You are on page 1of 19

EARTHQUAKE ENGINEERING AND STRUCTURAL DYNAMICS

Earthquake Engng Struct. Dyn. 2012; 41:61–79


Published online 4 May 2011 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/eqe.1118

Seismic risk assessment of interdependent critical infrastructure


systems: The case of European gas and electricity networks

Karmen Poljanšek∗, † , Flavio Bono and Eugenio Gutiérrez


European Laboratory for Structural Assessment, Institute for the Protection and Security of the Citizen,
Joint Research Centre of the European Commission, T.P. 480, I-21020 Ispra (VA), Italy

SUMMARY
We study the seismic vulnerability of the interdependent European gas and electricity transmission
networks from a topological point of view, whereby the electricity network depends on the gas network
through gas-fired power plants. First, we assessed the seismic response for each independent network;
then we analyzed the increased vulnerability due to their interdependency. We implemented a proba-
bilistic reliability model that encompasses the spatial distribution of both network structures and their
seismic hazard exposure using a Geographic Information System. We characterized the network interde-
pendency using the strength of coupling of the interconnections, together with the seismic response of the
independent—gas—network. We calculated the network fragility curves of the independent and dependent
networks in terms of various performance measures (connectivity loss, power loss, and impact on the
population) and found that the gas network is more seismically vulnerable than the electricity network.
The interdependency introduces an extra vulnerability to the electricity network response that decreases
with the extensiveness of the networks’ damage states. Damage was also evaluated at a local level in
order to identify the most vulnerable parts of the network, where it was found that the potential for the
highest power loss is located in southeast Europe. Copyright 䉷 2011 John Wiley & Sons, Ltd.

Received 21 April 2010; Revised 7 February 2011; Accepted 10 February 2011

KEY WORDS: earthquake hazard; critical interdependent infrastructure; network analysis; network
fragility curve; interoperability matrix; geographical spread of damage

1. INTRODUCTION

The issue of vulnerability of critical infrastructures has attracted considerable attention from both
the policy-making and academic spheres. Whereas policy making are concerned with how the lack
of performance of said infrastructure affects social wellbeing, researchers consider understanding
of the generic characteristics of real-world networks as one of the great challenges in, so-called,
complex system analysis. One of the most important mathematical methods used in complex system
analysis is the presentation of critical infrastructures as graphs. Graph theory has developed a broad
range of interesting results to study the vulnerability of certain classes of graphs: important steps
were taken by analyzing how the topology of different graph models (random [1], Small-world [2]
and Scale-free graphs [3]) changes due to certain kinds of ‘attacks’ [4–6]. By ‘attack’ we mean
the generic elimination of graph elements (i.e. the vertices and links) resulting from an intentional
plan (terrorist attack), a random process (human error) or natural processes (earthquakes, storms,

∗ Correspondence to: Karmen Poljanšek, European Laboratory for Structural Assessment, Institute for the Protection
and Security of the Citizen, Joint Research Centre of the European Commission, T.P. 480, I-21020 Ispra (VA),
Italy.
† E-mail: karmen.poljansek@gmail.com

Copyright 䉷 2011 John Wiley & Sons, Ltd.


62 K. POLJANŠEK, F. BONO AND E. GUTIÉRREZ

ageing). The effects on real-world networks of such attack strategies have been the subject of
recent research [7–10].
If we turn our attention to transportation and lifeline utility systems (water management,
energy, and communication systems) we realize that real infrastructure networks extend over large
geographical areas and are composed of many elements, which, most importantly, are physically
connected engineering structures. Each type of network can be treated as an entity in its own
right; however, this is not the whole picture: real-world networks are interconnected, either by
physical, operational, or social ties. So, in reality, critical infrastructures are not many but rather
one mega-infrastructure that encompasses all our daily activities.
In order to manipulate data on such complex and geographically distributed infrastructures, the
use of Geographic Information Systems (GIS) has become widespread. GIS is a convenient tool
for matching the spatial characteristics of infrastructures to the spatial variability of the hazard
demand on the overall system. Another key item is the possibility to parse this information into a
connectivity matrix that can be analyzed using graph-theoretic methods. Finally, having numerically
processed the network, GIS can again be used for effective visualization of the results.
Network elements (i.e. engineering structures) exposed to natural hazards can suffer physical
damage. Thus, the response of the infrastructure network is strongly dependent on the physical
vulnerabilities of its constituent assets which, in turn, are dependent on their structural character-
istics. A widely used method to assess the vulnerability of those individual assets to a given type
of hazard is the use of structural fragility curves.
One of the sources of a network’s complex behaviour is its connectivity; indeed, the topology
of the connection dictates how damage may spread through the network. In general, three types of
failure propagation can be diagnosed that eventually define how the vulnerability of each element
influences the vulnerability of the network as a whole. First, we may have failure of one element
which is independent of the failure of others but, notwithstanding individual failures, changes the
topology overall. Second, the network can be treated as a dynamical system that carries load flows,
so that changes in the topology or increased system load can trigger load redistributions in such
a manner that overloading of the elements causes additional—cascading—failures (in this case
the failure of one element is dependent on the failure of other element(s) in the network). Third,
regarding interconnected infrastructure networks, the failure of one element is dependent on the
failure of element(s) in the other network and the coupling links between them. In this paper
only the first (element-by-element) and the third (interconnections) types of failure propagation
will be addressed. In order to assess the final vulnerability of the network, different performance
measures, like connectivity loss [11], have been used to describe the network’s response, its
damage states and, eventually, the definition of network fragility curves. In order to evaluate the
network’s performance measure in a probabilistic manner, the fragility curves of the elements are
used in tandem with probabilistic numerical approaches (such as Monte Carlo simulations). This
combination of methods—which is the basis of our method here—was, for example, applied to
the assessment of seismic vulnerability of existing county-sized electricity networks [12–14] and
also seismic vulnerability of the two interdependent county-sized networks (water distribution and
electricity networks) in [15].
Clearly, developing an analysis for an all-enveloping mega-infrastructure is not yet feasible,
but we can take some more steps into understanding how (at least) two types of large continental
infrastructure networks depend on each other, and how their interdependence affects their aggregate
vulnerability. More specifically, what we address in this paper is how an earthquake not only
explicitly generates vulnerabilities in a given network (here the European Electricity transmission
network), but also how the vulnerability of another network on which it is partially dependent
(here the European Gas Transmission network) induces a second, implicit, vulnerability by virtue
of their interconnections. We are therefore primarily concerned with the probabilistic approach of
assessing the topological vulnerability without considering the flow dynamics of the systems.
Our paper starts off by explaining the procedure for assembling GIS information in order to
compile the interconnected graph of the European gas and electricity networks; then we discuss
the theoretical background of network analysis in order to define the performance measures used.

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
ASSESSMENT OF CRITICAL INFRASTRUCTURE SYSTEM 63

We then present the type of data we use, and later we deal with the mathematical formulation of
the probabilistic reliability model, the seismic response of the networks and their interdependency
behaviour. Finally, we present the results of our case study—the European Interconnected Gas
and Electricity Transmission networks—from two perspectives: first the global network measures
of (independent and dependent) network fragility, and then the geographical distribution of their
local fragility.

2. EUROPEAN INTERCONNECTED ENERGY NETWORK

2.1. Description of data sets used in the study


A key component of our work is the assembly of data sets of the European gas and electricity
transmission networks. The source data are gathered from the Platt’s GIS features data sets [16]
of the European gas and electricity infrastructures. The interconnected network we present here is
the result of processing the GIS data into a connectivity matrix.
The high-voltage electricity network consists of transmission lines with voltage greater or equal
to 220 kV, substations and power plants. The gas transmission network consists of gas pipelines
with a diameter greater or equal to 15 in., LNG terminals and compressor stations. Power plants
are divided into two groups by fuel type: the gas-fired power plants and all the other types (hydro,
nuclear, etc.). Here we consider gas-fired power plants as the common elements of both the gas and
electricity networks, whereas all the other types of power plants are connected to the electricity
network only. For the full description of how the raw Platt’s data set is parsed and compiled we
refer to [17].

2.2. Sinks and sources


We classified the vertices of both networks (i.e. the 3231 and 10508 vertices in the European gas and
electricity networks, respectively) into source, sink, and transmission classes (Table I). For the gas
network, the sources are vertices located in the countries of the exploitable gas fields (142 vertices)
and the LNG terminals (21 vertices), whereas sinks are gas-fired power plants. In the electricity
network, on the other hand, all power plants are source vertices; but, in addition to the 998 gas-fired
power plants (which cover approximately 20% of the total electricity generation capacity), there
are 4383 power plants using other fuels. Conversely, sink vertices are those substations that deliver
power into the electricity distribution network (see [17]). Thus, altogether, there are 3708 such
distribution substations.
The compiled European gas and electricity networks are, in the first instance, undirected; but here
we have assumed that the primary directions of flows (electricity power or natural gas) are from
sources to sinks; therefore, we have introduced into the networks’ topology an extra information
field, namely, directedness of those links which are adjacent either to a source (directed out) or
a sink (directed in). By introducing directedness as a key functionality of the network, we force
the shortest-path-finding algorithm (Section 3) to disregard unrealistic solutions. For example, the
shortest path from source to sink that goes through another source is, in our case, not admissible
because power plants do not have a transmission function.
Electricity sink vertices are regarded as the entry points into the electricity distribution network
(i.e. that other network that delivers electricity to consumers). This functionality was implemented

Table I. Division of vertices in the European gas and electricity networks.


Source vertices Transmission vertices Sink vertices
Gas network 163 2070 998
Electricity network 5381 1419 3708

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
64 K. POLJANŠEK, F. BONO AND E. GUTIÉRREZ

for Europe using the, so-called, Thiessen polygon method [18] (by partitioning the area around each
distribution substation into convex polygons). Using the geographical distribution of the population
in [19] we assign the population contained in each polygon to the corresponding distribution
substation.

3. NETWORK ANALYSIS

Graph theory is one of the most immediate mathematical methods to describe and manipulate
network data sets. Here, we deal with mixed graphs since our compiled networks contain directed
(arcs) and undirected (edges) links. A mixed graph G = (V, E, A) is composed of a vertex set V ,
where N denotes the number of vertices, an edge set E (of unordered pairs of vertices), and an
arc set A (of ordered pairs of vertices). The graphical connectivity of the network model defines
its topological structure, and is represented by the N × N connectivity matrix C, where ci j = 1 if
(i → j) ∈ A or ci j = c ji = 1 if (i, j) ∈ E and 0 otherwise. Having such a presentation of a network
model we can solve the problem of finding the shortest path [20] from one vertex to another
(shortest in the sense of the least number of links). The existence of the path between two vertices
verifies the connection between them. Such paths are jeopardized whenever the network topology
changes due to the elimination of vertices or links from the network as a result of some attack.
Next, we discuss how such attacks affect the network’s performance.

3.1. Parameters of a network’s response to element failures: performance measures


We measure a network’s performance by its capability of carrying flow (i.e. the existence of paths)
from the sources through the transmission vertices to the sinks. We considered metrics that describe
the damage to a network in terms of the loss of service due to the disconnections between sources
and sinks. We used various descriptive parameters of loss. First, we counted the lost connec-
tions and then we weighted the importance of the lost connections depending on the additional
attributes that are inherent to the network: for example, for the electricity network we consider
the nominal power of the power plants and the population attached to each distribution substation.
All the performance measures (connectivity loss, power loss, and impact on the population) are
presented as a relative value in the range 0–1, reflecting no damage and complete disconnection,
respectively.
Furthermore, performance measures are the attributes of the whole network and as such can
define the network (as opposed to the structural) damage states. However, due to a dearth of
substantiated technical evidence to justify correlation trends between the extent of the network
damage and network condition, we followed prior examples in the literature [15]; where 20,
50, and 80% performance measure levels have been assigned to define minor, moderate, and
extensive network damage states, respectively. When damage states are presented as the probability
estimation, they can be applied as a surrogate measure for risk indicators; however, they reflect
the response of the network under the chosen hazard.
Connectivity loss: The concept of connectivity loss has already been presented in [11]. It
measures the average reduction in the ability of sinks to receive flow from sources. Although,
in [11], it is assumed that in the initial state each sink is connected to all the sources, this is not
always the case in real networks. Therefore, in order to apply the same concept to fragmented
networks as well, we have to first count the number of the sources connected to the ith sink in the
i
original (undamaged) network Nsource,orig i
, and then in the damaged network Nsource,dam . Finally,
we calculate the modified version of connectivity loss CL as:
 i 
Nsource,dam
CL = 1− i
, (1)
Nsource,orig
i

where   denotes averaging over all sink vertices.

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
ASSESSMENT OF CRITICAL INFRASTRUCTURE SYSTEM 65

Power loss: Upgrading the connectivity loss with the size of the power plant (in MW) to which
sink vertices are still connected to, we can define the power loss PL as the next performance
measure
 
i
Pdam
PL = 1− i
, (2)
Porig
i

where i
Porig and i
Pdam are the sums of the power of all the power plants connected to the ith
distribution substation in the original and damaged networks, respectively.
Impact factor on the population: Attaching to each distribution substation i the assigned popu-
lation POPi we can assess the overall share of the population affected by the power loss. We
assume that the share of the population receiving power from the ith distribution substation equals
its normalized power supply Pdam i /P i
orig in the damaged network. We can eventually define the
impact factor on the population IP as:
 i
Pdam
i=1, ...,ND i
POPi
Porig
I P = 1− . (3)
POPall
Here, ND is the number of distribution substations in the electricity network and POPall is the
population in the Thiessen polygon area supplied by the electricity network.

4. REQUIRED DATA FOR THE PROBABILISTIC EVALUATION OF


THE NETWORK RESPONSE

We must bear in mind that the vulnerability analysis of an engineering structure is concerned with
the response of the structure under different levels of seismic load. Furthermore, if we consider
conducting a risk analysis, we must combine the probability of occurrence of the phenomena
(i.e. the load induced by the earthquake event) and the probabilistic evaluation of the consequences
(i.e. the response of the affected structures). We must therefore collect such statistical data (from
published and available sources) for all the elements of the infrastructure network, whereas their
location and characteristics are already in the GIS database [16].
We use the European seismic hazard map for the probabilistic and spatial characteristics of peak
ground acceleration (PGA), which, at present, is only available for the 475-year return period and
10% of probability of exceedance in 50 years of exposure time [21]. In addition to this, we need
the hazard curve for the location of each element of our network—or at least the seismic hazard
maps for different return periods (different hazard levels). However, because none of these have
been compiled at a European continental scale, we multiplied the values of the PGA from the
existent seismic hazard map by a general factor, called the PGA factor (PGA factors lower than 1
correspond to return periods shorter than 475 years—lower hazard levels—and PGA factors more
than 1 correspond to return periods longer than 475 years—higher hazard levels). As a rule of thumb
the PGA factors 1.25 and 2.5 correspond to 1000-year and 10 000-year return period, respectively.
We must point out that these estimations are based on seismic hazard maps for different hazard
levels found for Italy [22] and Slovenia [23]. Thus, we cannot prove that the rule described is valid
for the whole of Europe; even so, we get a feel for the general PGA factor in terms of the return
periods.
However, the application of the probabilistic seismic hazard map introduces inconsistences [24]
because it is improbable that the seismic intensities at geographically different sites occur simul-
taneously. Thus, the probabilistic seismic hazard map is applicable when assessing the risk to one
network element, while in the present case of spatial structures, it would be much more correct
to apply the spatial distributions of seismic intensity. However, the latter vary with the seismic
source and one would need an extensive earthquake catalogue to consider different earthquake
scenarios [25]; unfortunately, such a catalogue has not yet been made for Europe.

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
66 K. POLJANŠEK, F. BONO AND E. GUTIÉRREZ

In addition, we were obliged to disregard the influence of the local soil type on the seismic risk
because neither the soil type associated with the facilities nor the geological map as a GIS data
set for the whole of Europe was generally known.
Next, for the seismic vulnerabilities (expressed as a conditional probability of exceeding a certain
damage state at a given PGA) of the elements of the infrastructure network, we use the appropriate
fragility curves. Our source for the fragility curves of important elements of the utility systems is
HAZUS-MH [26], a standardized methodology for multihazard potential loss estimation that was
proposed for application in the United States.
Fragility curves for the facilities modeled as vertices (electric substation, power plants, and
compressor station) are dependent on the PGA ground parameter. They are classified according to
the functional capacity (except in the case of compressor stations, whose vulnerability is similar to
that of pumping plants in oil systems), seismic design considerations, and damage states. Electric
substations are classified according to their voltage rating (low voltage <150 kV, medium voltage
150–350 kV, and high voltage >350 kV). Power plants are classified according to the generation
capacity in MW under normal operations (small power plants 200 MW and medium/large power
plants >200 MW). We have chosen the fragility curves assuming that the structures’ design is in
accordance with the seismic provisions, and we have also assumed that failure in our analysis
corresponds to an extensive damage state; i.e. the range of damage that is characterized by loss of
basic functions and exceeds the short-term recovery time.
Conversely, the link elements (which in our case are the gas pipelines and electricity transmission
lines whose vulnerabilities in the electricity network were not considered) are already spatially
distributed structures. According to empirical studies in [27] the damage of such structures is better
correlated with peak ground velocity (PGV) due to seismic wave propagation, and permanent
ground displacement (PGD) due to ground failure. Unfortunately, seismic hazard maps for PGD
and PGV values are not available at the European level, so we used a conversion from PGA to
PGV suggested by Wald et al. [28]. In this way, we could consider damage due to seismic wave
propagation, whereas we had to factor out damage caused by ground failure. Fragility curves for
pipelines are based on the empirical functions of repair rates (repairs/km) and the assumption that
ruptures follow a Poisson process [26]. We adopted the repair rate function from [29] for a buried
and brittle pipeline (although most gas pipelines are typically made of ductile steel, we classify
all the pipelines as brittle because the database has no information on the type of joining between
the pipe segments). We defined the failure (i.e. a damage state corresponding to a non-functional
situation) of the pipeline segment with the occurrence of at least one rupture (breaks or leaks).
A similar approach was adopted in [15].
The HAZUS data sets are, in most cases, more adapted to the design criteria and hazard exposure
in the U.S.A. Therefore, it is evident that we are bound to encounter difficulties when applying the
same data sets to Europe. In particular, the major concerns are the unavailability of seismic hazard
maps for different hazard levels and the seismic fragility curves applicable to European facilities
and design standards. Nevertheless, the feasibility of performing the analysis at the European
level might, at least, motivate the compilation and collation of the missing data sets that are
geographically specific to Europe.

5. PROBABILISTIC RELIABILITY MODEL

The probabilistic reliability model we propose can deal with spatially distributed entities that
are identified by topological, geographical, and sometimes even geometrical features. Here, these
entities are, on the one hand, infrastructure networks and, on the other, the natural disaster. The
reliability of spatially distributed systems is jeopardized into two ways: one concerns how a
localized failure in the network propagates spatially through the network, and another is the fact
that an earthquake (being a spatially distributed phenomenon) affects the network over a wide area;
all of which is therefore likely to cause the simultaneous failures of many elements. Our main
interest is how the seismic damage of each of the vulnerable network’s elements affects the overall

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
ASSESSMENT OF CRITICAL INFRASTRUCTURE SYSTEM 67

Figure 1. Propagation of probabilities of elements’ failure through the analysis.

network performance. In order to apply the probabilistic reliability model as a risk analysis tool,
the Monte Carlo simulations serve to examine how failure probabilities of the singular elements
are propagated through the network analysis (Figure 1), and to assess the probabilities of different
outcomes (network damage states) for the whole infrastructure network.
Every element of a network is exposed to differing seismic demands (depending on its geograph-
ical location on the seismic hazard map) and differing seismic responses (depending on its dynam-
ical characteristics as represented by its fragility curves). Once we have determined which elements
have failed, we can generate a damaged network. It is presumed that only the topological deteri-
oration of the overall connectivity, due to the failure of any of the element(s), compromises the
functionality of the entire network. But, were it possible to construct a flow model, one could
expand the analysis to detect also the failures due to dynamic processes in the flow regime. Thus,
in our study, the extent to which the damaged network is capable of operating functionally is, in
principle, assessed solely by the network connectivity analysis.
We can say that our probabilistic reliability model consists of a seismic-performance network
analysis of damaged networks generated using Monte Carlo simulations. The basic steps of the
model’s algorithm are:
(1) Choose the seismic hazard level.
(2) Identify the location of all the vertices of the network and assign the PGA value to this
location from the seismic hazard map.
(3) Find the probability of exceeding the limit damage state for all the vulnerable elements of
the network from the relevant fragility curves.
(4) Execute the random number generator with the uniform probability distribution between
0 and 1 for all the elements. An element is damaged when the random number drawn is
smaller than the probability attained in step 2.
(5) Define the damaged network: the elimination of each considered element is dependent
merely on its location (PGA), dynamical characteristics that reflect the seismic response
(fragility curves) and a random number.
(6) Perform the network analysis to obtain the connectivity characteristics of the damaged
network.
(7) Determine the performance measures (connectivity loss, power loss, and impact on the
population).
Steps 4–7 perform one trial in the Monte Carlo simulation; repeating these steps we are executing
more trials. One series of trials gives us one set of performance measures that corresponds to one
hazard level. Each set of performance measures is then presented as the complementary cumulative
distribution function, where the probability of exceeding a given value of the performance measure
is calculated as the ratio of the number of trials where the performance measure exceeds this value
according to the number of all the trials.

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
68 K. POLJANŠEK, F. BONO AND E. GUTIÉRREZ

In order to construct the network fragility curves, one series of trials must be performed at several
PGA factors (reflecting different hazard levels). Only then does the response data satisfactorily
capture the phenomenon behaviour. After all the series of trials are completed, we introduce the
notion of network damage states (Section 3.1). The probabilities of exceedance of a certain damage
state at each hazard level (characterized by the maximum PGA in the network) become the input for
the calculation of network fragility curves, i.e. fitting the probabilities to a cumulative lognormal
distribution function. So, one network fragility curve describes one network damage state, whereas
the Monte Carlo simulations executed at one hazard level contributes only one point (probability)
to the fragility curve of each damage state.

5.1. Results of independent networks


In order to compare the independent network fragility curves for the gas and electricity networks,
they are presented in terms of connectivity loss. We consider nine hazard levels, applying the
method with the PGA factors (0.2, 0.4, 0.6, 0.8, 1, 1.25, 1.5, 2, 2.5 and 0.4, 0.6, 0.8, 1, 1.25, 1.5, 2,
2.5, 3 for the gas and electricity networks, respectively), and extract three levels of damage state.
In the case of the gas network, 10 000 Monte Carlo simulations were executed, whereas in the case
of the electricity network we performed 1000 simulations in one series of trials. The analysis of
the electricity network is computationally much more demanding because the electricity network
has not only five times more vertices, but also more than 30 times more source vertices.
Figures 2(a) and 3(a) show the complementary cumulative distribution functions for different
hazard levels. The results follow the trend that the probability of exceedance of a certain value
of connectivity loss increases with the hazard level. These results are the basis for the network
fragility curves of different damage states shown in Figures 2(b) and 3(b). We observe that the
more extensive a damage state, the lower is the probability of its occurrence at any given PGA.
This is more evident with the shift of the network fragility curves of the more extensive damage
state to the right. Next, we observe the steep part of the curves, where significant changes in
probabilities occur at small variations of hazard levels; however, the more extensive the damage
state, the gentler is the slope.
Furthermore, these results show that the gas network is more vulnerable to earthquake hazard than
the electricity network. Notice that at the hazard level of 475-year return period the performances of
the networks differ considerably. In the gas network the minor (20% CL) and moderate (50% CL)
damage states would certainly be achieved, and the extensive (80% CL) damage state would occur
with 44.4% probability of exceedance. Conversely, in the electricity network the minor damage
state would occur with only 6.2% probability of exceedance, and the moderate and extensive
damage states would not even arise.

Figure 2. Results of Monte Carlo simulations in the case of European gas network: (a) presented for
different hazard levels as complementary cumulative distribution functions and (b) summarized in network
fragility curves for different damage states.

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
ASSESSMENT OF CRITICAL INFRASTRUCTURE SYSTEM 69

Figure 3. Results of Monte Carlo simulations in the case of European electricity network: (a) presented
for different hazard levels as complementary cumulative distribution functions and (b) summarized in
network fragility curves for different damage states.

We impute the reason for this difference in vulnerability to the different geographical distribution
of the source vertices in each of the networks, and to the high vulnerability of the long pipelines.
Probabilities of failure of the vertices are always a reflection of the seismic hazard map (Figure 4).
This is because the fragility curves of the vertices are mostly dependent on the PGA of the vertex
location. This is not the case for links (pipes), where the fragility is not only dependent on the
PGA of the end vertices, but also on the length of the link itself. Therefore, in the gas network, the
pipelines (Figure 4(b)) that have the highest probabilities of failure do appear not only in the areas
with the high PGA values, but also in the source countries (Algeria, Turkey, and North Sea in
Norway). This is because the majority of gas pipelines in the source countries are very long, since
they must transport the gas directly from the gas fields to the areas of the high gas consumption.
It is precisely the disconnection of the gas fields that causes the fast rise in the connectivity loss
because the gas fields represent 87% of all source vertices, whereas the sources in the gas network
represent only 5% of all the vertices. As such, gas fields play an important role in the network
performance. On the other hand, in the electricity network (Figure 4(c)) this long-range peculiarity
is not present, so, first, only the vulnerability of vertices has been analysed and, second, we can
assume that the sources and the sinks are evenly distributed across Europe. Consequently, the
sinks mapped with high PGA regions will suffer heavier connectivity loss than those sinks within
the low PGA regions, but the range of loss is levelled out at the network level as a result of the
averaging process contained in Equation (1).

6. THE MODELLING OF NETWORK INTERDEPENDECY

In order to introduce interdependency between the networks (also known in the literature as
coupling behaviour), we will have to return to the element level of our model. In our case, the
electricity network is dependent on the gas network but only through gas-fired power plants. Strictly
speaking, gas-fired power plants are exposed to earthquake damage as well as to disturbances of
their gas supply. This additional vulnerability is applied directly on a gas-fired power plant when
it is considered as the vertex of the electricity network.
Rinaldi et al. [31] presented a universal multidimensional concept of the coupling behaviour,
which although a very generic description, could, on the other hand, be useful for the identifi-
cation of all possible interaction mechanisms. Among these, we find it possible to explore three
aspects that manage to capture the essential mechanism of coupled networks. These are: physical
interdependence, direction of the interaction, and strength of coupling, which we build into our
probabilistic reliability model using the interoperability matrix, similar to [15], as follows.

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
70 K. POLJANŠEK, F. BONO AND E. GUTIÉRREZ

Figure 4. (a) Seismic hazard map of Europe [21] for the 475-year return period, then relating
probabilities of failure of the elements considered in (b) the European gas network and (c) the
European electricity network (visualized by PAJEK [30]).

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
ASSESSMENT OF CRITICAL INFRASTRUCTURE SYSTEM 71

6.1. Theoretical concept of the interdependency


Generally, we try to capture bidirectional relationships, but for each direction of the coupling
behaviour we define one interoperability matrix. The size of the interoperability matrix is always
Nindependent × Ndependent , where Nindependent refers to the size of the independent network and
Ndependent is the size of dependent network. Two possible interoperability matrices arise by
switching the roles between dependent and independent networks. For example, I E/G is an
N G × N E interoperability matrix which captures the influence of the gas network on the electricity
network, and IG/E is an N E × N G interoperability matrix which captures the influence of the
electricity network on the gas network (N G —the size of the gas network, N E —the size of the
electricity network).
The interoperability matrix is a weighted version of inter-connectivity matrix. Inter-adjacent
vertices are specified with the type of dependency—in our case physical dependence where the
state of each infrastructure depends upon the material output of the other, but, of course, defined at
the element level. Since all vertices in one network usually do not couple to all vertices in the other,
the inter-connectivity matrix is very sparse. For the electricity network, we may consider that only
gas-fired power plants are dependent on the gas supply from the gas network. If the vertex i from
the gas network and the vertex j from the electricity network are adjacent, the (i, j) element in the
corresponding inter-connectivity matrix A E/G receives the value 1 and represents the directed link
between two networks. Conversely, A G/E is full of zeros because the supply of electrical energy
to run the compressor stations is not taken into account (many compressor stations run on gas).
So, the direction of interaction is only one way and, therefore, the gas network is independent,
whereas the electricity network is partially dependent on the gas network.
Finally, to complete the interoperability matrix, the weights are introduced to describe the
strengths of coupling, referring them directly to the interconnections. The strength of coupling
dep
is defined as the conditional probability of failure of the jth electricity vertex E j at the given
failure G i of the ith adjacent gas vertex:
dep
P(E j |G i ) = p E j|Gi for all i adjacent to j, (4)
dep
which can be tuned from complete independence P(E j |G i ) = 0, to complete dependence
dep
P(E j |G i ) = 1; simulating loose and tight coupling, respectively, on a plant-by-plant basis. Thus,
tightly coupled interactions are those that do not tolerate delay; for example, disturbances in the
gas supply would have an almost immediate effect on electrical power generation. On the other
hand, loose coupling implies the presence of slack in the system because the propagation of the
disturbances from one network to another is very slow. For example, a gas-fired power plant
could have local gas storage or could switch to alternative fuel. Such situations could also imply
different levels of loose coupling of the gas-fired power plants to the gas network.

6.2. Implementation of the probabilistic reliability model


To consider interdependency behaviour with the probabilistic reliability model, we need the updated
probabilities of failure of electricity vertices. In general, the failure of the jth vertex of the

Figure 5. Venn diagram of failure of the electricity vertex: (a) due to earthquake and
(b) due to the failure of the adjacent gas vertex.

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
72 K. POLJANŠEK, F. BONO AND E. GUTIÉRREZ

electricity network, denoted E j , can occur due to an earthquake or due to the failure of the gas
dep
supply, denoted E earth
j and E j , respectively. Therefore, the event E j (Figure 5(a)) can be defined
dep dep dep
as the union of the events E earth
j and E j : E j = E earth
j ∪ E j . Because E earth
j ∩ E j = 0, these
two events are not mutually exclusive and the probability of the electricity vertex failure can be
formulated as:
dep dep dep
P(E j ) = P(E earth
j ∪ E j ) = P(E earth
j )+ P(E j )− P(E earth
j ∩ E j ). (5)
dep
Since we know that E earth
j and E j are statistically independent, the joint probability of E earth
j
dep dep dep
and E j equals P(E earth
j ∩ E j ) = P(E earth
j )P(E j ), therefore we get:
dep dep dep
P(E j ) = P(E earth
j ∪ E j ) = P(E earth
j )+ P(E j )− P(E earth
j )P(E j ). (6)

Our next task is to calculate all the probabilities in Equation (6). First, P(E earth
j ) is determined
dep
from the fragility curves of the elements’ seismic vulnerability. Conversely, obtaining P(E j )
dep
is not so straightforward. We know that event E j will occur only after the occurrence of the
dep
failure of the adjacent gas vertex (denoted as event G i ). So, events E j and G i are statistically
dependent (Figure 5(b)). The relationship among the probabilities of their occurrences is defined by
dep
the conditional probability that expresses the probability of event E j given the occurrence of G i :
dep
dep P(E j ∩ G i )
P(E j |G i ) = . (7)
P(G i )
It is convenient for our analysis that we deal with the extreme case of the dependent events
dep
where one set (events G i ) contains the other (events E j ) as presented in Figure 5(b). Since
dep dep dep
the intersection of G i and E j is explicitly defined as E j ∩ G i = E j , the joint probability is
dep dep
expressed as P(E j ∩ G i ) = P(E j ). Now, we can simplify Equation (7) as:
dep dep
P(E j ) = P(E j |G i )P(G i ). (8)
dep dep
Eventually, for the statistical realization of the P(E j ), we need to know P(E j |G i ) and P(G i ).
dep
P(E j |G i ) is actually the strength of coupling p E j|Gi defined in the interoperability matrix,
whereas the calculation of probability P(G i ) is entirely related to the vulnerability of the inde-
pendent network (Section 6.3).

6.3. Gas-source supply stream fragility curves


The failure event of a gas vertex distributing gas G i , can be induced by an earthquake G iearth
or by the disconnection from the sources G iconnect . So, event G i is defined as the union of G i =
G iearth ∪ G iconnect . Since both events are not mutually exclusive (G iearth ∩ G iconnect = 0), they can
happen simultaneously, so the joint probability theoretically equals:

P(G i ) = P(G iearth ∪ G iconnect ) = P(G iearth )+ P(G iconnect )− P(G iearth ∩ G iconnect ). (9)

P(G iearth ) are determined from the elements’ seismic fragility curves, whereas the disconnection of
gas vertices from the sources P(G iconnect ) is impossible to compute analytically. The reason for this
is because P(G iconnect ) reflects the seismic response of the whole gas network (Figure 6), which is
associated with the probabilities of failure of other components in the gas network. However, the
probability P(G iearth ∪ G iconnect ) can be measured with the connectivity loss calculated as part of
the seismic-performance network analysis of the gas network using the Monte Carlo simulation.

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
ASSESSMENT OF CRITICAL INFRASTRUCTURE SYSTEM 73

Figure 6. Schema of gas-source supply stream of the gas-fired power plant.

Figure 7. The gas-source supply stream fragility curves for all gas-fired power plants (left) and the
European electricity network showing the probability of failure of gas vertices adjacent to gas-fired power
plants at hazard level of 475-year return period (right).

In that case we assume that event G iearth is considered as a 100% connectivity loss of the gas
vertex i. The probability of failure P(G i ) due to both causes is then determined as
i
NgasPP,dam
CL = 1−
i
i
, (10)
NgasPP,orig
i
where NgasPP,orig i
and NgasPP,dam are the number of sources connected to the ith gas vertex adjacent
to gas-fired power plant in the original and in the damaged network, respectively.
For each gas vertex adjacent to a gas-fired power plant the fragility curve is constructed in
terms of connectivity loss and local PGA (Figure 7). These fragility curves describe the seismic
vulnerability of the whole gas-source supply stream, while also taking into account the whole
topology of the gas network. The gas-source supply stream failure is defined by the exceedance
of the damage state defined by 80% of connectivity loss of the gas vertex i adjacent to a gas-fired
power plant.

6.4. Results of dependent networks


With the Monte Carlo simulation of element failures, we construct the dependent network fragility
curves for the dependent network which, in our case study, is the electricity network. Here, the
failure probability of the electricity vertices P(E j ), as required in the probabilistic reliability model

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
74 K. POLJANŠEK, F. BONO AND E. GUTIÉRREZ

Figure 8. Network fragility curves of the European electricity network in terms of connectivity loss (left),
power loss (middle) and impact factor on the population (right).

dep
(Section 5), is determined with P(E earth
j ), P(E j |G i ) and P(G i ). In contrast to the analysis of
the independent electricity network (Section 5.1), we present now three performance measures:
connectivity loss, power loss, and impact factor on the population. Likewise, we extract network
fragility curves for three damage states, and we consider three different values of strength of
dep dep
coupling to model the total independence (P(E j |G i ) = 0), the partial dependence (P(E j |G i ) =
dep
0.5), and the complete dependence (P(E j |G i ) = 1). In order to construct the damaged network,
the same strength of coupling is used for all gas-fired power plants. Finally, the network fragility
curves for the dependent network are presented in Figure 8.
We note from Figure 8 the small difference between the different performance measures. We
should stress that all presented metrics are correlated, because they are based on the same topo-
logical concepts, but they represent damage in alternative ways. In fact, the power loss and impact
on the population are only the weighted versions of the connectivity loss. As regards the weights,
almost 80% of power plants have nominal power rating less than 100 MW (max. 5706 MW), and
almost 60% of distribution substations cover less than 100 000 people (max. 3.35 million). Since
frequency distributions of the chosen weights are extremely skewed, we can only observe minor
changes in the final value of the performance measures.
Next, we found that the dependence of the electricity network on the gas network introduces an
extra vulnerability to the electricity network response under seismic hazard. This dependency effect
shifts the fragility curves slightly to the left; meaning that the higher strength of coupling induces
the chosen damage state with the same probability at the smaller PGA. For example, there is 50%
probability of exceeding the 20% (respectively 50, 80%) of connectivity loss occurring at 0.11 g
(respectively 0.09, 0.08 g) less PGA when increasing the strength of coupling for full range (0–1).
But the dependency effect does not interfere with the general behaviour of the network response;
i.e. the shapes of the network fragility curves are similar. So, the network fragility curves are very
steep for minor damage states, whereas the slopes of the fragility curves decrease when observing
more extensive damage states. Such behaviour was seen already in the case of the independent

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
ASSESSMENT OF CRITICAL INFRASTRUCTURE SYSTEM 75

electricity network. Consequently, the narrower curves and the gentler slopes indicate a smaller
dependency effect as the damage state increases.
The influence of coupling behaviour captured in the network fragility curves can be explained
vis-á-vis some topological characteristics of the dependent electricity network:
(1) Physical interdependency. We assumed one-to-one connection between two networks. For
each gas-fired power plant in the electricity network there is one adjacent vertex in the
gas network. The spread of damage from one network to the other is certainly depen-
dent on the number of connections among the networks. Thus, in the electricity network,
only gas-fired power plants directly suffer the disturbances in gas supply. Such physical
interdependency dictates slow damage propagation from one network to the other. This is
observed by the slight shift of the fragility curves, which is controlled by the strength of
coupling.
(2) Gas-fired power plants are a minor fraction of source vertices. Gas-fired power plants in
the European electricity network represent only 18.6% of all the source vertices. So, even
at high hazard levels, the damage due to the dependency effect cannot exceed 18.6% of
connectivity loss. Besides, sources in the European electricity network represent more than
half of all the vertices (51.2%). Based on the definition of the connectivity loss, we can say
that the higher the fraction of sources in the network, the smaller is the effect of failure of
one source on the connectivity loss.
(3) Gas-fired power plants are one-degree vertices. One-degree vertices do not lie on any path
between sinks and sources. If we eliminate a one-degree vertex we will not disconnect
any sink from its sources except in the case when that vertex is the source or sink itself.
Elimination of the gas-fired power plants raises the connectivity loss due to its role as the
source, but it does not change the topology and, consequently, neither does the response of
the transmission part of the electricity network.

6.4.1. Betweenness centrality attack versus seismic hazard and strength of coupling. In this chapter
we will explore the relation between two topological measures: betweenness centrality [20] and
connectivity loss measures.
The failure of gas-fired power plants results in an increase in connectivity loss, which is an
indication of the dependency effect of interconnected networks. However, there are other vertices
whose failure has a higher influence on the connectivity loss. These are vertices that are on the
path of many connections between sinks and sources which can be identified by the betweenness
centrality measure. The higher the betweenness centrality measure of a vertex, the more often
the vertex lies on the shortest paths between all pairs of vertices, whereas betweenness centrality
considers all the shortest paths, connectivity loss checks only the shortest paths between sources
and sinks; we conjecture that the elimination of the vertices with higher betweenness centrality
would accelerate the connectivity loss. Knowing that one-degree vertices (such as gas-fired power
plants) have betweenness centrality zero whereas all other vertices in the transmission part of the
electricity network have betweenness centrality greater than zero, then one will understand why the
direct earthquake failures of the high beetweenness centrality vertices mask the implicit increase
of the connectivity loss due to the dependency effect.
In order to support the above idea, we compare the response of the electricity network under
seismic hazard and interdependency effect to that under betweenness centrality attack. This attack
strategy is defined as successive removal of the vertex with the highest betweenness centrality
ranking, and where betweenness centrality is recalculated for each new damaged network iteration.
These results are presented in Figure 9 with the connectivity loss on the ordinate and the fraction
of the removed vertices on the abscissa. But note, we consider seismic hazard level and strength
of coupling as the parameters of the third and the fourth parametric dimensions.
In all the situations presented in Figure 9, the connectivity loss increases with the fraction of
removed vertices. So far, the fastest increase appears in the case of the betweenness centrality
attack because it is topologically biased toward the most important vertices for the connectivity of
the network. This is not the case in the seismic hazard because the elimination of the vertices has

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
76 K. POLJANŠEK, F. BONO AND E. GUTIÉRREZ

Figure 9. The European electricity network: comparison of connectivity loss between the betweenness
centrality attack and seismic hazard (average values of all the simulations in one series) with different
strength of coupling (left) and vertex betweenness centrality (right).

no relation with the connectivity of the network, but it is conditioned by the geographical position
of the vertices on the seismic hazard map.
Next, we can follow trends of connectivity loss along increasing hazard level or along the
increasing strength of coupling. Notice that the connectivity loss increases faster with increasing
hazard levels than with higher strength of coupling. Moreover, at the lower hazard levels the
increasing strength of coupling causes a higher increase in the connectivity loss than at the higher
hazard level. Both of the above are the main arguments of what was stated earlier, namely that
earthquake failures mask the increase in the connectivity loss due to the dependency effect. Besides,
we can also confirm that the average number of the removed vertices increases not only with the
hazard levels, but also with the strength of coupling.

6.4.2. Geographical spread of damage. Until now, we have observed the performance of the
network as one macroscopic structure. We notice that the averaging procedure incorporated into
the definitions of the performance measures filters out extreme damage concentrated on a certain
geographical location. Yet, using GIS we can relate the results with the Thiessen polygons (areas
covered by each of the distribution substations) which become the final object of the analysis and
visualization.
We calculated the power loss for each of the distribution substations of the electricity network
and we assigned its value to the corresponding Thiessen polygon. We have calculated results for
different hazard levels and different strengths of coupling but have presented them as the average
value of all the simulations executed in one series. Furthermore, we multiply the population in
each Thiessen polygon by the distribution substation’s power loss, and so obtain the absolute value
of the population affected for each of the Thiessen polygons. Finally, the spatial distribution of
both performance measures, power loss of the distribution substations, and the affected population
(Figure 10) shows the geographical spread of damage at a European-wide level. Here, in contrast
to the overall damage assessment, different performance measures reveal the vulnerability levels of
different parts of the network. The areas of highest power loss are in southeast Europe (as expected
according to the seismic hazard map), and within those areas the highest impact on the population
is located in the urban zones. We also notice that extreme values for the affected population do not
coincide with the highest values of power loss. It appears, for example, in the Netherlands, where
a locally not so branched electricity transmission network, generates large Thiessen polygons of
high population density; it is for this reason that a large population is assigned to only a few
distribution substations. To summarize, the results for the affected population are a combination

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
ASSESSMENT OF CRITICAL INFRASTRUCTURE SYSTEM 77

Figure 10. Comparison of the influence of the strength of coupling and PGA factor on the power loss and
the affected population for all distribution substations of the European electricity network.

of the population density, the size of the Thiessen polygons and all the factors that influence the
power loss of a given distribution substation.

7. CONCLUSIONS

A GIS-based probabilistic reliability model was developed in order to generate network fragility
curves of spatially distributed interconnected network systems subjected to natural hazards. The
method was successfully applied to encompass the geographic distributions of both the infras-
tructure and the natural hazard; specifically, we analyzed the interconnected European gas and
electricity transmission networks in such a manner that the gas-fired power plants form the physical
connections between the two types of networks. The network interdependency model combines
gas-source supply stream fragility curves and strength of coupling in order to follow (in a topo-
logical sense) the propagation of failures resulting from seismic vulnerability of the gas network
and how they affect the topology of the electricity network.
We have analysed these failures in terms of performance measures of network response; i.e.
connectivity loss, power loss, and impact factor on the affected population, which are based on

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
78 K. POLJANŠEK, F. BONO AND E. GUTIÉRREZ

topological metrics. These metrics have been used for macroscopic damage evaluation (condensed
in the form of network fragility curves), and also for the local examination of the damage status of
each and every electricity distribution substation in the electricity network. The network fragility
curves for independent networks highlight the higher seismic vulnerability of the gas network
compared to that of electricity. The reason is because gas fields are connected to the main gas
network along long (hence vulnerable) pipelines. The dependence of the electricity network on
the gas network introduces an additional seismic vulnerability of the electricity network over and
above the explicit seismic vulnerability of its structural components. This dependency effect causes
a slight shift in the network fragility curves toward the smaller hazard levels without changing
their shape. The high slopes of the network fragility curves are an intrinsic network characteristic,
be it independent or otherwise. These slopes decrease with increasing network damage states and,
consequently, we can observe a decrease in the dependency effect. Thus, dependent network fragility
curves capture many topological characteristics of the electricity network: that the propagation of
damage is slow due to one-to-one connection between gas-fired power plants and their adjacent
gas distribution vertices; that because gas-fired power plants cover circa 20% of power generation
capacity for Europe as a whole, their failure cannot cause a power loss major than this (even at
the extreme cases of hazard levels); and finally, that the elimination of the gas-fired power plants
cannot reduce the connectivity of the transmission part of the electricity network because these
power plants correspond to one-degree vertices. For the same reason, seismic-induced failures
of the other elements of electricity networks mask the damage that results from the dependency
effect; moreover, the masking effect becomes even more pronounced with increasing hazard levels.
Furthermore, a geographical distribution of the damage at the European level shows, as expected
from the seismic hazard map, that the highest direct damage is located in southeast Europe.
Nevertheless, it also reveals that the impact factor on the population is highest in the Netherlands
that is possibly due to the local (chain-like) topology of the electricity transmission network
combined with a high population density.
However, we cannot conclude from our analysis that the apparent low vulnerability dependence
of electricity on gas-fired generation is so clear cut. For example, the recent geopolitical crises
between Russia and Ukraine highlighted another coupling mechanism between the gas and the
electricity systems. So, if the gas supply is cut off, some direct users (heating devices at home or
industry) of natural gas can switch their energy source. As a consequence, the switch to electricity
would cause an extra load in the electricity system. Such extra vulnerabilities, although outside
the scope of our structurally biased hazard analysis can, in principle, be equally well studied using
the same probabilistic and GIS techniques described above.
Modeling the vulnerability of widely dispersed critical infrastructure networks has become a
necessary but difficult task due to the geographic extent, large number of connected components
and interdependencies between networks. Our approach provides a probabilistic estimate of the
performance of infrastructures exposed to catastrophic events, and can be used to identify the
most vulnerable parts of the network. Such network models can not only be used for design and
planning of infrastructures, but also for better allocation of funds to mitigate their vulnerability to
natural hazards. It is worth mentioning that only growing awareness and knowledge of the existing
vulnerabilities in critical infrastructure systems will eventually increase the investments needed to
increase resilience.

ACKNOWLEDGEMENTS
The results presented in this paper are based on the work supported by MANMADE Project contract
no. 043363 of the 6th EU Framework Programme (FP6) dealing with diagnosing vulnerability, emergent
phenomena, and volatility in man-made networks.

REFERENCES
1. Bollobas B. Random Graphs. Academic Press: London, 1985.
2. Watts DJ, Strogatz SH. Collective dynamics of ‘small-world’ networks. Nature 1998; 393(6684):440–442. DOI:
10.1038/30918.

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe
ASSESSMENT OF CRITICAL INFRASTRUCTURE SYSTEM 79

3. Barabási AL, Albert R. Emergence of scaling in random networks. Science 1999; 286(5439):509–512. DOI:
10.1126/science.286.5439.509.
4. Albert R, Jeong H, Barabási AL. Error and attack tolerance of complex networks. Nature 2000; 406(6794):378–482.
DOI: 10.1038/35019019.
5. Holme P, Kim BJ, Yoon CN, Han SK. Attack vulnerability of complex networks. Physical Review E 2002;
65(5):056109. DOI: 10.1103/PhysRevE.65.056109.
6. Callaway DS, Newman MEJ, Strogatz SH, Watts DJ. Network robustness and fragility: percolation on random
graphs. Physical Review Letters 2000; 85(25):5468–5471. DOI: 10.1103/PhysRevLett.85.5468.
7. Dall’Asta L, Barrat A, Bathelemy M, Vespignani A. Vulnerabilty of weighted networks. Journal of Statistical
Mechanics 2006; P04006. DOI: physics/0603163.
8. Petreska I, Tomovski I, Gutierrez E, Kocarev L, Bono F, Poljansek K. Application of modal analysis in assessing
attack vulnerability of complex networks. Communications in Nonlinear Science and Numerical Simulation 2010;
15(4):1008–1018. DOI: 10.1016/j.cnsns.2009.05.0022009.
9. Cohen R, Erez K, Ben-Avraham D, Havlin S. Breakdown of the Internet under intentional attack. Physical Review
Letters 2001; 86(16):3682–3685. DOI: 10.1103/PhysRevLett.86.3682.
10. Cohen R, Erez K, Ben-Avraham D, Havlin S. Resilience of the Internet to random breakdowns. Physical Review
Letters 2000; 85(21):4626–4628. DOI: 10.1103/PhysRevLett.85.4626.
11. Albert R, Albert I, Nakarado GL. Structural vulnerability of the North American power grid. Physical Review E
2004; 69(2):025103, 4. DOI: 10.1103/PhysRevE.69.025103.
12. Ang AHS, Pires J, Villaverde R. Seismic reliability of electric power transmission systems. In The 12th
International Conference on Structural Mechanics in Reactor Technology, Kussmaul K (ed.). Elsevier Publishers
B.V.: Amsterdam, 1993.
13. Shinozuka M, Cheng TC, Feng MQ, Mau ST. Seismic performance analysis of electric power systems. MCEER
Research Progress and Accomplishments 1997–1999; 61–69.
14. Vanzi I. Structural upgrading strategy for electric power networks under seismic action. Earthquake
Engineering and Structural Dynamics 2000; 29(7):1053–1073. DOI: 10.1002/1096-9845(200007)29:71053::AID-
EQE9543.0.CO;2-X.
15. Dueñas-Osorio L, Craig JI, Goodno BJ. Seismic response of critical interdependent networks. Earthquake
Engineering and Structural Dynamics 2007; 36(2):285–306. DOI: 10.1002/eqe.626.
16. Platts. Available from: http://www.platts.com, 2008.
17. Poljansek K, Bono F, Gutierrez E. GIS-based method to assess seismic vulnerability of interconnected
infrastructure: a case of EU gas and electricity networks. JRC Scientific and Technical Reports. EUR 24275 EN.
Luxembourg (Luxembourg): Publications Office of the European Union, 2010. JRC57064. DOI: 10.2788/71352.
18. Environmental Systems Research Institute. ArcGIS 9—Geoprocessing Commands Quick Reference Guide. ESRI
Press: Redlands, CA, U.S.A., 2005.
19. LandScanTM Global Population Database. Oak Ridge National Laboratory: Oak Ridge, TN. Available from:
http://www.ornl.gov/landscan/, 2009.
20. Brandes U, Erlebach T. Network Analysis: Methodological Foundations. Springer: Berlin, Heidelberg, Germany,
2005.
21. Giardini D, Grünthal G, Shedlock KM, Zhang P. The GSHAP global seismic hazard map. In International
Handbook of Earthquake and Engineering Seismology, International Geophysics Series 81 B, Lee W, Kanamori H,
Jennings P, Kisslinger C (eds). Academic Press: Amsterdam, 2003; 1233–1239.
22. Meletti C, Montaldo V. Stime di pericolosità sismica per diverse probabilità di superamento in 50 anni: valori
di ag. Progetto DPC-INGV S1, Deliverable D2, 2007. Available from: http://www.esse1.mi.ingv.it/d2.html.
23. Environmental agency of the Republic of Slovenia. Available from: http://www.arso.si [23 February 2009].
24. Adachi T, Ellingwood BR. Comparative assessment of civil infrastructure network performance under probabilistic
and scenario earthquakes. Journal of Infrastructure Systems 2010; 16(1):1–10. DOI: 10.1061/(ASCE)1076-
0342(2010)16:1(1)).
25. Jayaram N, Baker JW. Efficient sampling and data reduction techniques for probabilistic seismic lifeline risk
assessment. Earthquake Engineering and Structural Dynamics 2010; 39(10):1109–1131. DOI: 10.1002/eqe.988.
26. FEMA. HAZUS-MH technical manual: earthquake model. Multi Hazard Loss Estimation Methodology. United
States Department of Homeland Security, Federal Emergency Management Agency, Washington, DC, 2003.
27. O’Rourke TD, Toprak S, Sano Y. Factors affecting water supply damage caused by the northridge earthquake.
Proceedings of the 6th US National Conference on Earthquake Engineering, Seattle, WA, June, 1998; 1–12.
28. Wald DJ, Quitoriano V, Heaton TH, Kanamori H. Relationships between peak ground acceleration, peak
ground velocity, and modified Mercalli intensity in California. Earthquake Spectra 1999; 15:557–564. DOI:
10.1193/1.1586058.
29. O’Rourke MJ, Ayala G. Pipeline damage due to wave propagation. Journal of Geotechnical Engineering (ASC E)
1993; 119(9):1490–1498. DOI: 10.1061/(ASCE)0733-9410(1993)119:9(1490).
30. Pajek 1.22. Available from: http://www.vlado.fmf.uni-lj.si/pub/networks/pajek/ [February 2008].
31. Rinaldi SM, Peerenboom JP, Kelly TK. Critical infrastructure interdependencies. IEEE Control Systems 2001;
21(6):11–25. DOI: 10.1109/37.969131.

Copyright 䉷 2011 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2012; 41:61–79
DOI: 10.1002/eqe

You might also like