Received 4 October 2022, accepted 18 October 2022, date of publication 21 October 2022, date of current version 26 October 2022.

Digital Object Identifier 10.1109/ACCESS.2022.3216377

Carbon Footprint vs Energy Optimization

in IoT Network Deployments
FRANCISCO-JOSE ALVARADO-ALCON , RAFAEL ASOREY-CACHEDA , (Member, IEEE),
ANTONIO-JAVIER GARCIA-SANCHEZ , AND JOAN GARCIA-HARO , (Member, IEEE)
Department of Information and Communication Technologies, Universidad Politécnica de Cartagena, 30202 Cartagena, Spain
Corresponding author: Rafael Asorey-Cacheda ([email protected])
This work was supported in part by the MCIN/AEI/10.13039/501100011033 under Grant PID2020-116329GB-C22 and Grant
TED2021-129336B-I00, and in part by the Fundación Séneca, Región de Murcia through Atento under Grant 20889/PI/18.

ABSTRACT We are witnessing the full integration of the Internet of Things (IoT) into many social
and economic sectors. Part of this unprecedented growth is due to the emergence of new communication
technologies such as Low Power Wide Area Networks (LPWAN), which have been the catalyst for previously
unfeasible smart applications. Efforts to optimize energy consumption in these types of networks have been
necessary to extend their lifetime. However, not much attention has been paid to the study and optimization
of the carbon footprints (CF) of these network deployments. In general, it has always been understood that
minimizing energy consumption should also minimize the carbon footprint. In this work, the carbon footprint
of a generic IoT network that uses renewable energy sources and communicates via LoRa is explored,
and an optimization framework is proposed. We have found that minimizing energy consumption and the
carbon footprint are two different things. In fact, we show that it is not possible to minimize the carbon
footprint without greater energy consumption, and vice versa. This is due to the placement of gateways in the
network. Our findings could be extrapolated to other networks with similar topologies. These results suggest
that a fresh perspective on the optimization of IoT networks is needed to seriously consider environmental
sustainability criteria that has been ignored up to now.

INDEX TERMS Carbon footprint, LPWAN, IoT networks, optimization.

I. INTRODUCTION consumption [3], [4], [6] or the Human Toxicity Parame-

A wireless IoT network deployment consists of a large group ter [4], among others. In this context, suitably planning IoT
of autonomous low-cost devices (end devices or nodes) dis- networks before deployment is a major factor in determining
tributed throughout a given arbitrary area. End devices moni- the viability and operation of the final application. This is
tor one or several environmental variables and submit their extremely complex due to the many possible configurations
sample values wirelessly to one or more gateways (sink and scenarios that can be established depending on the target
nodes), which, in turn, deliver the information to network objectives. Moreover, IoT network deployments produce a
servers for further processing according to a specific appli- significant environmental footprint. However, there is scarce
cation (e.g., precision agriculture, home automation, smart literature about the sustainability of IoT and, more explicitly,
cities, e-health, Industry 4.0, etc.). There is a huge number the carbon emissions associated with IoT networks. This
of real examples that have proven the IoT utility [1], [2], paper analyzes the different elements that generate a carbon
[3]. This architecture has been studied extensively to evaluate footprint in IoT network deployment and proposes an opti-
general performance capabilities, such as economic cost [4], mization framework to minimize the environmental impact of
maximum lifetime [2], [5], [6], packet loss [7], energy IoT networks. Through this optimization problem, we empir-
ically demonstrate (somewhat counter-intuitively) that min-
The associate editor coordinating the review of this manuscript and imizing the carbon footprint is one thing, and minimizing
approving it for publication was Eyuphan Bulut . energy consumption is another. In fact, minimizing the
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 10, 2022 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 111297
 F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

carbon footprint requires increasing the energy consumption footprint. Conversely, the empirical evidence shows that
of IoT devices. enabling IoT devices to consume more energy has a pos-
The concept of a carbon footprint refers to the total itive effect on reducing the network’s carbon footprint.
emissions of carbon dioxide (and sometimes other green- The rest of the paper is organized as follows. Section II
house gases) emitted directly and indirectly by an activity or reviews some relevant works on IoT network optimization
product [8], [9]. Recently, this concept has been attracting and some studies measuring the carbon footprint of an IoT
significant attention from the scientific community, as the network. Section III introduces the linear model and problem
emission of carbon dioxide and other greenhouse gases is a formulation proposed, and Section IV discusses the numer-
standardized measurement of environmental impact. There- ical results obtained. Finally, in Section V, we conclude the
fore, it is a simple way to understand the effect of a deter- paper.
mined process on the environment.
Given that most IoT networks take samples from the real II. RELATED WORK
world and need to collect data from selected places [1], [2], In the past few years, IoT networks powered by photo-
[3], [5], end device (or node) location is unavoidably bound voltaic panels have been shown to have great advantages. The
to where sensing is needed. Furthermore, one of the most improvement in performance obtained by an optimal setting
popular ways of connecting IoT nodes is through LPWANs. and deployment of the network, from gateway location to
In principle, LPWAN deployments follow a star or connected- the throughput achieved, has also been proven. The authors
star topology. Consequently, the network needs gateways, of [2] reported that, theoretically, a network fed with photo-
or sink nodes, that gather the information sent by the end voltaic energy would work indefinitely until the components
devices and forward it through long-range technology to wore out. They also concluded that these networks would
data centers to process the acquired data. The number and perform better if their nodes consumed more energy. This
placement of these gateways will greatly impact the perfor- was also confirmed in the study conducted in [4], where the
mance of the network. Many specifications are available for authors used an Integer Linear Programming (ILP) approach
LPWAN networks, including LoRa, Zigbee, and SigFox. The to understand network behavior with varying objective func-
operating process of LoRa will be detailed in Section III. tions or problem conditions. They assumed that (i) the nodes
Although those devices consume very little power, the were powered by either one of the two models of battery
deployment of these networks is increasing substantially. The presented in the article and (ii) the environmental waste
Information and Communications Technology (ICT) industry came from the Li-ion batteries of the end devices after all
is estimated to be responsible for at least 2–3% of Global their recharge cycles were over. The experiment was tested
Greenhouse Gas (GHG) emissions, and this will continue to with a discrete number of gateway locations available and
increase in the future [10], [11]. Even though there is not a maximum of 400 end devices connected to each gateway.
enough data to estimate the part of these emissions that are They concluded that increasing the number of gateways could
produced by IoT, their impact should not be disregarded [11], decrease the chemical waste up to a certain threshold and
[12]. The importance of optimizing the carbon footprint of showed the existence of a trade-off between the environmen-
IoT networks is undeniable. For this reason, we propose an tal footprint, energy, and cost. They also concluded that the
optimization framework to place sink nodes or gateways in use of long-life batteries achieved better performance in terms
LPWANs to minimize carbon emissions. of reduced chemical waste (as the lifetime of the node, and
Renewable energies are usually suggested as green alterna- therefore the network, is increased) at the expense of increas-
tives to feed IoT devices, especially in large outdoor deploy- ing the deployment cost of the entire network. However, their
ments. However, there are also carbon emissions associated contribution was focused on the configuration of a deployed
with the use of these so-called greener technologies, from network, not on planning the optimal deployment beforehand.
production to the end-of-life of their components [13], [14]. The authors of [5] also proposed a novel optimization
Furthermore, renewable power supplies are irregular, mean- framework to maximize the lifetime of the network. Their
ing that a battery is also required to provide continuous oper- approach to the problem was to assign renewable energy
ation. The optimization under study could reduce the carbon sources to the nodes with greater energy consumption.
emissions of IoT networks, even in a network completely fed In addition, based on the results offered by the optimization
by photovoltaic panels, as we assume in our model. problem, they designed an algorithm that maximized the
In this context, the main contributions of this paper are life-time of the network while minimizing the number of
summarized below: hops. To this end, they envisioned a mesh multi-hop topology,
• An optimization framework to minimize the carbon meaning that a set of end devices worked as sink nodes and,
footprint in IoT network deployments using LPWAN as such, their location could not be changed. In contrast, our
technology and applying renewable energy sources. paper investigates the effects of free gateway placement.
• An analysis of the trade-off between the carbon footprint Others approaches to minimize the energy consumption of
and energy consumption in IoT networks. IoT networks include [3], where the authors study the effect
• The empirical demonstration that minimizing energy of reducing the traffic of a ZigBee IoT network. They con-
consumption significantly increases the carbon cluded that an estimated annual saving of up to 99% can be

111298 VOLUME 10, 2022

F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

achieved by removing seemingly crucial identifier fields from other studies also considered other GHGs [9], [13]. Our paper
packets and reducing the frequency of data polling interval. follows the first approach and considers only CO2 emissions,
It is clear that a wide range of approaches can be made to as the equivalency of other GHGs depends on the number
optimize a LPWAN. However, there are different perspectives of years used to calculate global-warming potential. After
that have not been studied in detail, and energy is usually GHGs have been converted to their equivalence in CO2 , the
selected as optimization function over CF. Consequently, carbon footprint is measured in CO2 eq. Given that this paper
in this paper the gateway placement is studied because there is only considers CO2 , both units will be equivalent.
few literature about it, and it can offer a significant reduction The emissions of producing IoT nodes are calculated
in environmental impact if the carbon footprint is considered in [12]. However, as we consider the number of end devices
while deploying the network. a constant and their CF is invariable, it can be disregarded
The carbon footprint of wireless networks has also been during the ILP. Authors in [13] estimated the carbon emis-
studied by authors in [15]. They estimated the carbon sions of photovoltaic panels and the percentage of emissions
emissions associated with the deployment and use of a coming from carbon dioxide. In the work presented in [14],
Fourth Generation Long Term Evolution (4G LTE) network the authors calculated power consumption during the pro-
in six different demographic areas. They concluded that the duction process of a 1 kW photovoltaic system. Therefore,
‘‘annual carbon emissions generated by the larger ICT net- they calculated the carbon footprint by using the emissions
works catering for high density urban and suburban areas of energy production. They considered both the efficiency of
and comparatively greater (up to three orders of magnitude) a photovoltaic grid connected power station and the degra-
than those produced by smaller networks,’’ and most of these dation of this efficiency over time at the rate of 2% annually.
emissions came from the manufacturing of mobile phones. Using these figures, they deduced the amount of energy that a
Additionally, they discovered a linear correlation between 1 kW photovoltaic panel exposed to the effective illumination
annual carbon footprints and number of subscribers. How- time of 3000 h could produce. Finally, authors in [23] stated
ever, this link was not maintained with small ICT systems due that crystalline-silicon solar panels ‘‘dominate 80% of the
to the network’s less efficient operation. Consequently, their market globally,’’ and the authors in [24] and [25] concluded
conclusions cannot by applied to our work, and the findings that there was still not enough information to calculate the
of our work may be untrue for very large networks. Should average emissions from the disposal of photovoltaic panels,
a LPWAN with tens of thousands of nodes be considered, as there are many options for disposal and few panels have
new investigations would be necessary to understand the reached their end-of-life.
behaviour of its CF. Finally, regarding batteries, in [25], the authors assessed
There is vast literature addressing the consumption of IoT the life cycle of batteries. They obtained the carbon footprint
end devices and their energy models. Authors in [6], [16], for each of the phases of lithium iron phosphate batteries:
[17], [18], and [19] broke down the functions of an end raw materials, production, and use. And, according to [27]
device, determining the electric current needed to accom- and [28], lithium-ion cobalt based batteries offer 500 recharge
plish every task. Data from manufacturers is also avail- cycles.
able online [20]. The three major functions are transmitting,
receiving, and the idle state. In comparison, all the other tasks, III. MODEL AND FORMULATION OF THE PROBLEM
from sensing to computing, can be considered negligible. The network model to develop the optimization problem
On the other hand, gateways have greater energy demands is based on LPWAN topologies. For this purpose, the
as they must be continuously listening according to the LoRaWAN specification, one of the most popular LPWAN
long-range technology requirements (nodes are assumed technologies today, will be taken as a reference. LoRaWAN
to only transmit, without listening requirements). In [21], is a wireless communication technology based on a one-hop
the authors studied the power consumption of front-end radio system and designed to achieve long ranges while con-
LoRaWAN gateways, including backhaul wireless technol- suming little power. A LoRaWAN network uses a star-of-
ogy. They provided a simplified model for gateways, assum- stars topology consisting of three basic elements, as shown
ing the device is always-on with a fixed number of channels. in Figure 1: end devices, gateways, and a central network
In this case, power consumption depended on two factors: the server. End devices, which may be sensors or actuators, com-
LoRaWAN gateway vendor and the backhaul technology. municate with the network server through the backhaul of
Concerning the carbon emissions of gateways and end the gateways. End devices use the LoRa physical layer to
devices, we shall start by providing a formal definition of exchange messages with the gateway, whereas the gateway
the term carbon footprint since its definition seems to be a and the network server communicate using an IP-based pro-
point of debate among the scientific community. The work tocol stack.
presented in [8] defined a carbon footprint as ‘‘a measure of LoRaWAN comprises three communication classes:
the exclusive total amount of carbon dioxide emissions that is Class A, Class B, and Class C. Class A, also known as
directly and indirectly caused by an activity or is accumulated basic LoRaWAN, schedules the transmissions based on the
over the life stages of a product.’’ This approach has been behavior of the end device. In this operating mode, downlink
followed by some studies, such as the work in [22], while transmissions (i.e., from the network server to the end device)

VOLUME 10, 2022 111299

 F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

TABLE 1. Spreading factor and related configuration for LoRa devices in

Europe in the 868 MHz band.

between them is less than or equal to the transmission

range. Transmission power is fixed at 14 dBm (the maxi-
FIGURE 1. LoRaWAN basic architecture. mum allowed by the LoRa specification in Europe in the
frequency band of 868 MHz [20]), but the spreading factor
can be reduced to minimize energy consumption, or it can be
can only occur after an uplink transmission (i.e., from the end increased to extend transmission range. The spreading factor
device to the network server) has taken place. Since this class is a LoRa parameter that determines the bit rate. A spread-
offers the lowest energy consumption, all the end devices will ing factor of 7 (SF7) offers the highest data speed rate
work under these characteristics. of 5470 bit/s, whereas a spreading factor of 12 (SF12) only
Class B introduces additional functions. It supports addi- transmits at 250 bit/s. However, lower physical bit rates are
tional downlink transmission opportunities at prescheduled associated with increased sensitivity in the receiving antenna.
times. Class C allows downlink transmissions at any time, This information is outlined in Table 1.
except when the end device is transmitting. A Class C device Receiver sensitivity corresponds to the spreading factor,
consumes more power but is ideal for the gateways since they ranging from −124 dBm for a spreading factor of 7 to
are always listening. −137 dBm for a spreading factor of 12. Finally, without
Thus, based on the LoRaWAN topology specification, loss of generality, we consider the energy consumption of
we develop a linear programming formulation for the min- non-radio components, such as the sensing of end devices,
imization of the IoT network carbon footprint. Basically, as negligible.
we consider the IoT network as a set, N , of n connected nodes All devices (both end devices and gateways) are powered
randomly and uniformly distributed over a square surface. using a photovoltaic panel and a battery with enough capacity
We assume an end device density of d, meaning the area to operate the device for 24 hours. The photovoltaic panel is
of the experiment is n/d. We also assume that all the nodes big enough to power the device all year long given the number
transmit directly to one of the network gateways (or, in other of hours of solar incidence a year. With the combination
words, there must always be a gateway within the cover- of battery and panel, the device can work uninterruptedly
age area of any end device), which is always listening and 24 hours a day. For the sake of simplicity, we do not con-
backhauling. sider periods of time without sunlight longer than a night or
Regarding the gateways, both their total amount and place- seasonal daylight differences.
ment are parameters to be optimized. To this end, we cre- The gateway consumption model is based on [21], which
ate a mesh grid with a finite number of points (a set G of provides power consumption as a constant, regardless of
p points) where a gateway can be located and define a variable the number of end devices connected to it. This study is
vector G, whose size matches the number of possible gateway based on the assumption of pure aloha-based channel access,
location points. One of the two optimization variables is the which only considers 3 and 6 LoRa channel configurations
logical vector G, containing elements with a value of one if of 125 kHz each. We have selected a Lorrier LR2 gateway
there is a gateway situated in that point, or zero if the point is with backhaul implemented with LTE (worst-case scenario
empty. The sum of the elements of G indicates the number of out of those covered in [21]). This energy consumption will
gateways in the IoT network under consideration. The size be referred to as EG .
of G, that is, the number of points that make up the grid, We have calculated the carbon footprint as the sum of
is p. Although the number of potential gateway locations is the carbon emissions generated during the manufacturing
limited, a custom algorithm has been designed to determine of the photovoltaic panel and the battery for both the end
the optimal area for gateway location. More details on this devices and the gateways. Solar photovoltaic panel end-of-
issue are explained in Section III-B2. life management is an evolving field that requires further
The second variable of optimization is L, a logical matrix research and development, as mentioned in [24] and [25].
where Li,j takes the value of 1 if a connection is estab- As such, and given that the life expectancy of photovoltaic
lished between end device i-th and the gateway located at panels is about 25 years [13], [14], [23], [24], [28], we have
point j-th. The link can only be established if the distance disregarded the end-of-life emissions.

111300 VOLUME 10, 2022

F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

TABLE 2. Problem parameters. The second step is to model the energy consumption
of the end devices. Thus, Ei,j represents the link energy
consumption between end device i-th and point j-th for one
year, measured in kW h, while Li,j represents whether a link
between end device i-th and point j-th exists (recall that
a point j is a potential location for a gateway). Moreover,
Gj denotes whether there is a gateway located at point j.
From these considerations, it is easy to obtain the expressions
for the energy consumption of an end device and a gateway,
expressions (3) and (4), respectively.
p
n X
X
Li,j Ei,j (3)
i=1 j=1
Xp
Gj EG (4)
j=1

Combining expressions (1) to (4), the carbon footprint, CT ,

of the network can be obtained as formulated in (5). This
function expresses the carbon emissions in kgs for one year
(CO2 eq/year), and it is the optimization function of the
problem.
 
  X n X p p
Cp Cb  X
CT = + Li,j Ei,j + Gj EG 
Tp Ti ηp R
i=1 j=1 j=1
(5)
Then, the problem of gateway and link assignment is
equivalent to that of minimizing the carbon footprint of the
network, as follows:
minimize CT (6a)
p
A. LINEAR PROGRAMMING FORMULATION X
subject to: Li,j Gj = 1, ∀i ∈ N , (6b)
The first step is to model the total carbon emissions of a j=1
generic device (end device or gateway). Thus, expression (1)
allows us to obtain Kp , the carbon footprint of a photovoltaic Li,j ∈ {0, 1}, ∀i ∈ N , ∀j ∈ G, (6c)
panel for every kW h it produces during a year: Gj ∈ {0, 1}, ∀j ∈ G. (6d)
Cp Table 2 defines the parameters employed for the program-
Kp = (1)
Tp Ti ηp ming formulation. Expressions (6b) to (6d) are the problem
constraints for the carbon footprint problem:
where Cp represents the CF generated where producing a
• Expression (6b) ensures that every end device is con-
photovoltaic panel depending on its power. Tp and Ti repre-
nected to a gateway.
sent, respectively, the lifetime of the panel and the number
• Expressions (6c) and (6d) model that problem variables
of hours of solar incidence, and ηp is the efficiency of the
(Li,j and Gj ) can take zero or one logical value.
panel.
Note that expression (6b) is a non-linear constraint, as it
Expression (2) represents the carbon footprint of 365 rec-
multiplies two optimization variables. However, the problem
harge cycles of a battery that provides a total of 1 kW h for one
can be expressed as a mixed-integer linear programming
year, Kb . Cb represents the CF generated where producing
(MILP) optimization. To this end, expression (6b) must be
a battery depending on its capacity, and R its number of
rewritten. Consequently, it is subdivided into two different
recharge cycles.
constraints, namely expressions (7b) and (7c). The linear
Cb problem formulation is thus expressed as follows.
Kb = (2)
R
minimize CT (7a)
Then, the total emissions of a generic device with an annual p
X
consumption of 1 kW h would be Kp + Kb . Note that all the subject to: Li,j = 1, ∀i ∈ N , (7b)
equation parameters are defined in Table 2 j=1

VOLUME 10, 2022 111301

 F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

TABLE 3. Different settings under which the experiment has been done. changing the values selected would appropriately change the
CF of the network. However, variations in EG could result
in changes in the network deployment, and consequently the
network emissions would change unpredictably.
To solve the problem, the following considerations should
be made. The most important one is that the problem is
NP hard and if the number of potential locations for the gate-
ways is too wide (an overly dense grid of points), it would take
TABLE 4. Problem parameter values. too long to solve the problem. On the other hand, defining a
grid with fewer points would result in having less precision
than desired. Therefore, we propose a series of simplifying
steps to reduce the time necessary to solve the linear integer
problem. These steps are shown in Sections III-B1 and III-B2.

1) REDUNDANT POINTS
Let i and j be two points of the grid, where both are poten-
tial locations for a gateway. Then, if for each given node b
contained in N , ∀b ∈ N , it is true that either Eb,i < Eb,j or
n
X Eb,i = Eb,j = ∞. j would never be selected as a location for
Li,j < α Gj , ∀j ∈ G, (7c) a gateway (mathematically, it means that Gj = 0 in vector G)
i=1 because it would always be better to place it in i: for every
Li,j ∈ {0, 1}, ∀i ∈ N , ∀j ∈ G, (7d) node in the grid, either none of them can offer coverage,
Gj ∈ {0, 1}, ∀j ∈ G. (7e) or i is closer, therefore consuming less energy.
Narrowing down the size of the number of points where a
Expression (7b) ensures that every end device is connected gateway could be placed, p, by using this criterion has shown
to a potential gateway location point. Additionally, expres- to reduce the computational burden of the problem, mainly
sion (7c) represents that every link to a potential gateway by eliminating check points located on the borders of our
location point is made to a point where there is a gateway working area. Consequently, the processing time needed to
(meaning there is no connection to an empty point), and no solve the problem is shorter. An example is shown in Figure 2,
gateway receives more than α connections. Both expressions where the number of points has been reduced from 324 to 163.
together guarantee that every node is connected to a gateway
and that no gateway receives more than α connections. Given 2) REFINING THE GRID
that α is a very large number, the second condition is irrele- Even after removing redundant points, as explained above,
vant. it has been proven that problems with more than 60 nodes
Ultimately, the combination of both restrictions achieves are still hard to solve, as they require large amounts of com-
the same behavior as expression (6b): guaranteeing that every putational/memory resources and time. Therefore, we have
node is connected to a gateway. However, using expres- resorted to an additional simplifying strategy based on solv-
sion (6b) as a constraint would make it a non-linear optimiza- ing the optimization problem iteratively. In the first itera-
tion problem. tion, the starting point is a low density of potential gateway
locations that will increase in successive iterations. As a
B. EXPERIMENT DESIGN result of each iteration, areas to place the gateways will be
We have planned 40 different settings for the experiment by discarded and we will check whether the carbon footprint
varying the density of the nodes and the total area of the prob- is lower than in the previous iteration. The iterative process
lem. We selected 1 node every 500 m2 , 1000 m2 , 2000 m2 , will stop when the carbon footprint does not significantly
3000 m2 , and 4000 m2 as the possible end device densities, improve. Below, we provide more details about this iterative
and 6 × 104 m2 , 1.2 × 105 m2 , 1.8 × 105 m2 , 2.4 × 105 m2 , process.
3 × 105 m2 , 3.6 × 105 m2 , 4.2 × 105 m2 , and 4.8 × 105 m2 As previously indicated, we first solve the problem for
as the values for the area under evaluation, as shown in a considerable distance between points (e.g., 60 m). This
Table 3. We have studied every possible combination of these problem contains fewer points and, as such, is easier to solve.
two factors by running every scenario with 20 different seeds. Therefore, it produces an approximate location for every
The values given to the problem parameters, and the refer- gateway in the ideal network. Then, we start iterating around
ences from where they were taken, are expressed in Table 4. these points to refine the outcome. In this way, we calculate
Given that equations (2) and (3) model the energy source, the exact solution faster.
and every device is powered by the same technology, most To do so, we calculate the cost of the objective/
of the parameters included in Table 4 are directly or inversely optimization function of the network for every node when we
proportional to the total carbon footprint of the LPWAN, and move a gateway to an adjacent position. If the cost increases,

111302 VOLUME 10, 2022

F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

FIGURE 2. Removing of redundant points for an example with 16 end nodes and 324 points where a gateway could be located. In the second image,
161 points have been eliminated according to the criteria in Section III-B1. Circles represent the possible positions and crosses mark the placement of
each end node.

we are moving away from the optimal location. If the cost nodes and the area of network deployment. Figure 4 shows
remains the same, we have found a target area where the the emissions obtained for every optimization scenario tested
gateway could be placed. We run over this target location after the algorithm minimized the carbon footprint.
in all the four possible directions, calculating a squared area Increasing the density of end devices in the network
where the gateway could be placed. Then, we reduce the barely increases emissions, and thus, the carbon footprint.
distance between points to half and place the points only For instance, increasing the number of end nodes 4 times,
around the target area. Next, we eliminate the redundant from 1 for every 4000 m2 to 1 for every 1000 m2 , the carbon
points, as explained in Section III-B1, and we solve the MILP footprint increases by only 29.6% in the worst-case scenario,
again. We repeat these iterations until the cost of the objective i.e., an 300% increase in the number of nodes entails an
function is not reduced any more, as shown in Figure 3. increase in the CF of one order of magnitude lower. How-
‘‘Refining the grid’’ is an experience-based simplify- ever, enlarging the area covered by the network increases the
ing approach, thereby allowing us to solve time-consuming carbon footprint at a higher rate.
device-placing problems faster. For instance, in the example In order to make perceptible the reduction in carbon emis-
shown in Table 5, we can observe that the problem has been sions that can be achieved, we have calculated the CF for the
solved in only 4.67 s, on average, when the original distance same scenarios without optimizing the gateway placement.
was 60 m and ‘‘refining the grid’’ was applied. Using a fixed Instead, gateways have been reasonably placed forming a grid
distance between points has proven to take longer. Moreover, that guarantees that every point in the network has coverage,
if ‘‘refining the grid’’ is not used, setting a large distance and every end device has been connected to the closest gate-
between potential points leads to less precise solutions with way. Results are shown in Figure 5. Without optimization, the
higher carbon emissions, while placing the points closer CF grows rapidly for certain scenarios where the number of
means a more complex problem. gateways is significantly higher than needed.
In classical problems, one common optimization goal
IV. NUMERIC RESULTS has been to minimize energy consumption. Thus, to better
This section discusses our results and provides some exam- understand our model and assess the relevance of the carbon
ples to show the benefits of the optimization process. footprint optimization, we also formulate the energy opti-
There are two factors that have been considered when mization framework. Specifically, we apply the same con-
analyzing the numerical results: the density of the end device strains while minimizing the amount of energy consumed by

VOLUME 10, 2022 111303

 F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

FIGURE 3. Step by step carbon footprint emissions optimization example results for a scenario with 60 end device nodes. Note that since the fourth
iteration does not offer an improvement over the third iteration, we assume that the optimal cost has been found.

the end devices: concluded that energy optimization leads to greater waste
j=p
i=n X
since ‘‘optimizing each variable leads to different network
X configurations.’’ As will be shown below, optimizing energy
Li,j Ei,j (8)
consumption leads to a larger number of gateways placed
i=1 j=1
closer to the end devices.
Expression (8) denotes the optimization function. This As observed in Figure 6, we found that when mini-
allows us to confirm the findings in [4], whose authors mizing energy consumption by increasing the number of

111304 VOLUME 10, 2022

F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

TABLE 5. Comparison of the required time to complete optimization according to the experiment settings and whether the ‘‘refining the grid’’ technique
is used.

FIGURE 4. Carbon footprint optimization.

FIGURE 6. Carbon footprint obtained by optimizing (top) carbon

emissions or (bottom) energy consumption. Each plot represents the
average result for each of the 40 different scenarios tested.

FIGURE 5. Full coverage carbon footprint (gateways deployed in a grid

layout).
per 1000 m2 , the carbon footprint is up to 4 times bigger if
we try to minimize energy consumption instead of carbon
emissions.
nodes 4 times, from 1 for every 4000 m2 to 1 for Another way of measuring emissions is by calculating the
every 1000 m2 , the carbon footprint increases 36.55% in the carbon footprint per node or end device, as represented in
worst-case scenario, slightly higher than in the previous case Figure 7. The fact that the carbon footprint remains constant
(29.6%). The larger coverage area in this case causes the for a given node density establishes that an increase in surface
carbon footprint to increase faster. So, Figure 6 shows that intensifies carbon emissions at the rate that the number of
for the largest studied scenario with a density of only 1 node nodes grows.

VOLUME 10, 2022 111305

 F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

FIGURE 7. Carbon footprint per end device obtained by optimizing (top)

carbon emissions or (bottom) energy consumption. Each plot represents FIGURE 8. Carbon footprint density obtained by optimizing (top) carbon
the average result for each of the 40 different scenarios tested. Notice emissions or (bottom) energy consumption. Each plot represents the
that the x axis has been inverted to better observe the plot. average result for each of the 40 different scenarios tested.

maximum energy that can be consumed by end devices. This

In Figure 7, we can see that increasing the surface constraint is expressed as follows:
covered by the network does not increase the carbon emis-
p
n X
sions per node, but it acts like a constant: in the worst- X
Li,j Ei,j < EM , (9)
case scenario, the carbon footprint only decreases 37.61%
i=1 j=1
from the maximum value for a given density of end devices
while minimizing the carbon footprint, or 19.39% while where EM is the maximum amount of energy that can be
minimizing energy consumption. Alternatively, measuring consumed
the carbon footprint per square meter (as illustrated in Figure 10 shows the results for two test scenarios with
Figure 8) reveals that an increase in the surface covered 240 nodes. These results show that it is not possible to
by the IoT network leads to a decrease in emissions per minimize the carbon footprint without allowing an increase
surface. in minimum energy consumption. This does not mean that
Therefore, we have validated the fact that minimizing carbon footprint minimization and energy minimization are
energy consumption leads to a bigger carbon footprint. On the opposites. In fact, the carbon footprint partially depends on
other hand, optimizing carbon emissions results in greater energy consumption. So, reducing energy consumption only
energy consumption, as shown in Figure 9. reduces the carbon footprint to a certain extent. From this
One of the main outcomes of this paper is that when point on, it is no longer possible to further reduce the carbon
using optimal deployments, there is a trade-off between the footprint without increasing energy consumption. What hap-
carbon footprint and energy consumption. This means that pens in this case is that to further reduce the carbon footprint,
minimizing one of the two parameters will increase the other. it is necessary to eliminate network equipment (gateways)
Thus, we have studied the same ILP adding an additional and force the nodes to transmit at greater distances, increasing
restriction: minimizing the carbon footprint while limiting the energy consumption to a lesser extent than the reduction in the

111306 VOLUME 10, 2022

F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

FIGURE 11. Average number of gateways placed when optimizing (top)

FIGURE 9. Energy consumed by the end devices, obtained by optimizing carbon emissions or (bottom) energy consumption. Each plot represents
(top) carbon emissions or (bottom) energy consumption. Each plot the average result for each of the 40 different scenarios tested.
represents the average result for each of the 40 different scenarios tested.

surface covered by the network or the density of the end

devices. These variances in the carbon footprint and energy
consumption are caused by changes in the network configura-
tion. The first and most obvious change affects the number of
gateways placed in the covered area. The average number of
gateways required to achieve optimization in each experiment
is represented in Figure 11.
As can be observed, minimizing energy consumption
instead of carbon emissions in the IoT network means deploy-
ing many more gateways at their corresponding places.
Accordingly, other key parameters in the LPWAN network
have noticeably decreased, like the total time that the network
is transmitting.
FIGURE 10. Trade-off between the optimal carbon footprint and energy
To represent the total transmission time of the network,
for an IoT network of 240 nodes. we will use Time-on-Air (ToA). This parameter represents
the percentage or time that the end devices are transmitting
carbon footprint achieved. This can be verified in Figure 10, data, as shown in Figure 12. Since there may be several gate-
where a minimal reduction in energy consumption allows ways working in different channels and different spreading
significant reductions in the carbon footprint. factors, the transmission time can exceed 100%. As previ-
The general behavior of energy consumption is similar ously explained, minimizing carbon emissions leads to fewer
to the carbon footprint, growing if we increase either the gateways, but also to an increased distance between end nodes

VOLUME 10, 2022 111307

 F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

is the most sustainable option from the environmental point

of view. However, the reduction in energy consumption not
only produces an increase in the carbon footprint, but this
relationship is non-linear. As a general conclusion, allowing
end device nodes to consume more energy in an IoT network
can contribute to reducing its carbon footprint.
As future work, we are focusing on cases where a portion
of the end nodes are allowed to use third end nodes as relays
instead of having to be directly in the reception range of a
gateway. Even a small portion might lead to a substantial
reduction in the carbon emissions of the network. More-
over, the MILP problem could be turned into a very reliable
instrument by modifying it to consider other effects, such as
3D terrain or a more complex model to characterize gateway
consumption.

REFERENCES
[1] R. K. Kodali, K. Y. Borra, and H. J. Domma, ‘‘An IoT based smart parking
system using Lora,’’ in Proc. Int. Conf. Cyber-Enabled Distrib. Comput.
Knowl. Discovery (CyberC), Oct. 2018, pp. 151–1513.
[2] H. Sharma, A. Haque, and Z. A. Jaffery, ‘‘Maximization of wireless
sensor network lifetime using solar energy harvesting for smart agriculture
monitoring,’’ Ad Hoc Netw., vol. 94, Nov. 2019, Art. no. 101966.
[3] A. Mukherjee, S. Misra, N. S. Raghuwanshi, and S. Mitra, ‘‘Blind entity
identification for agricultural IoT deployments,’’ IEEE Internet Things J.,
vol. 6, no. 2, pp. 3156–3163, Apr. 2019.
[4] M. Rady, J. Georges, and F. Lepage, ‘‘Can energy optimization lead to
economic and environmental waste in LPWAN architectures?’’ ETRI J.,
vol. 43, no. 2, pp. 173–183, Apr. 2021.
[5] R. Asorey-Cacheda, A. García-Sánchez, F. García-Sánchez,
J. García-Haro, and F. González-Castano, ‘‘On maximizing the lifetime
of wireless sensor networks by optimally assigning energy supplies,’’
Sensors, vol. 13, no. 8, pp. 10219–10244, Aug. 2013.
[6] A. Khalifeh, Z. Sakkijha, H. Qaderi, O. Ghatasheh, A. Issa, A. Shatat,
FIGURE 12. Time-on-Air measured in seconds of transmission per natural
second when optimizing (top) carbon emissions or (bottom) energy H. Jaber, R. Alwardat, and S. Alhaj-Ali, ‘‘An energy efficient WSN
consumption. Each graph represents the average result for each of the implementation for monitoring and critical event detection,’’ in Proc. 2nd
40 different scenarios tested. IEEE Middle East North Afr. Commun. Conf. (MENACOMM), Nov. 2019,
pp. 1–6.
[7] L. Lei, Y. Kuang, X. S. Shen, K. Yang, J. Qiao, and Z. Zhong, ‘‘Optimal
and gateways, thus forcing the use of a lower transmission bit reliability in energy harvesting industrial wireless sensor networks,’’ IEEE
rate. This means the spreading factor employed is increased Trans. Wireless Commun., vol. 15, no. 8, pp. 5399–5413, Aug. 2016.
and, as a consequence, the ToA is also increased. [8] T. Wiedmann and J. Minx, ‘‘A definition of ’carbon footprint,’’’ in Eco-
logical Economics Research Trends, vol. 1. Commack, NY, USA: Nova,
2008, pp. 1–11.
V. CONCLUSION [9] M. Lenzen, Y.-Y. Sun, F. Faturay, Y.-P. Ting, A. Geschke, and A. Malik,
The optimal planning of IoT networks is a complex problem ‘‘The carbon footprint of global tourism,’’ Nature Climate Change, vol. 8,
no. 6, pp. 522–528, Jun. 2018.
due to the openness of their deployment and final settings [10] A. Abdelli, L. Mokdad, J. Ben Othman, and Y. Hammal, ‘‘Dealing with
to fulfill the goals of a particular IoT application. In this a non green behaviour in WSN,’’ Simul. Model. Pract. Theory, vol. 84,
paper, we justify and develop an optimization framework to pp. 124–142, May 2018.
minimize the carbon footprint in IoT network deployments. [11] C. Freitag, M. Berners-Lee, K. Widdicks, B. Knowles, G. Blair, and
A. Friday, ‘‘The climate impact of ICT: A review of estimates, trends and
The impact on the environment of these types of networks by regulations,’’ 2021, arXiv:2102.02622.
optimizing their carbon footprint when LPWAN technology [12] T. Pirson and D. Bol, ‘‘Assessing the embodied carbon footprint of IoT
and renewable energy sources are in use is also explored. edge devices with a bottom-up life-cycle approach,’’ J. Cleaner Prod.,
vol. 322, Nov. 2021, Art. no. 128966.
Interestingly, we have found and confirmed that in an IoT [13] Y. Fu, X. Liu, and Z. Yuan, ‘‘Life-cycle assessment of multi-crystalline
network, energy optimization does not equal carbon footprint photovoltaic (PV) systems in China,’’ J. Cleaner Prod., vol. 86,
optimization. In fact, energy minimization has proven to incur pp. 180–190, Jan. 2015.
[14] X. Guo, K. Lin, H. Huang, and Y. Li, ‘‘Carbon footprint of the photovoltaic
in significantly higher carbon emissions. Therefore, we can
power supply chain in China,’’ J. Cleaner Prod., vol. 233, pp. 626–633,
accomplish a more sustainable IoT network deployment if the Oct. 2019.
network planning takes gateway placement into account to [15] D. Ruiz, G. San Miguel, J. Rojo, J. G. Teriús-Padrón, E. Gaeta,
minimize carbon emissions. M. T. Arredondo, J. F. Hernández, and J. Pérez, ‘‘Life cycle inventory
and carbon footprint assessment of wireless ICT networks for six demo-
This result is significant since the classical literature has graphic areas,’’ Resour., Conservation Recycling, vol. 176, Jan. 2022,
always assumed that the optimization of energy consumption Art. no. 105951.

111308 VOLUME 10, 2022

F.-J. Alvarado-Alcon et al.: Carbon Footprint vs Energy Optimization in IoT Network Deployments

[16] B. Martinez, M. Montón, I. Vilajosana, and J. D. Prades, ‘‘The power of RAFAEL ASOREY-CACHEDA (Member, IEEE)
models: Modeling power consumption for IoT devices,’’ IEEE Sensors J., received the M.Sc. degree in telecommunication
vol. 15, no. 10, pp. 5777–5789, Oct. 2015. engineering (major in telematics) and the Ph.D.
[17] Q. Wang, M. Hempstead, and W. Yang, ‘‘A realistic power consumption degree (cum laude) in telecommunication engi-
model for wireless sensor network devices,’’ in Proc. 3rd Annu. IEEE neering from the Universidade de Vigo, Spain, in
Commun. Soc. Sensor Ad Hoc Commun. Netw. (SECON), vol. 1, Sep. 2006, 2006 and 2009, respectively. He was a Researcher
pp. 286–295. with the Information Technologies Group, Univer-
[18] T. Bouguera, J.-F. Diouris, J.-J. Chaillout, R. Jaouadi, and G. Andrieux,
sity of Vigo, Spain, until 2009. From 2008 and
‘‘Energy consumption model for sensor nodes based on LoRa
2009, he was also a Research and Development
and LoRaWAN,’’ Sensors, vol. 18, no. 7, p. 2104, Jun. 2018, doi:
10.3390/s18072104. Manager at Optare Solutions, a Spanish Telecom-
[19] L. Casals, B. Mir, R. Vidal, and C. Gomez, ‘‘Modeling the energy perfor- munications Company. From 2009 to 2012, he held an Ángeles Alvariño
mance of LoRaWAN,’’ Sensors, vol. 17, no. 10, p. 2364, 2017. position at Xunta de Galicia, Spain. From 2012 to 2018, he was an Associate
[20] (2019). MultiTech mDot Long Range LoRa Module’s Datasheet. Professor at the Centro Universitario de la Defensa en la Escuela Naval
Accessed: Apr. 18, 2022. [Online]. Available: https://www.multitech.com/ Militar, Universidade de Vigo. He is currently an Associate Professor at the
documents/publications/data-sheets/86002171.pdf Universidad Politécnica de Cartagena, Spain. He is the author or coauthor
[21] H. H. R. Sherazi, G. Piro, L. A. Grieco, and G. Boggia, ‘‘When renewable of more than 60 journals and conference papers, mainly in the fields of
energy meets LoRa: A feasibility analysis on cable-less deployments,’’ switching, wireless networking, and content distribution. He was a Visiting
IEEE Internet Things J., vol. 5, no. 6, pp. 5097–5108, Dec. 2018. Scholar at New Mexico State University, USA, from 2007 to 2011, and
[22] C. Stoll, L. Klaaßen, and U. Gallersdörfer, ‘‘The carbon footprint of Universidad Politécnica de Cartagena, from 2011 to 2015. His research
bitcoin,’’ Joule, vol. 3, no. 7, pp. 1647–1661, 2019. interests include content distribution, high-performance switching, peer-to-
[23] M. S. Chowdhury, K. S. Rahman, T. Chowdhury, N. Nuthammachot, peer networking, wireless networks, and nano-networks. He received the
K. Techato, M. Akhtaruzzaman, S. K. Tiong, K. Sopian, and N. Amin, Best Master Thesis Award for his M.Sc. degree and also received the Best
‘‘An overview of solar photovoltaic panels’ end-of-life material recycling,’’
Ph.D. Thesis Award for his Ph.D. degree.
Energy Strategy Rev., vol. 27, Jan. 2020, Art. no. 100431.
[24] Y. Xu, J. Li, Q. Tan, A. L. Peters, and C. Yang, ‘‘Global status of recycling
waste solar panels: A review,’’ Waste Manage., vol. 75, pp. 450–458, ANTONIO-JAVIER GARCIA-SANCHEZ received
May 2018. the M.S. degree industrial engineering from the
[25] Y. Liang, J. Su, B. Xi, Y. Yu, D. Ji, Y. Sun, C. Cui, and J. Zhu, ‘‘Life cycle Universidad Politécnica de Cartagena (UPCT),
assessment of lithium-ion batteries for greenhouse gas emissions,’’ Resour., Spain, in 2000, and the Ph.D. degree from the
Conservation Recycling, vol. 117, pp. 285–293, Feb. 2017. Department of Information Technologies and
[26] J.-M. Durand, M. João-Duarte, and P. Clerens, ‘‘Joint EASE/EERA rec- Communications (DTIC), UPCT, in 2005. Since
ommendations for a European energy storage technology development 2001, he has been with the Department of Informa-
roadmap towards 2030,’’ in Proc. Joint Eur. Assoc. Storage Energy (EASE),
tion Technologies and Communications (DTIC),
2013.
[27] D. Deng, ‘‘Li-ion batteries: Basics, progress, and challenges,’’ Energy Sci.
UPCT, where he is currently a Full Professor, and
Eng., vol. 3, no. 5, pp. 385–418, 2015. the Head of the DTIC. He is the (co)author of more
[28] W.-H. Huang, W. J. Shin, L. Wang, W.-C. Sun, and M. Tao, ‘‘Strategy and than 100 conference and journal papers, 52 of them indexed in the Journal
technology to recycle wafer-silicon solar modules,’’ Sol. Energy, vol. 144, Citation Report (JCR). He has been the head of several research projects in
pp. 22–31, Mar. 2017. the field of communication networks and optimization, and he is currently a
reviewer of several journals listed in the ISI-JCR. He is also an inventor/the
co-inventor of 12 patents or utility models, and he has been a TPC member
or the chair in about 40 international congresses or workshops. He has
been a Visiting Scholar at Bologna University, Italy, in 2007, Wageningen
University, The Netherlands, in 2012, and Santiago de Cali University,
Colombia, in 2019. His research interests include the areas of wireless sensor
networks (WSNs), streaming services, artificial intelligence, the IoT, and
nanonetworks.

JOAN GARCIA-HARO (Member, IEEE) received

the M.S. and Ph.D. degrees in telecommunica-
tion engineering from the Universitat Politècnica
de Catalunya, Barcelona, Spain, in 1989 and
1995, respectively. He is currently a Pro-
fessor with the Universidad Politécnica de
Cartagena (UPCT). He has been a Visiting
FRANCISCO-JOSE ALVARADO-ALCON rece- Scholar at Queen’s University, Kingston, Canada,
ived the degree in telematics engineering from from 1991 to 1992, and Cornell University, Ithaca,
the Universidad Politécnica de Cartagena (UPCT), USA, from 2010 to 2011. He is the author or
where he is currently pursuing the M.S. degree. coauthor of more than 90 journal articles mainly in the fields of switching,
His research interests include wireless sensor net- wireless networking and performance evaluation. He served as the Editor-in-
works, the Internet of Things Technology, and Chief of the IEEE Global Communications Newsletter magazine, included
sustainable communications. He was awarded in the IEEE Communications Magazine, from April 2002 to December 2004.
the Premio Extraordinario de Bachillerato en la He has been the Technical Editor of the IEEE Communications Magazine,
Comunidad Autónoma de la Región de Murcia, from March 2001 to December 2011. He also received an Honorable Mention
in 2018, and a grant for Colaboration/Training in for the IEEE Communications Society Best Tutorial paper Award, in 1995.
Centres financed by Santander Bank.

VOLUME 10, 2022 111309