You are on page 1of 18

CHAPTER 7

SYSTEMS' ANALYSIS USING MATHEMATICAL


MODELLING

7.1. General Principles

Theoretically, there are two approaches for systems' analysis:

 The analytic method, consisting in dividing a system in its components, which are
afterwards analysed one by one.

 The systemic method, examining complex phenomena and processes as a whole, having
behaviour and properties, which do not belong to the system's components, but to their
interaction.

Systemic approach requires the study of the system's states, its organization and its dynamic. For
this study various models are being used which make it possible to obtain new information
concerning the structure, functionality and behaviour of the analysed system.

Any model constitutes in general a simplified representation of reality, reduced to its essential
characteristics. The main stages of the modelling are the following:

a) In the first stage the analysis of the original system S occurs and the establishment on this basis
of its essential properties.

For example, the following are being established: the system's initial states, the laws according to
which this evolution occurs, the links between the subsystems and their intensity etc. In fact, these
elements form the model M of the system S and the initial states (the initial conditions).

As the model must reflect reality with an acceptable accuracy, the model's formulation is followed
compulsory by its validation; the validation consists in the reproduction of some situations for which
one disposes of theoretical solutions or for which one has measured values of the state variables
or of the output variables.

1
b) In the second stage, through scientific investigations the study of model M takes place; thus,
new information concerning the model M is obtained that cannot be perceived by examining the
original system S;

c) In the third stage, the transfer over the original system of the new information obtained through
modelling takes place, thus resulting an increase of the degree of knowledge concerning the
modelled system.

By synthesizing the information gathered until a certain moment, the modelling leads at the same
time to the improving of the knowledge process. As a consequence, modelling implies an iterative
process of elaboration of new models, which are more and more perfected, and the model
(considered to be a satisfying representation of reality) is the result of some iteration covering each
and every time the cycle reality-model-validation-reality.

The models thus represent instruments for a better and better approximation of reality, to which
they tend to approach in an asymptotic way. The development of computers has led to increased
modelling performances, due to the fact that in a short time the examination of tens of variants is
possible, without affecting in any way the structure of the studied system. Direct experimentation
on the reality of some decision variants, except the fact that it implies high costs and lots of time,
may lead in certain cases to unwanted results, and the return to the situation previous to the
experiment is either dear, either impossible, due to some transformations which may take place.

7.2. Physical, analogical and mathematical models

The model may be defined as a conventional image of a phenomenon or process, built in such a
way as to reflect essential characteristics for the research's purpose. This specific feature results
from the modelling limits themselves; the models cannot reproduce the reality in all its complexity,
for in this case the model would be a substitute of reality itself (which study is, as it has been shown,
difficult, dear, and sometimes even impossible).

The concept of model was used for the first time in 1868 by Italian mathematician Bertrami, who
built an Euclidian model for non-Euclidian geometry. Previously, Descartes had put the bases of
the analytical geometry, conceived as a model for the Euclidian geometry. The new geometry could
be validated by the agreement of the results, which were obtained through the Euclidian, and the
new geometry. Subsequently, analytical geometry developed in a special way, outrunning the
Euclidian geometry's possibilities and creating the premises of the apparition of non-Euclidian
geometries.

The model, as schematic representation of the reality, is largely used in current practice, as well as
in the scientific activity. The concept of system itself was introduced through some elementary
concepts which are in fact models; one reminds thus of the input-output models (Figure 1.1) or of
the input-state-output models (Figure 1.4) for the schematic representation of the systems.

2
Any system may be modelled; moreover, a system's model is a system itself, being characterized
by input, states as well as by output. Once the model is established, it represents the original
system, any succession of the model's states being interpreted as a succession of the system's
states.

For practical study, physical and mathematical models are being used.

Physical models (also called imitative or iconic models) resemble physically to the object they
represent, but they are presented on a reduced scale. In this category we find, for example,
hydraulic or aerodynamic models, the models used in architecture etc. Special categories of the
physical models are the analogical models, which use a domain's properties, to model another
domain's properties; for example, the infiltration through earth dams was studied using electric
fields properties.

The abstract models (mathematical or symbolic models) represent unitary systems of variables and
mathematical relations destined to the analysis of a part of the reality, serving to the discovery of
new ways of organization and behaviour, which cannot be perceived by other means.

Real phenomena depend of a great number of variables, between which there are certain relations;
still, not all of these variables have the same importance. The mathematical model represents a
simplified mathematical description of a phenomenon or process, detaching from the large number
of variables, those that intervene essentially; thus a satisfying approximation of the reality on the
basis of a reduced number of variables and relations among them is realized.

The model's elements are mathematical relations, which may be algebraic, or differential equations
or systems of equations (reflecting the links between the system's components or their behaviour),
as well as inequalities, specifying the limits between which the variables' values may be situated.
The coefficients of these relations, known as models' parameters, depend on the investigated
system structure and generally are estimated during a calibration process.

The model's variables are of two types:

a. decision variables, representing unknown values, which are to be determined during the
mathematical modelling;

b. state variables, which characterize the system's evolution and which depend on the input
and the decision variables.

In current practice, one must find a balance between the wish to create a complex instrument for
investigation, which will describe as exactly as possible the evolution of the studied system and the
imperative that the model should be efficient, meaning to find the solutions in a reasonable time
and providing good results.

3
Practical experience has actually demonstrated that there cannot be only one mathematical model
to contain all the important aspects of the modelled system; the process which is used consists of
creating models specialized in solving some specific problems.

The complexity of the used model itself depends on the accuracy and volume of the input data in
the model. Like this, the use of some approximate data for an exact model usually leads to
important errors, the application of a simplified model being preferable, as it will require less primary
information. The precision of the model and the precision of the input data are thus in a relation of
tight interdependency. For example, if for the flood waves routing the geometry and the hydraulic
characteristics of the riverbed are known, the Saint-Vénant model may be used; otherwise,
hydrological models like Muskingum are indicated.

The elaboration of a model in general and of a mathematical model specifically, may be realized
either by starting from scientific knowledge already existent, either from measurements over some
phenomena and processes towards a theoretical approach. In other words, the model represents
an instrument of knowledge, a method of verifying the theory, but also an important stage in the
formulation of a new theory.

7.3. Classification of the mathematical models

In function of the criteria considered the following classification of mathematical models results:

a) From the point of view of the system's evolution, the models are:

1. in stationary regime (steady state in hydraulics);

2. in dynamic regime (non-steady state in hydraulics).

Generally, the systems have an evolution in time and they have a dynamic character. However,
there are situations in which one may use stationary models(for which the input variables, the
system's states and the output variables are constant, being not variable in time).

Let's consider for example an aquifer for which the input and output are supposed to be constant
(average multi-annual percolation, respectively multi-annual values of abstracted water through
waterworks). The stationary model allows to determine the hydraulic potential of the aquifer,
supposed invariant in time. Similar considerations may be made referring to the permanent
movement in natural waterbeds, channels or pipes.

The stationary model allows, thus the determination of the balance conditions of the system, which
is supposed to be stable, for constant inputs.

In case the variables depend on the time, the model is dynamic; variable input determine variable
output, as well as the evolution of the system's states. The variables describe trajectories, the links
between variables introducing links between the trajectories.

4
In general, the study of the system's dynamic is realized in the past-future direction, although in
principle the examination from present (or future) towards past is also possible.

b) Considering the mathematic nature of the relations used in the model's formulation, the
following models stand out:

1. linear models;

2. non-linear models.

The model is linear when the variables are at the first power and non-linear in a contrary case (at
least one relation contains a variable at a degree different from one).

In most of the cases (and especially for natural systems), the relation between cause and effect is
non-linear. However, for practical studies linear relations between variables are largely accepted,
due to the advantages involved by using the principle of the effects superposition. This principle is
based on the following two hypotheses:

 The system's output is proportional with the input's intensity (the proportionality property).

 In the case of a complex input, which may be decomposed in more partial inputs, the
system's output is obtained through the superposition of the partial outputs produced by
these partial inputs (the additivity property).

c) According the way of considering the time, the following models stand out :

1. discrete models;

2. continuous models.

It has been proved that the systems have a dynamic character, the input, output and the system's
states being functions of the time t. Obviously, the variable t is continuous belonging to the real
numbers, and thus the real systems are continuous versus time. The use of the continuous models
raises however special difficulties; that is why in practice, discrete models are used frequently, the
variable t being discretized. This approximation of continuous processes through discrete models
allows the use of numeric solving methods.

In hydrology and water management, as well as in the case of other technical systems, the
discretization of the time's axis in intervals, whose value depends on the analysed process, is
frequent. For example, in the case of high floods, this interval takes values from one to several
hours, in function of the river basin area; for the allocation of water resources from a reservoir, the
time step is usually one month. For regional aquifers, the discretization step may even reach 2-3
months.

d) According to the degree of knowledge of the analysed systems, the following models are
used:

5
1. physically based models (white-box models);

2. engineering models of various types:

i. input-output models (black-box);

ii. input-states-output elementary models;

iii. conceptual models (grey-box).


The physically based models describe rigorously a process or phenomenon through some
differential equations or equations with partial derivates; the solutions are obtained by analytical
or numeric integration. The physical processes of the runoff formation on slopes or the processes
of water's infiltration into the soil may be studied with such models. In fact, these models are
input-states-output physically based models. They reflect quite exactly the analysed phenomenon
or process, being characterized by complexity and implicitly by difficult solving; for the calibration
of the parameters of these models, experimental measurements are necessary.

Engineering models have as a purpose to obtain the output O(t) in function of the input I(t), without
a detailed study of the system's inner processes.

The function that transforms input into output within the framework of the black-box models type is
called kernel function or weighting function. The relations being used have different mathematical
expressions, as the variables may be discrete or continuous.

For continuous variables, the output O(t) is calculated using the convolution integral, called also
the Duhamel integral:

(7.1)

where U(t) represents the weighting function.

This integral has the following symmetry property:

(7.2)

meaning that:

(7.3)

In the case of discrete variables, the convolution integral becomes:

(7.4)

6
where Ui are the discretized values of the weighting function.

The most widespread weighting function in hydrology is the unit hydrograph, used in the modelling
of hydrological processes.

If within the input and the output functions are replaced with their Laplace transformations, the
notion of transfer function is used. In this case, the relation between input and output is written as
it follows:

(7.5)

where:

ai parameters attached to the output values;

s a complex variable;

O*(s) the output value;

bi parameters characterizing the model's input;

I*(s) the input value.

The ratio between the Laplace transformations of the system's output and input:

(7.6)
=

represents, by definition, the transfer function of the system and it allows the transformation of the
input I*(s) into the output O*(s):

= (7.7)

We remind here that the Laplace transformation of a function f(t) is the function Φ(s), defined by
the integral:

(7.8)

The denomination of transfer function may be used even in the cases when the input and output
variables are expressed as real numbers; even the runoff coefficient, which is the ratio between
the effective precipitation and the total precipitation represents a transfer function. Any operator,
which if applied to the input transforms it into output, may be considered as a transfer function.

7
As input-states-output elementary models in hydrology the most common are those models for
which the runoff coefficient depends on the evolution of the soil's humidity, which characterizes the
system's state.

Conceptual models consist in the detailing of the processes that take place within the system, after
its decomposition into elements. The evolution from input-output or input-states-output elementary
models to conceptual models is actually the equivalent of the opening (or the lighting) of the black-
box. The system's components are linked in series or in parallel; each component transforms its
input in specific output. Conceptual models move the use of kernel functions or transfer functions
from system level to subsystems level, which allows them to be considered a generalization of the
previous models.

The name of conceptual models refers to the way in which the specific transformation mechanisms
of every component, as well as the links between components, are imagined or conceived.

Typical examples of conceptual models in hydrology are the reservoir type models. The water
coming from precipitations goes successively through a series of interconnected reservoirs (vegetal
layer, snow reservoir, soil layer, aquifer, hydrographical network), each reservoir being emptied
according to its own laws.

e) Concerning the way to consider the parameters' variability, the following stand out:

1. models with global (or lumped) parameters;

2. models with distributed parameters.

In the case of the models with lumped parameters, the values that characterize the system's
structure, the value of the input and output fluxes as well as the state values are constant for the
whole system, while for the models with distributed parameters, the same values are variable in
space.

The models with global parameters are used when:

 the system's degree of homogeneousness is high;

 the system's structure is relatively unknown, and the system's internal states either are not
accessible for measurements, either do not present practical interest.

Such models may be used for example for the discharge evaluation in a certain section on the river.
The input elements (precipitations, temperatures) or the system's states (the humidity of the soil)
are averaged on the river basin surface, its heterogeneous structure being ignored.

The simplifying hypotheses, which stand at the basis of the models with global parameters, lead to
a more reduced degree of accuracy, but the results obtained are in many situations satisfactory.

8
In case of high heterogeneity of the system, the parameters of the corresponding models may not
be considered constant, but variable in space, resulting models with distributed parameters; the
system is discretized spatially (and not functionally) in quasi-homogeneous areas, each area being
characterized by its own values of the parameters. There must be mentioned however that within
an area the parameters have a global character, just like the input data or the state variables that
have values averaged on each one of these areas.

The first models with distributed parameters in hydrologic sciences have been used for modelling
groundwater resources, the aquifers being characterized by a hydraulic conductivity variable in
space. For mathematical modelling, the study domain is discretized; each element of the mesh can
have at the limit a specific value of hydraulic conductivity, recharge from precipitation or of the
storage coefficient.

Subsequently, the models with distributed parameters have begun to be used for surface runoff
modelling as well. The river basins are divided in quasi-homogeneous areas (cultivated areas,
surfaces covered with forests, impermeable zones etc), each area having its own transformation
structure and its own parameters.

Due to the large number of parameters of the distributed models, the parameters have generally a
physical meaning and may be obtained insitu. The models with distributed parameters lead to a
greater precision of modelling, but they present difficulties from the point of view of parameters
evaluation.

f) Considering the system's knowledge degree, mathematical models are classified in:

1. deterministic models;

2. stochastic (or probabilistic) models.

The model is deterministic when a given input always produces the same output. The relations
between input and output may be described using physical laws, known at least at a level
corresponding to the modelling purpose; the variables of these models take deterministic values.

Some input elements have a reduced or insignificant influence over the output; only the sensitive
relations between input and output must be evaluated. From a practical point of view, this means
that only a relatively reduced number of input elements are considered as proper input, being linked
to the output through strong causal dependencies; the remaining input elements are neglected or
considered perturbations (noises), which produce deviations from the system's rigorously
deterministic behaviour.

If the noises are important, one must either enlarge the investigation (one or more significant input
have been neglected), either introduces a random (or stochastic) component.

The model becomes stochastic when probabilistic laws are being used and stochastic elements,
with known or determinable distribution, occur in the model.

9
Stochastic models may be divided into:

 Models for frequency analysis (statistic repartitions).

 Regression models.

 Stochastic models.

 Models with random coefficients.

 Models with constraints expressed in probability (chance constrained models).

The frequency analysis models are generally used in hydrology to evaluate the values
characterized by a given exceedance probability (or by the corresponding return period). For
dimensioning the spillways of a dam, exceptional high floods with rare frequency are being used.
To obtain the maximum discharge with a given exceeding probability, a statistic processing of
maximum yearly discharges is necessary, extrapolating the empiric repartition through theoretical
repartitions.

Regression models are used for checking the dependency or independency of two or more statistic
variables. If the variables are independent, they may be analysed separately as one-dimensional
repartitions. If the variables are dependent, it is important to evaluate the influence of a variable (or
of a group of variables) over the explained variable. This statistic processing of a special practical
importance is known as correlation or regression analyses.

Regression curves between two variables have the significance of some conditioned average
values. The intensity of the statistical dependence between variables is expressed by the
correlation coefficient in the case of a linear correlation and by the correlation ratio for a non-linear
correlation.

The correlations have a special importance in hydrology; one may give as examples:

 the rating curve (the H-Q correlation);

 the correlation between the evaporation coefficient and the altitude;

 the correlation between a high flood's time of increase and the aggregated variable
, where is the river's slope;

 the correlation between a high flood's total duration and one of the following variables

(simple or aggregated): L , or (where L is the river's length, Ib is the river


basin's slope, F is the basin surface, etc.).

The values of the explained variable present deviations compared to the average values
represented by the correlation curve; the more the respective values are closer to the curve, the
more the dependency between the variables. At the limit, if all the values are situated on the curve,
the link between the variables is deterministic.

10
A stochastic process represents an infinite row of statistic variables; a finite sample of this row is
called a time series and constitutes in fact a multidimensional statistic variable.

If all the statistic variables which form the series have the same distribution, the respective time
series constitutes a sample of a stationary stochastic process; if the components of the time series
have the same distribution law, but with different parameters (average value, dispersion), the
respective stochastic process is non-stationary.

Observations concerning a stochastic process may underline a general evolution tendency,


representing the series' deterministic component (also called systematic component or tendency)
to which a statistic component, due to some factors with random influence, is added.

The determinist component is generally formed by a polynomial tendency, slowly variable in time,
over which come seasonal components, which manifest themselves periodically; this period is
usually the day, month, season, year, but could also be groups of years or centuries.

The statistic component of the process is analysed after subtracting the determinist component
from the initial series; the residuals (the differences between the initial series and the determinist
component) are interpreted like a time series, extracted from a stationary stochastic process.

Stochastic processes generally and time series especially are largely used in hydrology. Thus, the
discharges may be interpreted as a Markov process; the artificial generation of hydrological values
has actually been used for a long time in practice (M. Fiering, 1967) based on the Markov model.

The hydrological data sequences registered in the past do not offer all possible cases for
dimensioning or establishing the operation rules for the water management works. The extension
of available data by artificial generation using Markov models or time series (respecting the basic
characteristics of the initial data: average value, coefficient of variation and asymmetry) is largely
practiced. These techniques do not lead to new information concerning the river's hydrology; they
only allow obtaining different scenarios of the discharges, keeping the initial information or its
greatest part.

The models with random coefficients are used when certain coefficients of the mathematical
models do not have a unique value, but take a range of values with different probabilities. Thus,
the values obtained from measurements are subject to errors; on the other hand, by their own
nature (discharges, costs, etc.) some variables have a stochastic character. By taking it into
account, the model becomes more realistic, but also more difficult to solve.

The models with constraints expressed in probability impose themselves when certain constraints
cannot be always satisfied. Such situations are frequently encountered in the waters management
field. Some objectives (water supply for users, flood control, protection of water's quality) may not
be always realized with certitude, but with a certain probability. From this point of view, a
deterministic model can be seen as a probabilistic model, whose relations are satisfied with a
probability of 100 %.

11
g) According to the number of the input and output components, the models may be
classified in:

1. monovariable;

2. multivariable.

One input and one output characterize monovariable models, while multivariable models have more
input and one or more output.

An example of monovariable model in hydrology is the use of average global precipitation at the
scale of a watershed, representing a unique input value in the hydrological models. If one
associates to each station a Thiessen polygon, then one obtains a multivariable model, with more
input values. Similar considerations may be made concerning the zoning of the aquifer's recharge.

h) From the parameters number point of view, the following models stand out:

1. non-parametrical models;

2. parametrical models.

Tables as well as curves or sets of curves, represented graphically or whose ordinates are defined
numerically belong to the category of non-parametrical models. One can give as examples of non-
parametrical models: the operation rules of a reservoir, the SSARR curves, rules for water
allocation etc. The weighted function defined by discrete values (for example the unit hydrograph)
is another example of non-parametrical model.

The input-output models for which the transformation of the input in output values is expressed
under analytical form as well as the input-states-output models, written as state equations, are
examples of parametrical models; the studied processes are described by differential equations, in
the case of the models with global parameters, or by equations with partial derivates, in the case
of the models with distributed parameters. The equations' coefficients represent the parameters,
which must be estimated; the more robust the model, the more a smaller number of parameters
must be calibrated (the principle of the parameters' parsimony). The transfer function, analytically
defined, is another example of parametrical model.

The weighting function defined by discrete values is a non-parametrical model. It may be


transformed in a relatively simple way in a parametrical model, looking for analytically defined
curves, which fit well the weighting function's ordinates.

If the analytical expression of the weighting function is known, its parameters may be determined
directly through optimisation.

i) According to the purpose of the modelling, one distinguishes models for:

1. system's simulation;

12
2. system's optimisation.

Simulation models have as a final purpose to predict the evolution of a system under the action of
some disturbing factors or the anticipation of some decisions' consequences. The concept of
simulation is currently used in the study of the dynamics of complex systems; simulation may be
imitative, analogical or mathematical (numeric).

Numeric simulation represents a technique for realizing experiments using mathematical models
and corresponding computer codes to describe the behaviour of a real system (or of some of its
components) during a certain period of time. Simulation models are especially flexible, allowing
one to consider a great number of variables and some very complex links between the system's
components; obviously, this detailing of the examined system leads to an important amount of
computations. Mathematical models in non-steady state for floods' routing or the ones for simulating
the groundwater movement may be mentioned here.

The simulation's main objectives are (Shanon, 1975):

 to describe the systems' behaviour;

 to propose hypotheses and theories for explaining the observed evolution;

 using these hypotheses and theories for predictions concerning the future evolution of the
system.

Optimisation models are used for improving the system's behaviour finding better operation rules
as well as for the best evaluation of the parameters occurring in the structure of mathematical
models.

From a mathematic point of view, the optimisation process consists in determining some extremes,
the system's variables and states satisfying some constraints of physical or functional order; as a
result, the classic mathematical analysis is replaced by techniques of mathematical programming
(linear, non-linear, dynamic programming, programming in integer numbers, Pontriaghin's
maximum principle etc.).

An optimisation model contains, aside from algebraic or differential equations reflecting the
dynamic of the system or the links between variables, constraints regarding their variation domain,
as well as an objective function (criteria function, purpose function, optimising function,
performance function, performance index etc.).

7.4. Identifying (building) a mathematical model

We remind that a mathematical model represents a set of relations between the analysed system's
physical variables, expressed as algebraic equations, differential equations or equations/systems
of partial derivates equations.

13
Establishing a system's mathematical model is known as the system's identification. The
identification may be realized:

 Theoretically.

 Experimentally.

 In a mixed way, theoretically and experimentally.

Theoretical identification has as a starting point that the physical laws (known a priori)
characterizing the analysed system, as well as a series of simplifying hypotheses (such as the
linearity of certain phenomena), are available.

Among the physical laws largely used in hydrology and water management, we remind here:

 Continuity equation, applied in steady or unsteady regime, to surface or groundwater, in


problems of water resources or water management.

 Energetic balance, in which the solar radiation, the radiation necessary for evaporation, as
well as the energy's variation from the considered mass intervene.

 Heat balance (from soil, air or water).

As a result of the theoretical analysis, one obtains in the most general case a system of differential
equations valid for the whole studied domain in the case of global parameters, respectively a
system of equations with partial derivates which characterize every infinitesimal element (or
discretization element) in the case of the systems with distributed parameters.

Theoretical modelling is used when one knows well the laws, which characterize the system's
dynamic. This type of modelling is also applied in the phase of a system's design (which does not
exist yet, but the equations which describe its dynamic are known); it is the case of many water
management systems, whose characteristics (installed powers, storage lakes capacity, discharges
that can be delivered, etc.) are determined during modelling.

In some cases, theoretical approaches lead to models of a great complexity, difficult to solve.
Generally, to make theoretical models usable one must simplify them, resorting to:

 Linearization of non-linear equations with partial derivates.

 Approximation through ordinary differential equations of the equations with partial


derivates.

 Reducing the order of ordinary differential equations.

To analytically obtain the solutions of the theoretical models is not possible in many cases and for
this one resorts to numeric solving. For the same theoretical model one may use different numerical
models: finite differences, finite elements, frontier elements etc. In a similar way, for the integration
of a differential / partial derivates equation one may use an explicit method (for which the unknown

14
- usually the system's output, is determined only in function of the states variable at the beginning
of each step of computation) or an implicit method (for which the output is evaluated in function of
the initial and final states, which requires an iterative process in the frame of each step).

The attenuation of a flood wave in a reservoir, the integration of the Saint-Vénant equations or the
integration of the groundwater movement equations in unsteady state may be given as examples.
An explicit numerical model requires a very small computation step to insure precision, but
mathematical solving of the obtained equations is commonplace. On the contrary, in the case of
an implicit numerical model, the number of computation steps is clearly inferior, because the
computation step is bigger than in the first case; still, the computation effort is important due to the
difficulties of numerical solving: iterative solving, solving of some systems of equations etc.

Theoretical models cannot be developed without the contribution of experiments and


measurements insitu, for checking the adequacy of the models, as well as to evaluate some of their
parameters.

The experimental identification of the systems has as a purpose to establish the mathematical
model on the basis of the input and output values, obtained by measurements; in general, these
are black-box models. Experimental models do not have as a purpose to establish the physical
laws that characterize the analysed phenomenon or process, but only to find a link between the
input and output data. The hydrological models of flood waves routing, the Horton infiltration model,
the unit hydrograph model, the Thornthwaite evapotranspiration model, etc., may be given as
examples.

In all of these cases, one do not measure the system's states variables, or this is not possible; as
a result, neither does the model describe the system's inner structure. The parameters of these
models are evaluated by optimisation, the method of the least squares being the most frequently
used.

Joint (mixed) identification constitutes a middle way, trying to avoid the difficulties, which
theoretical, as well as experimental modelling, has to face. In practice one starts in general from a
priori obtained knowledge concerning the analysed system, which makes easy the choice of the
model's structure (theoretical identification), for which however the parameters' values must be
determined experimentally.

Physically based models, as well as conceptual models belong to this category. In the case of the
physically based models the measurement of the states variables is possible, and the link between
input and output is based on known physical laws.

Conceptual models represent a generalization of the input-states-output elementary models; the


system is decomposed into sub-systems, from among which at least some have states that may
be characterized through states variables. The sub-systems represent modelled systems, either as
input-output models, either as input-states-output elementary models or physically based models.
Between the sub-systems there are links that define the system's structure and its evolution laws.

15
After the model's elaboration, its validation is necessary; for this, the system's answer is obtained
for simply cases, mostly when an analytical solution exists. If the results obtained through modelling
are close to the real ones, the model may be validated. In an opposite case, re-examination of the
accepted hypothesis at the model's elaboration is necessary.

As it has been shown, partial theoretical knowledge of the system allows one to choose the type of
model to characterize it; after the model's identification and validation, the calibration (evaluation
of the numeric values) of the parameters comes next.

Parameters' identification requires information about the system, usually obtained on the basis of
measurements having been done in the past, or more rarely, by experiments. In hydrology one
generally encounters the first case, when through the meteorological or hydrological network
precipitations, temperatures, water levels, velocities etc., are measured. The experiments
programming is possible especially in hydrogeology (for example the evaluation of hydraulic
conductivity, transmissivity or leakage using pumping tests).

The parameters' estimation is called calibration or identification; when using the model for different
simulations, one must not forget that the models' parameters have a limited validity, characterizing
the field of registered values. The supposition that the model will behave linearly outside of the
measurements' domain is generally false, although it is being used frequently; even the behaviour
of the system could be changed. Let's take, for example, the case of a groundwater work. After
parameters identification, the model may be used for simulation of the consequences of the various
abstracted discharges for different wells locations. If the drawdowns are very important, a river in
the area may now supply the aquifer, although in the initial situation the discharge exchanges could
have been from the aquifer to the river.

In a similar way things evolve in the case of an irrigated area, with great water losses from the
irrigation system. The effect is the rise of the phreatic levels and the apparition of new discharge
areas (natural depressions, either even in a water course whose riverbed is found above an argyle
pack and which in the initial situation was not hydraulically related to the aquifer).

After defining the model's structure and the parameters calibration, the validation of the
parameters follows. Usually, from the number of registered cases, some of the data is used for
calibration, and the rest for validation.

For example, for the forecast of the flood waves, with the calibrated parameters one tries to
reproduce other floods registered in the past. If the deviations between the modelled discharges
and those registered are acceptable, the model may be used for future simulations.

If the model behaviour is not satisfactory, resuming the whole cycle of mathematical modelling may
be necessary; this operation is usually carried out in an opposite direction to the modelling process,
suspecting errors in: the parameters values, the computer code, the numerical model, the
mathematical model, the model constraints, the physical laws taken into account, the chosen
variables for the system's characterization, or even in the basic data, which may be affected by
errors.

16
Finally, a last possibility may be the evolution in time of the system's parameters. For example, the
catchment's deforestation leads to the modification of the runoff characteristics.

Depending on the system's evolution, the parameters' identification is:

 Stationary, when the system's structure does not vary in time; thus the model's parameters
are constant, and the identification's purpose is an estimation as precise as possible of the
parameters.

 Adaptive, when the system's structure and thus its parameters as well vary in time; the
identification's purpose is to estimate the parameters' evolution.

Depending on the lag between the field measurements and the use of the data for calibration, the
identification is:

 At different time (off line).

 In real time (on line).

The first case is most often met in practice; usually, the measured data is stored in databases, from
where it is used when necessary.

Quite recently, identification in real time has begun to be used. The data analysis and parameters
identification takes place in parallel with the evolution of the physical process, so the obtained
results become immediately accessible for the regulation of the process. Real time processing is
used in the case of an adaptive identification, followed by a management of the same type of the
system.

The floods generation, their routing along the rivers network and the attenuation in reservoirs might
be given as an example of real time processing.

Summarizing what has been previously shown, the main steps in the identification of a system's
model are the following:

1. Collecting all available data;

2. Interpretation of the information contained in the data bases, understanding the system's
structure and its schematisation.

3. The model's identification, or the formulation of the mathematical model, consisting in


defining the laws that govern the analysed process or phenomenon. The chosen model
may be: stationary - dynamic, discrete - continuous in time, deterministic - stochastic, with
lumped - distributed parameters, linear - non-linear etc. As a result of this step a
model M(q) is obtained, where q is the vector of the unknown parameters;

4. The choice of the numeric model, which should lead in circumstances of quick convergence
and numeric stability to the solution of the analysed problem; for example, to solve an

17
equation with partial derivates one may use the method of finite differences or the method
of the finite element, the solving scheme may be of explicit or of implicit type, etc.

5. The mathematical model's testing and validation;

6. Programming, if it is possible, experiments insitu ;

7. Parameters' calibration, minimizing the difference between the system's and the model's
output (states). The identification may be realized: a) using optimisation models, which
allow to find parameters in a quite short time; b) by trial and error, modifying the parameters
until one obtains a good agreement between the model's and the system's output (states).
Although trial and error search seems primitive from a mathematic point of view, it is still
preferred by many specialists, because the identification process is controlled and the
resulted values respect the physical meaning of the parameters.

8. Validation of the parameters.

9. Utilisation of the model either for anticipating the consequences of different scenarios
(structure, decision variables, input values) concerning the future evolution of the system,
either for increasing the system's efficiency.

18

You might also like