and WT units so, for reliability rea-
sons, reserve consideration is crucial. As can be inferred from
Tables 1–3, the superiority of ITLBO over the other approaches is
substantiated by the achievement of better values for cost objec-
tive function. It is obvious that the SAPMS is a useful tool which
not only accelerates the conversions of the original TLBO, but also
enables it to alleviate the stagnation and escape from local optima.
Above tables confirm that the proposed ITLBO provides high qual-
ity solutions and faster convergence in a confidence manner in
comparison with the other cited methods in the area. To show Fig. 14. Generation levels of different units for the best compromise solution (S2) in
the effect of the proposed modification methods 1 and 2, the re- kW.
sults of the optimization of each separate modification operator
in different strategy cases (S1, S2 and S3) have been added to Tables
1–3, correspondingly. It is necessary to note that the basic idea be- basis of their previous experiences in the generated promising
hind this approach is to simultaneously select adaptively multiple solutions and applied to perform the modification operation. Be-
effective operators from the candidate modification pool on the sides, the ITLBO with SAPMS can better manage transition from
T. Niknam et al. / Applied Energy 99 (2012) 455–470 467

Table 4
Comparison of the results obtained by different scenarios for SOMGO1 (S1).

No. of scenarios Total operation costs (€ct) Relative error (%) Coefficient of variation (%) CPU time (min)
Mean SD 0.95% confidence intervals Expected VSS
1  30 272.9092 5.9814 2.1404 274.6370 4.8769 0.7794 0.4002 0.352
2  30 270.2148 6.5252 1.6511 274.5769 4.8169 0.6013 0.3118 0.684
3  30 270.2008 6.4397 1.3305 274.5701 4.8101 0.4846 0.2512 1.014
4  30 270.2118 7.03878 1.2594 274.5676 4.8076 0.4587 0.2378 1.332
5  30 270.2092 7.5787 1.2128 274.5686 4.8085 0.4417 0.2290 1.695
6  30 270.3120 8.1009 1.1835 274.5686 4.8085 0.4310 0.2234 1.988

Table 5
Comparison of the results obtained by different scenarios for SOMGO1 (S2).

No. of scenarios Total operation costs (€ct) Relative error (%) Coefficient of variation (%) CPU time (min)
Mean SD 0.95% confidence intervals Expected VSS
1  30 291.5444 6.1512 2.2012 293.5038 6.0239 0.7499 0.3852 0.390
2  30 289.0636 6.5743 1.6635 293.3664 5.8866 0.5670 0.2936 0.785
3  30 288.9976 6.6506 1.3740 293.2640 5.7842 0.4685 0.2426 1.212
4  30 289.0253 7.1904 1.2865 293.1851 5.7052 0.4388 0.2271 1.583
5  30 289.0339 7.6747 1.2282 293.2628 5.7829 0.4188 0.2168 1.942
6  30 289.1054 8.2284 1.2021 293.2633 5.7834 0.4099 0.2121 2.396

Table 6
Comparison of the results obtained by different scenarios for SOMGO1 (S3).

No. of scenarios Total operation costs (€ct) Relative error (%) Coefficient of variation (%) CPU time (min)
Mean SD 0.95% confidence intervals Expected VSS Risk
1  30 306.5437 6.7752 2.4245 308.5888 6.3120 2.0451 0.7857 0.4035 1.330
2  30 304.4576 7.2178 1.8263 308.5241 6.2474 4.0665 0.5919 0.3061 2.891
3  30 303.9253 7.1983 1.4872 308.4972 6.2204 4.5719 0.4821 0.2496 4.840
4  30 303.9592 8.0202 1.4350 308.4946 6.2179 4.5354 0.4652 0.2409 6.422
5  30 304.0785 8.9822 1.4374 308.4983 6.2216 4.4198 0.4659 0.2412 8.122
6  30 304.2799 9.7806 1.4288 308.5006 6.2239 4.2207 0.4631 0.2396 9.790

each generation to the next one in comparison with each separate that the POFs achieved by the proposed method are well-distrib-
modification operator. The optimal generation levels of the differ- uted and have satisfactory diversity which covers the entire POF
ent units for the case of DOMGO1 are shown in Figs. 6–8 for S1, S2, effectively. The optimal generation levels of the different units
and S3, correspondingly. In the first two strategies (S1 and S2), all for the best compromise solution are shown in Figs. 12–14 for S1,
units are in service, hence the MGCC has to buy at least minimum S2, and S3, correspondingly.
power output of all units (because they are in service and they
have to generate electricity power) even though this procedure 5.2. Stochastic single objective and multi-objective solutions of the
burdens extra cost. Since the MT units are expensive so, they are SMOMGO
restricted to their minimum value during most hours in S1. But
the procedure is different in S2 and S3 scenarios, since the discharge From the point of MG operation, the stochastic approach by con-
action of the battery is restricted so, the MGCC should buy from the sideration of the intermittent nature of system components and
MT unit. It is clear that in all strategies economical choice such as loads can provide a more accurate solution for determining the
PAFC and utility are dispatched as much as possible to satisfy the allowance and optimum costs and emissions. The stochastic single
system load demand. Also, as shown in these figures, battery saves objective SOMGO1&2 are firstly solved by usage of the scenario
low-price energy of the utility at light-load hours and releases this generation scheme described in Section 2.1. 1000 24-h scenarios
energy during the peak-load hours which prepares good efficiency [38] are generated and subsequently trimmed to six batches. The
for the MGCC. The MGCC purchases power from the utility at the number of scenarios is usually selected in a way that the coefficient
first hours in order to fulfill load demand, meantime the extra of variation becomes small i.e. between 0.1% and 1% [39]. Tables 4–
power is saved in battery. Meanwhile, the discharge procedure of 6 show the coefficient of variation for the aforementioned batches
battery is started form hour 9 (peak load) to fulfill the load demand for S1, S2, and S3, respectively. As can be seen from these Tables,
as well as selling of the power to upstream network at high price. the total operation costs are decreased with increasing the number
This routine is altered in the 17th–19th hours. Also, the battery is of the scenarios. The indexes such as the value of expected, mean,
completely discharged in hour 20. In state of DOMGO2 which re- Standard Deviation (SD), 95% confidence intervals [24], relative er-
fers to the emission objective function, above debate can also be ror [24], coefficient of variation [28], and the Value of the Stochastic
concluded. Solution (VSS) [40] for each batches are also shown in these Tables.
After determining the extreme points, the ITLBO is applied to The values of all indexes except SD are decreased slightly when the
optimize operation costs and emissions simultaneously in order number of samples is increased. In order to quantify the relative im-
to achieve POF. Figs. 9–11 show POF achieved in one single run pact of the uncertainty sources on the decision making of the MGCC,
for S1, S2, and S3, respectively. These POF are extracted in 0.018, the VSS can be implemented. This metric measures how much the
0.023, and 0.068 min, correspondingly. These figures demonstrate MGCC would be willing to spend cost and emission to know the
468 T. Niknam et al. / Applied Energy 99 (2012) 455–470

11 12
10

9
10
8

7
8
6

5
4 6

3
2 4
1

0
580 585 590 595 600 605 610 615 2

Fig. 15. PDF of total operation costs for the expected best compromise solution of
S1. 0
675 680 685 690 695 700 705 710 715

Fig. 16. PDF of total operation costs for the expected best compromise solution of
future realizations of the under consideration stochastic problem. S2.
The VSS is the difference between the optimal solutions of the
single objectives DMOMGO and SMOMGO. The additional cost
and emission caused by considering of the uncertainties. In the 11
other word, modeling the system uncertainties increases the 10
operation cost and emission values as well as CPU time because
9
of considering different most probable scenarios in the stochastic
framework instead of single scenario in the deterministic scheme. 8
Also, it should be noted that lower the value of confidence interval
7
shows the precise degree of the expected value. When the interval
is smaller, we expect the real time solution to be closer to expected 6

value [24]. In addition, the relative error is computed as 5


ðð95% confidence intervalÞ=expected valueÞ  100%. The smaller
4
the relative error, the more efficient the proposed method is
[24]. 3
It should be noted that the main disadvantage of the batches 2
with enormous scenarios is the great CPU computational time
1
requirement. With the help of fast and efficient computing tools
i.e. scenario reduction technique and the ITLBO algorithm, a huge 0
690 695 700 705 710 715 720 725 730 735
stochastic model can be easily solved. From observing Tables 4–6,
it can be found that the batch which includes 30 scenarios is Fig. 17. PDF of total operation costs for the expected best compromise solution of
computationally significantly faster the other batches while its S3.
indexes are also near the another batches. The values of the
above tables show that the major difficulty with scenario reduc-
tion from 180 to 30 leads to a little increase in the total opera- 1X Ns

tion costs. Although using the 30 scenario increase 0.025%, RIq ¼ ðfq;s ð½Ps ; U; fs Þ  fq;s ð½Ps ; Us ; fs ÞÞ; q ¼ 1; 2 ð48Þ
Ns s¼1
0.082%, and 0.029% operational costs of SOMGO1 and the CPU
time is decreased 82%, 84%, and 86% for S1, S2, and S3, respec- It is worthwhile to note that the smaller this criterion gets, the re-
tively. The tradeoff between the operation costs and the CPU exe- sults of the proposed stochastic method are more close to the deter-
cution time reveals the superiority of the proposed backward ministic results.
scenario reduction. Thus, the 30 scenarios are considered for In the optimization process, each scenario considers different
the SMOMGO problem. In this paper, the stochastic approach values for load, WT and PV powers output and market price. Also,
tries to find expected POF solutions. The stochastic solutions each scenario has different decision variables but the same state
may not be global optima solutions to the individual scenarios variables. The expected operation costs and emissions of these 30
but they are a robust and also located near global solutions scenarios are the output random variables which should be opti-
which this provides possible realizations of the uncertainties. mized simultaneously in the algorithm procedure. This procedure
To this end, the Risk Index (RI) is calculated and added to Table is solved by the efficient ITLBO method and so the concluded sto-
6 for S3. The RI is the average value of the difference between the chastic POFs are extracted and shown in Figs. 9–11 for S1, S2, and
objective function value of the proposed approach and the deter- S3, respectively. The total CPU times for these stochastic problems
ministic approach which is specified as an index to compare the are about 0.512, 0.673, and 1.977 min for S1, S2, and S3, respec-
performance of the stochastic approaches. It is based on the dif- tively. The PDF of the operation costs of the expected best compro-
ference between the objective function value of deterministic mise solutions for S1, S2, and S3 through scenarios are presented in
solution of the proposed problem corresponding to each scenario Figs. 15 and 16, correspondingly. As can be seen from these figures,
fq,s([Ps, Us, fs]) and the objective function value of the proposed the output random variables e.g. operation costs tend to have the
problem of each scenario which is obtained by unit scheduling same PDF as the input random variables, which in this study is a
from applying the proposed approach fq,s([Ps, U, fs]) and is defined normal distribution function. Therefore, a normal distribution is
as follows [29]: fitted to the scenario results using its mean and SD. However, it
T. Niknam et al. / Applied Energy 99 (2012) 455–470 469

Table 7
Results of different combinations of weight factors for SMOMGO (all strategies).

Combination wq S1 S2 S3
0 w1 = 1.00 ^f ¼ 274:6370 €ct
1
^f ¼ 293:5038 €ct
1
^f ¼ 308:5888 €ct
1
w2 = 0.00 ^f ¼ 613:3658 kg
2
^f ¼ 706:9867 kg
2
^f ¼ 735:3178 kg
2

1 w1 = 0.80 ^f ¼ 361:2429 €ct


1
^f ¼ 388:4731 €ct
1
^f ¼ 407:2744 €ct
1
w2 = 0.20 ^f ¼ 432:9583 kg
2
^f ¼ 615:1857 kg
2
^f ¼ 652:6287 kg
2

2 w1 = 0.50 ^f ¼ 597:8253 €ct


1
^f ¼ 697:1258 €ct
1
^f ¼ 713:2213 €ct
1
w2 = 0.50 ^f ¼ 309:4458 kg
2
^f ¼ 507:8045 kg
2
^f ¼ 564:6447 kg
2

3 w1 = 0.20 ^f ¼ 983:0139 €ct


1
^f ¼ 1045:9978 €ct
1
^f ¼ 1072:3867 €ct
1
w2 = 0.80 ^f ¼ 242:4756 kg
2
^f ¼ 438:4151 kg
2
^f ¼ 490:0273 kg
2

4 w1 = 0.00 ^f ¼ 1472:4627 €ct


1
^f ¼ $1410:0035 €ct
1
^f ¼ 1319:9998 €ct
1
w2 = 1.00 ^f ¼ 192:1161 kg
2
^f ¼ 400:4604 kg
2
^f ¼ 440:1177 kg
2

is well depicted from these figures that the discrete and non- Table 8
smooth behavior of the SMOMGO problem skews the PDF of the Optimization statistics for all strategies.
output random variables. This behavior is difficult to be predicted Problem No. of variables No. of constraints
in advance, because it depends on the all of the input random vari- S1
ables uncertainties (see Fig. 17). DOMGO1&2 96 192
The main goal of extracting the non-dominated solutions in the SOMGO1&2 2880 5760
proposed SMOMGO problem during the search process is obtaining DMOMGO 96 192
SMOMGO 2880 5760
of the best compromise solution based on the need of the system
operator. The MGCC tends to supply load with the lower cost as S2
DOMGO1&2 96 216
well as lower emission; therefore the MGCC should implement
SOMGO1&2 2880 6480
the proposed approach under uncertain environment to obtain DMOMGO 96 216
the reliable solutions. The MGCC can select different combinations SMOMGO 2880 6480
P
of weight factors in such a way that 2q¼1 wq ¼ 1. The MGCC termi- S3
nates the combination steps if the desired solution is found; other- DOMGO1&2 192 216
wise, it selects a new wq. This procedure continues until the MGCC SOMGO1&2 2976 6480
is satisfied with the solution according to the system operating DMOMGO 192 216
SMOMGO 2976 6480
conditions based on the objective functions. To inform the compro-
mise solution values associated with different combinations of
weight factors, Table 7 is offered. In this table, w1 and w2 are
and constraints for the whole case studies and strategies are listed
weight factors corresponding to the operation costs and emissions,
in Table 8.
respectively. In addition, the decision makers can bias preference
The advantages of the proposed framework have been investi-
through the search process perfectly according to the min–max ap-
gated in the entire previous sections and concluded in Section 6.
proach which is described in (36) and presented in flowchart. This
Beside, one of the major drawbacks of the proposed ITLBO method
procedure is developed to decide on the best candidate solutions
in comparison with the classical optimization method such as lin-
for the next optimization part of the proposed algorithm. In this
ear programming [11] and mixed-integer linear programming [17],
work, the lref,qq = 1, 2 is set to one for both objective functions to
is its execution burden that increases with the number of the DG
assign the same importance to each of them.
and storage units which are used in the micro-grid network. How-
It can be inferred from the results that modeling of the system
ever, it is noteworthy that these methods lead to non-optimal solu-
uncertainties will increase the operation cost and emission values
tion with great economic and emission loss. Although it is usually
because stochastic procedure considers different most probable
motivating to increase the computational performance of the
scenarios instead of one scenario (as the deterministic scheme). In-
method, but since the operation cost and emission are very impor-
deed, approaching to the real conditions of the power system in the
tant in optimal micro-grid operation, a reasonable increase in com-
ante-scheduling studies will cost some expenses which are expect-
putation time would cause no severe problem. In order to
able. In other words, SMOMGO will concurrently consider the most
overcome the drawback of mathematical methods in solution of
probable scenarios. Besides, using the proposed stochastic frame-
optimal operation, the proposed ITLBO algorithm has been em-
work, all 30 accepted scenarios according to their probability val-
ployed to solve the problem with practical modeling of operation
ues contribute into the output random variable results, whereas
constraints. It should be pointed out that another disadvantage of
the deterministic method relies on only one scenario. The 30 ac-
the meta-heuristic evolutionary algorithms is their less robustness
cepted scenarios capture more of the uncertainty spectrum of the
in achieving optimal solution. The simulation results demonstrate
power system, which is approximately four times more than that
the robustness and consistent of the proposed ITLBO approach
of the deterministic framework. So, the MOMGO results of the sto-
chastic framework are more realistic than the deterministic frame-
work results. 6. Conclusion
From the result of the case studies and strategies, it can be de-
duced that the proposed SMOMGO have benefit from the both as- In this paper, an ITLBO method has been developed and applied
pects, i.e. presentation a flexible method to compromise the to solve the SMOMGO with clean sources such as WT, PV, and PAFC
conflicting objectives, total operation cost and emissions minimi- for economic and emission operation problem by considering
zation, with several constraints specially SRRs as well as more effi- uncertainties including WT and PV units powers output, load de-
cient handling of system uncertainties. The number of variables mand, and market price over the 24 h study horizon. Also for get-
470 T. Niknam et al. / Applied Energy 99 (2012) 455–470

ting closer to real condition as well as the reliability reasons, a re- [11] Chakraborty S, Weiss MD, Simoes MG. Distributed intelligent energy
management system for a single-phase high frequency AC microgrid. IEEE
serve constraint is taken into account. The best advantage of the
Trans Ind Electron 2007;54:97–109.
proposed algorithm is quick transfer of the information between [12] Sortomme E, El-Sharkawi MA. Optimal power flow for a system of microgrids
agents which this gives more ability to the proposed algorithm in with controllable loads and battery storage. IEEE/PES Power Syst Conf Expos
finding the global optima irrespective of the complexity of the 2009:1–5.
[13] Hernandez-Aramburo CA, Green TC, Mugniot N. Fuel consumption
problem. Besides, a SAPMS is complemented to the original TLBO minimization of a microgrid. IEEE Trans Ind Appl 2005;41:673–81.
to cope with the drawback of premature convergence. This modifi- [14] Mohamed FA, Koivo HN. System modelling and online optimal management of
cation includes two powerful knowledge interaction strategies. microgrid with battery storage. In: Proc 6th international conf renew energies
power, quality; 2007.
Each individual according to a probability model chooses one of [15] Mohamed FA, Koivo HN. System modelling and online optimal management of
these methods to improve its knowledge. Since some RESs such microgrid using mesh adaptive direct search. Int J Electr Power Energy Syst
as WT, PV have intermittent characteristic, approaches to analyze 2010;32:398–407.
[16] Logenthiran T, Srinivasan D. Short term generation scheduling of a microgrid.
MGs would be stochastic rather than deterministic. The proposed In: Tec IEEE region 10th conf; 2009. p. 1–6.
approach shows how the proposed formulation works in compar- [17] Morais H, Kàdàr P, Faria P, Vale ZA, Khodr HM. Optimal scheduling of a
ison with an unreal-case-based deterministic technique. To take renewable micro-grid in an isolated load area using mixed-integer linear
programming. Renew Energy 2010;35:151–6.
the uncertainties into account, a two-stage scenario-based method [18] Dukpa A, Dugga I, Venkatesh B, Chang L. Optimal participation and risk
according to the Monte Carlo technique is implemented. Using the mitigation of wind generators in an electricity market. IET Renew Power Gener
LMCS rank-1, possible scenarios of power system operating states 2010;4:165–75.
[19] Akbari R, Hedayatzadeh R, Ziarati K, Hassanizadeh B. A multi-objective
are generated and a probability is assigned to each scenario. For
artificial bee colony algorithm. Swarm Evol Comput 2012;2:39–52.
a tradeoff between computation time and accuracy, a backward [20] Akbari R, Ziarati K. Multiobjective bee swarm optimization. Int J Innov Comput
scenario reduction technique is utilized. I (IJICIC) 2012;8:715–26.
The main goal of the MOP is focusing on the minimization of the [21] Pablo Ruiz A, Russ Philbrick C, Peter Sauer W. Modeling approaches for
computational cost reduction in stochastic unit commitment formulations.
operation cost and emission level in MGs without sacrificing the IEEE Trans Power Syst 2010;25:588–9.
realization of the system. Since the proposed algorithm profits [22] Rao RV, Savsani VJ, Vakharia DP. Teaching–learning-based optimization: A
the niching and fuzzy clustering methods so it can obtain a true novel method for constrained mechanical design optimization problems.
Comput Aid Des 2011;43:303–15.
and well-distributed set of Pareto-optimal solutions which gives [23] Agrawal S, Panigrahi BK, Tiwari MK. Multiobjective particle swarm algorithm
the MGCC several chances to select an appropriate power dispatch with fuzzy clustering for electrical power dispatch. IEEE Trans Evol Comput
plan according to environmental or economical considerations. The 2008;12:529–41.
[24] Wu L, Shahidehpour M, Li T. Stochastic security-constrained unit commitment.
capability and performance of the proposed approach is tested on a IEEE Trans Power Syst 2007;22:800–11.
typical MG including various DGs under three strategies. The PDF [25] Billinton R, Allan RN. Reliability evaluation of power systems. 2nd ed. New
of the output random variables are achieved for each strategy. Also, York: Plenum; 1996.
[26] Michalewicz Z. Genetic algorithms + data structures = evolution programs. 3rd
a comparison between the deterministic and the stochastic ap- ed. New York: Springer; 1996.
proach in the term of the VSS and RI is carried out. As a result, this [27] Wu L, Shahidehpour M, Li T. Cost of reliability analysis based on stochastic unit
study provides a method for the MGCC with the blueprint for con- commitment. IEEE Trans Power Syst 2008;23:1364–74.
[28] Billinton R, Li W. Reliability assessment of electric power systems using Monte
trol and management of the challenging SMOMGO problem
Carlo methods. New York: Plenum; 1994.
emerged from proliferation of RESs in MG systems [29] Siahkali H, Vakilian M. Stochastic unit commitment of wind farms integrated
in power system. Electr Power Syst Res 2010;80:1006–17.
[30] LanLan Z, Ling W, Xiuting W, Ziyuan H. A novel PSO-inspired probability-based
References binary optimization algorithm. Int Symp Inform Science Eng 2008:248–51.
[31] Sakawa M, Yano H. An interactive fuzzy satisfying method using augmented
minimax problems and its application to environmental systems. IEEE Trans
[1] Llaria A, Curea O, Jiménez J, Camblong H. Survey on microgrids: unplanned Syst Man Cybern 1985;15:720–9.
islanding and related inverter control techniques. Renew Energy [32] Malekpour AR, Tabatabaei S, Niknam T. Probabilistic approach to multi-
2011;36:2052–61. objective volt/var control of distribution system considering hybrid fuel cell
[2] Hawkes AD, Leach MA. Modelling high level system design and unit and wind energy sources using improved shuffled frog leaping algorithm.
commitment for a microgrid. Appl Energy 2009;86:1253–65. Renew Energy 2012;39:228–40.
[3] Lasseter RH. MicroGrids. IEEE Power Eng Soc Winter Meet 2002;1:305–8. [33] Azizipanah-Abarghooee R, Niknam T, Roosta A, Malekpour AR, Zare M.
[4] Giannoulis ED, Haralambopoulos DA. Distributed Generation in an isolated Probabilistic multiobjective wind-thermal economic emission dispatch based
grid: methodology of case study for Lesvos–Greece. Appl Energy on point estimated method. Energy 2012;37:322–35.
2011;88:2530–40. [34] Wong KP, Dong ZY. Differential evolution, an alternative approach to
[5] Moghaddam AA, Seifi A, Niknam T. Multi-operation management of typical evolutionary algorithm. In: Proc 13th int conf intel syst appl power syst;
micro-grids using particle swarm optimization: a comparative study. Renew 2005. p. 73–83.
Sust Energy Rev 2012;16:1268–81. [35] Niknam T, Narimani MR, Jabbari M, Malekpour AR. A modified shuffle frog
[6] Ren H, Zhou W, Nakagami K, Gao W, Wu Q. Multi-objective optimization for leaping algorithm for multi-objective optimal power flow. Energy
the operation of distributed energy systems considering economic and 2011;36:6420–32.
environmental aspects. Appl Energy 2010;87:3642–51. [36] Goldberg DE. Genetic algorithms for search. Optimization, and machine
[7] Niknam T, Taheri SI, Aghaei J, Tabatabaei S, Nayeripour M. A modified honey learning, reading. MA: Addison-Wesley; 1989.
bee mating optimization algorithm for multiobjective placement of renewable [37] Moghaddam AA, Seifi A, Niknam T, Pahlavani MR. Multi-objective operation
energy resources. Appl Energy 2011;88:4817–30. management of a renewable MG (micro-grid) with back-up micro-turbine/fuel
[8] Soroudi A, Ehsan M, Zareipour H. A practical eco-environmental distribution cell/battery hybrid power source. Energy 2011;36:6490–507.
network planning model including fuel cells and non-renewable distributed [38] Morales JM, Minguez R, Conejo AJ. A methodology to generate statistically
energy resources. Renew Energy 2011;36:179–88. dependent wind speed scenarios. Appl Energy 2010;87:843–55.
[9] Hessami MA, Bowly DR. Economic feasibility and optimisation of an energy [39] Breipohl A, Lee FN, Huang J, Feng Q. Sample size reduction in stochastic
storage system for Portland Wind Farm (Victoria, Australia). Appl Energy production simulation. IEEE Trans Power Syst 1990;5:984–92.
2011;88:2755–63. [40] Morales JM. Impact on system economics and security of a high penetration of
[10] Kazempour SJ, Parsa Moghaddam M, Haghifam MR, Yousefi GR. Electric energy wind power. PhD dissertation, Dept Electr Eng, Univ Castilla-La Mancha,
storage systems in a market-based economy: comparison of emerging and Ciudad Real, Spain; 2010.
traditional technologies. Renew Energy 2009;34:2630–9.

You might also like