Professional Documents
Culture Documents
net/publication/234136052
CITATIONS READS
154 2,462
1 author:
Reza Barati
Water Authority
106 PUBLICATIONS 970 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Reza Barati on 02 June 2014.
Abstract: The linear form of the Muskingum model has been widely applied to river flood routing. However, a nonlinear relationship
between weighted-flow and storage volume exists in most rivers, making the use of the linear Muskingum model inappropriate. On the
other hand, the application of the nonlinear Muskingum model suffers from hydrologic parameters estimation. The current study aims
at presenting the objective approach of the Nelder-Mead simplex (NMS) algorithm for the purpose of estimating the parameters of the
nonlinear Muskingum model. The performance of this algorithm is compared with other reported parameter estimation techniques together
with a historical example. Results of the implementation of this procedure indicate that the NMS algorithm is efficient for the estimating
parameters of the nonlinear Muskingum models. This algorithm is easy to be programmed, and it is quite efficient for finding an optimal
solution very quickly. Although this technique requires an initial guess for the parameter estimation, results of the sensitivity analysis of the
initial parameter values showed that in 84.8% of the cases, the optimum or near-optimum results are achieved. DOI: 10.1061/(ASCE)HE
.1943-5584.0000379. © 2011 American Society of Civil Engineers.
CE Database subject headings: Floods; Routing; Hydrologic models; Optimization; Algorithms.
Author keywords: Flood routing; Hydrologic models; Optimization; Algorithms.
Introduction plotted loop can be chosen as the correct value of X, and the slope
of the straight line fitted through the loop gives K.
The importance of flood routing in rivers has been vastly recog- Yoon and Padmanabhan (1993) shows that when the relation-
nized in hydrologic engineering practices. Field data scarcity often ship between ½XI t þ ð1 XÞOt and St is not linear, use of a non-
prevents the use of the 1D Saint-Venant equations of gradually linear model may be more appropriate. In previous research, two
varied unsteady flow to route floods in rivers. One of the hydrologic forms of nonlinear Muskingum models have been recommended
routing techniques, the Muskingum method (McCarthy 1938), for taking into account the nonlinearity (Gill 1978; Singh and
has been frequently used to route floods in natural channels and Scarlatos 1987; Mohan 1997; Luo and Xie 2010) as
rivers. In the linear Muskingum model, the following continuity
and storage equations are used: St ¼ K½XI t þ ð1 XÞOt m ð3Þ
Downloaded 17 Dec 2011 to 35.8.11.2. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
1=m
proposed an alternative search technique, derived from genetic ΔSt 1 St 1
¼ þ I ð6Þ
algorithm (GA), for estimating the parameters of the models of Δt 1X K 1X t
Eqs. (3) and (4) and reported a better performance than other meth-
ods proposed in the literature. GA is heuristic iterative search algo-
Also, the next accumulated storage can be expressed as
rithm that attempts to find the optimal solution in a given decision
space derived from a search algorithm that mimics Darwinian Stþ1 ¼ St þ ΔSt ð7Þ
evolution and survival of the fittest in a natural environment
(Cheng et al. 2002). Kim et al. (2001) applied the harmony search
The solution procedure for the nonlinear model of Eq. (3) includes
(HS) algorithm to the same parameter calibration problem. Their
the following steps (Tung 1985; Kim et al. 2001; Geem 2006):
results show that HS estimation was better than GA and also
Step 1. Determine values of the hydrologic parameters by applying
did not require that the initial values of the design parameters were
optimization techniques. For the practical example data and NMS
close to the optimum. However, this method requires careful atten-
technique, the parameter vector is (K ¼ 0:5175, X ¼ 0:2869,
tion for the harmony memory considering rate and pitch adjusting
and m ¼ 1:8681).
rate. Das (2004) estimated parameters for linear and nonlinear
Step 2. Compute the initial storage volume by Eq. (3), where the
Muskingum models by using the Lagrange multiplier. This method
initial flow rate at downstream section is the same as the initial flow
transforms the constrained parameter optimization problem into an
rate at upstream section. For example, St ¼ 166:61.
unconstrained problem. The Lagrange multiplier estimates optimal
Step 3. Compute the time rate of change of the channel storage by
parameters by using the first-order necessary condition for optima.
Eq. (6). For example, ΔSt =Δt ¼ 1:4019, where Δt ¼ 6.
The results obtained by this method were not as good as the results
Step 4. Calculate the accumulated storage at the next time by
obtained by other researchers (Luo and Xie 2010). Geem (2006)
Eq. (7). For example, Stþ1 ¼ 175:02.
recommended the Broyden-Fletcher-Goldfarb-Shanno (BFGS) al-
Step 5. Compute the magnitude of the outflow at the next time by
gorithm to find the values of the three parameters in the nonlinear
Eq. (5). For example, Otþ1 ¼ 22:42.
Muskingum model. This technique found the best parameter values
These steps carry out unit total number of discharge hydrograph
compared to previous results in terms of the sum of the square
ordinates (N) are calculated. A similar procedure may be developed
deviation between the observed and routed outflows. However,
for the model given by Eq. (4).
BFGS requires initial value assumption for the K, X, and m param-
eters. Das (2007) offered a chance-constrained optimization-based
model for Muskingum model parameter estimation. This model is
very complex and requires massive computation for parameter es-
Nelder-Mead Simplex Algorithm
timation of the Muskingum model (Luo and Xie 2010). Chu and The Nelder-Mead simplex algorithm for unconstrained optimiza-
Chang (2009) used the particle swarm optimization (PSO) tech- tion has been used extensively to solve parameter estimation and
nique to estimate parameters for the nonlinear Muskingum model. other problems since its inception by Nelder and Mead (1965).
PSO does not demand any initial estimate of values of any of the In the past several years, the NMS technique has been successfully
parameters. However, the result obtained by PSO was not as good applied, including in hydrological modeling. For example, the
as the results obtained by other methods such as HS and BFGS. U.S. Army Corps of Engineers has used the NMS algorithm in
Luo and Xie (2010) recommended an immune clonal selection al- Hydrologic Engineering Center-Hydrologic Modeling System
gorithm (ICSA) to find the values of the three parameters in the (HEC-HMS) software as a search method for optimization of the
nonlinear Muskingum model. ICSA is an intelligent algorithm that hydrologic parameters (HEC-HMS 2000). In this software, optimi-
can efficiently overcome the prematurity and slow convergence zation of the hydrologic parameters [such as initial abstraction
speed of the traditional evolution algorithm. However, the sensitiv- and curve number in soil conservation service (SCS) loss model;
ity analysis is required for determination of the algorithm param- moisture deficit, hydraulic conductivity, and wetting front suction
eters such as the clonal scale, mutation probability, and crossover in the Green and Ampt loss model; and time of concentration and
probability. storage coefficient in Clark’s unit hydrograph (UH) model] can be
In previous research, including hydrological modeling, the done. Ouria and Toufigh (2009) presented the application of the
Nelder-Mead simplex (NMS) algorithm has been successfully NMS algorithm for unconfined seepage problems. Abida (2009)
applied (HEC-HMS 2000; Abida 2009; Ouria and Toufigh applied the NMS technique to estimate parameters required for
2009). In this study, the parameter estimation for the nonlinear routing floods through open compound channels.
Muskingum model is presented by using the NMS algorithm. The The NMS algorithm shares similarities with evolutionary com-
results of this algorithm are compared with other reported tech- putation techniques such as GA and PSO (Fan et al. 2006; Lasheen
niques (Gill 1978; Tung 1985; Mohan 1997; Kim et al. 2001; Geem et al. 2009; Zahara and Kao 2009). This method attempts to min-
2006; Luo and Xie 2010) through a historical example. Because the imize a scalar-valued nonlinear function of n real variables by using
nonlinear Muskingum model in Eq. (4) is not as popular as the one only function values, without any derivative information (Lagarias
in Eq. (3) (Geem 2006; Luo and Xie 2010), this paper will focus on et al. 1998). Thus, the NMS technique is classified in the general
the latter model for comparison purposes. class of direct search methods. In the NMS algorithm, simplex is a
geometric shape in n dimensions of nonzero volume that is the con-
vex hull of n þ 1 vertices (Lagarias et al. 1998). The algorithm is
Routing Procedure of Nonlinear Muskingum Model started by setting up an initial simplex. Therefore, NMS technique
By rearranging Eq. (3), the routed outflow Ot can be declared as uses an initial guess of the user to produce the initial simplex. The
1=m initial guess is used as one of the vertices of the simplex. The
1 St X n remaining vertices of the initial simplex are found by adding
Ot ¼ I ð5Þ ΔL (%) to each component of the initial guess vector. The oper-
1X K 1X t
ations of the NMS algorithm rescale the simplex by using four
By substituting Eq. (5) in Eq. (1), the time rate of change of the basic procedures: reflection, expansion, contraction, and shrinkage
storage volume can be achieved as (see Fig. 1).
Downloaded 17 Dec 2011 to 35.8.11.2. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
b. Inside. If f r ≥ f nþ1 , compute the inside contraction point xcc
by
Fig. 2. The illustration of new points in search space of Nelder-Mead simplex algorithm
Downloaded 17 Dec 2011 to 35.8.11.2. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
the sum of the squared deviations between the calculated and procedures are evaluated by using the following measures: (1) the
observed outflows sum of the square of the deviation between the computed and ob-
served outflows (SSQ [L6 T 2 ]); (2) the sum of the absolute value of
X
N
the deviations between the computed and observed outflows (SAD
min SSQ ¼ fOt O
^ t g2 ð14Þ [L3 T 1 ]); and (3) the absolute value of the deviations of peak of
t¼1
computed and observed outflows (DPO [L3 T 1 ]). SAD and SSQ
where SSQ = sum of the square deviation of the observed outflow are measures of the bias and accuracy of the techniques, respec-
and computed outflow (L6 T 2 ); Ot = observed outflow at time tively (Mohan 1997). Because all of the methods have good results
t (L3 T 1 ); O
^ t = computed outflow at time t (L3 T 1 ); and N = total in terms of time to peak, the results of different methods are not
number of discharge hydrograph ordinates to be simulated. There- compared in this term. The performance criteria of the nonlinear
fore, the objective function of the NMS algorithm for the nonlinear Muskingum model for various parameter estimation methods and
model of Eq. (3) is given as the hydrologic parameters values are presented in Table 2. Because
SSQ is the objective function, it is the premier measure in Table 2.
N
X 1=m 2 Other measures derived from the data may perform differently.
1 St X
min SSQ ¼ Ot I The results indicate that SSQ from NMS give the lowest value
1X K 1X t
t¼1 (SSQ ¼ 36:76) than other methods such as HS, GA, HJ+DFP,
ð15Þ LSM, and ICSA and is as accurate as BFGS for the nonlinear model
of Eq. (3). If the effects of three measures are simultaneously con-
where I t = observed inflow at time t (L3 T 1 ); St = channel storage sidered, the algorithms such as GA, HS, BFGS, ICSA, and NMS
at time t (L3 ); and K, X, and m = design parameters. are challenged. GA gives lower SAD and DPO than other methods,
In the previous papers, different researchers have presented the but it shows far away SSQ from optimal solution than other algo-
routed outflows with a different number of digits after the decimal rithms. Between HS and ICSA, HS performed better because of
point while the results of the measures are sensitive to the number a lower value of SSQ. The comparison between NMS and BFGS
of digits after the decimal point. Barati (2010) indicates that the shows that NMS carries out better because of lower values of SAD
rounding off effects on measures was faded by rounding off with and DPO. Finally, NMS and HS are compared. HS resulted in a
three digits after decimal points. For true comparison of the results lower value of SAD, but NMS produced lower values of SSQ and
of different methods in Table 1, the routed outflows are represented DPO. However, the different subjects (such as the number of iter-
with three digits after decimal point accuracy. Columns 1–3 of ations for convergence, the time convergence, and the algorithm
Table 1 are the actual data (Wilson 1974); columns 4–10 show parameters) must be considered for comparison of different meth-
the routed flows by using Eq. (3), with the parameters estimated ods. For example, the comparison between NMS and HS shows
by LSM (Gill 1978), HJ+DFP (Tung 1985), GA (Mohan 1997), that NMS found the optimal solution in a maximum of 221 itera-
HS (Kim et al. 2001), BFGS (Geem 2006), ICSA (Luo and Xie tions, but HS needs about 4,500 iterations for convergence (Kim
2010), and NMS, respectively. Fig. 3 depicts these results. In ad- et al. 2001). Therefore, the algorithms such as NMS and BFGS
dition, the performance of NMS and the other parameter estimation have the shortest operating time for convergence rather than
Downloaded 17 Dec 2011 to 35.8.11.2. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Fig. 3. Difference between the routed and observed outflows for various methods
Table 2. Hydrologic Parameters Estimates and Comparative Statistics of The results of two group methods are sensitive to the algorithm
Example Data parameters or initial guess. The estimation of the algorithm param-
Parameters values Statistics values eters or initial values assumption for these techniques derived from
the data may be on the same or a different difficulty level.
SSQ SAD DPO
Furthermore, NMS finds better results [SSQ ¼ 245:58 for the
Method K X m ðm6 s2 Þ ðm3 s1 Þ ðm3 s1 Þ
parameter vector (K ¼ 0:4606, X ¼ 0:2287, and m ¼ 1:5012)]
LSM 0.0100 0.2500 2.3470 145.71 46.26 1.51 than [SSQ ¼ 1038:2 for the parameter vector (0.0500, 0.2200,
HJ+DFP 0.0764 0.2677 1.8978 45.62 24.92 0.30 2.3500)] of NONLR (Yoon and Padmanabhan 1993) and [SSQ ¼
GA 0.1033 0.2813 1.8282 38.24 23.00 0.71 264:8 for the parameter vector (0.1773, 0.2305, and 1.7105)] of GA
HS 0.0883 0.2873 1.8630 36.78 23.40 0.92 (Mohan 1997) for the nonlinear model of Eq. (4). Therefore, it can
BFGS 0.0863 0.2869 1.8679 36.76 23.47 0.91 be said that the performance of the NMS algorithm for parameter
ICSA 0.0884 0.2862 1.8624 36.80 23.40 0.90 estimation of nonlinear Muskingum models is as good as or better
NMS 0.0862 0.2869 1.8681 36.76 23.46 0.90 than the other methods.
Discussion
GA, HS, and ICSA. On the other hand, it is known that algorithms
such as GA, HS, and ICSA do not need initial value assumption for
the hydrologic parameters, but they do require careful considera- Discussion of Results—Effect of NMS Algorithm
tion when choosing the algorithm parameters, such as the crossover Coefficients
rate and mutation rate in GA; the harmony memory considering Four scalar parameters consist of coefficients of reflection (ρ), ex-
rate and pitch adjusting rate in HS; and the clonal scale, mutation pansion (χ), contraction (γ), and shrinkage (σ) must be indicated to
probability, and crossover probability in ICSA (Geem 2006; Barati define a complete NMS algorithm (Lagarias et al. 1998). On the
2010). In contrast, the methods such as HJ+DFP, BFGS, and NMS other hand, ΔL must be determined on the basis of the smallest
do require initial value assumption for K, X, and m parameters. number of function evaluations.
Downloaded 17 Dec 2011 to 35.8.11.2. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Table 4. Mean and Minimum Number of Function Evaluations for Different Values of χ, γ, σ over Various Vectors of Initial Guess
σ Mean Minimum
γ ¼ 0:25 γ ¼ 0:5 γ ¼ 0:75 χ¼2 χ¼3 χ¼4 χ¼2 χ¼3 χ¼4
a a a
0.25 — — 275 275 354 131 182 354
0.5 — — 229a 275a 354a 131 182 354
0.75 — — 267a 275a 354a 131 182 354
— 0.25 — 201 221 225a 127 128 148
— 0.5 — 195 210 241a 127 128 148
— 0.75 — 197 209 242a 127 128 148
— — 0.25 353 345a 364a 272 261 282
— — 0.5 349 373a 364a 272 261 282
— — 0.75 353 374a 359a 272 261 282
Note: ρ ¼ 1.
a
This combination of NMS parameters failed to find the optimal solution for at least one of the initial values assumption of the hydrologic parameters.
To decide the final values for these parameters and ΔL, the of NMS in terms of SSQ for the parameter vector (K ¼ 0:1,
performance of the NMS technique with different parameter X ¼ 0:3, and m ¼ 1:75) as initial guess. It can be seen that in most
choices was investigated. A problem encountered in this procedure of cases SSQ was optimal (SSQ ¼ 36:76) for ρ ¼ 1. On the basis of
was that the size of the initial simplex had an effect on the speed of these results and the number of function evaluations, ρ ¼ 1 was
convergence; therefore, several vectors of initial guess and several selected. Table 4 summarizes the mean and minimum numbers of
ΔL of the initial simplex were used (Nelder and Mead 1965). function evaluations required for convergence over 10 vectors of
For all combinations of ρ ¼ ð1=2; 2=3; and 1Þ; χ ¼ ð2; 3; and 4Þ; initial guess. In calculation of the mean of the number of function
evaluations, the cases that failed to find the optimal solution are
γ ¼ ð0:25; 0:5; and 0:75Þ; and σ ¼ ð0:25; 0:5; and 0:75Þ over 10
omitted. It can be seen from Table 4 that the mean and minimum
vectors of initial guess of the hydrologic parameters have been op-
timized. Sensitivity analysis of NMS parameters indicates that the
results are more sensitive than ρ parameter. For example, Table 3
summarizes the results of sensitivity analysis of the key parameters
Downloaded 17 Dec 2011 to 35.8.11.2. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
numbers of function evaluations were minimized for χ ¼ 2, in a range that its effective factors can arise. According to results of
γ ¼ 0:5, and σ ¼ 0:5. The best set of parameters (ρ ¼ 1, χ ¼ 2, former researchers for Wilson’s data and Eq. (3), the ranges of K, X,
γ ¼ 0:5, and σ ¼ 0:5) have the minimum of the number of function and m parameters are as follows: K ¼ 0:01–0:2; X ¼ 0:2–0:3; and
evaluations for convergence, and this set resulted the optimal sol- m ¼ 1:5–2:5 (Gill 1978; Tung 1985; Yoon and Padmanabhan
ution in most of vectors of the initial guess. 1993; Mohan 1997; Kim et al. 2001; Geem 2006; and Luo and
For investigation of the effect of ΔL on the results, the mean of Xie 2010).
the number of function evaluations for 10 vectors of initial guess as In the first way, base value parameters have been determined;
the change of ΔL is presented in Fig. 4. Also the mean and mini- then the parameters in a specified range are changed, and the model
mum of the number of function evaluations for various ΔL and is run for the new conditions. In all cases, only one parameter is
NMS parameters for five vectors of initial guess are summarized changed and the rest remain unchanged. Here the base values are
in Table 5. It can be said that the number of function evaluations the optimized values of the parameters. In Fig. 5 the results of the
is smallest for ΔL ¼ 5 (%). sensitivity analysis for the nonlinear Muskingum model of Eq. (3)
In conclusion, because the universal choices used in the NMS have been illustrated. Fig. 5 demonstrates that a unique set of values
technique are ρ ¼ 1, χ ¼ 2, γ ¼ 0:5, and σ ¼ 0:5 (Nelder and for the hydrologic parameters resulted in minimum SSQ for the
Mead 1965; Lagarias et al. 1998; HEC-HMS 2000; Yang et al. historical example studied.
2005; Zahara and Kao 2009), it can be said that NMS does not In the second way, for the initial values K ¼ ð0:01; 0:05; 0:1;
require the determination of the algorithm parameters for each case 0:15; and 0:2Þ, X ¼ ð0:2; 0:225; 0:25; 0:275; and 0:3Þ, and m ¼
study. In other words, this set of parameters is best set for shortest ð1:5; 1:75; 2:0; 2:25; and 2:5Þ, the parameters have been optimized.
operating time of convergence. This set of coefficients correspond The results of this section are separated as four groups consisting
to a symmetric reflect step, an expansion step that is twice the of optimum, near optimal, far away optimal, and infeasible. The
length of the reflect step, and contraction steps that are half the model will be optimal if the optimum amount (SSQ ¼ 36:76) is
length of the reflect step. In a similar manner, the shrink step halves achieved. The model will be near optimal providing the maximum
the distance from the best vertex to each of the other vertices in the value of error limited to SSQ ¼ 36:77. In the infeasible group,
simplex (see Fig. 2). the model cannot properly run. At last, if the model can be run and
maximum value of error is not in the near optimal group, the re-
Discussion of Results—Effect of Initial Guess sults will be far away optimal. The results indicate that 50 cases
were optimal (SSQ ¼ 36:76), 56 cases were near optimal (SSQ ¼
For investigation of the effect of the initial guess for the hydrologic 36:7681, 36.7683, 36.7687, 36.7688, and 36.7691), 4 cases were
parameters on the objective function, the sensitivity analysis takes far away optimal (SSQ ¼ 36:7750, 36.7885, 36.7906, and 37.709),
place in two ways. The sensitivity analysis is true when takes place and 15 cases were infeasible (Table 6). This sensitivity analysis is
Table 6. Sensitivity Analysis of the Initial Guess of the Hydrologic Parameters in Terms of SSQ
m K
X ¼ 0:2 X ¼ 0:225 X ¼ 0:25 X ¼ 0:275 X ¼ 0:3 0.01 0.05 0.1 0.15 0.2
1.5 — — — — Inf. Inf. N-Opt. N-Opt. N-Opt.
1.75 — — — — Inf. N-Opt. N-Opt. N-Opt. N-Opt.
2 — — — — Opt. N-Opt. N-Opt. N-Opt. N-Opt.
2.25 — — — — N-Opt. Opt. F.A.Opt. N-Opt. N-Opt.
2.5 — — — — N-Opt. Opt. Opt. Opt. N-Opt.
— 1.5 — — — Inf. Inf. N-Opt. Opt. Opt.
— 1.75 — — — Inf. N-Opt. N-Opt. N-Opt. Opt.
— 2 — — — Opt. Opt. F.A.Opt. Opt. N-Opt.
— 2.25 — — — N-Opt. N-Opt. Opt. Opt. Opt.
— 2.5 — — — N-Opt. N-Opt. Opt. N-Opt. Opt.
— — 1.5 — — Inf. Inf. Opt. N-Opt. N-Opt.
— — 1.75 — — Inf. N-Opt. Opt. Opt. Opt.
— — 2 — — Opt. Opt. F.A.Opt. N-Opt. N-Opt.
— — 2.25 — — N-Opt. Opt. N-Opt. Opt. Opt.
— — 2.5 — — N-Opt. N-Opt. Opt. Opt. N-Opt.
— — — 1.5 — Inf. Inf. Opt. N-Opt. N-Opt.
— — — 1.75 — Inf. N-Opt. Opt. Opt. Opt.
— — — 2 — Opt. N-Opt. N-Opt. N-Opt. Opt.
— — — 2.25 — Opt. Opt. F.A.Opt. Opt. Opt.
— — — 2.5 — N-Opt. Opt. N-Opt. N-Opt. N-Opt.
— — — — 1.5 Inf. Inf. Opt. N-Opt. Opt.
— — — — 1.75 Inf. N-Opt. Opt. Opt. Opt.
— — — — 2 N-Opt. Opt. Opt. Opt. N-Opt.
— — — — 2.25 Opt. N-Opt. N-Opt. N-Opt. N-Opt.
— — — — 2.5 Opt. N-Opt. Opt. Opt. N-Opt.
Note: Inf. = infeasible; Opt. = optimal; N.-Opt. = near optimal; and F.A.Opt. = far away optimal.
Downloaded 17 Dec 2011 to 35.8.11.2. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
values (Table 6); (2) NMS produces improved results of SAD
and DPO (Table 2); (3) NMS provides acceptable results for the
model of Eq. (4) if BFGS is divergent (Geem 2006); and (4) BFGS
is commonly selected for smooth nonlinear problems, but NMS
is chosen effectively for either smooth nonlinear or nonsmooth
problems. In addition to higher accuracy, when compared with
algorithms such as GA, HS, and ICSA, NMS does not need deter-
mination of algorithm parameters for each case study. Furthermore,
because these methods randomly search for the optimal solution,
some uncertainties are included in these algorithms (Luo and Xie
2010). Also NMS has a shorter calculation time.
The NMS algorithm produces good results, does not need the
derivatives of the function, and deals with a population of points
instead of a single point. Moreover, it is quite efficient in finding the
optimal solution very quickly, easy to program, and requires very
little storage and information for execution. Because of its simplic-
ity and robustness, the NMS technique is much more efficient than
alternative traditional methods (Lasheen et al. 2009). However, it is
important to note that the NMS algorithm requires initial value as-
sumptions for the hydrologic parameters. The sensitivity analysis
for the initial guess of parameters in a range that is available for the
study shows that in 84.8% of the cases the optimum results are
Fig. 6. Progression of solution with number of iterations achieved after running two time optimizations. Eventually, consid-
ering that NMS is modeled by the MATLAB program that is widely
available for engineers, the users are able to evaluate the parameters
without knowing the exact details of optimization technique. Also,
conducted in the same manner as Geem (2006) with the difference good results are yielded for nonlinear Muskingum models. This
that a wider range of parameters is considered. By comparing the algorithm is one of the efficient approaches for estimating the
results of sensitivity analysis of BFGS and NMS, it is concluded nonlinear Muskingum model parameters. Furthermore, it is antici-
that NMS produces better results. In addition, except in infeasible pated that the NMS algorithm can be successfully applied to other
cases, the results of NMS are always improved, but BFGS results problems in hydrologic engineering.
are divergent in some cases. Investigation of the results of the Because optimization algorithms have only been compared for
sensitivity analysis also shows that in optimal cases the lower and Wilson’s data, in future researches these methods can be compared
upper bounds for the maximum number of iterations of vectors of for several flood events with a nonlinear relationship between
the initial guess are limited to 71 and 221, respectively. However, weighted flow and storage volume. Therefore, the best algorithm
all cases can be done during only 1 s. In Fig. 6, the improvement can be determined in the field conditions.
procedure of the best solution in the NMS algorithm as a number of
iterations for the lower and upper bounds is presented. Fig. 6 clearly
demonstrates that NMS technique produces good results even Acknowledgments
for the value of SSQ of the initial guess distant from optimal
solution (SSQ ¼ 36:76). The writer appreciatively acknowledges the valuable comments
In real calculations, the initial values can be adapted by a trial- offered by the editors and anonymous reviewers in improving the
and-error procedure to minimize the value of error (Geem 2006). technical contents of this paper. The author would like to thank the
For improvement of the result of NMS, the model can be run at Young Researchers Club, Mashhad Branch, Islamic Azad Univer-
least twice. In this way, after running one-time optimization, the sity, Mashhad, Iran, for financially supporting this research.
values obtained for the hydrologic parameters are considered as
the initial guess for the next run. Thus, finding the optimal solution
can be increased. For the data of the study, the optimum results are References
achieved in 84.8% of the cases.
Abida, H. (2009). “Identification of compound channel flow parameters.”
J. Hydrol. Hydromech., 57(3), 172–181.
Conclusions Barati, R. (2010). “Discussion of ‘Parameter estimation for nonlinear
Muskingum model based on immune clonal selection algorithm’ by
In the past, several parameter estimation methods have been Jungang Luo and Jiancang Xie.” J. Hydrol. Eng., 16(4), 391–393.
recommended to determine the parameter values for the nonlinear Cheng, C. T., Ou, C. P., and Chau, K. W. (2002). “Combining a fuzzy
Muskingum methods. In the present study, the Nelder-Mead optimal model with a genetic algorithm to solve multi-objective
simplex algorithm has been proposed for estimating the hydrologic rainfall-runoff model calibration.” J. Hydrol. (Amsterdam), 268(1–4),
72–86.
parameters. It found the best parameter values in terms of the min-
Chu, H. J., and Chang, L. C. (2009). “Applying particle swarm optimization
imal sum of the square deviation between the observed and routed
to parameter estimation of the nonlinear Muskingum model.” J. Hydrol.
outflows. The performance of this approach was compared with Eng., 14(9), 1024–1027.
other common methods by using different criteria. For the model Das, A. (2004). “Parameter estimation for Muskingum models.” J. Irrig.
of Eq. (3), this method is more accurate than other ones (only in one Drain. Eng., 130(2), 140–147.
case is NMS as accurate as BFGS). However, for the following Das, A. (2007). “Chance-constrained optimization-based parameter esti-
reasons NMS performance is better than BFGS: (1) NMS produces mation for Muskingum models.” J. Irrig. Drain. Eng., 133(5),
improved results of sensitivity analysis of the initial parameter 487–494.
Downloaded 17 Dec 2011 to 35.8.11.2. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Fan, S. K., Liang, Y. C., and Zahara, E. (2006). “A genetic algorithm and McCarthy, G. T. (1938). “The unit hydrograph and flood routing.” Proc.,
a particle swarm optimizer hybridized with Nelder-Mead simplex Conf. of the North Atlantic Division, U.S. Army Corps of Engineers,
search.” Comput. Ind. Eng., 50(4), 401–425. New London, CT.
Geem, Z. W. (2006). “Parameter estimation for the nonlinear Muskingum Mohan, S. (1997). “Parameter estimation of nonlinear Muskingum models
model using the BFGS technique.” J. Irrig. Drain. Eng., 132(5), using genetic algorithm.” J. Hydraul. Eng., 123(2), 137–142.
474–478. Nelder, J. A., and Mead, R. (1965). “A simplex method for function
Gill, M. A. (1978). “Flood routing by Muskingum method.” J. Hydrol. minimization.” Comput J. (UK), 7(4), 308–313.
(Amsterdam), 36(3–4), 353–363. Ouria, A., and Toufigha, M. M. (2009). “Application of Nelder-Mead
Hydrologic Engineering Center. (2000). Hydrologic modeling system simplex method for unconfined seepage problems.” Appl. Math.
HEC-HMS: Technical reference manual, U.S. Army Corps of Engi- Modell., 33(9), 3589–3598.
neers, Davis, CA.
Singh, V. P., and Scarlatos, P. D. (1987). “Analysis of nonlinear Muskingum
Kim, J. H., Geem, Z. W., and Kim, E. S. (2001). “Parameter estimation of
flood routing.” J. Hydraul. Eng., 113(1), 61–79.
the nonlinear Muskingum model using harmony search.” J. Am. Water
Tung, Y. K. (1985). “River flood routing by nonlinear Muskingum
Resour. Assoc., 37(5), 1131–1138.
Lagarias, J. C., Reeds, J. A., Wright, M. H., and Wright, P. E. (1998). method.” J. Hydraul. Eng., 111(12), 1447–1460.
“Convergence properties of the Nelder–Mead simplex method in low Wilson, E. M. (1974). Engineering hydrology, MacMillan, Hampshire, UK.
dimensions.” SIAM J. Optim., 9(1), 112–147. Yang, W., Cao, W., Chung, T., and Morris, J. (2005). Applied numerical
Lasheen, A. A., El-Garhy, A. M., Saad, E. M., and Eid, S. M. (2009). methods using MATLAB, Wiley, Hoboken, NJ.
“Using hybrid genetic and Nelder-Mead algorithm for decoupling of Yoon, J. W., and Padmanabhan, G. (1993). “Parameter estimation of linear
MIMO systems with application on two coupled distillation columns and nonlinear Muskingum models.” J. Water Resour. Plann. Manage.,
process.” Int. J. Math. Comput. Simul., 3(3), 146–157. 119(5), 600–610.
Luo, J., and Xie, J. (2010). “Parameter estimation for nonlinear Muskingum Zahara, E., and Kao, Y. T. (2009). “Hybrid Nelder–Mead simplex search
model based on immune clonal selection algorithm.” J. Hydrol. Eng., and particle swarm optimization for constrained engineering design
15(10), 844–851. problems.” Expert Syst. Appl., 36(2), 3880–3886.
View publication stats Downloaded 17 Dec 2011 to 35.8.11.2. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org