Professional Documents
Culture Documents
Dam Monitoring International Conference PDF
Dam Monitoring International Conference PDF
Summary
The data collected from monitoring system of dams are usually processed by statistical
methods and very rare by artificial neural networks, although the late are frequently used in
economy and industry.
The paper analyses the performances of the artificial neural networks applied in the field of
concrete dams were comparable with those of statistical methods (EdF, CONDOR, with air
temperature integration).
The study is carried out on X (bank-bank), Y(upstream-downstream) direction of horizontal
displacements measured at two direct pendulums installed at Strâmtori-Firiza buttress dam
(H=51.50m)
The study pointed out that a artificial neural network adequat selected and trained has
similar performances with the most rafinated statistical methods.
Corresponding author
Prof. Adrian Popovici
Email: popovici@utcb.ro
Mobile: +40740206864
1. Introduction
Surveillance of dams behavior during first reservoir filling and operation is essential to
ensure their safety and to warn on some atypical phenomenon which can become dangerous for
construction safety.
Surveillance consists of two main activities: monitoring activity and visual inspection
carried out by verified personal [1] [2].
Monitoring activity is based on the data collected from monitoring devices (MD) equiping
the dam-foundation unitary system, data which are processed and analysed by deterministic,
statistical or artificial neural network methods.
In deterministic method is necessary to have a numerical model, ussual in finite elements,
of the dam-foundation system, capable to simulate the system response to environment action.
Statistical and artificial neural networks (ANN) methods are based on processing the long
time rounded data from monitoring devices (MD) in order to point out any abnormal behavior [3]
[4].
In engineering practice from dam’s field, very used are statistical methods (EdF, CONDOR
with air temperature integration etc.). Contrary, neural networks are very scarce applied in dam’s
field although they are frequently applied in other fields (economy, industry etc)
The scope of this paper is to comparatively analyze the performances of ANN versus
statistical methods in modelling the concrete dam’s response to environment actions.
Some practical applications of the above mentioned methods are performed in modelling
the pendulums displacement which are installed at Strâmtori-Firiza buttres dam (H=51.50m). The
application results pointed out that ANN have similar performance with the most rafinated
statistical methods and also have good capacity to signal any deviation from a normal response of
the dam to environment actions.
where
with NRL – normal retention level, H- reservoir depth in the day of the recording
and – dam’s height.
where
with the reference date for model and - recording data
age of structure (structure time aging influence)
f3(T)= a9 eT + a10 e-T
where ; eT modelling acceleration evolution and e -T – damped
evolution
The mathematical equation has the form:
where y is the parameter evaluated statistic and – the approximation error of the model.
CONDOR statistic model is an improved variant of the EdF model differing from it by a
more detailed of time influence. So, the time influence in CONDOR model is considered by the
following function:
where
The final equation of CONDOR model is as follows:
Groups of 10/11 recorded data are successively selected from long time parameter
recordings at monitoring device, constitute equations system which solved, gives the values, the
value of the y, a0 ...a10/a0...a11 equation coefficients. Final values of the coefficients are evaluated
minimising errors based on the algorithm of the smallest squares.
The model with air temperature integration uses directly the time recorded values of the air
temperature. This model is used especially for concrete dam displacements modelling, taking into
accent the different evolution of the air temperature from one to another year.
The hydrostatic pressure influence is modelled according to formula:
f1(h) = a1h + a2h2 + a3h3 + a4h4
where with CT talveg
elevation (or a reference evaluation which don’t influence
dam’s behabior), H- reservoir elevation in the day of displacement measurement, H b – dam
height.
The temperature influence is given by the sum of the series:
where pt is time step of temperature (in days) and nt the number of the time steps.
The hydrostatic pressure and structure time aging have similar expression as in statistical
models previously presented, as follows:
The artificial neural networks are networks of models of neurons connected through some
adjustable synapses. All the models of neural networks are based on the interconnection of some
element of calculation from a dense network of connection. Each unit of the process is capable to
carry out only simple calculation, but the network, as the whole, can have remarkable qualities in
the recognition of forms, the solution of problems for which it doesn’t have an algorithm but
learning from examples or from experience.
The structural model of neural networks is defined by topology of the network and the
scheme of the interconnection of the units.
The development of a neural network in monitoring the concrete dams consists in the
following stages:
- choice of input parameters (calendar date, reservoir level, air temperature etc)
- choice of output parameters (displacements, stresses etc)
- choice of transfer functions which is made function of network topology and number of
ituations for calibration.
In present study, as transfer function was chosen Sigmoid function having formula
and its derivative .
Sigmoid
Functia function
sigmoidala
1 0.3
0.8
0.2
0.6
1 d 1
x dx 1 e x
1 e
0.4
0.1
0.2
0 0
6 4 2 0 2 4 6
x
Fig.1 – Graphic representation of the Sigmoid Function.
- choice of the network topology which initially can be chosen according to the following
empiric relation:
The dam foundation system is equipped with several monitoring devices presented in
Table 2, which allow beside visual inspection to detect from initial phase any abnormal behavior
in the system and to take in due time appropriate intervention, if necessary.
Table 2
To date
Provided
No Type of device in use
(1965)
(2016)
Environment actions
1 Measuring level water staff 1 1
2 Reservoir water termometer 1 1
3 Air termometer 1 1
4 Pluviometer 1 1
Dam foundation system response
5 Teleformeters 140 121
6 Telepresmeters 10 9
7 Teletermeters in concrete 59 50
8 Direct pendulums 2 2
9 1 1
10 6 6
11 Dilatometers 16 16
12 Deformeters 11+6 17
13 Hydrometers pressure cells 11 11
14 Piezometer drillings 13 12
15 Manotriangulation network 1 1
Concerning the data presented in Table 2, it may be remarked the reliability of the devices
equipping Strâmtori-Firiza dam-foundation system.
The frequency of measurements at devices depends of the type of the measurement and the
safety state of the dam and is setted by artificial regulation. The dam behavior is framed in
normal and abnormal behavior, the late having three steps: attention, alert and allarm states.The
frequency of measurements gradually increases versus the safety state of the dam. Moreover, in
the case of abnormal behavior of the dam, there are special procedures provided in regulation that
the operational staff from the dam must follow step by step.
The behavior of the dam along of all operation years, from 1965 till 2017 may be
characterized as a normal one.
Fig.2 Longitudinal section through the Strâmtori-Firiza dam axis and cross sections through
no.1, no. 6, and no. 8 blocks.
Some results concerning EdF method application in processing the displacements measured
at direct pendulum from block no. 6 are shown in figures 3, 4 and 5.
In compliance with graphs from figure 3 may be remarked that computed values are very
closed to measured values. The series of computed displacements has smaller amplitudes than
measured displacements, because the air temperature influence is considered by a reasonal
function.
It may remark (fig 6 and 7) the values computed with CONDOR method are quasi
identical with their correspondents from EdF method, excepting aging influence which is linear
and with very small amplitude.
Fig.7 – Graphs with reservoir level season and age influences modelling Y displacements
from pendulum block no. 6 evaluated by CONDOR method.
Chronological series measured/ computed with air temperature integration method of the Y
pendulum displacements from block no. 6 are very closed one/ another (fig. 9), global coefficient
of correlation being 0,931 higher than those computed by EdF/CONDOR methods (0,931).
Standard deviation for chronological series computed by air temperature method is 0.770 versus
0.860 for measured chronogical series.
Figure 10 points out that reservoir level influence resulted in air temperature method is
higher than in EdF/CONDOR methods, reaching 4,4 mm versus 4,4 mm. The shapes of curves
displacement/ reservoir level are very closed.
The amplitudes of displacements given by air temperature variation are positioned between
6 mm and 10,5 mm (fig. 11), versus maximum 6 mm in methods with seism influence expressed
by trigonometric function.
Several testings were made concerning neural network scheme in order to find a scheme of
maximum rate. So, successively were tested different transfer functions being chosen. Sigmoidal
function network tipology with different hidden layer, number of neurons per layer, type of
interconnection (only forward), optimum density of interconnection.
Fig.13 - Structure of the artificial neural network used to predict pendulums displacements
Figure 13 illustrates the final scheme of the artificial neural network used to predict
pendulums displacements installed in blocks no. 6 and no. 1 from Strâmtori-Firiza dam. It
consists of two hidden layers, five input parameters (calendar date, hydrostatic level, daily
temperature, mobile average temperature – 3 days, mobile average temperature – 7 days) and two
output parameters (pendulum displacements on X and Y directions).
Fig.14 - Errors evolution during neuronal network block no. 6 training.
Network training settles neurons weight in system. This operation can lead to wrong
solution and in order to avoid it, it is recommended to display permanently the evolution of errors
or of global correlation ratio. The evolution of errors during neural network training of block no.
6 shows that after 1 106 cycles the errors reaches 2%.
Correlation measured/computed for neural network block no. 6 (fig. 15) is made with
normalized variables as a unique series, althoug there are two output nodes.
Fig.16 - Statistical data in neural network and weight of the main
input variable (T, h, t) – block no. 6.
In figure 16 are illustrated the final results of the neural network training block no. 6: node
1- bank- bank X direction and node 2- upstream- downstream Y direction, but the displacements
of pendulum on bank-bank direction are very small, reaching only 0,5 mm. Concerning the
weights of input parameters it may remark on X direction the maximum weight corresponds to
the hydrostatic pressure (47,64%) and on Y direction of temperature (77,41%)
Figura 18 - Contour lines of Y block no. 6 displacements function of reservoir level and time.
In figure 18 are presented contour lines of the displacements block no. 6, Y direction
function of time and reservoir elevation. It may remark the influence of hydrostatic pressure as
weight in displacements has increased in time.
Fig.19 - Contour lines of Y block no. 6 displacements function of daily average temperature and
reservoir level.
Contour lines of the displacements block no. 6, Y direction function of daily average
temperature and reservoir elevation show the temperature influence is higher when the reservoir
level is higher. This result is explained because of temperature acting on dam downstream face
while upstream face is covered by reservoir water with lower temperatures than air.
Fig.21 - Comparative analysis between reservoir elevation and block no. 6, Y direction,
displacements computed by neural network and statistical method
Fig. 21 illustrates a comparative analysis between neural network and statistical method
concerning displacements function reservoir eleavation block no. 6, Y direction. The shape of
curves is similar but the corresponding values are significantly different. It may remark that an
analyse of the dam-foundation system by finite element method leds to results closer to neural
network.
Data presented previously emphaseises the performances of statistical of statistical methods
and neural network in modelling the displacement of concrete dams have the same degree of
accuracy. Consequently is very justified the neural networks to be commonly applied in this field.
5. Concluding remarks
A summary of the results obtained in case studies carried out in the frame of this paper is
presented in Table 3.
Table 3
X direction Y direction
The resulting conclusion of the table available for concrete dams field is that an
approproate neural network perform equal to the most refinated statistical methods.
Typology of a neural network should be chosen taking into account the number of input
parameters with influence in response, number of hidden layers, density of the connections
between layers, number of neurons per layer etc. to build a high-performance network.
References
1. Popovici A. - Dams for Water Storage Vol. 2 (in Romanian). Editura Tehnicǎ, Bucureşti
2002.
2. Ilinca C.- Aplicarea reţelelor neuronale şi a modelelor statistice în supravegherea
construcţiilor hidrotehnice (in Romanian). Teza de doctorat, U.T.C.B. , Bucureşti, 2003.
3. Ilinca C., Popovici A. – Some aspects concerning use of statistical models in hydraulic
structures surveyllance.
4. Mateescu O, Ilinca C – Prediction of seepage and piezometry at Motru dam using neural
network. Proceedings Seventh Benchmark Workshop in Numerical Analysis of Dams
ICOLD- Committee on Computation and Design of Dams, Bucharest, 2003.
5. Ayvaz M.T. (Turkey), Popovici A. (Romania) – Predicting level variation ar the
downstream of Motru dam (Romania) using hybrid HS-Solver optimization algorithm.
BALWOIS-2012. International Conference in Water, Climate and Environment. Ohrid ,
Macedonia, June, 2012.
6. Ayvaz M.T. (Turkey), Popovici A. (Romania) – Predicting seepage flow through an
earthfill dam-foundation system using an artificial neural network based solution
approach. Application to Motru dam site (Romania). Proceedings 6 th International
Perspective on Water Resources; the Environment Conference, Izmir, Turkey, January,
2013.
7. Popovici A., Ilinca C.(Romania), Ayvaz T. (Turkey) – The performance of the neural
networks to model some response parameters of a buttress dam to environment actions.
9th ICOLD European Club Symposium, Venice, April, 2013
8. Bobocu D., Ilinca C., Popovici A.- Aspects concerning surveyllance activity of the Gura
Râului Dam- Romania.Third Congres on Large Dams organised by Macedonian
Committee on Large Dams, Struga, June 2013.
9. Bohuş C – Studii comparative asupra acurateţei unor modele de prelucrare a datelor din
monitorizarea barajelor de beton (in Romanian). Tezǎ de doctorat, U.T.C.B. , Bucureşti,
2017.
Legend of figures