You are on page 1of 14

Available online at www.sciencedirect.

com

ScienceDirect
Journal of Hydro-environment Research 10 (2016) 5063
www.elsevier.com/locate/jher

Research papers

Optimal operation of multi-reservoir hydropower systems using enhanced


comprehensive learning particle swarm optimization
Xueqing Zhang a,*, Xiang Yu a, Hui Qin b
a

Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong
b
School of Hydropower and Information Engineering, Huazhong University of Science and Technology, Wuhan, Hubei 430074, China
Received 21 November 2014; revised 25 April 2015; accepted 10 June 2015

Abstract
Metaheuristics are promising optimization algorithms for tackling reservoir-system operation. Comprehensive learning particle swarm optimization (CLPSO) is a state-of-the-art metaheuristic that is strong in exploration. Recently we have proposed enhanced CLPSO (ECLPSO) to
improve the exploitation performance of CLPSO. In this paper, we apply ECLPSO to the optimal operation of multi-reservoir hydropower systems.
Two novel strategies are proposed to handle the various physical and operational constraints. First, the outflow and storage volume constraints are
appropriately enforced to achieve a tradeoff between preserving diversity and facilitating convergence. Second, with the penalty function technique
adopted to penalize the constraint violations and convert the original constrained problem into an unconstrained one, the penalty factor is
dynamically adjusted in order to encourage exploration of the search space in the beginning and gradually guide the search to concentrate in the
feasible region. The short-term scheduling of a 4-reservoir hydrothermal power system and the long-term planning of Chinas Xiluodu
XiangjiabaThreegorges 3-reservoir hydropower system are studied. Experimental results demonstrate that ECLPSO helps to robustly derive
feasible high quality solutions for the two cases studied. The contribution to performance improvement by ECLPSO as well as the constraint
enforcement and penalty factor adjustment strategies are analyzed.
2016 International Association for Hydro-environment Engineering and Research, Asia Pacific Division. Published by Elsevier B.V. All rights
reserved.
Keywords: Optimal operation; Multi-reservoir hydropower system; Enhanced comprehensive learning particle swarm optimization

1. Introduction
Hydropower has become a critical source of electricity as it is
renewable, clean, and cheap. Most hydropower is generated from
plants constructed within reservoirs.A reservoir impounds water
to serve various purposes such as hydropower generation, flood
control, navigation, and/or water provisioning. A multi-reservoir
system consists of multiple reservoirs sited within the same river
basin. The cascaded reservoirs are hydraulically coupled as the
outflow of a reservoir constitutes part of the inflow into its
immediate downstream reservoir (Orero and Irving, 1998). The
optimal operation of the cascaded reservoirs is beneficial for the
overall development of the river basin.
The optimal operation of multi-reservoir hydropower
systems is challenging to solve because of the following

* Corresponding author. Department of Civil and Environmental


Engineering, The Hong Kong University of Science and Technology, Clear
Water Bay, Kowloon, Hong Kong. Tel.: +852 2358 8480; fax: +852 2358 1534.
E-mail address: zhangxq@ust.hk (X. Zhang).

factors: (1) the input (e.g. inflow) imprecision and uncertainties


need to be addressed (Simonovic, 1987); (2) the decisionmaking process is multi-stage and dynamic (Simonovic, 1987);
(3) the resulting optimization problem is often large-scale with
a lot of decision variables and constraints (Labadie, 2004); (4)
the hydraulic coupling among the reservoirs complicates the
problem (Orero and Irving, 1998); (5) because of the various
physical and operational constraints, it isnt easy to find a
feasible solution that satisfies all the constraints (Simonovic,
1987); (6) the hydropower performance model is usually nonlinear (El-Hawary and Christensen, 1979), could be nonconvex
(Tauxe et al., 1980), discontinuous and non-differentiable (Lyra
and Ferreira, 1995), and even mixed-integer (Wang and Zhang,
2012); and (7) sometimes multiple conflicting objectives are
considered (Lyra and Ferreira, 1995; Tauxe et al., 1980). The
input imprecision and uncertainties can be addressed implicitly
through forecast data, long/representative historical records,
and/or synthetically generated data (Labadie, 2004). The multiple objectives can be transformed into a single objective
through using techniques such as the weighting and -constraint
(Miettinen, 1999).

http://dx.doi.org/10.1016/j.jher.2015.06.003
1570-6443/ 2016 International Association for Hydro-environment Engineering and Research, Asia Pacific Division. Published by Elsevier B.V. All rights
reserved.

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

Various optimization algorithms have been applied to solve


the optimization problems related to the operation of
multi-reservoir hydropower systems. The algorithms are usually
classified into three general categories: traditional optimizers,
modern metaheuristics, and hybrid approaches. Traditional
optimizers, including linear programming, nonlinear programming,
dynamic programming, and optimal control theory, have rigorous
mathematical foundations (Labadie, 2004). A metaheuristic is
essentially a high level intelligent strategy to guide the search
toward the promising region. A lot of different metaheuristics
have been proposed, including simulated annealing, tabu search,
genetic algorithm, differential evolution, and particle swarm
optimization, just to name a few (Boussad et al., 2013). The
metaheuristics are inspired by different nature principles from
biology, ethology, or physics. Compared with traditional
optimizers, modern metaheuristics are significantly more flexible
as they dont require the objective and constraints to be
continuous, differentiable, linear, or convex, and they usually
can efficiently solve large-scale problems. Although a
metaheuristic might not find the exact optimum, it often can
find a near-optimal solution. In addition, metaheuristics can be
directly linked with simulation models (e.g. flood inundation
and water quality) without requiring simplifying any assumptions
in the models. Hybrid approaches try to combine the strengths
of different optimization algorithms. For example, Li et al.
(2012) applied a hybrid approach to the optimal operation of
Chinas Three Gorges cascaded hydropower system; the approach
first exploits incremental dynamic programming to narrow
down the search space, and then takes advantage of genetic
algorithm to solve the nonlinear and nonconvex problem.
Introduced in 1995 (Eberhart and Kennedy, 1995; Kennedy
and Eberhart, 1995), particle swarm optimization (PSO) is a
swarm-intelligence inspired metaheuristic simulating the movements of organisms in a bird flock or fish school. PSO solves an
optimization problem using a swarm of particles, with each
particle representing a candidate solution. Metaheuristics
(including PSO) need to achieve a balance between exploration
and exploitation. Exploration is the ability to search different
regions for locating a good solution, while exploitation is the
ability to concentrate the search around a small region for
refining a hopeful solution (Boussad et al., 2013). Comprehensive learning PSO (CLPSO) (Liang et al., 2006) is state-of-theart PSO variant that is strong in exploration but weak in
exploitation. We have recently proposed enhanced CLPSO
(ECLPSO) (Yu and Zhang, 2014) to improve the exploitation
performance of CLPSO. In this paper, we apply ECLPSO to the
optimal operation of multi-reservoir hydropower systems. Two
novel strategies are proposed to handle the various physical and
operational constraints. First, the outflow and storage volume
constraints are appropriately enforced to achieve a tradeoff
between preserving diversity and facilitating convergence.
Second, with the penalty function technique adopted to penalize
the constraint violations and convert the original constrained
problem into an unconstrained one, the penalty factor is
dynamically adjusted in order to encourage exploration of the
search space in the beginning and gradually guide the search to
concentrate in the feasible region. The short-term scheduling

51

of a 4-reservoir hydrothermal power system and the longterm planning of Chinas XiluoduXiangjiabaThreegorges
3-reservoir hydropower system are studied.
The rest of this paper is organized as follows. In
Section 2, a generalized problem formulation is presented.
Section 3 reviews the working principle of ECLPSO. Related
works on constraint handling mechanisms and the application
of PSO to optimal reservoir-system operation are discussed
in Section 4. Section 5 details the application implementation
of ECLPSO. In Section 6, the performance of ECLPSO
is evaluated on the selected cases. Section 7 concludes the
paper.
2. Generalized problem formulation
For the optimal operation of multi-reservoir hydropower
systems, the short/medium/long-term optimization problems
(Chen et al., 2013; Ji et al., 2014; Li et al., 2012, 2014a, 2014b;
Orero and Irving, 1998) have a similar generalized formulation.
A single-reservoir system could be regarded as the special case
of a multi-reservoir system when the number of reservoirs
degenerates to one. Before putting the generalized problem
formulation into perspective, decision variables, state variables,
and model parameters are defined as follows.
Decision variables:
Qi,t Power discharge rate of reservoir i in time step t
Si,t Spillage rate of reservoir i in time step t
State variables:
Vi,t Storage volume of reservoir i at the beginning of time step
t
Model parameters:
Ei,t Evaporation and seepage loss from the storage of reservoir
i in time step t
i, j Reservoir indices, starting from 1
Ii,t Natural inflow into reservoir i in time step t
N Number of reservoirs
Oi,t Outflow rate of reservoir i in time step t
Oi ,t Maximum outflow rate of reservoir i in time step t
Oi ,t Minimum outflow rate of reservoir i in time step t
Pi,t Power output of reservoir i in time step t
Pi ,t Maximum power output of reservoir i in time step t
Pi ,t Minimum power output of reservoir i in tim e step t
Qi ,t Maximum power discharge rate of reservoir i in time
step t
Ri,t Precipitation onto the surface of reservoir i in time step t
Si Maximum spillage rate of reservoir i
t Time step index, starting from 1
T Number of time steps, i.e. the planning horizon
Vi begin Initial storage volume limit of reservoir i
Vi ,t Maximum storage volume of reservoir i at the beginning
of time step t
Vi ,t Minimum storage volume of reservoir i at the beginning
of time step t
i The set of immediate upstream reservoirs with respect to
reservoir i

52

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

j,i Water transportation delay from reservoir j to its immediate downstream reservoir i
t Length of time step t
The problem essentially aims to schedule the power discharge and spillage rates of all the reservoirs in all the time
steps over the planning horizon. The optimization objective f is
usually expressed as a dynamic function, i.e.
N

Max (or Min) f =

i=1

t =1

i (Vi ,t , Qi ,t , Si ,t ) + i (Vi ,T +1 )

(1)

where i is lower level objective function (or numerical


simulation procedure); and i is the final term representing
the future estimated benefit (or cost) beyond the planning
horizon. The optimization objective could be different for
different reservoir systems and/or under different operation
contexts.
The problem basically is subjected to the following physical
and operational constraints.
(a) Outflow constraints:

0 Qi ,t Qi ,t , i = 1, 2, , N , t = 1, 2, , T

(2)

0 Si ,t Si , i = 1, 2, , N , t = 1, 2, , T

(3)

Oi ,t = Qi ,t + Si ,t , i = 1, 2, , N , t = 1, 2, , T

(4)

Oi ,t Oi ,t Oi ,t , i = 1, 2, , N , t = 1, 2, , T

(5)

The limits on the outflow rate Oi,t maintain desired streamflow


for flood protection, navigation, water demand satisfaction,
water quality control, and/or fish and wildlife maintenance.
Additional constraints may be imposed on the variation of the
outflow rates between consecutive time steps to facilitate
navigation (Wang and Zhang, 2012) and/or avoid the scouring
of the river banks and channels (Needham et al., 2000).
(b) Power output constraints:

Pi ,t Pi ,t Pi ,t , i = 1, 2, , N , t = 1, 2, , T

(6)

(O

Vi ,1 = Vi begin , i = 1, 2, , N

(7)

Vi ,t Vi ,t Vi ,t , i = 1, 2, , N , t = 2, 3, , T + 1

(8)

The explicit lower and upper bounds on the storage volume


Vi,t are assigned based on the dead storage limit and to
accommodate the needs for flood control, navigation, and
recreation. Note that Vi ,T +1 Vi ,T +1 Vi ,T +1 is the constraint on
the storage volume of reservoir i at the end of the planning
horizon. Often the final reservoir storage volume Vi,T+1 is
restricted to be equivalent to a fixed value (Orero and Irving,
1998) and in this situation Vi ,T +1 = Vi ,T +1 . Some works (Wang
and Zhang, 2012; Xiong, 1990) allowed an inequality
constraint on Vi,T+1.

j ,t j ,i

t j ,i )

ji

(9)

Ei ,t Oi ,t t , i = 1, 2, , N , t = 1, 2, , T
Here

ji

(O j ,t

j ,i

t j ,i ) determines the total immedi-

ate upstream inflow into reservoir i in time step t. For a more


accurate formulation, the diversion(s) of water from the reservoirs (Kumar and Reddy, 2007; Westphal et al., 2003) and the
withdrawal(s) of water along the river reaches may be considered; in such cases the diversion(s) and/or withdrawal(s) are
also treated as decision variables. In short-term operation, if
there is a long distance interconnection between a reservoir j
and its immediate downstream reservoir i, the water transportation delay j,i is often greater than the time steps length t
(Orero and Irving, 1998).
In operation problems considering water supply, constraints
on the water diversion(s) and/or withdrawal(s) are included
(Kumar and Reddy, 2007; Westphal et al., 2003). Some works
studied the coordination of interconnected hydrothermal power
systems (Amjady and Soleymanpour, 2010; Kumar and Naresh,
2007; Orero and Irving, 1998; Sinha et al., 2003; Yu et al.,
2007; Yuan et al., 2008). A hydrothermal power system has
both hydro and thermal installed capacities. The optimal operation of a hydrothermal power system thus inevitably needs to
take into account the constraints imposed on the thermal part.
For works that address the individual operation of each turbinegenerator, the turbine-generator related constraints such as prohibited zones, minimum up/down time, maximum startup
number, ramping capacity, and/or multi-vibration zones must
be considered (Cheng et al., 2012; Li et al., 2014a, 2014b;
Wang and Zhang, 2012). In case the input is a long collection of
naturalized historical records or synthetically generated data, a
constraint could incorporate a probability term and thus be
reliability-based in order to allow some acceptable degree of
violation (Ji et al., 2014; Wu et al., 2008).
The power output Pi,t is a nonlinear function of the power
discharge rate and net head (Ji et al., 2014; Li et al., 2014a,
2014b; Xiong, 1990), i.e.

Pi ,t = K i Qi ,t H i ,t

(c) Storage volume constraints:

(d) Water balance constraints:

Vi ,t +1 = Vi ,t + I i ,t + Ri ,t +

(10)

where Ki is the synthetic power output coefficient of reservoir i;


and Hi,t is the average net head of reservoir i in time step t. The
amount of power generated by reservoir i in time step t is given
by Pi ,t t .
The gross head is the difference between the forebay and
tailrace elevations. The net head is obtained by subtracting the
head loss from the gross head (Xiong, 1990), i.e.

H i ,t = Lforebay
Ltailrace
H iloss
i ,t
,t
i ,t
forebay
i ,t

(11)

is the average forebay elevation of reservoir i in


where L
time step t, and it is a function is geometry and average storage
is the average tailrace elevation of reservoir i in
volume; Ltailrace
i ,t
time step t, and it is dependent on the outflow rate of i and the
average forebay elevation of is immediate downstream
reservoir (if the two reservoirs are close in distance); and H iloss
,t
is the average head loss of reservoir i in time step t.

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

Orero and Irving (1998) fitted the power output Pi,t by a


quadratic function of the power discharge rate and storage
volume. If the operation of each individual turbine-generator is
addressed, then we need to model the relationship among the
turbine-generators power output, discharge rate, and water
head (Li et al., 2013, 2014a, 2014b; Wang and Zhang, 2012).
The maximum power discharge rate Qi ,t and the maximum
power output Pi ,t are both dependent on is water head.
3. Enhanced comprehensive learning particle swarm
optimization
In PSO, all the particles fly in the search space. Let there
be D decision variables, the swarm of particles move in
D-dimensional space. Each particle, denoted as p, is associated
with a position Posp = (Posp,1, Posp,2, , Posp,D), a flight velocity
Velp = (Velp 1, Velp,2, , Velp,D), and a fitness that indicates its
performance. PSO relies on iterative learning to find the
optimum. In each iteration, p adjusts its velocity according to
its previous velocity, its historical best position (i.e. personal
best position), and the personal best positions of its neighborhood particles.
Comprehensive learning PSO (CLPSO) (Liang et al., 2006)
encourages a particle to learn from different exemplars on different dimensions. In each generation, Velp and Posp are updated
as follows.

Vel p ,d = w Vel p ,d + c rd ( Exe p ,d Pos p ,d )

(12)

Pos p ,d = Pos p ,d + Vel p ,d

(13)

where d (1 d D) is the dimension index; w is the inertia


weight; Exep = (Exep,1, Exep,2, , Exep,D) is the guidance vector
of exemplars; c is the acceleration coefficient and usually
c = 1.5; and rd is a random number in [0, 1].
The dimensional velocity Velp,d is usually clamped to a positive value Veld . If Vel p ,d > Veld , then Velp,d is set as Veld ; or if
Vel p ,d < Veld , then Velp,d is set as Veld . Veld is an important
parameter that determines the resolution of the flying space
centered at the present particle position. If Veld is too large, the
particles might fly past good solutions; on the other hand, if
Veld is too small, the particles might not be able to explore
sufficiently the search space. Let Posd and Posd respectively
be the lower and upper bounds of the search space on the dth
dimension, Veld is usually set as an appropriate ratio of
Posd Posd .
The personal best position of particle p is denoted as
Pbestp = (Pbestp,1, Pbestp,2, , Pbestp,D). After Posp is updated,
Posp is evaluated and will replace Pbestp if it has a better fitness
value. When the algorithm terminates, the output is the best
solution out of all the particles personal best positions.
The inertia weight w linearly decreases to balance exploration and exploitation. Specifically, let k be the predefined
maximum number of generations, in each generation k, w is
updated according to Eq. (14).
k
(14)
w = w (w w )
k
where w and w are respectively the maximum and minimum
inertia weights. Usually w = 0.9 and w = 0.4.

53

The exemplar Exep,d can be Pbestp,d or Pbestq,d with q p.


The decision to learn whether from Pbestp,d or Pbestq,d depends
on a learning probability Lprob p . On dimension d, a
random number is generated. If the number is no less than
Lprobp, dimension d will learn from Pbestp,d; otherwise from
Pbestq,d. q is selected from a tournament procedure; first
two particles are randomly selected excluding p; then the
particles personal best fitness values are compared; finally, the
winner is q. If all the exemplars of a particle are its own
personal best position, CLPSO will randomly choose one
dimension to learn from some other particles personal best
position.
The learning probabilities are set such that each particle has
a different learning probability, thus the particles exhibit different levels of exploration and exploitation capabilities. CLPSO
empirically determines the Lprobp value for each particle p
according to Eq. (15).

10 ( p 1)
exp
1
Npar 1
(15)
Lprobp = Lprob + ( Lprob Lprob)
exp (10) 1
where Lprob and Lprob are respectively the maximum and
minimum learning probabilities, usually Lprob = 0.5 and
Lprob = 0.05 ; and Npar is the number of particles.
To ensure that a particle learns from good exemplars and to
minimize time wasted on poor directions, CLPSO allows the
particle to learn from the same exemplars until its fitness
value ceases improving for a certain consecutive number of
generations called refreshing gap g, then the exemplars are
re-determined. Usually g = 7.
The experimental results reported in Zhan et al. (2011)
demonstrate that CLPSO works excellently on many
complex multi-modal problems as it is able to preserve the
particles diversity and is strong in exploration.
However, CLPSO performs worse than other PSO variants
on unimodal and simple multi-modal problems because it is
poorer in exploitation and significantly lower in solution
accuracy.
Recently in Yu and Zhang (2014), we have proposed
enhanced CLPSO (ECLPSO). ECLPSO introduces two
enhancements, namely, perturbation based exploitation (PbE)
and adaptive learning probabilities (ALPs), to improve the
exploitation performance of CLPSO.
In each generation, the PbE enhancement determines
the normative knowledge which is the dimensional
lower/upper bounds relevant to all the personal best positions.
In other words, on each dimension d, Pbest d =

min { Pbest1,d , Pbest2,d , , Pbest Npar ,d }


and
Pbest d =
max { Pbest1,d , Pbest2,d , , Pbest Npar ,d } are calculated. If Eq.
(16) is true, the PbE enhancement updates Velp,d for each particle p according to Eq. (17) instead of following Eq. (12).
Pbestd Pbestd ( Posd Posd )
and

Pbestd Pbestd

(16)

54

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

Vel p ,d = wPbE Vel p ,d + c rd Exe p ,d

Pbestd + Pbestd

+
Exe p ,d Pos p ,d

(17)

where is the relative ratio; is the absolute bound; wPbE is the


inertia weight used exclusively for exploitation; and is the
perturbation coefficient. Empirical values chosen for and
are 0.01 and 2 respectively. The combined use of the relative
ratio and absolute bound ensures that the normative interval
Pbestd , Pbestd is indeed small enough. A small normative

interval usually indicates that the particles have reached an


equilibrium state and the interval hopefully contains the global
optimum on the corresponding dimension. Empirically, wPbE is
fixed at 0.5, and is randomly generated from a normal
distribution with mean 1 and standard deviation 0.65. Note that
that is clamped to 10 times of its standard deviation on both
sides of its mean for the purpose of preventing the use of too
large values. The PbE enhancement contributes to sufficient
exploitation around the small normative interval.
A small Lprob benefits exploration, whereas a large Lprob
contributes to exploitation. In each generation k, let Actk be the
number of the dimensions whose normative intervals have ever
satisfied the exploitation activation condition specified in Eq.
(16) before or just in k, Lprob is empirically determined
according to Eq. (18).

Lprob = Lprob + 0.25 + 0.45 log( D+1) ( Act k + 1)

(18)

Lprob is fixed at 0.05. Lprob is small (i.e. 0.3) when Actk = 0


and thus helps to explore the search space incorporating the
concern of avoiding premature convergence. Lprob increases
rapidly with the particles exploitation progress to facilitate
convergence. The learning probabilities are determined based
on the rankings of the personal best fitness values, as expressed
in Eq. (19).
Lprobp = Lprob + ( Lprob Lprob)
10 ( Rank p 1)
exp
1

Npar 1

exp (10) 1

(19)

where Rankp is the rank of particle p. If p has the best personal


best fitness value, then its rank Rankp is 1. The underlying
rationale of Eq. (19) is that for any two particles p and q with
the personal best fitness value of p better than that of q, Pbestp
often is better on more dimensions than Pbestq. Therefore, when
assigning learning probabilities to p and q, Lprobp needs to be
less than Lprobq as p would benefit more from learning from its
own personal best position than q.
4. Related works
4.1. Constraint handling mechanisms
The most popular technique to handle constraints is the use
of penalty functions. The penalty function technique worsens

the fitness of infeasible solutions to favor the evolution of


feasible solutions. The constraint violations are incorporated
into the objective function so that the original constrained
problem is converted into an unconstrained problem. Although
being simple, a penalty function requires an appropriate selection of the penalty factor which is dependent on the specific
problem being solved. Joines and Houck (1994) dynamically
adjust the penalty factors. The penalty factors are treated as
decision variables and evolved in He and Wang (2007). Some
alternative constraint handling techniques have been proposed,
including treating the constraints violation explicitly as one
more optimization objective and solving the multi-objective
optimization problem through handling the multiple objectives
directly (Aguirre et al., 2004), the feasibility-based rules (Deb,
2000), as well as heuristics/operators to relieve constraint violations, preserve the feasibility of candidate solutions, or
explore the boundaries of the feasible and infeasible regions
(Leguizamn and Coello Coello, 2007).
Afshar (2012) proposed heuristics to explicitly enforce the
outflow and storage volume constraints for a single reservoir.
The heuristics use the water balance equation to define a new
set of bounds for the decision variables. The heuristics were
further extended to the optimal operation of multi-reservoir
systems in Afshar (2013).
4.2. Application of PSO to optimal reservoir-system
operation
PSO has many variants and the variants differ in how to use
the neighborhood search experience for the particle velocity
update. The most commonly used PSO variant is global PSO
(GPSO) (Shi and Eberhart, 1998). In GPSO, the historical best
position out of the entire swarm (i.e. global best position) is
used to guide a particles flight on all the dimensions. GPSO is
the easiest to implement and converges fast, but it is liable to get
stuck in premature convergence. GPSO was applied to optimal
reservoir-system operation in Afshar (2012, 2013). Kumar and
Reddy (2007) proposed EMPSO through augmenting GPSO
with an elitist-mutation strategy. In EMPSO, the worst particles positions are replaced by mutations of the global best
position. EMPSO was applied to the Bhadra reservoir in India.
Zhang et al. (2013) proposed multi-elite guide PSO (MGPSO)
for the optimal operation of multi-reservoir systems. In
MGPSO, an external repository is used to store elites obtained
during the run; the personal and global best positions used in
GPSO are replaced by two different elitists randomly selected
from the external repository. Improved adaptive PSO (IAPSO)
was proposed in Zhang et al. (2014) and applied to the optimal
operation of the Three Gorges Dam and Xiluodu Dam in China.
IAPSO is the same as GPSO except that it adaptively adjusts the
inertia weight based on the convergence speed and the swarms
diversity. Yuan et al. (2008), Amjady and Soleymanpour
(2010), and Yu et al. (2007) applied PSO to the short-term
operation of hydrothermal power systems. In Amjady and
Soleymanpour (2010) and Yuan et al. (2008), a particle learns
not only from the personal and global best positions, but also
from the historical worst position out of the entire swarm. A
local PSO (LPSO) (Kennedy and Mendes, 2002) was used in

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

Yu et al. (2007). In a LPSO, a static social topology (e.g. ring


and pyramid) is constructed and the historical best position out
of a particles immediately connected neighborhood (i.e. local
best position) is used for the particles velocity update.
5. Application implementation of enhanced
comprehensive learning particle swarm optimization
5.1. Representation and initialization of each particle
For each particle p, its position Posp, as shown in Eq. (20), is
a N T matrix representing candidate reservoir outflow rates.
Each dimension of the position is randomly initialized from
Oi ,t , Oi ,t .

O1,1, O1,2, , O1,T

O2,1, O2,2, , O2,T

Pos p =

O , O , , O
N ,T
N ,1 N ,2

(20)

The power discharge rate Qi,t and the spillage rate Si,t are
determined from the outflow rate Oi,t as follows. If Oi ,t Qi ,t ,
then Qi,t = Oi,t and Si,t = 0; otherwise, Qi ,t = Qi ,t and Si,t = Oi,t Qi,t.
5.2. Enforcement of the outflow and storage volume
constraints
The dimensional position Posp,d updated using Eq. (13) is
likely to be outside the feasible range Oi ,t , Oi ,t . Clamping

Posp,d to the nearest bound of Oi ,t , Oi ,t or randomly

reinitializing Posp,d within Oi ,t , Oi ,t both might negatively


impact the solution quality. Clamping might cause the search to
lose diversity, while reinitialization might lead the search to
some unfavorable region. In CLPSO (Liang et al., 2006) and
ECLPSO (Yu and Zhang, 2014), Posp,d is allowed to be outside
the feasible range. If Posp,d is far from Oi ,t , Oi ,t , then its return
to the feasible range would be slow and the exploration process
would possibly be seriously affected. Accordingly, Posp,d is
allowed to be outside Oi ,t , Oi ,t but isnt allowed to be far from

Oi ,t , Oi ,t . Here Posp,d is said to be far from Oi ,t , Oi ,t if Eq. (21)

is true.

Vi ,new
t

new
i ,t

Vi ,t ,

=
max Vi ,new
I i ,t Ri ,t

t +1

Vi ,t ,

=
min Vi ,new
I i ,t Ri ,t

t +1

Pos p ,d < Oi ,t min { (Oi ,t Oi ,t ), }


or

Pos p ,d > Oi ,t + min { (Oi ,t Oi ,t ), }

j ,i

ji

for Pos p ,d > Oi ,t + min { (Oi ,t Oi ,t ), }, Posp,d is randomly

reinitialized within Oi ,t , Oi ,t + min { (Oi ,t Oi ,t ) , }. As we


can see, after the reinitialization, Posp,d is still infeasible but
close to the feasible range. In addition, during the PbE and
when enforcing the storage volume constraints, a Posp,d outside
Oi ,t , Oi ,t is appropriately reinitialized/clamped.

As can be seen from Eq. (17) and Eq. (13), the perturbation
might lead Posp,d to jump out of Oi ,t , Oi ,t . The global optimum
is highly likely to be located around the normative interval
Pbestd , Pbestd . To avoid wasting time searching outside the

feasible range Oi ,t , Oi ,t , the dimensional velocity Velp,d updated


using Eq. (17) is further modified as follows.

Vel p,d

Oi ,t + rPbE ( Exe p,d Oi ,t ) Pos p,d


if Pos p,d + Vel p,d < Oi ,t

= Exe p,d + rPbE (Oi ,t Exe p,d ) Pos p,d , if Pos p,d + Vel p,d > Oi ,t

Vel p,d
otherwise

(22)

where rPbE is a random number in [0, 1].


The storage volume constraints are explicitly enforced using
a heuristic introduced in Afshar (2013). The reservoir indices
are ordered through the breadth first search sequence of their
hydraulic coupling relationships; i.e. an upstream reservoir has
a smaller index than its immediate downstream reservoir. Incrementally starting from reservoir 1 and ending at N, the storage
volume constraints for each reservoir i are sequentially
enforced as follows.
Step (1) Reversely starting from time step T+1 and
ending at 2, sequentially calculate the new bounds for is
storage volume in time step t according to Eq. (23) and
Eq. (24).

ji

j ,i

t = T +1

t j ,i ) + Ei ,t + Oi ,t t , Vi ,t , otherwise

if

(O j ,t

(21)

Remember that and have already been defined in


the
PbE
enhancement
of
ECLPSO.
For
Pos p ,d < Oi ,t min { (Oi ,t Oi ,t ), } , Posp,d is randomly
reinitialized within Oi ,t min { (Oi ,t Oi ,t ), } , Oi ,t ; while

if

(O j ,t

55

(23)

t = T +1

t j ,i ) + Ei ,t + Oi ,t t , Vi ,t , otherwise

(24)

56

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

Step (2) Incrementally starting from time step 1 and ending at


T, sequentially calculate the deviation of is storage volume
in time step t+1 according to Eq. (25), modify is outflow
rate in time step t according to Eq. (26) and update the
storage volume if the deviation isnt zero.

Vi ,t +1 Vi ,new
if Vi ,t +1 < Vi ,new
t +1,
t +1

new

Dev = Vi ,t +1 Vi ,t +1, elsse if Vi ,t +1 > Vi ,new


t +1

0,
otherwise

(25)

if Oi ,t + Dev < Oi ,t
Oi ,t ,

Oi ,t = Oi ,t ,
else if Oi ,t + Dev > Oi ,t , if Dev 0

Oi ,t + Dev, otherwise

(26)

Note that in Eq. (26), Oi,t is clamped to the nearest bound of


Oi ,t , Oi ,t if the deviation isnt zero and the outflow rate is

outside the feasible range.


To sum, infeasible outflow rates are appropriately treated in
order to achieve a tradeoff between preserving diversity and
facilitating convergence.
5.3. Adjustment of the penalty factor
The penalty function technique similar to the one introduced
in Amjady and Soleymanpour (2010) is adopted to handle all
the constraints. For each constraint as expressed in Eq. (5), the
violation is calculated according to Eq. (27).

1 Oi ,t , if
O +
i
v (Oi ,t , Oi , Oi ) =
Oi ,t
1, if

Oi +

Oi ,t < Oi
(27)

Oi ,t > Oi

where is a predefined, very small positive number almost


close to zero. is used only if the bound is zero and the bound
is likely to be violated. For the other constraints as expressed in
Eq. (2), Eq. (3), Eq. (6) and Eq. (8), the violations are calculated
respectively using v (Qi ,t , 0, Qi ,t ), v (Si ,t , 0, Si ), v ( Pi , Pi ,t , Pi ,t ),
and v (Vi ,t , Vi ,t , Vi ,t ). Accordingly, the total constraints violation
is calculated according to Eq. (28).
N

vtotal =

v (Q , 0, Q ) + v (S , 0, S )
i=1

i ,t

i ,t

i ,t

t =1

v (Vi ,t , Vi ,t , Vi ,t )

t =2
(28)
The original constrained problem is converted to an unconstrained problem optimizing the following objective:
+ v (Oi ,t , Oi ,t , Oi ,t ) + v ( Pi , Pi ,t , Pi ,t ) +

T +1

Fig. 1. Growth patterns of the two penalty factor adjustment strategies.

where is the penalty factor and is a positive number.


The penalty factor is set as a fixed value in Amjady and
Soleymanpour (2010). In this paper, we propose to adjust
dynamically. In each generation k, the penalty factor is
dynamically adjusted according to Eq. (30).

10(k 1)
exp
1
k 1
exp(10)1

(30)

where and are respectively the minimum and maximum


penalty factors. is far larger than . As can be seen in Eq.
(30), is affected by the generation counter k; initially is
small to facilitate the search of different regions and later
grows sufficiently large to guide the search to concentrate in the
feasible region.
Joines and Houck (1994) also proposed to adjust the penalty
factors dynamically, however in a linear manner. If the Joines
Houck strategy was used, then was dynamically adjusted in
each generation k as below.

= k

(31)

Fig.
1
plots
the
two
growth
terms
(exp(10(k 1) (k 1))1) (exp(10)1)
and
k
respectively
critical
to
( )
the two penalty factor adjustment strategies, with k =15000
and =1014 . It can be observed from Fig. 1 that the penalty
factor in our proposed strategy increases significantly less
slowly than that in the JoinesHouck strategy until the end of
the run and are considerably larger than that in the Joines
Houck strategy during the end of the run.

Max h = f vtotal, if the original objective f is a maximization objective

Min h = f + vtotal, otherwise

(29)

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

57

Table 1
ECLPSO variants with different configurations.
ECLPSO
variant

Description

1
2
3
4
5
6
7
8
9
10
11

ECLPSO
Only the PbE enhancement is invoked
Only the ALPs enhancement is invoked
No enhancement is invoked, i.e. CLPSO
Any Posp,d updated according to Eq. (13) and outside Oi ,t , Oi ,t is clamped to the nearest bound of Oi ,t , Oi ,t
Any Posp,d updated according to Eq. (13) and outside Oi ,t , Oi ,t is reinitialized within Oi ,t , Oi ,t using the strategy introduced in Zhang and Li (2007)
Any Posp,d updated according to Eq. (13) and outside Oi ,t , Oi ,t is allowed to be far from the feasible range
In the PbE enhancement, Eq. (22) is not invoked to repair feasibility
The penalty factor is fixed at
The penalty factor is fixed at
The penalty factor is dynamically adjusted using the JoinesHouck strategy

5.4. Flow chart

6.1. Case 1: 4-reservoir hydrothermal power system

The flow chart of the application implementation is similar


to that detailed in (Yu and Zhang, 2014) with the following
several exceptions:

The 4-reservoir hydrothermal power system consists of four


cascaded reservoirs and an aggregated thermal power plant. The
hydraulic coupling relationships of the reservoirs are illustrated
in Fig. 2. Let the power output of the thermal power plant in
time step t be Pt thermal , the load demand in time step t be Pt load ,
and the power transmission loss in time step t be Pt loss , Pt thermal

(1) In Step 12, if the updated dimensional position is infeasible, then it will be appropriately treated.
(2) In Step 14, we dont check whether the position Posp lies
within the search space. Instead, in this step, the storage
volume constraints are explicitly enforced, and the power
discharge and spillage rates are determined.
(3) In Step 15, the fitness value of each particle p is now
h(Posp) according to Eq. (29).
(4) In Step 16, if the problem is a maximization problem, the
condition judged is h(Posp) > h(Pbestp); otherwise, the
condition is h(Posp) < h(Pbestp).
6. Experimental studies
Two representative cases are studied, with case 1 being the
short-term scheduling of a 4-reservoir hydrothermal power
system introduced in Orero and Irving (1998) and case 2 being
the long-term planning of Chinas XiluoduXiangjiaba
Threegorges 3-reservoir hydropower system.
The following performance issues are investigated: (1) how
the PbE and ALPs enhancements as well as the constraint
enforcement and penalty factor adjustment strategies help
improve the algorithm performance; and (2) how ECLPSO
performs compared with other literature algorithms. For the
first issue, ECLPSO variants with different configurations are
evaluated. The ECLPSO variants are described in Table 1. For
the ECLPSO variants, the algorithm parameters take the recommended empirical values stated in Section 3 unless otherwise specified. Each ECLPSO variant is tested for 25 runs
independently. The experiment platform is a notebook computer with an Intel Core i5-3210M 2.5 GHz CPU, 6.14 GB
DDR3-1600 memory, and 64-bit Windows 7 operating system.
The following performance metrics are used to evaluate the
algorithms performance: (1) the mean, standard deviation
(SD), best, and worst of the solutions objective values; (2) the
average violation; (3) the average execution time; and (4) the
average number of exploitation valid dimensions (EVDs).

is determined by Pt thermal = Pt demand + Pt loss


needs to satisfy the following constraint.

Pthermal Pt thermal Pthermal

N
i=1

Pi ,t . Pt thermal

(32)

where Pthermal and Pthermal are respectively the minimum and


maximum power outputs of the thermal power plant. The
thermal power plant consumes fossil fuels and the fuel cost
Fuelt incurred in time step t is a quadratic function of Pt thermal .
The optimization objective is to minimize the total fuel cost
summed over all the time steps. The planning horizon is one day
with hourly time steps. The water transportation delay between
a reservoir and its immediate downstream reservoir takes
several hours. Case 1 assumes that there is no spillage, hence
the power discharge rate equals to the outflow rate. The

Fig. 2. Hydraulic coupling relationships of the reservoirs in the 4-reservoir


hydrothermal power system.

58

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

Table 2
Results of the ECLPSO variants on case 1.
ECLPSO variant

1
2
3
4
5
6
7
8
9
10
11

Fuel cost f
(in $)
Mean

SD

Best

Worst

922,327
922,330
922,364
922,376
922,768
922,460
922,459
922,358
922,323
922,468
922,336

1.80
3.34
61.76
60.82
355.70
171.37
182.75
45.14
5.79
99.83
3.90

922,324
922,325
922,325
922,328
922,324
922,326
922,366
922,330
922,303
922,358
922,329

922,332
922,339
922,464
922,464
923,581
922,998
923,130
922,472
922,328
922,654
922,344

hydropower performance model is a quadratic function of the


power discharge rate and storage volume.
For the ECLPSO variants, the swarm size Npar = 40, the
minimum penalty factor = 2729, the maximum penalty
factor = 1014 , the maximum dimensional velocity
Veld = 0.08 ( Posd Posd ) , and the maximum number of generations k =15000. ECLPSO is compared with literature algorithms including GA (Orero and Irving, 1998), IFEP (Sinha
et al., 2003), RCGA in Kumar and Naresh (2007), LWPSO in
Yu et al. (2007), and MAPSO in Amjady and Soleymanpour
(2010). The performance data of the literature algorithms are
obtained from the corresponding papers.
Table 2 lists the results of the ECLPSO variants with different configurations. Table 3 gives the performance comparison
of ECLPSO with the literature algorithms. Figs. 35 respectively depict the power discharge rates, storage volumes and
power outputs of the best solution obtained by ECLPSO.
6.2. Case 2: Chinas XiluoduXiangjiabaThreegorges
3-reservoir hydropower system
The XiluoduXiangjiabaThreegorges 3-reservoir hydropower system consists of the Xiluodu Dam (XLD), Xiangjiaba
Dam (XJB), and Three Gorges Dam (TGD) located on the
Yangtze River in China. With a length of about 6300 km, the
Yangtze River is Chinas longest river and the worlds third
longest river. The Yangtze River originates from the Qinghai
Tibet Plateau and flows eastward before debouching into the
East China Sea near Shanghai. The Yangtze River drains a wide
basin of 1.8 million km2, nearly one fifth of Chinas territory.
Fig. 6 shows the Yangtze River drainage basin and the XLD
XJBTGD system. The stretches above Yichang form the upper

Avg. violation vtotal

Avg. execution time (in s)

Avg. EVDs

0
0
0
0
0
0
0
0
1.67E-3
0
8.88E-18

7.30
6.85
5.93
5.86
7.60
7.56
6.53
6.95
7.29
6.97
7.08

35
29
50
29
44
44
33
41
33
31
33

reach of the Yangtze River, being more than 4500 km long. The
TGD is located at Sandouping near Yichang. The XJB is located
near Yibing, about 700 km upstream from Yichang and 150 km
downstream from the XLD. The TGD, XLD, and XJB are
respectively Chinas first, second, and third largest hydropower
stations. The XLDXJBTGD system plays a vitally important
role in developing and harnessing the water resources of the
Yangtze River as the system provides comprehensive benefits
such as hydropower generation, flood control, navigation, sediment control, and water provisioning.
The XLD, XJB and TGD all have seasonal regulation capabilities. Table 4 lists the detailed specifications of the three
reservoirs. The flood season is from June to September. The
TGDs forebay elevation decreases to the flood control limit
level (i.e. 145 m) during June 1 and June 10. Until September
10, the TGD maintains its forebay elevation at 145 m in order to
vacate enough storage for the incoming flood. Then the TGD
starts to impound water and the forebay elevation gradually
increases to the normal pool level (i.e. 175 m) at the end of
October or the beginning of November. From November to May
of the following year, the TGD is operated at high forebay
elevations that are no less than the drought season control level
(i.e. 155 m). The TGDs forebay elevation should be no more
than 155 m at the end of May. The XLD and XJB work similarly
with the TGD, i.e. operate at low forebay elevations in the flood
season, refill in the flood recession period, and gradually draw
off water in the drought season. The XLD starts to refill in early
September, while the XJB normally starts impounding water in
mid-September.
In this case, we study the long-term planning of the XLD
XJBTGD system. The planning horizon is a water year from

Table 3
Performance comparison of ECLPSO and literature algorithms on case 1.
Algorithm
Fuel cost f (in $)

Avg. violation vtotal


Avg. execution time (in s)

Best
Mean
Worst

ECLPSO

GA

IFEP

RCGA

LWPSO

MAPSO

922,324
922,327
922,332
0
7.30

932,734
936,969
939,734
>0.1
1,200

930,130
930,290
930,882

1,033.20

925,940
926,120
926,539

57.52

925,384
926,353
927,240

82.90

922,421
922,544
923,508

64.00

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

59

Fig. 3. Power discharge rates of the best solution obtained by ECLPSO on case
1.

June 1 to the following years May 31 with 10-day time steps;


thus there are 36 time steps in total. The observed natural inflow
records in the water year 20052006 are used and are depicted
in Fig. 7. The natural inflows into the XJB arent shown in Fig. 7
and are around 1% of those into the XLD. The optimization
objective is to maximize the total hydropower generated from
the three reservoirs over the planning horizon. The hydropower
performance model is calculated according to Eq. (10). The
surface precipitation, evaporation, and seepage loss are
ignored. The water transportation delay between any two neighboring reservoirs is far less than that of 10-day time steps. The
XLDs tailrace elevation is affected by its outflow rate and the
immediate downstream reservoir XJBs forebay elevation. This

Fig. 5. Power outputs of the best solution obtained by ECLPSO on case 1.

case is formulated based on realistic operation guidelines. Firm


power requirements are considered. The minimum outflow rate
requirements could vary for different time steps. For example,
the TGDs minimum outflow rates are 10,000 m3/s during mid
and late September, 8000 m3/s in October, and 6000 m3/s in the
drought season. Additional constraints are imposed on the
variation of each reservoirs consecutive forebay elevations.
Take the TGD again as an example, when the reservoirs
forebay elevation gradually decreases from 175 m to 155 m in
the drought season, the daily decrease is suggested to be no
more than 0.6 m.
ECLPSO is compared with literature algorithms such
as GPSO (Shi and Eberhart, 1998) and LPSO (Kennedy
and Mendes, 2002). For the ECLPSO variants, GPSO,
= 6.25E9,
= 1012 ,
and
LPSO,
Npar = 40,
Veld = 0.06 ( Posd Posd ) , and k = 7000 .
Table 5 lists the results of the ECLPSO variants. Table 6
gives the performance comparison of ECLPSO with GPSO and
LPSO. Figs. 810 respectively depict the outflow rates, forebay
elevations, and power outputs of the best solution obtained by
ECLPSO.
6.3. Analysis of the experimental results

Fig. 4. Storage volumes of the best solution obtained by ECLPSO on case 1.

The PbE and ALPs enhancements: As can be observed


from Table 2 and Table 5, ECLPSO performs significantly
better than CLPSO on the two cases in terms of the mean
objective result. The SD objective result of ECLPSO is much
smaller than that of CLPSO on case 1, and around the same on
case 2. ECLPSO variant 2 far surpasses CLPSO in terms of the
mean, SD, and worst fuel costs on case 1. It is noted that
ECLPSO variant 2 peers with CLPSO in terms of the hydropower generation performance results on case 2; this is because

60

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

Fig. 6. The Yangtze River drainage basin and the XiluoduXiangjiabaThreegorges 3-reservoir hydropower system.

min { ( Posd Posd ), } = 2 m3 s in this case, and the


release decisions around the PbE activated dimensions lead to
very similar hydropower generation results. On both cases,
ECLPSO outperforms ECLPSO variant 2, and ECLPSO variant
3 beats CLPSO, in terms of the mean objective result. The mean
and worst fuel costs of ECLPSO are quite better than those of
ECLPSO variant 3 on case 1. The observations demonstrates
that: (1) the PbE enhancement, through introducing sufficient
diversity into the exploitation process, is strong in solution
refinement; (2) the ALPs enhancement, with the particles
learning probabilities adaptively adjusted according to the
rankings of the personal best fitness values and the particles
exploitation progress, helps to facilitate convergence; and (3)
using the ALPs enhancement alone might not significantly
benefit exploitation, as indicated from the results on case 1.
Therefore, the integration of the PbE and ALPs enhancements
is an appropriate way to significantly facilitate convergence and
enhance the exploitation performance of CLPSO.
The enforcement of the outflow constraints: It can be
observed from Table 2 and Table 5 that ECLPSO variant 5,
ECLPO variant 6, and ECLPSO variant 8 are quite worse than

ECLPSO in terms of the mean, SD, and worst objective results


on case 1, and perform similarly on case 2. ECLPSO significantly outperforms ECLPSO variant 7 concerning the mean,
best, and worst objective result on both cases. The observations
verify that: (1) an infeasible outflow rate needs to be appropriately treated; (2) clamping, reinitialization, and allowing an
infeasible outflow rate to stay far from the feasible range all
might negatively impact the solution quality; and (3) it is beneficial to allow infeasible outflow rates close to the feasible
range and repair feasibility in the PbE enhancement.
The adjustment of the penalty factor: The results in Table 2
and Table 5 indicate that ECLPSO variant 9 doesnt perform
worse than ECLPSO in terms of the mean, best, and worst objective results on the two cases. However, the average violation
results of ECLPSO variant 9 are nonzero and unsatisfactory on
the two cases. In contrast, ECLPSO and ECLPSO variant 10 are
able to derive absolutely feasible solutions in all the runs, and
ECLPSO variant 11 can find solutions with rather small violations, on both cases. The mean and worst objective results of
ECLPSO variant 10 and ECLPSO variant 11 are significantly
worse than those of ECLPSO on the two cases. Hence, the

Table 4
Specifications of the XiluoduXiangjiabaThreegorges 3-reservoir hydropower system.
Reservoir
8

Total storage capacity (in 10 m )


Crest elevation (in m)
Normal pool level (in m)
Flood control limit level (in m)
Drought season control level (in m)
Dead pool level (in m)
Installed capacity (in MW)
Firm power output (in MW)

Xiluodu Dam

Xiangjiaba Dam

Three Gorges Dam

115.73
605
600
560

540
13,860
3,395

51.63
384
380
370

370
6,448
2,009

393.00
185
175
145
155
145
22,500
4,990

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

61

Table 6
Performance comparison of ECLPSO and literature algorithms on case 2.
Algorithm
Hydropower generation f
(in 107 kWh)

Mean
SD
Best
Worst

Avg. violation vtotal


Avg. execution time (in s)

Fig. 7. Natural inflows for case 2.

observations testify that the dynamic adjustment of the penalty


factor according to Eq. (30) is appropriate; the penalty factor is
small in the beginning to encourage exploration of the search
space, grows slowly until the end of the run to intentionally guide
the search to concentrate in the feasible region, and increases
sharply to a very large value during the end of the run to help
obtain a feasible solution. Compared with our proposed penalty
factor adjustment strategy, the JoinesHouck strategy grows too
quickly until the end of the run and doesnt increase to a large
enough value at the end of the run, thus the JoinesHouck strategy cannot sufficiently explore the search space and the final
solution obtained might not be absolutely feasible.
The execution time: For the ECLPSO variants, their execution time mainly differ based on the invoking of the PbE
enhancement, the invoking of the ALPs enhancement, and the
efforts spent on treating infeasible outflow rates. Note that the
time consumed by the PbE enhancement depends on when Eq.
(16) becomes true on each dimension and how the dimensional
normative interval changes. As can be seen from Table 2 and
Table 5, the average execution time of ECLPSO is slightly more

ECLPSO

GPSO

LPSO

18,378
8.18
18,395
18,361
0
6.68

18,283
112.42
18,513
18,023
0.31
6.29

18,290
31.77
18,345
18,219
6.66E-17
6.28

than that of CLPSO on the two cases, indicating that the PbE
and ALPs enhancements as well as the constraint enforcement
strategy trade the performance improvement with some slight
computation cost.
Comparison of ECLPSO with other literature algorithms: As
can be observed from Table 3, on case 1 ECLPSO knocks off all
the other literature algorithms concerning the mean, SD, best,
and worst fuel costs. The average execution time of ECLPSO is
also significantly better than that of the other literature algorithms on case 1. The results given in Table 6 show that on case
2 ECLPSO performs quite better than GPSO and LPSO with
regard to the mean, SD, best, and worst hydropower generation
results. GPSO even cannot find a feasible solution in some runs
on case 2 because GPSO is liable to get stuck in premature
convergence. The average execution time of ECLPSO is just
slightly more than that of GPSO and LPSO on case 2. On both
cases the literature algorithms compared arent robust in performance, i.e. the worst objective result is significantly different
from the best objective result. In contrast, ECLPSO is rather
robust. ECLPSO is able to robustly derive high quality feasible
operation policies for multi-reservoir hydropower systems
using a reasonable amount of time because of the following
reasons: (1) the comprehensive learning strategy allows each
particle to learn from different exemplars on different dimensions, hence ECLPSO is good at preserving the particles diversity and avoiding premature convergence; (2) the PbE
enhancement lets ECLPSO to achieve high performance
exploitation; (3) the ALPs enhancement facilitates convergence; (4) the outflow and storage volume constraints are
appropriately enforced to achieve a tradeoff between preserving

Table 5
Results of the ECLPSO variants on case 2.
ECLPSO variant

1
2
3
4
5
6
7
8
9
10
11

Hydropower generation f
(in 107 kWh)
Mean

SD

Best

Worst

18,378
18,367
18,378
18,365
18,376
18,379
18,360
18,376
18,379
18,341
18,352

8.18
7.22
6.51
8.29
8.89
4.82
12.13
8.42
7.59
11.62
10.79

18,395
18,377
18,388
18,379
18,385
18,384
18,378
18,396
18,394
18,366
18,368

18,361
18,350
18,364
18,352
18,352
18,367
18,343
18,361
18,364
18,316
18,332

Avg. violation vtotal

Avg. execution time (in s)

Avg. EVDs

0
0
0
0
0
0
0
0
1.87E-3
0
8.24E-10

6.68
6.62
5.96
5.90
6.61
6.61
6.58
6.40
6.65
6.56
6.61

44
39
41
38
42
39
44
44
41
40
41

62

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

Fig. 8. Outflow rates of the best solution obtained by ECLPSO on case 2.

Fig. 10. Power outputs of the best solution obtained by ECLPSO on case 2.

diversity and facilitating convergence; (5) the penalty factor is


appropriately adjusted during the run in order to gradually
guide the search toward the feasible region; and (6) the PbE and
ALPs enhancements as well as the constraint enforcement strategy incur just a slight computation burden. Accordingly,
ECLPSO together with the constraint enforcement and penalty
factor adjustment strategies proposed in this paper is recommended as a promising optimization framework for the operation planning of multi-reservoir hydropower systems.
Discussions: As Fig. 5 shows, the hydropower subsystem of
case 1 tries to generate more power during the peak-load hours.
However, as the initial storage volumes of the reservoirs are
small and the hourly natural inflows into the reservoirs are also
small (Orero and Irving, 1998), the hydropower outputs are

miserly as compared with the thermal power outputs. As listed


in Table 2 and Table 5, the average EVDs results of ECLPSO
are respectively 35 and 44 on case 1 and case 2. The problem
dimensions are 96 for case 1 and 108 for case 2. This noted
phenomenon that not all the dimensions can be activated for the
PbE is caused by the dynamic nature of decision making and
the hydraulic coupling among the reservoirs. The objective
space is highly multi-modal.
7. Conclusions and future work
An optimization framework for the operation of multireservoir hydropower systems has been proposed in this paper.
The framework adopts our recently proposed algorithm
ECLPSO as the optimization tool because ECLPSO is well
balanced in exploration and exploitation. The outflow and
storage volume constraints are appropriately enforced to
achieve a tradeoff between preserving diversity and facilitating
convergence. The penalty function technique is employed to
convert the original constrained problem into an unconstrained
one. The penalty factor is adjusted dynamically in order to
encourage exploration of the search space in the beginning and
gradually guide the search to concentrate in the feasible region.
Experiments on two cases have demonstrated that the framework is able to help derive feasible high quality operation
policies robustly. In our future work, the exploration and exploitation performance of ECLPSO will be further improved and
ECLPSO will be adapted and applied to more multi-reservoir
hydropower systems with the operation being discontinuous
and/or mixed-integer.
Acknowledgement

Fig. 9. Forebay elevations of the best solution obtained by ECLPSO on case 2.

This work was supported by the Public Benefit Special


Research Fund of the Ministry of Water Resources of the Peoples Republic of China (201201017). The work of Hui Qin was
supported by the National Natural Science Foundation of China
(51209008).

X. Zhang et al. / Journal of Hydro-environment Research 10 (2016) 5063

References
Afshar, M.H., 2012. Large scale reservoir operation by constrained particle
swarm optimization algorithms. J. Hydro-environ. Res. 6 (1), 7587.
Afshar, M.H., 2013. Extension of the constrained particle swarm optimization
algorithm to optimal operation of multi-reservoir system. Int. J. Electr.
Power Energy Syst. 51, 7181.
Aguirre, A.H., Rionda, S.B., Coello Coello, C.A., Lizrraga, G.L., Montes,
E.M., 2004. Handling constraints using multiobjective optimization
concepts. Int. J. Numer. Methods Eng. 59 (15), 19892017.
Amjady, N., Soleymanpour, H.R., 2010. Daily hydrothermal generation
scheduling by a new modified adaptive particle swarm optimization
technique. Electr. Power Syst. Res. 80 (6), 723732.
Boussad, I., Lepagnot, J., Siarry, P., 2013. A survey on optimization
metaheuristics. Inf. Sci. (Ny) 237, 82117.
Chen, J.-H., Guo, S.-L., Li, Y., Liu, P., Zhou, Y.-L., 2013. Joint operation and
dynamic control of flood limiting water levels for cascade reservoirs. Water
Resour. Manag. 27 (3), 749763.
Cheng, C.-T., Shen, J.-J., Wu, X.-Y., 2012. Short-term scheduling for
large-scale cascaded hydropower systems with multivibration zones of high
head. J. Water Resour. Plann. Manage. 138 (3), 257267.
Deb, K., 2000. An efficient constraint handling method for genetic algorithms.
Comput. Methods Appl. Mech. Eng. 186 (2), 311338.
Eberhart, R.C., Kennedy, J., 1995. A new optimizer using particle swarm
theory. International Symposium on Micromachine and Human Science
3943.
El-Hawary, M.E., Christensen, G.S., 1979. Optimal Economic Operation of
Electric Power Systems. Academic Press.
He, Q., Wang, L., 2007. An effective co-evolutionary particle swarm
optimization for constrained engineering design problems. Eng. Appl. Artif.
Intell. 20 (1), 8999.
Ji, C.-M., Zhou, T., Huang, H.-T., 2014. Operating rules derivation of Jinsha
reservoirs system with parameter calibrated support vector regression.
Water Resour. Manag. 28 (9), 24352451.
Joines, J.A., Houck, C.R., 1994. On the use of non-stationary penalty functions
to solve nonlinear constrained optimization problems with GAs. IEEE
Congress on Evolutionary Computation 579584.
Kennedy, J., Eberhart, R.C., 1995. Particle swarm optimization. International
Conference on Neural Networks 19421948.
Kennedy, J., Mendes, R., 2002. Population structure and particle swarm
performance. IEEE Congress on Evolutionary Computation 16711676.
Kumar, D.N., Reddy, M.J., 2007. Multipurpose reservoir operation using
particle swarm optimization. J. Water Resour. Plann. Manage. 133 (3),
192201.
Kumar, S., Naresh, R., 2007. Efficient real coded genetic algorithm to solve the
non-convex hydrothermal scheduling problem. Int. J. Electr. Power Energy
Syst. 29 (10), 738747.
Labadie, J.W., 2004. Optimal operation of multireservoir systems:
state-of-the-art review. J. Water Resour. Plann. Manage. 130 (2), 93
111.
Leguizamn, G., Coello Coello, C.A., 2007. A boundary search based ACO
algorithm coupled with stochastic ranking. IEEE Congress on Evolutionary
Computatio, IEEE 165172.
Li, F.-F., Wei, J.-H., Fu, X.-D., Wan, X.-Y., 2012. An effective approach to
long-term optimal operation of large-scale reservoir systems: case study of
the Three Gorges system. Water Resour. Manag. 26 (14), 40734090.
Li, F.-F., Shoemaker, C.A., Wei, J.-H., Fu, X.-D., 2013. Estimating maximal
annual energy given heterogeneous hydropower generating units with
application to the Three Gorges System. J. Water Resour. Plann. Manage.
139 (3), 265276.

63

Li, L.-P., Liu, P., Rheinheimer, D.E., Deng, C., Zhou, Y.-L., 2014a. Identifying
explicit formulation of operating rules for multi-reservoir systems using
genetic programming. Water Resour. Manag. 28 (6), 15451565.
Li, X., Li, T.-J., Wei, J.-H., Wang, G.-Q., Yeh, W.W.-G., 2014b. Hydro unit
commitment via mixed integer linear programming: a case study of the
Three Gorges project, China. IEEE Trans. Power Syst. 29 (3), 12321241.
Liang, J.J., Qin, A.K., Suganthan, P.N., Baskar, S., 2006. Comprehensive
learning particle swarm optimizer for global optimization of multimodal
functions. IEEE Trans. Evol. Comput. 10 (3), 281295.
Lyra, C., Ferreira, L.R.M., 1995. A multiobjective approach to the short-term
scheduling of a hydroelectric power system. IEEE Trans. Power Syst. 10 (4),
17501755.
Miettinen, K., 1999. Nonlinear Multiobjective Optimization. Kluwer Academic
Publishers.
Needham, J.T., Watkins, D.W., Lund, J.R., Nanda, S.K., 2000. Linear
programming for flood control in the Iowa and Des Moines rivers. J. Water
Resour. Plann. Manage. 126 (3), 118127.
Orero, S.O., Irving, M.R., 1998. A genetic algorithm modelling framework and
solution technique for short term optimal hydrothermal scheduling. IEEE
Trans. Power Syst. 13 (2), 501518.
Shi, Y.-H., Eberhart, R.C., 1998. A modified particle swarm optimizer. IEEE
Congress on Evolutionary Computation 6973.
Simonovic, S., 1987. The implicit stochastic model for reservoir yield
optimization. Water Resour. Res. 23 (12), 21592165.
Sinha, N., Chakrabarti, R., Chattopadhyay, P., 2003. Fast evolutionary
programming techniques for short-term hydrothermal scheduling. IEEE
Trans. Power Syst. 18 (1), 214220.
Tauxe, G.W., Inman, R.R., Mades, D.M., 1980. Multiple objectives in reservoir
operation. J. Water Resour. Plan. Manag. Div. 106 (1), 225238.
Wang, J.-W., Zhang, Y.-C., 2012. Short-term optimal operation of hydropower
reservoirs with unit commitment and navigation. J. Water Resour. Plann.
Manage. 138 (1), 312.
Westphal, K.S., Vogel, R.M., Kirshen, P., Chapra, S.C., 2003. Decision support
system for adaptive water supply management. J. Water Resour. Plann.
Manage. 129 (3), 165177.
Wu, J.-K., Zhu, J.-Q., Chen, J.-Q., Zhang, H.-L., 2008. A hybrid method for
optimal scheduling of short-term electric power generation of cascaded
hydroelectric plants based on particle swarm optimization and
chance-constrained programming. IEEE Trans. Power Syst. 23 (4),
15701579.
Xiong, M., 1990. Short-term generation scheduling in a hydrothermal power
system (Doctoral dissertation). Durham University, UK.
Yu, B.-H., Yuan, X.-H., Wang, J.-W., 2007. Short-term hydro-thermal
scheduling using particle swarm optimization method. Energy Convers.
Manag. 48 (7), 19021908.
Yu, X., Zhang, X.-Q., 2014. Enhanced comprehensive learning particle swarm
optimization. Appl. Math. Comput. 242, 265276.
Yuan, X.-H., Wang, L., Yuan, Y.-B., 2008. Application of enhanced PSO
approach to optimal scheduling of hydro system. Energy Convers. Manag.
49 (11), 29662972.
Zhan, Z.-H., Zhang, J., Li, Y., Shi, Y.-H., 2011. Orthogonal learning particle
swarm optimization. IEEE Trans. Evol. Comput. 15 (6), 832847.
Zhang, Q.-F., Li, H., 2007. MOEA/D: a multiobjective evolutionary algorithm
based on decomposition. IEEE Trans. Evol. Comput. 11 (6), 712731.
Zhang, R., Zhou, J.-Z., Ouyang, S., Wang, X.-M., Zhang, H.-F., 2013. Optimal
operation of multi-reservoir system by multi-elite guide particle swarm
optimization. Int. J. Electr. Power Energy Syst. 48, 5868.
Zhang, Z.-B., Jiang, Y.-Z., Zhang, S.-H., Geng, S.-M., Wang, H., Sang, G.-Q.,
2014. An adaptive particle swarm optimization algorithm for reservoir
operation optimization. Appl. Soft Comput. 18, 167177.

You might also like