Professional Documents
Culture Documents
Parameter Estimation of The Nonlinear Muskingum Flood-Routing Model Using A Hybrid Harmony Search Algorithm
Parameter Estimation of The Nonlinear Muskingum Flood-Routing Model Using A Hybrid Harmony Search Algorithm
net/publication/235994718
CITATIONS READS
78 327
3 authors:
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Research on sustainability in water, food, environment, energy & infrastructure View project
All content following this page was uploaded by Halil Karahan on 29 May 2014.
Abstract: In this paper, a hybrid harmony search (HS) algorithm is proposed for the parameter estimation of the nonlinear Muskingum
model. The BFGS algorithm is used as local search algorithm with a low probability for accelerating the HS algorithm. In the proposed
technique, an indirect penalty function approach is imposed on the model to prevent negativity of outflows and storages. The proposed
algorithm finds the global or near-global minimum regardless of the initial parameter values with fast convergence. The proposed algorithm
found the best solution among 12 different methods. The results demonstrate that the proposed algorithm can be applied confidently to
estimate optimal parameter values of the nonlinear Muskingum model. Moreover, this hybrid methodology may be applicable to any con-
tinuous engineering optimization problems. DOI: 10.1061/(ASCE)HE.1943-5584.0000608. © 2013 American Society of Civil Engineers.
CE Database subject headings: Flood routing; Parameters; Hydrologic models; Optimization; Algorithms.
Author keywords: Flood routing; Parameters; Hydrologic models; Optimization.
Introduction at close intervals are not available. In this case, lumped hydrologic
flood-routing models may be used because they can cope with
Flood routing is required for proper management and design sparse spatial data, allowing the estimation of the outflow discharge
of many environmental and water resource projects (Singh and as a function of the inflow discharge and of the geometric and
Scarlatos 1987). Flow routing may be considered an analysis to hydraulic properties of the selected river channel. In general,
trace the flow through a hydrologic system, given the input. The two types of hydrologic models for flood routing can be used (Reed
difference between lumped and distributed system routing is that 1984). The first is based on the conservation of mass and uses a
in a lumped system model, the flow is calculated as a function conceptual relationship between storage and discharge in place
of time alone at a particular location; whereas in a distributed sys- of the dynamic equation. The Muskingum method is an example.
tem, routing the flow is calculated as a function of space and time The second is established on the theory of linear systems employ-
throughout the system. Routing by lumped system methods is ing convolution (Keefer and McQuivey 1974; Singh 1988).
sometimes called hydrologic routing, and routing by distributed The hydraulic methods are typically more accurate than hydro-
system methods is sometimes referred to as hydraulic routing logic methods, but hydraulic methods are more complicated
(Chow et al. 1988).
than hydrologic methods. The statement is generally true if one
Hydraulic methods of routing involve the numerical solutions of
reduces the hydrologic model to a linear time invariant lumped
either the convective-diffusion equations or the one-dimensional
model. Nonetheless, several hydrologic models are derived by in-
Saint-Venant equations of gradually varied unsteady flow in open
tegrating in space the original hydraulic point equations, and they
channels. A great deal of literature is available on the flood-routing
can be viewed as a finite-element approach to the integration of the
process. For instance, the Hydrologic Engineering Center’s River
hydraulic differential equations. This is the case of the well-known
Analysis System (HEC-RAS) [U.S. Army Corps of Engineering
Muskingum-Cunge (MC) model (Cunge 1969), which although
(USACE) 2002] and MIKE-11 [Danish Hydrological Institute
(DHI) 2003] are able to solve the full Saint-Venant equations that based on a linear Muskingum allows the parameters to vary in time
represent the most complete approach to describe flow regimes according to the average flow characteristics in the reach, in such a
in natural channels. However, these models are not suitable to way as to mimic the behavior of a hydraulic parabolic model. More
serve the purpose of flood routing in places for which detailed recently, Todini (2007) corrected this model to preserve the mass
topographical surveys of channel cross sections and roughness and to correctly reproduce the Muskingum equations. The new ap-
proach, the Muskingum-Cunge-Todini (MCT), was shown to reach
1
Professor, Dept. of Civil Engineering, Pamukkale Univ., TR-20017, the same level of approximation of more sophisticated hydraulic
Denizli, Turkey (corresponding author). E-mail: hkarahan@pau.edu.tr models over a wide range of flow regimes. In MC or MCT, the
2
Lecturer, Dept. of Civil Engineering, Pamukkale Univ., TR-20017, model parameters do not require an estimation procedure (they
Denizli, Turkey. all depend on the friction coefficient, which is the only unknown
3
Assistant Professor, Dept. of Energy IT, Gachon Univ., Seongnam, element), while they are updated in time as a function of the mean
South Korea. state variables in the river reach (Todini 2007).
Note. This manuscript was submitted on May 12, 2011; approved on
On these bases, the hydrologic models, relatively simple to im-
February 13, 2012; published online on February 15, 2012. Discussion per-
iod open until August 1, 2013; separate discussions must be submitted for plement and reasonably accurate, have to be preferred in practical
individual papers. This paper is part of the Journal of Hydrologic Engi- applications (Haktanir and Ozmen 1997). The Muskingum method
neering, Vol. 18, No. 3, March 1, 2013. © ASCE, ISSN 1084-0699/2013/ is a simple flood-routing technique widely used for field applica-
3-352-360/$25.00. tions (Karahan 2009).
x11 x12 · ·· x1N−1 obtain the search direction, where d = search direction vector.
6 7
6 x21 x22 · ·· x2N−1 x2N fðX2 Þ 7 4. Compute optimum step size αk ¼ α to minimize
6 7
6 . .. .. .. .. .. 7 gðXðkÞ þ αdðkÞ Þ.
HM ¼ 6 6 . . . . . . .
7
7 5. Update the design as Xðkþ1Þ ¼ XðkÞ þ αdðkÞ .
6 7
6 xHMS−1 xHMS−1 · ·· xHMS−1 xHMS−1 fðXHMS−1 Þ 7 6. Update the Hessian approximation for the cost function as
4 1 2 5
N−1 N
HMS Hðkþ1Þ ¼ HðkÞ þ DðkÞ þ EðkÞ
HMS
x1 x2HMS HMS
· ·· xN−1 xN fðXHMS
Þ
ð5Þ where the correction matrices DðkÞ and EðkÞ are given as
5. Calculate the next outflow (Otþ1 ) using Eq. (6). If the next
outflow has a negative value, apply to the penalty factor by Application to Data Set Given by Wilson
using Eq. (9b).
6. Repeat steps 2–5 for all times. The parameter estimation technique for the nonlinear Muskingum
model using HS-BFGS is applied to the data set given by Wilson
(1974). The data reported by Wilson are known to present a non-
Numerical Applications linear relationship between weighted discharge and storage
(Al-Humoud and Esen 2006) and are used extensively in the liter-
To fairly test the reliability of the proposed HS-BFGS procedure ature as a benchmark problem. To determine the suitable HMS and
and to compare it with other different methods available in the lit- pc values, a sensitivity analysis was carried out using a thorough
erature, two example are solved through the developed model with data set given by Wilson with the pc ¼ ð0; 0.005; 0.01; 0.05Þ for
the data given by Wilson (1974) and River Wye December 1960 each HMS value. Statistical values (best, worst, mean, and standard
Flood data [Natural Environment Research Council (NERC) deviation of the SSQ values) of the parameters, SSQs, iteration
1975]. The objective function to be minimized within the HS number, and CPU times obtained by different pc values for each
algorithm ðfÞ is the sum of the square of the deviations between HMS value are presented in Tables 1–3. The number of successful
observed and computed outflows as follows: experiments with the different HMS and pc values are given in
X Table 4.
Minimize∶ SSQ ¼ ½Ot − Ôt ðK; χ; mÞ2 ð10Þ As is shown in Tables 1 and 4, the global optimum is not
t
obtained for HMS ¼ 5 when a pure HS algorithm is used
ðpc ¼ 0Þ. In this case, the best SSQ value is the near-global opti-
where Ot denotes the observed outflow; Ôt denotes the computed
mum. However, the global optimum is found when HS-BFGS is
outflow; and K; χ; m = design variables. The ranges of the
used ðpc ≠ 0Þ. The global optimum is found 93 times for
three parameters used in the applications are selected as
pc ¼ 0.005, 97 times for pc ¼ 0.01, and 99 times for pc ¼ 0.05.
K ¼ 0.01 − 1.0, χ ¼ 0.01 − 0.5, and m ¼ 1.0 − 3.0. There are a
As is shown in Table 4, the global optimum is found for HMS ¼
large number of parameter vectors, P~ ¼ ðK; χ; mÞ, producing neg-
10 whether a classical HS algorithm or an HS-BFGS algorithm is
ative outflows in these ranges. These vectors are called infeasible
used. The global optimum is found once for pc ¼ 0 and 100 times
vectors. For example, P~ ¼ ð0.7011; 0.3332; 2.0783Þ is an infea-
for all the other pc values.
sible vector and gives negative outflows for the data set given
As is shown in Table 4, the global optimum is not obtained for
by Wilson (1974). These infeasible vectors lead to the termination
of the optimization process. The stopping condition can be given as HMS ¼ 20 when a pure HS algorithm is used ðpc ¼ 0Þ. In this
(Ali and Törn 2004; Karahan 2011) case, the best SSQ value is the near-global optimum. However,
the global optimum is found when HS-BFGS is used ðpc ≠ 0Þ.
jfbest − fworstj ≤ ε ð11Þ Global optimum is found 100 times for all the other pc values.
From Table 3, it can be inferred that the HS-BFGS results are better
than the pure HS results. As is shown in Tables 1–3, increasing
Iter ≤ MaxIter ð12Þ HMS and pc values produces a decrease of the standard deviation
of SSQs and an increase of the CPU times. The best standard
where ε = a small number, used as 1E − 12 in both HS and BFGS deviation is obtained as 1.02E − 9 when HMS is assumed to be
algorithms; Iter = iteration number; MaxIter = maximum iteration 20 and pc is assumed to be 0.05. In this case, the related CPU time
number (selected as 50,000 for preventing the premature conver- is 15.39 s. If HMS is assumed to be 10 and pc is assumed to be
gence); fbest is the best fitness value in the harmony memory; 0.005, the standard deviation is obtained as 5.85E − 8. The related
and fworst is the worst fitness value in the harmony memory. CPU time is only 0.0872 s. Standard deviations of these two cases
To evaluate the model performance, SSQ and E, the modified are similar, but the required CPU time for the second case is much
coefficient of efficiency (Nash and Sutcliffe 1970; Karahan and smaller than the first one. So, selection of the second case is more
Ayvaz 2005), are used: reasonable than the first one. The HMS and pc values are selected
P as 10 and 0.005 for this problem in terms of standard deviation of
jOt − Ôt ðK; χ; mÞj SSQ values and required CPU time. From Tables 1–4, it can be
E¼1− t P ð13Þ
t jOt − Ōj
concluded that the proposed hybrid algorithm is very fast and
robust according to the HS algorithm.
where Ō = mean of the observed outflows in a given flood Table 5 shows the comparison of the best SSQ for Wilson’s data
data set. obtained from various techniques. As is shown in Table 5, the SSQ
Table 2. Statistical Values of the Parameters for 100 Runs, SSQs, Iteration Number, and CPU Times Obtained by Different pc Values for HMS ¼ 10
Iteration number
pc Statistical values K χ m SSQ HS BFGS Total CPU (s)
0 Best 0.086219 0.286909 1.868166 36.767892 3,407 0 3,407 0.7805
Worst 0.041298 0.281345 2.031977 49.747990 50,000 0 50,000 11.3872
Mean 0.073289 0.285700 1.908900 38.637592 26,467 0 26,467 6.0132
Standard deviation 1.46E−02 1.44E−03 4.66E−02 2.79Eþ00 21,020 0 21,020 4.7724
0.005 Best 0.086249 0.286917 1.868088 36.767888 174 1 175 0.0872
Worst 0.086241 0.286916 1.868109 36.767889 491 1 492 0.1668
Mean 0.086248 0.286916 1.868089 36.767888 758 3 761 0.3324
Standard deviation 2.02E−06 5.14E−07 5.19E−06 5.85E−08 510 2 513 0.2144
0.01 Best 0.086249 0.286917 1.868088 36.767888 1,699 18 1,717 1.0743
Worst 0.086246 0.286918 1.868094 36.767889 394 2 396 0.1592
Mean 0.086248 0.286916 1.868090 36.767888 757 8 764 0.4500
Standard deviation 1.88E−06 4.66E−07 4.86E−06 5.17E−08 510 6 516 0.2964
0.05 Best 0.086249 0.286917 1.868087 36.767888 1,182 59 1,241 1.9631
Worst 0.086244 0.286916 1.868099 36.767889 161 5 166 0.2177
Mean 0.086249 0.286917 1.868089 36.767888 828 41 869 1.4199
Standard deviation 7.83E−07 1.00E−07 2.01E−06 8.39E−09 611 30 640 0.9545
Table 3. Statistical Values of the Parameters for 100 Runs, SSQs, Iteration Number, and CPU Times Obtained by Different pc Values for HMS ¼ 20
Iteration number
pc Statistical values K χ m SSQ HS BFGS Total CPU (s)
0.0 Best 0.086464 0.286927 1.867534 36.768042 21,455 0 21,455 4.8746
Worst 0.030211 0.277797 2.101898 62.746185 50,000 0 50,000 11.3585
Mean 0.073048 0.285532 1.911363 39.211602 47,992 0 47,992 10.8918
Standard deviation 1.59E−02 1.96E−03 5.65E−02 4.73Eþ00 7,374 0 7,374 1.6716
0.005 Best 0.086249 0.286917 1.868087 36.767888 4,529 22 4,551 1.7309
Worst 0.086242 0.286916 1.868104 36.767889 1,379 4 1,383 0.4369
Mean 0.086248 0.286916 1.868089 36.767888 2,144 11 2,154 0.8582
Standard deviation 1.12E−06 2.25E−07 2.88E−06 2.03E−08 1,362 7 1,369 0.4890
0.01 Best 0.086249 0.286917 1.868087 36.767888 4,209 36 4,245 2.2042
Worst 0.086246 0.286916 1.868094 36.767888 1,021 6 1,027 0.4498
Mean 0.086249 0.286917 1.868088 36.767888 2,715 27 2,742 1.4475
Standard deviation 6.37E−07 9.79E−08 1.64E−06 3.98E−09 2,036 22 2,057 1.0301
0.05 Best 0.086249 0.286917 1.868087 36.767888 9,884 495 10,379 15.3947
Worst 0.086248 0.286917 1.868090 36.767888 867 33 900 1.3991
Mean 0.086249 0.286917 1.868088 36.767888 3,170 157 3,327 5.1207
Standard deviation 4.10E−07 3.86E−08 1.05E−06 1.02E−09 2,189 111 2,299 3.4673
NL-LSM (Yoon and 0.06 0.27 2.36 156.4399 scheme given by the authors for the parameters obtained by PSO to
Padmanabhan 1993) prove his claim. Chu and Chang (2009) found the best SSQ value as
S-LSM (Gill 1978) 0.01 0.25 2.347 145.6945 36.89 by using the PSO algorithm. The ranges of the parameters
LMM (Das 2004) 0.0753 0.2769 2.2932 130.4872
HJ+DFP (Tung 1985) 0.0764 0.2677 1.8978 45.612
given by the authors are K ¼ 0.01–0.20, χ ¼ 0.2–0.3, and
GA (Mohan 1997) 0.1033 0.2813 1.8282 38.2363 m ¼ 1.5–2.5. Their value of χ ¼ 0.3330 is greater than 0.3. Also,
ICSAa (Luo and Xie 0.0884 0.2862 1.8624 36.8026 the authors used the PSO algorithm with a different Muskingum
2010) model, which utilizes an average inflow in Eq. (4). So, their result
NMSb (Barati 2011) 0.0862 0.2869 1.8681 36.76 was not compared in Table 5.
DEc (Xu et al. 2011) 0.5175 0.2869 1.868 36.77 Computed outflows of the Wilson data that were obtained by
HS (Kim et al. 2001) 0.0883 0.2873 1.863 36.7829 different methods are given in Table 6. As is shown in Table 6,
PSF-HS (Geem 2011) 0.0864 0.2869 1.8677 36.768
computed outflows from HS-BFGS are better than the other
BFGS (Geem 2006) 0.0863 0.2869 1.8679 36.7679
HS-BFGS 0.086249 0.286917 1.868088 36.767888 methods. The comparison of the observed and computed hydro-
a
graphs of the Wilson data is presented in Fig. 1. As is shown in
Luo and Xie (2010) reported that the SSQ value was 35.64, but this value is
Fig. 1, the computed hydrograph is well suited to the observed
not correct for the parameters given by the authors. As reported by Geem
(2011), the correct SSQ value is 36.8026. Hydrograph.
b
Barati (2011) reported that the SSQ value was 36.76, but this value is not Selection of the penalty constants is a difficult task and is prob-
correct for parameters given by the authors. The correct SSQ value is lem dependent. In Table 7, the statistical values of the parameters
calculated as 36.7679. (for 100 runs) and SSQs, obtained by different λ values for
HMS ¼ 10, are presented. As is shown in Table 7, the best standard
c
Although the DE algorithm uses the numerical routing procedure given by
Geem (2006), a different K value is found than that provided by Geem deviation of SSQ values is obtained for λ ¼ 1;000. From Table 7, it
(2006) and the other algorithms given in the literature.
can be concluded that SSQ values are less sensitive to λ values. The
penalty constant can be selected as 1,000 for this problem. The se-
(36.767888) obtained by HS-BFGS is better than all techniques lected λ values for this problem may not suitable for different opti-
given in the literature. As is clearly shown in Table 2, there is mization problems. They must be determined experimentally for
an extremely small difference between the best solution and the any optimization problem.
Table 6. Comparison of the Observed and Best Computed Outflows for Wilson Data
Time (h) I t (cms) Ot (cms) NL-LSM S-LSM LMM HJ+DFP GA ICSA HS PSF-HS BFGS DE NMS HS-BFGS
0 22 22 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0
6 23 21 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0 22.0
12 35 21 22.8 22.8 22.8 22.4 22.4 22.4 22.4 22.4 22.4 22.4 22.4 22.4
18 71 26 29.3 29.7 29.6 26.7 26.4 26.6 26.6 26.6 26.6 26.6 26.6 26.6
24 103 34 37.9 39.3 38.7 34.8 34.2 34.4 34.4 34.5 34.5 34.5 34.5 34.5
30 111 44 45.7 48.0 47.0 44.7 44.2 44.2 44.1 44.2 44.2 44.2 44.2 44.2
36 109 55 56.0 58.4 57.8 56.9 57.0 56.9 56.8 56.9 56.9 56.9 56.9 56.9
42 100 66 65.5 67.5 67.6 67.7 68.2 68.1 68.1 68.1 68.1 68.1 68.1 68.1
48 86 75 73.5 75.1 75.8 76.3 77.2 77.1 77.1 77.1 77.1 77.1 77.1 77.1
54 71 82 79.8 80.7 81.9 82.2 83.3 83.3 83.3 83.3 83.3 83.3 83.3 83.3
60 59 85 83.2 83.5 85.0 84.7 85.7 85.9 85.9 85.9 85.9 85.9 85.9 85.9
66 47 84 83.2 83.0 84.4 83.5 84.2 84.5 84.5 84.5 84.5 84.5 84.5 84.5
72 39 80 80.9 80.1 81.4 79.8 80.2 80.5 80.6 80.6 80.6 80.6 80.6 80.6
78 32 73 75.6 74.5 75.2 73.3 73.3 73.6 73.7 73.7 73.7 73.7 73.7 73.7
84 28 64 68.5 67.0 67.1 65.5 65.1 65.3 65.4 65.4 65.4 65.4 65.4 65.4
90 24 54 59.4 57.8 57.2 56.5 55.8 55.9 56.0 56.0 56.0 56.0 56.0 56.0
96 22 44 49.4 47.6 46.5 47.5 46.7 46.6 46.7 46.7 46.7 46.7 46.7 46.6
102 21 36 38.7 37.0 35.5 38.7 38.0 37.8 37.8 37.8 37.7 37.8 37.8 37.7
108 20 30 29.1 27.7 26.4 31.4 30.9 30.5 30.5 30.5 30.5 30.5 30.5 30.4
114 19 25 22.3 21.6 21.1 25.9 25.7 25.3 25.3 25.2 25.2 25.2 25.2 25.2
120 19 22 19.1 19.0 19.0 22.1 22.2 21.8 21.8 21.7 21.7 21.7 21.7 21.7
126 18 19 19.0 19.0 19.0 20.2 20.3 20.0 20.0 20.0 20.0 20.0 20.0 20.0
Table 7. Statistical Values of the Parameters for 100 Runs, SSQs, Obtained by Different Values for HMS ¼ 10
λ Statistical values K χ m SSQ
1 Best 0.086249 0.286917 1.868087 36.767888
Worst 0.086247 0.286916 1.868091 36.767888
Mean 0.086249 0.286917 1.868087 36.767888
Standard deviation 2.90E−07 9.91E−08 7.44E−07 1.35E−09
10 Best 0.086249 0.286917 1.868087 36.767888
Worst 0.086247 0.286916 1.868093 36.767888
Mean 0.086249 0.286917 1.868087 36.767888
Standard deviation 2.55E−07 6.31E−08 6.54E−07 1.54E−09
100 Best 0.086249 0.286917 1.868087 36.767888
Worst 0.086249 0.286916 1.868088 36.767888
Mean 0.086249 0.286917 1.868087 36.767888
Standard deviation 1.87E−07 1.14E−07 4.80E−07 1.35E−09
1,000 Best 0.086249 0.286917 1.868087 36.767888
Worst 0.086248 0.286917 1.868090 36.767888
Mean 0.086249 0.286917 1.868087 36.767888
Standard deviation 1.72E−07 8.11E−08 4.43E−07 4.95E−10
10,000 Best 0.086249 0.286917 1.868087 36.767888
Worst 0.086251 0.286917 1.868082 36.767888
Mean 0.086249 0.286917 1.868087 36.767888
Standard deviation 2.61E−07 6.56E−08 6.69E−07 1.22E−09
1,000,000 Best 0.086249 0.286917 1.868087 36.767888
Worst 0.086251 0.286917 1.868082 36.767888
Mean 0.086249 0.286917 1.868087 36.767888
Standard deviation 2.61E−07 6.56E−08 6.69E−07 1.22E−09
Note: Bolded numbers indicate that the best standard deviation of SSQ values is obtained for λ ¼ 1;000.
600
400
References
Al-Humoud, J. M., and Esen, I. I. (2006). “Approximate methods for the
200 estimation of Muskingum flood routing parameters.” Water Resour.
Manage., 20(6), 979–990.
Ali, M. M., and Törn, A. (2004). “Population set-based global optimization
0
0 20 40 60 80 100 120 140 160 180 algorithms: Some modifications and numerical studies.” Comput. Oper.
Time (h) Res., 31(10), 1703–1725.
Downloaded from ascelibrary.org by Pamukkale University on 06/13/13. Copyright ASCE. For personal use only; all rights reserved.