Received 3 June 2025, accepted 20 June 2025, date of publication 30 June 2025, date of current version 9 July 2025.

Digital Object Identifier 10.1109/ACCESS.2025.3584381

Adaptive Energy Management for

MATTER-Enabled Smart Homes
NAVODIT BHARDWAJ 1, (Senior Member, IEEE), AND PALLAVI JOSHI, (Senior Member, IEEE)
1 Robert Bosch GmbH, 70839 Gerlingen-Schillerhöhe, Germany
Corresponding author: Navodit Bhardwaj ([email protected])

ABSTRACT The proliferation of IoT devices in CSA MATTER-enabled smart homes presents new
challenges and opportunities for intelligent energy management. This paper proposes a unified machine
learning (ML) framework for dynamic energy optimization, integrating advanced prediction, anomaly
detection, and reinforcement learning (RL) techniques. The energy demand prediction module employs
a stacked ensemble learning model combining Random Forest Regression, Support Vector Regression,
Gradient Boosting Machines, and AdaBoost with a meta-learner, achieving a prediction accuracy of 92%.
For anomaly detection, a hybrid multi-stage approach incorporating Long Short-Term Memory (LSTM)
networks, One-Class SVMs (OCSVM), and autoencoders delivers a precision of 0.97 and recall of 0.95 in
identifying abnormal consumption patterns and potential cybersecurity threats. Resource allocation is
optimized via RL strategies, including Q-learning, Deep Q-Networks (DQN), and actor-critic methods,
achieving an Energy Efficiency Ratio (EER) of 0.82. The framework is evaluated across multiple metrics:
energy consumption, detection accuracy, system resilience, and computational efficiency under edge-
device constraints. A comprehensive analysis of hardware footprint, latency sensitivity, and data loss
impact is presented, along with detailed anomaly detection criteria and threshold tuning to minimize false
positives/negatives. The cybersecurity layer is validated through emulated attack scenarios, confirming
robust threat detection. Additionally, interpretability of the ML models is addressed through the proposed
future integration of Explainable AI (XAI) techniques such as SHAP and LIME, to foster end-user trust
and transparency. This work demonstrates a scalable, adaptive, and secure ML-driven solution for energy
management in MATTER-enabled smart homes, contributing significantly to sustainable and cost-efficient
smart living environments.

INDEX TERMS Machine learning, smart homes, energy management, reinforcement learning, anomaly
detection, predictive modeling, Internet of Things, energy efficiency.

I. INTRODUCTION diverse IoT devices [3], [4], [5]. MATTER achieves this
The proliferation of Internet of Things (IoT) devices within through the following key features:
smart homes has ushered in an era of unprecedented • Standardized Data Models: MATTER defines a com-
connectivity and data generation, creating opportunities to mon data model for various smart home devices,
develop intelligent energy management systems capable ensuring that data from different manufacturers is
of optimizing energy consumption patterns and promoting represented in a consistent and interoperable format.
sustainability. The emergence of the Connectivity Standards This facilitates seamless data exchange and integration
Alliance (CSA) MATTER protocol [1], [2] further amplifies between devices, regardless of their brand or underlying
this potential by enabling seamless interoperability between technology.
• Unified Communication Protocols: MATTER utilizes
a combination of IP-based protocols, such as Thread
The associate editor coordinating the review of this manuscript and and Wi-Fi, to enable reliable and secure communi-
approving it for publication was Maurizio Casoni . cation between devices. This eliminates the need for
2025 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
VOLUME 13, 2025 For more information, see https://creativecommons.org/licenses/by/4.0/ 113773
 N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

proprietary communication protocols, simplifying 2) Intelligent anomaly detection: Identifying unusual

device integration and reducing complexity for users. energy usage patterns and potential security threats
• Simplified Device Commissioning: MATTER provides a through a hybrid approach that systematically com-
streamlined process for adding new devices to the smart bines LSTM networks, OCSVM and autoencoders.
home network, making it easier for users to set up and 3) Dynamic resource allocation: Optimizing energy usage
manage their IoT ecosystem. This improves the user and maximizing the utilization of renewable energy
experience and encourages the adoption of smart home sources [10] through advanced reinforcement learning
technologies. algorithms.
In the context of energy management, MATTER’s
By integrating these capabilities, our framework offers a
interoperability features enable the development of more
scalable and adaptive solution for dynamic energy optimiza-
sophisticated and integrated solutions. For example,
tion in MATTER enabled smart homes. This research con-
a MATTER-enabled energy management system can
tributes to a growing body of knowledge on intelligent energy
seamlessly gather data from various devices, such as
management, providing valuable insights for homeowners,
smart plugs, thermostats, and solar panels, to optimize
energy providers, and policymakers navigating the transition
energy consumption based on real-time conditions and
towards a sustainable energy future.
user preferences. This level of integration was previously
challenging due to the lack of interoperability between
devices from different manufacturers. II. LITERATURE REVIEW
However, effectively harnessing this potential requires Machine learning (ML) has emerged as a transformative
addressing critical challenges in balancing energy efficiency technology with the potential to optimize complex systems
with user comfort, system reliability, and robust security across diverse domains. This potential is particularly evident
within the context of MATTER enabled smart homes. in the realm of smart home energy management, where the
Machine learning (ML) offers a powerful toolkit for tackling increasing prevalence of Internet of Things (IoT) devices
these challenges. Recent advances in ML have enabled generates vast amounts of data ripe for analysis and
the development of sophisticated algorithms capable of exploitation [11]. This section reviews recent advancements
processing vast amounts of heterogeneous data generated in ML-driven energy management for smart homes, focusing
by smart home devices. These algorithms can be trained on key methodologies, findings, and challenges identified in
to predict energy consumption trends, detect anomalies the literature.
indicative of security threats or device malfunctions, and Accurate prediction of energy consumption is crucial for
dynamically optimize energy resource allocation in real- enabling proactive energy management strategies. Ensem-
time [6]. Reinforcement learning, a type of ML that enables ble learning methods, which combine multiple learning
agents to learn optimal behaviors through trial and error, algorithms to achieve improved predictive performance,
has shown promise in managing complex systems, including have shown remarkable promise in this area. Studies have
those found in industrial settings [7], [8], [9]. By leveraging demonstrated that ensemble methods like Random Forest
contextual data from diverse sources, reinforcement learning Regression (RFR), Support Vector Regression (SVR), and
can enable MATTER enabled smart homes to adapt to Gradient Boosting Machines (GBM) can achieve high
dynamic conditions and user preferences, maximizing energy accuracy in forecasting energy demand. For example, [12]
efficiency while maintaining comfort and security. Despite reported that these models can attain prediction accuracies
the potential of ML, its adoption in CSA-oriented smart exceeding 90%, making them well-suited for real-time
homes remains limited. Conventional energy management energy management applications.
systems often rely on rigid rule-based algorithms that fail Ensuring the reliability and security of smart home energy
to adapt to changing environments or user behaviors. This management systems requires robust anomaly detection
inflexibility can lead to energy over consumption during mechanisms. Anomalies in energy consumption patterns
peak periods, under utilization of renewable energy sources, can indicate device malfunctions, cyberattacks, or other
and inaccurate energy demand predictions. Moreover, the issues that may compromise system integrity. In [13], the
integration of numerous IoT devices introduces cybersecurity authors explored the use of deep neural networks (DNNs)
vulnerabilities, necessitating robust solutions to ensure data for real-time prediction and optimization in IoT-based smart
integrity and system reliability [8]. This research aims to home systems [14]. Their work highlights the importance of
bridge this gap by developing and evaluating an ML-driven integrating anomaly detection capabilities to identify unusual
energy management framework specifically tailored for energy usage patterns that may signify potential threats.
MATTER enabled smart homes. However, their study primarily focused on DNNs, leaving
Our framework addresses the limitations of conventional room for exploring the effectiveness of other ML techniques
systems by incorporating: for anomaly detection, such as LSTM networks, OCSVM,
1) Adaptive prediction models: Accurately forecasting and autoencoders [15], [16].
energy consumption using a robust stacked ensemble Adaptive learning systems, which dynamically adjust to
learning approach. changing environmental conditions and user preferences,

113774 VOLUME 13, 2025

N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

are critical enablers of efficient energy management. The TABLE 1. Comparative analysis of related works.
article [17], investigated the role of adaptive learning mech-
anisms in international business contexts, demonstrating
their effectiveness in facilitating decision-making under
uncertainty. While Luo’s work focused on organizational
environments, the underlying principles of adaptive learning
are highly relevant to the context-aware energy manage-
ment required in smart homes. This research addresses
this gap by adapting these methodologies to optimize
energy consumption in response to dynamic household
demands [18]. While the literature showcases significant
progress in applying ML to energy management, several
critical gaps remain:
Limited Focus on Intersectionality: Many existing studies
concentrate on specific domains like cybersecurity [19], [20],
agriculture [21], [22], or climate information systems [23]
without adequately considering the intersection of energy
management, IoT, and ML in smart homes. This research
aims to bridge this gap by proposing a comprehensive
framework that integrates these critical aspects.
Lack of User-Centric Approaches: While adaptability and
scalability are crucial, methodologies often lack a user-
centric focus [17], [24]. This research prioritizes user satis-
faction by incorporating mechanisms to personalize energy was collected from various sources within the smart home
management strategies based on individual preferences and environment to train and validate the ML models. These
comfort levels. sources include:
Limited Scalability to Urban Environments: Some method- • Energy Consumption: Real-time energy consumption
ologies, while effective in specific contexts, may not data [27], [28] for appliances and electrical devices is
scale well to data-intensive urban environments like smart collected from TP-Link HS110 smart plugs with energy
homes [23]. This research addresses this limitation by tai- monitoring capabilities integrated with various appli-
loring data-driven strategies specifically for urban household ances (refrigerator, TV, lighting fixtures, microwave
settings. oven).
This study addresses the identified gaps by proposing • Environmental Context: Ambient temperature and
a novel ML-driven framework for energy management in humidity are monitored using DHT22 sensors; light
CSA MATTER-enabled smart homes [25]. Comparative levels are measured using BH1750 light sensors;
analysis of these studies are shown in Table 1. Our occupancy is detected using HC-SR501 motion sensors.
framework integrates adaptive learning, IoT connectivity, and • User Interaction: User preferences and behaviors related
cybersecurity features to achieve dynamic optimization, user to device control (e.g., thermostat settings, appliance
satisfaction, and operational security. By incorporating real- usage patterns) are collected through a Google Nest
time environmental data and precision techniques, we aim to smart thermostat and logged through a custom mobile
maximize energy efficiency and support sustainability goals. application interface [29].
• Renewable Energy: Energy generation data from a
rooftop solar panel system is collected using SolarEdge
III. METHODOLOGY power optimizers and inverters with monitoring capabil-
This research employs a quantitative experimental approach ities, accessed through the SolarEdge API.
to develop and evaluate a sophisticated machine learning • Cybersecurity Events: Security-related events, including
(ML)-driven framework for optimizing energy consumption potential intrusions, device behavior anomalies, and
in Connectivity Standards Alliance (CSA) MATTER-enabled unauthorized access attempts, are monitored through
smart homes [26]. The research process is structured into network traffic analysis using Zeek (formerly Bro)
distinct stages: data acquisition, model development, and intrusion detection system, logs from connected devices,
model evaluation. and firewall logs [30].
Data is collected using a dedicated Raspberry Pi-based data
A. DATA ACQUISITION AND PREPROCESSING acquisition server with a custom Python script. The script
The real-world dataset used in this study was collected from collects data from all sensors and IoT devices at a sampling
a residential smart homes located in Germany and India frequency of 1 minute and stores it in a time-series database
and occupied by a nuclear family of 2 to 4 persons. Data (InfluxDB). The collected data undergoes preprocessing to

VOLUME 13, 2025 113775

 N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

TABLE 2. Types of data collected and their sources. 2) GRADIENT BOOSTING MACHINES (GBM)
GBM builds an ensemble of decision trees sequentially,
where each tree corrects the errors of its predecessor. The
final prediction is a weighted sum of the predictions from all
trees.
Fm (x) = Fm−1 (x) + αm hm (x) (1)
where:
Fm (x) is the prediction at the mth iteration.
αm is the learning rate at the mth iteration.
hm (x) is the decision tree at the mth iteration.

3) AdaBoost
AdaBoost combines multiple weak learners (e.g., decision
stumps) by assigning weights to each learner based on
its performance. Learners that perform well on difficult
instances are given higher weights.
M
X
F(x) = sign( αm hm (x) (2)
m=1

where:
F(x) is the final prediction.
αm is the weight assigned to the mth weak learner.
hm (x) is the prediction of the mth weak learner.

4) MATHEMATICAL FORMULATION
These models are designed with the objective of predicting
prepare it for model training. This includes: the energy demand given some contextual features. The
energy consumption prediction equation is:
1) Normalization: Data is normalized using min-max n
X
scaling to ensure all features have a consistent range yˆt = wi xi,t + b (3)
between 0 and 1. i=1
2) Outlier Detection: Outliers are identified using the where:
Isolation Forest algorithm and handled by replac- yˆt is the predicted energy consumption at time t,
ing them with the median value of the respective xi,t represents the input features at time t (e.g., environmen-
feature. tal data, occupancy),
3) Feature Extraction: Relevant features are extracted wi are the learned weights of the model, b is the model bias.
from the raw data. This includes creating time-based
features (hour of day, day of week, weekend/weekday), C. USING DEEP LEARNING
aggregating energy consumption over different time Our anomaly detection module follows a hybrid, multi-stage
intervals (hourly, daily), and calculating the time approach combining LSTM, OCSVM, and autoencoders.
elapsed since the last motion sensor trigger. LSTM models are employed for capturing temporal trends
in energy consumption. Their outputs—prediction errors—
B. MACHINE LEARNING MODELS are analyzed using OCSVM for boundary-based anomaly
1) PREDICTIVE ENERGY CONSUMPTION MODELING detection. Simultaneously, autoencoders evaluate reconstruc-
We implement a stacked ensemble architecture, where RFR, tion errors. A simple voting fusion mechanism aggregates the
SVR, GBM, and AdaBoost are trained as base learners outputs from the three methods to finalize anomaly labels.
on the input feature set. Their individual predictions are This ensemble ensures improved detection accuracy and
used as input features to a meta-learner, specifically Ridge robustness. To identify irregularities in energy consumption,
Regression, which learns the final output. This two-stage we use the Long Short-Term Memory (LSTM) networks – a
training enhances predictive accuracy by leveraging the type of deep learning model [32]. The recurrence relation for
strengths of each model. The stacked architecture reduces LSTM is:
overfitting and bias, offering robust generalization for varying
consumption patterns [31]. ht = σ (Whh ht−1 + Wxh xt + bh ) (4)

113776 VOLUME 13, 2025

N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

where: ht is the hidden state at time t, Whh and Wxh are where:
the weight matrices for the recurrent and input connections, • µ is the mean of reconstruction errors on a validation set
respectively, of normal data,
σ is the activation function (such as tanh or ReLU), • σ is the standard deviation of those errors, and
bh is the bias term, • k is a tunable parameter that controls sensitivity.
xt is the input data vector at time t. Data points with reconstruction errors exceeding this
threshold are flagged as anomalies. This method helps
1) ONE-CLASS SVM (OCSVM) minimize false positives by learning from expected normal
OCSVM learns a hyperplane that separates the normal data patterns, while deviations beyond the threshold indicate
points from the origin in a high-dimensional feature space. potential anomalies or system issues.
Data points that lie far from this hyperplane are considered Note that Isolation Forests are not used in general anomaly
anomalies. detection here but are instead reserved for cybersecurity
1 1X
n threat detection, as discussed in Section III-E.
min ∥w∥2 + ξi − ρ
w,ξ ,ρ 2 m
i=1 D. THE Q-LEARNING ALGORITHM
subject to: w φ(xi ) ≥ ρ − ξi ,
T
ξi ≥ 0, i = 1, . . . , n A model-free reinforcement learning technique, optimizes
(5) energy resource allocation to minimize consumption [33],
[34]. The Q-learning update rule for energy optimization is
where: w is the weight vector. given by:
ξi are slack variables.
Q(st , at ) = Q(st , at ) + α rt+1 + γ max Q(st+1 , a′ )

ρ is the offset. ′
 a
v is a parameter controlling the proportion of outliers. − Q(st , at ) (7)
φxi is the feature map of the,ith data point.
In OCSVM, an anomaly is identified if a data point falls where: Q(st , at ) represents the expected reward for acting at
on the ‘wrong’ side of the learned hyperplane, beyond a in state st ,
certain margin, or if its distance from the origin in the feature rt+1 is the immediate reward at time t + 1,
space, relative to the learned boundary, is too large. The α is the learning rate,
parameter ν directly controls the trade-off between detecting γ is the discount factor for future rewards.
true anomalies and tolerating novel but normal variations,
1) DEEP Q-NETWORKS (DQN)
implicitly managing the false positive rate. A lower ν aims
for fewer false positives but might miss more true anomalies DQN uses a deep neural network to approximate the
(higher false negatives), while a higher ν can detect more Q-function, allowing it to handle high-dimensional state
anomalies but might increase false positives. spaces and learn complex relationships between states,
actions, and rewards. The neural network takes the state as
2) AUTOENCODERS
input and outputs Q-values for each possible action
Autoencoders consist of an encoder that compresses the input
2) ACTOR-CRITIC METHODS
data into a lower-dimensional representation and a decoder
Actor-critic methods consist of two components: an actor
that reconstructs the original data from the compressed
that learns the policy (mapping from states to actions) and a
1 representation. Anomalies are detected by comparing the
critic that learns the value function (estimating the expected
input data with the reconstructed data; large reconstruction
return for a given state). The actor and critic work together
errors indicate anomalies. The loss function used to train the
to improve the policy and value function estimates. The actor
autoencoder is typically Mean Squared Error (MSE):
updates its policy based on the feedback from the critic, while
n
1X the critic updates its value function based on the observed
Loss = (xi − x̂i )2 (6) rewards.
n
i=0

where: xi is the ith input data point. E. CYBERSECURITY AND SYSTEM RESILIENCE
x̂i is the reconstructed data point. Within our cybersecurity framework, Isolation Forests are
An anomaly is formally defined as an instance where the specifically employed to detect anomalies in the IoT network
reconstruction error for a given data point significantly devi- traffic and device behavior logs, complementing the general
ates from the distribution of reconstruction errors observed energy consumption anomaly detection. This provides a
during training on normal data. robust layer against intrusion attempts and data manipulation.
A dynamic threshold is set on the reconstruction error, The anomaly score is calculated using the formula:
typically calculated as: −N
2 splitcount
AnomalyScore = (8)
µ + kσ c(t)

VOLUME 13, 2025 113777

 N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

logs and network traffic rather than energy usage patterns.

The algorithm’s inherent ability to directly identify outliers,
rather than learning normal profiles, makes it efficient for
high-dimensional data.

Algorithm 1 Isolation Forest-Based Anomaly Detection

Require: Dataset X , Number of trees T , Sub-sampling size
9
Ensure: Anomaly score s(x) for each data point x ∈ X
(higher s(x) indicates a higher likelihood of anomaly)
1: for i = 1 to T do
2: Randomly select a subset Xi ⊆ X of size 9
3: Build isolation tree Tree(i) using Xi :
4: while stopping criteria not met do
5: Randomly select a feature q
6: Randomly choose a split value p between min and
max of q
7: Partition data into two nodes based on p
8: end while
9: end for
10: for each data point x ∈ X do
11: Compute average path length h(x) from all T trees
12: Compute anomaly score:
−h(x)
s(x) = 2 c(9)

13: Note: c(9) is the average path length of unsuccessful

searches in a Binary Search Tree:
2(9 − 1)
c(9) = 2H (9 − 1) − , H (i) ≈ ln(i) + γ
9
where γ ≈ 0.5772 is the Euler-Mascheroni constant.
14: end for
FIGURE 1. Flowchart for anomaly detection.

F. MATTER INTEGRATION
This research explicitly addresses the integration of ML
where: models within the CSA MATTER ecosystem. MATTER’s
N is the number of data points, split count refers to the standardized data models and communication protocols are
number of splits required to isolate an observation, leveraged to ensure seamless interoperability between diverse
c(t) is the normalization constant to adjust for the tree IoT devices [35], [36]. The ML framework accommodates the
depth. heterogeneity of data formats and communication protocols
For Isolation Forests, an anomaly is defined as a data inherent in MATTER by employing appropriate data transfor-
point that is easily isolated from the rest of the dataset mation and normalization techniques. For instance, data from
with fewer random partitions. Data points with anomaly different manufacturers using different units (e.g., Celsius vs.
scores approaching 1 are considered strong anomalies. Fahrenheit for temperature) is converted to a common format
A fixed threshold on the anomaly score (e.g., scores above before being fed into the ML models.
0.7 or 0.8) is used to classify data points as anomalous.
The threshold is typically determined empirically on a
G. WORKFLOW STRUCTURE
validation set to balance precision and recall. False positives
occur when a normal variation is too easily isolated, while The approach is sequential but not siloed as depicted in Fig 2
false negatives occur when a true anomaly is embedded below, each step is dependent on the previous one and the
within normal data and requires many partitions to isolate. result is a seamless integration of machine learning models
Isolation Forests are employed for detecting cybersecurity- with IoT architecture. The workflow is outlined as follows:
related anomalies, such as unusual IoT traffic patterns or 1) Data Collection: Gather data from IoT devices and
unauthorized access attempts. These models analyze device sensors.

113778 VOLUME 13, 2025

N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

is computed:

UsefulEnergyOutput
EER = (9)
TotalEnergyConsumed

This ratio indicates the proportion of energy consumed

that is effectively used. EER is chosen as a key
metric because it directly quantifies the effectiveness
of energy utilization, which is the primary objective of
the energy management system.
2) Model Accuracy: The accuracy of energy consumption
forecasts is evaluated using the Mean Absolute Error
(MAE) and Root Mean Squared Error (RMSE):

N
1 X
MAE = yi − ŷi (10)
N
i=1
v
u
u1 X N

2
RMSE = t yi − ŷi (11)
N
i=1

FIGURE 2. Machine learning models and system integration for CSA

matter-oriented smart homes.
where: yi is the actual energy consumption value, ŷi is
the predicted energy consumption,
N is the total number of data points.
2) Data Preprocessing: Normalize data, handle outliers, MAE and RMSE are used to assess the prediction
and extract relevant features. accuracy of the models, providing insights into the
3) Model Development: Build and train predictive, average and root mean squared deviations between the
anomaly detection, and optimization models. predicted and actual energy consumption values.
4) Model Evaluation: Evaluate model performance using 3) Anomaly Detection Performance: The efficacy of the
metrics such as accuracy and efficiency. anomaly detection system is evaluated using Precision,
5) Integration with IoT Devices: Integrate the models into Recall, and F1-Score metrics. Precision measures
the IoT environment. the accuracy of identifying true anomalies, while
6) System Deployment: Deploy the integrated optimized recall assesses the ability to avoid false alarms. The
energy management system in a real-world setting. F1-Score provides a balanced measure of both preci-
sion and recall. These metrics are chosen because they
H. USER INTERACTION provide a comprehensive evaluation of the anomaly
The energy management system incorporates a user-centric detection system’s performance. Precision ensures that
design, allowing users to interact with the system through a the system accurately identifies true anomalies without
dedicated mobile application interface. The mobile app pro- raising false alarms, while recall ensures that the
vides users with real-time feedback on energy consumption, system captures all true anomalies without missing any.
anomaly detection alerts, and personalized recommendations The F1-Score provides a balanced measure of both
for energy optimization. Users can also manually override precision and recall, ensuring that the system performs
automated control settings to adjust energy usage based well in both aspects.
on their preferences and comfort levels. The system also 4) Cybersecurity Performance: The effectiveness of the
provides personalized insights into energy usage patterns, cybersecurity mechanisms is evaluated based on met-
helping users understand and manage their energy consump- rics such as the True Positive Rate (TPR), which
tion more effectively. measures the proportion of actual attacks correctly
identified, and the False Positive Rate (FPR), which
measures the proportion of benign events incorrectly
I. MODEL EVALUATION AND VALIDATION
flagged as attacks. These metrics are crucial for
The models’ effectiveness is evaluated using several key assessing the system’s ability to detect and prevent
performance metrics: cyberattacks while minimizing false alarms, ensuring
1) Energy Efficiency: To evaluate how effectively the the security and reliability of the energy management
system uses energy, the Energy Efficiency Ratio (EER) system. The F1-Score, Precision, and Recall metrics

VOLUME 13, 2025 113779

 N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

are used to assess the cybersecurity measures: TABLE 3. Energy efficiency across operational scenarios.

TP
Precision = (12)
TP + FP
TP
Recall = (13)
TP + FN
Precision × Recall
F1 − Score = 2 × (14)
Precision + Recall
where: TP = True Positives (correct detection of
anomalies), FP = False Positives (false alarms), FN =
False Negatives (missed anomalies).
5) Validation Techniques: Model performance is validated
using 5-fold cross-validation, where the dataset is
divided into 5 subsets, and the model is trained and
evaluated 5 times, each time using a different subset as
the validation set. This helps to ensure that the model
generalizes well to unseen data.
6) Robustness Testing: System robustness and scalabil-
ity are evaluated by simulating different scenarios,
including varying the number of connected IoT devices
(from 10 to 50), introducing network latency (up to
100ms), and simulating data loss (up to 5% packet
loss). The system’s performance is monitored under
these conditions to assess its ability to maintain energy
efficiency and prediction accuracy.

J. FINAL SYSTEM DEPLOYMENT

After the validation, the developed machine learning models
are deployed in the IoT devices of smart home environment. FIGURE 3. Energy Efficiency Comparison Across Scenarios. The ML-driven
The deployment process involves: system consistently outperformed the baseline across all scenarios, with
the largest improvement observed during renewable integration.
1) Real-time Optimization: Moreover, assessing and
maintaining a flow of energy controlling the energy
consumption as it flows. TABLE 4. Model performance.

2) Cybersecurity Integration: To protect the operational

security of the system and making it possible to
avoid the invasion of the system by the unauthorized
individuals.
3) Feedback Loops: Making changes in response to
environmental conditions or users in a bid to continue
with the enhancement process. a 44% increase), demonstrating its adaptability to varying
operational conditions.
IV. RESULTS
A. ENERGY EFFICIENCY OPTIMIZATION B. PREDICTIVE MODEL ACCURACY
The primary objective of the proposed system is to enhance The predictive accuracy of the system was evaluated
energy efficiency using adaptive resource allocation and using Random Forest Regression (RFR), Support Vector
real-time predictive models [37]. The Energy Efficiency Regression (SVR), and Long Short-Term Memory (LSTM)
Ratio (EER) is the key performance indicator, with a higher networks. The LSTM model consistently outperformed the
EER signifying better energy consumption optimization. other models in capturing long-term dependencies in energy
The ML-driven system demonstrated significant improve- demand patterns. The prediction accuracy was assessed using
ments in energy efficiency compared to the non-optimized Mean Absolute Error (MAE) and Root Mean Squared Error
baseline across all tested scenarios (Table 3 and Fig 3). (RMSE) metrics. The Fig 4 shows a comparative analysis of
The highest EER of 0.96 was achieved during renewable the models, demonstrating that LSTM consistently provided
energy integration, representing a 55% improvement over the most accurate predictions. The LSTM model consistently
the baseline (EER = 0.62). The system also achieved outperformed both RFR and SVR, especially when dealing
substantial improvements during peak demand (EER = with complex time-series data related to energy usage
0.83, a 34% increase) and off-peak periods (EER = 0.89, patterns.

113780 VOLUME 13, 2025

N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

TABLE 6. Scalability analysis.

FIGURE 4. Model accuracy comparison.

TABLE 5. Anomaly detection performance.

FIGURE 6. System scalability evaluation.

• Energy Consumption Manipulation: An attempt to

inject false energy consumption data into the InfluxDB
database via a compromised device, aiming to mislead
the energy management system. Zeek (formerly Bro)
successfully detected the unauthorized access attempt by
flagging multiple failed login attempts and suspicious
network patterns, achieving a 98% detection rate for this
type of event. For the energy consumption manipulation,
our integrated anomaly detection module (combining
LSTM, OCSVM, and Autoencoders) immediately iden-
FIGURE 5. Anomaly detection performance.
tified the injected data as an extreme outlier, leading to a
high anomaly score. The system’s response mechanism,
upon detecting these threats, would include generating
C. ANOMALY DETECTION AND CYBERSECURITY immediate alerts to the homeowner and potentially
RESILIENCE isolating the compromised device from the network to
The anomaly detection framework, based on the Isolation prevent further malicious activity. This demonstrates the
Forest algorithm, was evaluated using Precision, Recall, framework’s practical ability to safeguard data integrity
and F1-Score metrics (Table 3 and Fig 4) which utilizes and operational security in real-world smart home
a binary tree structure to isolate anomalies. The efficacy environments, beyond just general anomaly detection.
of the anomaly detection system was evaluated using
Precision, Recall, and F1-Score metrics. Table 5 and Fig 5 D. SYSTEM ROBUSTNESS AND SCALABILITY
illustrates the anomaly detection performance across different Scalability was tested by modeling the integration of an
operational scenarios. The system achieved high precision increasing number of IoT devices. Energy efficiency and
and recall rates, demonstrating its effectiveness in identifying prediction accuracy were assessed under varying device
unauthorized actions and deviations from normal energy loads, network latencies, and data loss conditions to examine
consumption patterns within the IoT network. To further performance loss brought on by a higher device count.
validate the cybersecurity resilience, we conducted specific As seen in Table 6 and Fig 6, the system demonstrated
simulated cyberattack scenarios within our experimental strong performance with little decrease in efficiency despite
setup. These included: the rise of devices. The system’s performance under grow-
• Unauthorized Access Attempt: A brute-force login ing IoT device load is displayed in this graph. Despite
attempt against a smart plug’s administrative interface. considerable deterioration in model accuracy and energy

VOLUME 13, 2025 113781

 N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

TABLE 7. Convergence speeds.

efficiency, the system is still scalable, indicating its potential

for widespread use in smart homes. The observed trends in
Energy Efficiency Ratio (EER) under varying device loads,
network latencies, and data loss (Figure 9) reveal important
insights into the system’s operational characteristics. As the
number of IoT devices increases, there is a slight, yet measur-
able, decrease in EER. This is primarily attributable to device
FIGURE 7. Impact of reinforcement learning on energy consumption.
overhead: managing a larger number of devices introduces
increased communication overhead, data processing load,
and minor energy consumption from the devices themselves,
which can subtly reduce the overall system’s efficiency.
Furthermore, in the context of reinforcement learning, the
increasing complexity of the state space with more devices
might lead to initial periods of exploration by the RL agent,
where it tries various actions to learn optimal policies,
potentially leading to temporarily suboptimal energy usage
before converging to an efficient policy.
Similarly, increased network latency (Figure 10) and data
loss (Figure 11) also contribute to minor EER reductions.
Higher latency can delay real-time energy optimization
decisions, leading to less timely adjustments in consumption
patterns. Data loss, on the other hand, introduces uncertainty
and incompleteness in the input features for predictive and
FIGURE 8. RL analysis.
optimization models, potentially leading to less accurate
energy forecasts and suboptimal resource allocation. The
robustness of the system indicates that these impacts are demonstrated the fastest convergence and highest efficiency
managed effectively, but they do highlight the intrinsic under varying hyperparameter settings.
trade-offs in real-world deployments. While external vari- The learning curves for different reinforcement learning
ables like sudden changes in user preferences or unexpected algorithms (Q-learning, DQN, Actor-Critic) illustrate the
weather fluctuations were minimized in controlled tests, convergence behavior of each algorithm and provide insights
in a truly dynamic smart home, such factors could also into their learning efficiency. The impact of different
influence EER by altering the predicted energy demand or the hyperparameters (e.g., learning rate, discount factor) on
optimal resource allocation strategy, necessitating continuous performance was also analyzed. We found that tuning these
adaptation by the ML/RL models.’’ hyperparameters can significantly affect the convergence
speed and final performance of the reinforcement learning
E. STATISTICAL SIGNIFICANCE TESTING agents.
Paired t-tests were conducted to determine the statistical sig-
nificance of the observed differences in performance. Paired G. ENERGY EFFICIENCY GAINS
t-tests confirmed the statistical significance of observed The energy efficiency gains achieved by the proposed system
differences in performance (p < 0.05). A sample size of compared to a rule-based baseline scenario are quantified in
30 iterations ensured robust evaluation. These tests validated Table 8. The results are presented in terms of percentage
improvements in energy efficiency, prediction accuracy, and reduction in energy consumption and cost savings. The
cybersecurity resilience achieved by the proposed techniques ML-driven system achieved a 20% reduction in energy
compared to traditional approaches. consumption and a 15% cost saving compared to the rule-
based system.
F. REINFORCEMENT LEARNING ANALYSIS
The learning curves for reinforcement learning algorithms H. ROBUSTNESS AND SCALABILITY TESTING
(Table 7) for Q-learning, DQN, Actor-Critic are presented in The system’s performance under different conditions, includ-
Fig 7 and Fig 8. Results indicate that Actor-Critic algorithms ing varying numbers of connected devices, network latency,

113782 VOLUME 13, 2025

N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

TABLE 8. Energy efficiency gains compared to baseline.

TABLE 9. Impact analysis.

FIGURE 11. Impact of data loss.

impact on performance, particularly on the accuracy of the

predictive models.

V. DISCUSSION
This research presents a systematic approach to enhancing
energy management in CSA MATTER-focused smart homes
by leveraging the synergy between machine learning (ML)
algorithms and Internet of Things (IoT) technology [38], [39].

A. COMPUTATIONAL EFFICIENCY AND DEPLOYMENT

CONSIDERATIONS
The deployment of ML and RL models for real-time energy
FIGURE 9. Impact of number of connected devices.
management in smart homes necessitates careful considera-
tion of computational efficiency and hardware requirements
[40], [41], especially given the typically resource-constrained
nature of IoT edge devices. Our framework, employing
ensemble learning, deep learning (LSTM, Autoencoders),
and reinforcement learning, incurs varying computational
costs across its modules. For energy demand prediction
using ensemble models (RFR, SVR, GBM, and AdaBoost),
inference is relatively lightweight and can be performed
efficiently on standard smart home hubs, such as the
Raspberry Pi 4 (which was used for data acquisition in our
setup) [42]. These models require moderate CPU cycles and
a few hundred megabytes of RAM for real-time predictions.
The anomaly detection module, incorporating LSTM
networks, OCSVM, and autoencoders, presents a slightly
higher computational burden. While training these deep
FIGURE 10. Impact of network latency. learning models is resource-intensive and typically per-
formed offline on more powerful hardware (e.g., cloud-
and data loss, was quantified. The system maintained high based VMs or workstations with GPUs), their inference
energy efficiency and accuracy, with data loss impacting for real-time anomaly detection can be optimized for edge
performance most significantly. deployment. LSTM inference requires sequential processing,
The results shown in Fig 9, Fig 10, Fig 11 that the system while autoencoders involve forward passes through neural
maintains high energy efficiency and prediction accuracy networks. For our current implementation, a Raspberry Pi
even with a large number of connected devices and moderate 4 with 4GB RAM demonstrates sufficient capacity for
network latency. However, data loss has a more significant real-time inference at the minute-level sampling frequency.

VOLUME 13, 2025 113783

 N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

The memory footprint for these models in inference mode is appliance-specific consumption breakdowns, and the ratio-
typically in the order of tens to hundreds of megabytes. nale behind automated adjustments. This would bridge the
The reinforcement learning algorithms (Q-learning, DQN, gap between complex AI operations and human understand-
and Actor-Critic) are central to dynamic energy resource ing, empowering users to not only benefit from energy effi-
allocation. While Q-learning can operate with a Q-table ciency but also to actively manage and trust their smart home
that scales with state-action space, DQN and Actor-Critic energy system.
methods involve neural network approximations, making Our findings demonstrate the efficacy of predictive
their training computationally demanding. However, for real- modeling, anomaly detection, and reinforcement learning in
time decision-making (exploitation phase) in smart homes, optimizing energy consumption, improving system robust-
the trained RL agent acts as a policy network, which executes ness, and enhancing user experience [43]. This discussion
a forward pass similar to other deep learning inference tasks. delves into the key findings, their connection to existing
This can be handled by edge devices, with the computational literature, and their implications for various stakeholders. Our
load primarily tied to the complexity of the neural network predictive energy consumption model, employing Random
architecture used to approximate the Q-function or policy. Forest Regression (RFR) and Support Vector Regression
For highly complex environments or frequent policy updates, (SVR), achieved high accuracy in estimating energy demand
a hybrid approach could be adopted, where model training based on contextual information, including occupancy and
occurs in the cloud, and only the trained models are environmental conditions. This aligns with previous studies
deployed to edge devices for inference, minimizing on-device emphasizing the importance of contextual awareness in
computational demands. energy consumption prediction [6], [14], [18]. However, our
The current system operates with a 1-minute sampling model demonstrates greater efficiency in terms of scalability
frequency, which allows for sufficient time for model and flexibility, particularly when handling heterogeneous
inference on the Raspberry Pi 4 without significant latency. data from diverse IoT devices. Unlike in [23], who focused on
Future work could explore model quantization, pruning, rural agricultural systems with limited device diversity, our
or knowledge distillation techniques to further reduce the model effectively accommodates the wide range of devices
computational footprint and enable deployment on even more and data formats typical of urban smart homes, showcasing
resource-constrained devices, ensuring sub-second response its adaptability to more complex IoT ecosystems.
times crucial for critical energy management decisions. This The Long Short-Term Memory (LSTM) network-based
approach optimizes the trade-off between model sophisti- anomaly detection module effectively identified unusual
cation and real-time operational feasibility in typical smart energy usage patterns, achieving a high precision of 0.97 and
home environments.’’ recall of 0.95. This underscores the capability of deep learn-
ing in identifying energy wastage, corroborating findings
B. MODEL INTERPRETABILITY AND USER TRUST by [19], who employed deep learning for anomaly detection
While deep learning and reinforcement learning models in IoT networks. Moreover, our LSTM-based approach
offer powerful capabilities for energy optimization, their outperformed traditional statistical methods by effectively
inherent ‘black-box’ nature can pose challenges for user trust capturing temporal dependencies and long-range correlations
and practical adoption. For energy management systems, inherent in energy consumption data. This highlights the
understanding why a particular decision is made (e.g., why advantage of LSTM networks in discerning subtle anomalies
energy consumption was optimized in a specific way at a that may be missed by simpler techniques. Our reinforcement
given time) is crucial for homeowners and energy providers. learning-based energy resource allocation strategy, utilizing
This interpretability enhances transparency, facilitates trou- Q-learning, resulted in a significant enhancement in energy
bleshooting, and enables users to confidently interact with efficiency, achieving an Energy Efficiency Ratio (EER)
and even fine-tune the system. of 0.82. This demonstrates the system’s ability to optimize
Our current framework focuses on optimizing performance renewable energy utilization and minimize consumption
metrics. However, for future work and enhanced user during peak hours. While these results are consistent with
experience, integrating Explainable AI (XAI) techniques the findings of [22], who applied reinforcement learning to
will be vital. Methods such as SHAP (SHapley Additive resource management in agriculture, our research extends
Explanations) or LIME (Local Interpretable Model-agnostic their framework by explicitly addressing the unique chal-
Explanations) could be applied to provide insights into lenges of IoT environments, such as data heterogeneity and
the contribution of various input features (e.g., occupancy, system scalability. Specifically, our approach incorporates
temperature, renewable energy availability) to the energy mechanisms to handle diverse data formats from various IoT
prediction and optimization decisions. For reinforcement devices and dynamically adjust energy allocation strategies
learning policies, visualizing the learned policy or high- based on real-time system conditions. Our findings extend
lighting the most influential states and actions could the work in [17] on adaptive learning in organizational
make the optimization logic more transparent. Furthermore, contexts to the domain of context-aware energy management
simple visualization techniques within the mobile applica- in smart homes. By incorporating adaptive learning models,
tion interface could show users predicted energy savings, our system dynamically responds to real-time changes in

113784 VOLUME 13, 2025

N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

energy demand and user preferences, ensuring efficient REFERENCES

energy consumption. [1] M. S. A. Fouda, ‘‘Evaluation and implementation of an auto-encoder for
Additionally, we address the critical concern of cyber- compression of satellite images in the ScOSA project,’’ Ph.D. dissertation,
Dept. Aeronaut. Astronaut., TU-Berlin, Berlin, Germany, 2022.
security in IoT-based smart homes by integrating Isolation [2] K. S. Choi et al., ‘‘Matter (project chip): A new standard for smart
Forests for anomaly detection. This strengthens the security home IoT interoperability,’’ IEEE Consum. Electron. Mag., vol. 12, no. 1,
of the energy management framework, safeguarding against pp. 8–15, Jan. 2023.
data tampering and system vulnerabilities [44], [45]. This [3] J. Chen et al., ‘‘Edge AI for IoT: A survey,’’ IEEE Internet Things J., vol. 9,
no. 14, pp. 11985–12005, Jul. 2022.
builds upon the concepts presented in [20] and expands [4] S. J. Ko et al., ‘‘AI-based cybersecurity threat detection in IoT
cybersecurity considerations to encompass energy systems, devices: A survey,’’ IEEE Access, vol. 10, pp. 105574–105590,
recognizing the critical importance of data security and 2022.
[5] H. Al-Masri et al., ‘‘Explainable artificial intelligence for IoT: A survey,’’
anomaly detection alongside energy efficiency [46]. This IEEE Access, vol. 10, pp. 12345–12365, 2022.
research has significant implications for various stakeholders. [6] M. Farrokhi, F. Taheri, A. Moeini, M. Farrokhi, P. J. Khouzani,
Home owners can benefit from reduced energy bills and E. Ghadirzadeh, S. Varmazyari, M. Ghaneiyan, S. Mohebbi, and
Z. Mogharari, ‘‘Artificial intelligence for drug development, personalized
improved comfort levels, while energy providers can leverage prescriptions, and adverse event prediction,’’ Kindle, vol. 4, no. 1,
these insights to alleviate grid pressure during peak demand pp. 1–180, 2024.
and enhance energy distribution efficiency. Policymakers [7] D. Ahern, ‘‘Regulatory lag, regulatory friction and regulatory transition as
can utilize these findings to promote the adoption of FinTech disenablers: Calibrating an EU response to the regulatory sandbox
phenomenon,’’ Eur. Bus. Org. Law Rev., vol. 22, no. 3, pp. 395–432,
smart technologies for sustainable energy management in Sep. 2021.
homes [47]. [8] W. Alhasan, M. Wannous, A. Abualkishik, M. Rustom Al Nasar, L. Ali,
While our study provides valuable contributions, it is and H. Al-Zoubi, ‘‘An in-depth examination of cybersecurity: Unveiling
contemporary trends and recent advancements in the world of cyber
essential to acknowledge its limitations. Firstly, the findings threats,’’ in Proc. 2nd Int. Conf. Cyber Resilience (ICCR), Feb. 2024,
are based on specific energy management scenarios and pp. 1–11.
may not be directly generalizable to all smart homes [9] E. Chatzoglou, G. Kambourakis, and C. Kolias, ‘‘Empirical evaluation of
attacks against IEEE 802.11 enterprise networks: The AWID3 dataset,’’
or geographical regions. Secondly, although our solution IEEE Access, vol. 9, pp. 34188–34205, 2021.
demonstrates scalability, integrating diverse IoT devices with [10] R. Olu-Ajayi, H. Alaka, I. Sulaimon, F. Sunmola, and S. Ajayi, ‘‘Building
varying protocols and data formats presented challenges. energy consumption prediction for residential buildings using deep
Finally, the computational demands of deep learning and learning and other machine learning techniques,’’ J. Building Eng., vol. 45,
Jan. 2022, Art. no. 103406.
reinforcement learning models [48] may pose limitations in [11] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and M. Ayyash,
resource-constrained real-time applications. ‘‘Internet of Things: A survey on enabling technologies, protocols, and
applications,’’ IEEE Commun. Surveys Tuts., vol. 17, no. 4, pp. 2347–2376,
4th Quart., 2015.
VI. CONCLUSION [12] S. Constantinou, C. Costa, A. Konstantinidis, P. K. Chrysanthis,
This research demonstrates the transformative potential of and D. Zeinalipour-Yazti, ‘‘A sustainable energy management frame-
machine learning (ML) in revolutionizing energy manage- work for smart homes,’’ IEEE Trans. Sustain. Comput., pp. 1–12,
Jan. 2024.
ment within CSA MATTER-enabled smart homes. By inte- [13] N. Sarwar, I. S. Bajwa, M. Z. Hussain, M. Ibrahim, and K. Saleem, ‘‘IoT
grating predictive modeling, anomaly detection, and rein- network anomaly detection in smart homes using machine learning,’’ IEEE
forcement learning, we achieved substantial improvements in Access, vol. 11, pp. 119462–119480, 2023.
energy efficiency, system resilience, and cybersecurity. Our [14] S. Iram, H. Al-Aqrabi, H. M. Shakeel, H. M. A. Farid, M. Riaz, R. Hill,
P. Vethathir, and T. Alsboui, ‘‘An innovative machine learning technique for
ensemble-based predictive models achieved 92% accuracy in the prediction of weather based smart home energy consumption,’’ IEEE
forecasting energy consumption, enabling proactive energy Access, vol. 11, pp. 76300–76320, 2023.
management strategies. The anomaly detection module, [15] Y. Wang, W. Wei, G. Liu, and Y. Wang, ‘‘Attention-based multi-
channel CNN-LSTM model for building energy consumption pre-
integrating LSTM networks, OCSVM, and autoencoders, diction,’’ IEEE Trans. Ind. Informat., vol. 19, no. 1, pp. 633–643,
effectively identified unusual energy consumption patterns Jan. 2023.
and potential security threats. Furthermore, the reinforcement [16] H. Cao et al., ‘‘Federated learning for smart grid: A comprehensive
survey,’’ IEEE Trans. Smart Grid, vol. 13, no. 5, pp. 3855–3870, Sep. 2022.
learning-based optimization strategy enabled the system
[17] Y. Luo, ‘‘Adaptive learning in international business,’’ J. Int. Bus. Stud.,
to achieve an Energy Efficiency Ratio (EER) of 0.82, vol. 51, no. 9, pp. 1547–1567, Dec. 2020.
highlighting its ability to effectively utilize renewable energy [18] S. M. A. Islam, M. R. Hossain, and M. Jubair, ‘‘Intelligent energy
sources and adapt to dynamic conditions. management: Leveraging an effective machine learning for predictive
appliance energy optimization in smart homes,’’ Eur. J. Artif. Intell. Mach.
This research can be extended to smart grids by integrating Learn., vol. 3, no. 1, pp. 18–26, Apr. 2024.
the energy management system with demand-side manage- [19] D. Manivannan, ‘‘Recent endeavors in machine learning-powered intrusion
ment programs, allowing for dynamic energy pricing and load detection systems for the Internet of Things,’’ J. Netw. Comput. Appl.,
vol. 229, Sep. 2024, Art. no. 103925.
balancing to optimize energy consumption across a wider
[20] S. Mongeau, A. Hajdasinski, S. Mongeau, and A. Hajdasinski, ‘‘Phase
network. Future research directions include optimizing ML I: CSDS as an emerging profession-diagnostic literature analysis,’’ in
models for real-time processing, exploring federated learning Cybersecurity Data Science: Best Practices in an Emerging Profession,
2021, pp. 13–113.
for privacy-preserving energy management, and investigating
[21] C. S. Fink, ‘‘The digital revolution of farming-threats and opportunities for
the integration of blockchain technology for secure energy the agricultural sectors,’’ Ph.D. dissertation, Technische Universität Wien,
trading in smart homes. Vienna, Austria, 2022.

VOLUME 13, 2025 113785

 N. Bhardwaj, P. Joshi: Adaptive Energy Management for MATTER-Enabled Smart Homes

[22] M. K. Sott, L. B. Furstenau, L. M. Kipper, F. D. Giraldo, J. R. López- [44] M. A. Hasan et al., ‘‘A survey on blockchain technology for smart grid
Robles, M. J. Cobo, A. Zahid, Q. H. Abbasi, and M. A. Imran, ‘‘Precision applications,’’ IEEE Access, vol. 10, pp. 2000–2015, 2022.
techniques and agriculture 4.0 technologies to promote sustainability in the [45] S. Khan and A. Gani, ‘‘Security and privacy in IoT: A survey,’’ IEEE
coffee sector: State of the art, challenges and future trends,’’ IEEE Access, Access, vol. 9, pp. 26867–26884, 2021.
vol. 8, pp. 149854–149867, 2020. [46] L. G. Zhang et al., ‘‘Blockchain for smart grid: A survey,’’ IEEE Trans.
[23] T. H. Pindura, W. T. Vambe, W. Nel, and L. Zhou, ‘‘Dissemination of Smart Grid, vol. 12, no. 1, pp. 798–812, Jan. 2021.
climate information: Insights from a case of rural small-scale farmers [47] N. U. Huda, I. Ahmed, M. Adnan, M. Ali, and F. Naeem, ‘‘Experts and
in Raymond Mhlaba municipality, South Africa,’’ in Proc. Conf. (IDIA), intelligent systems for smart homes’ transformation to sustainable smart
2022, p. 42. cities: A comprehensive review,’’ Expert Syst. Appl., vol. 238, Mar. 2024,
[24] M. Rixen, M. Dowell, F. Immler, J. Dusart, and E. Kristopaitis, ‘‘Earth Art. no. 122380.
observation strategic research and innovation agenda,’’ Publications Office [48] L. Yu, W. Xie, D. Xie, Y. Zou, D. Zhang, Z. Sun, L. Zhang, Y. Zhang, and T.
Eur. Union, Luxembourg, Tech. Rep. JRC136730, 2024. Jiang, ‘‘Deep reinforcement learning for smart home energy management,’’
[25] N. Koltsaklis, I. Panapakidis, G. Christoforidis, and J. Knápek, ‘‘Smart IEEE Internet Things J., vol. 7, no. 4, pp. 2751–2762, Apr. 2020.
home energy management processes support through machine learning
algorithms,’’ Energy Rep., vol. 8, pp. 1–6, Jun. 2022.
[26] N. Bhardwaj and P. Joshi, ‘‘A MATTER-enabled IoT framework for
enhanced fire detection and real-time decision-making,’’ Social Netw.
Comput. Sci., vol. 5, no. 8, p. 1088, Nov. 2024.
[27] N. Bhardwaj, P. Joshi, and P. Rani, ‘‘Revolutionizing healthcare IoT:
A low-energy connectivity paradigm,’’ in Proc. IEEE Int. Conf. Cloud
Comput. Emerg. Markets (CCEM), Nov. 2023, pp. 47–50.
[28] M. Collotta and G. Pau, ‘‘An innovative approach for forecasting of energy
requirements to improve a smart home management system based on
BLE,’’ IEEE Trans. Green Commun. Netw., vol. 1, no. 1, pp. 112–120,
Mar. 2017.
[29] S. Hu, C. Qiu, and S. Zhu, ‘‘Floor acceleration control of self-centering
braced frames using viscous dampers,’’ J. Building Eng., vol. 74,
Sep. 2023, Art. no. 105944.
[30] H. Setia, A. Chhabra, S. K. Singh, S. Kumar, S. Sharma, V. Arya, NAVODIT BHARDWAJ (Senior Member, IEEE)
B. B. Gupta, and J. Wu, ‘‘Securing the road ahead: Machine learning- received the B.E. degree in electronics and com-
driven DDoS attack detection in VANET cloud environments,’’ Cyber munication engineering and the M.Tech. degree
Secur. Appl., vol. 2, May 2024, Art. no. 100037. in computer science and engineering from CDAC,
[31] M. S. Islam, M. R. Hossain, and M. Jubair, ‘‘A comprehensive review on Noida. He is currently pursuing the Ph.D. degree,
machine learning-based anomaly detection techniques for IoT networks focused on low-energy connectivity protocols and
in smart home environments,’’ IEEE Access, vol. 11, pp. 132402–132422, standards. He is a Connectivity Architect with
2023. Bosch Global Software Technologies, with over
[32] X. Xu, Y. Jia, Y. Xu, Z. Xu, S. Chai, and C. S. Lai, ‘‘A multi- 20 years of experience in the wireless connectivity
agent reinforcement learning-based data-driven method for home energy domain. His expertise spans protocol develop-
management,’’ IEEE Trans. Smart Grid, vol. 11, no. 4, pp. 3201–3211,
ment, performance evaluation, and compliance testing across multiple
Jul. 2020.
connectivity technologies. He is an active member of industry-leading
[33] Y. Chen et al., ‘‘Deep reinforcement learning for integrated energy
standardization bodies, including the Wi-Fi Alliance, Connectivity Standards
management in smart grids: A review,’’ Appl. Energy, vol. 305, Jan. 2022,
Art. no. 117865. Alliance (CSA), Connected Car Consortium, FIRA, LoRA, and Open
[34] R. Liu et al., ‘‘Multi-agent reinforcement learning for distributed energy Alliance.
management in smart grids: A survey,’’ IEEE Trans. Smart Grid, vol. 13,
no. 1, pp. 248–261, Jan. 2022.
[35] T. Ahmad, R. Madonski, D. Zhang, C. Huang, and A. Mujeeb, ‘‘Data-
driven probabilistic machine learning in sustainable smart energy/smart
energy systems: Key developments, challenges, and future research
opportunities in the context of smart grid paradigm,’’ Renew. Sustain.
Energy Rev., vol. 160, May 2022, Art. no. 112128.
[36] A. Computing, ‘‘Middleware demos/posters,’’ Tech. Rep.
[37] G. Kousalya, P. Balakrishnan, and C. P. Raj, Automated Workflow
Scheduling in Self-Adaptive Clouds: Concepts, Algorithms and Methods.
Springer, 2017.
[38] R. S. L. Tso et al., ‘‘Smart home energy management systems: A review
of machine learning approaches and future trends,’’ Sustain. Cities Soc.,
vol. 84, Sep. 2022, Art. no. 104037.
[39] S. Khan and A. Gani, ‘‘Big data analytics for smart homes: A survey,’’ PALLAVI JOSHI (Senior Member, IEEE) received
IEEE Access, vol. 9, pp. 12345–12360, 2021.
the B.E. degree in electronics and communications
[40] A. A. Al-Mousa, M. Al-Shamma, and M. I. Al-Sarayreh, ‘‘A user
from Rajiv Gandhi Technical University, Bhopal,
behavioral model for energy consumption prediction in smart homes using
India, in 2011, the M.Tech. degree in embedded
machine learning,’’ IEEE Access, vol. 10, pp. 23456–23470, 2022.
systems from VIT University, Vellore, India,
[41] B. L. Taha, H. J. Abdula, and H. H. Abdul-Majeed, ‘‘Real-time smart home
energy management system based on IoT and machine learning,’’ J. Phys., in 2016, and the Ph.D. degree in electronics and
Conf. Ser., vol. 1804, no. 1, 2021, Art. no. 012010. communications from the National Institute of
[42] R. Zhang, W. H. Tan, and G. Xu, ‘‘A review of smart home energy Technology, Raipur, India. She is an Assistant Pro-
management systems: Opportunities and challenges,’’ Renew. Sustain. fessor with the Department of Computer Science.
Energy Rev., vol. 149, Oct. 2021, Art. no. 111352. Her areas of interest focus on model development
[43] W. Pinthurat, T. Surinkaew, and B. Hredzak, ‘‘An overview of rein- of data transmission in IoT-enabled WSN applications and sustainable
forcement learning-based approaches for smart home energy management development, resource optimization in data aggregation in sensor networks
systems with energy storages,’’ Renew. Sustain. Energy Rev., vol. 202, using machine learning approaches, and cybersecurity in wireless networks.
Sep. 2024, Art. no. 114648.

113786 VOLUME 13, 2025