Mathematics and Computers in Simulation 59 (2002) 319–327

A hybrid chaotic genetic algorithm for short-term

hydro system scheduling
Xiaohui Yuan∗ , Yanbin Yuan, Yongchuan Zhang
Department of Hydropower and Automation Engineering, Huazhong University of Science and Technology,
Wuhan, Hubei 430074, PR China
Received 1 April 2001; accepted 10 May 2001

Abstract
This paper proposes a novel hybrid chaotic genetic algorithm (HCGA) to solve the short-term generation schedul-
ing of hydro system. The integration of chaotic sequence and genetic algorithm with a new self-adaptive error
back-propagation mutation operator are developed, which can overcome premature and increase the convergence
speed. Simulation results have demonstrated that the proposed approach is feasible and effective for the applications.
© 2002 IMACS. Published by Elsevier Science B.V. All rights reserved.
Keywords: Chaos; Genetic algorithm; Short-term optimal dispatch; Hydroelectric system

1. Introduction

From an economic point of view, an efficient hydropower scheduling plays an important role in the
planning and operation of a power system. The purpose of short-term hydro system scheduling is to
find the optimum hourly generation of hydro units by utilizing the limited water resource in a schedule
horizon so as to optimize the total benefit of hydro generated energy, while satisfying various constraints.
Cascaded hydroelectric plants are related to each other in both power and hydraulic aspects, so the
short-term optimal dispatch of cascaded hydroelectric plants is a large-scale, dynamic with time delay
and complicated constrained nonlinear optimization problem.
Various approaches have been proposed to solve the short-term optimal schedule problem of hydro
generation system. The main methods include dynamic programming [1], maximum principle [2], net-
work flow method and linear programming [3,4], nonlinear programming [5], functional analysis [6],
mathematical decomposition and artificial neural networks [7,8]. These methods have one or another
drawback such as dimensionality difficulty, large memory requirement or inability to handle a nonlinear
cost function.
∗
Corresponding author. Tel.: +86-2787-5527-67; fax: +86-2787-5439-92.
E-mail address: [email protected] (X. Yuan).

0378-4754/02/$ – see front matter © 2002 IMACS. Published by Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 8 - 4 7 5 4 ( 0 1 ) 0 0 3 6 3 - 9
320 X. Yuan et al. / Mathematics and Computers in Simulation 59 (2002) 319–327

Nomenclature
htj net head of hydro plant j at time t
Ijt natural inflow into reservoir j at time t
j plant index, j = 1, 2, . . . , N
N number of hydro plants
Pjt power generation of hydro plant j at time t
Pj min minimum power generation of hydro plant j
Pj max maximum power generation of hydro plant j
Qtj water discharge of hydro plant j at time t
Qj min minimum water discharge of hydro plant j
Qj max maximum water discharge of hydro plant j
Sjt water spillage of hydro plant j at time t
t time index, t = 1, 2, . . . , 24
Vjt water volume of reservoir j at the end of hour t
Vj min minimum water volume of reservoir j
Vj max maximum water volume of reservoir j
begin
Vj initial storage volume of reservoir j
Vjend final storage volume of reservoir j
τ j −1 water travel time from plant j − 1 to plant j

In recent years, a new optimization method called genetic algorithm (GA) has aroused intense interest,
due to the flexibility, versatility and robustness in solving optimization problems, which above conven-
tional optimization methods find difficult. It also has been given much attention to apply the short-term
generation scheduling of hydro system [9]. However, there exist some flaws on GA. For example, it tends
to converge prematurely and takes a large number of iterations to reach the global optimal solution, the
optimization may get stuck at a local optimum. In order to overcome these flaws, it is necessary to adopt
some improvements on GA to speed up the convergence and heighten the effectiveness of GA.
The chaotic optimization method based chaos theory has recently come into being as a new type of
random search algorithm. Taking advantage of the universality, randomicity and sensitivity dependence
on initial conditions, it is more likely to acquire the global optimum solution. Although the chaotic op-
timization method reduces the search space by carrier wave and thus speeds up search, it makes no us of
the experiential information previously acquired. As a result, search effect of chaotic optimization has its
own limitation.
In order to overcome the shortcomings of both chaotic optimization method and GA, this paper presents
a new optimization approach, i.e. hybrid chaotic genetic algorithm (HCGA) by integration of chaotic
sequence and GA with a new self-adaptive error backpropagation mutation operator. Taking advantage
of the sensitivity dependence on initial conditions makes it unlikely for the new population to trap
simultaneously into the same local solution, while keep evolving without break. The optimum solution
will be updated and premature convergence is restrained, thanks to the intervention of chaotic sequence
and the introduction of new mutation operator. Finally, application of this algorithm to the short-term
scheduling of cascaded hydroelectric system, simulation results have demonstrated the feasibility and
effectiveness of the proposed approach for the practical applications.
 X. Yuan et al. / Mathematics and Computers in Simulation 59 (2002) 319–327 321

2. Mathematical model for short-term cascaded hydro system scheduling

To formulate the problem and its solution mathematically, the following notation used in this paper is
first introduced:

2.1. Objective function and constraints

Suppose the object studied is mainly power generation for cascaded hydroelectric systems. The short-
term hydro scheduling problem can be stated as to find the water release from each reservoir and through
each powerhouse over all the planning time intervals so as to maximize the total power generation
while satisfying various constraints. Mathematically, the generation scheduling problem may be stated
as follows:
N


24
max f (V , Q) = max Pjt (Qtj , Vjt , Vjt+1 ) (1)
j =1 t=1

subject to the following constraints.

1. Hydro plant power generation limits
Pj min ≤ Pjt ≤ Pj max (2)
2. Hydro plant discharge limits
Qj min ≤ Qtj ≤ Qj max (3)
3. Reservoir storage volumes limits
Vj min ≤ Vjt ≤ Vj max (4)
4. Initial and terminal reservoir storage volumes
begin
Vj1 = Vj , Vj25 = Vjend (5)
5. Water dynamic balance equation with travel time
t−τ t−τ
Vjt = Vjt−1 + Ijt + Qj −1j −1 + Sj −1j −1 − Qtj − Sjt (6)

2.2. Power generation characteristics

Power output of a hydro plant is a nonlinear function of effective reservoir head, turbine discharge
and generating efficiency Pjt = f (ηjt , Qtj , Vjt ). By approximating the plant generating characteristic at
maximum head by a concave fourth degree polynomial function of turbine discharge and by adding a
linear correction term for head losses, the following model is obtained:
Pjt = a0j + a1j Qtj + a2j (Qtj )2 + a3j (Qtj )3 + a4j (Qtj )4 − Hjt (7)
Head losses Hjt are modeled as a function with maximum head as reference
Hjt = βj htj Qtj (8)
322 X. Yuan et al. / Mathematics and Computers in Simulation 59 (2002) 319–327

In case power production depends on the upper reservoir level only, then
htj = hmax j − htj (9)

If power production depends on upper reservoir level as well as the level of the neighboring downstream
reservoir, htj is evaluated as

htj = hmax j − htj + hti − hmin i (10)

where a0j , a1j , a2j , a3j , a4j , βj are constants, hmax j denote maximum water head at reservoir j, hmin i
denote minimum water head at reservoir i, index i relates to the downstream reservoir.
Reservoir water heads are approximated as piecewise linear functions of the reservoir storage.
htj = kjt + αjt Vjt = min{kji + αji Vji }, i = 1, 2, . . . (11)

3. Hybrid chaotic genetic algorithm (HCGA)

3.1. Chaotic sequence

Chaos is apparently an irregular motion, seemingly unpredictable random behavior exhibited by a

deterministic nonlinear system under deterministic conditions. Roughly speaking, a nonlinear system is
said to be chaotic if it exhibits sensitive dependence on initial conditions and has an infinite number of
different periodic responses. We must distinguish here between so-called random and chaotic motions.
The former is reserved for problems in which we truly do not know the input forces or we only know
some statistical measures of the parameters. Chaotic is reserved for those deterministic problems for
which there are no random or unpredictable inputs or parameters. It may happen that small differences
in the initial conditions produce very great ones in the final phenomena. A small error in the former will
produce an enormous error in the latter.
The chaotic sequence can usually be produced by the following well-known one-dimensional logistic
map defined by
tk+1 = λtk (1 − tk ); tk ∈ (0, 1), k = 0, 1, 2, . . . (12)

where tk is the value of the variable t at the kth iteration, tk in the interval [0,1], λ is a so-called bifurcation
parameter of the system.
For certain values of the parameter λ, of which λ = 4 is one, the above system exhibits chaotic behavior.

3.2. Implementation steps of hybrid chaotic genetic algorithm

Consider the following optimization problem:

min f (xi ); xi ∈ [ai , bi ], i = 1, 2, . . . , n (13)

where xi is an unknown decision variable, f(x1 , x2 , . . . , xn ) is the objective function, ai and bi are the
bounds of xi , i is the variable numbers index, i = 1, 2, . . . , n.
 X. Yuan et al. / Mathematics and Computers in Simulation 59 (2002) 319–327 323

3.2.1. Initialization
The m initial values yj,i (0) ∈ [0, 1] are randomly generated in the range 0 and 1. And then use above
logistic map (12) to produce l chaotic variables.

{yj,i (k)}; j = 1, 2, . . . , m, i = 1, 2, . . . , n, k = 1, 2, . . . , l

3.2.2. Generating family members xj,i (k) from the chaotic variables yj,i (k)
xj,i (k) = ai + (bi − ai )yj,i (k) (14)

3.2.3. Floating-point number encoding

Since the short-term optimization dispatch of hydro system is a large-scale complex nonlinear con-
strained optimization problem, the floating point numbers encoding technique is appropriate to ap-
ply. In the floating-point numbers representation each chromosome vector is coded as a vector of
floating point numbers of the same length as the solution vector. Each element is initially selected
as to be within the desired domain. In addition, the floating-point numbers representation is capa-
ble of representing quite large domains (or cases of unknown domains). Also, it is easier to handle
constraints.
Considering the features of short-term generation scheduling of cascaded hydroelectric system, encoded
parameters adopt plants’ water discharge Qtj (j = 1, 2, . . . , N; t = 1, 2, . . . , 24). Each chromosome
string contains N × 24 genes to represent the solution for the hourly water discharge schedule of the N
plants in 24 h scheduling horizon.

3.2.4. Selection
According to the fitness value f (xj,i (k)), an individual of the highest fitness value which had been
seek out from each family has chance to become the parent individual.

3.2.5. Crossover
New parent individuals are produced by crossover operation. Uniform arithmetical crossover are usually
used for the floating-point numbers encoding individuals, i.e. the offspring individuals are produced by the
linear combination of the parent individuals. Suppose two parent individuals that have been selected from
the tth generation population are svt = v1 , v2 , . . . , vn , swt = w1 , w2 , . . . , wn , respectively, offspring
individuals that are produced by parent individuals are

svt+1 = βsvt + (1 − β)swt , swt+1 = βswt + (1 − β)svt (15)

where β is a constant between 0 and 1.

3.2.6. Mutation
Mutation is an effective operator to increase and retain the population diversity, and is meanwhile an
efficient method to escape the local optimum solution and to overcome the premature convergence. A
large scale of mutation is good for acquiring the optimum solution in extensive search, but the search is
rough and the solution precision is poor. On the other hand, if the precision is satisfactory, the solution
will be got stuck at a local optimum or take too long time to converge. In view of overcoming these
flaws, this paper adopts the self-adaptive error backpropagation mutation operator. The mutation rules
324 X. Yuan et al. / Mathematics and Computers in Simulation 59 (2002) 319–327

are shown as follows [10]:

xji (k + 1) = xji (k) + ηxji (k) + αsxij (k)
xji (k) = [xjbest (k) − xji (k)] |N(0, 1)|
sxij (k + 1) = ηacci (k)xji (k) + αsxij (k) (16)

1 if the current update has improved cost
acci (k) =
0 otherwise
where η is the learning rate, α the momentum rate, N(0, 1) denotes the normal distribution, |·| denotes
an absolute value, xj,i (k) is the amount of change in an individual, sxij (k) is the evolution tendency of
previous evolution, xji (k) is the jth variable of an ith individual at the kth generation.

3.2.7. Fitness function and constraint handling

GA adopts fitness value to direct the search. It means, for an individual with larger fitness value, it
will appear in the next generation with larger possibility. In HCGA the value of Qtj satisfies its definition
domain, so the water discharge constraint function is automatically satisfied. Other constraints may be
handled through penalty function. Then the constrained optimization problem may be transformed to
unconstrained optimization. After some treatment for constraints, the following objective functions can
be obtained:


3
min F = −f + σi φi (17)
i=1

N

 25 
φ1 = V − V 1  (18)
j j
j =1

N


24
 t 
φ2 = V − V lim 
j j
j =1 t=1


 V if Vjt > Vj max
 j max
Vjlim = Vj min if Vjt < Vj min (19)


Vt otherwise
j

N


24
 t 
φ3 = P − P lim 
j j
j =1 t=1


 P if Pjt > Pj max
 j max
Pjlim = Pj min if Pjt < Pj min (20)


Pt otherwise
j

where σ i and φ i are the ith penalty coefficient and penalty functions, respectively.
 X. Yuan et al. / Mathematics and Computers in Simulation 59 (2002) 319–327 325

The choice of the penalty term can be significant, for, if the penalty term is too harsh, infeasible strings
that carry useful information for the GA, but lie outside the feasible region will largely be ignored and
their information lost while if the penalty term is not strong enough, the GA may search only among
infeasible strings, and miss out on the feasible solutions. Therefore, this paper chooses penalty factor
σi = 1/T , T = αT , where α is a constant between zero and one. σ i chosen this way absorbs the idea of
simulated annealing and forces T to decrease gradually, that is, σ i increases gradually which thus ensures
the satisfaction of constraints with the process of the evolution.
Aiming at the above objective functions, the fitness function FIT may be stated as follows:
FIT = Fmax − F + K(Fmax − Fmin ) (21)
where Fmax , Fmin are, respectively, the maximum and the minimum of the objective functions for present
populations; K the control parameter and its value choice ranges from 0.01 to 0.1.

4. Numerical experiments

To investigate the effectiveness of the algorithms described above, in this section we present two
examples to illustrate the application of the algorithms.

4.1. Example 1

This example investigates the convergence speed and solution accuracy of the proposed approach on
the following well-known benchmark test function.
The Rosenbrock function is given in (22),
min f = 100 × (x1 − x2 )2 + (1 − x1 )2 ; −2.048 ≤ xi ≤ 2.048, i = 1, 2 (22)
The global optimum of the Rosenbrock function resides inside a long, narrow, and parabolic-shaped
flat valley, which is difficult to follow.
Fig. 1 shows the evolutionary process of the proposed method.
From Fig. 1, the proposed method converged to the global optimum faster than the classical GA
when applied to the above Rosenbrock function. While GA found a solution with an objective function
value near 10−5 after 1000 generations, the proposed algorithm converged to a solution around 10−10

Fig. 1. Evolutionary process.

326 X. Yuan et al. / Mathematics and Computers in Simulation 59 (2002) 319–327

Fig. 2. Hourly hydro plant power generations.

Fig. 3. Hourly hydro plant discharge trajectories.

within 200 generations. It is because that in HCGA exists chaotic sequence and a new self-adaptive error
back-propagation mutation operator to direct search in the evolutionary process, the convergence speed
is very faster than classical GA.

4.2. Example 2

In this example, a hydro system with two reservoirs is used to demonstrate the feasibility and effective-
ness of the proposed approach. The upstream reservoir is long-term adjustment while the downstream
reservoir is daily adjustment reservoirs. The scheduling period is 24 h, with one hour time intervals. Each
experiment was run 20 times, starting with a different random initial population, the best result is selected
as the final result. The hourly hydro plant power generation and discharge trajectories are given in Figs. 2
and 3, respectively.

5. Conclusion

This paper presents a new HCGA. HCGA is thus formed by introducing the chaotic sequence and
presenting a new self-adaptive error back-propagation mutation operator in the evolutionary process of
GA. The algorithm not only retains the generality of GA, but also improves convergence speed and
solution accuracy. The proposed method is applied to the short-term optimal dispatch of cascaded hydro
 X. Yuan et al. / Mathematics and Computers in Simulation 59 (2002) 319–327 327

system. Simulation results show that the proposed approach is feasible and effective for the large-scale
constrained nonlinear optimization problem.

References

[1] L. Christiano, R. Luiz, A multiobjective approach to the short-term scheduling of a hydroelectric power system, IEEE
Trans. PAS 10 (4) (1995) 1750–1754.
[2] M. Papageorgiou, Optimal multi reservoir network control by the discrete maximum principle, Water Resources Res. 21 (2)
(1985) 1824–1830.
[3] M.R. Piekutowski, T. Litwinowicz, R.J. Frowd, Optimal short-term scheduling for a large-scale cascaded hydro system,
IEEE Trans. PAS 9 (2) (1994) 805–811.
[4] Q. Xia, N. Xiang, S. Wang, B. Zhang, M. Huang, Optimal daily scheduling of cascaded plants using a new algorithm of
non-linear minimum cost network flow concept, IEEE Trans. PWRS 3 (3) (1988) 929–935.
[5] H. Habibollahzadeh, J.A. Bubenko, Application of decomposition techniques to short term operation planning of
hydro-thermal power system, IEEE Trans. PWRS 1 (1) (1986) 41–47.
[6] S.A. Soliman, G.S. Christensen, Application of functional analysis to optimization of variable head multi reservoir power
system for long term regulation, Water Resources Res. 22 (6) (1986) 852–858.
[7] N. Ernan, G. Xiaohong, L. Renhou, Scheduling hydrothermal power systems with cascaded and head-dependent reservoirs,
IEEE Trans. PAS 14 (3) (1999) 1127–1132.
[8] R. Naresh, J. Sharma, Hydro system scheduling using ANN approach, IEEE Trans. PAS 15 (1) (2000) 388–395.
[9] S.O. Orero, M.R. Irving, A genetic algorithm modeling framework and solution technique for short term optimal
hydrothermal scheduling, IEEE Trans. PAS 13 (2) (1998) 501–518.
[10] C. Doo-Hyun, O. Se-Young, A new mutation rule for evolutionary programming motivated from backpropagation learning,
IEEE Trans. Evol. Comput. 4 (2) (2000) 188–190.

Xiaohui Yuan was born in Hubei, China, on 18 December 1971. He received his BS and MS degrees in hydropower engineering
from Huazhong University of Science and Technology (HUST) in 1993 and 1996, respectively. He is currently a PhD student
in the hydropower and automation engineering department in HUST. His current research interests include the application of
evolutionary computation techniques to optimal short-term scheduling in hydro system.

Yanbin Yuan was born in Hubei, China, on 1 January 1970. He received his BS, MS and PhD degrees in earth probing and
information technology from China University of Geosciences in 1992, 1995 and 2000, respectively. He is currently a post
doctor in the hydropower and automation engineering department in Huazhong University of Science and Technology. His
current research interests include the application of geographic information system (GIS) techniques to economical operation in
hydro system and digital valley.

Yongchuan Zhang was born in Henan, China on 1 March 1935. He is presently an Academician of Chinese Academy of
Engineering and a professor in the hydropower and automation engineering department in HUST. His major areas of research
include digital cascaded hydroelectric system and its optimal dispatching.