Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

On Inferring how resources are shared in IoT

ecosystems; a Graph Theoretic Approach


N. Kouvelas∗ , V. Balasubramanian∗ , A. G. Voyiatzis† , R. R. Prasad∗ , D. Pesch‡
∗ ES,
EEMCS, TU Delft, the Netherlands,
† SBA
Research, Austria,‡ Nimbus Research Center, CIT, Ireland
N.Kouvelas, V.Balasubramanian, R.R.VenkateshaPrasad@tudelft.nl, avoyiatzis@sba-research.org, Dirk.Pesch@cit.ie

Abstract—The Internet of Things (IoT) is an enabler the criticality of their specific functions. They also face
of the digital transformation dictating new needs and severe constraints regarding their energy sources, relying
trends in the domains of business and technology. mostly on (slow recharging) batteries as human interven-
Ecosystems of IoT devices are often organized in net-
works, using wireless technology and sharing access tion is often not possible [1], [2]. In these settings, energy
infrastructure. These networks are used to monitor a harvesting, for example by means of renewable sources
wide range of systems, from simple household activities (e.g., solar panels) or piezoelectric materials, can provide
to fully-interconnected smart cities. In many usage sce- a complementary source of energy. Existing approaches
narios, the IoT devices are resource-constrained. Thus, focus on extending the active operation of a specific device
energy scavenging is utilized to meet their expanding
longevity requirements. In this paper, we study the and supporting the overall network longevity [1].
local resource dynamics of IoT devices in an ecosystem, We consider the case of ecosystems of IoT devices. We
i.e., a set of different IoT devices that co-exist in spa- use this term to describe a set of IoT devices with different
tiotemporal level to coordinate the use of available com- capabilities and energy requirements that co-exist in space
mon resources for their individual goals. To this end, and time. Different devices might belong to different local
we model an ecosystem of IoT devices as a time-varying
graph and provide a theoretical foundation for resource networks performing a different set of tasks. However, all
distribution using Graph Theory. We show that simple these IoT devices coordinate the use of available common
graph-theoretic metrics, such as, the clustering coeffi- resources (e.g., energy or spectrum) for their individual
cient and degree distribution, can provide rich informa- aims. A smart neighborhood is an example ecosystem of
tion about the priority policy that is followed for the IoT devices (e.g., smart buildings, smart traffic lights,
distribution of resources among different IoT devices.
We take the case of micro grids; with some nodes having smart parking lots). The smart neighborhood covers its
harvesting potential and smart meters measuring the energy needs by drawing power from the micro grid. It
current consumption/generation and being connected also produces and shares energy locally and thus reduces
to the control unit. We use this notion in our example or offloads the costs. Each IoT device or network of devices
use-case, appropriating this to micro-grids with enough can employ its own policy regarding energy management,
harvested energy. Even one link per node can describe
an ecosystem as a connected component with more than as previously addressed in the literature. Once a device
60% of its total energy needs covered. Additionally, the is fully charged, it stops harvesting energy. However, such
nodes presenting harvesting potential are formed into an interconnected mesh of IoT devices has the potential
unipartite graphs of affiliation networks. Studying their to collect excess energy and redistribute it inside the
clustering coefficient we infer the priority policy that ecosystem.
ia applied when excess energy is shared within their
ecosystem. In this paper, we will characterize in an abstract way
how the excess of energy (and generally any resource)
I. Introduction is distributed between the devices of an IoT ecosystem
by employing Graph Theory. Since we target a micro
The Internet of Things (IoT) realizes the vision of grid application as an example, we use the term ‘energy’
an interconnected future, where any tangible ‘object’, or to also represent any other resources. For example, if,
‘thing’, can be part of the Internet regardless of its loca- instead of energy, computational burden is the resource
tion. The recent advances in communication technology to be shared, then the heavy computational tasks can be
enable dynamic and distributed systems furthering the allocated to certain IoT devices in order for others, the
cause of the IoT. Examples of contemporary IoT-based prioritized ones, to improve the longevity of their batteries.
applications include monitoring of the indoor air quality Using graph-formations and graph-metrics, we can infer
of buildings [1] or the cargo and structural health of which IoT devices carry the highest computational burden
transportation means in logistics [2]. The IoT devices (e.g., inside an IoT ecosystem and under which prioritization
sensors, actuators, and communication gateways) must strategy. The above reveals the advantage of choosing
work seamlessly and unattended for long periods, given Graph Theory as a method of explaining interactions (e.g.,
978-1-4673-9944-9/18/$31.00
c 2018 IEEE energy or computations transfer) among members (e.g.,

760
Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on November 14,2022 at 15:03:41 UTC from IEEE Xplore. Restrictions apply.
each set of devices in a network. Specifically in [6], devices
are grouped inside the ecosystem, forming disjoint sets.
Among these sets, only a few are monitoring continuously,
while the rest are duty cycled, and thus network lifetime
is extended. Within the same scope, in [5] the authors or-
ganize devices in cooperative groups to reduce the number
of transmitted messages and thus the traffic in the system.
In [8], traffic reduction inside the network is approached
from the scope of reducing the computational burden of
data compression. A ‘random walk’ algorithm for data
gathering is proposed in accordance with compressive
sensing. Tree-graphs are designed for data gathering in [9].
Loads of sensed data are balanced by using probabilistic
Fig. 1: Abstraction level: ecosystem of IoT devices sharing methods among the leaves of the tree networks. Using
common resources. Study case: micro grid of consumers simulation results, both [8] and [9] demonstrate that the
and producers sharing excess energy. lifetime of their networks is extended. In [3], it is claimed
that if two devices can communicate directly (if their
transmit signals overlap), they should share the same
IoT devices) of communities/graphs (e.g., micro-grid, IoT amount of resources, and thus be considered as one node.
ecosystem). Due to the high level of abstraction that After forming graphs under this scheme, they proceed on
Graph Theory provides, we do not need to focus on a resource allocation using the graph coloring method. The
particular resource. Instead, we can define unified models impact of node failures in an ecosystem of device nodes is
of resource sharing that enclose several systems and types analyzed in [10]. Mobile and stationary relays are proposed
of resources. In Fig. 1, we see the analogy of an ecosystem to reconnect the segmented parts of faulty graphs and
of IoT devices seen as a graph and an actual micro grid enable data transfer. Carrying the principles of the above
ecosystem. studies, a generalized graph theoretic ecosystem of IoT
The main contributions of this paper are the following: devices is modeled. By utilizing it we define how resources
(i) we present the analogy of an ecosystem of IoT devices are shared among devices.
to a graph; (ii) we use graph metrics (clustering coefficient
and node degree) and graph formations (bipartite graphs III. System Model
and giant components) to characterize the ecosystem re- Let us consider a graph G = (|v|, |e|) with |v| nodes
garding the policies used to prioritize different groups of and |e| links, where v = {v1 , v2 , ..., vv } are the nodes
nodes when the allocation of excess energy occurs inside; and e = {e1 , e2 , ..., ee } are the links. Additionally, let us
and (iii) we evaluate our theoretical study using fine- assume an ecosystem including IoT devices which coexist
grained consumption and generation data from an actual spatiotemporally. The devices exchange messages through
ecosystem of households which uses smart meters. controllers to distribute the excess energy they harvest in
the ecosystem, assuming that power transmission among
II. Related Works them is feasible. In this work, we do not go into the
In the literature, ecosystems of IoT devices, mostly used specifics of energy (or resource) distribution. Instead, we
for sensing the ambient environment, are represented as only focus on the modeling of such systems in general with
graphs to serve the following goals: (i) to organize the graphs and assume it is feasible to share resources once the
distribution of resources (from batteries and renewable solutions to the abstract sharing problems are found.
sources of energy) inside the network [3]; (ii) to route the The ecosystem can be considered as a time-varying
data inside the network efficiently [4], [5]; (iii) to achieve graph. The nodes of this graph are the entities related
monitoring coverage with the least amount of active to the energy transfer process, i.e., controllers and IoT de-
sensor-devices [6], [7]; (iv) to optimize data gathering [8], vices. The IoT devices that produce excess energy during
[9], and (v) to tolerate device-node failures [10]; All the a timespan are called provider (or producer) nodes and the
above approaches aim to reduce energy consumption in an rest are the receiver (or consumer) nodes, which are energy
ecosystem, in order to prolong its lifetime. In [4], a routing deficient. A device can be either a provider or a receiver at
protocol, namely FAF-EBRM, is designed to define the different timespans. A graph link between a provider and
most efficient path for the sensed data to be delivered a receiver denotes the transfer of excess energy between
from a source-node to any sink-node in a network. To these two nodes. A graph link between a provider/receiver
achieve this, before every ‘hop’, the link weights (defined and a controller denotes a communication link for message
during the graph creation) and the energy needed for data exchange purposes. The existence and formation of the
forwarding are checked. In [6] and [7], bipartite graphs are links depend on the way the excess energy is shared
designed according to the targets that are monitored by inside the ecosystem. This is illustrated using real data

761
Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on November 14,2022 at 15:03:41 UTC from IEEE Xplore. Restrictions apply.
we defined the deficiency and excess of energy for every
household through this data. To achieve convergence and
hence the stability of energy allocation, we smoothed the
daily (and hourly) differences in energy by averaging the
measurements over predefined weekly time intervals. Our
final dataset provides an example case of an IoT device
ecosystem, where each household optimizes independently
the balance between deficiency and excess of energy. Every
household was modeled as a graph node, either provider or
(a) Week 47 (b) Week 48 receiver at any time interval, according to the total balance
of energy generation and consumption as derived from the
Fig. 2: Nodes changing roles in a time-varying graph aggregated smart meter readings. Fig. 1 depicts visually
representing a micro grid. an abstract representation of the studied system.

V. Graph metrics for evaluation and results


in Fig. 2, where we observe that the links connecting
different nodes of the graph vary between two consecutive A. Node Degrees
timespans (weeks), as the energy transfer relations change The degree of a node, vi , dvi , is the number of links that
from time to time among the provider and receiver nodes. connect vi with other nodes [12], [13]. We consider only
Note that in our model there are no self-loops because self- time-varying graphs that contain links depending on the
transferring of the excess energy has no physical meaning. flow of resources (transferred energy in our case study),
Under the aforementioned assumptions, we proceed in the which occurs between providers and receivers. In a time-
following sections to study and characterize the energy varying graph and for a given time interval T , the quantity
allocation behavior in an ecosystem of IoT devices using dTvi = x denotes that if i is a deficient device (node), then
graph-theoretic approaches. i accepted excess of energy from x providers. If i is a
IV. Evaluation Model provider, then i transfers excess of energy to x deficient
nodes. If vi is a deficient node and x = 0, then the excess
The energy transactions in our evaluation model occur energy was depleted before it could be shared to vi .
in one of the three ways: (i) FIFO, prioritizing from the
In Fig. 3, we present the degree distribution for 263
highest deficient nodes to the lowest; (ii) FIFO, prioritizing
nodes without energy harvesting potential when prior-
from the lowest deficient nodes to the highest; and (iii)
itizing randomly, for low, or for high deficiency. The
randomly. We opted to use a real-world case for our
impact of prioritization is reflected on the degree of the
evaluation model rather than relying on synthetic datasets
nodes that received the excess energy. High degree-values
because of the following two reasons: (a) our model illus-
imply that the corresponding nodes received energy from
trates actual scenarios of micro grids, where the system
many providers and low degree-values imply the opposite.
energy is dynamic, measured using smart meters, and
We note that the policies designed to prioritize the low-
communicated over an IoT ecosystem; and (b) our model
deficient nodes (red) do not present extremities regarding
possesses a large scale of data over a long period of time.
degree distribution. This is because the energy needs of
We utilize the information from the Dataport database,
the prioritized nodes are easily covered and, as a result,
the world’s largest source of disaggregated customer en-
many nodes received excess energy. On the other hand, the
ergy data [11]. We selected the Pecan Street dataset,
priority policies that help the highly-deficient nodes (blue)
a neighborhood of smart households in Austin, Texas,
manage to cover the needs of a relatively small number of
USA. This is a real-world case of a highly-scalable system,
them. Thus, only a few nodes with resource constraints
where Internet-enabled smart meters sense the consump-
attract a higher number of links. For example, in week
tion of household appliances and energy harvesters. The
42, prioritizing the highly deficient nodes covers only four
smart meters provide measurements at high frequencies
of them with a degree-range of [30, 5]. Prioritizing the
– one reading every few minutes up to several readings
low deficient nodes covers over 40 nodes, which present
per minute. The dataset we selected comprises of 443
a degree-range of [5, 1] (cf. the zoomed part of week 42 in
households, out of which 180 generate energy using solar
Fig. 3). The random-selection service policy results (green)
panels1 . The availability of excess energy is not always
lie in between the two aforementioned policies.
guaranteed. We used the aggregated consumption and
We conclude that by computing the degree distribution
generation data (in kW) from the smart meters of all the
of the nodes that receive excess energy in an ecosystem of
households in the ecosystem during one year, discarding
IoT devices, and by comparing the results with those of
households with data of less than 300 days. Further,
a random priority policy, we can infer that the ecosystem
1 the datasets we queried are available at https://dataport.cloud/ prioritizes a certain type of nodes (e.g., the high or the
data/database of https://dataport.pecanstreet.org low deficient ones).

762
Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on November 14,2022 at 15:03:41 UTC from IEEE Xplore. Restrictions apply.
Fig. 4: Ratios of: (i) nodes of the largest connected com-
ponent over total nodes (blue), (ii) nodes with more than
Fig. 3: Weekly degree distribution for different energy 60% of their needs covered over total number of deficient
sharing priority policies, from week 40 (a) to week 43 (d). nodes (red), and (iii) links per node (black).

B. Giant Components
A connected component in a graph is a group of nodes
in which it is possible to go from any node to any other
node of the same group by using the group links. A
Giant Component (GC) is a connected component that
contains the majority of graph nodes in its corresponding
graph [12], [14], [13]. Assuming that the graph expands
by involving more nodes and links randomly, the GC
grows rapidly, until the whole graph has become one GC. Fig. 5: Links per node (x-axis) against the following: (i)
This occurs because, for every new node or link that is nodes of the largest connected component over total nodes
added, the possibility of connecting with the existing GC is (blue) and (ii) deficient nodes with more than 60% of their
higher than with any other non-GC graph component. The needs covered over total number of nodes in need of energy
GC, being the majority leader regarding nodes, attracts (red).
new nodes/links with higher probability than any other
connected component in the graph.
In our time-varying graph representation of ecosystems, This is attributed to the specific nature of link creation
the formation and size of a GC relate to the ratios of (i.e., ‘resource providers → receivers’). Other graph-types
excess energy over deficit for a number of providers and require higher link-per-node values to form a GC [12].
receivers. When these ratios are high, the graph is highly- Hence, we conclude that in an ecosystem of IoT devices
connected and thus, a GC can be observed. We isolate with low energy requirements and large amounts of har-
the 180 nodes of our dataset that can harvest energy. vested energy, a ratio of one link per node is sufficient to
Every week (our time interval), these nodes can be ei- connect the resulting graph into a GC and thus, cover
ther providers or receivers, depending on their available adequately the needs of deficient nodes, even if excess
energy. We apply random energy sharing to simulate the energy is transferred randomly.
creation of connected components. In Fig. 4, we relate the Deficiency coverage is strongly related to an increase
number of nodes of the largest connected component with in the fraction of links in the graph. The effect of link
the number of links in the resulting graph. We observe creation in the formation of a GC and energy sharing
that the existence of a GC is affected strongly by the is depicted in Fig. 5. Every deficient node is covered, at
number of links and in particular, by the fact that a link least over 60%, when the graph is connected in one GC.
representing energy transfer cannot exist between nodes of We arbitrarily chose 60% as a coverage percentage which
the same type (e.g., two energy providers). Thus, although guarantees that the deficient node is able to maintain its
the energy transfer is random, the GC does not expand basic functionalities (e.g., sensing, computing).
randomly, but according to the type of nodes and the We conclude that the ratio of nodes of the connected
amount of excess energy. When the ratio of links over the component over the total nodes grows exponentially when
total number of nodes is below 0.7, the nodes of the largest increasing the number of links in the graph. Furthermore,
connected component are less than 30% of the total nodes the fraction of nodes that are covered grows exponentially
of the graph. There is no GC, but rather several small when increasing the size of the largest connected com-
connected components. On the other hand, for higher link- ponent, until this becomes a GC. After this point, the
per-node values, GCs are formed. When these values are GC increases rapidly to connect the whole graph, covering
over 0.95, the study case ecosystem is fully-connected. adequately the needs of every deficient node.

763
Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on November 14,2022 at 15:03:41 UTC from IEEE Xplore. Restrictions apply.
C. Clustering coefficient
In a time-varying graph, it is important to know its level
of completeness when measuring robustness against node
failures. An undirected graph is characterized complete (or
full-mesh) when all the possible links that can be created
between the nodes exist. The completeness of a graph
is measured by defining the individual completeness of
each node inside it [13]. The level of completeness for (a) High deficiency concept (b) Low deficiency concept
a node vi of a graph G = (|v|, |e|) is the Clustering
Coefficient, denoted as CCvi . Assuming a node vi of a Fig. 6: Comparison of clustering coefficient for different
graph has n neighbors of degree higher than one (i.e., priority policies in energy sharing.
nodes directly connected with vi , which are connected
also to other nodes), CCvi is defined as the ratio of the TABLE I: Average clustering coefficient of nodes produc-
number of links connecting the n neighbors, |e(n) |, over ing excess energy and deficient ones from Fig. 6.
the total number of possible links that could connect the High Deficiency Low Deficiency
n neighbors, n(n − 1)/2 [13]: CCC 0.5 0.778
CCP 1 0.778
2|e(n) |
CCvi = ∈ [0, 1]. (1)
n(n − 1) subgraphs around the few deficit nodes of high degree;
while other deficit nodes that remain uncovered are not
To measure the CCG of the whole graph G, one must
considered for averaging clustering coefficients of Eq. (1).
consider all the individual clustering P coefficients of its
|v| In contrast, low-deficient node priority policies (e.g., the
nodes. Thus, for G = (|v|, |e|), CCG = i=1 CCvi . If the
ones presented in Fig. 6b) augment the diversity of nodes
result is |v|, the whole graph is complete. A high CCG
receiving energy. Thus, there are no nodes of high degree
implies that the graph maintains its functionality even
because of the low needs of the nodes that are served. The
after disconnecting several nodes because the remaining
degrees of the energy providers are increased as they can
nodes are able to operate over the available links.
cover the needs of a higher number of deficient nodes and
Considering our case study ecosystem, let the graph
form links with them, sharing their excess of energy more
represent energy allocation to nodes using different pri-
efficiently. We evaluated the insights provided by the case
ority policies. We use the average clustering coeffi-
described in Fig. 6 using our study case ecosystem. As
cient of provider and receiver nodes, defined as CCP
confirmed by Fig. 7, turning from a policy that prioritizes
and CCC respectively. Through this, we can gain in-
the high-deficient nodes (blue) to a policy prioritizing the
sights on the priority policies, as explained below. As-
low-deficient ones (red), the CCC increases and the CCP
sume a time-varying graph (simplified) of excess energy
decreases.
providers P ={p1 , p2 , p3 } with their individual excess
being Ee ={Ee1 , Ee2 , Ee3 }, deficient nodes C ={c1 , c2 , c3 } D. Affiliation networks
with their individual deficiency being Ed ={Ed1 , Ed2 , Ed3 },
Nodes sharing a similar characteristic can be grouped
and one controller, where the following conditions are met:
into a complete graph and become part of a larger network,
3
X called affiliation network [8], [7]. Two graphs co-exist
Ed1 > Ee3 > Ee2 > Ed2 > Ee1 > Ed3 and Ed1 > Eep . in an affiliation network: the first is a bipartite graph,
p=1
describing actions between pairs of nodes and the second is
Under a simple energy allocation strategy, we assume the
providers transfer energy in the order p1 , p2 , p3 , which is
arbitrary. One policy that prioritizes the highly deficient
nodes is depicted in Fig. 6a. Another policy aiming at
the low deficient ones is depicted in Fig. 6b. The black
links in Fig. 6 denote the connection with the controller
while the red links denote energy transfer from providers
to deficient nodes. The clustering coefficient results for the
case of Fig. 6 are summarized in Table I. We observe that
the policies prioritizing the highly-deficient nodes achieve
higher values of CCP and lower values of CCC compared
to the ones prioritizing the low-deficient nodes. This can be
explained by the differences in degree distribution, which Fig. 7: Comparison of weekly Clustering Coefficient of
can be observed in Fig. 3. We conclude that by prioritizing producers and consumers in the micro grid ecosystem, for
high-deficient nodes, the energy providers form complete different consumer priority policies.

764
Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on November 14,2022 at 15:03:41 UTC from IEEE Xplore. Restrictions apply.
(a) High Deficiency (b) Low Deficiency

Fig. 9: Provider unipartite graphs for Week 42 according


Fig. 8: The affiliation network created by the deficit nodes to different affiliation networks.
{c1, c2, c3}, which receive energy by the providers {a...h}.
as defined in Eq. (1). For a node vi ∈ P the following
cases are possible: (i) vi serves c1 → CCvi = 1, (ii) vi
a unipartite graph, grouping nodes that perform the same serves c2 → CCvi = 1, and (iii) vi serves c1 and c2 →
actions [14]. An ecosystem of IoT devices can be mod- CCvi ≤ 1. The first two cases are obvious due to the nature
eled using affiliation networks. A link between two nodes of affiliation networks [14]. For the third case, CCvi = 1
describes the resource flow from a provider to a receiver only if P f 1 = P f 2 = ∅, which is not the case. As it
(or communication with the controller), i.e., forming the is implied by the above, CCGC depends on Gc1 ∩ Gc2 .
bipartite graph. However, graph links can also be used to The higher the providers-intersection (i.e., the number of
link two providers to signify their affiliation in covering the common provider nodes sending excess energy both to c1
energy needs of the same node (or communicating with and c2 ), the higher the total clustering coefficient of the
the same controller), i.e., forming the unipartite graph. corresponding unipartite graph. This can be extended for
We recall that one deficient node can receive energy from a number of providers sharing their excess energy with a
multiple providers, especially when it is highly-deficient. number of deficient nodes. Thus, assuming a unipartite
The ‘bipartite → unipartite’ representation of an IoT graph GC = Gc1 ∪ Gc2 ∪ · · · ∪ Gcn , for P energy providers
device ecosystem with three energy-receiving nodes and sharing excess with N ={c1 , c2 , ..., cn } deficient nodes,
eight providers is depicted in Fig. 8. Assuming all the CCGC = 1 only if Gc1 ∩ Gc2 ∩ · · · ∩ Gcn = P . The value
energy providers cover the same deficient nodes, their of CCGC decreases with any decrease of the intersection
resulting unipartite graph would be complete, because result (i.e., different nodes receive excess energy).
CC = 1 [14]. To illustrate this, we consider the three We argue that the difference in the values of clustering
unipartite subgraphs of Fig. 8, namely Gc1 , Gc2 , and Gc3 . coefficient between the unipartite graphs of two priori-
These subgraphs are observed when focusing only on the tization strategies is analogous to the difference in the
bipartite subgraphs created by the energy providers that number of nodes that receive energy in each of the two
link with the same deficient node (c1, c2, or c3). However, strategies. This is observed in Fig. 9, where we present the
the unipartite graph that is created by the union of these unipartite graph created by the providers of our dataset
three subgraphs, GC = Gc1 ∪ Gc2 ∪ Gc3 , has a lower during a certain timespan (namely, Week 42). The lowest
clustering coefficient. This is because the graph is created value of clustering coefficient is seen in the unipartite
by groups of energy providers covering different deficient graph of Fig. 9b, where we observe the highest number of
nodes. Thus, they are affiliated for different reasons or low-degree nodes (light green), i.e., many deficient nodes
actions. are covered. The highest value of clustering coefficient is
Theorem 1 As the number of nodes that receive energy observed in the graph of affiliated energy providers sharing
increases in a bipartite graph (e.g., following a priority their excess with high deficient nodes (cf. Fig. 9a). We
policy focused on the low-deficient) the corresponding uni- observe in Fig. 10 that when the difference in the number
partite graph has lower clustering coefficient. of those who received energy is small between the policies
Proof Without loss of generality, consider an affiliation (e.g., Week 41), so is the difference in the clustering
network similar to the one of Fig. 8, but with two deficient coefficient results in the provider-unipartite graph.
nodes, c1 and c2 , served by a group P of excess energy
providers. Covering the energy needs of a certain node is VI. Discussion
an action (or service); a fraction of providers, P f 1 ∈ P , Ecosystems of IoT devices with resource sharing capa-
serves node c1 , another fraction, P f 2 ∈ P , serves node c2 , bilities can be modeled as time-varying graphs. Within this
and the rest, P − P f 1 − P f 2 , serve both c1 and c2 . The perspective, two important attributes of the ecosystem can
total clustering coefficient, CCGC , of the unipartite graph be derived, namely the strategy that is followed in terms
of the affiliation network, GC = Gc1 ∪ Gc2 , depends on of resource sharing and the balance between deficiency
the individual clustering coefficient of every provider node and excess of available resources. In the context of this

765
Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on November 14,2022 at 15:03:41 UTC from IEEE Xplore. Restrictions apply.
VII. Conclusions
Ecosystems of IoT devices are principal components for
interconnected societies. Actors/nodes of such ecosystems
coexisting in space and time can cooperate/link to build
ephemeral/time-varying relationships in order to achieve
a particular goal, e.g., sustainability. We took the example
of micro grids, where IoT devices measure the energy
consumption/generation of each household and the excess
energy is bartered among them. Our work reveals some
Fig. 10: Average clustering coefficient of the unipartite interesting attributes of the ecosystems of IoT devices
graphs (left). Number of deficient nodes that received when looked carefully through the lens of Graph Theory.
energy (right). We showed how to model such ecosystems as time-
varying graphs and affiliation networks. We then related
important graph metrics, such as, degree, clustering coef-
work, although we considered one type of network that has ficient to define the balance between excess/deficiency of
energy harvesting and sharing capability (micro grid), our energy inside these ecosystems and to infer the policy that
idea can be used in various other resource sharing ecosys- is used for energy sharing. The use of real-world data offers
tems, without connecting the term ‘resources’ strictly with additional validation for our theoretical constructs. We
the energy. In the use case of this paper, the energy sharing emphasize that this was only an example case of applying
strategy is revealed by the node degree distribution (cf. Graph Theory for energy sharing in ecosystems of IoT
Fig. 3) and the clustering coefficient (cf. Fig. 6, Fig. 7, devices. Our approach is generic and can be applied to
Fig. 10) with which we can see the groups of nodes that a multitude of resources and ecosystems.
are prioritized for receiving energy inside a network. The
energy balance is brought into the picture by our study on Acknowledgements
the existence of Giant Components, as connected graphs SCOTT http://www.scott-project.eu has received fund-
tend to present high ratios of excess energy over deficiency ing from the Electronic Component Systems for Euro-
pean Leadership Joint Undertaking, grant agreement No
(cf. Fig. 4, Fig. 5). Further, creating graphs of affiliated 737422. This joint undertaking is supported from the EU
energy providers allows to relate groups inside a network Horizon 2020 research and innovation program and Aus-
regarding sharing with common deficient nodes and, as tria, Spain, Finland, Ireland, Sweden, Germany, Poland,
a result, offers a broader view on the internal energy Portugal, Netherlands, Belgium, Norway.
dynamics of an ecosystem of IoT devices (cf. Fig. 8, Fig. 9). References
Hitherto, published literature on graph-theoretic ap- [1] F. Touati et al., “Feasibility of air quality monitoring systems
proaches for IoT networks utilize graph links for modeling based on environmental energy harvesting,” May 2015.
communication links between nodes, with the aim of data [2] M. Becker et al., “Logistic applications with wireless sensor
networks,” June 2010.
transfer [4], [9], [8]. We provided a new look at this [3] M. Hajiaghayi et al., “Efficient and practical resource block
topic, using graph links for modeling additionally the allocation for lte-based d2d network via graph coloring,” August
resource exchange among different nodes (e.g., IoT de- 2013.
[4] D. Zhang et al., “An energy-balanced routing method based
vices or whole households operating as energy “islands”). on forward-aware factor for wireless sensor networks,” February
Communication-based approaches consider mostly graphs 2014.
of static-in-time links, signifying an always-present com- [5] M. Garcia et al., “Saving energy and improving communications
using cooperative group-based wireless sensor networks,” Au-
munication link if two devices are in (wireless) range and gust 2011.
transmit power permits so [3], [10], [5]. We use time- [6] M. Cardei and D.-Z. Du, “Improving wireless sensor network
varying graphs where new links are created between nodes lifetime through power aware organization,” May 2005.
[7] H. Rowaihy et al., “Measurement and analysis of online social
at each time interval according to an agreed resource networks,” May 2007.
sharing policy among the actors of the ecosystem. This [8] H. Zheng et al., “Data gathering with compressive sensing in
approach allows us to study the ecosystem dynamics for wireless sensor networks: A random walk based approach,”
January 2015.
long periods and use historical observations to characterize [9] J. S. He et al., “Constructing load-balanced data aggregation
the distribution of energy within it. Furthermore, bipartite trees in probabilistic wireless sensor networks,” July 2014.
graphs are used in the literature to relate (monitoring) sen- [10] M. Younis et al., “Topology management techniques for tolerat-
ing node failures in wireless sensor networks: A survey,” January
sor devices (source nodes) to their corresponding groups of 2014.
target nodes [6], [7]. Here, source nodes and target nodes [11] O. Parson et al., “Dataport and nilmtk: A building data set
change dynamically, according to the energy balance of ev- designed for non-intrusive load monitoring,” December 2015.
[12] R. Kumar, J. Novak, and A. Tomkins, “Structure and evolution
ery node. Thus, the corresponding time-varying unipartite of online social networks,” August 2006.
graphs of affiliated energy providers (not considered in the [13] A. Mislove et al., “Measurement and analysis of online social
literature until now) are used to characterize the resource networks,” October 2007.
[14] M. E. J. Newman, D. J. Watts, and S. H. Strogatz, “Random
sharing inside an ecosystem of IoT devices. graph models of social networks,” February 2002.

766
Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on November 14,2022 at 15:03:41 UTC from IEEE Xplore. Restrictions apply.

You might also like