1

Generative AI for Energy Harvesting Internet of

Things Network: Fundamental, Applications, and
Opportunities
Wenwen Xie, Geng Sun, Jiahui Li, Jiacheng Wang, Hongyang Du, Dusit Niyato, Fellow, IEEE,
Octavia A. Dobre, Fellow, IEEE

Energy harvesting technology that collect energy from the

arXiv:2408.08496v1 [cs.NI] 16 Aug 2024

Abstract—Internet of Things (IoT) devices are typically pow-

ered by small-sized batteries with limited energy storage capacity, environment to power or store for future use can provide a
requiring regular replacement or recharging. To reduce costs promising solution for IoT networks [2]. One feasible method
and maintain connectivity in IoT networks, energy harvesting
technologies are regarded as a promising solution. Notably, due is to harvest energy from renewable natural sources such as
to its robust analytical and generative capabilities, generative solar, wind, and vibrational energy. The benefit of this method
artificial intelligence (GenAI) has demonstrated significant po- is energy friendliness, which reduces the burden on the envi-
tential in optimizing energy harvesting networks. Therefore, we ronment and promotes the development of green technology
discuss key applications of GenAI in improving energy harvesting and sustainability. However, the supply of renewable natural
wireless networks for IoT in this article. Specifically, we first
review the key technologies of GenAI and the architecture of energy can be affected by weather conditions and time, which
energy harvesting wireless networks. Then, we show how GenAI may lead to instability of the network. Unlike renewable natu-
can address different problems to improve the performance of the ral energy sources, radio frequency (RF) energy is more stable,
energy harvesting wireless networks. Subsequently, we present making RF energy transmission and collection technology an
a case study of unmanned aerial vehicle (UAV)-enabled data alternative method for powering the next generation of wireless
collection and energy transfer. The case study shows distinctively
the necessity of energy harvesting technology and verify the networks.
effectiveness of GenAI-based methods. Finally, we discuss some RF energy can be broadly classified into two types. One
important open directions. is ambient RF, which refers to the RF signals that are not
Index Terms—Generative AI, energy harvesting, UAV, diffusion intended for RF energy transfer but for other purposes such
model, optimization. as data transmission. The other one involves dedicated RF
sources that can be used to deliver energy to network nodes
where a reliable energy supply is necessary. Because the
I. I NTRODUCTION dedicated RF source is fully controllable, it is suited for
With the rapid development of the Internet of Things for supporting network applications with quality of service
(IoT), it has spawned a number of domains such as smart (QoS) requirements. Moreover, the dedicated RF sources can
transportation, healthcare and environmental monitoring [1]. include mobile units like unmanned aerial vehicles (UAVs) that
In these domains, a large number of smart devices or sensors are capable of periodically moving and delivering RF energy
are typically deployed to collect surrounding environment to network nodes, thereby improving the quality of energy
information for further analysis and decision-making, which is transfer service. [3].
crucial for the intelligent control. However, these IoT devices Note that optimizing energy harvesting wireless networks
are typically energy-constrained in practice, which greatly from the perspectives of resource allocation and network
limits their availability. It is worth noting that traditional deployment is crucial for enhancing overall network perfor-
methods of manually replacing batteries and wired charging mance. However, conventional optimization strategies, such as
increase the maintenance effort and operating costs. Therefore, convex optimization, have notable limitations. Firstly, tradi-
it is especially important to find a new solution to provide IoT tional optimization lacks adaptability, which makes it strug-
devices with continuous energy supply to ensure the reliable gle to effectively address dynamic optimization in energy
network operation. harvesting wireless networks. Secondly, due to the inherent
uncertainties and unpredictability in energy harvesting wireless
W. Xie and J. Li are with the College of Computer Science networks, handling situations where network states or param-
and Technology, Jilin University, Changchun 130012, China (e-mail:
[email protected], [email protected]). eters are unknown poses a significant challenge for traditional
G. Sun is with the College of Computer Science and Technology, Jilin Uni- optimization. Therefore, identifying a more suitable technique
versity, Changchun 130012, China, and also with the College of Computing for energy harvesting wireless networks is of paramount
and Data Science, Nanyang Technological University, Singapore 639798 (e-
mail: [email protected]). importance.
J. Wang, H. Du and D. Niyato are with the College of Computing and Data With the recent popularity of ChatGPT and Sora, gener-
Science, Nanyang Technological University, Singapore 639798 (e-mail: ji- ative artificial intelligence (GenAI) technology has garnered
[email protected], [email protected], [email protected]).
Octavia A. Dobre is with the Faculty of Engineering and Applied Science, widespread attention, particularly in the realm of multimedia
Memorial University, Canada (e-mail: [email protected]). content generation. It is worth noting that the capabilities of
 2

GenAI can be attributed to its following characteristics: 1) Solar Energy: Solar energy harvesting typically involves
• Data Augmentation: GenAI can generate highly similar converting sunlight into electricity using photovoltaic cells,
and authentic data by learning training samples, which which is highly effective in sunny outdoor environments.
helps to alleviate data scarcity. For example, GenAI can Consequently, solar energy harvesting is widely applied in
produce more dataset to train UAV trajectory optimiza- remote sensing, environmental monitoring, and agricultural
tion. IoT systems.
• Latent Representation Learning: GenAI can learn the 2) Wind Energy: Wind energy harvesting typically employs
distribution of training data in order to mine the latent wind turbines to convert the kinetic energy of wind into
features and patterns of the data, which is conducive electricity, which is particularly suitable for areas with abun-
to dimensionality reduction of high-dimensional data. dant wind energy resources. Moreover, wind power generation
For instance, GenAI can extract most important channel devices can be installed near wireless network base stations or
features of UAV network environments. nodes to provide them with sustained power supply.
• Knowledge Transfer: GenAI can transfer the learned 3) Other Energy: Other forms of renewable energy also
knowledge among domains by unifying the distribution hold potential in energy harvesting for wireless networks.
of newly generated data to the given data. For exam- For example, thermoelectric generators convert temperature
ple, GenAI can apply the energy harvesting knowledge differences into electricity, while piezoelectric materials gen-
learned in urban environments to other environments. erate electricity through mechanical vibrations. These energy
Thanks to these characteristics, GenAI is highly suitable sources are commonly found in industrial environments and
for energy harvesting wireless networks. For instance, GenAI can provide power support for industrial wireless sensors
can predict the effects of different antenna configurations on widely.
energy harvesting and network coverage, thereby identifying The most notable advantage of renewable natural energy
the optimal antenna design solutions [4]. Moreover, the robust is its environmental friendliness. However, renewable natural
analytical capability of GenAI can be employed to estimate energy still has limitations, such as geographic constraints or
channel conditions, which is crucial for energy harvesting instability due to environmental conditions. The emergence of
wireless networks [5]. Although integrating GenAI into energy RF energy harvesting offers a new approach to alleviate such
harvesting wireless networks offers significant advantages, limitations, especially dedicated RF sources.
there exist several issues that require further investigation.
Motivated by this, we provide a comprehensive tutorial to B. RF Energy Harvesting
present important applications of GenAI on energy harvesting
RF energy harvesting allows wireless devices to harvest
wireless networks. The contributions of our work can be
energy from RF signals for their information processing and
summarized as follows:
transmission. This technique becomes a promising solution
• We first present fundamentals of common GenAI models.
for energy-constrained wireless networks. Note that the RF
Subsequently, we discuss energy harvesting technologies
sources can be classified into two types, i.e., dedicated RF
based on renewable natural source and RF energy source
sources and ambient RF sources.
with their issues that can be addressed by GenAI tech-
• Dedicated RF Sources: Dedicated RF sources are specifi-
nologies. Furthermore, we demonstrate the basic archi-
tecture of an energy harvesting wireless network. cally designed and implemented to provide RF energy for
• We explore how GenAI can address an optimization prob-
energy harvesting purposes. These sources are often tai-
lems in energy harvesting wireless networks from multi- lored to deliver consistent and predictable energy output
ple perspectives, i.e., channel estimation, relay topology at certain frequencies and power levels to ensure efficient
design, antenna design, secure cognitive radio communi- energy transfer to the network devices.
• Ambient RF Sources: Ambient RF sources refer to the
cation, and renewable energy generation forecasting.
• We introduce a case study on UAV energy harvesting and
naturally occurring or existing RF signals in the envi-
information transfer. Experimental results demonstrate ronment, which are not specifically designed for energy
the effectiveness of GenAI in energy harvesting wireless harvesting. These sources include RF emissions from
networks. various communication and broadcasting systems, such as
TV, radio towers, and WiFi routers. It is worth noting that
II. OVERVIEW OF E NERGY H ARVESTING ambient RF sources can be also regarded as renewable
Energy harvesting wireless networks represent a significant energy sources.
advancement for wireless communication, where devices can Both dedicated RF sources and ambient RF sources play
self-sustain by harvesting energy from their environment. an important role in RF energy harvesting wireless networks.
Specifically, dedicated RF sources provide efficient and con-
A. Renewable Natural Energy Harvesting trollable energy transfer, and they are suitable for the applica-
Renewable natural energy refers to energy sources that tions that require a stable energy supply. Moreover, ambient
can be naturally replenished and sustainably utilized [6], [7]. RF sources utilize existing RF signals to provide a flexible
Currently, renewable natural energy is a significant energy energy harvesting method for low-power devices, and they
source for rechargeable wireless devices, and common types are suitable for a wide range of urban and portable device
of renewable energy are as follows: applications.
 3

energy harvesting are elaborated. Moreover, we introduce

Renewable how GenAI can address an optimization problems in energy
Energy Source harvesting wireless networks.
Information
Renewable Transmission Zone
Energy Flow A. Overview of GenAI Models
Information Flow The emergence of GenAI has led AI technology to a
Network Device whole new stage and brought new perspectives to a number
of areas, including media content generation, system design,
RF Energy and wireless network optimization. In the following, several
Harvesting Zone common GenAI models are described in detail.
1) Generative Adversarial Network (GAN): GANs consist
Information of two neural networks: a generator and a discriminator. The
Gateway generator creates artificial data from random noise, while the
RF Energy Flow
discriminator attempts to distinguish between real and artificial
data. Through adversarial training, the generator network
generates high-quality data. GANs excel at generating highly
RF Energy realistic data but suffer from training instability and mode
Harvesting Zone
collapse that the generator keeps producing limited similar
RF Energy data patterns. GANs are primarily used in image generation,
Source data augmentation, and network optimization.
2) Diffusion Model: Diffusion models gradually add noise
Fig. 1. Energy Harvesting Architecture. The information gateway is respon- to data through a forward diffusion process until it becomes
sible for data transmission with the network devices. The energy transmitter
is responsible for delivering energy to the network devices, where the energy pure noise, and then denoise it to recover the data through a
sources include renewable natural energy and RF energy. reverse diffusion process. Note that the reverse process can use
neural networks to denoise step-by-step. Diffusion models can
produce high-quality data, making them suitable for image and
C. Energy Harvesting Architecture
speech generation. Moreover, diffusion models can be utilized
As shown in Fig. 1, the general infrastructure-based archi- to improve the policy network of deep reinforcement learning
tecture of energy harvesting wireless network consists of the (DRL), which has been applied to several domains and has
following three components: shown superior performance.
• Information Gateway: Information gateways, such as base 3) Variational Autoencoder (VAE): VAEs consist of an
stations and relay devices, can serve as communication encoder and a decoder. The encoder maps input data to a
hubs in energy harvesting wireless networks. Moreover, probabilistic distribution of latent variables, while the decoder
they are responsible for managing and coordinating the reconstructs the data from these variables. VAEs are trained
communication processes across an entire network. In- by minimizing reconstruction error and the Kullback Leibler
formation gateways typically have continuous and fixed (KL) divergence between the latent distribution and a standard
power supplies and can provide power support to wireless normal distribution. They excel at creating continuous latent
devices. spaces and explicitly modeling generation probabilities, mak-
• Energy Source: The energy source acts as the energy ing them suitable for image generation and anomaly detection.
provider in the energy harvesting wireless network, sup- 4) Transformer: The Transformer model consists of mul-
plying the necessary wireless power to network nodes. tiple stacked encoder and decoder layers, each containing
Note that, in some special cases, the information gateway self-attention mechanisms and feed-forward neural networks.
and the RF energy source can coexist on the same device. The encoder transforms input sequences into context-aware
• Network Nodes/Devices: Network nodes refer to user representations, while the decoder generates the next sequence
devices within the network, responsible for executing element based on the output and the previous output of the
specific application tasks. They can obtain sustainable decoder. Transformers handle long-range dependencies well,
power supplies from the energy transmitter to extend their offer high training efficiency, and have a flexible structure,
operational time, and they can also communicate with the making them widely used in natural language processing tasks
information gateway. such as machine translation and text generation.
Note that infrastructure-less energy harvesting network has
a similar architecture to the infrastructure-based one shown in
Fig. 1, except that the network nodes communicate directly B. Reasons for GenAI on Energy Harvesting Optimization
with each other. Most current work on energy harvesting wireless network
focuses on how to design energy harvesting devices or systems
III. G ENAI- ENABLED E NERGY H ARVESTING N ETWORK from hardware perspectives [8]–[10]. It is equally important
In this section, we first present several common GenAI to improve the performance of the entire energy harvesting
models. Subsequently, the reasons for using GenAI to optimize wireless network by optimizing the allocation of certain key
 4

resources. In this case, traditional methods such as evolu- under challenging low signal-to-noise ratio (SNR) conditions.
tionary algorithm and convex optimization have the following Moreover, at an SNR of 5 dB, the CGAN channel estimator
limitations in energy harvesting wireless networks: consistently surpasses existing CE solutions, achieving an 82%
• Poor Real-time Performance: In dynamic environments, improvement over CRLD and MMSE estimators.
traditional optimization methods (e.g., evolutionary algo- 2) Relay Topology Design: Relay-based structures are par-
rithm) are difficult to quickly respond and adapt to envi- ticularly suitable for energy harvesting wireless networks.
ronmental changes, thus affecting real-time performance First, cooperative relaying techniques enhance network effi-
optimization. ciency and reliability by using relay nodes to overcome fading
• Strong Model Dependency: Convex optimization requires and attenuation. Second, energy-rich nodes can transmit data
an accurate mathematical model, and the uncertainty and on behalf of energy-constrained nodes to maximize overall
dynamic changes in the actual network environment make network efficiency. Therefore, through proper deployment and
it difficult to describe the model accurately. energy management of relay nodes, the network can address
Note that GenAI has powerful analyzing ability, generating energy harvesting uncertainties, extend node lifetimes, and
ability, and dynamic environment adaptability, which provides improve data transmission reliability.
new approaches to solve the limitations above. Therefore, we The authors in [11] considered an uplink transmission
demonstrate how to improve the performance of an energy scenario in an relay-based IoT sensor network, where each
harvesting wireless network from different perspectives by wireless node self-powers by harvesting energy from power
using GenAI techniques. beacons (PBs) and stores remaining energy in a limited-
capacity rechargeable battery. The authors aimed to optimize
the topology matrix among nodes to maximize the transmis-
C. Roles of GenAI in Energy Harvesting IoT Networks sion bits of all nodes. Given the lack of datasets and the nonlin-
1) Channel Estimation: The recent proliferation of con- ear nature of the optimization problem, the authors proposed
nected devices has led to a substantial surge in spectrum an unsupervised relay topology algorithm based on a VAE.
utilization and energy consumption. As a solution, ambient Specifically, a random latent vector is input into the VAE, and
backscatter communication (AmBC) has garnered consider- subsequently the analytical and generative capabilities of the
able attention. Note that an AmBC node needs to harvest VAE decoder are utilized to generate the topology matrix. To
energy to supply its backscatter operations, i.e., circuit energy evaluate quality of the topology matrix generated by the VAE,
consumption. In this case, channel estimation is crucial in the authors introduced a packet-tracking evaluation method
AmBC, since it can help to optimize energy utilization. Specif- inspired by ray-tracing techniques. This method can accurately
ically, by accurately estimating the channel state, key resources evaluate the output of the VAE, providing appropriate guidance
such as frequency allocation can be dynamically adjusted to for VAE training.
extend the working life of wireless devices. However, tradi- The VAE-based topology approach was compared with tra-
tional channel estimation methods, including the techniques ditional schemes such as the direct connect scheme, minimum
such as pilot-based estimators, blind estimators, and semi- spanning tree (MST) scheme, and greedy scheme, as well
blind estimators, typically face several issues such as requiring as the optimal brute-force solution. The experiment results
extensive prior knowledge, being limited by scenario changes, indicate that the MST scheme exhibits insufficient adaptability
and exhibiting sensitivity to model matching. Therefore, ad- to various IoT network configurations, and the brute-force
vanced machine learning algorithms for channel estimation are search algorithm has excessively long computation times.
of great interest. In contrast, the VAE-based topology approach demonstrates
In this context, a novel channel estimation algorithm based better performance than all other considered approaches across
on conditional GAN (CGAN) for AmBC was proposed [5]. all node counts and network configurations, being the closest
The CGAN estimator can estimate the wireless channel status to the optimal solution.
and quality by accurately capturing complex features and 3) Antenna Design: Antennas are the core components
utilizing stronger reflection paths. This capability allows the of energy harvesting devices, which means that the antenna
CGAN estimator to adapt well to the changes in the relative design is critical in energy harvesting wireless networks. By
strength of reflected signals, resulting in superior channel adjusting the antenna design in terms of frequency selectivity,
estimation performance. Specifically, the proposed channel directivity, and structure, we can optimize signal reception and
estimation method based on CGAN applies specific channel minimize energy loss and interference. Consequently, an ideal
state information (CSI) as a condition to GAN, and the and well-considered antenna design can maximize the ability
generator network generates synthetic channel data that closely of devices to harvest RF energy from the environment.
resembles real-world conditions. This synthetic data in turn The authors in [4] investigated to enhance the RF energy
facilitates accurate estimation of AmBC channel coefficients. harvesting and power transfer with GAN-optimized antenna
The performance of the proposed CGAN estimator is eval- design. Inspired by the quasi-Yagi antenna (QYA) model, the
uated by comparing it with the well-known methods such as authors aimed to develop a cube-shaped three-dimensional
least squares (LS), minimum mean-squared error (MMSE), (3D) antenna optimized for 915 MHz by analyzing various
blind, deep residual learning denoiser (CRLD), and convolu- 2D QYA configurations. First, to enhance robustness of the
tional neural network (CNN). The results show that CGAN analysis, the authors employed GAN to expand the dataset.
channel estimator excels in normalized MSE (NMSE), even Specifically, the generator network generates synthetic QYA
 5

data, while the discriminator network engages in adversarial 5) Renewable Energy Generation Forecasting: Since the
training to optimize the performance of the generator network. renewable energy is an intermittent generation, it is subject
Moreover, GAN was also used to optimize the design of to external factors such as weather variations and changes in
the 3D QYA. A well-trained GAN can effectively capture user behavior, which often bring new challenges to the energy
the complex relationships between QYA parameters and QYA harvesting solutions. In this context, accurate forecasting of
structure, generating target designs based on specified criteria. power generation in advance is essential to improve the
To strengthen this process, the authors integrated the GAN operational efficiency of the energy harvesting systems, which
with the CST Studio Suite for virtual performance testing of can be used in wireless networks.
GAN-generated designs to enable timely adjustments of the A VAE-based method for solar power generation forecasting
GAN model, thereby enhancing energy capture from ambient was proposed [13]. VAE can effectively reduce the dimen-
and dedicated RF sources. sionality of parameters while retaining important data features,
The experimental results show that the minor deviation which is highly beneficial for enhancing the computational ef-
between the GAN and CST Studio Suite results is within an ficiency and generalization capability of the prediction model.
acceptable range, underscoring the reliable performance of the The proposed VAE-based process involves mapping high-
antenna. Moreover, compared to other 2D QYA designs, the dimensional data to a low-dimensional latent space and then
GAN-designed 3D antenna exhibits superior energy harvesting reconstructing the data from the latent space. Therefore, by
capability. optimizing the reconstruction error and the probability distri-
4) Secure Cognitive Radio Communication: Cognitive ra- bution of latent variables, VAE can learn a low-dimensional
dio, through dynamic spectrum management and spectrum representation of the data. Building on the VAE, a Bayesian
sharing, effectively alleviates spectrum scarcity issues, ensur- bidirectional long short-term memory (BiLSTM) network was
ing efficient network operation. Given that cognitive radio incorporated for time series prediction. BiLSTM can capture
devices frequently engage in spectrum sensing and dynamic temporal dependencies and bidirectional information in the
adjustments, traditional battery power typically fails to meet data, further improving prediction accuracy.
their long-term operational needs. Energy harvesting can pro- Compared with benchmark methods, the VAE-based ap-
vide continuous power support for these devices. Moreover, proach shows significant improvement in parameter reduction.
the benefit of energy harvesting can be further enhanced by Specifically, in six months of solar power generation data, the
optimizing the allocation of resources in the energy harvesting number of weight parameters was reduced from 764224 to
network. 2022, achieving a quantized parameter reduction of 97.35%
In this case, the authors in [12] developed the GAN-based and a computation time improvement of 37.93%.
DRL algorithm to enhance the physical security in energy
harvesting cognitive radio system. In the considered system, D. Lesson Learned
the communication security of the secondary user (SU) can
be tampered by eavesdroppers, prompting the deployment of a From the applications above, we can find that GenAI im-
cooperative jammer to secure wireless transmissions. Notably, proves the performance of the energy harvesting IoT networks
both the SU and the jammer can harvest energy from the from the following perspectives.
received RF signals emitted by the primary transmitter (PT), • Data Augmentation: Leveraging the generative capabili-

with batteries capable of storing energy for extended periods. ties of GenAI, it can be used to produce data that closely
The authors employed deep Q-network (DQN) algorithm to resembles the training samples, thereby addressing the
maximize the average minimum secrecy rate and minimize issue of dataset scarcity [4].
the average maximum secrecy outage probability (SOP). To • Solution Exploration: GenAI can leverage its powerful

better estimate the state-action values, GAN was introduced learning and reasoning capabilities to find solutions for
to learn the optimal state-action value distribution for DRL. optimization problems [5], [11]. Additionally, GenAI can
Specifically, the generator network G is responsible for cre- be employed to enhance other algorithms, such as DRL,
ating the estimated state-action value distribution, while the to improve their performance [12].
′
target generator network G produces the target state-action • Dimensionality Reduction: GenAI can extract the latent

value distribution. The role of the discriminator network D is representation variables of data to perform dimensionality
to differentiate between the values produced by G and those reduction while retaining essential information [13].
′
from G .
The experimental results indicate that, compared to the case IV. C ASE S TUDY: G ENAI- ENABLED E NERGY
with one eavesdropper, the secrecy rate achieved by DQN H ARVESTING W IRELESS N ETWORK
decreases by approximately 78% and the SOP increases by In this section, we present a case study to evaluate the
67% when there are ten eavesdroppers. This suggests that the performance of GenAI on energy harvesting wireless network.
DQN approach is highly sensitive to the number of eavesdrop-
pers. In contrast, the GAN-powered DQN reduces the secrecy
rate by 20% and improves the SOP by 67%. Moreover, the A. Scenario Description
maximum average SOP achieved with GAN-powered DQN 1) Motivation: To improve online data analysis and
closely matches the theoretical value of the closed expression, decision-making in IoT networks, it is essential to effectively
which validates the effectiveness of resisting eavesdropping. gather the latest information produced by IoT devices. To
 6

Environment
Critic

Experience Replay Memory

UAV Location

Experience
Space
Target Critic Networks

State
Critic Networks
AoI of Devices

Tuple
Soft
Energy of Devices

Mini-batch
Update

Update
Energy UAV Trajectory
𝑸’𝟏 𝑸’𝟐

Action
𝑸𝟐

Space
𝑸𝟏
Data Transfer Device Scheduling
Collection Evaluate Improve
Charging Time
Policy Policy
Actor
+1 +1

Actor Network
Diffusion Model-based Actor Network

Gaussian

Target
Action
State 𝑠

Noise
Soft
+1 𝑷𝑻 ... 𝑷𝒕 ... 𝑷𝟏
1 +1 Update
Action 𝑎 Reverse Diffusion Chain

Fig. 2. The framework of case study. The UAV is dispatched to charge a set of IoT devices and collect data from them to minimize AoI. Diffusion model is
used to improve the actor network of TD3 algorithm to generate high quality decisions.

precisely measure information freshness, the concept of Age where the inverse process of the diffusion model is used to
of Information (AoI) has been introduced [14]. AoI is defined improve the actor network of TD3. Specifically, instead of the
as the duration from when the most recent update packet is traditional Multi-Layer Perception (MLP), the actor network
created at the source until it is received at the destination. Note can be regarded as a denoiser, which starts from Gaussian
that enhancing the power supply for energy-constrained IoT noise and progressively recovers the optimal decision solution
devices is essential for improving AoI performance in IoT net- according to the environment conditions.
works. However, traditional methods of energy supply, such as
battery replacements or wired connections, are impractical and C. Simulation Settings
labor-intensive. Note that using UAVs for energy transmission • Environment and Tools: The server used for the simula-
has emerged as a promising solution to overcome the energy tion is equipped with an Intel i9 13900K CPU and an
limitations of IoT devices. UAVs offer flexibility and mobility, NVIDIA RTX 4090 GPU, and the RAM is a 32 GB
allowing them to reach remote areas that traditional power DDR4. Moreover, the simulation code is implemented by
sources might not be able to access. Therefore, UAVs can Pytorch.
provide stable energy transfer in an on-demand basis, reducing • Simulation Parameters: The critic network and the de-
downtime for IoT devices and ensuring the freshness of data noising network of the diffusion model both include two
collection and continuity of communication. hidden layers, each containing 128 neurons. Moreover,
2) Problem Formulation: As shown in Fig. 2, multiple IoT the activation function for the critic network is Mish, and
devices equipped with energy buffers are randomly distributed the activation function for the final layer of the denoising
within the target area to sense surrounding environment. A network is Tanh. The size of experience replay buffer and
UAV is moving within the target area, responsible for charging mini-batch are set to 106 and 128, respectively. Moreover,
the IoT devices to ensure their activity and receiving data the learning rate is configured to 0.0003.
from the IoT devices. Note that only one IoT device can
upload data during each time slot. We aim to minimize the
D. Performance Analysis
AoI by optimizing the trajectory of UAV, the scheduling of
IoT devices, and the charging time. The AoI convergence curves for various algorithms are
illustrated in Fig. 3. We can find that the AoI of IoT de-
vices becomes higher when the UAV do not provide energy
B. Proposed Design transfer services. This is due to the fact that uploading data
Due to the mobility of the UAV, our considered scenario consumes energy, and IoT devices have limited energy storage.
involves highly dynamic nature, which makes the traditional Therefore, when the energy buffer becomes empty, they are
optimization methods difficult for handling such problems. unable to continue uploading data, leading to an increased
Notably, DRL enables agents to adapt their behavior quickly AoI. Moreover, we can observe that the proposed DTD3 al-
through trial-and-error learning, thus effectively adapting to gorithm achieves better AoI performance and converges more
the changing environment of UAV-assisted IoT systems. rapidly and stably than the traditional TD3 algorithm. This
Among the many DRL algorithms, the twin delayed deep performance improvement can be attributed to the diffusion
deterministic policy gradient (TD3) algorithm stands out for model. First, the analytical ability of diffusion model allows
its excellent stability [15]. for a better exploration of the latent relationship between
To further enhance the performance of the TD3 algorithm, states and actions. Second, by leveraging the generative and
we propose a diffusion model-based TD3 (DTD3) algorithm, data augmentation abilities of the diffusion model, DTD3
 7

 VI. C ONCLUSION

'7' In this article, we have introduced how GenAI optimizes the
 7' energy harvesting wireless network. Specifically, we have first
'7'1R(QHUJ\+DUYHVWLQJ presented the fundamentals of GenAI, different types of energy

source, and the architecture of energy harvesting network.
$YHUDJH$R,

 Then, we have discussed several GenAI-based methods to

address various issues of energy harvesting wireless network.
 Subsequently, we have conducted a case study on UAV-
enabled energy transfer and data collection to demonstrate the
 importance of energy harvesting technology in the IoT network
and validate the effectiveness of GenAI. Finally, three key
 future directions have been shown that can further improve
GenAI on energy harvesting wireless network. We hope that
 this article can inspire researchers to propose more GenAI-
     
7UDLQLQJHSLVRGH based methods in IoT network.

Fig. 3. The AoI convergence of different algorithms. R EFERENCES

[1] A. Khanna and S. Kaur, “Internet of things (iot), applications and
challenges: A comprehensive review,” Wirel. Pers. Commun., vol. 114,
no. 2, pp. 1687–1762, 2020.
can produce diverse high-quality actions, thus improving the [2] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks
exploration of the action space compared to the traditional with RF energy harvesting: A contemporary survey,” IEEE Commun.
TD3. Surv. Tutorials, vol. 17, no. 2, pp. 757–789, 2015.
[3] L. Liu, K. Xiong, J. Cao, Y. Lu, P. Fan, and K. B. Letaief, “Average AoI
minimization in UAV-assisted data collection with RF wireless power
transfer: A deep reinforcement learning scheme,” IEEE Internet Things
V. F UTURE D IRECTIONS J., vol. 9, no. 7, pp. 5216–5228, 2022.
[4] Y. A. Nando and W.-Y. Chung, “Enhancing rf energy harvesting and
wireless power transfer with gan-optimized 3d quasi-yagi antenna,” in
In this section, we present future directions about GenAI- WPTCE, 2024, pp. 454–458.
powered energy harvesting wireless network. [5] S. Zargari, C. Tellambura, A. Maaref, and G. Y. Li, “Deep conditional
generative adversarial networks for efficient channel estimation in ambc
systems,” IEEE Transactions on Machine Learning in Communications
and Networking, 2024.
A. GenAI-based Distributed Energy Beamforming [6] M. H. Ahmadi, M. Ghazvini, M. Alhuyi Nazari, M. A. Ahmadi,
F. Pourfayaz, G. Lorenzini, and T. Ming, “Renewable energy harvesting
Distributed energy beamforming allows multiple dispersed with the application of nanotechnology: A review,” International Journal
of Energy Research, vol. 43, no. 4, pp. 1387–1410, 2019.
energy sources to work together like an antenna array by [7] D. K. Sah and T. Amgoth, “Renewable energy harvesting schemes in
simultaneously directing RF energy towards a specific energy wireless sensor networks: A survey,” Inf. Fusion, vol. 63, pp. 223–247,
harvester. In this case, GenAI can perform coordination of 2020.
[8] S. Guo, Y. Shi, Y. Yang, and B. Xiao, “Energy efficiency maximization in
the distributed carriers in phase and frequency so that the RF mobile wireless energy harvesting sensor networks,” IEEE Trans. Mob.
signals can be constructively combined at the receiver. Comput., vol. 17, no. 7, pp. 1524–1537, 2018.
[9] L. Yang, Y. J. Zhou, C. Zhang, X. M. Yang, X.-X. Yang, and C. Tan,
“Compact multiband wireless energy harvesting based battery-free body
area networks sensor for mobile healthcare,” IEEE Journal of Electro-
B. GenAI-based Inference Management magnetics, RF and Microwaves in Medicine and Biology, vol. 2, no. 2,
pp. 109–115, 2018.
Interference is a common factor affecting wireless networks. [10] A. Georgiadis and A. Collado, “Improving range of passive RFID tags
utilizing energy harvesting and high efficiency class-e oscillators,” in
Currently, most research focuses on how to avoid and mitigate EUCAP, 2012, pp. 3455–3458.
interference. In energy-harvesting wireless networks, lever- [11] K. Chung and J. Lim, “Machine learning for relaying topology: Opti-
aging GenAI to devise intelligent scheduling strategies that mization of IoT networks with energy harvesting,” IEEE Access, vol. 11,
pp. 41 827–41 839, 2023.
transform harmful interference into useful energy represents a [12] R. Lin, H. Qiu, J. Wang, Z. Zhang, L. Wu, and F. Shu, “Physical-layer
promising and intriguing direction for future exploration. security enhancement in energy-harvesting-based cognitive internet of
things: A GAN-powered deep reinforcement learning approach,” IEEE
Internet Things J., vol. 11, no. 3, pp. 4899–4913, 2024.
[13] D. Kaur, S. N. Islam, M. A. Mahmud, M. E. Haque, and A. Anwar,
C. GenAI-based Trade-off Strategy Design “A VAE-bayesian deep learning scheme for solar power generation
forecasting based on dimensionality reduction,” Energy and AI, vol. 14,
For RF-powered devices, where the transmission power p. 100279, 2023.
[14] H. Hu, K. Xiong, G. Qu, Q. Ni, P. Fan, and K. B. Letaief, “AoI-minimal
is generally low, employing multiple antennas can enhance trajectory planning and data collection in UAV-assisted wireless powered
transmission efficiency. However, increasing the number of IoT networks,” IEEE Internet Things J., vol. 8, no. 2, pp. 1211–1223,
antennas also leads to higher power consumption. GenAI 2021.
[15] S. Fujimoto, H. van Hoof, and D. Meger, “Addressing function ap-
can be utilized to develop suitable optimization strategies proximation error in actor-critic methods,” in Proceedings of the 35th
that balance transmission efficiency and power consumption, International Conference on Machine Learning, ICML, vol. 80, 2018,
making it well-suited for dynamic environments. pp. 1582–1591.