COMMISSION INTERNATIONALE
DES GRANDS BARRAGES
-------
VINGT-SIXIÈME CONGRÈS DES
GRANDS BARRAGES
Autriche, juillet 2018
-------
ASSESSMENT OF WATER FLOW MEASUREMENT IN A ZONED DAM USING
ARTIFICIAL NEURAL NETWORK MODELS
Ricardo C. SANTOS
Geotechnics Department, LABORATÓRIO NACIONAL DE ENGENHARIA CIVIL
Juan T. MATA
Concrete Dam Department, LABORATÓRIO NACIONAL DE ENGENHARIA
CIVIL
PORTUGAL
1. INTRODUCTION
This paper concerns about the assessment of the hydraulic behavior of a
50 m height zoned dam used for industry water supply, located in Alentejo
region, in Portugal. The original draining system of the embankment is composed
of a sub-vertical filter, located downstream of the central core, and by a drainage
blanket placed over the foundation, in the deeper area of the valley, which, in
turn, discharges to a downstream drainage toe. After first filling of the reservoir,
resurgences and artesian pressures in the terrain at downstream of the dam
body were identified in the left bank. Following that abnormal behavior, the
reservoir water level was restricted, and an inverted filter was built over the
affected area, to limit occurrence of internal erosion and heave. Additionally, to
monitor the hydraulic behavior of the dam and of the inverted filter, several
piezometers and a few flowmeters were installed.
Based on the monitoring data (piezometers, flowmeters and reservoir water
level) recorded in the last 30 years, a neural network (NN) model was established
for the hydraulic behavior prediction of the zoned dam. The results of this study
show that NN models can be a useful tool for the prediction of the flow rate
measured in the flowmeters.
2. SAFETY OF ZONED DAMS
Dams are water-retaining structures that are built to provide water for
human consumption, irrigating, generating hydroelectric power and use in
industrial processes. They are critical structures for the continuation of life and
providing public safety. Statistics from the International Commission on Large
Dams (ICOLD) [1,2] reveals that embankment dams represent the most common
dam construction type, totalizing about 3/4 of all the existing dams, and, of those,
88% are earthfill dams and 12% are rock-fill dams. Zoned dams account with
about 53% of the world population of embankment dams. The most common
zoning profiles are zoned earthfill (36%), zoned earth and rock-fill (9%), and
central core earth and rock-fill (8%). The main failure modes associated with
these type of structures are caused by overtopping (due to inadequate spillway
capacity or malfunction of gates), by piping, and by slope instability. Failure
modes by piping and overtopping together are responsible, in similar importance,
for more than 92% of the total failures [3].
Embankment dam failure and accident statistics [3] show clearly that high
average frequencies of failure are associated with dams using profile types
without zoning and with inherently poor control of seepage and pressures in the
embankment and foundation. On the contrary, embankment dams with
downstream rock-fill zones have a particularly low incidence of failure, mainly
because they are less likely to progress to breaching if piping erosion initiates
compared with dams with earthfill materials in the downstream zones.
In some case studies of zoned dams, leaks appeared at downstream of the
embankment, meaning that they bypassed the dam’s internal drainage system.
Typically, these pathologies appear during first filling or during the first years of
exploration of the reservoir. These incidents are usually associated to an
inadequate treatment of the foundation, or due to the presence of pervious soil
formations or interconnected cracks in a rock formation, not detected during
geological prospection of the construction site.
In this paper, we describe the history of a zoned dam, built in the 80s, in
which leaks appeared at downstream of the dam during first filling. The remedial
actions to overcome the abnormal hydraulic behavior of the foundation are also
indicated. Finally, we use a neural network, considering as input parameters the
reservoir level and piezometric data, to predict the flowrates measured in a
flowmeter located near the leakage area.
3. CASE STUDY
3.1. DAM DESCRIPTION
Figure 1 shows a satellite view of the dam in concern. It is a zoned earthfill
dam located in the south center region of Portugal, in particular, in Setúbal
district, in Alentejo. The dam crest, following a curve in the longitudinal axis, has
an extension close to 2700 m, and a maximum height above foundation of about
52 m (45 m above streambed). The completion of the dam dates the year 1980,
and the current dam owner is APA (Agência Portuguesa do Ambiente). The
dam’s reservoir aims mainly the supply of water to industry. The reservoir has an
effective storage of 27000 ML.
Fig. 1
Satellite view of the dam (source: Google maps)
Figure 2 shows the typical cross section of the embankment. The crest is
defined at level 70 m above sea level and the normal water level (NWL) of the
reservoir was initially established from design at 68.3 m. The upstream shell of
the embankment has a slope of 1(V):3(H), whereas the downstream shell has a
slope of 1(V):2.4(H). The core is made of selected weathered schist materials
and has a slope of 2(V):1(H) and 2.5(V):1(H) at the upstream and downstream
sides, respectively. The internal drainage system is composed by sub-vertical
filter/drains, located at downstream and upstream of the impervious core, a
horizontal drainage blanket, located in the downstream shell and above the
foundation, and a downstream toe drain, below level 41.5 m.
Fig. 2
Cross sectional profile of the embankment of the dam
The dam has a siphon type spillway, located on the left bank, and a bottom
outlet, composed by a conduit, under the embankment (in the deepest zone of
the valley).
The geological setting for the embankment is characterized by
metamorphic rock formations mainly of schist and greywacke. Preparation works
involved slush-grouting of the core trench and removal of superficial deposits
beneath the shell zones.
3.2. THE MONITORING SYSTEM OF THE DAM
The initial monitoring system of the dam was composed of survey points, to
measure vertical displacements of the embankment, inclinometers, to measure
the horizontal displacements in the embankment, hydraulic piezometers in the
embankment and foundation, to measure pore pressure, and a flowmeter located
in the bottom of the valley, to assess the seepage water through the
embankment and foundation. Figure 3(a) represents the majority of those
monitoring devices.
Legend of hydraulic piezometers
M – survey points; I - inclinometers (a) (b)
Fig. 3
Partial representation of the monitoring system:
(a) Initial monitoring system; (b) Piezometers installed in the left bank after
abnormal hydraulic behaviour of the foundation
Deficient hydraulic behavior of the dam foundation occurred short after first
filling of the dam reservoir. In particular, high pore pressures formed in the
natural terrain downstream of the embankment, mainly, at the left bank side.
Artesian pressures installed in the natural terrain, and, thus, to improve the safety
conditions and the monitoring system, an inverted blanket filter and an installation
of additional 66 piezometers at downstream of the dam were conducted,
respectively. Figure3 (b) shows the location of the majority of these piezometers.
In addition, to measure the seepage water collected by the inverted blanket filters
a set of flowmeters was installed. In this paper, we evaluate the usefulness of
neural network models to predict the hydraulic behavior of such complex dam
site. In particular, we correlate the monitoring data from one of the additional
piezometers with the data from the closest flowmeter, and confirm that the
prediction from the neural network is more accurate than that from a simple linear
correlation.
3.3. DATA EVALUATED WITH THE NEURAL NETWORK MODELS
The data used in the model corresponds to a period between March 2003
and August 2018, resulting in more than 120 observations per variable. The time
evolution of the reservoir water level, the uplift pressure in the P30, P31 and P17
piezometers (in profile E, shown in Fig 3(b)), and the seepage in the drain MC4
are presented in Fig. 4. The samples were collected every month. Among the
different loads acting on dams, it is usual to distinguish, as the most important
ones for structures in normal operation, the hydrostatic pressure variation, Fig. 4.
Missing data resulted from lack of inspections.
Fig. 4
Seepage, uplitf pressure and wáter level measured between 2004 and 2016
3.4. TRADITIONAL ANALYSIS BASED ON MULTIPLE LINEAR REGRESSION MODELS
The structural response of, for instance, the displacement in any point of
the dam, is strongly related to the corresponding variation in the water level in the
reservoir. The observations presented in Fig. 3 will be used for the computation
of the models presented in this work.
In this case study, the multiple linear regression (MLR) model with the best
performance for the seepage of the MC4 drain yHST,MC4 was obtained as the sum
of the hydrostatic pressure term β1 × h (where h is the reservoir water level
height), the uplift pressure terms in the upstream zone β 2 × P31 and the uplift
pressure terms in the downstream zone β3 × P17 e to represent the effect of the
head loss variation along the dam and its foundation. The time effect did not
seem to have a significant importance in the period examined by this study. The
regression coefficients of the quantitative models obtained are β1 = 1.09,
β2 = -3.05 and β3 = 2.4, with k=-257.8; being the MLR model represented trough
the following equation:
yHST,MC4 = β1 × h + β2 × P31 + β4 × P17 +k
the residuals were obtained through the difference between the observed
horizontal displacement and the corresponding predicted value obtained through
the MLR model.
4. METHODOLOGY FOR SEEPAGE CONTROL
4.1. INTRODUCTION
Seepage measurement devices are installed at the downstream of the dam
to measure the amounts of seepage through, around, or under dams. Drain
outlets are commonly used as seepage measurement points. The rate of flow is
measured either individually in each drain or for a combination of drains grouped
by specific zones of the foundation and, in this case, it will also include the
infiltrated water through the dam body.
Modelling the hydraulic behaviour of dam foundations is a non-linear
problem and a very complex task, mainly because of the discontinuities of the
foundation mass and the lack of information concerning the hydraulic properties
of foundation mass and the discontinuities characteristics.
As referred before, in order to study the application of the neural network
technique in the seepage control of case study, drain MC4 was selected. This
measurement point is located at downstream of the embankment (in profile E,
Figure 3(b)) and collects the water in the vicinity, from the inverted filter placed
above the natural terrain. The reservoir water level and the pressures measured
in the P31 and P17 were considered as inputs for the ANN model.
4.2. THEORETICAL CONCEPTS - ARTIFICIAL NEURAL NETWORK MODEL
Artificial Neural Networks are computational methods inspired on the
efficiency of the brain process (Patterson, 1996). The increasing interest for that
area derives from the learning ability of these models, which relate the variables
without imposing relationships among them. For the creation of these models, it
is necessary to define the architecture of the neural networks and to go through a
training process with examples (the monitoring data).
Multilayer Perceptron networks have units arranged in layers. The first layer
receives the inputs and the last layer produces the outputs. The middle layers
have no connection with the external world, and are called hidden layers. A unit
is an operator with inputs and outputs, associated with a transfer function,
interconnected by synaptic connections.
Each unit in one layer is connected to every unit on the next layer. The
information is constantly "feed forward" from one layer to the next.
Multilayer Perceptrons learn by an iterative process, adjusting the weights
so as to be able to correctly learn the training data and hence, after the testing
phase, to predict unknown data. Fig. 5 illustrates a generic example of a
Multilayer Perceptron neural network, with an input layer having input
parameters, one hidden layer, l, with Q processing elements and an output layer,
L, with M outputs.
Fig. 5
Multilayer Perceptron neural network designations
The parameters have the following meaning:
xip
- input network i , from pattern p ;
P - number of patterns;
L - output layer;
l - hidden layer;
N - number of inputs in input layer;
Q - number of processing elements in the hidden layer;
M - number of processing elements in the output layer;
wijl
- synoptic weight between input network i from layer l 1 at processing
element j from layer l ;
slj, p
- activation value at processing element j from layer l , from pattern p ;
f jl
- activation function at processing element j from layer l .
yil , p
- output unit i , from layer l , from pattern p ;
First, the neural networks were trained with a small number of processing
elements in the hidden layer and the performance over the training set was
determined. Then, the procedure was repeated with one more processing unit in
the hidden layer and the performance was again calculated. After a sufficient
number of architectures were trained, the best architecture was chosen.
4.3. APPLICATION AND RESULTS
The ANN model used in this case study consists of an input layer with 3
input parameters (the reservoir water level and the uplift pressures measures in
the piezometer P31 and P17), an output layer (to represent the seepage
measured in the MC4) and one hidden layer. Every neuron in the network is fully
connected with each neuron of the next layer. A hyperbolic tangent transfer
function has been chosen to be the activation function for the hidden layer and
the linear function for the output layer. The generalized backpropagation delta
learning rule algorithm was used in the training process. The chosen network
architecture has shown the best results, considering all the tested networks, from
3 until 30 neurons, at the hidden layer. To find the optimum result, 5 initializations
of random weights and a maximum of 5000 iterations were performed on each
ANN architecture.
Figure 7 shows the evolution of the seepage at the drain MC4 (black color),
and the predicted values from the MLR model (red color) and the ANN model
(green color).
Measurements, ANN
prediction and QI
prediction alog time
The good fit shown by the models can be seen by comparing observations
with the values predicted by the models (Figures 6 and 7) and by the analysis of
Table 1, where the values of the maximum residual and the values of the
standard deviation of the residuals, , for the MLR and the ANN models are
presented.
Table 1
Performance parameters for the
MLR and the ANN models
Model Max of the
residual residual
(l/min) (l/min)
MLR 13.9 4.5
ANN 7.0 2.2
Fi
g. 7
Model predictions vs measurements
5. FINAL REMARKS
After identification of the main inputs and outputs for the neural network
models and after establishing the methodologies for the learning and
generalization procedures, the Multilayer Perceptron neural networks may
become a new useful tool to support the safety control of seepage in zoned
dams.
In this study, the achieved performance of neural network models to predict
the flow rate in drain MC4 was very good when compared with traditional multiple
linear regression models. Despite the good results presented for MC4, it is
necessary to carry out further studies including the results obtained from other
individual drains, and it is not dispensable a global evaluation of the dam safety
on the basis of a comprehensive analysis of the various observed parameters.
Finally, it is always wise to carry out the safety evaluation simultaneously with
various models rather than with just one model.
ACKNOWLEDGEMENTS
We would like to thank APA - Portuguese Environment Agency (dam
owner), by given us permission to publish the dam monitoring data results.
REFERENCES
[1] ICOLD 1995. Dam failures statistical analysis. Bulletin 099-1995,
International Commission on Large Dams (ICOLD), Paris.
[2] ICOLD 2003. World register of dams, International Commission on Large
Dams (ICOLD), Paris.
[3] Foster, M., Fell, R., and Spannagle, M. 2000. The statistics of embankment
dam failures and accidents. Canadian Geotechnical Journal, 37: 1000-
1024.
SUMMARY
This paper concerns about the assessment of the hydraulic behavior of a
50 m height zoned dam used for industry water supply, located in Alentejo
region, in Portugal. The original draining system of the embankment is composed
of a sub-vertical filter, located downstream of the central core, and by a drainage
blanket placed over the foundation, in the deeper area of the valley, which, in
turn, discharges to a downstream drainage toe.
In this study, the achieved performance of neural network models to predict
the flow rate in a drain was very good when compared with traditional multiple
linear regression models. Despite the good results presented, it is necessary to
carry out further studies including the results obtained from other individual
drains, and it is not dispensable a global evaluation of the dam safety on the
basis of a comprehensive analysis of the various observed parameters. Finally, it
is always wise to carry out the safety evaluation simultaneously with various
models rather than with just one model.
