ISBN 978-952-5726-10-7 Proceedings of the Third International Symposium on Computer Science and Computational Technology(ISCSCT 10) Jiaozuo, P. R.

China, 14-15,August 2010, pp. 283-288

Parameter Optimization of Multi-tank Model with Modified Dynamically Dimensioned Search Algorithm
Xiao-Lan Huang 1, Jun Xiong 2
Institute of Porous Media Mechanics,Wuhan Polytechnic University, Wuhan 430023, China E-mail: hxl_huang0226@sina.com 2 Social and Environmental Engineering Department,JP Business Service Corporation, Tokyo 135-8541, Japan E-mail: xiongjun79@gmail.com AbstractMulti-tank model is proposed for heavy rainfall infiltration simulation since it can represent the nonlinear transport behavior and give predictions very quickly. On the other hand, because its parameter space is high dimensional, it is difficult to obtain optimal parameters using existing methods. A new optimization approach called modified dynamically dimensioned search (MDDS) is developed for parameters calibration of this new model. It is based on heuristic global search and its adjustment is achieved by dynamically and randomly reducing the number of searching dimensions. Multi-tank model with 32 parameters is applied to an actual case. According to the results, better agreements between observations and calculation results are obtained compared with genetic algorithm. It is clarified that this model is a helpful tool in prediction of water table and stability factor of the slope.
Index Termsmulti-tank model; parameter optimization; modified dynamically dimensioned search; groundwater table prediction; slope stability factor also the solutions are not so good. Therefore, in order to facilitate calibration process, the development of a new method is needed. Basically, during the studies of estimating parameters, the calibration function of multi-tank model is replaced with a nonlinear optimization function, for example, the most popular way is to minimize the errors between calculations and measurements. In this article, water table observations are adopted as criterion function. Using genetic algorithm to find the solution of tank models parameters, if it is regarded as an inverse problem, it is an ill-posed problem without uniqueness of the solution especially when parameters are over 30. At the same time because too many parameters will lead to distribution order increase with geometric series, it also presents a significant computational burden. Actually if we consider obtaining results in the limited number of model evaluations, the idea of achieving global optimality will become unreasonable in most automatic calibration process. Therefore for high dimension optimization problems, a better multipoint random optimization method is necessary, and in this article a new approach called modified dynamically dimensioned search is provided as one of such good algorithm that focused on identifying good calibration results when calculation time is limited.
1

I. INTRODUCTION
It is noted that heavy rainfall may cause landslide and incur significant damage to the local area. In order to estimate the slope stability before collapse happens, the first step is to provide accurate water table forecasting. Sugawara and Takahashi [1,2] proposed a simple tank model to describe the long-term stream flow of specified drainage area. Because it can represent the non-linear stream flow behavior, it is widely used for long-term runoff analysis. Recently it is gradually developed to estimate the groundwater fluctuation of slope during rainstorm. In real cases complicate tank distributions are required, which results in high dimensional parameter space. Because of many parameters, it is difficult to properly identify those based on observed data. In previous literatures some researches tried to find solutions for tank model with 4 tanks (16 parameters) using many methods: Kobayashi and Maruyama[3] applied Powells conjugate direction method to the problem. Watanabe[4] suggested Newtons method. Yasunaga[5] and Hino[6] tried sequential estimation using Kalman filter. Tanakamaru[7], Suzuki[8] tried to use genetic algorithm (GA) as an efficient search procedure. In this paper, multi- tank model is more complicate: four series of tanks are introduced, and 32 parameters have to be retrieved from observations by use of optimization functions. For such high dimensional parameter space, GA is very time-consuming and
2010 ACADEMY PUBLISHER AP-PROC-CS-10CN007

II.

MULTI-TANK MODEL AND ITS PARAMETER DETERMINATION

Tank model is composed of one or several series of tanks with some outlets on the side and bottom in each tank. The basic concept is illustrated as Fig.1. From this figure it is obvious that rainwater in the tank will accumulate until water level becomes over than Z, then lateral flow happens. Outflow through the side outlets represents components of the total discharge due to the immediate or delayed response to the rainfall. Flow through the bottom holes means the portion of infiltrating flow and does not contribute to the surface flow directly.
Rainfall q, a q: outflow
WL

a: discharge coefficient Z: heights of lateral outlets WL: water level in the tank

Figure 1. Basic concept of one tank

283

A. Schematic figure of multi-tank model in the slope In this study, in order to monitor the water table of the slope and estimate the slope stability factor during rainfall, four series of multi-tanks are distributed as shown in Fig-2, which is designed to simulate rainfall infiltration process and runoff responses. One of these interconnected tanks is set at the highest position to serve as the base point, which is followed by another two at the medium position and the last one at the lowest position. In this way, they can be designed to evaluate the major groundwater behavior of the slope.
r r
r r
q1 a1
W L3

lower tanks, then groundwater table GWTi at a specific time t is calculated by the following equation:
GWTi (t ) = GWTi (0) + WLbot i (t ) i

(3)

Where: is effective porosity of the soil where tanks are set. GWTi(0) is the reference groundwater level at initial state, which must be small enough based on available observations. Generally the reference groundwater levels in four lower tanks are defined as smaller than the lowest water table in the history. B. Definition of optimization function In order to find appropriate solution, optimization function is necessary. Here this evaluation functions is to minimize the errors between calculations and measurements, which is defined as the following equation.

Z1

q3 a3
W L5

W L1 h11 b1

I1

q8 a8

q5 a5
Z8 W L7

Z3

b2

I2

Z5

b4
I4

b3

q2 a2
Z2 W L2

I3

q4 a4
Z4
W L4

J XS =

q9 a9

W L8
Z22 h 9

a10 a10

q6 a6 q7 a7

W L6

1 M

i =1

(Qc (i ) Qo (i )) 2 Qc (i )

(4)

Top tank
Z6
Z7

Z43 h 10

Middle tank

Bottom tank

Figure 2. Parameter distribution of multi-tank model

Where: Qo(i) is water table observation of four lower tanks; Qc(i) means calculation results; M is the number of observed data. Generally there are 32 parameters that needed to be estimated by use of the developed new optimization method. C. Definition of modified dynamically dimensioned search The dynamically dimensioned search algorithm[9] (DDS) is a novel and simple stochastic single-solution method, and it is based on heuristic global search algorithm that was developed for the purpose of finding good global solutions within the specified maximum function evaluation limit. In short, the algorithm searches globally at the start of the search and becomes more and more local as the number of iterations approaches the maximum allowable number of function evaluations. The adjustment from global to local search is achieved by dynamically and randomly reducing the number of dimensions in the neighborhood. The decision variables in automatic calibration are the model parameters, and the dimension being varied is the number of model parameter, which is changed to generate a new search neighborhood. Candidate solutions are created by perturbing the current solution values randomly selected dimensions only. Its perturbations magnitudes are sampled from a normal distribution N(0,1).
The DDS algorithm is unique relative to current other random optimization approaches because of the way that neighborhood is dynamically adjusted by changing the dimension of the search. For example if GA is adopted, the final solutions will be different for every calculation process. The only algorithm parameter to set in DDS is the scalar neighborhood size perturbation parameter (r) that defines the random perturbation size as a fraction of the decision variable range. An initial value of the r parameter is set as 0.3. Because multi-tank models parameter space dimension is over 30, in order to improve searching efficiency, with the calculation process going on, r will reduce step by step, the minimal value 284

The new models flow patterns are shown as in Fig.2. According to water balance theory at the specific time, its discrete flow equation can be evaluated:
WL1 (t ) = WL1 (t 1) + r (t ) q1 (t ) I1 (t ) WL2 (t ) = WL2 (t 1) + I1 (t ) q2 (t ) WL3 (t ) = WL3 (t 1) + r (t ) + q1 (t ) q3 (t ) I 2 (t ) WL4 (t ) = WL4 (t 1) + I 2 (t ) + q3 (t ) q4 (t ) WL5 (t ) = WL5 (t 1) + r (t ) + q3 (t ) q5 (t ) I 3 (t ) WL6 (t ) = WL6 (t 1) + I 3 (t ) + q4 (t ) q6 (t ) q7 (t ) WL7 (t ) = WL7 (t 1) + r (t ) + q5 (t ) q8 (t ) I 4 (t ) WL8 (t ) = WL8 (t 1) + I 4 (t ) + q6 (t ) + q7 (t ) q9 (t ) q10 (t )

(1)

Where: r(t) is rain intensity (mm/day), ai is the coefficients of runoff from the side hole of the tank; bi means coefficients of seepage from the bottom hole; Zi represents the height of the runoff on the side of tanks; and qi is seepage runoff volume from the side of the tank; WLi (t ) represents water level in ith tank at time t;. Here the lateral flow discharge qi(t) and vertical seepage volume Ii(t) at one specific time assumed to be proportional to its corresponding tank water level, which can be evaluated by equation (2):
I i = bi WLtop i qi = ai (WLi Z i ), qi 0, (i = 1 6) q7 = a7 (WL6 Z 7 ), q7 0 q8 = a8 (WL7 Z8 ), q8 0 q9 = a9 (WL8 Z 9 ), q9 0 (WL8 q10 = a Z10 ), q10 0 WL , WL , WL and 10 four Water levels in lower tanks: 2 4 6 WL8 are considered to be related to groundwater table of the slope. After obtaining water levels of all the four

(2)

Figure 3. Configuration of multi-tank model in the slope

of r is 0.05, which is different from Tolson[9]. This initial sampling region size is designed to allow the algorithm to escape regions around poor local minima. In the final stage because the current solution is close to final results, in order to avoid big perturbation, the value of r must decrease. And also in order to accelerate the speed of convergence, its update algorithm (STEP5) is also modified: if the change in objective function value is positive, the new solution is accepted with a certain probability; whereas in previous literatures it is abandoned simply. The calculation process of modified dynamically dimensional search (MDDS) algorithm is provided as the follows: STEP1. Define inputs: z Neighborhood perturbation size parameter: r(the default initial value is 0.3); z Vectors of lower, xmin, and upper, xmax, and initial solution, x0 = [x1,x2,,xm]. STEP2. Set counter to 1, i=1, and evaluate objective function F at initial solution, F(x0): z Fbest = F(x0), and xbest = x0 STEP3. Randomly select J of the m parameters for inclusion in neighborhood, {N}. STEP4. For j=1,,J decision variables in {N}, perturb x best j using a standard normal random variable: N(0,1), reflecting at decision variable bounds if necessary:

if (0.05<Pn.and.Pn<0.1) r_val=Pn if (Pn<0.05) Pn=0.05; r_val=0.05

III. CASE STUDIES ON THE ACTUAL SLOPE USING MULTITANK MODEL

Based on previously suggested procedures, multi-tank model is applied to the slope along Japanese national road No.12 in Yamagata Prefecture to simulate fluctuations of groundwater table induced by during rainfall period.
A. Outline of the slope
From the boring survey results, it is revealed that in the slope, weathered rock is about 3 to 10 meter thick. With the history of collapses, it was regarded that it is urgent to determine its water table fluctuations and evaluate its stability. As illustrated in Fig-3, it is the configuration of multi-tank model in the slope. Top tank (tank2) is assumed on the top hill(x=325m); bottom tank (tank8) lies on the lowest part of the slope (x=7m). It is found that the surface layer is almost homogeneous, therefore middle tank distribution is relatively easy: totally there are two series. Because a mound exists near the foot of slope, it has big influence on rainfall infiltration, Tank6 is set at the boundary of the mound (x=98m). Tank4 is set at Boring No.B2-707 (x=225m) to serve as slope lateral flow simulation and help to provide more accurate water table forecasting. The porosities near the four parts are 0.09, 0.12, 0.16 and 0.15 respectively. As aforementioned, tank models parameters are difficult to be directly measured by experiments, therefore, in order to evaluate model parameters, historical data of rainfall and ground water table are required. In this case, there are four observation borings along this cut-slope: boring No.B3-D2(x=310m), B2-707 (x=225m), B2-706(x=165m), and B3-D5(x=98m) are drilled to monitor the ground water table at the four locations. Rainfall intensities and ground water tables of 100 days have been recorded for parameters optimization.

STEP5. Evaluate F(xnew) and update current best solution if necessary:

If F(xnew)Fbest, update new best solution: F F(xnew) and xbestxnew z If F(xnew)>Fbest new best exp( (F -F ) f(j)) >random(Pn) z
best

and

FbestF(xnew) and xbestxnew

STEP6. Update iteration count, i=i+1, and check stopping criterion:

z z If i= Maxiter, STOP, print output (e.g: Fbest and xbest ) Else go to STEP3

The only parameter r is defined as the following lines: Pn decreases with the increase of the number of function evaluations (NF is maximum number of function evaluation; i is the current calculation step): Pn=1.0-dlog(dfloat(i))/dlog(dfloat(NF)
if(0.3<Pn.) r_val=0.30 if (0.2<Pn.and.Pn<0.3) r_val=Pn if (0.1<Pn.and.Pn<0.2) r_val=Pn 285

B. Analytical conditions
The parameters of multi-tank model can be identified by the reproduction of observed hydrographs assuming that basic watershed characteristics remain unchanged during the observation of events. In the runoff analysis, because the studied slope area is fairly small in size, the travel time is considered to be relatively short. The data used for the analysis are the 100 daily measurements of precipitation (shown as Fig. 4 between 1994-5-27 and 1994-9-4) and water tables averaged

over catchment of the slope. By using multi-tank model, water tables can be obtained; and using optimization method of modified dynamically dimensioned search algorithm, parameters are retrieved. Here there are totally 32 parameters; their lower and upper bounds of the search for parameters are listed in Table1. The bounds of search are set based on the result of an application of the three-series tank model. The maximum number of function evaluation (Maxiter) is 4000, and it is considered large enough for practical purposes.
80. 0 Rai nI nt ensi t y( m m ) 60. 0 40. 0 20. 0 0. 0
1994- 5- 27 0: 00 1994- 6- 21 0:00 1994- 7- 16 0: 00 1994- 8-10 0: 00 1994-9- 4 0: 00

but with the calculation process going on, they go down quickly. At the same time, because of its stochastic nature of modified dynamically dimensioned search, the perturbation is very big at the initial stage, which means the algorithm can escape from regions around poor local minima. Eventually, because the current solution is close to final results, in order to avoid big perturbation, perturbation amplitude becomes smaller and smaller gradually, the current solution is close to the final results.
TABLE2 OPTIMIZATION SOLUTIONS BY MDDS
a(1)=0.126 a(2)=0.107 a(3)=0.36 a(4)=0.113 a(5)=0.318 a(6)=0.122 a(7)=0.119 a(8)=0.386 a(9)=0.099 a(10)=0.092 b(1)=0.05 b(2)=0.05 b(3)=0.478 b(4)=0.468 Z(1)=0.397 Z(2)=0.033 Z(3)=0.064 Z(4)=0.080 Z(5)=0.016 Z(6)=0.213 Z(7)= 0.011 Z(8)=0.011 Z(9)=0.302 Z(10)= 0.011 Unit/ m WL0 (1)=0.012 WL0 (2)=0.173 WL0 (3)=0.013 WL0 (4)=0.207 WL0 (5)=0.015 WL0 (6)<0.069 WL0 (7)=0.015 WL0 (8)=0.017 Unit/ m

El apsed t i m e( day)

Figure 4. Daily rainfall intensities TABLE1. PARAMETER BOUNDS

0.0<a(1)<0.6 0.0<a(2)<0.6 0.0<a(3)<0.6 0.0<a(4)<0.6 0.0<a(5)<0.6 0.0<a(6)<0.6 0.0<a(7)<0.6 0.0<a(8)<0.6 0.0<a(9)<0.6 0.0<a(10)<0.6 0.0<b(1)<0.6 0.0<b(2)<0.6 0.0<b(3)<0.6 0.0<b(4)<0.6
Funct ion Val ue

0.005<Z(1)< 0.4 0.005<Z(2)< 0.4 0.005<Z(3)< 0.4 0.005<Z(4)< 0.4 0.005<Z(5)< 0.4 0.005<Z(6)< 0.4 0.005<Z(7)< 0.4 0.005<Z(8)< 0.4 0.005<Z(9)< 0.4 0.005<Z(10)< 0.4 Unit/ m

Function Value

0.0< WL0 (1)<0.5 0.0< WL0 (2)<0.5 0.0< WL0 (3)<0.5 0.0< WL0 (4)<0.5 0.0< WL0 (5)<0.5 0.0< WL0 (6)<0.5 0.0< WL0 (7)<0.5 0.0< WL0 (8)<0.5 Unit/ m

40 30 20 10 0 0 500 1000 150 0 2000 2500 3000 3500 40 00 E va lua tion N umbe r
1.5

C. Analytical results The analysis results of by dynamically dimensioned search are shown in Table 2. It is found that lateral coefficients of upper tanks become bigger and bigger from top to bottom (a(1)= 0.126, a(3)= 0.360, a(5)= 0.318, a(8)= 0.386); the exception is a(5), which lies in the existence of the mound. At the same time, corresponding height of lateral outflow hole become smaller and smaller (Z(1)= 0.397, Z(3)= 0.064, Z(5)= 0.016, Z(8)= 0.011), which means the top part is dry, and it needs more infiltrating water until surface runoff happens. On other hand, at the bottom, surface flow comes into being easily (Z(1) is much bigger than Z(8)). All these phenomenon are consistent with reality. The initial water levels in four upper tanks also have the same order: water level in lower positions has bigger values. From this table, it is demonstrated that multi-tank model can still provide good forecasting of surface stream flow and modified dynamically dimensioned search is very good method that calibrate reasonable parameters.
The final calculation error JXS is represented by 0.05812. Objective function values plotted against the number of function evaluation is shown in Fig 5. From the figures, it is clear that at the beginning objective function values are over 50, 286

0.5

0 0 5 00 1000 1 5 00 2 00 0 2 50 0 30 0 0 3500 4 0 00 E va lu at io n N u m b er

Figure 5. Objective function value against number of function evaluations

Differences of water levels at tank2 and 8 between observations and calculation results are shown as Fig.6 and Fig.7 in the 100 days (from May 27th, 1994 to Sept. 4th, 1994), from which it can be concluded that in the two sites, water levels of observations and calculation results are similar, especially for tank8, they are almost the same. From the two figures, it is illustrated that the default settings of the neighborhood perturbation test functions (r) produced good results across the multi-tank model calibration problem. Therefore, the strategy for r seems reasonable and suggests its validity for most future application.

285 W at er tabl e(m ) Tank2(Obs ) 280 275 Tank2(MD DS)

sharply, which also demonstrates that rainfall has great influence on slope stability.

265 W at er t abl e( m )

T a n k 4 ( Ob s ) T a n k 4 ( MDDS )

270 0 10 20 30 40 50 60 70 80 90 100 El aps ed t i m e(day)

260

T a n k 4 ( GA )

255

Figure 6. Comparison between optimal results and observations at tank2

235 Tank 8( Obs ) Tank 8( MD DS) 230

250 0 20 40 60 80 100

E l a p s e d t i me ( d a y )

Fig.8 Comparison between MDDS and GA at tank4

245 W at ert abl e( m ) 240 235 230 0 20 40 60 80 10 0 El a ps ed ti me (da y) Ta nk 6(Obs ) Ta nk 6(MDDS) Ta nk 6(GA)

225

220 0 10 20 30 40 50 60 70 80 90 100

Figure 7. Comparison between optimal results and observations at tank8

In the aforementioned parameters calibration procedures of tank model, Tanakamaru (1993), Suzuki(1999) and others has tried to use genetic algorithm (GA) as an efficient search procedure, where totally there are 16 parameters. Here, in order to compare the effectiveness of MDDS and GA for parameters estimation of multi-tank model, GA will also be tried for parameter estimation of multi-tank model. For multi-tank model, totally there are 32 parameters, in order to improve the quality of the results, multiple populations has been employed, which is known as the migration, or island model. Each subpopulation is evolved over generations by a traditional GA, within each subpopulation it is necessary to perform functions such as selection, crossover and reinsertion. And from time to time individuals migrate from one subpopulation to another. Comparisons of water levels between observations and calculation results (both MDDS and GA) at two positions are illustrated in Fig.8 and Fig.9 during the 100 days (from May 27th, 1994 to Sept. 4th, 1994), from which it can be concluded that in the two sites, results of MDDS is very good; while water tables obtained from GA is bad: obviously it is trapped in the local optimal solution. Reproductions of water levels at four locations are shown from Fig.6 to Fig.9), from which it can be concluded that in all four sites, water levels of observations and calculation results from our new strategy are similar, especially for tank6 and tank8, they are almost the same. After getting the water tables of the several locations during heavy rainfall, the instantaneous groundwater lines can be estimated quickly by spline interpolation method. Then the slope stability factor is calculated with the division methods commonly. In the division method, there are many methods such as Fellenius, Bishop, Janbu and Spencer. Fig.10 is the results of stability analysis of this slope using Bishop method during the analysis period: it is clear that when rain is big, its stability factor decreases

Figure 9. Comparison between MDDS and GA at tank6

St abi l i t y f act or 1. 5 1. 45 1. 4 1. 35 1. 3 1. 25 1. 2 0 20 40 60 80 100

El apsed day Figure 10. Slope stability factor changes during rainfall

IV. CONCLUSIONS The paper focused on the behavior of rainfall infiltration process and aimed to develop a simple and quick analytical tool to evaluate underground water table and slope stability factor during heavy rainfall. The insights gained through this study are summarized as the following: According to water balance (tracking flows of water into and out of the particular hydrologic system of interest), a multi-connected tank model that can reproduce the rainwater movement behavior was developed. A new stochastic single-solution method called modified dynamically dimensioned search was adopted to identify optimal solutions. This method is robust for parameter calibration with high space, since it generated relatively good solutions without requiring any algorithm parameter adjustments. Compared with genetic algorithm, the new method could find relatively good solutions in a shorter time. Multi-tank model was applied to the actual slope. Its consistency with field data was confirmed, and its practicability was proved. Meanwhile, its algorithm is very simple and thus it can be easily coded in any programming language. Although this study focused on multi-tank model, the results are just as relevant to all
287

environmental simulation modelers who is calibrating parameters of a computationally demanding model. Although multi-tank model has the advantage of predicting water table fluctuations quickly, it can not give information of infiltration process in the unsaturated zone. The authors are planning to work on the development of unsaturated tank model; therefore the stability for shallow landslide can be assessed during by rainfall. Combined with a new accurate rain gauge, the methodology will be evolved further into an assessment system for correctly predicting the hazards of rainfall that may lead to slope failure. In addition, it is necessary to test the modification and improve the ability of MDDS to locate the exact global optimum or implement a parallelized version of this algorithm. REFERENCES
[1] M. Sugawara, E. Ozaki, I. Watanabe, etc, Method of automatic calibration of tank model(second report), Research notes of the National Research Center for Disaster Prevention, Japan, Vo17,43-89,1978.( in Japanese) [2] K. Takahashi, Y. Ohnish, J. XIONG and T. Koyama, Tank Model and its Application to Groundwater Table

[3]

[4]

[5]

[6]

[7]

[8]

[9]

Prediction of Slope, Chinese Journal of Rock Mechanics and Engineering, Vol.27(12): 2501-2508, 2008. S. Kobayshi and T. Maruyama, Search for the coefficients of the reservoir model with the Powers conjugate direction method, Trans. JSIDRE, No.65, pp.42-47, 1976. K. Watanabe, Refinements to parameter optimization in the tank model, Proceedings of the 33rd Japanese Conference on Hydraulic, pp.55-60, 1989. (in Japanese) T. Yasunaga, K. Jinno and A. Kawamura, Change in the runoff process into an irrigation pond due to land alteration, Proceedings of the 36th Japanese Conference on Hydraulics, Japan, 1992: 629-634(in Japanese). M. Hino, Prediction of hydrologic system by Kalman Filter, Proc. of Japan Society of Civil Engineers, 1974, No.221: 39-47. (in Japanese) H. Tanakamaru, Parameter identification of tank model with the genetic algorithm, Annuals of Disaster Prevention Research Institute, No.36B-2, 1993:231-239 (in Japanese). K. Suzuki, H. Momota, H. Takahashi, Statistical study of tank model identification by genetic algorithm, Japanese Journal of Hydroscience and Hydraulic Engineering, 1999, VOL.17, No.1 May, 11-19. B.A.Tolson, C.A. Shemaker, Dynamically dimensioned search algorithm for computationally efficient watershed model calibration,Water Resources Research, VOL.43, W01413, 2007.

288