You are on page 1of 12

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO.

5, SEPTEMBER 2014

2347

Parallel Self-Adaptive Differential Evolution


Algorithm for Solving Short-Term Hydro
Scheduling Problem
Arnel Gloti, Student Member, IEEE, Adnan Gloti, Peter Kitak, Member, IEEE, Joe Pihler, Member, IEEE,
and Igor Tiar, Member, IEEE

AbstractIn order to optimize hydro power plants generator scheduling according to 24-h system demand, a parallel
self-adaptive differential evolution algorithm has been applied.
The proposed algorithm presents a novel approach to considering
the multi-population and utilization of the preselection step for
the improvements of the algorithms global search capabilities.
A preselection step with the best, middle, and worst populations
individuals establishes the new trial vectors. This algorithm has
been verified on two different models. The first one consists of
eight power plants with real parameters, and the second one
consists of four power plants, mostly used as a test model in
scientific papers. The main goal of the optimization process is to
satisfy system demand for 24 h with a decreased usage of water
quantity per electrical energy unit. The initial and final states of
the reservoirs must also be satisfied.
Index TermsAlgorithms, dispatching, hydroelectric power
generation, optimization methods, parallel algorithms.

Minimal output power of hydro plant .


Number of hours of the scheduling period.
Total number of hydro power plants.
Variables
Sum of inflows to hydro plant in hour .
Flow through turbine of hydro plant in hour .
Spillage of hydro plant in hour .
Reservoir volume of hydro plant in hour .
Output power of hydro plant in hour .
Flow through old riverbed to hydro plant in hour .
B. Differential Evolution
Parameters

NOMENCLATURE

Population size.
Number of parameters.

A. Hydro Power Plant Model

Crossover control parameter.

Parameters
Natural inflow to hydro plant in hour .
Required biological minimum flow of hydro plant
.

Difference factor.
Variables
Generation through algorithm steps.
Value of th vector.

Maximal reservoir volume of hydro plant .


Minimal reservoir volume of hydro plant .
Maximum allowed change of reservoir .
Maximal output power of hydro plant .

Trial vector of th vector in generation


C. Objective Function
Parameters
Weight for objectives.

Minimal water discharge of hydro plant .


Maximal water discharge of hydro plant .
Manuscript received August 01, 2013; revised August 26, 2013, November
06, 2013, December 13, 2013, December 23, 2013, and December 23, 2013; accepted January 21, 2014. Date of publication February 07, 2014; date of current
version August 15, 2014. Paper no. TPWRS-00992-2013.
A. Gloti, P. Kitak, J. Pihler, and I. Tiar are with University of Maribor,
Faculty of Electrical Engineering and Computer Science, Institute of Power
Engineering, Maribor 2000, Slovenia (e-mail: arnel.glotic@uni-mb.si; peter.
kitak@uni-mb.si; joze.pihler@uni-mb.si; ticar@uni-mb.si).
A. Gloti is with the HSE Group (Holding Slovenske elektrarne d.o.o.), Ljubljana 1000, Slovenia (e-mail: adnan.glotic@hse.si).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TPWRS.2014.2302033

Correction factor for objective function.


Variables
Objective function.
Demand energy in hour
Sum of optimal energy production in hour .
I. INTRODUCTION
HE topic of this paper is the optimization of cascade
hydro power plants (HPPs) according to the system
demand. A significant portion of publications within this field
covers the hydrothermal power system optimization, where

0885-8950 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2348

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 5, SEPTEMBER 2014

the goal is to reduce thermal cost by hydroelectric generation scheduling [1][7]. Power generation in todays market
conditions requires continuous rationalization and profit maximization. In this context, the motivation behind these works
was to optimize the operations of HPPs. Therefore, the main
goal is to satisfy the system demand with HPPs by optimal
generator scheduling in order to decrease water quantity per
produced electrical energy unit (WQPEU) and decrease or even
eliminate water spillage. In order to create an optimization
algorithm appropriate for real-world and practical applications,
respectively, and for use in hydro generating companies, the
algorithm should have fast convergence time with the ability
to satisfy 24-h system demand in respect to all operational
constrains and requirements.
The optimization of cascade HPP reservoirs scheduling,
where downstream plant operation depends on upper plants
operation in addition to a large solution space, is known as a
complex problem. Within the last 25 years, a wide range of optimization methods has been applied for solving this problem.
In order to achieve optimal production, several methods can
be implemented [7]. These can be generally classified into two
main groups: deterministic and heuristic methods. Deterministic methods generally arrive at the same final solution through
the same sequence of solutions, while heuristic methods can be
constructive (build a solution piece by piece) or improving (take
a solution and alter it to find a better solution). Deterministic
methods include Lagrangian relaxation [8][10] and Benders
decomposition-based methods [7], mixed-integer programming
[11], [12], dynamic programming [13], [14] and sequential
quadratic programming (SQP) [15], [16]. In [15], by means of
the Lagrangian relaxation, with the hydro scheduling problem
split into a sequence of smaller and easier to solve subproblems
and afterwards, the nonlinear problem is successfully solved
by using the SQP method. The impact of combining the SQP
method with an evolutionary algorithm in the process of determining the optimal daily self-scheduling is shown in [16].
Particle swarm optimization (PSO) [1], [17], Genetic algorithms [18][20], evolutionary programming [21], and
differential evolution algorithms [4], [5], [22][28] are classified as heuristics methods. The differential evolution (DE)
algorithm [22] is an efficient and robust global optimization
algorithm, and therefore it has been selected in this paper as an
appropriate optimization technique. Short-term optimization
using DE [23] has been used on four cascade HPPs, where the
best objective value has been reached after 2000 generations.
The authors in [23] highlighted the difficulty of setting the adequate control parameters and therefore applied the chaos theory
[26] in order to improve the performance of the algorithm. The
modified DE presented in [5] includes a penalty factor during
the objective function evaluation, which preserves the satisfied
final reservoirs levels of four cascade HPPs. In order to find
optimal system cost and emission, the algorithm [5] was first
tested on four cascade HPPs and one equivalent thermal plant
and on the four cascade HPPs and three thermal plants. In
[24], the authors combined the advantages of the two modified
DE algorithms, where the grouping and shuffling operation is
carried out over the population periodically. The goal of this
approach was to find the best reservoir scheduling of two HPPs,
but without satisfying the final states of the reservoirs.

The capability of finding the global optimum by using DE


is closely connected with the proper selection of DE control pa, the differential factor
rameters [27] like the population size
and crossover constant
. To avoid this uncertainty, the
self-adaptive DE has been proposed in [25], [28]. The authors
in [29] applied dynamical differential factor combined with
the modified version of mutation strategy where population diversity information was utilized in order to control the search
capability and prevent the algorithm being stuck within a locally optimal solution. As soon as the population diversity was
equal or lower than that prescribed, the algorithm automatically
replaced several worst individuals.
The DE algorithms performance can be also improved by
parallelization [30]. The parallelization of the DE can be carried out on three levels [31]: the objective function evaluation
level (masterslave model), the population level (island or
migrant models), and the elements level (cellular model). The
parallelization on objective evaluation level is generally used
in applications where the objective function evaluation is the
most time consuming element within the optimization process
and the elements level parallelization is generally used for
massive parallel machines [31]. Therefore the parallelization
on the population level is used in this paper in similar way as in
[31][33]. The proposed parallel differential evolution (PDE)
uses self-adaptive parameters and consists of subpopulations
, where denotes the number
with an equal population size
of CPUs. After all subpopulations have been executed with
the mutation, crossover, and evaluation step, the common
selection step is followed by using the novel best-middle-worst
(BmW) strategy, where an optimal balance is achieved between
convergence time, population diversity, and global optimum.
Parallel self-adaptive differential evolution (PSADE) has been
verified on two models, as shown in Fig. 1. It is compared with
the different parallelization of the DE algorithm and with the
classic DE algorithm by using different parameter sets. The
PSADE is also compared with the SQP method. The power
system demand for both models must be satisfied by optimal
generation schedule followed by the decreased usage of water
quantity per produced electrical energy unit m MWh and
decreased or eliminated water spillage.
This paper is organized as follows. Mathematical models
are presented in Section II. The proposed PSADE with BmW
strategy are given in Section III. Simulation and results are
presented in Section IV, with conclusion given in Section V.
II. MATHEMATICAL MODELS OF HYDRO POWER PLANTS
This paper considers two different mathematical models.
Model I [Fig. 1(a)] is a real parameters model consisting of
eight cascade HPPs located in Slovenia and Model II [Fig. 1(b)]
as recently used in several papers [1][5]. The Model II consists
of four HPP with details presented in [2] and [5], and it uses
fewer restrictions in comparison to Model I, which uses real
restrictions and requirements which must be fulfilled. The total
inflow for each HPP for Model I in the observed hour is the
sum of the upstream plant flow through the turbine, the spillage,
and the natural inflow:
(1)

GLOTI et al.: PARALLEL SELF-ADAPTIVE DIFFERENTIAL EVOLUTION ALGORITHM FOR SOLVING SHORT-TERM HYDRO SCHEDULING PROBLEM

2349

by considering that the reservoirs values in the observed hour


of -th plant must be between the minimal and the maximal
level
(5)
of each individual reservoir . The water discharge on each HPP
must also be within the minimal and maximal allowed water
discharge of th HPP as
(6)
Model I also includes maximum allowed change
of the th
HPP reservoir level in hour compared with
as:
(7)
In cases where the inflow
exceeds the maximal allowed flow
through the turbines of the th HPP in hour and the reservoir
volume reaches the maximal allowed level, then the spillage
is unavoidable:
(8)
The hydro generator output power is expressed as

(9)
where
represents the hydropower generation
coefficients. The input data and coefficients for Model I were
obtained from the company Dravske elektrarne Maribor and
HSE Group, respectively. The input data and coefficients for
Model II can be found in [2] and [5].

Fig. 1. (a) Model I with eight HPPs. (b) Model II with four HPPs.

The last two HPPs in Model I are canal-based types where the
flow merges with the riverbed flow at the end of the canal. For
the old riverbed flow, time delay to the downstream plant must
be considered. These last two HPPs also have the required biological minimum flow which must always be provided to the
riverbed and it is different for winter and summer time. Therefore, the total inflow of the last HPPs in Model I is
(2)
is the time delay from entry of old riverbed to the
where
downstream HPP and
is the water flow through the old
riverbed, which is usually used at a time when the reservoir 6
is on the maximum storage level and the water discharge from
HPP 6 has already reached the maximum allowed discharge.
The total inflow for last two HPPs in Model II consist of the
sum of upstream HPPs discharges as

III. APPLICATION OF THE MODIFIED DE ALGORITHM


The PSADE algorithm has been used as parallel implementation of the multipopulation algorithm. The flow chart of the
proposed algorithm is shown in Fig. 2.
A. Initialization and Implementation of the Considered
Problem
The initialized population for the presented mathematical
models consists of the values for the hydro discharges of each
individual models HPP and the individual hours during the
observed day (24 h). For instance, the Model I with eight HPPs
and within observed intervals over a 24-h period with 1-h
time steps represents a problem dimension
of 192 (96 for
Model II). Therefore, the initial population is composed of
-dimensional vectors as follows:

(10)
(3)
(4)
The inflow water in both models can be used for charging reservoirs up to the maximal reservoirs volume
or used in
combination with the flow gained from discharging reservoirs,

where
is the set of hydro discharges of the th HPP which
contains randomly chosen values
. The
initial population is distributed on CPU cores as shown in Fig. 3.
The mutation and crossover step is run on each core, and the
new trial members are created by using self-adaptive parame. Each th place in population contains water
ters
and

2350

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 5, SEPTEMBER 2014

Fig. 2. Flow chart of parallel implementation of multipopulation DE.

Fig. 4. Step-by-step procedure of PSADE algorithm.

compared with a target member. Members with better objective


function values obtain positions in the target population.
B. Mutation and Crossover With Self-Adaptive Parameters
DE algorithm control parameters selection requires certain
effort in order to find proper values for individual optimization
problems. In order to avoid this, the authors of [25] proposed a
DE algorithm with self-adaptive parameters and
. In this
paper, the idea of self-adaptive parameters for DE from [25] is
expanded as

if
elseif
Fig. 3. HPP model integration into PSADE algorithm.

otherwise
(11)

discharges for all HPPs. The discharges for each -th HPP are
collected and used as input data of th HPP.
The generator output of the th HPP in hour is calculated as
in (9); afterwards, the evaluation step is followed. The step-bystep procedure of the evaluation step is shown in Fig. 4.
When the evaluation step is finished by all cores, the trial population is collected and sorted according to the members objective values, from minimum to maximum. The new trial population is formed by the proposed BmW strategy which is described
below. Afterwards, each trial member from trial population is

elseif
otherwise
(12)
where
,

,
, and

,
are randomly generated

GLOTI et al.: PARALLEL SELF-ADAPTIVE DIFFERENTIAL EVOLUTION ALGORITHM FOR SOLVING SHORT-TERM HYDRO SCHEDULING PROBLEM

numbers from interval [0, 1],


and
difference factor for each th vector (10) in generation

2351

is the
.

C. Subpopulations Evaluation Step


trial
The subpopulation on each CPU core consists of
vectors which are evaluated by using the objective function. In
this paper, the objective function is composed of three separate
objective functions
which are merged into a single
objective function (22) by using the weighted sum method
[34]. This method transforms a set of three objectives of a given
problem into a single objective by using weights , , and
which are defined by the user in proportion to the individual
objectives importance. Each weight can take any number from
interval [0, 1], but the sum of the weights should be equal to 1.
When the value of 1/3 is set as the weight for each individual objective function, the chances of functions being minimized are
equal for all. A set of different weights leads to a set of different
solutions. Therefore, a different set of weights has been tested
and the set that obtained the best result by all objectives, has
been chosen as authors appropriate selection for weights ,
, and , which were 0.65, 0.25, and 0.1. Objective evaluates agreement between the system demand
and production
obtained by PSADE as follows:

Fig. 5. Impact of (a) the sensitivity of equation depends on (b) demand and
production.

the weight
is set low (sets 17 in Fig. 6(b)), the priority to
the largest mismatch of the system demand is decreased and
therefore the average mismatch is adequately larger. The best
agreement is achieved when both weights are set to 0.5 [set 10
in Fig. 6(b)].
The second objective in

(13)
(14)

(17)

(15)

(16)
is the hydro
where is the sensitivity of the equation, and
scheduling time step and the basic time interval, respectively.
The sensitivity of the equation has an impact on the function
value in (14), as shown in Fig. 5(a), where the system demand
through 100-h period has been set to 10 MWh and the production set to increase from 0 to 10 MWh with 0.1 MWh step
through a 100-h period as shown in Fig. 5(b). It is obvious that
bigger equation sensitivities have higher and faster impacts
on those parts, where the differences between system demand
and production are smaller. The authors set the sensitivity of
equation and time-step
to value 1. The mismatch between
system demand and production for each hour is calculated by
using (14). The larger mismatch on time interval [1, ] is obtained by using (15) and then the obtained value is used in (16).
This helps the objective function not just to follow the average
mismatch value, but also to follow the higher mismatch. The
and
are weights and both have been set as 0.5 to ensure balance between the largest mismatch of system demand founded
in hour and the average mismatch value for time interval, with
the exception of hour .
and
(example: set 1 in
The impact of weights set
Fig. 6(a) is
and
) on the final agreement between system demand and production obtained by the
PSADE and applied to Model II is shown in Fig. 6(b). When

evaluates the water quantity used per produced electrical energy, which means that the total water discharge obtained by
HPPs units is divided by the total produced energy. Because the
second objective is normalized between 0 and 1, the total maximal energy of th HPP unit is divided by the total maximal energy of th HPP unit and used in (17). The third objective shown
in

(18)
is complementary to the second objective and it evaluates the
water spillage. The restriction (6) can be integrated directly
within the mathematical model by cutting the parameters with
values outside the allowed boundaries or it can be considered
through objective function (22) by using correction factor
for the objective function, as calculated in

(19)
(20)
(21)

2352

Fig. 6. Impact of (a) 19 different weights set


runs.

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 5, SEPTEMBER 2014

and

to (b) final agreement between the system demand and the production, averaged over ten independent

Fig. 7. Correction matrix for objective function.

The separate correction factors obtained from (19) actually


present a correction matrix, as shown in Fig. 7.
The influence of the correction factor on the objective function value is greater when the individual is closer to the lower
allowable parameters bounds and the impact is smaller when
the individual is closer to zero.
and
are weights and both
have been set as 0.5 to ensure the balance between the largest
deviation from minimal output power across all HPPs obtained
in (20) and averaged deviation for the whole time interval ,
with the exception of hour .
Correction factor
obtained from (21) is applied onto a
has a significant impact
single objective function (22). The
on the process of keeping or moving exceeding parameters into
the allowable interval. When the output power on each th HPP
in each hour during time-interval satisfies the output power
restrictions, the correction factor will be equal to 1 and therefore
will not have any impact on objective function:
(22)
D. Forming New Trial Vector by BmW Strategy
After all of the subpopulations have been processed by the
mutation, crossover, and evaluation steps on each CPU core,
the subpopulations are sent to the master CPU core. The master
CPU core collects sub-populations into a temporary population
which is -times larger in comparison to the original initial population, and where denotes the number of CPU core. Afterwards, the master CPU core sorts the temporary population according to their objective function values, from best to worst, as
shown in Fig. 8.

Fig. 8. Forming the new trial vector by BmW strategy.

After the population is collected and sorted, the new trial popis created by the proposed BmW strategy. The trial
ulation
population has the original size of the initial population and it
consists of
best,
middle, and
worst
and
are
vectors from the temporary population. The ,
user-selected weights. Their selection influences the algorithms
global search capability, as shown in Section IV.
E. Selection
In the selection step, vectors from the new trial population
are compared with the target (parent) vectors. When the trial
vector has a lower objective value compared to its parent it replaces one, otherwise the parent retains its position in the population. After the selection step is finished and the stopping criterion has not been met, the obtained population is forwarded
to the CPU cores. The process repeats all steps shown in Fig. 2
and Fig. 8 until the stopping criteria is met.
IV. SIMULATION AND RESULTS
The proposed PSADE was implemented in MATLAB environment using an Intel Core i7 3.5-GHz and 32-GB RAM

GLOTI et al.: PARALLEL SELF-ADAPTIVE DIFFERENTIAL EVOLUTION ALGORITHM FOR SOLVING SHORT-TERM HYDRO SCHEDULING PROBLEM

2353

TABLE I
RESULTS OBTAINED BY DIFFERENT METHODS FOR MODEL I IN COMPARISON WITH REAL DATA FROM SCADA

Number of function evaluations.

Water quantity per produced electrical energy unit

Fig. 9. Final agreement between the system demand and the production obtained by different optimization methods and applied to Model I.
TABLE II
OPTIMAL HYDRO SCHEDULING OBTAINED WITH PSADE FOR MODEL I

Deviation between total hydro generation obtained by PSADE and SCADA

computer. The performance of the proposed algorithm has been


tested on two test models.
A. Test Model I
For the optimization process, the real system demand data
from SCADA [35] have been used as the reference for the
Model I. The effectiveness of the PSADE algorithm is evident
from Table I, where the results obtained by different optimization methods are shown in comparison with data from
SCADA. The classic DE and SQP method dissatisfied the

system demand, but the PSADE algorithm has successfully


satisfied it in regard to the total water discharge in comparison
with SCADA. The final agreement between the system demand
and the production obtained by those algorithms is shown in
Fig. 9. The total production of all HPPs in each individual hour
obtained with the PSADE algorithm is compared in Table II
with the system demand from SCADA.
The stopping criterion for the optimization algorithm was the
maximal number of generations (1500). The weights
for forming the new trial vectors
by BmW strategy have an

2354

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 5, SEPTEMBER 2014

TABLE III
EXPERIMENTAL RESULTS, AVERAGED OVER TEN INDEPENDENT RUNS OF PROPOSED PSADE
WITH DIFFERENT SETS OF WEIGHTS FOR THE BMW SELECTION STEP FOR MODEL I

Number of function evaluations.

Water quantity per produced electrical energy unit

Fig. 10. Objective function values through generations obtained with the proposed PSADE algorithm and different control parameters (Model I) (
,
,
,
).

impact on the algorithms capabilities of reaching global solutions. Therefore the different sets of weights have been tested
and the results are shown in Table III. The best objective function value during the tests were obtained by weight sets
,
and
, and population size
.
By using these settings for Model I the result of the optimization
process satisfied system demand with more than 4.2
of the water in reservoirs saved in comparison with the manual
operation of the dispatching personnel. According to the average water quantity used per produced electrical energy unit,
the saved water quantity is equivalent to 176.2 MWh of energy.
The disagreement between production according to PSADE algorithm and demand from SCADA was negligible. The performance of the PSADE algorithm with the self-adaptive parameters (11) and (12) in comparison with the self-adaptive parameters from [25] and the set of various fixed control parameters
are shown in Fig. 10.
In order to verify the proposed method for parallelizing the
DE algorithm, the comparison between the proposed algorithm
PSADE and the other three methods from [30][32] are shown
in Fig. 11. The constant denotes migration rate, which determines how many of the best individuals from one sub-population will replace the same number of the worst individuals from
other subpopulations.
The constant defines the migration interval which denotes
that trough each generation, migration will be set up. Other two

Fig. 11. Objective function values through generations obtained with the proposed PSADE algorithm and different parallelization types (Model I).

TABLE IV
RESULTS OBTAINED BY DIFFERENT METHODS FOR MODEL II
COMPARISON WITH DATA FROM [5]

number of function evaluations.


energy unit

IN

water quantity per produced electrical

techniques of parallelization DE algorithm are used from [31]


and [32]. The best results are obtained by using the proposed
algorithm.
B. Test Model II
The results obtained by different optimization methods in
comparison with the system demand from [5] are shown in
Table IV. The optimal solution is obtained by the PSADE
algorithm, where the optimal hydro schedule shown in Table V
water in reservoirs, compared to
saved more than 228.000
[5] as shown in Table IV. This saving is equal to 200.5 MWh
of electrical energy. To evaluate performance of proposed
algorithm the standard test model combined of four HPPs,
has also been used. The reference system demand obtained

GLOTI et al.: PARALLEL SELF-ADAPTIVE DIFFERENTIAL EVOLUTION ALGORITHM FOR SOLVING SHORT-TERM HYDRO SCHEDULING PROBLEM

2355

TABLE V
OPTIMAL SCHEDULING OBTAINED WITH PSADE FOR MODEL II

TABLE VI
HYDRO STORAGE VOLUMES FOR MODEL II OBTAINED WITH OPTIMAL SCHEDULING,
TOTAL HYDRO GENERATION, SYSTEM DEMAND FROM [5] AND DEVIATION IN %

Deviation between total hydro generation obtained by PSADE and system demand [5]

Fig. 12. Objective function values through generations with different control
,
parameter of the proposed PSADE implemented on Model II (
,
,
).

from [5], where authors show water discharges and reservoirs


volumes, are used for comparison purposes in Table VI. The

final reservoir levels from [5] are also satisfied and shown in
Table VI. To optimize the hydro scheduling for Model II the
PSADE algorithm has also been applied. The comparison of
results obtained for different methods for setting the algorithms
control parameters is shown in Fig. 12. The comparisons of
different methods for parallelization [30][32] are shown in
Fig. 14. The maximum generation number for Model II was set
at 1200. The best objective function value was obtained with
weights set
,
and
, and
.
The weight selection impact for the novel BmW strategy is
shown in Table VII.
The classic DE control parameters for both models were set
,
,
, and strategy to DE/rand/1.
to:
The SQP method for both models has been used with MATLAB
Optimization Toolkit by default settings, where the different
random initial points have been used, and the best obtained optimization result for Model II is shown in Table IV. Comparison
of the final agreement between each the system demand and the
production obtained by PSADE, classic DE and SQP is shown
in Fig. 13, separately for each individual hour. The associated

2356

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 5, SEPTEMBER 2014

Fig. 13. Final agreement between the system demand and the production obtained by different optimization methods and applied to Model II.

TABLE VII
EXPERIMENTAL RESULTS, AVERAGED OVER TEN INDEPENDENT RUNS OF THE PROPOSED PSADE ALGORITHM
WITH DIFFERENT SETS OF WEIGHTS FOR THE BMW SELECTION STEP FOR MODEL II

Number of function evaluations.

Water quantity per produced electrical energy unit

Fig. 14. Objective function values through generations for different parallelization of PSADE and DE algorithm implemented on Model II.

reservoirs volumes at optimal hydro scheduling and the disagreements between total hydro energy production and system
demand in [5] are shown in Table VI.
V. CONCLUSION
The main goal of this research was to modify the standard
DE algorithm in order to improve its performance when applied to solve short-term hydropower scheduling optimization

problem. This means that the algorithm should be capable of


satisfying system demand and the fulfillment of a given scheduling plan, respectively within the context of the optimal production allocation by minimizing the used water quantity per
produced electrical energy unit. Therefore, the presented parallel self-adaptive differential evolution algorithm has been developed and tested on two different models, one real-parameter
HPP model and one standard test model mostly used in scientific
publications. The experimental results for both models showed
that the novel approach for DE algorithm parallelization along
with the proposed BmW strategy and the introduced correction
matrix, significantly improved the performance of the standard
DE algorithm. The BmW strategy is used in preselection step,
and it controls the balance between global search capability and
population diversity, while the correction matrix has a significant impact on the process of keeping or moving exceeding parameters into the allowable intervals. The proposed algorithm
also uses a novel approach for the self-adaptations of DE control parameters, which improves the algorithms performance.
The PSADE algorithm, therefore, outperformed the original algorithm as well as the other modified algorithms used in this
study. According to the obtained results it can be concluded
that the proposed algorithm enables significant savings when
applied to hydro scheduling problem. It is also evident that the
PSADE convergence time is appropriate for the usage of the algorithm for real-world and practical applications, respectively.

GLOTI et al.: PARALLEL SELF-ADAPTIVE DIFFERENTIAL EVOLUTION ALGORITHM FOR SOLVING SHORT-TERM HYDRO SCHEDULING PROBLEM

REFERENCES
[1] J. Zhang, J. Wang, and C. Yue, Small population-based particle swarm
optimization for short-term hydrothermal scheduling, IEEE Trans.
Power Syst., vol. 27, no. 1, pp. 142152, Feb. 2012.
[2] S. O. Orero and M. R. Irving, A genetic algorithm modeling framework and solution technique for short term optimal hydrothermal
scheduling, IEEE Trans. Power Syst., vol. 13, no. 2, pp. 501518,
May 1998.
[3] S. Soares, C. Lyra, and H. Tavares, Optimal generation scheduling
of hydrothermal power system, IEEE Trans. Power App. Syst., vol.
PAS-99, no. 3, pp. 11071118, Mar. 1980.
[4] K. K. Mandal and N. Chakraborty, Differential evolution technique-based short-term economic generation scheduling of hydrothermal system, Electr. Power Syst. Res., pp. 19721979, 2008.
[5] L. Lakshminarasimman and S. Subramanian, Short-term scheduling
of hydrothermal power power system with cascaded reservoirs by using
modified differential evolotuion, Proc. Inst. Electr. Eng.Gener.,
Transm. Distrib., vol. 153, pp. 693700, 2006.
[6] N. Sinha, R. Chakrabarti, and P. K. Chattopadhyay, Fast evolutionary
programming techiques for short-term hydrothermal scheduling,
IEEE Trans. Power Syst., vol. 18, no. 1, pp. 214220, Feb. 2003.
[7] I. A. Farhat and M. E. El-Hawary, Optimization methods applied for
solving the short-term hydrothermal coordination problem, Electr.
Power Syst. Res., pp. 13081320, 2009.
[8] S. Al-Agtash and S. Renjeng, Augmented Lagrangian approach to
hydro-thermal scheduling, IEEE Trans. Power Syst., vol. 13, no. 4,
pp. 13921365, Nov. 1998.
[9] H. Yan, P. B. Luh, X. Guan, and P. M. Rogan, Scheduling of hydrothermal power systems, IEEE Trans. Power Syst., vol. 8, no. 4,
pp. 13921400, Nov. 1993.
[10] C. Li, E. Hsu, A. J. Svoboda, C. Tseng, R. Tseng, and R. B. Johnson,
Hydro unit commitment in hydro-thermal optimization, IEEE Trans.
Power Syst., vol. 12, no. 2, pp. 764769, May 1997.
[11] O. Nilsson and D. Sjelvgren, Mixed-integer programming applied to
short-term planning of a hydro-thermal system, IEEE Trans. Power
Syst., vol. 11, no. 1, pp. 281286, Feb. 1996.
[12] A. J. Conejo, J. M. Arroyo, J. Contreras, and F. A. Villamor, Selfscheduling of a hydro producer in a pool-based electricity market,
IEEE Trans. Power Syst., vol. 17, no. 4, pp. 12561272, Nov. 2002.
[13] S. Chang, C. Chen, C. Fong, and L. P. B., Hydroelectric generation
scheduling with an effective differential dynamic programming algorithm, IEEE Trans. Power Syst., vol. 5, no. 3, pp. 737743, Aug. 1990.
[14] J. Tang and P. B. Luh, Hydrothermal scheduling via extended differential dynamic programming and mixed coordination, IEEE Trans.
Power Syst., vol. 10, no. 4, pp. 20212027, Nov. 1995.
[15] E. C. Finardi, E. L. Silva, and C. Sagastizabal, Solving the unit commitment problem of hydro plants via Lagrangian Relaxation and Sequential Quadratic Programming, Computational Appl. Math, vol. 24,
pp. 317341, 2005.
[16] X. Yuan, Y. Wang, J. Xie, X. Qi, H. Niw, and A. Su, Optimal selfscheduling of hydro producer in the electricity market, Energy Convers. Manag., vol. 51, pp. 25232530, 2009.
[17] P. Chen, L. Chen, A. Liu, and H. Chen, Application of particle swarm
optimization to hydro generation scheduling, in Proc. Int. Conf. Energy and Environ. Technol., Guangxi, 2009, pp. 541544.
[18] E. Gil, J. Bustos, and H. Rudnick, Short-term hydrothermal generation scheduling model using a genetic algorithm, IEEE Trans. Power
System, vol. 18, pp. 12561264, 2003.
[19] C. E. Zoumas, A. G. Bakirtzis, J. B. Theocharis, and V. Petridis, A
genetic algorithm solution approach to the hydrothermal coordination
problem, IEEE Trans. Power System, vol. 19, pp. 13561364, 2004.
[20] P. Chen and H. Chang, Genetic aided scheduling of hydraulically
coupled plants in hydro-thermal coordination, IEEE Trans. Power
System, vol. 11, pp. 975981, 1996.
[21] P. C. Yang, H. T. Yang, and C. L. Huang, Scheduling short-term hydrothermal generation using evolutionary programming techniques,
Proc. Inst. Electr. Eng.Gener., Transm. Distrib., vol. 143, pp.
371376, 1996.
[22] R. Storn and K. Price, Differential evolutionA simple and efficient
Heuristic for global optimization over continuous spaces, J. Global
Optimization, pp. 341359, 1997.
[23] X. Yuan, Y. Zhang, L. Wang, and Y. Yuan, An enhanced differential
evolution algorithm for daily optimal hydro generation scheduling,
Comput. Math. With Applic., pp. 24582468, 2008.

2357

[24] Y. Li and J. Zuo, Optimal scheduling of cascade hydropower system


using grouping differential evolution algorithm, in Proc. Int. Conf.
Comput. Sci. Electron. Eng., 2012, pp. 625629.
[25] J. Brest, S. Greiner, B. Boskovic, M. Mernik, and V. Zumer, Selfadapting control parameters in differential evolution: A comparative
study on numerical benchmark problems, IEEE Trans. Evol. Computat., vol. 10, no. 6, pp. 646657, Dec. 2006.
[26] R. Caponetto, L. Fortuna, S. Fazzino, and M. G. Xibilia, Chaotic sequences to improve the performance of evolutionary algorithms, IEEE
Trans. Evol. Computat., vol. 7, no. 3, pp. 289304, Jun. 2003.
[27] R. Gamperle, S. D. Muller, and P. Koumoutsakos, a parameter study
for differential evolution, in Proc. Int. Conf. Adv. Intell. Syst., Evol.
Computat., 2002, pp. 293298.
[28] A. K. Qin, V. L. Huang, and P. N. Suganthan, Differential evolution
algorithm with strategy adaptation for global numerical optimization,
IEEE Trans. Evol. Computat., vol. 13, no. 2, pp. 398417, Apr. 2009.
[29] J. Cheng and G. Zhang, Improved differential evolutions using a dynamic differential factor and population diversity, in Proc. Int. Conf.
Artif. Intell. Computat. Intelligence, 2009, pp. 402406.
[30] E. Alba and M. Tomassini, Parallelism and evolutionary algorithms,
IEEE Trans. Evol. Computat., vol. 6, no. 5, pp. 443462, Oct. 2002.
[31] D. Zaharie and D. Petcu, Parallel implementation of multi-population
differential evolution, in Proc. 2nd Workshop Concurrent Inf. Process.
Computing, Romania, 2003, pp. 223232.
[32] C. Y. Wu and K. Y. Tseng, Truss structure optimization using adaptive
multi-population differential evolution, Structural Multidisciplinary
Optimization, vol. 4, pp. 575590, 2010.
[33] C. H. Liang, C. Y. Chung, K. P. Wong, and X. Z. Duan, Parallel optimal reactive power flow based on cooperative co-evolutionary differential evolution and power system decomposition, IEEE Trans. Power
Syst., vol. 22, pp. 249257, 2007.
[34] A. M. Jubril, O. A. Komolafe, and K. O. Alawode, Solving multiobjective economic dispatch problem via semidefinite programming,
IEEE Trans. Power Syst., vol. 28, no. 3, pp. 20562064, Aug. 2013.
[35] A. G. Bruce, Reliability analysis of electric utility SCADA systems,
IEEE Trans. Power Syst., vol. 22, no. 1, pp. 249257, Feb. 2007.
Arnel Gloti (S12) was born in Slovenj Gradec,
Slovenia, in 1987. He received the B.S. degree
in electrical engineering from the University of
Maribor, Maribor, Slovenia, in 2011, where he is
currently working toward the Ph.D. degree at the
Institute for Power Engineering.
His current research interest includes multi-objective optimization in the application of evolutionary
techniques for electrical power system.

Adnan Gloti was born in Slovenj Gradec, Slovenia,


in 1980. He received the B.S. and Ph.D. degrees
in electrical engineering from the University of
Maribor, Maribor, Slovenia, in 2006 and 2011,
respectively.
Since 2011, he has been with the Holding
Slovenske elektrarne d.o.o., Ljubljana, Slovenia,
within the R&D Department. His special fields of interest are electrical power network and optimization
methods.

Peter Kitak (M09) was born in Ptuj, Slovenia,


in 1974. He received the B.S. and Ph.D. degrees
in electrical engineering from the University of
Maribor, Maribor, in 1999 and 2006, respectively.
Since 2002, he has been a Faculty Member with
the University of Maribor, Maribor, Slovenia. His
special fields of interest are modeling insulators and
switchgears, numerical calculation of electromagnetic fields, and optimization methods.

2358

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 5, SEPTEMBER 2014

Joe Pihler (M09) was born in Ptuj, Slovenia, in


1955. He received the B.S., M.S., and Ph.D. degrees
in electrical engineering from the University of
Maribor, Maribor, Slovenia, in 1978, 1991, and
1995, respectively.
Currently he is a Full Professor with the University of Maribor, Maribor, Slovenia, where his special
fields of interest are switching devices and optimization methods.

Igor Tiar (M00) was born in Maribor, Slovenia, in


1949. He received the B.S. degree in electrical engineering from the University of Ljubljana, Ljubljana,
Slovenia, in 1975, and the M.S. and Ph.D. degrees
from the University of Maribor, Maribor, Slovenia,
in 1981 and 1993, respectively.
In 1977, he joined the University of Maribor,
Maribor, Slovenia, as an Assistant, where he is
currently a Full Professor. His special fields of
interest are also optimization methods.