Generative AI for Energy Harvesting Internet of Things Network: Fundamental, Applications, and Opportunities

Solar energy harvesting typically involves converting sunlight into electricity using photovoltaic cells, which is highly effective in sunny outdoor environments. Consequently, solar energy harvesting is widely applied in remote sensing, environmental monitoring, and agricultural IoT systems.

Wind energy harvesting typically employs wind turbines to convert the kinetic energy of wind into electricity, which is particularly suitable for areas with abundant wind energy resources. Moreover, wind power generation devices can be installed near wireless network base stations or nodes to provide them with sustained power supply.

Other forms of renewable energy also hold potential in energy harvesting for wireless networks. For example, thermoelectric generators convert temperature differences into electricity, while piezoelectric materials generate electricity through mechanical vibrations. These energy sources are commonly found in industrial environments and can provide power support for industrial wireless sensors widely.

The most notable advantage of renewable natural energy is its environmental friendliness. However, renewable natural energy still has limitations, such as geographic constraints or instability due to environmental conditions. The emergence of RF energy harvesting offers a new approach to alleviate such limitations, especially dedicated RF sources.

GANs consist of two neural networks: a generator and a discriminator. The generator creates artificial data from random noise, while the discriminator attempts to distinguish between real and artificial data. Through adversarial training, the generator network generates high-quality data. GANs excel at generating highly realistic data but suffer from training instability and mode collapse that the generator keeps producing limited similar data patterns. GANs are primarily used in image generation, data augmentation, and network optimization.

Diffusion models gradually add noise to data through a forward diffusion process until it becomes pure noise, and then denoise it to recover the data through a 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 speech generation. Moreover, diffusion models can be utilized to improve the policy network of deep reinforcement learning (DRL), which has been applied to several domains and has shown superior performance.

VAEs consist of an encoder and a decoder. The encoder maps input data to a probabilistic distribution of latent variables, while the decoder reconstructs the data from these variables. VAEs are trained by minimizing reconstruction error and the Kullback Leibler (KL) divergence between the latent distribution and a standard normal distribution. They excel at creating continuous latent spaces and explicitly modeling generation probabilities, making them suitable for image generation and anomaly detection.

The Transformer model consists of multiple stacked encoder and decoder layers, each containing self-attention mechanisms and feed-forward neural networks. The encoder transforms input sequences into context-aware representations, while the decoder generates the next sequence element based on the output and the previous output of the decoder. Transformers handle long-range dependencies well, offer high training efficiency, and have a flexible structure, making them widely used in natural language processing tasks such as machine translation and text generation.

The recent proliferation of connected devices has led to a substantial surge in spectrum utilization and energy consumption. As a solution, ambient backscatter communication (AmBC) has garnered considerable attention. Note that an AmBC node needs to harvest energy to supply its backscatter operations, i.e., circuit energy consumption. In this case, channel estimation is crucial in AmBC, since it can help to optimize energy utilization. Specifically, by accurately estimating the channel state, key resources such as frequency allocation can be dynamically adjusted to extend the working life of wireless devices. However, traditional channel estimation methods, including the techniques such as pilot-based estimators, blind estimators, and semi-blind estimators, typically face several issues such as requiring extensive prior knowledge, being limited by scenario changes, and exhibiting sensitivity to model matching. Therefore, advanced machine learning algorithms for channel estimation are of great interest.

In this context, a novel channel estimation algorithm based on conditional GAN (CGAN) for AmBC was proposed [5]. The CGAN estimator can estimate the wireless channel status and quality by accurately capturing complex features and utilizing stronger reflection paths. This capability allows the CGAN estimator to adapt well to the changes in the relative strength of reflected signals, resulting in superior channel estimation performance. Specifically, the proposed channel estimation method based on CGAN applies specific channel state information (CSI) as a condition to GAN, and the generator network generates synthetic channel data that closely resembles real-world conditions. This synthetic data in turn facilitates accurate estimation of AmBC channel coefficients.

The performance of the proposed CGAN estimator is evaluated by comparing it with the well-known methods such as least squares (LS), minimum mean-squared error (MMSE), blind, deep residual learning denoiser (CRLD), and convolutional neural network (CNN). The results show that CGAN channel estimator excels in normalized MSE (NMSE), even under challenging low signal-to-noise ratio (SNR) conditions. Moreover, at an SNR of 5 dB, the CGAN channel estimator consistently surpasses existing CE solutions, achieving an 82% improvement over CRLD and MMSE estimators.

Relay-based structures are particularly suitable for energy harvesting wireless networks. First, cooperative relaying techniques enhance network efficiency and reliability by using relay nodes to overcome fading and attenuation. Second, energy-rich nodes can transmit data on behalf of energy-constrained nodes to maximize overall network efficiency. Therefore, through proper deployment and energy management of relay nodes, the network can address energy harvesting uncertainties, extend node lifetimes, and improve data transmission reliability.

The authors in [11] considered an uplink transmission scenario in an relay-based IoT sensor network, where each wireless node self-powers by harvesting energy from power 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 transmission bits of all nodes. Given the lack of datasets and the nonlinear nature of the optimization problem, the authors proposed an unsupervised relay topology algorithm based on a VAE. Specifically, a random latent vector is input into the VAE, and subsequently the analytical and generative capabilities of the VAE decoder are utilized to generate the topology matrix. To evaluate quality of the topology matrix generated by the VAE, the authors introduced a packet-tracking evaluation method inspired by ray-tracing techniques. This method can accurately evaluate the output of the VAE, providing appropriate guidance for VAE training.

The VAE-based topology approach was compared with traditional schemes such as the direct connect scheme, minimum spanning tree (MST) scheme, and greedy scheme, as well as the optimal brute-force solution. The experiment results indicate that the MST scheme exhibits insufficient adaptability to various IoT network configurations, and the brute-force search algorithm has excessively long computation times. In contrast, the VAE-based topology approach demonstrates better performance than all other considered approaches across all node counts and network configurations, being the closest to the optimal solution.

Antennas are the core components of energy harvesting devices, which means that the antenna design is critical in energy harvesting wireless networks. By adjusting the antenna design in terms of frequency selectivity, directivity, and structure, we can optimize signal reception and minimize energy loss and interference. Consequently, an ideal and well-considered antenna design can maximize the ability of devices to harvest RF energy from the environment.

The authors in [4] investigated to enhance the RF energy harvesting and power transfer with GAN-optimized antenna design. Inspired by the quasi-Yagi antenna (QYA) model, the authors aimed to develop a cube-shaped three-dimensional (3D) antenna optimized for 915 MHz by analyzing various 2D QYA configurations. First, to enhance robustness of the analysis, the authors employed GAN to expand the dataset. Specifically, the generator network generates synthetic QYA data, while the discriminator network engages in adversarial training to optimize the performance of the generator network. Moreover, GAN was also used to optimize the design of the 3D QYA. A well-trained GAN can effectively capture the complex relationships between QYA parameters and QYA structure, generating target designs based on specified criteria. To strengthen this process, the authors integrated the GAN with the CST Studio Suite for virtual performance testing of GAN-generated designs to enable timely adjustments of the GAN model, thereby enhancing energy capture from ambient and dedicated RF sources.

The experimental results show that the minor deviation between the GAN and CST Studio Suite results is within an acceptable range, underscoring the reliable performance of the antenna. Moreover, compared to other 2D QYA designs, the GAN-designed 3D antenna exhibits superior energy harvesting capability.

Cognitive radio, through dynamic spectrum management and spectrum sharing, effectively alleviates spectrum scarcity issues, ensuring efficient network operation. Given that cognitive radio devices frequently engage in spectrum sensing and dynamic adjustments, traditional battery power typically fails to meet their long-term operational needs. Energy harvesting can provide continuous power support for these devices. Moreover, the benefit of energy harvesting can be further enhanced by optimizing the allocation of resources in the energy harvesting network.

In this case, the authors in [12] developed the GAN-based DRL algorithm to enhance the physical security in energy harvesting cognitive radio system. In the considered system, the communication security of the secondary user (SU) can be tampered by eavesdroppers, prompting the deployment of a cooperative jammer to secure wireless transmissions. Notably, both the SU and the jammer can harvest energy from the received RF signals emitted by the primary transmitter (PT), with batteries capable of storing energy for extended periods. The authors employed deep Q-network (DQN) algorithm to maximize the average minimum secrecy rate and minimize the average maximum secrecy outage probability (SOP). To better estimate the state-action values, GAN was introduced to learn the optimal state-action value distribution for DRL. Specifically, the generator network $G$ is responsible for creating the estimated state-action value distribution, while the target generator network $G^{{}^{\prime}}$ produces the target state-action value distribution. The role of the discriminator network $D$ is to differentiate between the values produced by $G$ and those from $G^{{}^{\prime}}$.

The experimental results indicate that, compared to the case with one eavesdropper, the secrecy rate achieved by DQN decreases by approximately 78% and the SOP increases by 67% when there are ten eavesdroppers. This suggests that the DQN approach is highly sensitive to the number of eavesdroppers. In contrast, the GAN-powered DQN reduces the secrecy rate by 20% and improves the SOP by 67%. Moreover, the maximum average SOP achieved with GAN-powered DQN closely matches the theoretical value of the closed expression, which validates the effectiveness of resisting eavesdropping.

Since the renewable energy is an intermittent generation, it is subject to external factors such as weather variations and changes in user behavior, which often bring new challenges to the energy harvesting solutions. In this context, accurate forecasting of power generation in advance is essential to improve the operational efficiency of the energy harvesting systems, which can be used in wireless networks.

A VAE-based method for solar power generation forecasting was proposed [13]. VAE can effectively reduce the dimensionality of parameters while retaining important data features, which is highly beneficial for enhancing the computational efficiency and generalization capability of the prediction model. The proposed VAE-based process involves mapping high-dimensional data to a low-dimensional latent space and then reconstructing the data from the latent space. Therefore, by optimizing the reconstruction error and the probability distribution of latent variables, VAE can learn a low-dimensional representation of the data. Building on the VAE, a Bayesian bidirectional long short-term memory (BiLSTM) network was incorporated for time series prediction. BiLSTM can capture temporal dependencies and bidirectional information in the data, further improving prediction accuracy.

Compared with benchmark methods, the VAE-based approach shows significant improvement in parameter reduction. Specifically, in six months of solar power generation data, the number of weight parameters was reduced from 764224 to 2022, achieving a quantized parameter reduction of 97.35% and a computation time improvement of 37.93%.

To improve online data analysis and decision-making in IoT networks, it is essential to effectively gather the latest information produced by IoT devices. To precisely measure information freshness, the concept of Age of Information (AoI) has been introduced [14]. AoI is defined as the duration from when the most recent update packet is created at the source until it is received at the destination. Note that enhancing the power supply for energy-constrained IoT devices is essential for improving AoI performance in IoT networks. However, traditional methods of energy supply, such as battery replacements or wired connections, are impractical and labor-intensive. Note that using UAVs for energy transmission has emerged as a promising solution to overcome the energy limitations of IoT devices. UAVs offer flexibility and mobility, allowing them to reach remote areas that traditional power sources might not be able to access. Therefore, UAVs can provide stable energy transfer in an on-demand basis, reducing downtime for IoT devices and ensuring the freshness of data collection and continuity of communication.

As shown in Fig. 2, multiple IoT devices equipped with energy buffers are randomly distributed within the target area to sense surrounding environment. A UAV is moving within the target area, responsible for charging the IoT devices to ensure their activity and receiving data from the IoT devices. Note that only one IoT device can upload data during each time slot. We aim to minimize the AoI by optimizing the trajectory of UAV, the scheduling of IoT devices, and the charging time.