⚡ electrical engineering

Energy-Efficient IRS-Assisted Green Cellular Networks With Ground Base Station Sleep Mode

This paper proposes a hybrid deep reinforcement learning framework that jointly optimizes intelligent reflecting surface phase shifts, ground base station sleep modes, user association, power allocation, and aerial base station positioning to maximize energy efficiency in IRS-assisted aerial-terrestrial cellular networks.

Original authors: Vala Saleh, Zahra Amiri, Mohsen Eslami, Kamran Kazemi

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Vala Saleh, Zahra Amiri, Mohsen Eslami, Kamran Kazemi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a bustling city where the lights in every building are always on, even when the rooms are empty. This is how our current cell phone networks work: ground-based towers (Base Stations) stay awake and consume massive amounts of electricity 24/7, even when very few people are using them. This is expensive and wasteful.

This paper proposes a smarter, "green" way to run these networks by combining three key technologies: Sleeping Towers, Flying Drones, and Magic Mirrors.

Here is how the authors explain their solution, broken down into simple concepts:

1. The Problem: The "Always-On" Waste

Think of the ground towers as streetlights. Currently, they are programmed to stay bright all night, regardless of whether anyone is walking below. When traffic is low, these towers are still burning energy, which drives up costs and pollution.

2. The Solution: A Three-Part Team

The authors suggest a team of three distinct technologies working together to save energy without dropping your phone call.

The Sleepers (Ground Base Stations):
Instead of staying awake all night, the ground towers are given a "sleep mode." If a tower sees that very few people are nearby, it turns itself off (or dims to a low-power state) to save energy.
- The Risk: If a tower goes to sleep, people in that area might lose signal.
The Flyers (Aerial Base Stations):
To catch the people left behind when a tower sleeps, a drone (or flying base station) hovers overhead. Think of this drone as a mobile spotlight. It can fly to exactly where the people are, filling in the coverage gaps left by the sleeping towers. Because it can move, it doesn't need to be everywhere at once; it only goes where it's needed.
The Magic Mirrors (Intelligent Reflecting Surfaces - IRS):
This is the most unique part. Imagine a wall covered in thousands of tiny, smart mirrors. These aren't just glass; they are computer-controlled. If a signal from a tower or drone is blocked by a building or is too weak, these "mirrors" catch the signal, bounce it off at the perfect angle, and shoot it directly to your phone.
- The Benefit: This makes the signal stronger without needing to turn up the volume (power) on the tower or drone. It's like using a mirror to reflect sunlight into a dark room instead of turning on a bright lamp.

3. The Brain: The "Smart Coach" (AI)

The big challenge is coordinating all these moving parts. You have to decide:

Which ground towers should sleep?
Where should the drone fly?
How should the "magic mirrors" angle themselves?
How much power should everyone use?

Doing this math manually is impossible because the network changes every second. The authors created a Hybrid AI Coach (using Deep Reinforcement Learning) to make these decisions in real-time.

The Discrete Coach (DDQL): This part of the AI makes "Yes/No" decisions. Should Tower A sleep? Should User B connect to the Drone?
The Continuous Coach (DDPG): This part makes "fine-tuning" decisions. Exactly how much power should the drone use? What exact angle should the mirrors be set to?

These two coaches work together to maximize Energy Efficiency (getting the most data done for the least amount of electricity).

4. What They Found

The authors ran computer simulations to test this idea. Here is what happened:

The System Learned: At first, the AI was confused, but after training, it figured out the perfect balance.
More Sleep, Less Power: The system learned to turn off many ground towers (sleep mode) because the drone and the mirrors could handle the load.
Better Performance: Even though they used less electricity, the users actually got faster data speeds. This is because the "Magic Mirrors" cleaned up the signal, and the drone moved to the best spot.
Beating the Competition: When compared to systems without mirrors, without drones, or without sleep mode, their "Hybrid Team" was the clear winner, using significantly less energy to do the same job.

Summary

The paper claims that by letting ground towers sleep, using drones to fill the gaps, and using smart mirrors to bounce signals around obstacles, we can build a cellular network that is much cheaper to run and greener for the environment, all while keeping your phone connection strong. The "Smart Coach" AI is the key to making all these pieces work together without human intervention.

Technical Summary: Energy-Efficient IRS-Assisted Green Cellular Networks With Ground Base Station Sleep Mode

Problem Formulation
This research addresses the challenge of maximizing energy efficiency (EE) in beyond-fifth-generation (B5G) and sixth-generation (6G) wireless networks. The specific scenario involves a downlink cellular network integrating three key components: Ground Base Stations (GBSs), a single Aerial Base Station (ABS), and Intelligent Reflecting Surfaces (IRSs). The core problem arises from the need to balance energy savings with Quality of Service (QoS) requirements. While GBS sleep modes reduce static power consumption, they can create coverage gaps and degrade performance, particularly for cell-edge users. Conversely, ABSs offer flexible coverage but introduce platform energy costs (e.g., hovering), and IRSs enhance signal propagation but require joint optimization of phase shifts.

The authors formulate this as a mixed-integer non-convex optimization problem. The objective is to maximize the ratio of total network throughput to total power consumption. The optimization variables include:

Discrete variables: GBS activation states (sleep/active) and user association (which transmitter serves which user on which subchannel).
Continuous variables: Transmit power allocation for GBSs and the ABS, the horizontal positioning of the ABS, and the phase shifts of IRS elements.

The problem is constrained by power budgets, QoS (minimum data rate) requirements, unit-modulus constraints for IRS elements, and the explicit modeling of co-channel interference between active terrestrial and aerial transmitters. The non-convex nature of the problem, driven by fractional objectives, binary variables, and coupled interference terms, renders conventional optimization methods computationally intractable.

Methodology
To solve this complex optimization problem, the authors propose a Hybrid Deep Reinforcement Learning (DRL) framework. This approach decomposes the decision space into discrete and continuous components, leveraging the strengths of two distinct algorithms:

Double Deep Q-Learning (DDQL): Used to handle discrete decision-making tasks. Specifically, the DDQL agent determines the GBS sleep-mode indicators ( $\alpha_m$ ) and user association variables ( $\beta_{k,m,l}, \beta_{k,a,l}$ ). This module learns the optimal network topology and activation patterns to minimize unnecessary energy usage while maintaining coverage.
Deep Deterministic Policy Gradient (DDPG): Used to optimize continuous control variables. The DDPG agent adjusts the ABS horizontal position, transmit power allocation for both ground and aerial nodes, and the IRS phase-shift matrices. This module adapts to dynamic channel conditions and interference patterns.

The system is modeled as a Markov Decision Process (MDP). The state space includes channel state information, current GBS activity, user associations, power levels, ABS position, and IRS configurations. The reward function is designed to maximize energy efficiency while imposing penalties for QoS violations (users failing to meet minimum rate requirements) and constraint violations (e.g., power limits).

Key Contributions
The paper outlines the following specific contributions:

Framework Development: The design of an energy-efficient network framework that jointly integrates IRS assistance, ABS mobility, and GBS sleep modes.
Interference-Aware Modeling: The creation of a system model that explicitly accounts for co-channel interference between terrestrial and aerial transmissions within an orthogonal resource allocation scheme.
Optimization Formulation: The formulation of a mixed-integer non-convex problem that jointly optimizes sleep indicators, user association, power, positioning, and phase shifts.
Hybrid Algorithm: The proposal of a hybrid DRL solution combining DDQL and DDPG to efficiently navigate the high-dimensional, mixed-variable decision space.
Performance Validation: Demonstration that the proposed approach improves energy efficiency while satisfying QoS constraints, outperforming conventional benchmarks.

Numerical Results
The proposed framework was evaluated through numerical simulations involving 3 GBSs, 6 users, and 2 IRSs within a 500m radius area. The results were compared against three benchmark schemes: a system without IRSs, a system without an ABS, and a system where all GBSs remain active (no sleep mode).

Convergence: The hybrid DRL framework demonstrated stable convergence, with energy efficiency increasing from approximately $3 \times 10^6$ bit/J to between $5.5$ and $6.5 \times 10^6$ bit/J after training.
Trade-offs: The system successfully achieved a dual improvement: increasing the average user data rate while simultaneously reducing total power consumption.
Sleep Mode Effectiveness: The learned policy progressively reduced the number of active GBSs over training episodes. The results indicate that IRS-assisted links and ABS support allow the network to maintain QoS even with fewer active ground stations.
Comparative Performance: The proposed method achieved the highest energy efficiency among all compared schemes. The performance gap with the "No-IRS" and "No-ABS" baselines highlighted the specific benefits of signal reflection enhancement and aerial flexibility, respectively.

Significance and Claims
The paper claims that the integration of ABS flexibility, IRS-assisted propagation, and intelligent sleep-mode control enables a highly efficient and adaptive network operation. The authors assert that their hybrid DRL approach provides a promising solution for addressing energy efficiency challenges in future wireless networks by effectively balancing coverage, interference management, and energy consumption. The study concludes that the proposed framework significantly improves energy efficiency compared to schemes lacking IRS assistance, ABS support, or sleep-mode optimization, without compromising user data rates.

The authors maintain a modest scope regarding future work, noting that future research directions include extending the framework to scenarios with multiple ABSs, dynamic IRS deployment strategies, and more realistic mobility and traffic models.