Hindawi

International Journal of Rotating Machinery

Volume 2017, Article ID 5142825, 9 pages
https://doi.org/10.1155/2017/5142825

Research Article
Optimal Sizing for Wind/PV/Battery System Using Fuzzy
𝑐-Means Clustering with Self-Adapted Cluster Number

Xin Liu,1 Hong-Kun Chen,1 Bing-Qing Huang,2 and Yu-Bo Tao3

1
School of Electrical Engineering, Wuhan University, Wuhan 430072, China
2
Shandong Electric Power Research Institute, Jinan 250001, China
3
State Grid Jiangsu Electric Power Maintenance Branch Company, Nanjing 210013, China

Correspondence should be addressed to Hong-Kun Chen; [email protected]

Received 24 April 2017; Revised 14 July 2017; Accepted 16 August 2017; Published 26 September 2017

Academic Editor: Yuanzheng Li

Copyright © 2017 Xin Liu et al. This is an open access article distributed under the Creative Commons Attribution License, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Integrating wind generation, photovoltaic power, and battery storage to form hybrid power systems has been recognized to be
promising in renewable energy development. However, considering the system complexity and uncertainty of renewable energies,
such as wind and solar types, it is difficult to obtain practical solutions for these systems. In this paper, optimal sizing for a
wind/PV/battery system is realized by trade-offs between technical and economic factors. Firstly, the fuzzy 𝑐-means clustering
algorithm was modified with self-adapted parameters to extract useful information from historical data. Furthermore, the Markov
model is combined to determine the chronological system states of natural resources and load. Finally, a power balance strategy is
introduced to guide the optimization process with the genetic algorithm to establish the optimal configuration with minimized cost
while guaranteeing reliability and environmental factors. A case of island hybrid power system is analyzed, and the simulation results
are compared with the general FCM method and chronological method to validate the effectiveness of the mentioned method.

1. Introduction chronological methods. The autoregressive moving average

(ARMA) is utilized to model the uncertainties of wind
Hybrid power systems (HPS) [1, 2], especially those depen- generation, photovoltaic (PV) power, and load in [3, 7, 8].
dent on renewable energy generations (REGs), such as solar However, methods for parameters estimation of ARMA are
photovoltaic (PV), together with wind turbine generations always somewhat cumbersome. Reference [9] put forward an
(WTGs), have been regarded as the most promising config- efficient approach for sizing optimization in a stand-alone
urations for remote areas power supply, since it is neither HPS with Hybrid Big Bang-Big Crunch algorithm. Reference
economical nor practical for delivering power over long [10] analyzed the different results obtained by four heuris-
distances. Although these clean energies provide significant tic algorithms; nevertheless the uncertainties of renewable
contributions and opportunities, the unpredictable nature [3] energies have not been considered detailedly. Reference [11]
of these resources has posed serious challenges to power suggested a method for technical and economic optimization
systems [4, 5]. In the context of remote HPS, the greatest in an isolated PV system; the solar radiation classification and
obstacle is to maintain power balance, because the adjustable power supply reliability calculation are performed hourly.
capacity depends merely on REGs and batteries [6]. Hence, However, only the cluster corresponding to the minimum
the dynamic characteristic of the wind speed and solar solar radiation is selected, which may not be suitable in the
irradiance, together with the power management of batteries, context of HPS. The other investigations are based on gross
should be investigated to obtain practical configurations for chronological data [12, 13] where the computation time is
HPS. always too unbearable. In [14], the traditional fuzzy 𝑐-means
In previous literature, various methods have been intro- (FCM) is adopted, which divides the data of wind speed, solar
duced for sizing optimization of HPS. The stochastic nature of radiation, and load evenly. Thus the inherent characteristics
the REGs has been investigated with several probabilistic and of the data are handled in a somewhat arbitrary way.
2 International Journal of Rotating Machinery

This proposed methodology will be complementary to 45

the previous studies and take a step further. First of all, time 40
series analysis [15] was used to describe the characteristic of

Power delivered to grid (kW)

35
hourly wind and solar and load data with FCM, the function
of which is to group the elements of data sets that have 30
analogous characteristics. Considering that FCM is sensitive
25
to the initialization number of clusters [16], a parameters
self-adaptive method is introduced to optimize the initial 20
state. Furthermore, the Markov model [17] is combined to 15
obtain the system scenarios of HPS. Then the correlation
and time dependency of data sets are maintained with the 10
time-dependent clusters of the renewable generations and 5
load power consumption. The optimal sizes for WTGs, PV,
0
DG, and batteries are determined with the genetic algorithm 0 2 4 6 8 10 12 14 16 18 20
(GA), in which a power balance strategy is designed to ensure Hub height wind speed (m/s)
that the capitalized and operational costs are minimized
and the reliability requirements, CO2 emission, and batteries Figure 1: Power curve of IEC 61400-12 standard.
constraints are preserved at the same time.
The remaining parts of the article can be demonstrated
𝑡
in the following manner. The models of the components in The overall wind power 𝑃𝑊 can be derived with (3), and
𝑡
HPS and the technique of FCM with self-adapted clustering the total wind energy 𝐸𝑊 can be obtained with (4):
number are introduced in Section 2. Section 3 presents the
objective function and constraints for optimal sizing method 𝑡 𝑡
𝑃𝑊 = 𝑁𝑊 × 𝑝𝑊 , (3)
in HPS. Section 4 utilizes a case of stand-alone hybrid system
located in Hainan, China, to verify the advantage of the 𝑡 𝑡
𝐸𝑊 = 𝑃𝑊 × Δ𝑡, (4)
proposed methodology, where the comparison between the
self-adapted FCM model and the traditional model with where 𝑁𝑊 is the number of wind turbines and Δ𝑡 is the time
chronological data is analyzed. In Section 5, conclusions are step.
summarized and the relationship between reliability and cost
is discussed.
2.1.2. PV System. For each PV panel, the output power can be
obtained [10] with
2. Models of the Components in HPS
𝐺
𝑝PV = 𝑝𝑝V𝑟 × ( ) × [1 + 𝛼𝑇 (𝑇𝑐 − 𝑇Ref )] , (5)
2.1. The Components in HPS 𝐺Ref
2.1.1. WTGs Generation System. The output power of each where 𝑝PV (𝑡) denotes the PV module power at time 𝑡. 𝑃𝑝V𝑟
WTG [10, 18] is obtained by (1), and the power curve of IEC means the rated power, 𝐺 means the solar irradiance, and
61400-12 standard is displayed in Figure 1. 𝑅Ref means the referenced solar irradiance, in 1000 W/m2 .
𝑇Ref means the referenced temperature on the surface of
{ 0, 0 ≤ V ≤ Vci , panels, which can be set to be equal to 25∘ C. 𝛼𝑇 means the
{
{
{
{ temperature-coefficients of PV panels, and it can be set to
{ 3
{𝑎V (𝑡) − 𝑏𝑝𝑟 , Vci ≤ V ≤ V𝑟 , be equal to −3.7 × 10−3 . The temperature 𝑇𝑐 of each cell is
𝑝𝑊 (𝑡) = { (1)
{
{ 𝑝𝑟 , Vci ≤ V ≤ Vco ,
deduced with
{
{
{
{ (𝑇stc − 20)
{0, Vco ≤ V. 𝑇𝑐 = 𝑇air + (( ) × 𝑅) , (6)
800
The parameters 𝑎 and 𝑏 are calculated by
where 𝑇air is the atmospheric temperature. 𝑅 means the
solar irradiance and 𝑇stc means the standard operation cell
𝑝𝑟
𝑎= , temperature and set to be 25∘ C.
(V𝑟3 − Vci3 ) 𝑡
The overall wind power 𝑃PV can be obtained with (7), and
(2) 𝑡
the total wind energy 𝐸PV can be calculated with (8):
Vci3
𝑏= ,
(V𝑟3 − Vci3 ) 𝑡
𝑃PV 𝑡
= 𝑁PV × 𝑝PV , (7)
𝑡 𝑡
where V denotes the rated wind speed, Vci and Vco are, 𝐸PV = 𝑃PV × Δ𝑡, (8)
respectively, the cut-in wind speed and cut-off wind speed.
𝑝𝑟 means the rated power of WTGs. where the number of PV panels is given with 𝑁PV .
International Journal of Rotating Machinery 3

2.1.3. Battery Bank and the Power Balance Strategy. To since the solution is locally minimized of the object function,
accommodate the stochastic behavior of PV and wind which validates the effectiveness of Step 4 in Figure 3.
resources, battery banks are widely utilized for hybrid power In [14], the cluster number for WTGs is selected by evenly
systems. The power balance strategy is mainly based on dividing the range between the cut-in and cut-off wind speed.
the flexibility of batteries and diesel generations. The diesel The selection of cluster number of PV and load is the same.
generator in [14] is the only adjustable power generation However, the inner uncertain nature of the wind resource
without consideration of storage devices. Thus the power may be disregarded by this means, and the accuracy of this
balance strategy is limited to only one pattern, namely, the method may be reduced significantly.
diesel generators run to make up for the power shortage of In this paper, the cluster numbers for WTGs, PV, and
renewable energies. To maximize the utilization of REG and load are obtained via the method proposed in Section 3.1.
minimization of diesel generation, a power balance strategy More specifically, the wind speed V, solar irradiance, and
is illustrated in Figure 2. load power can be divided into 𝐶WT , 𝐶PV , and 𝐶LD clusters
coherently. The cluster centers 𝑃𝑊𝑐 , 𝑃PV𝑐 , and 𝑃LD𝑐 are the
2.2. FCM with Self-Adaptive Cluster Number. In this section, representative in this cluster, namely, the representative state
FCM clustering is modified to identify the operation state of wind speed, solar radiation, and load.
of HPS. The calculation complexity can be significantly
optimized considering the number of states will be much less 2.2.3. The Markov Stochastic Process. In the analysis of a
than the 8760 h in the chronological methods. Traditional stochastic process, the Markov chain is an effective method
FCM clustering algorithm can only deal with a prescribed to relate the probability of a state with the frequency of
data set with clustering number given in advance, which is the corresponding event. The operation states are 𝑃WT (𝑐WT ),
not flexible in the context of large data sets. A new validity 𝑃PV (𝑐PV ), and 𝑃LD (𝑐LD ), where 𝑐WT , 𝑐PV , and 𝑐LD are the state
function [19] is introduced to construct the proportion of indices obtained by the proposed method in Section 3.1. Take
compactness and divergence; thus the cluster number can be a wind farm with four Markov states as an example, shown in
obtained according to the given data set. Figure 4. The state transfer probability and failure rate among
different states are given with 𝜆 and 𝜇, respectively.
2.2.1. FCM Clustering for HPS. The FCM clustering algo-
If 𝑝𝑐 , 𝑐 = 1, 2, . . . , 𝐶, are probabilistic for 𝐶 states (𝐶 is the
rithm established by Dunn and then further improved by
cluster number), then they should satisfy
Bezdek has been used extensively.
The given data set is divided into 𝑐 clusters relating to 𝐶
some given criterions to optimize an objective function. The ∑𝑝𝑐 = 1,
problem can be formulated as 𝑐=1 (11)
𝑐 𝑛 𝑡
[𝑇] [𝑝1 , . . . , 𝑝𝐶] = 0.
min 𝐽𝑚 (𝑈, 𝑉) = ∑ ∑ 𝑢𝑖𝑗𝑚 𝑑𝑖𝑗2 ,
𝑖=1 𝑗=1 The diagonal elements 𝑇𝑐,𝑐 are equal to the negative-
(9) sum of each off-diagonal element at column 𝑐 of transition
𝑐
∑𝑢𝑖𝑗 = 1, 1 ≤ 𝑗 ≤ 𝑛. probability matrix 𝜑, and the elements in other positions
𝑖=1 correspond, respectively, to 𝜑.
An effective partition of a given data set should be 𝜑𝑖,𝑗 = 𝑃 (𝑋𝑡+1 = 𝑗 | 𝑋𝑡 = 𝑖) ,
divergent and compacted. The degree of compactness and
divergence are evaluated with a clustering validity function. ∑ 𝜑 = 1,
Hence, a new validity function is adopted in
𝑛𝑖,𝑗 (12)
󵄩󵄩 󵄩2 𝜑𝑖,𝑗 = ,
∑𝑐𝑖=1 ∑𝑛𝑗=1 󵄩󵄩V𝑖 − 𝑥󵄩󵄩󵄩 / (𝑐 − 1)
𝑢𝑖𝑗𝑚 𝑁𝑖
𝐿 (𝑐) = 󵄩 󵄩2 , (10)
∑𝑐𝑖=1 ∑𝑛𝑗=1 𝑢𝑖𝑗𝑚 󵄩󵄩󵄩󵄩𝑥𝑗 − V𝑖 󵄩󵄩󵄩󵄩 / (𝑛 − 𝑐) 𝑖 ≠ 𝑗,
where 𝑚 is the fuzzy weighting index and greater than 1 and where 𝑛𝑖,𝑗 denotes the transition number from state 𝑖 to 𝑗, and
𝑐 means the clustering number. 𝑛 is the subscript index of 𝑁𝑖 means the number of states 𝑖. Rate of departure (RD) is
the data set 𝑋, in which 𝑥𝑗 (𝑗 = 1, 2, . . . , 𝑛) is each pattern. the modulus of the diagonal elements; the frequency (𝐹𝑖 ) and
𝑥𝑗 (𝑗 = 1, 2, . . . , 𝑛) is the clustering prototype. ‖ ⋅ ‖ means the duration for state 𝑖 (𝐷𝑖 ) can be formulated as
statistical distance of the very data set. The central vector of
the overall data is given with 𝑥 = (1/𝑛) ∑𝑐𝑖=1 ∑𝑛𝑗=1 𝑢𝑖𝑗𝑚 𝑥𝑗 . 𝐹𝑖 = 𝑝𝑖 ∑ 𝜑𝑖,𝑗 , 𝑖 = 1, 2, . . . , 𝑐,
𝑗=1,𝑗=𝑖̸
(13)
2.2.2. Clustering Procedure. Then, the FCM algorithm with 1
self-adapted clustering number is outlined with the proposed 𝐷𝑖 = , 𝑖 = 1, 2, . . . , 𝑐.
∑𝑗=1,𝑗=𝑖̸ 𝜑𝑖𝑗
validity function 𝐿(𝑐). The clustering procedure is illustrated
in Figure 3. The system states of HPS are mainly determined by the
The partition matrix 𝑈(0) is set as an initial condition; combination of WT and PV states, which also determine the
merely two local values of 𝐿(𝑐) are required to be compared state of DG and batteries.
4 International Journal of Rotating Machinery

No
t
Start E＂；Ｎ,ＧＣＨ ⩽ E＂；Ｎ ⩽ E＂；Ｎ,Ｇax

Data input Yes

Batteries in discharge state
t t−1
PW + P０６ ≥ P，＄ E＂；Ｎ = E＂；Ｎ × (1 − ) −
t t t
(E，＄ − EPV − EW ) × ＂；Ｎ
No
Yes

＃／2 ≤ ＃／2Ｇ；Ｒ
t
E＂；Ｎ,ＧＣＨ ⩽ E＂；Ｎ ⩽ E＂；Ｎ,Ｇax ，０３０ ≤ ，０３０Ｇ；Ｒ

Yes Yes
Batteries in charge state
Diesel generation operates
t t−1
E＂；Ｎ = E＂；Ｎ × (1 − ) + EtDG = E，＄
t t
− E０６ − EWt
− No
No t
(E０６ t
+ EW t
− E，＄ ) × ＂；Ｎ t−1
E＂；Ｎ × (1 − ) + E＂；Ｎ,ＧＣＨ

LPSP and ＃／2 calculation

Wind and solar curtailment
t t t t
E＝ＯＬＮ = E０６ + EW − ELD −
t−1
E＂；Ｎ,Ｇ；Ｒ + E＂；Ｎ × (1 − )/＂；Ｎ End

Figure 2: Power balance strategy of HPS.

Start

Initialization
c > 2,  > 0, k = 0, V(0) , L (1)

Partition matrix

uij(k) = 1/ ∑cr=1 (dij(k) /drj

(k) 2/(m−1)
)

Prototypes calculation
m m
i(k+1) = ∑nj=1 (uij(k) ) xj / ∑nj=1 (uij(k) )

 (k+1) 
V
 − V(k)  < 
No
Yes
k= k+1

L(c − 2) No
< L(c − 1)
L(c)
Yes

Stop

Figure 3: Self-adapted clustering procedure.

International Journal of Rotating Machinery 5

3.2.1. Power Balance in the Given Time Resolution. The

foundation of the sizing optimization problem is the power
P2
12 balance. The power balance strategy in this paper is illustrated
32 in Figure 2 (power balance strategy of HPS). The power
balance equation is

21 23 𝑁𝑥.𝑖.𝑡

31 ∑ 𝑃𝑖.𝑡.𝑗 = 𝑃𝐿.𝑡 , (19)
𝑗=1
P1 P3
13
43
where 𝑁𝑥.𝑖.𝑡 means output power of each kind of DG in the
41
given time interval 𝑡, the length of which is associated with
14
the length of the planning period and is set to be 1 h in this
34
paper.
P4

3.2.2. The Minimum and Maximum Scale of the DGs

Figure 4: Four-state Markov model. 𝑁𝑖 min ≤ 𝑁𝑖 ≤ 𝑁𝑖 max , (20)

where the subscripts max and min mean the maximum and
3. Sizing Optimization minimum restraints for DG scales, respectively.
3.1. Objective Functions. In the proposed optimal sizing
methodology of HPS, the main goal is to determine the 3.2.3. Battery Constraints. At time 𝑡, the value of charge level
amounts of each kind of DG. For a given system load, the for each battery bank should satisfy
objective can be set for the overall cost optimization.
𝐸Bat,min ⩽ 𝐸Bat (𝑡) ⩽ 𝐸Bat,max . (21)
Min 𝐶𝐹 (𝑥) = min ∑ 𝑁𝑖 𝐶𝑖,𝑗 , (14)
where subscript 𝑖 denotes the type of generations, namely, The battery banks capacity (𝑆Bat ) sets the maximum
WT, PV, DG, and batteries, and 𝑗 contains the cost of the unit, value of charge level (𝐸Bat,max ) and the depth of discharge
installation, and fuel consumption of the HPS. (DOD) determines the minimum value of charge level for the
The fuel consumption of the DGs using fossil fuel is given batteries (𝐸Bat,min ).
with
𝐸Bat,min = (1 − DOD) × 𝑆Bat . (22)
𝑓 (𝑃𝑑 ) = 𝑎 + 𝑏 ⋅ 𝑃𝑑 + 𝑐 ⋅ 𝑃𝑑 2 . (15)
Then the cost model of a diesel generator is 3.2.4. Reliability Index. LPSP (loss of power supply probabil-
ity) is selected for its simplicity, which can be derived from
𝐶𝑑 = (𝑐1 ⋅ 𝑃𝑑𝑟 + 𝑐2 ⋅ 𝑃𝑑𝑟 + 𝑓 (𝑃𝑑 ) ⋅ 𝑐3 ) ⋅ 𝑁𝑑 . (16) the division of total time of power unbalance and the studied
For the DGs using renewable energy resources like WG period with
and PV, the operation cost can be ignored. For the DGs using
fossil fuel, the operation cost should be accumulated in the ∑TP
𝑡=1 (𝐸Load (𝑡) − 𝐸Gen (𝑡))
LPSP = . (23)
studied period. The combustion of fossil fuel will contribute ∑𝑇𝑡=1 𝐸Load (𝑡)
to the emission of CO2 and gaseous pollutants. The ramping
characteristics of diesel generation are neglected in the article The subscript “max” signifies the maximum limits.
since the time resolution is set to be one hour.
In (17), the fuel consumption cost 𝐶𝑑 can be obtained. 3.2.5. Environmental Factors. The CO2 emission of the DGs
8760
using fossil fuel can be obtained with (24), when considering
fuel
𝐶𝑑sum = ∑ 𝐶𝑑fuel (𝑃𝑑 (ℎ)) . (17) it in the total period, and should be constrained by a
ℎ=1 maximum CO2 max .
By means of FCM clustering, it can be reduced to CO2 = 𝑑 + 𝑒 ⋅ 𝑃𝑑 + 𝑓 ⋅ 𝑃𝑑2 (24)
fuel 󸀠 󸀠 󸀠
𝐶𝑑sum = 𝐶𝑑fuel (𝑃𝑑 (𝑐 )) × 𝐹 (𝑐 ) × 𝐷 (𝑐 ) × 8760, (18) CO2 ≤ CO2 max
. (25)
where 𝑃𝑑 can be obtained with the proposed power balance In this article, 𝑑, 𝑒, and 𝑓 of a 30-kW diesel generator are
strategy. set to be 0.028144, 0.001728, and 0.0000017.

3.2. Constraints. On the basis of normal operation in the 3.3. Using Genetic Algorithm to Get Optimal Solution. The
stand-alone HPS, in order to associate the reliability factor genetic algorithm (GA) is chosen to solve the sizing problem
and environmental factors, the main constraints of the considering its ability to obtain a globally optimal solution
proposed methodology are as follows. for optimization problems. It is inspired by the process
6 International Journal of Rotating Machinery

25
40

20 35

Temperature (∘ C)
Wind speed (m/s)

30
15

25
10
20

5 15

0 10
0 1 2 3 4 5 6 7 8 9 10 11 12 0 1 2 3 4 5 6 7 8 9 10 11 12
Month Month

Figure 5: Wind speed. Figure 7: Temperature.

1.2 350

300
1

250
Illumination (kW/Ｇ2 )

0.8
Load power (kW)

200
0.6
150
0.4
100

0.2 50

0 0
0 1 2 3 4 5 6 7 8 9 10 11 12 0 1 2 3 4 5 6 7 8 9 10 11 12
Month Month

Figure 6: Illumination rate. Figure 8: Load profile.

of biological evolution, namely, crossover, mutation, and of WTGs (𝑃𝑊) and PV (𝑃PV ). Then the cluster numbers for
selection. The population of individual solutions is repeatedly WT, PV, and load are obtained via the method proposed in
modified with a “fitness” function, typically related to the Section 2.1. The cluster centers Vc, Gc, and Lc are, respectively,
objective function. the representative states for wind speed, solar irradiance, and
In this article, the optimization problem (14)∼(25) is load, illustrated in Tables 1–3.
handled with GA, where the variables 𝑁WT , 𝑁PV , 𝑁DG , and According to the simulation results from self-adaptive
𝑁BAT are linked to form the gene strings in the state variable FCM, the optimal cluster numbers for wind WT, PV, and LD
(chromosome), and (14) is set to be the fitness function. are 10, 5, and 8. Namely, each WT has 10 states, each PV has 5
states, and the overall load has 8 states. There are 80 possible
4. Results states for the renewable energy generations as a whole. It can
be noted here that the scenario of HPS has been significantly
4.1. Case Introduction. The data from an island in Hainan simplified.
province of China are used to analyze the proposed problem. The operation scenario considered here is greatly simpli-
There are abundant wind and solar resources in this island. fied to be the aggregation of clustering states. The outputs
of DGs and batteries are also determined by these states,
4.2. Simulation Results and Analysis. The monthly average according to the power balance strategy demonstrated before.
wind speed, solar irradiation, temperature, and load power For a new state, the probability is the multiplication of
profile consumption are shown in Figures 5–8. state probabilities for every individual WTG, PV, and LD. The
Firstly, the wind speed V, illumination 𝐺, and temperature frequency 𝐹 and duration time 𝐷 for a new scenario can be
𝑇 are imported to the HPS model to obtain the power output obtained likewise.
International Journal of Rotating Machinery 7

Table 1: Markov model of WT/35 kW.

3.2
State (𝑐1) Pw/kW ps RD 𝐹/(oc⋅h−1 ) 𝐷 (h)
1 0.20 0.08 0.47 0.03 2.15 3
2 3.90 0.09 0.65 0.05 1.53
3 7.48 0.15 0.52 0.08 1.92 2.8

Cost (M$)
4 10.99 0.08 0.82 0.07 1.22
5 14.56 0.08 0.59 0.05 1.69 2.6
6 18.08 0.09 0.62 0.12 1.61
7 21.86 0.09 0.85 0.08 1.18 2.4
8 25.89 0.06 0.72 0.06 1.39
9 30.19 0.09 0.84 0.08 1.20 2.2
10 34.87 0.19 0.76 0.07 1.32
2
0 5 10 15 20 25 30 35 40 45
Table 2: Markov model of PV/200 W. Iteration performance
State (𝑐1) 𝑃PV /W ps RD 𝐹/(oc⋅h−1 ) 𝐷 (h) Self-adapted FCM
1 1.37 0.61 0.37 0.22 2.71 Basic FCM
2 83.14 0.10 0.62 0.06 1.63 Chronology data
3 37.23 0.13 0.44 0.06 2.29
Figure 9: Iteration performance comparison.
4 134.05 0.09 0.59 0.05 1.69
5 188.58 0.07 0.26 0.02 3.82
and the computation burden and CPU time can be
Table 3: Markov model of load. significantly reduced, shown in Table 4.
State (𝑐1) 𝑃PV /W ps RD 𝐹/(oc⋅h−1 ) 𝐷 (h) (2) Compared to the traditional FCM based method,
1 135.66 0.16 0.67 0.11 1.49 the clustering numbers of data sets are inherently
2 83.78 0.18 0.56 0.10 1.77 obtained and optimized, which increases the prob-
3 156.19 0.17 0.74 0.13 1.34
ability to obtain a global optimum solution. While
the traditional method simply makes the partition by
4 217.17 0.07 0.82 0.05 1.23
uniform division, the scenarios selection is somewhat
5 182.49 0.10 0.63 0.07 1.60 arbitrary and disregards the inner stochastic charac-
6 63.81 0.17 0.22 0.04 4.46 teristic of the data sets. Thus the basic FCM based
7 111.11 0.13 0.51 0.06 1.96 method finally obtains a locally optimized solution.
8 268.08 0.03 0.28 0.01 3.61
In the proposed method, the benefits of the renewable
energies are thought to be the reduction of CO2 and the
Then the results from chronology-based, traditional FCM improvement of LOSP, which has been set as constraint to the
based, and the self-adapted clustering number based GAs are very problem. Considering the impacts of different reliability
compared; results are illustrated in Table 4. GA is capable index on the investment cost, let CO2 max be 30000 kg/y. The
of realizing global optimization but cannot guarantee it. The overall investment cost grows higher as the reliability request
chronological-based method requires 8760 iterative loops, (LPSPmax ) increases, as shown in Figure 10. The impact of
and the traditional FCM based method needs 920 (=23 ∗ 4 ∗ CO2 is similar to the LPSP index.
10) iterative loops, and the proposed method needs 400 (=10 Actually, the benefits of the reduction of CO2 and the
∗ 5 ∗ 8) iterative loops. improvement of LOSP are negatively related to the cost
It should be noted that the wind energy is superfluous optimization procedure. We have modified the discussion on
in the winter nights. The output power of PV panels is this issue in our revised manuscript.
redundant in the summer daytimes.
It can be found that chronology-based method is the 5. Conclusion
most time-consuming due to complicate loops. With regard
to the overall cost, the proposed method still has advantages. In this paper, a novel method utilizing the self-adapted
Figure 9 shows the iteration performance for the mentioned FCM clustering combined with the Markov model and GA
algorithms. is proposed to determine the best mix of HPS. A power
The proposed self-adapted FCM clustering model is balance strategy is also designed to guide the optimization
superior to traditional method in two aspects: process. The self-adapted FCM clustering can handle the
stochastic characteristics of REGs, and the Markov model can
(1) Compared to chronological-based methods, with the significantly reduce the operational scenarios of REGs. The
reduction of investigated data set, the number of proposed method has comparable competitive overall cost,
system scenarios of HPS can be significantly reduced, and it can be concluded that the benefits of the reduction of
8 International Journal of Rotating Machinery

Table 4: Performance comparison.

Method 𝑁𝑤 𝑁PV 𝑁𝑑 𝑁bat Cost/M$ CO2 /kg Time/s

Proposed method 15 126 62 864 2,03M 556,000 27.5
Basic FCM 10 96 56 956 2.20M 542,000 27.8
Chronology-based 13 104 64 886 2.06M 560,000 126

3.5 power system of Solar/Wind/Energy Storage,” IEEE Transac-

tions on Sustainable Energy, vol. 5, no. 2, pp. 363–371, 2014.
[4] R. R. Barton, “Presenting a more complete characterization of
uncertainty: Can it be done,” in Proceedings of the in Proceedings
of the Simulation Society research workshop (INFORMS ’07 ), pp.
3 26–60, 2007.
Cost (M$)

[5] E. Ela and M. O’Malley, “Studying the variability and uncer-

tainty impacts of variable generation at multiple timescales,”
IEEE Transactions on Power Systems, vol. 27, no. 3, pp. 1324–1333,
2012.
2.5
[6] A. Keane, L. F. Ochoa, C. L. T. Borges et al., “State-of-the-
art techniques and challenges ahead for distributed generation
planning and optimization,” IEEE Transactions on Power Sys-
tems, vol. 28, no. 2, pp. 1493–1502, 2013.
2 [7] A. Mohamed Abd El Motaleb, S. Kazim Bekdache, and L.
0 5 10 15 20 25 30 35 40 45 A. Barrios, “Optimal sizing for a hybrid power system with
Iteration performance wind/energy storage based in stochastic environment,” Renew-
able and Sustainable Energy Reviews, vol. 59, pp. 1149–1158, 2016.
，０３０Ｇ；Ｒ = 0.006
[8] E. Koutroulis, D. Kolokotsa, A. Potirakis, and K. Kalaitzakis,
，０３０Ｇ；Ｒ = 0.005
“Methodology for optimal sizing of stand-alone photovoltaic/
，０３０Ｇ；Ｒ = 0.004
wind-generator systems using genetic algorithms,” Solar Energy,
Figure 10: Sensitivity analysis of LPSPmax and cost. vol. 80, no. 9, pp. 1072–1088, 2006.
[9] S. Ahmadi and S. Abdi, “Application of the Hybrid Big Bang-
Big Crunch algorithm for optimal sizing of a stand-alone hybrid
PV/wind/battery system,” Solar Energy, vol. 134, pp. 366–374,
CO2 and the improvement of LOSP are negatively related to 2016.
the cost optimization procedure. [10] A. Maleki and F. Pourfayaz, “Optimal sizing of autonomous
The future work will include the following: hybrid photovoltaic/wind/battery power system with LPSP
technology by using evolutionary algorithms,” Solar Energy, vol.
(1) Improving the clustering model to further study the 115, pp. 471–483, 2015.
correlation among renewable resources [11] K. Benmouiza, M. Tadj, and A. Cheknane, “Classification
(2) Adding the local and global control strategy to the of hourly solar radiation using fuzzy c-means algorithm for
power balance analysis process optimal stand-alone PV system sizing,” International Journal of
Electrical Power and Energy Systems, vol. 82, pp. 233–241, 2016.
(3) Extending the proposed method to the operation plan [12] M. Mosadeghy, R. Yan, and T. K. Saha, “A time-dependent
stage of the HPS. approach to evaluate capacity value of wind and solar PV
generation,” IEEE Transactions on Sustainable Energy, vol. 7, no.
1, pp. 129–138, 2016.
Conflicts of Interest
[13] L. Xu, X. Ruan, C. Mao, B. Zhang, and Y. Luo, “An improved
The authors declare that there are no conflicts of interest optimal sizing method for wind-solar-battery hybrid power
regarding the publication of this paper. system,” IEEE Transactions on Sustainable Energy, vol. 4, no. 3,
pp. 774–785, 2013.
[14] Y.-Y. Hong and R.-C. Lian, “Optimal sizing of hybrid wind/PV/
References diesel generation in a stand-alone power system using markov-
based genetic algorithm,” IEEE Transactions on Power Delivery,
[1] R.-J. Wai, S. Cheng, Y.-F. Lin, and Y.-C. Chen, “Installed capacity vol. 27, no. 2, pp. 640–647, 2012.
selection of hybrid energy generation system via improved [15] S. N. Gopiya Naik, D. K. Khatod, and M. P. Sharma, “Analytical
particle-swarm-optimisation,” IET Generation, Transmission approach for optimal siting and sizing of distributed generation
and Distribution, vol. 8, no. 4, pp. 742–752, 2014. in radial distribution networks,” IET Generation, Transmission
[2] J. G. McGowan and J. F. Manwell, “Hybrid wind/pv/diesel and Distribution, vol. 9, no. 3, pp. 209–220, 2015.
system experiences,” Renewable Energy, vol. 16, no. 1-4, pp. 928– [16] X. Yejun, “Fitzwilliam College,” in Proceedings of the 10th
933, 1999. International Conference on Computer Science &amp; Education
[3] A. Arabali, M. Ghofrani, M. Etezadi-Amoli, and M. S. Fadali, (ICCSE’ 15), pp. 22–24, Fitzwilliam College, Cambridge Univer-
“Stochastic performance assessment and sizing for a hybrid sity, Cambridge, UK, 2015.
International Journal of Rotating Machinery 9

[17] A. S. Dobakhshari and M. Fotuhi-Firuzabad, “A reliability

model of large wind farms for power system adequacy studies,”
IEEE Transactions on Energy Conversion, vol. 24, no. 3, pp. 792–
801, 2009.
[18] L. Chen, H. Chen, Z. Shu, G. Zhang, T. Xia, and L. Ren,
“Comparison of inductive and resistive SFCL to robustness
improvement of a VSC-HVDC system with wind plants against
DC fault,” IEEE Transactions on Applied Superconductivity, vol.
26, no. 7, pp. 1–8, 2016.
[19] Y. Li and F. Yu, “A new validity function for fuzzy clustering,”
in Proceedings of the International Conference on Computational
Intelligence and Natural Computing (CINC ’09), vol. 1, pp. 462–
465, IEEE, June 2009.