User selection and dynamic power allocation in the SWIPT-NOMA relay system

Xianzhong Xie, Min Li, Zhaoyuan Shi, Hong Tang, Qian HuangEURASIP J. Wirel. Commun. Netw 2021 · DOI

Abstract

This paper introduces USDPA, a novel joint optimization framework for user selection and dynamic power allocation in 5G-and-beyond relay systems. By integrating Deep Neural Networks (DNN) for user access decisions and Deep Reinforcement Learning (DRL) for power management, the proposed system addresses the critical challenges of spectral efficiency and energy constraints in SWIPT-NOMA networks. Simulation results demonstrate that USDPA significantly enhances system throughput and energy efficiency, outperforming traditional non-selective and static power allocation methods. Notably, the scheme achieves a sum rate nearly 50% higher than standard benchmarks in high-demand scenarios while maintaining stable convergence.

Introduction

Modern wireless networks face a dual challenge: the explosion of connected devices requiring high spectral efficiency and the limited battery life of mobile terminals necessitating green energy solutions. To address these, researchers have turned to Non-Orthogonal Multiple Access (NOMA) and Simultaneous Wireless Information and Power Transfer (SWIPT). NOMA allows multiple users to share the same frequency/time resource, while SWIPT enables relay nodes to harvest energy from radio signals to power their transmissions. However, combining these technologies creates a complex "non-convex, non-linear" optimization problem. Traditional mathematical methods are often too slow or fail to find global optima in dynamic, real-time environments.

System Model

The study considers a SWIPT-NOMA relay system consisting of a base station (BS), a decode-and-forward relay, and multiple destination nodes ($D_1$ to $D_N$). The relay employs a Power Splitting-based Relaying (PSR) strategy to harvest energy and forward information.

The SWIPT-NOMA relay model illustrating the energy harvesting and information forwarding process between the base station, relay, and destination nodes.

Proposed Methodology: The USDPA Scheme

The authors propose the User Selection and Dynamic Power Allocation (USDPA) scheme, which decomposes the optimization problem into two AI-driven stages:

  1. User Selection (DNN): A four-layer Deep Neural Network takes channel state information as input. Since standard DNNs output continuous values, the authors implemented a unique mapping rule. This rule generates several candidate access vectors, allowing the system to adaptively choose the most efficient group of users for the current channel conditions.
  2. Power Allocation (DQN): For each candidate user group, a five-layer Deep Q-Network (DQN) determines the optimal power allocation factors. The objective is to maximize the "sum rate" (total data speed) while ensuring every selected user meets a minimum Quality of Service (QoS) threshold ($R_{th}$).
The USDPA architecture featuring integrated DNN and DQN units for joint optimization of user access and power levels.

The system utilizes integrated training where the best pair of access and power actions is used to calculate a reward, which then updates both networks simultaneously via a replay memory.

Key Insights and Results

Simulations conducted in TensorFlow 2.0 reveal several critical insights. One of the most significant findings is the "Selectivity Advantage": in high-demand scenarios, letting fewer users connect reduces interference so significantly that the total system throughput increases compared to "All-User Access" models.

When the QoS requirement was set to 0.3 bits/s/Hz, USDPA achieved a sum rate 47.8% higher than the standard "All Users + DQN" approach and 178% higher than simple average power allocation schemes. Furthermore, the system consistently maintained a higher number of Successful Communication Users (NSCUs). The DRL approach proved stable, with the mapping scheme and power allocation stabilizing within 4,000 training slots, demonstrating viability for real-time applications.

Conclusion and Future Work

USDPA provides a robust, computationally efficient framework for managing next-generation relay systems. While the current model assumes single-antenna nodes and unidirectional traffic, it establishes a strong foundation for future research. The authors plan to extend this AI framework to multi-antenna (MIMO) configurations and bidirectional relaying systems to better reflect the complexity of 6G cellular networks.

Summary updated

BibTeX
@article{xie2021user,
  author = {Xianzhong Xie and Min Li and Zhaoyuan Shi and Hong Tang and Qian Huang},
  title = {{User selection and dynamic power allocation in the SWIPT-NOMA relay system}},
  journal = {EURASIP J. Wirel. Commun. Netw.},
  year = {2021},
  doi = {10.1186/s13638-021-01998-0},
}
Figure