IET Generation, Transmission & Distribution

Research Article

Sequence-based differential evolution for ISSN 1751-8687

Received on 25th August 2018
Revised 30th April 2019
solving economic dispatch considering Accepted on 15th May 2019
E-First on 27th June 2019
virtual power plant doi: 10.1049/iet-gtd.2018.6432
www.ietdl.org

Yude Yang1,2, Bori Wei1,2, Zhijun Qin1

1School of Electrical Engineering, Guangxi University, Nanning, People's Republic of China
2Guangxi Key Laboratory of Power System Optimization and Energy Technology, Guangxi University, Nanning, Guangxi, 530004, People's
Republic of China
E-mail: [email protected]

Abstract: The economic dispatch (ED) of distributed energy resources (DERs) is an important and challenging problem in smart
grid operation. Different frameworks are proposed for solving this issue, among which the virtual power plant (VPP) is
considered as a promising means for ED in smart grids with DERs. This study presents a novel optimisation algorithm entitled
sequence-based differential evolution (SDE) for solving generation dispatch among several DERs with VPPs. In the proposed
method, a novel and efficient initialisation technique, namely sequence-based deterministic initialisation, is used to generate
high-quality initial population. Besides, a self-adaptation mechanism is utilised to eliminate the difficulty of tuning the problem-
specific control parameters of the algorithm. Subsequently, the SDE is applied for solving the ED model to obtain the optimal
power generation share of the VPP. Case studies compare the solutions obtained by the proposed method with several state-of-
the-art algorithms in the existing literature. It is validated the proposed method is advantageous for ED with VPPs in better
solution accuracy, less computational burden, and faster convergence.

1 Introduction sell energy to market or store in storage system based on the

exchange of information (e.g. market prices).
1.1 Background Since the concept and framework of VPP were proposed it has
The share of renewable power (wind and solar) in smart grids is drawn great attention worldwide and become an active area of both
increasing rapidly in recent years, driven by the lack of fossil fuel research and industry applications. Several nations have developed
and the concern for reducing environmental pollution. The power VPPs to counter an increasing penetration of RESs in power
generated by renewable energy sources (RESs) will play an systems and the increasingly distributed nature of the smart grids.
increasingly important role in smart grids because of small From 2001 to 2005, Spain and Germany jointly implemented the
pollutant emission, high efficiency, and less transmission cost. project to establish a virtual fuel cell power plant [4]. From 2005 to
However, the intermittent nature of RES also supposes a tough 2008, in order to ensure the sustainable development of the
challenge for the operation of smart grids, namely, maintaining European power system, a plan to establish a flexible electricity
power imbalance between generation and consumption is much network for managing large VPPs and distributed energy sources is
more challenging [1]. Therefore, one of the most challenging issues proposed by eight countries including the United Kingdom, France,
of the smart grid is its ability to facilitate the integration of RES, and Spain [5]. From 2009 to 2012, Germany and Denmark
controlling the power generation of RES within a small time proposed the sustainable energy and open networks program, a
interval [2]. Nonetheless, the variability and uncertainty of RES, distributed and integrated market considering electric vehicle
along with the stochastic feature of appliance usage of end-users, established [6]. From 2010 to 2015, Europe finished
makes maintaining the power balance between supplied sides and WEB2ENERRGY program, implemented intelligent energy
demand sides even more difficult. Over the last few years, management and power distribution automatic technology in smart
researchers took effort to find an efficient way to integrate the power distribution [7]. In August 2016, AGL Energy, Australia's
RES, leveraging its advantages while mitigating its disadvantages. largest privately owned and operator of renewable energy assets,
The concept and framework of virtual power plant (VPP) proposed announced the launch of the world's largest VPP, ultimately
by Awerbuch and Preston in 1997 has been widely acknowledged involving 1000 connected batteries installed in homes and
as a practical and feasible operational strategy for optimal dispatch businesses in South Australia, providing 5 MW of peaking capacity
of distributed resources including renewable energy. and offering customers the opportunity to save on their energy
The operation of VPP is briefly outlined as follows. A VPP is bills. In May 2017, the first set of ‘source net load smart grid’
defined as a cluster of generating units, storage systems, and system for large-scale clean energy consumption successfully put
demands, which are aggregated and operate as a unique entity with into operation in Jiangsu province, China.
the objective of optimising the energy resources [3]. The idea is
outlined with respect to the VPP in Fig. 1. In practice, VPPs turn 1.2 Related work and research gap
the original complex fully concentrated grid into a distributed
manner among several distributed units, each one of which has a Literature review shows that different methods have been proposed
control management centre, namely energy management system for dealing with the optimisation problems related to different
(EMS), which gathers all the needed information and performs the fields of VPP. The mainly concerned area of prior research is about
balancing task. At any moment the power supply and demand the strategy of a VPP to participate in energy markets. Sarker et al.
imbalance occurs, the EMS sends message to the market and buy [8] propose a bidding strategy for an EV aggregator participating in
energy from it, and determine disconnected some controllable both day-ahead and reserve markets, and compensating EV owners
loads when it is unavoidable. In the opposite case, if more energy for degradation is considered. In [9], the optimal bidding strategy
than required is produced, then the EMS will ascertain whether to of a commercial VPP participates in the day-ahead electricity
market was developed, uncertainties in the distributed energy

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215 3202
© The Institution of Engineering and Technology 2019
 aspects. Nonetheless, the application of global optimisation
methods in this research area needs further investigation.
Furthermore, we note that, in existing studies, the formulation of
economic dispatch (ED) with VPP mainly defined as a day-ahead
optimisation model with the time resolution as 1 h. Practically, the
characteristics of renewable power generation, energy storage
system, the intermittency of renewable generation, demands
uncertainty and dynamic market pricing in the intra-hour time scale
have not be thoroughly addressed.

1.3 Paper contribution and organisation

On viewing the above research gaps, we aim to develop a
comprehensive ED model considering VPPs in a more detailed
time scale to achieve an optimal operational strategy for VPPs.
Towards realising this vision, we conducted the relevant research
work of this article. The main contribution of the study in this
paper, i.e. an effective evolutionary algorithms based on
Fig. 1  Idea of the modelled VPP differential evolution (DE) called sequence-based DE (SDE), have
been proposed for solving the complex power generation task of
resource (DER) production and load consumption, in real-time VPPs.
market prices and in the bids of other market participants were DE, which was first proposed by Storn and Price in 1997 [23],
considered. Reference [10] develop a VPP participates in both day- has advantages such as simple structure, accuracy, efficiency, and
ahead and real-time markets with the sequential bid. In [11], the robustness [24, 25]. In this paper, we put forward corresponding
profit of a VPP participates in spinning reserve markets was modifications to improve the performance of the algorithm after
assessed by considering different risk levels for real-time carefully considering the effect of different stages on the algorithm.
imbalances. A VPP was developed as a service-centric aggregator In them, a new initialisation technique, which is capable of
for both DERs and distribution system operator in [12]. Nguyen et covering the entire search space is applied for generating a more
al. [13] presents a bidding strategy for VPP participates in both the reliable initial population, therefore, the stability of the results
regular electricity market and the intraday demand response provided by the algorithm is improved, which is very important for
exchange market, the VPP can be offer demand response services. the metaheuristic evolutionary algorithm; and a self-adaptive
The impact of incorporating energy storage devices into a VPP for tuning mechanism is introduced for eliminating the difficulty of
performing generating system adequacy was assessed in Reference selecting the best control parameters. Then, an improved version
[14]. DE is thus generated. Furthermore, a new constraint handling
Another critical aspect of the VPP operation is, however, the strategy is proposed for handling various constraints without
uncertainty in the area of RES's generation, loads, price, etc. The adding any additional goal on the objective function and to prevent
EMS performs the balancing task based on the assumption that at premature convergence by introducing a diversity strategy. Indeed,
each time, the RESs have the ability to supply the demanded the proposed measures have a positive and reliable effect on the
power. If the aforementioned assumption cannot be achieved, there convergence of the algorithm and the quality of the solution
may be an imbalance created between supply and demand, provided by the algorithm. Results of the proposed algorithm are
resulting in economic losses and even serious accidents. Amongst compared with other algorithms that are recently published. The
current literature in this area, we can observe that the most numerical analysis of results confirm the main strong points of the
important components of a VPP scheduling that can be uncertain proposed method, which are as follows:
are wind generation, solar generation, load demand, and market
price [15]. There are some research articles that focus on identifing (i) Accuracy: the accurateness of solution is attested by the
the uncertainty analysis methods and which one of the methods is benchmark results provided by other three famous algorithms.
more beneficial and unbeneficial from the view of EMS [16–20]. (ii) Fast convergence: the proposed algorithm defeats some well-
To successfully maintain VPP's operation, power producers in known algorithms in both the iteration number and the calculation
the pool should generate power at a competitive tariff rate and thus time.
achieve minimum cost. Hence, the scheduling problem is one of (iii) Easy implementation: this characteristic makes the proposed
the core issues of VPP in the smart grid, and optimal scheduling algorithm easy to be mastered, and in particular, can be successful
method is the perfect answer to this question. In order to deal with in large-scale fields.
the complexity of the power system, including the intermittent (iv) Large scope of application: this remarkable feature makes it
nature of RES and load demand uncertainty, optimisation method possible to quickly apply the proposed method to deal with other
for solving nonlinear and non-convex problems quickly and with optimisation problems in the power system.
high quality is currently a most interesting and challenging area of
research in VPP environment. Many literature works have already In this paper, our work mainly focuses on internal optimisation of
discussed VPP's power scheduling problem. A stochastic adaptive VPP, the purpose of VPP in the market environment is to verify
optimisation approach (an off-line solution method) for the power that VPP participating in power market can improve its operational
scheduling problem of a VPP participating in the day-ahead and efficiency. Real-time market optimisation scheduling is not the
the real-time energy markets is proposed in [3]. In [21], microgrid focus of our manuscript research. We hereby justify our proposed
scheduling with multi-energy systems has been formulated as a as follows: (i) in the internal dispatching optimisation of VPPs, the
rolling horizon Markov decision process (MDP) and the rollout computational time required for optimal dispatching is enough to
algorithm is applied for solving the large state and decision space support the meta-heuristic algorithm to perform enough operations.
of MDP. In Crisostomi et al. [22], an efficient algorithm called (ii) From a statistical point of view, the meta-heuristic algorithm
Additive-Increase-Multiplicative-Decrease algorithm was used to can guarantee stable and satisfactory optimisation solution under
share the power generation task of VPP in distributed fashion; the condition of multiple operations.
however, it requires the existence of gradient of objective function, The rest of this paper is organised as the follows: Section 2
which may be limited because of the nonlinear characteristics presents the scheduling problem formulation of modelled VPP.
aroused by practical features such as prohibited operating zones, Section 3 introduces the proposed methodology. Simulation results
valve point effect etc. in the actual power system. According to the and discussion are shown in Section 4, followed by the conclusions
previous survey, we remark that the power scheduling problem of and future work in Section 5.
VPPs participating in the market has been studied from different

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215 3203
© The Institution of Engineering and Technology 2019
 2
2 Mathematical formulations of ED with VPP C(Pcg, i(th)) = αi(Pcg , i(th)) + βiPcg, i(th) + γi

min
(3)
The VPP aims to integrate synchronous scheduling of different + λi(ρisin(Pcg , i (th) − Pcg, i(th)))
resource pools, engage the market as a whole, and obtain the best
economic benefits when providing appropriate power balance. In where, λi($) and ρi(rad/MW) are the valve-point coefficients of the
other words, the distributed generation resources in the region are min
physically interconnected to the upstream distribution networks, generator i, Pcg , i (th) is the minimum power output of the ith

and the VPP has been empowered to participate in the electricity conventional thermal unit in time slot th.
market on behalf of all participants in the entire region. As a The cost function for the jth CHP unit is modelled as below
representative of the overall interests of all resource providers, the
VPP trades with the market and is bound to face challenges from 1
Cchp, j(th) = cNG ⋅ ⋅P (t ) ⋅ τ (4)
other competitors. Thus, the VPP should optimise its scheduling ηchp, j chp, j h
strategy to enhance its own advantages in the market and maximise
the entire portfolio's benefit, subject to various operational 1 2
cNG ⋅ ⋅P (t ) = achp, j × Pchp
constraints. ηchp, j chp, j h ,j
(5)
In the context of this paper, we make the following assumptions
on the operation of the VPP. First, the distributed generation units +bchp, j × Pchp, j + cchp, j .
in the VPP are centrally controlled. That is to say, despite that the
DERs are distributed geographically, the region being considered where cNG and ηchp, j are the natural gas price per amount of energy
are treated as an entity, and the DERs in one region are dispatched and electrical efficiency of CHP, respectively. In the actual
according to the scheme obtained by the proposed algorithm. operating environment, the cost function of the CHP unit is
Second, the VPP can obtain the required data (e.g. price and load), dependent on heat production and power production,
and that relevant time-series forecast techniques are available [26]. simultaneously, which can be expressed as follows [29]:
Third, modelled VPP participates in an electricity market
environment with a dynamic price, which is not influenced by its Cchp, j(th) = a j(P j)2 + b jP j + c j + d j(H j)2 + e jH j + f jH jP j (6)
scheduling strategy. In this environment, the transmission loss of
the power grid is borne by the system operators. Cho, k = ak(Hk)2 + bkHk + ck (7)

2.1 Objective function where a j, b j, c j, d j, e j, and f j are the coefficients of jth CHP unit.
The objective of the ED problem is to minimise the operational Cho, k is the cost of kth heat-only unit for producing H MWth heat
cost of the VPP in the time horizon H with a time step size τ, and and ak, bk, and ck are its coefficients.
here H = 24 h and τ = 5 min. Thus, in this paper, the objective The operational cost Csto, s(th) of sth storage system for interval τ
function formulated with the generation cost of the conventional can be modelled as below
generation unit, the operational costs of the RES (expressed in the
form of generation cost), the fuel cost of combined heat and power Csto, s(th) = csto, s ⋅ Psto, s(th) ⋅ τ (8)
(CHP), the operation cost of the heat-only unit, the operational cost
of storage unit, and energy exchange cost of market can be written where csto, s is the average operational cost of storage per unit
as power.
H Ncg Nchp Nsto
2.2 Power balance constraints
Min: ∑ ∑ Ccg i(th) + ∑ Cchp j(th) + ∑ Csto s(th)
, , ,
th = 1τ i = 1 j=1 s=1 For the stable operation of VPP, it must safeguard the generated
(1) power in accordance with the load demand power at each th. Thus,
Nho Nres
+ ∑ Cho, k(th) + ∑ Cres r(th) + PEM(th) ⋅ cEM(th) ⋅ τ
, the power balance equation can be formulated as
k=1 r=1
Ncg Nchp Nsto
where Ncg, Nchp, Nho, Nsto, and Nres are the number of conventional ∑ Pcg i(th) + ∑ Pchp j(th) + ∑ Psto s(th)
, , ,

generation unit, CHP unit, heat-only unit, storage station, and RES i=1 j=1 s=1
(9)
units, respectively; Ccg, i, Cchp, j, Csto, s, Cho, k, and Cres, r represents Nres

the generation cost of ith conventional generation unit, the + ∑ Pres, r(th) + PEM(th) = Pl(th) .
generation cost of jth CHP unit, the generation cost of kth heat- r=1
only unit, the operate cost of sth storage unit and the generation
cost of rth RES unit, respectively; PEM is the power exchanged where Pchp, Psto, and Pres are power generation of CHP, storage and
toward electricity market and cEM is the market price; th denotes RESs, respectively; Pl is the load demand power.
current time slot.
In the following equations, a more detailed description of the 2.3 Heat balance constraints
different components of the objective function is given. The cost Nchp Nho
equations for each renewable resource are modelled as a quadratic
function, which is commonly used in the VPP literature [22, 27, ∑ Hchp j(th) + ∑ Hho k(th) = HD(th) .
, , (10)
j=1 k=1
28]. It can be shown as follows:
2
where PD is the thermal demand of the system.
C(Pres, r(th)) = αrPres , r (th) + βr Pres, r (th) + γr . (2)
2.4 Power generation limits
where C is the hourly cost in currency unit per hour, αr($/MW2),
βr($/MW), γr($) are the cost coefficients of the rth RES unit. In realistic power systems, the generation of resources should be
The cost function of the conventional thermal unit becomes within its generation range.
non-convex by adding a sinusoidal term to the quadratic objective
function when considering the valve-point effects and can be • Capacity limits of the conventional thermal unit
modelled as min max
Pcg , i (th) ≤ Pcg, i(th) ≤ Pcg, i (th) (11)

3204 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215
© The Institution of Engineering and Technology 2019
 min max based on the power levels of the storage system at which the
where Pcg , i and Pcg, i are the lower and upper bounds for the
power output of the ith conventional thermal unit. scheduling starts.
• Capacity limits of CHP units When using the simplified model
in [22], the power production range constraints of CHP unit can 2.5 Power flow constraint
be expressed as N
PGk − PDk − V k ∑ Y ktV tcos(δk − δt − θkt);
min max
Pchp , j(th) ≤ Pchp, j(th) ≤ Pchp, j(th) (12) t=1
N
When considering a more practical situation, as done in QGk − QDk − V k ∑ Y ktV tcos(δk − δt − θkt);
Document 12, the production range constraints of the CHP unit t=1
can be expressed as [30–32] min max (21)
PGk ≤ PGk ≤ PGk ;
min max
min
Pchp max
, j(H j)(th) ≤ P j(th) ≤ Pchp, j(H j)(th) (13) QGk ≤ QGk ≤ QGk ;

min max
V kmin ≤ Qk ≤ V kmax;
Hchp , j(P j)(th) ≤ H j(th) ≤ Hchp, j(P j)(th) (14)
Pktmin ≤ PGk ≤ Pktmax;
min max
where Pchp , j(H j) and Pchp , j(H j)
are the lower and upper bounds
where N is the total number of buses; PGk and QGk are the active
for the jth CHP unit which is the function of generated heat H j;
min max power and reactive power generation of the kth bus, respectively;
Hchp , j and Hchp, j are heat generation limits of the jth CHP unit PDk and QDk are the active and reactive power load demand of the
which is the function of generated power P j. kth bus, respectively; V k is the magnitude of the voltage of the kth
• Production limits of the heat-only units bus and δk is the phase angle of the voltage of the kth bus; Y kt and
θkt are the magnitude and phase angle of branch admittance
Hkmin(th) ≤ Hk(th) ≤ Hkmax(th) (15)
between the kth bus and the tth bus, respectively. ∙max and ∙min are
the minimum and maximum limitation of the corresponding
where Hkmin and Hkmax are the minimum and the maximum limit variables, respectively; Pkt is the real power flow between the kth
for the heat output of the kth heat-only unit in MWth. bus and the tth bus.
• Production limits of the renewable units and the market To sum up, the proposed simplified model, in which the
objective function is convex can be expressed as below
min max
Pres , i(th) ≤ Pres, i(th) ≤ Pres, i(th) (16)
H Ncg Nchp Nsto

We note that the production of renewable units is uncertain. In Min: ∑ ∑ Ccg i(th) + ∑ Cchp j(th) + ∑ Csto s(th)
, , ,
th = 1τ i = 1 j=1 s=1
this paper, we model the output of renewable units using the
uncertainty set (or confidence interval), which describe the upper Nho Nres (22)
and lower bound of the uncertain variables. The proposed method + ∑ Cho, k(th) + ∑ Cres r(th) + PEM(th) ⋅ cEM(th) ⋅ τ
,
can also be applied to solve scenario-based stochastic optimisation k=1 r=1
problems and will be addressed in our future work. s.t. (2), (4), (5), (7) − (12), and (15) − (21) .
The exchange power of the market in each time period instant
depends on the transmission capability of the connection line. Also, another more realistic and the objective function is non-
Thus, there is a limit to where the market can produce/consume convex, and can be written as
power
H Ncg Nchp Nsto
min
PEM max
(th) ≤ PEM(th) ≤ PEM (th) (17) Min: ∑ ∑ Ccg i(th) + ∑ Cchp j(th) + ∑ Csto s(th)
, , ,
th = 1τ i = 1 j=1 s=1

where Pλmin max

, i and Pλ, i (λ = res, EM) are the minimum and maximum
Nho Nres (23)
generation of the ith resource. + ∑ Cho, k(th) + ∑ Cres r(th) + PEM(th) ⋅ cEM(th) ⋅ τ
,
k=1 r=1
In this paper, the energy storage systems, which aims at making
effective use of the power energy and mitigating the impact of the s.t. (3), (6), (7), (8), (11), and (13) − (21) .
intermittency of the renewable energy resources, are defined as
devices which are able to transfer, store, and produce energy at the 3 Solution methodology
required time [33]. Considering the efficiency of charge and
discharge, both the power level at the previous time step (th − τ) In this section, the improved version of DE used in this paper is
and the power have been injected into storage determine the power described in detail. DE is known as high-efficiency optimisation
level at each time th algorithm and has been successfully applied to solve many
complex optimisation problems, especially in areas where
traditional algorithms are limited. However, it also has some
Psto, i(th) ⋅ τ
Esto, i(th − τ) − defects such as the inability to guarantee convergence to the global
ηdis, i optimum, the need for adjusting parameters based on problems,
Esto, i(th) = Psto, i(th) > 0; (18) and available domain information is often ignored. The key to
avoid these bad situations mentioned is the maintenance of
Esto, i(th − τ) − Psto, i(th) ⋅ τ ⋅ ηch, i,
population diversity, the selection of control parameters, the design
Psto, i(th) ≤ 0. of constraint handling mechanism, etc. Special enhancement in
redesigning the DE algorithm to address these challenges is
min
Psto, i(th) ≥ Psto discussed in the following sections.
, i(th) (19)

max
Psto, i(th) ≤ Psto 3.1 Sequence-based deterministic initialisation
, i(th) (20)
The DE is categorised as the population-based algorithm. For these
The power capacity boundaries of the storage system have been methods, an individual of the population represents a point that
described in (18), (19). The initial values Esto, 0 can be assigned moves in the search space with the optimisation process. The
structure of each individual, in this paper, can be described as

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215 3205
© The Institution of Engineering and Technology 2019
 Xi = [PEM, i, Pcg, 1, i, . . . , Pcg, Ncg, i, Pchp, 1, i, . . . , Pchp, Nchp, i,
Psto, 1, i, . . . , Psto, Nsto, i, Pres, 1, i, . . . , Pres, Nres, i, Hho, 1, i, . . . , (24)
Hho, Nho, i, Hchp, 1, i, . . . , Hchp, Nchp, i]; i = 1, 2, . . . , N .

where N is the total number of population.

It is well known that the performances of evolutionary
algorithms are influenced by the quality of their initial populations.
So, the number of new initialisation techniques recently proposed
has dramatically increased. In them, a new initialisation technique,
named sequence-based deterministic initialisation, has been proven
to have excellent performance [34]. In [34], their work focuses on
unconstrained problems of linear and convex programming. This
initialisation technology is first applied to solve the complicated
nonlinear and non-convex power system optimal scheduling
problem with various constraints in our research. The details
process of the sequence-based deterministic initialisation can be
depicted as below.
Initially, the search domain [x − x] (in this paper,
max max max max max max min min min
x = [Pcg , i , Pchp, j, Pres, r , Psto, s, PEM , Hho, j ], x = [Pcg, i , Pchp, j, Pres, r ,) is
min min min
Psto, s, PEM , Hho, j] Fig. 2  Algorithm 1: population initialisation
divided into n + 1 segment vectors. The interval (I) of each
segment vector can be determined using
→ x−x
I = . (25)
n

Then, the kth segment vector (Sk, ∀k = 1, 2, . . . , n + 1) can be

generated as
→
x k = 1;
→ → →
Sk = S k−1 + I 2 ≤ k ≤ n; (26)
x k = n + 1.

Subsequently, a set of sequences can be generated as described in

Algorithm 1 (see Fig. 2).
Considering the constraints (13) and (14), heat generated by jth
CHP unit (Hchp, j, i) of ith individual in the population can be
generated as

H chp chp, min chp,

j, i = H j (P j, i ) + rand( . ) × (H chp
j
, max
(Pchp ,
j, i )
(27)
−H chp
j
, min
(Pchp
j, i )) .

where rand( . ) is a random number between 0 and 1.

The maximum number of points generation Gmax of this
technique is

Gmax = (n + 1) × ((D × n) + 1) . (28)

where D is the dimensionality of the model.

In this paper, set n = 10 (generates (10 + 1) × ((D × 10) + 1)
individuals) and 100 initial individuals selected from the initial
population pool as the final initial population. As selecting the best
solutions from a large population of individuals may reduce
diversity, especially for multimodal problems, and to gain benefits
from the proposed method, the selected individuals could be taken
from different areas of the search space. Thus, we can get high- Fig. 3  Algorithm 2: a general framework for DE/rand/1 algorithm
quality initial solutions without overly reducing randomness.
Although DE has been widely considered as an optimisation
3.2 Differential evolution method with stability, reliability, and small computing burden,
some potential pitfalls sometimes leave it in an awkward situation
Two search operators, one named mutation and other one called such as convergence to an inappropriate position (i.e. it can become
crossover are used in DE. First, a mutation solution is generated by stuck in a local optimum). Furthermore, the existing mechanism of
mutation operator. Then, a child solution is generated by the diversifying populations does not have the ability to help it escape
mutation solution and pre-determined parent solution using a from the local optimum. Therefore, a new powerful mutation and
crossover operator. Finally, between child solution and parent crossover mechanism, which mainly enhances the ability of the
solution, one with better fitness is chosen as offspring solution. The algorithm to exploit the unknown area of the search space, is
main steps of the general DE/rand/1 are presented in Algorithm 2 introduced to solve the aforediscussed problems, thus increasing
(see Fig. 3). the ability of the DE to eventually be enhanced. Furthermore, when

3206 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215
© The Institution of Engineering and Technology 2019
the DE is applied to solve constrained optimisation problems, its k k k
k+1
Ym , if (FC(Y m ) ≤ FC(Xm ));
performance depends largely on control parameters and it is Xm = (32)
k
generally difficult to select the best control parameters based on the Xm , otherwise .
problem [35]. Thus, in this paper, a self-adaptive method is applied
→ →
for selecting the most suitable control parameter, which is It is important to point out that the value of F and C r are also
described below. updated with the optimisation process. If the new solution is better
→ →
than the current solution, then, F m = Fm and.C rm = Crm
3.2.1 Mutation and crossover mechanism: Since mutation has
been applied to the algorithmic process, many mutation operators
have been proposed [36]. Unfortunately, there exists no single 3.3 Constraint handling strategy
optimal solution to all problems. We have noticed that the The ED problem is a nonlinear constrained optimisation problem,
optimisation mechanism of DE itself is simple and efficient, even which contains a large amount of equality and inequality
adding additional strategies that increase search power will not constraints. So, it is obvious that an appropriate constraint handling
have an unacceptable negative impact on the computational mechanism can improve the performance of the algorithm.
efficiency of the algorithm. Therefore, a new powerful mutation Generally, there are two strategies to deal with constraints of the
strategy that contains two mutation operators is considered in this ED problem, one is to use a penalty function which is achieved
paper. The main effect of the first mutation operator is to speed up through adding an extra objective function for punishing violations
the convergence, and the second one is to diversify the population. of constraints on the original objective function, and the other way
And a crossover operator named binomial, which has been proven is to generate solutions that satisfy all constraints by some
to have excellent performance [37], is considered. The generation strategies and maintain the feasibility of the solution in the
process of offspring is as described below. optimisation process so that optimisation is only done in the
First, in the mutation events, three individuals v1 to v3 obtained feasible region. The first method is simple and can maintain
from the population are randomly selected as v1 ≠ v2 ≠ v3 ≠ m. population diversity but not adequate for handling constraints. The
k
Consequently, a child individual Y m is generated as second method has high efficiency in finding a feasible solution but
lose population diversity. In this paper, a new constraints handling
Xvk1 + Fm(Xvk2 − Xvk3), mechanism can avoid their disadvantages while retaining
advantages has been proposed for handling various constraints,
if rand1 ≤ Crm and rand2 ≤ 0.5; which process will be described in detail next.
k k
Ym = Xm + Fm((Xvk1 − Xvk3) + (Xbesk
k
− Xvk2)), (29)
(1). Arrange individual (xi; i = 1, 2..., N) into a matrix form as
if rand1 ≤ Crm and rand2 > 0.5;
k
Xm , otherwise . x11 x12 ... x1D
⋮ ⋮ ⋱ ⋮
where, rand1 and rand2 are random numbers generated from a x = xi1 xi2 ... xiD (33)
uniform distribution in the interval [0, 1]. Fm is the amplification
⋮ ⋮ ⋱ ⋮
factor and Crm is the crossover rate. They should be fixed values,
but picking the optimum values for a specific problem is tricky. xn1 xn2 . . . xnD
Thus, a self-adaptation strategy was introduced to select the most (2). Randomly select a solution and check whether the solution
appropriate value, as described in the next subsection. satisfies system constraints. If a variable ( xij, j = 1, 2,..,D) is out of
the boundary, then
3.2.2 Self-adaptive tuning control parameters: For each
individual in the search space, with two control parameters (F and xij(th) = xij, min(th), if xij(th) ≤ xij, min(th) . (34)
→ →
Cr). In the beginning, F ∈ N(0.5, 0.1) and C r ∈ N(0.5, 0.1).
N(0.5, 0.1) means a normal distribution whose mean equals to 0.5 xij(th) = xij, min(th), if xij(th) ≥ xij, min(th) . (35)
and standard deviation is 0.1. Consequently, Fm and Crm in (29) are
generated as described below (3). To make solutions satisfy equality constraints. Here, first,
check if the solution meets the power balance constraint. The
→ → → feasibility of a solution is checked as
F v1 + rand1(F v2 − F v3), if (rand2 < δ);
Fm = (30) Np
rand3, otherwise .
∑ (Pi) − (Pload + Ploss) = ΔP < ε . (36)
→ → → i=1
Crv1 + rand4(Crv2 − Crv3), if(rand5 < δ);
Crm = (31)
rand6, otherwise . where, Np is the total number of the power generation unit, ε is a
tolerance limit factor, the value of ε, in this paper, from a larger
→ → initial value gradually reduced to a small final value set to 10−8.
where F mλ and C rmλ (λ = 1, 2, 3) are parameters of the
→ → The way ε changes can be given as
corresponding individual ( Xvkλ) in F and C r, respectively; and
randμ ∈ (0, 1) (μ = 1, 2, . . . , 6), are generated using uniform ε(0) = ϕ(xθ) = εθintial; (37)
distribution of 0–1. The value of δ, in this paper, set to 0.75
according to the test, appropriate range of Fm and Crm is 0.1 to 1, cp
k
ε(0) 1 − , 0 < k < T c;
so, if their value is outside this range, it is truncated to 0.1 and 1, ε(k) = Tc (38)
respectively [38].
The mechanism followed by a greedy selection process is such 10−8, k ≥ Tc .
k
that, the best one between the current individual (Xm ) and the new
k
where xθ is the top θth individual and θ = 0.4N. cp is a control
individual (Y m), will replace the position of the current individual parameter of the θ level and set to five in this paper. With the
and become the new offspring. The process can be written as number of iterations k increases to the control generation T c, The θ
level has been updated. There are no solutions that violate the
constraints in the population when the control generation is
reached, and the value of T c is 150 in this paper.

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215 3207
© The Institution of Engineering and Technology 2019
Table 1 Initial values for the controlling parameters of 3.1). If ΔH > 0, randomly select an heat generation unit Hr to
algorithms reduce its own heat production until Hr = Hrmin or ΔH > 10−8.
Algorithm Parameter Value Repeat this step until all units are selected or ΔH < 10−8.
PSOw topology fully connected
3.2). If ΔH < 0, randomly select an heat generation unit Hr to
cognitive and social 1.5, 1.5
constants increase its own heat production until Hr = Hrmax or ΔH > 10−8.
inertial weight linearly decreases from 0.9 Repeat this step until all units are selected or ΔH < 10−8.
to 0.4
population size 30 (4). Return a feasible x to the algorithm.
CBA loudness A change using the sinusoidal
As this proposed constraint handling strategy does not place any
map
priority on a unit, it will be helpful for maintaining the diversity of
pulse emission rate R updated as the iterations solutions expected in evolutionary algorithms including the
progress proposed algorithm. Also, the ε-constrained method for handling
α,γ 2.3, 0.9 the equality constraints avoid the risk of missing some solutions
population size 30 which have good objective function values but marginally violate
MFA random movement factor α monotonically decreased the constraints, enhance the convergence rate of the algorithm by
from 0.5 to 0 both increasing the population diversity and boundary searching
light absorption coefficient γ 1 ability of the algorithm.
attractiveness parameter 1 and 0.2
βmax and βmin 3.4 Implement SDE to ED problem
population size 30 Implementing the SDE for solving ED problem can be briefly
described via the following steps:

(1). Generate a sequence-based deterministic initial population as

described in Section 3.1.
(2). Make each individual satisfy constraints according to the
constraint handling strategy reported in section 3.3.
(3). Calculating the value of the objective function of all
individuals using the formula in (1).
(4). Generate new solutions using the mechanisms depicted in
Sections 3.2 and 3.3.
(5). Update offspring population based on the selection process
depicted in Section 3.2.2.
(6). Modify the solutions to satisfy the constraints using the
proposed constraint handling strategy.
(7). Check the stopping criterion. If the maximum number of
iterations has been reached, stop; otherwise, go to Step3.
Fig. 4  Schematic topology of the tested network
4 Simulation results and discussion
If the value of the power deviation is larger than the present In order to evaluate the performance of the proposed algorithm, it
value, a slack unit Ps (s = 1, 2, . . . , Np) that choose randomly from was compared here to three well-known effective algorithms, that
the unit poor was used to balance the power deviation follow the is, particle swarm optimisation (PSO) with inertia weight (PSOw)
following rules [39], chaotic bat algorithm (CBA) [40], and modified firefly
algorithm (MFA) [41]. The PSO is chosen because it is the most
Np widely used optimisation algorithm with great success, so it is the
Ps = (Pload + Ploss) − ∑ (Pi); (39) best comparative algorithm. The rest of the algorithms, BA and FA
i = 1(i ≠ s) are of the best recently proposed population-based optimisation
algorithms. Since the proposed algorithm SDE is an improved
if Ps(th) < Psmin(th), Ps(th) = Psmin(th) version of DE, hence, we prefer to use the improved version of the
(40) algorithms for comparisons. The parameters of each algorithm
else if Ps(th) > Psmax(th), Ps(th) > Psmax(th) end mentioned are shown in Table 1.
The superiority of the SDE is demonstrated through eight
If the power balance constraint is still not satisfied typical benchmark functions in the first case study. Then, the
( ΔP − ε > 10−8), similarly, one unit from the remaining units is effectiveness of the SDE for solving ED problem in smart grid is
randomly selected as the slack generator to balance the power illustrated with two test systems (Fig. 4 depicts the schematic
deviation. This process continues until all units are selected or topology of the tested network.): the second case study, a microgrid
( ΔP − ε < 10−8). including three wind plants, two PV plants and one CHP system
Then, update the boundary of Hchp, j(th) according to the operates in isolated mode; the third case study, a VPP has three
constraint (14) and reset the value of Hchp, j(th) according to (34) wind plants, two PV plants, one CHP system and one storage
system as a unit participates in market.
and (35). Next, check whether the solution satisfies the heat
The ability to consistently obtain good optimisation results, will
balance constraint
demonstrate the efficiency of the algorithms. The robustness of the
Nh proposed algorithm will be validated from the results of 50
∑ (Hi) − HD = ΔH . (41) independent runs for each case study. The simulations and
experimental test are carried out on MATLAB (R2013a)
i=1
environment using a desktop machine, which CPU is Intel
where Nh is the total number of the thermal generation unit. Core(TM) i7 processor with 3.6 G-Hz clock frequency and 8 GB of
RAM.
If ΔH > 10−8, then proceed to the following steps.

3208 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215
© The Institution of Engineering and Technology 2019
4.1 Simulation results with benchmark functions microgrid system has three wind plants, two PV plants, and one
CHP system is utilised. The parameters of DERs are shown in
This section demonstrates the excellent performance of the SDE Table 4. In this case, Ncg = 0, Nchp = 1, Nho = 0, Nsto = 0, Nres = 5,
over the other most well-regarded and recent algorithms through
NEM = 0, and the corresponding objective function is (22).
eight typical benchmark functions, which are listed in Table 2. In
the table, dimensions are the number of the variables, search range The maximum generation of each wind DER was randomly
denotes the range of the variables, and f min indicates the optimum. selected from the real wind turbine data provided by the National
Renewable Energy Laboratory (NREL) [43]. The method presented
In addition, the two-dimensional (2D) versions of each benchmark
in [22] refers to the maximum generation profile of each PV. The
functions are depicted in Fig. 5. We aim to assess the quality of the
load profiles were from random selection for a period of 24 h,
final solution and the convergence speed of each algorithm after a
according to [44]. We decided to sample the required power
fixed number of iterations is reached. Table 3 reports the
profiles and the maximum generation of the DERs every 5 min,
comparison results of SDE with advanced evolutionary algorithms.
and assumed their values are constant at intervals. The reason why
We conclude from this table that the proposed method, which gives
we adopted such a time interval instead of the universally used 24
competitive results, is very effective. Its performance surpasses that
intervals a day is that the real system executes the ED at an interval
of the listed algorithms. In short, SDE nearly provides the optimum
of 5 min.
solutions for all the test functions with the fastest convergence rate,
The errors and estimated operation costs of the 50 independent
which means that SDE outperforms other listed algorithms.
trials found by the investigated algorithms are shown in Figs. 6a
and b, respectively. The errors are defined as
4.2 Simulation results for optimised power generation in N
microgrid using SDE e(i) = ∑k =g 1 Pk − Pl − Ploss(i), and i ∈ [1, 50]. Two criteria are used
in selecting the optimal solution, one is
In this subsection, the ability of SDE to provide optimum N N H
∑k =g 1 Pk(th) ≤ ∑k =g 1 Pk, max(th), the other is ∑th = 1τ e(i)(th) ≤ 10−8.
generation scheduling of DERs in the smart grid are tested. A

Table 2 Benchmark functions

Benchmark function Name Dimensions Search range f min
n Sphere 1.2 30 [−100, 100] −80
F1(x) = ∑ (xi + 40)2 − 80
i=1
n Sphere 30 [−100, 100] 0
F2(x) = ∑ (xi)2
i=1
n n Schwefel 2.22 30 [−10, 10] 0
F3(x) = ∑ xi + ∏ xi
i=1 i=1
n Schwefel 30 [−500, 500] −12,569.5
F4(x) = − ∑ [xi × sin( xi )]
i=1
n−1 Rosenbrock 30 [−30, 30] 0
F5(x) = ∑ [100(xi + 1 − xi2)2 + (xi − 1)2]
i=1
n Rastrigin 30 [−100, 100] 0
F6(x) = ∑ [xi2 − 10cos(2πxi) + 10]
i=1

1
n
1
n Ackley 30 [−32, 32] 0
n i∑
F7(x) = − 20exp −0.2 xi2 − exp ∑ cos 2πxi + 20 + e
=1
ni=1
n n Griewank 30 [−600, 600] 0
1 xi
4000 i∑
F8(x) = (xi)2 − ∏ cos +1
=1 i=1 i

Fig. 5  2D versions of benchmark functions

(a) F1, (b) F2, (c) F3, (d) F4, (e) F5, (f) F6, (g) F7, (h) F8

Table 3 Final optimisation results for benchmark functions

Algorithm
Test function SDE MFA CBA
Minimum Average Standard Minimum Average Standard Minimum Average Standard
F1 −8.000 × 1001 −8.000 × 1001 0.0000 −79.99998 −79.99556 7.272 × 10−04 −73.11186 −67.66678 1.546 × 1001
F2 0.0000 0.0000 0.0000 3.439 × 10−08 4.770 × 10−08 5.956 × 10−09 1.837 × 10−03 2.333 × 10−01 5.678 × 10−01
F3 0.0000 0.0000 0.0000 3.659 × 10−04 5.281 × 10−04 1.043 × 10−05 1.117 × 10−03 2.148 × 10−02 5.249 × 10−02
F4 −12227.15 −11303.96 3.754 × 1002 −11049.50 −10525.03 2.837 × 1002 −10872.20 −10630.73 6.023 × 1001
F5 0.0000 0.0000 0.0000 2.178 × 1001 2.214 × 1001 5.684 × 10−01 2.843 × 1001 1.321 × 1002 1.136 × 1002
F6 0.0000 0.0000 0.0000 7.963 1.156 × 1001 2.612 9.742 × 1001 1.462 × 1002 3.463 × 1001
F7 0.0000 0.0000 0.0000 4.026 × 10−05 5.047 × 10−05 3.937 × 10−06 2.354 × 10−02 3.455 × 10−01 5.674 × 10−01
F8 0.0000 0.0000 0.0000 7.982 × 10−06 9.796 × 10−06 4.386 × 10−05 4.972 × 10−03 1.833 × 10−01 1.645 × 10−01

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215 3209
© The Institution of Engineering and Technology 2019
 
Algorithm
Test function PSOw LSA [42] BSA [42]
Minimum Average Standard Minimum Average Standard Minimum Average Standard
F1 −79.85673 −79.14578 4.157 × 10−02 — — — — — —
F2 1.854 × 10−06 2.567 × 10−05 4.527 × 10−05 1.061 × 10−19 4.810 × 10−08 3.401 × 10−07 1.432 9.967 9.812
F3 1.375 × 10−03 3.641 × 10−03 2.843 × 10−03 2.217 × 10−07 3.680 × 10−02 1.562 × 10−01 4.551 × 10−01 1.197 5.285 × 10−01
F4 −8942.432 −8459.370 3.248 × 1002 −9193.912 −8001.388 6.691 × 1002 −10357.64 −9615.769 2.530 × 1002
F5 1.276 × 1001 5.942 × 1001 5.424 × 1001 5.600 × 10−01 6.428 × 1001 4.375 × 1001 2.083 × 1002 4.715 × 1002 2.311 × 1002
F6 1.427 × 1001 4.124 × 1001 9.456 4.079 × 1001 6.276 × 1001 1.491 × 1001 5.318 × 1001 6.575 × 1001 8.029
F7 9.132 × 10 −04 2.863 × 10 −03 2.415 × 10 −04 8.730 × 10 −08 2.686 9.108 × 10 −01 6.551 × 10 −01 3.153 1.491
F8 2.894 × 10−09 1.758 × 10−02 2.352 × 10−02 2.224 × 10−16 7.241 × 10−03 6.753 × 10−03 6.793 × 10−01 1.068 8.686 × 10−02
—, data not available.
Bold and Bold-italic values will give a more intuitive understanding of the advantages of the proposed algorithm over other algorithms.

Table 4 Parameters of each DERs

Plant ai bi ci
wind plant 1 0.0027 17.83 4.46
wind plant 2 0.0028 17.54 4.45
wind plant 3 0.0026 17.23 4.44
solar plant 1 0.0055 29.30 4.45
solar plant 2 0.0055 29.58 4.46
CHP 0.0083 75.73 5.21

the total generation cost of 1481.34 ($); whereas DE, PSOw, CBA
and MFA achieves 1593.351, 1507.975, 1498.68, and 1494.384
($), respectively. Based on the generation cost of the day, the
ranking of the algorithms is SDE, MFA, CBA, PSOw and DE.
Again according to the distribution of error and the distribution of
optimised Fcost at each trial, which is illustrated in Fig. 9. It can be
dawn that SDE provides most optimum results with the most
stable, robust, and accurate, compared to other methods. The
average computational time for a single run of SDE at each period
is 2.37 s, much lower than 5 min. This illustrates that the
computing efficiency of SDE can meet the efficiency required by
real-time algorithm. Table 5 reports the generation cost at each
Fig. 6  Estimated operation costs and corresponding power mismatch with Hour × 12
hour CostHour = ∑th = (Hour − 1) × 12 + 1 Cost(th) (Hour = 1, 2, . . . , 24) of
50 trials of the investigated algorithms for case 2
(a) Error of Power mismatch, (b) Estimated operation costs
the whole day provided by mentioned algorithm. From this picture,
we can clearly obtain information about the difference in
generation cost required by the tested algorithms at each hour.

4.3 Simulation results for optimised power generation in VPP

using SDE
In this case study, a VPP is constructed based on the microgrid
system used in case-2. In this case, Ncg = 0, Nchp = 1, Nho = 0,
Nsto = 1, Nres = 5, NEM = 1, and the corresponding objective
function is (22). In VPP, the storage system and the electricity
market will participate in maintaining the power balance for each th
power production/consumption. We assume that the capacity of
Fig. 7  SDE solution manages to balance demand and provide power
power exchange between the market and VPP is restricted and
during the whole day (case 2)
there is set as 500 kW. Also, the maximum capacity of storage
charging/discharging at each time step is set as 400 kW with
The 33rd trial of SDE is selected as the optimal solution based on respect to constraints (10)–(12). Furthermore, a revised version of
the criteria, and the detail results of sharing the power generation the load profile is shown in Fig. 10, with the maximum power that
are reported in Fig. 6 with the maximum power that could be could be generated by all renewable DERs, the price of market and
generated by each DERs in each interval. Fig. 7 depicts that the storage for each th. From this picture, we can know that the
SDE solution manages to balance consumption and production of
maximum generation of RES cannot meet the requested power in
power for the whole day. It can be seen intuitively that the power
some periods. This situation embodies the important role of storage
balance constraint has been satisfied.
and market in VPP operation. In addition, two operation modes are
Fig. 8 shows the convergence properties of the SDE with DE,
chosen to demonstrate both the impact of the storage system and
PSOw, CBA, and MFA when the best results are obtained from 50
the effectiveness of the proposed algorithm. Mode1 is VPP
independent trials. We can observed from this figure that SDE operation without the storage system. Mode2 is VPP operation
gives the same convergence speed as MFA at the initial stage, with the storage system. In the VPP operation, resource in the poor
which is significantly better than PSOw, CBA, and DE. It also can is dispatched based on the objective function (1), the constraints
be seen that SDE settles at about 100 iterations and finally provides
3210 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215
© The Institution of Engineering and Technology 2019
(2)–(8). The operational cost of the storage system is referred in [5] high. Once the load demand is satisfied, the excess power will be
and here it is assumed that the modelled price maintain a constant sold to the market. The storage system charges during the time
value. periods when the market prices are low and when the generation of
The rules of choosing the optimal solution are the same as in RES is more than the required power. Moreover, the storage has
case 2, and the optimised schedules of resources provided by SDE also been charged during the time periods at which the market
are reported in Fig. 11. According to Figs. 10 and 11, the schedules price was higher than the storage price. This is because the amount
of the resources in the VPP pool depend on the market prices and of power that could be generated by storage in the later period is
the storage prices at each time step. The generation of VPP DERs determined by the number of power that was injected into storage
are more during the time periods at which the market prices are in the previous periods. Thus, the storage is charged during the
time periods when the market prices are high, in order to be able to
sell the previously stored power in the market or meet the load
demand when the market prices are higher for minimising total
cost. Fig. 12 clear display of such a practice does yield greater
benefits, increase micro costs in some time periods in exchange for
a sharp reduction in costs during other periods. Finally, the
generation of the CHP depends both on the total power that could
be generated by other resources and the prices.
Subsequently, we compare the performance indices of our
proposed method with other algorithms in terms of convergence
speed and global optimality. The convergence curves in Figs. 13a
and b are recorded for the trial where the minimum operation cost
Fig. 8  Convergence curves of the investigated algorithms (case 2) Fcost of Mode1 and Mode2 are provided, respectively. From
Fig. 13a, it can be observed that the investigated algorithms still
maintain the convergence rate and the most competitive solution
ranking consistent with what happened in case 2. The final
minimum cost value provided by SDE, MFA, CBA, PSOw, and DE
are 1585.42, 1607.98, 1620.01, 1635.00, and 1686.91 ($),
respectively. In Mode2, the problem is difficult to solve because of
the existence of constraints (10)–(12). Note that the generation
limit of the storage system in any period depends on the stored
power of the previous periods. Hence, finding the global optimal
solution of this system is harder and thus the performance of the
algorithm is tested.
From Fig. 13b, it can be seen that the investigated algorithms
converged to the minimum operation cost in more than 500
iterations, much larger than in Mode1 (<150 iterations). This
indicates that the computed cost is increased due to the increase in
Fig. 9  Detail of SDE shares the power production for case 2 the complexity of the problem. We also can observe that the MFA
has a faster convergence rate compared to that of the SDE;

Table 5 Estimated cost at each hour provided by the investigated algorithms

Hour Algorithm
DE PSOw CBA MFA SDE
1 55.3889 52.46771 52.48136 52.50293 52.44403
2 54.95305 50.7591 52.11873 50.7476 50.67691
3 51.4174 49.39592 49.68487 49.38894 49.36837
4 49.09357 49.05453 49.04481 49.71958 49.04184
5 61.85644 52.02783 52.19909 52.02653 52.02653
6 52.59333 52.58193 52.56673 53.17361 52.54944
7 60.38452 55.2904 55.28108 55.77636 55.09955
8 64.43757 57.31854 58.61125 57.33723 56.9141
9 59.94798 57.29089 57.31214 56.71952 56.45126
10 61.35654 56.95628 57.30629 57.4039 56.27716
11 61.64301 57.33354 58.33414 57.98189 56.24718
12 64.08487 57.98534 57.06252 57.58296 56.2666
13 63.07444 59.88348 58.48415 58.65837 56.90329
14 64.62493 58.51693 58.95484 58.978 58.10038
15 64.75549 59.17579 58.66327 58.958 58.15453
16 66.29486 64.74419 59.93551 59.92593 59.87642
17 65.8719 62.26609 61.54945 61.68401 61.54493
18 83.50562 72.55942 74.45309 72.91911 72.38298
19 90.08189 92.08034 88.2599 86.8824 86.8824
20 94.10895 96.91758 94.10895 94.10895 94.10895
21 93.60391 87.37512 88.73993 86.10716 86.10716
22 82.66499 79.36909 79.36909 79.50619 79.36909
23 68.42967 68.30265 65.84109 67.52221 65.83575
24 59.17739 58.32215 58.31792 58.77224 58.29577
Total cost/$ 1593.351 1507.975 1498.68 1494.384 1481.34

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215 3211
© The Institution of Engineering and Technology 2019
Fig. 10  Total power generation of wind and solar, load demand power, Fig. 13  Convergence curves of the investigated algorithms (case 3)
market price, and storage price (a) Mode1, (b) Mode2

Table 6 Comparison of cost performance among the

investigated algorithms
Algorithm Min_Cost Ave_Cost Max_Cost Std . dev .
Mode1
SDE 1585.4223 1587.4621 1588.9025 1.0036
MFA 1607.9835 1610.3550 1612.8413 1.3726
CBA 1620.0159 1624.4757 1629.4517 3.2066
PSOw 1635.0066 1643.7177 1650.6881 3.3114
DE 1686.9145 1696.0830 1709.3174 5.8984
Mode2
SDE 1493.7112 1494.9374 1496.1843 0.6052
MFA 1550.5280 1550.9995 1551.4762 0.2413
CBA 1547.8719 1559.9428 1580.6077 7.9754
PSOw 1565.5407 1574.5470 1584.3808 5.6682
DE 1596.9368 1600.0328 1603.5042 1.4791
Bold-italic values will give a more intuitive understanding of the advantages of the
proposed algorithm over other algorithms.

verified the capabilities and suitability of SDE. Comparison of the

final minimum cost results of two Mode indicates that when the
storage system is integrated into the VPP, the operation cost of the
VPP during 24 h cut down around 92 $. Table 6 lists the
comparison results of the investigated algorithms. In Table 5,
Min_Cost, Ave_Cost, Max_Cost, and Std . dev . are the minimum,
Fig. 11  Detail of SDE shares the energy production for case 3 average, maximum operation costs during the optimisation, and
(a) Mode1, (b) Mode2 standard deviation, respectively. It can be observed that the SDE
offers the best result in terms of all the indicators in Mode 1 over
50 independent runs, and in Mode 2 still achieved it except in terms
of Std . dev . However, it is the second best in terms of Std . dev .
These indicates that the SDE furnish the most optimum and
consistent solutions.

4.4 Simulation results for a classical test system of CHP ED

using SDE
In this case study, the proposed algorithm is implemented and
tested on a classic test system for ED problem, which is developed
in [32, 45]. In this system, Ncg = 26, Nchp = 12, Nho = 10, Nsto = 0,
Nres = 0, NEM = 0, and the corresponding objective function is
(23). Considering both valve-point effects of convention unit and
Fig. 12  Value of the estimated cost at each hour of the day provided by the feasible heat–power operation regions (heat–power dependency
optimal solution characteristics) of the CHP unit, which make the objective function
highly non-linear, non-convex, non-smooth. Then, the proposed
however, this is not an advantage in the MFA, since the minimum method will demonstrate its capability in solving the large-scale
operation cost provided by the MFA is much higher than the ED problem, handling nonlinear, non-convex, non-smooth
corresponding value of the SDE, which makes clear that the MFA characteristics related to the objective function and constraints and
has been prematurely trapped in a local minima. The phenomenon achieving an effective solution with regard to other previous
of being unable to withstand the confusion of the local minima also algorithms. The power and heat demands are both considered and
happened in DE, CBA, and PSOw. The ranking based on minimum equal to 4700 MW and 2500 MWth, respectively.
cost of the algorithms is SDE, CBA, MFA, PSOw, and DE. The Table 7 provides a comparison between SDE and the other
corresponding cost value are 1493.71, 1550.52, 1547.87, 1565.54, addressed methods for solving the problem. SDE achieved a more
and 1596.93 ($), respectively. We can find it very interesting that efficient solution than other methods. This table clearly shows the
the performance of CBA is better than MFA in this mode study, effectiveness and good searching capability of the proposed
which reveals that the ability of the algorithm may be limited as algorithm to provide the best solution than the other recent
conditions change. Fortunately, the proposed algorithm has algorithms. Fig. 14 shows the convergence curve of SDE, which
achieved excellent performance under all test conditions, which

3212 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215
© The Institution of Engineering and Technology 2019
Table 7 Comparison of performance for case 4.4
Algorithm Min_Cost Ave_Cost Max_Cost
CPSO [32] 119,708.8818 NA NA
TVAC-PSO [32] 117,824.8956 NA NA
TLBO [45] 116,739.3640 116,756.0057 116,825.8223
OBTLBO [45] 116,579.2390 116,613.6505 116,649.4473
GSO [46] 116,452.3150 116,463.6522 116,476.7524
OGSO [46] 116,403.3311 116,412.6214 116,423.9803
GSA [47] 117,266.6810 NA NA
TVAC-GSA-PSO [48] 116,393.4034 116,398.2042 116,404.6097
BA [49] 118,620.8913 118,644.5330 118,713.3496
ABC [49] 117,980.2510 118,017.6625 118,053.4593
BA-ABC [49] 116,250.0176 116,265.3372 116,279.4550
CSA-BA-ABC [49] 115,770.3910 115,781.6813 115,791.0402
IGA-NCM [50] 115,648.364 115,743.7491 115,793.1151
SDE 115,645.5875 115,690.4521 115,725.3845
Bold-italic values will give a more intuitive understanding of the advantages of the proposed algorithm over other algorithms.

Fig. 14  Convergence curve of SDE for case 4.4

Table 8 Details of the best solution provided by SDE for case 4.4
Unit Output Unit Output Unit Output Unit Output
P1 538.5904 P16 299.5442 P31 10.0452 H34 75
P2 299.2634 P17 159.8080 P32 35.0827 H35 104.9783
P3 299.3728 P18 110.2584 P33 81.0999 H36 75.0503
P4 109.8997 P19 109.9852 P34 40.0065 H37 40.0183
P5 159.8322 P20 109.9476 P35 81.0056 H38 20.1066
P6 109.8586 P21 109.8581 P36 40.0800 H39 469.3142
P7 159.7535 P22 109.8466 P37 10.0277 H40 60
P8 109.8402 P23 40.2248 P38 35.2649 H41 59.9985
P9 109.8905 P24 40.1918 H27 104.9282 H42 120
P10 40.1179 P25 55.1540 H28 75.0176 H43 120
P11 40.2550 P26 55.0731 H29 104.9652 H44 470.6422
P12 55.1391 P27 81.1576 H30 75.0735 H45 60
P13 55.0753 P28 40.0143 H31 40.0284 H46 60
P14 538.7613 P29 81.2376 H32 20.0511 H47 119.9714
P15 299.3844 P30 40.0519 H33 104.856 H48 120
total power = 4700 (MW) total heat = 2500 (MWth)

demonstrated the ability of SDE to fast converge. The detailed objective function, four test functions (F1, F4, F5, and F6) in section
optimisation results provided by the SDE are listed in Table 8. 4.1 are selected based on their mathematical characteristics, where
F1 and F5 are the unimodal problems and F4 and F6 are multimodal
5 Effect of algorithmic design strategies on the problems.
overall performance An initial population of each function was generated using
uniform and sequence-based solutions, which is depicted in
In this section, we analyse the effect of two algorithmic design Fig. 15. For the proposed initialisation method, n was set to a value
strategies on the overall performance of the proposed method, of 10, which means that the population size was 3311 for all
namely the initialisation technique, and mutation strategies. function. Note that the objective function values were plotted based
on the sequences in which the individuals were generated by the
5.1 Effect of initialisation technique initialisation methods. It can be clearly seen from the figure that
the method we adopted was able to obtain much better results for
To demonstrate that the proposed initialisation technique is able to all problems.
provide useful information about the behavioural pattern of the

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215 3213
© The Institution of Engineering and Technology 2019
Fig. 15  Fitness values of the initial population using different initialisation methods for F1, F4, and F6 with 30D, in each subfigure, are with n = 10
(a) F1, (b) F4, (c) F5, (d) F6

Table 9 Detail of best solution provided by SDE for case 4.4

Function Optimal Sequence-based Uniform
F1 −8.0000 × 1001 −8.0000 × 1001 1.6616 × 1004
F4 −1.2569 × 1004 −1.0955 × 1004 −0.3796 × 1004
F5 0.0000 2.9000 × 1001 1.2537 × 1008
F6 0.0000 0.0000 4.7899 × 1004

converge and finally provided unsatisfactory results. It is gratifying

that the dilemma situation has been greatly improved in the DE
with the combined mutation operator, which provides extremely
smooth convergence plot and shows that it is able to quickly escape
from the current poor position and obtain satisfactory results at the
end of evolutionary process. In conclusion, the proposed combined
mutation operator is able to greatly enhance the performance of
algorithm.

6 Conclusions
One of the most challenging tasks of a VPP is to participate in the
Fig. 16  Effect of combined mutation in DE for solving ED problem (x-axis market of the smart grid is the optimal scheduling of power
in log scale) generation among different resources. This objective of this task is
to allow for fast changing in power generation of RES to maintain
To compare the quality of the initial solutions obtained by the power balance while minimising the cost of VPP operation. A
different methods, Table 9 listed the average fitness values of the perfect solution to this ambitious objective needs a robust and well-
best fitness values in the initial population for 30 runs. From the organised optimiser, which can effectively cope with the dynamic
table, it is clear that the quality of initial population obtained by our changes of the network. The work done in this paper is summarised
method much better than those obtained from the uniform method. as below
In conclusion, the initial technique we adopted in this paper indeed
does provide a high-quality initial population. (i) A new metaheuristic algorithm SDE is proposed for minimising
the cost of VPPs operation by integrating sequence-based
5.2 Effect of mutation strategy deterministic initialisation and parameters self-adaptive mechanism
into DE.
In this subsection, we analyse the performance of the first mutation (ii) A new equality constraint handling mechanism designed to bias
operator, the second mutation operator, and combined mutation the optimisation towards the feasible region and thus enhances
operator used in (29) with DE. A sample of 1000 iterations with convergence rate.
self-adaptive algorithm parameters mechanism for all variants and (iii) We have illustrated the effectiveness of the proposed method
the convergence characteristics for the optimal objective values of with eight typical benchmark functions.
each variant to solve the non-linear and non-convex constrained (iv) We have successfully applied SDE to deal with dynamic ED
optimisation problem in subsection 4.4 are shown in Fig. 16. problem of a microgrid with the aim of generation cost
Although the three algorithms variants have a fast convergence minimisation, and demonstrated the effectiveness of using SDE to
rate in the early evolutionary process, DE with the first mutation solve dynamic ED problem of VPPs participate in the market.
operation quickly falls into the dilemma of requiring multiple (v) We have successfully demonstrated that the suggested
iterations to escape from local optima due to lack of population algorithm has superior performance, to solve the large-scale, highly
diversity and able to obtain poor solutions at the end of iterations. nonlinear, non-convex, and high-complex ED problem in the smart
This dilemma situation is not so serious in DE with the second grid.
mutation operator because its mechanism is beneficial to
population diversity, and this is why we choose it for enriching The analysis of simulation results demonstrates that the proposed
population diversity; however, needs to iterate many times to methods exhibit superior performances in solving dynamic ED

3214 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215
© The Institution of Engineering and Technology 2019
problem, compared to previously proposed state-of-the-art methods [24] Neri, F., Tirronen, V.: ‘Recent advances in differential evolution: a survey and
experimental analysis’, Artif. Intell. Rev., 2010, 33, (1-2), pp. 61–106
(MFA, CBA, PSOw, DE). Facing the foreseeable trend of smart [25] Das, S., Suganthan, P.N.: ‘Differential evolution: A survey of the state-of-the-
grid operation towards fast automatic control, the performance of art’, IEEE Trans. Evol. Comput., 2011, 15, (1), pp. 4–31
the proposed method proves that it is more qualified than many [26] Jing, Q., Brinsmead, T.S., Reedman, L.: ‘Optimal scheduling of distributed
traditional solutions to become a priority. In future work, we intend energy resources as a virtual power plant in a transactive energy framework’,
IET Gener. Transm. Distrib., 2017, 11, (13), pp. 3417–3427
to apply the proposed algorithm to solve other optimisation [27] Parisio, A., Glielmo, L.: ‘A mixed integer linear formulation for microgrid
problems of smart grid, in which, multiobjective optimisation economic scheduling’. 2011 IEEE Int. Conf. on Smart Grid Communications
problems is an attractive future direction. (SmartGridComm), Brussels, Belgium, 2011, pp. 505–510
[28] Askarzadeh, A.: ‘A memory-based genetic algorithm for optimization of
power generation in a microgrid’, IEEE Trans. Sustain. Energy, 2018, 9, (3),
7 Acknowledgment pp. 1081–1089
[29] Mehdinejad, M., Mohammadi-Ivatloo, B., Dadashzadeh-Bonab, R.: ‘Energy
This work was supported in part by the National Natural Science production cost minimization in a combined heat and power generation
Foundation of China (Grant No. 51767001). systems using cuckoo optimization algorithm’, Energy Effic., 2016, 10, (1),
pp. 1–16
[30] Subbaraj, P., Rengaraj, R., Salivahanan, S.: ‘Enhancement of combined heat
8 References and power economic dispatch using self adaptive real-coded genetic
algorithm’, Appl. Energy, 2009, 86, (6), pp. 915–921
[1] Ott, A.L.: ‘Evolution of computing requirements in the PJM market: past and [31] Khorram, E., Jaberipour, M.: ‘Harmony search algorithm for solving
future’. IEEE PES General Meeting, Providence, RI, USA, July 2010, pp. 1–4 combined heat and power economic dispatch problems’, Energy Convers.
[2] Ott, A.L.: ‘Development of smart dispatch tools in the PJM market’. 2011 Manag., 2011, 52, (2), pp. 1550–1554
IEEE Power and Energy Society General Meeting, San Diego, CA, USA, July [32] Mohammadi-Ivatloo, B., Moradi-Dalvand, M., Rabiee, A.: ‘Combined heat
2011, pp. 1–3 and power economic dispatch problem solution using particle swarm
[3] Baringo, A., Baringo, L.: ‘A stochastic adaptive robust optimization approach optimization with time varying acceleration coefficients’, Electr. Power Syst.
for the offering strategy of a virtual power plant’, IEEE Trans. Power Syst., Res., 2013, 95, (1), pp. 9–18
2017, 32, (5), pp. 3492–3504 [33] Khazali, A., Kalantar, M.: ‘Spinning reserve quantification by a stochastic-
[4] European virtual fuel cell power plant management summary report [eb/ol], probabilistic scheme for smart power systems with high wind penetration’,
13 August 2012, available at: htpp:ec.europa.eu/energy/efficiency/ Energy Convers. Manag., 2015, 96, pp. 242–257
industry/doc/euvpp.pdf/ [34] Elsayed, S., Sarker, R., Coello Coello, C.A.: ‘Sequence-based deterministic
[5] Zakeri, B., Syri, S.: ‘Electrical energy storage systems: a comparative life initialization for evolutionary algorithms’, IEEE Trans. Cybernet., 2017, 47,
cycle cost analysis’, Renew. Sust. Energy Rev., 2015, 42, pp. 569–596 (9), pp. 2911–2923
[6] Sahraei-Ardakani, M., Blumsack, S., Kleit, A.: ‘Estimating zonal electricity [35] Elsayed, W.T., Hegazy, Y.G., El-bages, M.S., et al.: ‘Improved random drift
supply curves in transmission-constrained electricity markets’, Energy, 2015, particle swarm optimization with self-adaptive mechanism for solving the
80, pp. 10–19 power economic dispatch problem’, IEEE Trans. Ind. Inf., 2017, 13, (3), pp.
[7] Project interim report[r/ol], 20 February 2013, available at: http:// 1017–1026
www.web2energy.com/ [36] Elsayed, S., Hamza, N., Sarker, R.: ‘Testing united multi-operator
[8] Sarker, M.R., Dvorkin, Y., Ortega-Vazquez, M.A.: ‘Optimal participation of evolutionary algorithms-ii on single objective optimization problems’. 2016
an electric vehicle aggregator in day-ahead energy and reserve markets’, IEEE IEEE Congress on Evolutionary Computation (CEC), Vancouver, BC,
Trans. Power Syst., 2016, 31, (5), pp. 3506–3515 Canada, 2016, pp. 2966–2973
[9] Kardakos, E.G., Simoglou, C.K., Bakirtzis, A.G.: ‘Optimal offering strategy [37] Sarker, R.A., Elsayed, S.M., Ray, T.: ‘Differential evolution with dynamic
of a virtual power plant: a stochastic bi-level approach’, IEEE Trans. Smart parameters selection for optimization problems’, IEEE Trans. Evolutionary
Grid, 2016, 7, (2), pp. 794–806 Comput., 2014, 18, (5), pp. 689–707
[10] Rahimiyan, M., Baringo, L.: ‘Strategic bidding for a virtual power plant in the [38] Elsayed, S.M., Sarker, R.A., Essam, D.L., et al.: ‘Testing united multi-
day-ahead and real-time markets: A price-taker robust optimization operator evolutionary algorithms on the cec2014 real-parameter numerical
approach’, IEEE Trans. Power Syst., 2016, 31, (4), pp. 2676–2687 optimization’. 2014 IEEE Congress on Evolutionary Computation (CEC),
[11] Dabbagh, S.R., Sheikh-El-Eslami, M.K.: ‘Risk assessment of virtual power Beijing, People's Republic of China, 2014, pp. 1650–1657
plants offering in energy and reserve markets’, IEEE Trans. Power Syst., [39] Al Bahrani, L.T., Patra, J.C.: ‘Multi-gradient pso algorithm for optimization
2016, 31, (5), pp. 3572–3582 of multimodal, discontinuous and non-convex fuel cost function of thermal
[12] Koraki, D., Strunz, K.: ‘Wind and solar power integration in electricity generating units under various power constraints in smart power grid’,
markets and distribution networks through service-centric virtual power Energy, 2018, 147, pp. 1070–1091
plants’, IEEE Trans. Power Syst., 2018, 33, (1), pp. 473–485 [40] Adarsh, B.R., Raghunathan, T., Jayabarathi, T., et al.: ‘Economic dispatch
[13] Nguyen, H.T., Le, L.B., Wang, Z.: ‘A bidding strategy for virtual power plants using chaotic bat algorithm’, Energy, 2016, 96, pp. 666–675
with intraday demand response exchange market using stochastic [41] Yang, X.-S., Hosseini, S.S.S., Gandomi, A.H.: ‘Firefly algorithm for solving
programming’, IEEE Trans. Ind. Appl., 2018, 54, (4), pp. 3044–3055 non-convex economic dispatch problems with valve loading effect’, Appl. Soft
[14] Bagchi, A., Goel, L., Wang, P.: ‘Adequacy assessment of generating systems Comput., 2012, 12, (3), pp. 1180–1186
incorporating storage integrated virtual power plants’, IEEE Trans. Smart [42] Shareef, H., Ibrahim, A.A., Mutlag, A.H.: ‘Lightning search algorithm’, Appl.
Grid, 2019, 10, (3), pp. 3440–3451 Soft Comput., 2015, 36, pp. 315–333
[15] Nosratabadi, S.M., Hooshmand, R.-A., Gholipour, E.: ‘A comprehensive [43] National sustainable energy laboratory (NREL), available at: https://
review on microgrid and virtual power plant concepts employed for pfs.nrel.gov/main.html
distributed energy resources scheduling in power systems’, Renew. Sust. [44] Aloini, D., Crisostomi, E., Raugi, M., et al.: ‘Optimal power scheduling in a
Energy Rev., 2017, 67, pp. 341–363 virtual power plant’. 2011 2nd IEEE PES Int. Conf. and Exhibition on
[16] Zamani, A.G., Zakariazadeh, A., Jadid, S.: ‘Day-ahead resource scheduling of Innovative Smart Grid Technologies (ISGT Europe), Manchester, UK, 2011,
a renewable energy based virtual power plant’, Appl. Energy, 2016, 169, pp. pp. 1–7
324–340 [45] Roy, P.K., Paul, C., Sultana, S.: ‘Oppositional teaching learning based
[17] Peik-Herfeh, M., Seifi, H., Sheikh-El-Eslami, M.K.: ‘Decision making of a optimization approach for combined heat and power dispatch’, Int. J. Electr.
virtual power plant under uncertainties for bidding in a day-ahead market Power Energy Syst., 2014, 57, (5), pp. 392–403
using point estimate method’, Int. J. Electr. Power Energy Syst., 2013, 44, (1), [46] Basu, M.: ‘Combined heat and power economic dispatch using opposition-
pp. 88–98 based group search optimization’, Int. J. Electr. Power Energy Syst., 2015, 73,
[18] Pandžić, H., Morales, J.M., Conejo, A.J., et al.: ‘Offering model for a virtual pp. 819–829
power plant based on stochastic programming’, Appl. Energy, 2013, 105, pp. [47] Beigvand, S.D., Abdi, H., Scala, M.L.: ‘Combined heat and power economic
282–292 dispatch problem using gravitational search algorithm’, Electr. Power Syst.
[19] Sowa, T., Krengel, S., Koopmann, S., et al.: ‘Multi-criteria operation Res., 2016, 133, pp. 160–172
strategies of power-to-heat-systems in virtual power plants with a high [48] Beigvand, S.D., Abdi, H., Scala, M.L.: ‘Hybrid gravitational search
penetration of renewable energies’, Energy Proc., 2014, 46, pp. 237–245 algorithm-particle swarm optimization with time varying acceleration
[20] Tajeddini, M.A., Rahimi-Kian, A., Soroudi, A.: ‘Risk averse optimal coefficients for large scale chped problem’, Energy, 2017, 126, pp. 841–853
operation of a virtual power plant using two stage stochastic programming’, [49] Murugan, R., Mohan, M.R., Christober, A.R.C., et al.: ‘Hybridizing bat
Energy, 2014, 73, pp. 958–967 algorithm with artificial bee colony for combined heat and power economic
[21] Lan, Y., Guan, X., Wu, J.: ‘Rollout strategies for real-time multi-energy dispatch’, Appl. Soft Comput., 2018, 72, pp. 189–217
scheduling in microgrid with storage system’, IET Gener. Transm. Distrib., [50] Dexuan, Z., Li, X.K.S., Ouyang, H.: ‘Solving the combined heat and power
2016, 10, (3), pp. 688–696 economic dispatch problems by an improved genetic algorithm and a new
[22] Crisostomi, E., Liu, M., Raugi, M., et al.: ‘Plug-and-play distributed constraint handling strategy’, Appl. Energy, 2019, 237, (13), pp. 646–670
algorithms for optimized power generation in a microgrid’, IEEE Trans.
Smart Grid, 2014, 5, (4), pp. 2145–2154
[23] Storn, R., Price, K.: ‘Differential evolution–a simple and efficient heuristic for
global optimization over continuous spaces’, J. Global Opt., 1997, 11, (4), pp.
341–359

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 15, pp. 3202-3215 3215
© The Institution of Engineering and Technology 2019