Numge PDF
Numge PDF
Numge PDF
ABSTRACT: Several numerical models have been developed for seepage analysis in embankment dams.
However, these methods require a specification of the initial and boundary conditions and the spatial distribution
of hydraulic parameters which are not easily measured. Therefore, data modeling tools that are able to capture
and represent complex input/output relationships, such as Artificial Neural Networks (ANNs), are potential tools
for obtaining more accurate results. The principal objective of this study was to train a neural network model to
predict water levels in piezometers in an embankment dam using Dadin Kowa dam in Gombe State, Nigeria as a
case study. A feedforward 3-layer network employing the backpropagation algorithm for network learning was
used. The model results show that predicted water levels compared satisfactorily with those measured on the field.
This study offers insight to the effectiveness of ANNs in monitoring seepage flow through an embankment dam.
479
Xu et al. (2003) designed a hydraulically optimal earth
dam cross section.
Generally, numerical methods require a compli-
cated technique for the solution of model equations,
determination of initial and boundary conditions as
well as the spatial distribution of hydraulic parameters.
The complexity of the phenomena governing seepage
through an embankment dam and the lack of infor-
mation concerning the boundary conditions makes it
Figure 1. Dam section showing some of the piezometers.
very difficult to build a finite element deterministic
model. In practical situations, satisfying all the data
needs of a comprehensive FEM is seldom available
due to time and budgetary constraints. Furthermore, 2 MATERIALS AND METHODS
there is always a problem of convergence and insta-
bility in the numerical solution of the highly nonlinear 2.1 Case study
differential equation of the physics based model. These The Dadin Kowa dam was used as a case study to
factors reduce the accuracy of the model in predicting carry out this investigation. It is located in North-
seepage flow. Eastern Nigeria in Yamaltu-Deba Local Government
Tayfur et al. (2005) developed a FEM model along- in Gombe State. It lies on Lat.N10 18 and Long E
side an Artificial Neural Network model, to predict 11 31 . The dam is located 5 km north of the village
seepage path through the body of Jeziorsko dam, of Dadin Kowa on River Gongola in Gombe State. Its
Poland. The piezometer water levels computed by the major basin is the Benue River basin. It has a capac-
models satisfactorily compared with those measured ity of 2 885 000 000 m3 and area 300 000 000 m2 . The
by the piezometers. However, the model results also reservoir extends northwards for a distance of 65 km
revealed that the ANN model performed as good as approximately to where the Borno extension railway
and in some cases better than the FEM model. In recent line crosses the Gongola River in Ashaka.
time, both in research and practical applications, neu- It is a rockfill embankment which consists of a
ral networks have proven to be a very powerful method central core of silty clay, supported upstream and
of mathematical modeling. downstream by rockfill shells in which the maxi-
mum aggregate size ranges from 300 to 1200 mm. The
1.2 Artificial Neural Networks (ANNs) embankment is 42 m high with a length of 520 m and
Artificial neural networks are powerful data mod- a base width of 230 m. Figure 1 shows a schematic
eling tools that are able to capture and represent representation of a section of the dam.
complex input/output relationships. The motivation The transition zone between the core and the shell
for the development of neural network technology consists of 3 m of granular filter formed of a mix-
stemmed from the desire to develop an artificial sys- ture of crushed rock and natural sand as shown. The
tem that could perform intelligent tasks similar to embankment is founded on river sand (up to 18.5 m
those performed by the human brain. The true power thick) at the river section, and on bedrock (a sandstone
and advantage of neural networks lies in their ability with occasional mudstone intercalations) at the banks.
to represent both linear and non-linear relationships The bedrock is overlain by a thick alluvial deposit, up
and in their ability to learn these relationships directly to 70 m wide along the dam axis, which consists of
from the data being modeled. Traditional linear mod- medium to coarse sand with discontinuous zones rich
els are simply inadequate when it comes to modeling in gravel and/or silt. The river sand which has an in
data that contains non-linear characteristics. situ hydraulic conductivity of about 0.00010.1 cm/s
Over the last few years, the use of Artificial Neu- (before densification) attains a maximum thickness
ral Networks (ANN) has increased in many areas of of about 18.5 m along the dam axis (Chido-Amajuoyi
Engineering. In particular, ANNs have been applied and Ofoegbu, 1987).
to many geotechnical engineering problems and some
2.2 Methodology
degree of success has been recorded. A review of liter-
ature reveals thatArtificial Neural Networks have been To create the network prediction model, the NNTOOL
recently employed for the solution of many hydraulic, application in MATLAB was used. Water levels for
hydrologic and water resources problems ranging from twenty-five piezometers in the embankment sections,
rainfall runoff (Tokar and Johnson 1999) to sedi- the alluvial foundation and the bedrock were collected
ment transport (Jain 2001; Nagy et al. 2002) to solute from the Dadin Kowa dam from 6 September, 1983
transport (Aziz and Wong 1992; Lu et al. 1998). In to 21 October, 1998. This much data was required in
embankment dams, ANNs have been used to predict order to select the year which contains all the pos-
seepage path in a homogenous earthfill dam in Poland sible variations in water rise in the upper reservoir.
(Tayfur et al. 2005). They have also been used sat- Therefore, water levels from the period of 17 October,
isfactorily to estimate peak outflow from breached 1989 to 10 September, 1990 was selected and used
embankments (Amini et al. 2011; Hoosyaripor and to calibrate the network as this showed a wider varia-
Tahershamsi 2012). tion in the reservoir level. Similarly, water levels from
480
Figure 2. Network topology. X Reservoir water level,
Y Tailwater level, Z Identification of piezometer, O
Piezometer water level. Figure 3. Training process.
481
Figure 7. Correlation Coefficient for Piezometers in the
Validation Run.
Figure 5. Correlation Coefficient for Piezometers in the
Training Run.
482
The field data was used to calibrate and validate the Elinwa, A.U. & Chida-Amajuoyi, G.U. 2000. Geotechni-
ANN model. The model performance was evalu- cal instrumentation programme in Dadin Kowa dam
ated by carrying out a regression analysis on the project pore water pressure analysis. African Journal
predicted model results with the measured water of Enviromental Studies 1(1), 123126.
Honjo, Y., Giao, P.H. & Naushashi, P.A. 1995. Seepage
levels. The coefficient of determination (R2 ) for analysis of Tarbela dam (Pakistan) using finite element
the training and validation runs was 0.997 and method. International Journal Rock Mechanics. Min. Sci.
0.994 respectively. Similarly, the slope of regres- Geomech., 32(3), 131A.
sion line obtained from the analysis was 0.995 and Hooshyaripor, F. & Tahershamsi, A. 2012. Comparing the
0.99 respectively. This suggests that neural network Performance of Neural Networks for Predicting Peak
model selected successfully captured the relation- Outflow from Breached Embankments when Back Propa-
ship between the reservoir, tailwater levels and the gation Algorithms Meet Evolutionary Algorithms, Inter-
piezometer water levels. national Journal of Hydraulic Engineering, 1(6) 5567.
Further analysis revealed that although the overall International Committee on Large Dams, 1987, Dam safety
guidelines, Bulletin 59, ISSN 0534-829.
performance of the model was successful, the model International Committee on Large Dams, 1995, Dam Failures
performed better in some piezometers than others. Statistical Analysis, Bulletin 99, ISSN 0534-8293
The ANN model predicted water levels in the Jain, S.K. 2001. Development of integrated sediment rating
piezometers using only the reservoir and tailwater curves using ANNs. Journal of Hydraulic Engineering,
levels which are easily measured field data. It is 127(1), 3037.
therefore a simpler method for monitoring seepage Lu, R.-S., Lai, J.-L. & Lo, S.-L. 1998. Predicting solute trans-
in embankment dams fer to surface runoff using neural networks. Water Sci.
Technology, 38(10), 173180.
This study provides more insight into the excellent McCully P. 2007. Dams can pose security risk to Africa.
approximating capability of the neural network as it Retrieved February 2009 from http://news.mongobay.
was used to capture the relationship between upstream com/2007/1004-irn.html
and downstream water levels in an embankment and Nagy, H.M., Watanabe, K. & Hirano, M. 2002. Predic-
water levels in the piezometers. It must be noted how- tion of sediment load concentration in rivers using using
ever, that the ANN model is a data-driven black box artificial neural network model. Journal of Hydraulic.
Engineering., 128(6), 588595.
model which does not reveal any explicit relationship Naouss, A.W. & Najjar, Y.M. 1996. Seepage design charts
between input and output variables and hence, does not for flat bottom dams resting on heterogenous media. Iner-
provide any useful insight to the solution of the physi- national Journal Rock Mech. Min. Sci. Geomech. Abstr.,
cal problem. In addition, its performance is extremely 33(3), 136A. Sharma, H.D. 1991. Embankment dams.
sensitive to the data used during testing and calibra- Oxford and IBH Publishing Company. New Delhi.
tion. In other words, the model is only as accurate as Tayfur, G., Swiatek, D., Wita, A., & Singh, V. P. 2005. Case
the data used in calibration. study: Finite element method and artificial neural network
Furthermore, the inability of neural networks to models for flow through Jeziorsko earthfill dam in Poland.
extrapolate beyond the range of data used for train- Journal Hydraulic Engineering 131(6), 431440.
Tokar, A.S & Johnson, P.A. 1999. Rainfall-runoff model-
ing is widely acknowledged. Hence, ANN models are ing using artificial neural networks. Journal Hydrologic
usually site-specific and can only be used for cases for Engineering, 4(3), 232239.
which it was trained. Tien-Kuen, H. 1996. Stability analysis of an earth dam under
The satisfactory prediction of the piezometer water steady state seepage. Comput. Struct., 58(6), 10751082.
levels indicates that the ANN model, if trained with Turkmen, S., Ozguler, E., Taga, H. & Karaogullandarinda
adequate and sufficient data, can be used to verify the N. 2002. Seepage problems in the karstic limestone
piezometer water levels in an embankment dam. From foundation of the Kalecik Dam. Engineering Geology
a practical point of view, it could serve as a tool for Amsterdam, 63(34), 247257.
the detection of anomalies in the course of infiltrated Van Genuchten, R. 1979. Calculating the unsaturated
hydraulic conductivity with a new closed form analyti-
water and seepage hence enabling the planning and cal model. Research Rep., 1978-WR-08, Water Research
implementation of technical and economic remedial Program Department of Civil Engineering, Princeton
stability measures. Univ.
Xu, Y.-Q., Unami, K. & Kawachi, T. 2003. Optimal
hydraulic design of earth dam cross-section using
REFERENCES saturated-unsaturated seepage flow model. Advanced.
Water Resources., 26(1), 17.
Amini, B.A., Nourani, V., & Hakimzadeh, H. 2011. Appli- Yager, R.R., & Filev, D.P. 1994. Essential of Fuzzy modeling
cation of Artificial Intelligence tools to Estimate Peak and control, Wiley, New York.
Outflow from Earth Dam Breach. International Journal Zienkiewicz, O.C. & Cheung, Y.K. 1965. Finite Eements in
of Earth Sciences and Engineering. 4(06), 243246 . the Solution of field problems. The Engineer, London.
Aziz,A.R.A., & Wong, K.F.V. 1992. Neural network approach
to the determination of aquifer parameters. Groundwater
30(2), 164166.
Chido-Amajuoyi, G.U. & Ofoegbu, G.I. 1987. An analysis
of water pressure and deformation data for the Dadin
Kowa Dam. 9th Regional Conference for Africa on Soil
Mechanics and Foundation Eng., September 1987.
483