SPECIAL SECTION ON ADVANCES ON HIGH PERFORMANCE

WIRELESS NETWORKS FOR AUTOMATION AND IIOT

Received 15 May 2023, accepted 31 May 2023, date of publication 5 June 2023, date of current version 14 June 2023.
Digital Object Identifier 10.1109/ACCESS.2023.3283208

ML-RPL: Machine Learning-Based Routing

Protocol for Wireless Smart Grid Networks
CARLOS LESTER DUENAS SANTOS 1 , (Graduate Student Member, IEEE),
AHMAD MOHAMAD MEZHER 1 , JUAN PABLO ASTUDILLO LEÓN 2 ,
JULIAN CARDENAS BARRERA 1 , EDUARDO CASTILLO GUERRA 1 , (Senior Member, IEEE),
AND JULIAN MENG 1 , (Senior Member, IEEE)
1 Electrical and Computer Engineering Department, University of New Brunswick, Fredericton, NB E3B 5A3, Canada
2 School of Mathematical and Computational Sciences, Yachay Tech University, Urcuqui 100119, Ecuador
Corresponding author: Carlos Lester Duenas Santos ([email protected])
This work was supported in part by the Atlantic Canada Opportunities Agency under Project 208437, and in part by the NSERC
Collaborating Research and Development under Project CRDPJ537347-18.

ABSTRACT This research explores the potential of Machine Learning (ML) to enhance wireless commu-
nication networks, specifically in the context of Wireless Smart Grid Networks (WSGNs). We integrated
ML into the well-established Routing Protocol for Low-Power and Lossy Networks (RPL), resulting in an
advanced version called ML-RPL. This novel protocol utilizes CatBoost, a Gradient Boosted Decision Trees
(GBDT) algorithm, to optimize routing decisions. The ML model, trained on a dataset of routing metrics,
predicts the probability of successfully reaching a destination node. Each node in the network uses the model
to choose the route with the highest probability of effectively delivering packets. Our performance evaluation,
carried out in a realistic scenario and under various traffic loads, reveals that ML-RPL significantly improves
the packet delivery ratio and minimizes end-to-end delay, making it a promising solution for more efficient
and responsive WSGNs.

INDEX TERMS Machine learning, wireless smart grid networks, neighbourhood area networks (NAN),
routing protocol for low-power and lossy networks (RPL).

I. INTRODUCTION appliances. Smart meters collect data generated by HANs and

Communication technologies play an important role in the transmit it through the NAN to the utility data centers. NANs
current transformation of the electricity infrastructure. The connect smart meters and other utility equipment such as line
operation of the new power grids, known as smart grids, heav- sensors, reclosers, bank capacitors, smart switches, battery
ily relies on communication systems. Incorporating devices storage, and EV charge stations with data aggregation points
with communication capacities into the grid allows the util- (DAPs), also named collectors. Finally, WANs represent the
ities to achieve a high level of control over the grid, making last segment of the smart grid communication network, serv-
its operation more reliable, efficient, and secure. ing the critical role of linking the data aggregation points with
From the communication point of view, the smart grid the utility control centers. The primary function of WANs is
networks can be divided into three segments or subnetworks to establish a seamless connection between these two compo-
as it is shown in Fig. 1: Home Area Network (HAN), Neigh- nents of the smart grid network.
borhood Area Network (NAN), also referenced as Field Area The three subnetworks have their own requirements and
Network (FAN), and Wide Area Network (WAN). Home are important for the correct functioning of the smart grid
Area Networks are deployed inside homes and act as the com- network. However, the NAN has attracted the attention of
munication infrastructure to interconnect sensors and smart academia and industry since it supports the transmission
of a considerable volume of data and distributes necessary
The associate editor coordinating the review of this manuscript and control signals between many end devices and utility control
approving it for publication was Stefano Scanzio . centers [1].
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 57401
 C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

NANs usually cover large areas and can interconnect hun-

dreds of thousands of devices, which challenges the reliability
of the network. Communication systems deployed in NANs
include wired and wireless technologies or a combination of
them. However, due to the advantages of wireless technolo-
gies, they are becoming more popular for NANs. Two groups
of wireless technologies are relevant in this domain, RF mesh
technologies, which include 802.15.4 [2] and 802.11s [3],
and the second group considered Low Power Wide Area
Networking (LPWAN), which includes Sigfox [4], LoRa [5],
and NB-IoT [6].
This work has focused on one of the biggest challenges of FIGURE 1. Smart grid network [1].
mesh technologies: the way the packets are routed. Mesh net-
works heavily rely on the performance of the utilized routing
protocol. One of the well-known routing protocols for Wire-
• We conduct simulations using smart meter locations
less Smart Grid Networks (WSGNs) is the Routing Protocol
from a real deployment in the city of Montreal to eval-
for Low-Power and Lossy Networks (RPL) [7]. However,
uate the performance of our proposed routing strategy.
numerous studies have pointed out that RPL suffers from
Our simulation scenario was designed to be realistic and
issues that limit its efficiency and domain of applicability [8].
reflective of actual network conditions in wireless smart
We present in this article a new routing strategy over RPL
grid networks. To our knowledge, only one previous
based on Machine Learning (ML) predictions, which we have
study [11] has used real-world topology data for eval-
named ML-RPL. In recent years, ML techniques have shown
uation purposes in a simulation environment.
promising results to solve different problems in communi-
cation networks, and routing is one of them. In this case, The rest of the paper is organized as follows. Section II
we have selected CatBoost, a gradient-boosting algorithm, discusses some related works. A background on RPL parent
to improve the routing decisions in RPL. The predictions selection is provided in Section III. The proposed solution is
made with the algorithm are used to calculate the path cost presented and described in detail in Section IV. The perfor-
from the source to the destination. Nodes choose the routes mance of the new routing strategy is evaluated in Section V,
with the highest probability of delivering the packets to the and the conclusions and future works are addressed
destination. in Section VI.
While our proposed routing strategy uses Catboost for
prediction, it is important to note that this approach can be II. RELATED WORK
extended to other ML algorithms as well. We chose Catboost This section shows how ML-based solutions have gained
for its strong performance in previous studies [9], [10] and its interest in solving traditional network issues in wireless net-
ease of integration into our simulation platform. However, it is works. We mainly cover solutions for routing algorithms,
not the focus of this work to perform an in-depth comparison but other investigations related to congestion control and
of different ML algorithms in our routing strategy. traffic classification are also highlighted. We also show recent
In order to assess the effectiveness of our proposed improvements in RPL and the approaches they have followed.
approach, we conducted simulations using actual smart meter The work in [12] proposes DTMR, a decision tree-based
locations in the city of Montreal, with traffic patterns repre- multi-metric routing protocol for vehicular ad hoc networks
sentative of those found in smart grid applications. Our new (VANETs) that aims to make more intelligent forwarding
ML-based routing mechanism provided notable improve- decisions. The authors obtained their targeted dataset from
ments to both, packet delivery ratio (PDR) and end-to-end multiple simulations runs over different urban VANET sce-
delay across a range of traffic loads. narios and evaluated the importance of different routing met-
rics by applying regularization. Subsequently, only the most
relevant routing metrics were incorporated into DTMR for
A. CONTRIBUTION AND ORGANIZATION
forwarding decisions. However, the authors did not exploit
Based on our findings in this paper, the following contribu-
the inherent capability of decision tree-based models to deter-
tions can be outlined:
mine feature importance directly, which could potentially
• We propose a novel routing strategy based on ML to have resulted in a different ranking of routing metrics. Addi-
improve the routing decisions of RPL. To the best of our tionally, while the evaluation focuses on achieving an optimal
knowledge, previous research has not explored the use protocol design, the proposed solution is not benchmarked
of ML algorithms to improve RPL’s routing decisions. against other routing protocols for VANETs, which repre-
• We present a method based on simulations to generate sents the primary limitation of this work. The same authors
a dataset for collecting routing information in wireless present GraTree in [9], another ML-based multimetric routing
smart grid networks. protocol. For this paper, they employed CatBoost to train the

57402 VOLUME 11, 2023

C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

ML model using the same dataset they obtained in [12]. Gra- proposed routing algorithm outperforms other solutions in
Tree is compared to an existing solution not based on ML and the literature, the experiments conducted for a small network
to DTMR in terms of packet delivery ratio (PDR), end-to-end of 50 nodes raise concerns about the algorithm’s complex-
delay, and overhead. GraTree performs better than the other ity. The algorithm comprises multiple components, such as
routing protocols except for the delay measure. Although a grid-based network construction with various levels of
the routing solution presented in the paper targets VANETs, unequal grids, grid head selection, a transmission scheduling
the comprehensive description of their ML-based approach, mechanism, two different objective functions, and modifica-
including data collection, processing, learning phase, and tions to the trickle timer algorithm, in addition to employing
integration into the simulator, makes the methodology adapt- the Adam Deep Neural Network mentioned above.
able to other types of networks. Another work focused on improving RPL using ML tech-
Authors in [13] address the network congestion control niques is presented in [16]. In this case, a new parent selec-
problem in smart grids. They present a new congestion con- tion strategy is proposed to choose the best parent when
trol mechanism based on feed-forward neural networks for two or more candidates have the same ranking in the RPL
smart grids. The purpose of the work is to guarantee the destination-oriented routing tree. The core of this technique is
different Quality of Service (QoS) requirements for different to use Random Forest (RF) [17], for feature importance anal-
smart grid applications. The proposed mechanism requires ysis. The parent selection strategy is designed based on the
source nodes to decide whether to transmit new data packets importance of each feature, which are routing metrics, such as
generated by the applications based on the current state of expected transmission count, mac losses, channel utilization,
the network. This state is characterized by the value of the and throughput. The routing metric importance is then used to
channel utilization factor and by the packet buffer occupation. assign weights to a forwarding score function, which aids in
The evaluation of the proposed mechanism shows signifi- determining the optimal candidate parent among all potential
cant improvements in terms of throughput, transit time, and options. The primary limitation of this strategy is the static
quality of service differentiation. However, the grid topology assignment of weights, which prevents the parent selection
considered in this research work does not represent a real strategy from being more adaptable to different load condi-
smart grid deployment, which would provide a more realistic tions. Despite this limitation, the simulation results presented
test for congestion management. Future works must take into in the paper demonstrate significant improvement in terms
account smart meter positions from real deployments. of PDR compared to standard RPL implementations across
An Enhanced Tree Routing Based on Reinforcement various network sizes.
Learning for Wireless Sensor Networks (WSN) is presented Aside from the last two investigations mentioned above,
in [14]. The aim of the protocol is to identify the most suitable to the best of our knowledge, more recent approaches to
parent node within a tree topology by utilizing empirical improve RPL did not apply ML techniques to enhance the
network data gathered via Q-learning. The authors establish a routing decisions. For example, in a recent enhancement
state space, action set, and reward function based on multiple of RPL named Weighted Random Forward RPL (WRF-
cognitive metrics, and subsequently determine the optimal RPL) [18], the authors tackle the load balancing problem by
parent node through an iterative process. Simulation out- combining the energy remaining in the nodes and the number
comes indicate that the proposed method outperforms exist- of possible parents that a node has. These two routing metrics
ing techniques, such as the linear weighted sum-based parent are the base of a weighted random selection algorithm used to
selection algorithm, in aspects like end-to-end delay, packet choose the best next-hop candidate. Even though WRF-RPL
delivery ratio, and energy consumption. The analysis does not improves the network’s lifetime and the PDR, the solution
address the potential routing overhead incurred by the algo- ignores other important routing metrics related to link quality.
rithm due to the cycle detection mechanism implemented, In addition, considering only the number of parents may not
which requires each node to send a join request message accurately reflect the actual load on a node. Some nodes with
containing its list of child nodes to the candidate parent node. fewer parents might experience higher traffic or processing
Furthermore, the paper does not examine the scalability of demands, while others with more parents might be underuti-
the proposed method in larger network scenarios, which may lized. This can lead to an imbalanced load distribution and
present additional challenges and limitations. potentially reduced network performance.
In an attempt to improve the RPL protocol, authors Authors in [19] stand for QWL-RPL. They analyze the
in [15] use Adam Deep Neural Network (ADNN). The work RPL protocol under a heterogeneous traffic pattern and pro-
addresses the problem of routing overhead, packet losses, and pose a new protocol based on the queue and workload-based
load imbalanced in RPL. They state that the parent selection condition. The queue condition is obtained by counting the
policy must be changed based on the type of packet required number of packets in the queue, and the workload is measured
to transmit in order to achieve a better quality of service in at each node by counting the number of transmitted packets at
the network. ADNN is used to classify the packets consid- the MAC layer during fixed periods of time. The first metric
ering packet header information such as packet length, time would be an indicator of congestion, and the second one of
to live, payload length, and payload content. Although the traffic load. Thus, the node chooses as its preferred parent the

VOLUME 11, 2023 57403

 C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

one less congested and with less traffic load. The combination The rules of how each node determines the preferred parent
of the metrics is simply the sum of them. The most significant are guided by the Objective Function (OF). The Internet
improvement observed with QWL-RPL is a routing overhead Engineering Task Force (IETF) has standardized two OFs
reduction. After examining the results and the protocol’s per- for this purpose: the Objective Function Zero (OF0) defined
formance throughout all the experiments, the authors suggest in RFC6552 [22], and the Minimum Rank with Hysteresis
the potential benefits of incorporating ML into the routing Objective Function (MRHOF) defined in RFC6719 [23]. OF0
protocol. This would allow the protocol to possess self- utilizes the hop count as a routing metric to identify the
learning and self-adaptive features, facilitating a shift from optimal parent among candidate neighbors, making it ideal
a rule-based routing approach to a learning-based one. for selecting the nearest node to the DODAG root as the
The possibility of exploiting the multi-topology routing preferred parent.
feature of the RPL standard has also received some attention. In the case of MRHOF, the OF has been designed to
The authors in [20] investigate the use of multiple RPL find the paths with the smallest path cost while preventing
instances in Wireless Sensor Networks (WSNs). However, excessive network churn. It does so by using two mecha-
they utilize the two standard RPL implementations, and no nisms. First, it finds the minimum path cost, and second,
modifications are proposed for them. The implementation it switches to the node that offers the shortest path cost if
based on the hop count metric is used for the instance of peri- the path cost through this node is less than the current path
odic and non-critical traffic, and the implementation based cost by at least a given threshold. This second mechanism is
on the expected transmission count (ETX) is used for the called ‘‘hysteresis’’ [23]. By default, the OF uses the expected
instance of high-critical data traffic. Another work in the transmission count as a routing metric to calculate the path
same direction is presented in [21]. This work focuses on cost. The ETX is a measure of the quality of a link in terms of
the use of multiple RPL instances to ensure the Quality of reliability. Low values indicate a link is more reliable and vice
Service (QoS) provision for different traffic classes. Also, this versa. The path cost from a node to the DODAG root is the
investigation addresses the single routing metric problem in sum of the ETX of each link in the path. If multiple candidate
RPL, and a new parent selection framework based on a multi- parents share the same path cost, other tie-breaking criteria
attribute decision-making approach is proposed. The results might be used, which is implementation dependent. One of
show that the multi-topology routing approach improves the the effective alternatives for tie-breaking is presented in [16].
QoS provision in the network. However, the total length of
each simulation may not be sufficient to capture the long-term IV. MACHINE LEARNING-BASED ROUTING DESIGN
behavior and performance of the proposed routing approach In an RPL network, each node recognizes its neighbor nodes
under various network conditions. by DIO messages received from them. Thus, every time a
In summary, while some works covered in the literature node receives a DIO message, it must update the candidate
review have shown promising results, there remain challenges parent set and select the preferred parent considering the
such as routing metrics selection, the use of realistic network routing metrics defined for that purpose. In our strategy,
topologies, scalability, dynamic adaptability, and balancing we follow the same approach, but instead of combining the
the trade-offs between performance improvements and added metrics directly as many previous investigations in the litera-
complexity. Building on these insights, our proposed machine ture have done, we use the metrics to calculate the probability
learning-based routing protocol addresses these challenges to of reaching the node that sent the DIO. To calculate the prob-
further enhance the routing performance for WSGNs. ability, the nodes use an ML model that has as inputs features
based on certain routing metrics and outputs a probability on
whether the DIO-sender node can be successfully reached.
III. RPL PARENT SELECTION BACKGROUND Let’s designate the positive event of reaching the DIO-
To provide basic knowledge for understanding our proposed sender as class label y = 1, and the p(y = 1|x) is the
machine learning-based routing protocol for smart grid net- probability that a particular sample belongs to class 1 given
works, it is essential to first explore the parent selection its features x. Then, the path cost of node k to reach the DIO-
process in RPL. As our design builds upon and enhances sender (e.g. node m) could be expressed by the following
the RPL protocol, this section aims to provide the necessary equation:
background on RPL’s parent selection mechanism.
RPL is a distance vector routing protocol that is adapted path_cost(k, m) = 1 − p(y = 1|x) (1)
to a variety of Low-Power and Lossy Networks (LLN). The As we are interested not only in the path cost to reach node
protocol constructs a multihop routing tree rooted at a single m but also in the path cost to the destination (i.e., DODAG
6LoWPAN Border Route (6LBR) by forming a destination- root), the total path cost of node k to reach the DODAG root
oriented directed acyclic graph (DODAG) [7]. When a new through node m would be the sum of the path cost to reach
node joins an RPL network, it selects a parent node (default node m, plus the node m’s path cost to reach the root. That is
route) based on the DODAG information that it receives expressed in Eq.( 2):
from its neighbors through DAG information object (DIO)
messages. C(k, m) = path_cost(k, m) + C(m, Pˆm ) (2)

57404 VOLUME 11, 2023

C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

where Pˆm is the preferred parent of node m. It is worth noting

that node m must advertise its path cost so that node k can
calculate the path cost through node m.
Nodes keep a record of the total path costs through all
possible candidate parents. The node that offers the lowest
path cost to the destination is then selected as the preferred
parent. In general, If S is the set of n candidate parents of
node k, the best alternative parent or preferred parent Pˆk is
given by:
Pˆk = argmin{C(k, n)} (3)
n∈S
As in the MRHOF version of RPL, a node switches to
a new parent only if the new minimum calculated total
path cost is smaller than the current total path cost through
the preferred parent by the parent_switch_threshold. The
value of parent_switch_threshold serves as a hysteresis factor
that aims to prevent unnecessary and inefficient parent node
changes.
Algorithm 1 describes the entire strategy for parent selec-
tion. The algorithm starts having a set of candidate parents,
followed by computing the total path cost to the destination
via each candidate. Then, the algorithm determines the can-
didate with the lowest path cost and switches to a new parent
if the threshold decision is satisfied.

Algorithm 1 Pseudo-Code of the Parent Selection Strategy

Require: S, a set of n candidate parents of node k. Pk ,
current_preferred_parent. ζ , parent_switch_threshold.
for each n ∈ S do FIGURE 2. Roadmap to build the ML model for ML-RPL.

Calculate total cost to the root through n, C(k, n).

end for
Pˆk = argmin{C(k, n)} provides the positions of 335,297 smart meters distributed
n∈S over an area of approximately 431 square kilometers in the
if C(k, Pˆk ) ≤ (C(k, Pk ) − ζ ) then city of Montreal. The large scale of this deployment made it
Pk ← Pˆk infeasible to simulate the entire scenario using any network
end if simulation tool. Thus, the simulations for data collection were
return Pk run separately over different areas, as illustrated in Fig. 3.
The simulations were carried out using the network simulator
The parent selection strategy described in Algorithm 1 OMNeT++ [24].
is the core of our ML-based routing proposal, ML-RPL. In each simulation, the smart meters sent packets through
Since the path cost calculation in the strategy depends on random routes to the collector. Our focus was to capture
the predictions made by an ML model, the necessary steps to routing metrics at the moment a node was either transmitting
develop the ML model are presented in Fig. 2. The following or forwarding a packet. This allowed us to collect routing
subsections describe each step of the process. metrics related to the link between sender and receiver and
others that give information related to the status of the nodes.
A. DATA COLLECTION AND PROCESSING In summary, the gathered routing metrics are:
The first step in Fig. 2 focuses on data collection and the • Hop count. Distance to the destination based on the
subsequent processing required to obtain a representative number of hops.
dataset for model training and evaluation. In practice, con- • Expected Transmission Count (ETX). Calculated with
structing a good dataset can be challenging as it is heavily respect to the meter that is supposed to receive the
reliant on the type of data being observed. To address this packet. The ETX is calculated on a link according to the
challenge, it was imperative to collect data using a real smart following expression:
grid network, as there were no pre-existing datasets avail-
able for our proposal. Accordingly, representative simulation 1
ETX = (4)
scenarios were created, considering a public database that Df · Dr

VOLUME 11, 2023 57405

 C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

TABLE 1. Simulation settings for data collection.

FIGURE 3. Smart meters deployment in the city of montreal.

meters varied from 5 minutes up to 1 hour. Also, the number

Df is the measured probability that a packet is received of smart meters in the scenarios was changed from 50 to 300.
by the neighbor and Dr is the measured probability that The communication channel was modeled by using the
the acknowledgment packet is successfully received. Log-Normal Path Loss Model as suggested in [26]. The value
The ETX is also smoothed by using an exponential for the path-loss exponent α and the shadowing deviation σ
weighted moving average (EWMA) filter, making it are taken from [27], which have been used in previous related
robust to sudden link condition changes: works like [28]. Other simulations setting are given in Table 1.
ETXnew = α · ETXcurrent + (1 − α) · ETXprior (5) To summarize the data collection, Fig. 4 shows a simplified
example of the process. For instance, if node 5 sends a packet
the value of α is implementation dependent, and it has to the collector, represented by node C, the hop statistics are
been set to 0.8 as in other implementations [25]. stored in a repository with the structure shown at the bottom
• MAC losses. The average percentage of frame losses in of Fig. 4. The columns of the repository corresponding to
the MAC Access Control (MAC) layer. These losses are packet sending time, source, packet identifier, packet event
computed as the sum of losses resulting from the max- (packet sent, forwarded, or received at the destination), and
imum number of frames retransmitted, the maximum the values of the different routing metrics at each hop.
number of extra backoff times reached, and queue drops. The raw data obtained from the simulations have to be
• Density. Number of neighbors within the transmission processed and organized in a structured way. To this end,
range of the meter that is supposed to receive the packet. it was used Pandas [29], a library developed in the Python
• Channel utilization. Percentage of the time the channel programming language for the purpose of manipulating and
is busy at the receiver side. analyzing data. Once the data is processed, the dataset used
• Throughput. Frames per second transmitted by the to train the ML model looks like in Fig. 5. Each sample in
meter that is supposed to receive the packet. the dataset represents a unique packet transmitted over the
• Queue utilization (Qu). Queue utilization of the meter network, recorded hop by hop. The features are the rout-
that is supposed to receive the packet. It is computed by ing metrics recorded at each hop, and the classes represent
each node as: whether the packet was received at the next hop or not
Number of packets in queue of node k (r = 1, if the packet was received successfully, r = 0 oth-
Qu(k) = (6)
Total queue size of node k erwise). A summary of the characteristics of the dataset that
• Received Signal Strength Indicator (RSSI). Average will be used for training is shown in Table 2.
of RSSI values of the packets received from the node
that is supposed to receive the packet. B. ML MODEL SELECTION
All the routing metrics except for Density and the Hop The second step in our roadmap involves the selection of the
count were smoothed using an EWMA filter in a similar way ML model that will be used for predicting the best parent
as the ETX. In order to obtain a large number of samples cov- for forwarding packets at each hop. We are presented with
ering a wide range of values of these routing metrics, different a binary classification problem where we want the ML algo-
simulation conditions were considered. For example, to have rithm to learn when the packets are received successfully or
different congestion levels, the sending interval of the smart not at the next hop.

57406 VOLUME 11, 2023

C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

make forwarding decision tasks, and the results encouraged

its use in this work. Catboost can also be exported to C++,
which makes it easy to integrate it into the network simulator.
CatBoost is built upon the theory of decision trees and
gradient boosting. Gradient boosting ML technique was first
described in [30]. The main idea of boosting is to sequentially
combine many weak models to create a solid competitive
predictive model. When decision trees are used as base learn-
ers, a new tree is added at each step of the process. The
decision trees are fitted sequentially, so the fitted trees learn
from the former trees’ errors to minimize the value of the
loss function. Adding a new tree to existing ones continues
until the selected loss function is no longer minimized or a
maximum number of trees is reached. Algorithm 2 describes
the boosting process with decision trees as base learners.

Algorithm 2 GBDT Algorithm

Require: (xi , yi )ni a set of data, where xi are input val-
ues, and yi , i ∈ {1 . . . n} are the expected output values.
L(yi , F(x)), a differentiable Loss Function. Learning rate,
α. M , maximum number ofP decision trees.
Initialize: F0 (x) = argmin ni=1 L(yi , γ )
γ

for m = 1 to M do
1: Compute rim = − ∂L(yi ,F(xi ))
h i
FIGURE 4. Example of data collection hop by hop and the repository. ∂F(xi ) for i =
F(x)=Fm−1 (x)
1, . . . , n
2: Fit a regression tree to the rim values and create
terminal regions Rjm , for j = 1 . . . Jm
1 . . . Jm , compute
3: For j = P
γjm = argmin xi ∈Rjm L(yi , Fm−1 (xi ) + γ )
γ
4: Update Fm (x) = Fm−1 (x) + α · Jj=1 γm 1(x ∈ Rjm )
Pm
end for
Output: FM (x)

Two critical algorithmic introduced in CatBoost are the

FIGURE 5. Dataset appearance.
implementation of ordered boosting, a permutation-driven
alternative to the classic algorithm, and an innovative algo-
TABLE 2. Dataset summary. rithm for processing categorical features. Both techniques
were created to fight a prediction shift caused by a spe-
cial kind of target leakage present in all currently existing
implementations of gradient boosting algorithms [31]. Also,
CatBoost does not follow similar gradient boosting models in
the growing procedure of the decision trees. Instead, it grows
oblivious trees, meaning that the same splitting criterion is
used across an entire level of the tree. Such trees are balanced,
less prone to outfitting, and allow to speed up the executions
at the testing time significantly [32].

C. MODEL OPTIMIZATION AND VALIDATION

There are many ML algorithms for classification problems. Prior to initiating the training process, the dataset is parti-
In our case, we have chosen CatBoost, a member of the tioned into two subsets, with 80% allocated for training and
family of Gradient Boosted Decision Trees (GBDT) algo- 20% for testing. The initial phase of the training process
rithms. This algorithm has already been used successfully in involves adjusting the primary hyperparameters of the model.
networking. In [9], it is compared to other ML algorithms to These hyperparameters, such as the learning rate, the number

VOLUME 11, 2023 57407

 C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

TABLE 3. Hyperparameter search space.

FIGURE 6. Receiver operating characteristic (ROC) curve.

of trees in the ensemble, and the maximum depth of each tree, It is seen that the RSSI metric has the most influence on
and others, are predetermined model parameters that are not the model output, followed by the ETX and the throughput.
learned from the data. To do so, we have used a technique Contrarily, the density, and the queue utilization show low
named Grid Search, which is implemented in the Python significance in the model output. In order to complement
library scikit-learn [33]. The grid search technique involves the previous analysis, we have applied Recursive Feature
exploring the optimal parameter values within a specified Elimination (RFE). RFE is a feature selection method that
parameter grid. This means that a range of values is assigned recursively removes the least important feature until the spec-
for each hyperparameter, which is then evaluated through ified number of features is reached. In Fig 8, we have plotted
the grid search to determine the most suitable combination. how the AUC of the model varies versus the number of
Table 3 summarizes the hyperparameters search space. features selected. The figure shows that utilizing more than
In order to see how the Catboost model performs after the six metrics does not result in any improvement in the ROC
hyperparameter optimization, we show the Receiver Oper- metric. As a consequence, the ML model does not consider
ating Characteristic (ROC) curve in Fig. 6. The ROC curve either the density or the queue utilization when making rout-
is one of the most well-known performance metrics for ML ing decisions.
classifiers [34]. The area under the curve (AUC) is a measure The developed ML model has to be integrated into the
of the ability of a classifier to distinguish between classes and network simulator, step 3 in Fig. 2. Since the model was
is used as a summary of the ROC curve. The higher the AUC, initially developed using Python, we converted the Python-
the better the performance of the model at distinguishing based model to its equivalent C++ implementation. The
between the positive and negative classes. Fig. 6 shows five C++ model is incorporated into the OMNeT++ simulator
curves and their mean because K-fold Cross-Validation has by linking the implementation with the rest of the modules
been used. K-fold Cross-Validation is a technique where the and components of the simulator. The next section is dedi-
dataset is randomly split into k folds without replacement, cated to testing the model as part of the routing protocol in
where (k − 1) folds are used for the model training and one the network simulator.
fold is used for testing. This procedure is repeated k times,
and the ROC curve was recorded at each time. The mean value V. PERFORMANCE EVALUATION
of ROC shown in Fig. 6 is 0.91, which means that our model We present the performance evaluation of our proposed
achieves good class separation performance. ML-RPL in this section. Our approach is compared to one
We are also interested in analyzing the relevance of of the standard RPL implementations (MRHOF) and also
each feature in the model. By identifying the routing met- to RPL+, the proposal presented in [16]. Three different
rics that are most relevant, the proposed routing algorithm experiments varying the network load were conducted to
can improve efficiency. Catboost has implemented different compare the routing algorithms. Next are the details of the
methods to calculate the feature importance after model train- simulation settings, and later the results of each experiment
ing, so we can directly get the individual importance values are presented.
for each of the input features. The result of this analysis is
presented in Fig 7. The values are normalized so that the sum A. SIMULATION SETTINGS
of the importance of all features is equal to 100. A higher To compare the routing algorithms, we utilized the
value of the importance indicates a larger average change to OMNeT++ simulator in a realistic scenario extracted from
the prediction value, if this feature is changed. the city of Montreal. The scenario consisted of a set of

57408 VOLUME 11, 2023

C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

FIGURE 7. Feature importance analysis.

FIGURE 9. Simulation scenario.

smart grid application. It refers to the usage information that

smart meters collect and must report periodically to utilities.
AE is the second application taken into consideration. Alarms
can happen at any time and are sent randomly during the sim-
ulation time by a percentage of smart meters. AE can include
events such as measurement failure, system restart, system
memory full, configuration errors, etc. The other application
FIGURE 8. ROC_AUC score versus number of features.
considered has been PQ. Examples of Power Quality events
include but are not limited to leading/lagging power, voltage
TABLE 4. Simulation parameters. fluctuations, imbalance in energy flow, harmonics, and volt-
age sags and swells. Table 5 presents how these applications
have been configured in each experiment.
As shown in Table 5, experiment 1 has the lowest traffic
load. For experiment 2, the three applications are transmitted,
but only 25% of the smart meters transmit AE and PQ. Exper-
iment 3 has the highest traffic load, where the MR traffic is
doubled, and 50% of the smart meters generate AE and PQ
traffic.

B. SIMULATION RESULTS
Fig. 10 shows the PDR measured at the collector for the first
experiment. Recall that the PDR expresses the ratio between
200 smart meters and one collector, as depicted in Fig. 9. the number of successfully received packets at the destination
The channel, physical layer, and MAC layer have the same and the total number of transmitted packets. In this experi-
characteristics as was described in the data collection process, ment, the smart meters transmit only MR packets. It can be
Table 1. Each simulation executed corresponds to 1.5 hours of seen that ML-RPL performs better than MRHOF and RPL+
network operation, and the results presented are the average by 7% and 5%, respectively. Overall, ML-RPL is able to
of 10 simulations per experiment with a confidence interval achieve 94% of PDR on average.
of 95%. The simulation settings can be found in Table 4. In terms of end-to-end delay, the comparison among the
Regarding network traffic, typical smart grid applications three routing protocols is shown in Fig 11. The highest delay
such as Meter reading (MR), Alarm events (AE), and Power corresponds to MRHOF since it is based on the ETX metric
Quality (PQ) have been considered. MR is the most basic which measures the most reliable paths not necessarily the

VOLUME 11, 2023 57409

 C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

TABLE 5. NAN applications transmitted over the SG per experiment.

FIGURE 10. Packet delivery ratio in experiment 1.

shortest ones. The average delay of RPL+ is almost half of

MRHOF’s value. It is worth noting that RPL+ takes into
account the hop count metrics to build the routes from nodes
to the root. Later other metrics are applied to decide among
candidates with the same hop distance, so the paths with
minimum distance to the destination are always chosen. This
is why better results with respect to MRHOF were expected.
ML-RPL was found to be slightly better than RPL+ with
an average difference in delay of only 60 ms. However, it is
worth noting that ML-RPL exhibits greater deviation from
FIGURE 11. End-to-end delay in experiment 1.
the mean delay value, with some individual simulation runs
experiencing delays exceeding 600 ms. ML-RPL is more
of an intermediate solution between MRHOF and RPL+.
It considers the hop count metric alongside other metrics to
predict the optimal path. As a result, the weight assigned
to other metrics may sometimes result in longer paths hav-
ing a better probability of successful delivery. Nevertheless,
ML-RLP guarantees successful delivery to the destination.
In the second experiment, we added AE and PQ traffic.
Fig 12 depicts the PDR for each routing variant. We can
observe that ML-RPL achieves the best PDR for the 3 traffic
categories. However, for MR, the PDR fell 3% with respect
the experiment 1 when there was no more traffic simulta-
neously. The other routing algorithms were also unable to
achieve the same PDR for MR traffic as in experiment 1.
Nevertheless, for RPL+ the reduction was only 1%. The
difference in favor of our machine learning-based routing
protocol is more notable for the other traffic cases. For AE, FIGURE 12. Packet delivery ratio in experiment 2.
the PDR of ML-RPL is 6% and 9% better than the PDR
observed with RPL+ and MRHOF, respectively. Regarding
the third traffic type, PQ, the 94% of PDR achieved by The packet delay for the second experiment is shown in
ML-RPL is better by 11% with respect MRHOF and 4% with Fig 13. It can be seen again that MRHOF has the worst
respect to RPL+. average packet delay. It is also seen that ML-RPL achieves

57410 VOLUME 11, 2023

C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

three applications. Since the traffic load was increased, the

transmission attempts face more contention, which results
in longer back-off times and more re-transmission, so the
delay increases in consequence. It is worth noting that there
is a spike in the PQ application when ML-RPL is used. The
average value is slightly bigger than even MRHOF, but we
know from Fig. 14 that more smart meters are able to reach
the destination when using ML-RPL. Some of those smart
meters are further away from the collector, which results in
an increase in the average end-to-end delay.
In summary, the improvement observed in our ML-based
routing proposal compared to the other protocols can be
attributed to several key factors. Firstly, we emphasize the
data-driven approach employed in our design. In our specific
context, this means that the parent selection process is guided
FIGURE 13. End-to-end delay in experiment 2. by insights extracted from a previously generated dataset.
This allows our routing protocol to make more accurate deci-
sions when selecting the most suitable node for forwarding
packets.
In addition, the incorporation of an ML model into our
routing protocol design enables greater adaptability com-
pared to traditional approaches. Our proposal has demon-
strated its ability to adapt to varying network conditions and
to learn the complex relationships between diverse features.
This adaptability allows our ML-based routing protocol to
make more accurate predictions even in the presence of
network load variations, leading to improved performance
metrics such as PDR for all test scenarios.

C. ROUTING OVERHEAD
The consumption of significant network resources and reduc-
tion of efficiency are major concerns in wireless networks due
to routing overhead. This section will evaluate the routing
FIGURE 14. Packet delivery ratio in experiment 3. overhead of our suggested routing protocol and compare it
with MRHOF and RPL+.
Routing overhead is defined as the number of control
slightly better delay that RPL+ for MR traffic as in experi- messages generated by a routing protocol to maintain network
ment 1, but this is not the case for the rest of the traffic cases. connectivity and routing tables. In RPL, this involves DIO
Fig. 14 shows the PDR results for the third experiment. and DAO messages transmitted to keep network connectiv-
In this experiment, we increased the network load by decreas- ity. Fig. 16 shows the average number of control messages
ing the sending interval of the MR application by 50% and generated in each experiment by each protocol.
doubling the percentage of the smart meters sending alarm As traffic load increases, the number of control messages
reports and power quality events. The main goal is to see tends to increase for all routing protocols. This is because
how the new proposal routing protocol behaves when the net- nodes change their preferred parent more frequently when
work load increases. The PDR shows that ML-RPL decreases traffic load increases, looking for the best alternative. These
slightly for each traffic category with respect to experiment 2. changes generate extra control messages to notify neigh-
The reduction in PDR is 1% for MR, 4% for AE, and 4% bors about the change, which can also cause neighbors to
for PQ. The PDR of RPL+ and MRHOF decreases as well change parents based on the new information. When nodes
for 2 out of 3 traffic types with respect to experiment 2. change parents too frequently, it can cause network instability
It was expected as the traffic load increases, the PDR would and potential loops. Thus, there is a clear trade-off between
decrease. However, ML-RPL can keep the PDR above 90% changing to the best alternative candidate parent at the price
for each traffic type, which is a significant advantage of our of increasing network overhead and instability, and keeping
strategy based on ML techniques. the current parent even when it is not the best option.
The end-to-end delay for the last experiment is depicted Fig. 16 shows that ML-RPL has slightly less overhead
in Fig. 15. It shows a similar pattern as in experiment 2, but than MRHOF. This is due to our protocol’s consideration
all the routing variants have higher values of delay for the of additional factors in the state of the link and the state

VOLUME 11, 2023 57411

 C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

protocol strikes a good balance between the other two routing

protocols.

VI. CONCLUSION AND FUTURE WORK

We have presented a novel proposal of a machine learning-
based routing protocol for Wireless Smart Grid Networks
named ML-RPL. The proposal focuses on enhancing the
parent selection strategy of one of the well-known protocols
in this type of networks, the Routing Protocol for Low-Power
and Lossy Networks. We started our design by gathering a
large amount of simulation data from a real deployment of
smart meters that helped us to build a representative dataset.
Each feature in the dataset corresponds to a value of a routing
metric that characterizes the link or the state of the nodes
measured hop by hop as the packets travel from source to
FIGURE 15. End-to-end delay in experiment 3. destination. This dataset served to train the ML algorithm
and is later used to make the routing decisions. In addition,
a feature importance analysis was conducted to use, by the
protocol, only the routing metrics more relevant.
The performance evaluation of the new proposal showed
significant improvements with respect to one of the standard
RPL implementations in terms of packet delivery ratio and
end-to-end delay. The proposal also overcame an RPL mod-
ification (RPL+) that uses Random Forest to improve the
parent selection strategy. When comparing to RPL+, it can be
seen that ML-RPL is consistently better in terms of PDR, and
in some cases, the end-to-end delay achieved is also better.
However, as more packets arrive at the destination when
ML-RPL is used, some of those packets followed longer paths
and consequently increase the average end-to-end delay with
respect to RPL+.
In general, the results show a significant potential of the
ML techniques to be applied in solving or improving net-
FIGURE 16. Overhead per experiment. working tasks. The capacity of learning from a dataset that
contains previous events and later use that knowledge has
been the key for better results. For future work, we are
of the possible candidate parents, rather than relying solely strongly considering to incorporate Quality of Service (QoS)
on the ETX routing metric, as MRHOF does. Consequently, criteria in the decision-making process to guarantee the QoS
variations in the ETX value do not always prompt a parent of some applications with specific requirements. In addition,
change in our protocol. we plan to extend our routing strategy to other ML algorithms
Compared to RPL+, ML-RPL incurs more overhead. This and conduct a detailed comparison among them.
difference in overhead can be attributed to certain key differ- REFERENCES
ences between the two protocols that affect their operation [1] J. P. Astudillo León, C. L. Duenas Santos, A. M. Mezher, J. Cardenas
and performance. RPL+ is an extension of the RPL standard Barrera, J. Meng, and E. Castillo Guerra, ‘‘How does the selection of wire-
less technology impact the performance of the smart grid? A simulation
that uses the hop count metric as its base. It also takes
approach,’’ in Proc. 19th ACM Int. Symp. Perform. Eval. Wireless Ad Hoc,
into account other metrics such as ETX, throughput, MAC Sensor, Ubiquitous Netw. New York, NY, USA: Association for Computing
losses, and channel utilization, in a tie-breaking equation to Machinery, 2022, pp. 67–74.
decide among candidates with the same ranking in the routing [2] IEEE Standard for Local and Metropolitan Area Networks—Part 15.4:
Low-Rate Wireless Personal Area Networks (LR-WPANS) Amendment 3:
tree, but the weights assigned to each of them are fixed. Physical Layer (PHY) Specifications for Low-Data-Rate, Wireless, Smart
In contrast, ML-RPL is based on a Gradient boosted decision Metering Utility Networks, Standard 802.15.4g-2012, 2011.
tree algorithm, which is a non-linear model that can capture [3] IEEE Draft Standard for Information Technology-Telecommunications and
Information Exchange Between Systems-Local and Metropolitan Area
complex interactions between features, resulting in a more Networks-Specific Requirements—Part 11: Wireless LAN Medium Access
dynamic routing protocol with wider ranges of prediction and Control (MAC) and Physical Layer (PHY) Specifications-Amendment 10:
higher variability in parent selection. Despite the increase Mesh Networking, Standard P802.11bb/D5.0, Dec. 2022.
[4] U. Raza, P. Kulkarni, and M. Sooriyabandara, ‘‘Low power wide area
in overhead, network performance is not degraded; in fact, networks: An overview,’’ IEEE Commun. Surveys Tuts., vol. 19, no. 2,
it is beneficial based on the PDR. Therefore, our innovative pp. 855–873, 2nd Quart., 2017.

57412 VOLUME 11, 2023

C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

[5] What is it ? A Technical Overview of LoRa, LoRa Alliance, San Ramon, [27] D. Cypher, NISTIR 7761 NIST Priority Action Plan 2 Guidelines for
CA, USA, 2015. Assessing Wireless Standards for Smart Grid Applications NIST Priority
[6] K. Mekki, E. Bajic, F. Chaxel, and F. Meyer, ‘‘Overview of cellular Action Plan 2 Guidelines for Assessing Wireless Standards for Smart Grid
LPWAN technologies for IoT deployment: Sigfox, LoRaWAN, and NB- Applications, document NISTIR 7761, Feb. 2011, pp. 1–104.
IoT,’’ in Proc. IEEE Int. Conf. Pervasive Comput. Commun. Workshops [28] G. Rajalingham, Y. Gao, Q.-D. Ho, and T. Le-Ngoc, ‘‘Quality of service
(PerCom Workshops), Mar. 2018, pp. 197–202. differentiation for smart grid neighbor area networks through multiple RPL
[7] R. A. T. Winter, P. Thubert, A. Brandt, J. Hui, R. Kelsey, P. Levis, K. Pister, instances,’’ in Proc. 10th ACM Symp. QoS Secur. Wireless Mobile Netw.,
R. Struik, and J. P. Vasseur, RPL: IPv6 Routing Protocol for Low-Power Sep. 2014, pp. 17–24.
and Lossy Networks, document RFC 6550, 2012. [29] W. McKinney, ‘‘Data structures for statistical computing in Python,’’ in
[8] B. Ghaleb, A. Y. Al-Dubai, E. Ekonomou, A. Alsarhan, Y. Nasser, Proc. 9th Python Sci. Conf., S. van der Walt and J. Millman, Eds., 2010,
L. M. Mackenzie, and A. Boukerche, ‘‘A survey of limitations and pp. 51–56.
enhancements of the IPv6 routing protocol for low-power and lossy net- [30] J. H. Friedman, ‘‘Greedy function approximation: A gradient boosting
works: A focus on core operations,’’ IEEE Commun. Surveys Tuts., vol. 21, machine,’’ Ann. Statist., vol. 29, no. 5, pp. 1189–1232, Oct. 2001.
no. 2, pp. 1607–1635, 2nd Quart., 2019. [31] L. Prokhorenkova, G. Gusev, A. Vorobev, A. V. Dorogush, and A. Gulin,
[9] L. Lemus Cárdenas, J. P. Astudillo León, and A. M. Mezher, ‘‘GraTree: ‘‘CatBoost: Unbiased boosting with categorical features,’’ in Proc. Adv.
A gradient boosting decision tree based multimetric routing protocol Neural Inf. Process. Syst., vol. 31, 2018, pp. 1–11.
for vehicular ad hoc networks,’’ Ad Hoc Netw., vol. 137, Dec. 2022, [32] A. Veronika Dorogush, V. Ershov, and A. Gulin, ‘‘CatBoost: Gradient
Art. no. 102995. boosting with categorical features support,’’ 2018, arXiv:1810.11363.
[10] J. P. Astudillo León, L. J. de la Cruz Llopis, and F. J. Rico-Novella, [33] G. Hackeling, Mastering Machine Learning With Scikit-Learn. Birming-
‘‘A machine learning based distributed congestion control protocol for ham, U.K.: Packt Publishing Ltd, 2017.
multi-hop wireless networks,’’ Comput. Netw., vol. 231, Jul. 2023, [34] A. Tharwat, ‘‘Classification assessment methods,’’ Appl. Comput. Infor-
Art. no. 109813. mat., vol. 17, no. 1, pp. 168–192, Jan. 2021.
[11] G. Iyer, P. Agrawal, E. Monnerie, and R. S. Cardozo, ‘‘Performance
analysis of wireless mesh routing protocols for smart utility networks,’’ in
Proc. IEEE Int. Conf. Smart Grid Commun. (SmartGridComm), Oct. 2011,
CARLOS LESTER DUENAS SANTOS (Graduate
pp. 114–119.
Student Member, IEEE) received the B.Sc. degree
[12] L. Lemus Cárdenas, A. M. Mezher, J. P. Astudillo León, and M. A. Igartua,
‘‘DTMR: A decision tree-based multimetric routing protocol for vehicular in telecommunication and electronics and the
ad hoc networks,’’ in Proc. 18th ACM Symp. Perform. Eval. Wireless Ad M.Sc. degree in telecommunication from the Uni-
Hoc, Sensor, Ubiquitous Netw., 2021, pp. 57–64. versity ‘‘Marta Abreu’’ of Las Villas, Santa Clara,
[13] J. P. Astudillo León, F. J. Rico-Novella, and L. J. de la Cruz Llopis, Cuba, in 2008 and 2015, respectively. He is cur-
‘‘Predictive traffic control and differentiation on smart grid neighborhood rently pursuing the Ph.D. degree with the Emera
area networks,’’ IEEE Access, vol. 8, pp. 216805–216821, 2020. and NB Power Research Center for Smart Grid
[14] B.-S. Kim, B. Suh, I. J. Seo, H. B. Lee, J. S. Gong, and K.-I. Kim, Technologies, Electrical and Computer Engineer-
‘‘An enhanced tree routing based on reinforcement learning in wireless ing Department, University of New Brunswick,
sensor networks,’’ Sensors, vol. 23, no. 1, p. 223, Dec. 2022. Canada. His research interests include smart grid communications, routing
[15] A. Kumar and N. Hariharan, ‘‘DCRL-RPL: Dual context-based routing and protocols, network security, and machine learning.
load balancing in RPL for IoT networks,’’ IET Commun., vol. 14, no. 12,
pp. 1869–1882, Jul. 2020.
[16] C. L. Duenas Santos, J. P. Astudillo León, A. M. Mezher, J. Cardenas AHMAD MOHAMAD MEZHER received the
Barrera, J. Meng, and E. Castillo Guerra, ‘‘RPL+: An improved parent
M.S. degree in signals and systems from the
selection strategy for RPL in wireless smart grid networks,’’ in Proc. 19th
Central University ‘‘Marta Abreu’’ of Las Villas,
ACM Int. Symp. Perform. Eval. Wireless Ad Hoc, Sensor, Ubiquitous Netw.,
2022, pp. 75–82.
Santa Clara, Cuba, in 2011, and the Ph.D. degree in
[17] S. Raschka and V. Mirjalili, Python Machine Learning: Machine Learning
network engineering from Universitat Politècnica
and Deep Learning with Python, Scikit-Learn, and TensorFlow. Birming- de Catalunya (UPC), Barcelona, Spain, in 2016.
ham, U.K.: Packt Publishing Ltd, 2017. He currently holds the McCain Postdoctoral Fel-
[18] P. D. Acevedo, D. Jabba, P. Sanmartín, S. Valle, and E. D. Nino-Ruiz, lowship of innovation with the Electrical and Com-
‘‘WRF-RPL: Weighted random forward RPL for high traffic and energy puter Engineering Department, University of New
demanding scenarios,’’ IEEE Access, vol. 9, pp. 60163–60174, 2021. Brunswick (UNB). His research interests include
[19] A. Musaddiq, Y. B. Zikria, and S. W. Kim, ‘‘Routing protocol for low- smart grid communications, vehicular ad hoc networks, data privacy, and
power and lossy networks for heterogeneous traffic network,’’ EURASIP load forecasting. He was a recipient of a FI-AGAUR Fellowship from the
J. Wireless Commun. Netw., vol. 2020, no. 1, pp. 1–23, Dec. 2020. Generalitat de Catalunya and the Social European Budget.
[20] W. Mardini, S. Aljawarneh, and A. Al-Abdi, ‘‘Using multiple RPL
instances to enhance the performance of new 6G and Internet of Every-
thing (6G/IoE)-based healthcare monitoring systems,’’ Mobile Netw. Appl., JUAN PABLO ASTUDILLO LEÓN received the
vol. 26, no. 3, pp. 952–968, Jun. 2021. Ph.D. degree (Hons.) in network engineering
[21] K. S. Bhandari, I. Ra, and G. Cho, ‘‘Multi-topology based QoS- from Universitat Politècnica de Catalunya (UPC),
differentiation in RPL for Internet of Things applications,’’ IEEE Access,
Spain, in 2020. He was a Postdoctoral Fellow with
vol. 8, pp. 96686–96705, 2020.
the University of New Brunswick (UNB), Canada.
[22] E. P. Thubert, Objective Function Zero for the Routing Protocol for Low-
He worked as an Undergraduate with the Univer-
Power and Lossy Networks (RPL), document RFC 6552, 2012, pp. 5–48.
sidad Politècnica Salesiana (UPS), Ecuador. He is
[23] P. L. O. Gnawali, The Minimum Rank With Hysteresis Objective Function,
document RFC 6719, 2012. currently a full-time Professor with Yachay Tech
[24] OMNeT++ Discrete Event Simulator. Accessed: Jan. 31, 2023. [Online]. University, Urcuqui, Ecuador, and a Postgradu-
Available: https://omnetpp.org/ ate Professor with Pontificia Universidad Católica
[25] H.-S. Kim, H. Cho, H. Kim, and S. Bahk, ‘‘DT-RPL: Diverse bidirectional del Ecuador (PUCE), Sede Manabí, Ecuador. He has been involved in
traffic delivery through RPL routing protocol in low power and lossy several national and international projects and has authored international
networks,’’ Comput. Netw., vol. 126, pp. 150–161, Oct. 2017. publications in conferences and journals. His research interests include the
[26] K. Kritsis, G. Z. Papadopoulos, A. Gallais, P. Chatzimisios, and application of artificial intelligence in wireless multi-hop networks and the
F. Théoleyre, ‘‘A tutorial on performance evaluation and validation development of the IoT smart services. He received the Best Ph.D. Thesis
methodology for low-power and lossy networks,’’ IEEE Commun. Surveys Prize of Information and Communication Technologies from Escola de
Tuts., vol. 20, no. 3, pp. 1799–1825, 3rd Quart., 2018. Doctorat, UPC.

VOLUME 11, 2023 57413

 C. L. D. Santos et al.: ML-RPL: ML-Based Routing Protocol for Wireless Smart Grid Networks

JULIAN CARDENAS BARRERA received the JULIAN MENG (Senior Member, IEEE) received
Ph.D. degree in electrical engineering from Uni- the Ph.D. degree in electrical engineering from
versidad Central Marta Abreu de Las Villas, Queen University, Kingston, ON, Canada, in 1993.
Santa Clara, Cuba, in 1994. He is currently an His research interests include adaptive signal esti-
Associate Professor with the Department of Elec- mation, nonlinear signal processing, renewable
trical and Computer Engineering and the NB energy, and intelligent systems. He is a registered
Power Industrial Research Chair with the Emera Professional Engineer of the Association of Pro-
and NB Power Research Center for Smart Grid fessional Engineers and Geoscientists NB.
Technologies, University of New Brunswick. His
research interests include digital signal processing,
the optimization of renewable energy systems, energy forecasting, and smart
grid communications. He is a registered Professional Engineer with the
Province of New Brunswick, Canada.

EDUARDO CASTILLO GUERRA (Senior Mem-

ber, IEEE) received the B.Sc. degree in electrical
engineering and the M.Sc. degree in telecommu-
nication from the University of ‘‘Marta Abreu’’
Las Villas, Santa Clara, Cuba, in 1992 and
1996, respectively, and the Ph.D. degree in elec-
trical engineering from the University of New
Brunswick in 2003. He is currently a Professor
with the Department of Electrical and Computer
Engineering, University of New Brunswick. His
research interests include modeling digital systems, artificial intelligence,
digital signal and speech processing, the optimization of renewable energy
systems, digital circuit and sensor design, voice authentication, and secure
communications. He is a registered Professional Engineer in the Province of
New Brunswick, Canada.

57414 VOLUME 11, 2023