arXiv CSopen access

Large Language Model Assisted Intent-Based Satellite-Integrated Access and Backhaul FWA for Rural Areas

Large Language Model Assisted Intent-Based Satellite-Integrated Access and Backhaul FWA for Rural Areas

arXiv:2607.21272v1 [cs.NI] 23 Jul 2026

Anselme Ndikumana1 , Kim Khoa Nguyen 1 , Adel Larabi2 , and Mohamed Cheriet1 1 Synchromedia Lab, École de Technologie Supérieure, Université du Québec, QC, Canada {anselme.ndikumana, kim-khoa.nguyen; Mohamed.Cheriet}@etsmtl.ca 2 GAIA Montreal, Ericsson Canada {adel.larabi}@ericsson.com

Rural areas exhibit low population density and highly variable connectivity needs shaped by both household usage and field operations such as planting, harvesting, and mining. These field activities often occur in isolated locations requiring temporary connectivity, whereas rural households depend on fixed broadband. During intensive outdoor activities, household fixed networks may remain underutilized, resulting in inefficient resource use and unnecessary energy consumption. The coexistence of residential and field-based communication demands creates substantial spatial and temporal fluctuations that the current rural network cannot effectively adapt to. Limited visibility into user mobility, activity patterns, and intent makes it difficult for operators to coordinate temporary and fixed networks. To address these underexplored challenges, we propose an AI-driven IntentAware Satellite–Integrated Access and Backhaul (IAB) approach to connect rural areas. In our proposal, a large language model (LLM) translates users’ intents into explicit network requirements. Guided by these inferred requirements, we develop a dynamic satellite-IAB-based Fixed Wireless Access (FWA) network approach that jointly optimizes temporary field connectivity and fixed broadband access to maximize energy efficiency while satisfying the data rate requirement. The formulated optimization problem is solved using a two-stage Benders decomposition approach. The simulation results show that our approach significantly reduces energy consumption while maximizing energy efficiency. Index Terms—Intent-aware network, satellite communication, integrated access and backhaul, energy efficiency, large language model, fixed wireless access

I. Introduction A. Background and Motivations Rural areas play an essential role in national and global economies, being home to major land-based industries such as agriculture, forestry, and natural resource extraction [1]. The ongoing digital transformation of these industries increasingly depends on advanced automation and intelligent systems, including autonomous tractors, This work was supported by NSERC (under project ALLRP 566589-21) and InnovÉÉ (INNOV-R program) through the partnership with Ericsson and ECCC. We thank the Ericsson’s Montréal GAIA team for their constructive and helpful comments, which have significantly improved the quality and clarity of this manuscript (corresponding author: Anselme Ndikumana).

unmanned aerial vehicles (drones), and collaborative robots. These devices require communication networks to support data exchange, remote monitoring, and real-time control. Unlike urban regions that benefit from dense, permanent communication infrastructure, rural areas are characterized by low population density, wide geographic dispersion, and highly variable traffic demand. Furthermore, field operations, such as planting, harvesting, logging, and mining, are often conducted in isolated locations where connectivity is unavailable. Consequently, field activities require on-demand or temporary networks that can be deployed and activated dynamically when needed. In contrast, rural households primarily rely on fixed networks, such as Fixed Wireless Access (FWA) networks [2], for daily broadband connectivity. In 5G FWA, houses are equipped with Customer Premises Equipment (CPE) with antennas mounted on their roofs, wirelessly connected to fixed cellular base stations [3]–[5]. However, during outdoor activities, these fixed networks may remain underutilized, resulting in inefficient resource use and unnecessary energy consumption. The coexistence of residential and field-based communication network needs creates distinct spatial and temporal fluctuations in network demand. Despite recent advances in rural broadband access, the current network infrastructure cannot adapt to dynamic changes in connectivity requirements. Furthermore, network operators typically have limited visibility into population mobility and activity patterns [6], making it difficult to coordinate between fixed and temporary networks. This gap motivates the development of an intelligent, context-aware approach capable of predicting or interpreting user needs in rural areas to dynamically adjust network configurations and resource allocation. Recent progress in intent-based networking (IBN) [7] has opened new possibilities for adaptive network management. In an intent-aware system, users or applications can express high-level operational objectives, referred to as intents, which the network then interprets to provision appropriate connectivity and quality-of-service (QoS) levels automatically. For example, in a rural area, a farming cooperative could express an intent such as “farm operations from 9:00 AM to 4:00 PM involving drones and robots,” prompting the network to establish temporary high-throughput, low-latency connections in

the corresponding area during that time window. Applying such intent awareness in rural networks can significantly improve the coordination between temporary and fixed networks. Existing rural connectivity solutions, such as FWA, are limited in flexibility, coverage, and adaptability. Integrated Access and Backhaul (IAB) [8] has been included in FWA to extend coverage and capacity. The IAB network consists of an IAB donor and multiple IAB nodes. The IAB donor functions as the central base station and is directly connected to the core network. The remaining base stations, referred to as IAB nodes, establish wireless backhaul links with the IAB donor. In an IAB-based FWA scenario, CPE can connect to either the IAB donor or the IAB nodes via wireless access links [4]. Each IAB node is composed of a Distributed Unit (DU) and a Mobile Termination (MT) unit. The DU serves downstream CPEs and subordinate IAB-MTs, while the MT enables the node to operate as a relay by connecting to its parent IAB-DU. In contrast, the IAB donor integrates both Distributed Unit (DU) and Control Unit (CU) functionalities. However, the IAB donor requires a fiber-optic backhaul to the core network, which is not economically feasible in rural areas. Furthermore, fiber-optic backhaul is often unreliable due to the complex, variable terrain in some rural areas, such as agriculture or mining zones. Hence, there is a need for non-terrestrial network backhauling, such as satellite communication [9], to support terrestrial networks in rural areas to ensure seamless connectivity across diverse operational contexts.

C. Contributions

To address the challenges outlined above, this paper proposes an Artificial Intelligence (AI)- driven, intent-aware Satellite–IAB–based FWA approach to connect rural areas. The framework leverages intent-based networking principles to intelligently coordinate the activation, deactivation, and configuration of IAB functionality at terrestrial and satellite nodes based on user activities and connectivity demands in rural areas. The main contributions of this work are summarized as follows: • We propose a mapping approach that converts user intent into network vocabulary and, subsequently, into QoS requirements, using a Large Language Model (LLM) as an AI tool. This enables the automatic establishment of fixed broadband access in residential areas and temporary network coverage in relevant field areas when field activities begin, and the deactivation of unnecessary functions once activities end to minimize energy consumption and release unused resources while maintaining fixed broadband access. • Based on the derived network requirements from intents, we formulate a user admission optimization problem as a 0–1 integer linear program. This approach efficiently resolves conflicts between multiple intents and network requirements, supporting both temporary and fixed network deployments. • We propose a satellite-assisted IAB network that supports temporary field connectivity alongside FWA. The IAB functions can be dynamically activated or deactivated to maximize energy efficiency (i.e., minimize energy consumption) while meeting delay and data rate requirements. B. Challenging Issues • For network setup and configuration in the satelliteassisted IAB-based FWA network, we formulate an Despite significant fixed network progress for rural areas, optimization problem to maximize energy efficiency several key challenges hinder the realization of intent-aware while satisfying delay and data rate requirements. and adaptive connectivity in rural areas: Then, we propose a two-stage Benders decomposition • FWA [10] provides an affordable solution for broad[15] as the solution approach. band connectivity, but is generally limited to one-hop The remainder of this paper is organized as follows. access. Its coverage footprint is inadequate for large, Section II reviews the existing literature on rural comremote areas such as agricultural or forestry sites, and munication systems, intent-based networking, and satelliteits static nature limits deployment flexibility. assisted IAB architectures. Section III presents the system • IAB [11] extends FWA coverage through multi-hop model and details the proposed intent-aware satellitewireless relays. However, it relies on a fiber-optic assisted IAB-based FWA. Section IV discusses the problem backhaul link to the IAB donor, which may be formulation. Section V describes the solution approach. impractical in temporary and remote rural areas. Section VI presents performance evaluations and discusses • Microwave backhaul [12] can be effective under clear the effectiveness of the proposed framework. Finally, Secline-of-sight (LoS) conditions to replace fiber-optic, but tion VII concludes the paper and outlines directions for its reliability deteriorates when LoS is obstructed by future research. forests, terrain undulations, or vegetation [13]. These limitations reduce its applicability for establishing II. Literature Review temporary networks in rural areas. • Satellite communication provides wide-area coverage Rural Area Connectivity: Rural and remote regions and inherent mobility [14], making it a strong can- often face persistent connectivity challenges due to low didate for replacing microwave backhaul to connect population density, difficult terrain, and limited infrasthe IAB donor to the core network. Nevertheless, tructure investment. Traditional cellular networks and integrating satellite links into terrestrial IAB-FWA fixed broadband solutions are economically unattractive for is unexplored in the existing literature. operators in these areas, resulting in digital exclusion [12],

[13]. To address this issue, FWA has been explored as a Table I Summary of key notations. cost-effective alternative to fiber deployment. FWA enables broadband-like performance using wireless backhaul, but Notation Definition its coverage is often limited to a single hop between T Time slots, t = 1, . . . , T the base station and end users [3], [10], [16]. To extend U Set of terminals ( i.e., UEs/CPEs) coverage beyond FWA’s capabilities, IAB has emerged as a S Set of satellites IAB nodes N Set of terrestrial IAB nodes promising 5G solution [11]. IAB enables multi-hop connecE Set of directed links tivity via wireless backhaul links, improving deployment M Set of user terminal intents flexibility and reducing reliance on fiber. However, in the V A set of network vocabulary Q set of network QoS IAB network, the donor still needs fiber-optic connectivity mu User intent at time t t to access the internet. Consequently, maintaining reliable du Data rate requirement of terminal u t multi-hop links remains a technical challenge for rural Cti,j Link capacity between satellites i and j deployments [13]. Ctj,k Link capacity between terrestrial IAB station k with satellite m-IAB node j Satellite-Assisted IAB Network: Satellite communications set Satellite visibility for forming link at time t offer a complementary solution to the limitations of terresb1 Intent to network vocabulary mapping function trial backhaul. The integration of Low Earth Orbit (LEO) b2 Network vocabulary to QoS mapping function satellite constellations with terrestrial networks provides L̃i,j Delay for intersatellite link xu User admission variable t wide-area coverage, mobility, and global accessibility [17]. L̃max Maximum delay Unlike traditional geostationary systems, LEO satellites Rmax Maximum data rate offer lower latency and higher throughput, making them Ei Energy consumption at IAB node i ∈ S ∪ N MEC server activation decision variable at IAB node i ait suitable for backhaul support in dynamic or temporary zte Activation of link e at time t rural operations. Several studies have investigated satelliteassisted 5G networks [18], [19], in which satellites serve as backhaul providers for remote base stations or IAB stations [20], [21]. However, while satellite-assisted IAB terrestrial integration, and intent-based management, the networks enhance coverage, they introduce new challenges, coordination between temporary field networks (e.g., during including link intermittency, increased handover complexity, agricultural or forestry operations) and fixed residential and delay variations due to satellite motion. Addition- networks to minimize energy consumption while satisfyally, resource management across satellite and terrestrial ing data rate requirements remains largely unexplored. segments requires joint optimization of radio, backhaul, Moreover, current related works lack mechanisms that and computational resources to achieve seamless service dynamically interpret user or operator intent to control continuity. Furthermore, rural activities, such as seasonal when and where to activate network functions, scale farming [22] or mining using connected devices, create capacity, and ensure energy efficiency. This gap motivates temporally varying connectivity demands that require the development of an Intent-Aware Satellite-Assisted IABbased FWA for rural areas. By leveraging user intents, adaptive network configurations. Intent-based Networking for Dynamic Resource Allo- such as farm operation schedules or machinery usage cation: The concept of IBN has gained attention as a patterns, the network can proactively deploy temporary paradigm that allows users or applications to express high- connectivity, dynamically scale backhaul links, and delevel goals or intents, which are automatically translated activate network functions when activities cease. This into low-level network configurations [7]. Intent-based intent-driven coordination between satellite backhaul and management frameworks have been applied in data centers IAB terrestrial nodes enables both energy-efficient and and enterprise networks to simplify automation, improve on-demand connectivity for rural operations. Overall, agility, and reduce operational complexity. In mobile and the proposed Intent-Aware Satellite-Assisted IAB-based edge networks, intents can express service requirements FWA introduces a new paradigm for rural connectivity such as latency bounds, bandwidth guarantees, or reliability by combining intent-based automation, dynamic resource levels [23], [24]. Recent studies have explored intent- management, and hybrid satellite–terrestrial networking. aware orchestration for next-generation networks [25]. For It enables a flexible, context-driven, and energy-efficient example, [26] proposed intent-driven network slicing to communication network that supports both temporary field allocate resources based on service-level intents dynamically. and residential broadband services. Rural regions exhibit dynamic population movements III. System Model and activity-dependent connectivity needs, which require context-aware intent interpretation and adaptive resource In Fig. 1, we consider a system model with one gateway orchestration across terrestrial and non-terrestrial segments as an IAB donor, multiple Low Earth Orbit (LEO) satellites [27]. The coexistence of residential and field-based network as IAB nodes, and terrestrial fixed and mobile IAB nodes. demands in an intent-based network has not been explored In other words, each LEO satellite has on-board processing, in existing literature for rural areas. i.e., a regenerative satellite, that provides uplink and Research Gap and Motivation: Although prior studies downlink services to terrestrial IAB nodes and the donor. have addressed FWA and IAB architectures, satellite- We denote S as a set of satellite IAB nodes and N as a set

Orbit segment

m-DU m-DU

m-DU

m-MT

MT

UE m-DU

m-IAB node

m-IAB node

m-IAB node

UE

m-MT

m-MT

DU

IAB node

m-MT

m-IAB node

CPE

CPE

Fixed network

Temporary network

CPE DU

CU

INO

Fiber

IAB donor Weak link Strong link Laser inter-satellite link

Figure 1. Illustration of our system model for a rural area.

the orbital segment. Depending on the location of the field area requiring temporary network access and the residential areas requiring an FWA network, either circular or elliptical INO: Intent-based Network Orchestrator orbits can be considered. For instance, in polar regions, elliptical orbits are preferred over circular ones because Intent to Network Vocabulary Mapping their high inclination enables them to pass over the poles, providing better coverage. In this work, we focus on a LLM segment of the orbit [28] and LEO satellites. Here, we Network Vocabulary to QoS Requirements remind you that the LEO satellite can have an elliptical Mapping INO orbit, in which the satellite altitude varies between perigee (the closest point) and apogee (the farthest point). If Conflict Detection both the perigee and apogee altitudes of the satellite lie and Resolution within the LEO range (100-2000 km), then the satellite is considered to be in a LEO elliptical orbit. The choice Network Setup and Configuration between a circular and an elliptical orbit for the IAB station in our solution can be made based on the rural area’s geographic location. Furthermore, we consider T to be the Figure 2. Illustration of INO-Intent-based Network Orchestrator. period during which the satellite travels in the rural area and is slotted. Therefore, we define set ∈ {0, 1} as visibility of terrestrial fixed and mobile IAB nodes. Each fixed IAB of satellite s ∈ S for forming link e ∈ E at time t during node comprises a DU and an MT unit. On the other hand, the pass arc, where E is the set of directed links. The key each mobile IAB node comprises a mobile Distributed Unit notations used in this paper are shown in Table I. (m-DU) and a mobile Mobile Termination (m-MT) unit. IV. Problem Formulation The DU/ m-DU serves both CPEs and downstream MTs/mMTs. The MT/m-MT allows the node to operate as a relay As shown in Fig. 2, in this section, we discuss the problem when connecting to its parent DU/m-DU. Furthermore, formulation for Intent-to-network QoS mapping and conflict the IAB donor includes both a DU and a CU. Unless stated resolution, and optimization to maximize energy efficiency otherwise, we use the term IAB station to refer to either an for establishing the network shown in Fig. 1. IAB node, m-IAB node, or an IAB donor. The terrestrial IAB stations serve user terminals, i.e., user devices (UEs) A. Intent Profiling and Customer Premises Equipment (CPE), in rural areas under intent-driven demands. Unless stated otherwise, we Here, we assume that the FWA exists in rural areas use the term terminal to refer to either a UE or a CPE. and intent helps to improve it and establish a temporary We denote U as a set of user terminals and M as a set of network. For establishing temporal and improving FWA user intents. network shown in Fig. 1, we assume each user of a terminal In serving user terminals U in rural areas, we consider u ∈ U in residential expresses an intent at an abstract level Application Layer for Intent Profiling Intent 3 Intent 2 Intent 1

Intent M

for time t and location l, denoted by mut ∈ M, where M = {1, 2, . . . , M } represents the set of all possible intents. The intents are sent to the Intent-based Network Orchestrator (INO) via the application layer. This process is called intent profiling. Here are some examples of user intents: • User 1: Real-time monitoring of soil moisture sensors with minimal energy usage for a specific location and time. • User 2: Streaming HD farm activities to the farm office, farm location, and time. • User 3: Secure remote control of mining equipment, mining site location, and time.

corresponding to all vocabulary elements derived from the user’s intent mut as follows: [ Qut = b2 (v). (4) v∈Vtu

This is an example of intent-to-network QoS mapping using b2: • User 1: {q1 , q2 , q5 } (Latency ≤ 50 ms, Packet loss ≤ 1%, Uptime ≥ 99% ). • User 2: {q3 , q2 , q5 } (Throughput ≥ 20 Mbps, Packet loss ≤ 1%, Uptime ≥ 99%). • User 3: {q1 , q4 , q4 , q5 } (Latency ≤ 50 ms, Encryption enabled, Packet loss ≤ 1%, Uptime ≥ 99%).

B. Intent to Network Vocabulary Mapping D. Composite Mapping of Intent to QoS Requirements We assume the INO has a network vocabulary, where By combining the above mapping functions b1 and b2 , V = {1, 2, . . . , V } denotes the set of vocabulary elements. we define a composite function that directly associates user The following are examples of the network vocabulary: v1 : intents with network QoS requirements: low latency, v2 : energy efficiency, v3 : high throughput, v4 : secure connection, v5 : reliable connectivity. b = b2 ◦ b1 : M → Q. (5) When intent reaches the INO at the IAB donor, the INO must search the network vocabulary that matches Therefore, for each user terminal u ∈ U, the final QoS u the user’s intent to understand the user’s requirements. requirement derived from its intent mt is given by [ Therefore, each user intent needs to be matched with one b(mut ) = b2 (v). (6) or more network vocabulary elements using the following v∈b1 (mu t) mapping function: b1 : M → V.

(1)

For each user terminal u ∈ U, the corresponding network vocabulary elements for its intent mut can be expressed as: Vtu = b1 (mut ),

(2)

where Vtu ⊆ V represents the subset of network vocabulary elements associated with intent mut . This is an example of intent-to-network-vocabulary mapping using b1 : • User 1: {v1 , v2 , v5 } (low latency, energy efficiency, reliable connectivity). • User 2: {v3 , v5 } (high throughput, reliable connectivity). • User 3: {v1 , v4 , v5 } (low latency, secure connection, reliable connectivity). C. Network Vocabulary to QoS Requirements Mapping We consider Q = {1, 2, . . . , Q} as the set of network Quality of Service (QoS) needed in rural areas, such as latency, throughput, reliability, and energy efficiency. Each network vocabulary element v ∈ V is mapped to one or more QoS parameters by the following function: b2 : V → Q.

E. Conflict Detection and Resolution Conflicts may arise when multiple users simultaneously request network services with QoS requirements that cannot be jointly satisfied within the same time interval at the same location in a rural area. To model this, we define a conflict-free indicator function between any two users u ∈ U and w ∈ U as:  u w  1, if both b(mt ) and b(mt ) c(u, w) = (7) can be simultaneously satisfied,   0, otherwise. To handle the above conflicts, we introduce a binary user-admission variable xut for each user u, where xut = 1 indicates that the user’s intent can be satisfied in the current scheduling period. Otherwise, xut = 0. Since the network’s QoS requirements are numerous, we focus here on delay and data rate requirements. Other requirements, such as packet loss and error rate, will be considered for future work. Then, we formulate the following user admission optimization as a 0-1 integer linear program: max x

(3)

Hence, for an extracted vocabulary element v from an intent, the associated QoS parameters are Qv = b2 (v), where Qv ⊆ Q. Here, we assume that the network provider has the network vocabulary-QoS requirements mapping. For each user terminal u ∈ U, the overall QoS requirements can be obtained by aggregating the QoS sets

s.t.

U X u=1 U X u=1 U X

xut dut xut ≤ Rmax ,

∀t, (8)

Lut xut ≤ L̃max ,

∀u,

u=1 xut + xw ∀ 1 ≤ u < w ≤ U, t ≤ 1 + c(u, w), u w xt , xt ∈ {0, 1}, ∀u, w ∈ U.

F. Network Setup and Configuration After identifying the total number of users the network can accommodate and the network’s QoS requirements, the next step is to set up and configure the network. We consider satellites as mobile IAB nodes, and terrestrial mobile and fixed IAB nodes, where m-DU, m-MT, DU, and MT can be deployed as Virtual Network Functions (VNFs) [29]. VNFs require a computer to host them. Here, VNFs are software implementations that run on general-purpose computing hardware, often referred to as commercial offthe-shelf (COTS) servers. Here, we consider a small Multiaccess Edge Computing (MEC) server. In other words, VNFs can be activated when needed. When they are not needed, they can be disabled to save energy. Each satellite with active VNFs can be assigned a specific area called a footprint, based on its orbit segment. Here, we assume the satellite’s orbit segment is known and predictable. During the satellite pass, the satellite can serve terrestrial mobile and fixed IAB nodes within its footprint. We model the power consumption at IAB station i at time t as: φit = Υibase + Υit νti , (9)

Satellite 1

Link quality based on RSSI

where Rmax denotes the maximum data rate the network can provide at a time t and dut is the data rate required for each user terminal u based on intent analysis. Furthermore, we define Lut as the latency requirement for each user terminal u based on intent analysis. L̃max represents the maximum tolerable end-to-end latency, accounting for all user terminals that need to be connected to the network. The constraint xut + xw t ≤ 1 + c(u, w) ensures that conflicting users are not admitted simultaneously. This means that if users u and w are in conflict, i.e., c(u, w) = 0, they cannot both be active at the same time; one must be denied based on the time of arrival of the intent. In the absence of conflict, both users can be accepted. The objective function maximizes the total number of users that the network can admit, subject to network capacity and end-to-end latency constraints.

Handover threshold

Satellite 2

Strong link for satellite 1

Strong link for satellite 2

Time

Figure 3. Link quality and handover illustration.

where U i is the number of user terminals connected to IAB i,v node i, λ̃i,v t is packet arrival rate, and µ̃t is service rate at time t. We define a decision variable ait ∈ {0, 1} for activating an MEC server with VNFs at IAB node i ∈ S ∪ N in slot t, where ∆t is the slot duration, such that: ( 1, if νti ̸= 0, activate MEC server with VNFs, i at = 0, otherwise. (12) Then, we define the energy consumption of IAB node i as follows: E i = (Υibase + ait Υit νti )∆t. (13)

In other words, when IAB functions are not activated, there is no need to use the MEC server. Therefore, when ait = 0, ait Υit νti ∆t = 0 and the energy consumption when the IAB node with MEC server and VNFs disabled in timeslot t becomes E i = Υibase ∆t. Inter-satellite backhaul: Let us consider two interconnected LEO satellites, i.e., m-IAB nodes, denoted as i and j. We assume these satellites maintain perfect antenna alignment and that the intersatellite link experiences negligible interference from signals transmitted by other preceding or succeeding LEO satellites. Consequently, the intersatellite link capacity between the LEO satellites i and j can be expressed as follows: i  where Υbase is the baseline power of IAB station i ∈ S ∪ N . Pti (Di,j )−κ Gjt  i,j i,j i C = B log 1 + , ∀e = (i → j) ∈ E We define Υt as the power of activating the MEC server 2 t σ2 with VNFs at time t such that: (14) χ̃ where Pti is the transmission power when the IAB node X Υit = ψbase + ψχi , (10) with MEC server and VNFs is active. Di,j is the distance χ=1 between the satellites i and j and −κ represents the path i,j where χ̃ denotes the number of VNFs deployed at the loss exponent. Furthermore, we define de delay L̃ for MEC server, including m-DU, m-MT, DU, and MT. We intersatellite link as: P u u define ψχi as the power consumption associated with an u∈U i,j xt dt ∆t i,j L̃ = (15) i,j active VNF, while ψbase represents the baseline power Ct consumption when the MEC server is powered on but not i,j processing any traffic. Furthermore, the power consumption where U is the set of terminals using intersallelite link of VNF depends on the number of terminals connected between satellites i and j for transmission duration ∆t. Link between satellite IAB and terrestrial IAB nodes: to it. Therefore, we define νti as the traffic load at VNFs We assume that satellite passes are known and predictable deployed in the MEC server at IAB station i such that: based on publicly available satellite orbital datasets, such Ui X λ̃i,v as those provided by CelesTrak [30]. As shown in Fig. 3, t i νt = , (11) i,v the terrestrial IAB station, such as the IAB node, m-IAB µ̃ u=1 t

node, and IAB donor, continuously measures the received signal strength indicator (RSSI) of available satellites in its area. At any given time, the link with the highest RSSI is selected as the strong link, while the remaining links are maintained as weak links. When the RSSI of the current strong link falls below a predefined handover threshold, the IAB station reclassifies this link as a weak link and promotes the link with the highest RSSI to become the new strong link for handover. This dynamic link selection mechanism is similar to the handover mechanism discussed in [31]. Furthermore, IAB stations need to inform the IAB donor about its available links so that the IAB donor maintains details of all strong and weak links to reach each IAB station. This mechanism enables seamless connectivity and improves link robustness under time-varying satellite visibility conditions. Therefore, we consider the handover delay L̃j,k h between terrestrial IAB station k with satellite j (m-IAB node), which is defined as the delay associated with the handover process and is given by: L̃j,k h = (1 − Ph )Tr + Ph Th ,

(16)

where Bu is the channel bandwidth, Gut is the channel gain, and σ 2 is the noise power. Furthermore, we define the latency of the access link as follows: L̃u,k =

xut dut ∆t Ctu,k

.

(21)

We formulate the following optimization problem to maximize energy efficiency, specifically the data rate per energy consumption: P du P u∈U t i max (22) ait ,set ,du t i∈S∪N E s.t.

Ctu,k ≥ xut dut , ∀u ∈ U, t ∈ T , X xut dut ≤ set Ctj,k ∀ t ∈ T ,

(22a) (22b)

u∈U j,k

X

xut dut ≤ set Cti,j , ∀e ∈ E, t ∈ T ,

(22c)

u∈U i,j set ≤ ait , set ≤ ajt , ∀e = (i → j) ∈ E, (22d) U X j,k L̃u,k + L̃j,k + L̃i,j ≤ L̃max , ∀u ∈ U. (22e) h + L̃ u=1

where Ph denotes the probability of a successful handover. The constraints in (22a), (22b), and (22c) ensure that the Specifically, Ph = 1 if the handover succeeds and Ph = 0 links meet the user terminals’ data rate requirements, given otherwise. The parameter Th represents the time required the satellites’ visibility. Constraint (22d) guarantees that to complete a successful handover, while Tr denotes the the satellites are activated as m-IAB nodes when they are additional time incurred to retry the handover in the event visible. Constraint (22e) ensures that the total latency is of failure. Furthermore, we define the capacity of the link satisfied. between terrestrial IAB station k and satellite m-IAB node j as follows: V. Solution Approach !
−κ j In this section, we present a detailed solution approach P Dj,k Gkt Ctj,k = B k,j log2 1 + t , (17) for intent-to-QoS mapping, conflict resolution, and the 2 σ optimization problem to maximize energy efficiency. where Dj,k is the distance between terresterial IAB station k with satellite j. The distance Dj,k can be expressed as follows: (κj + ξ j ) sin β j Dj,k = (18) cos θj where κj is Earth’s radius and ξ j is the orbit height, i.e., the altitude from a point in a rural area directly below satellite j. We define β j as the coverage angle and θj as the elevation angle. Considering the capacity of the link Ctj,k and terminals using terrestrial IAB station k connected to satellite j, we can define delay as follows: P u u j,k x d ∆t j,k L̃h = u∈U j,kt t , (19) Ct where U j,k is the user terminals connected to satellite j via IAB station k in transmission duration ∆t. Access link between terrestrial IAB station and terminals: The capacity of the access link between terminal u and terrestrial IAB station k is defined as follows:  P k Gu  Ctu,k = akt B u log2 1 + t 2 t , (20) σ

A. LLM and Optimization Approach for Intent-to-QoS Mapping and Conflict Resolution The mapping functions b1 : M → V and b2 : V → Q are unknown because the intents are expressed at an abstract level using natural language. Therefore, we can use AI models, such as LLMs [32], to learn these functions from data. We choose LLMs over other AI frameworks because they can analyze user intents expressed at an abstract level by interpreting natural-language intents and translating them into executable network configurations or policies [33]. In other words, the LLM model can help us automatically learn how user intents map to network vocabulary and corresponding QoS requirements. For b1 : M → V, the first LLM learns a semantic mapping between user intents and the network vocabulary. Given an intent expressed as a high-level operational objective, the process begins with sentence segmentation. Specifically, each sentence g u from the user intent mut and each vocabulary term v ∈ V are encoded into embedding vectors using a pre-trained sentence embedding model h1 (·), yielding sentence embedding ou = h1 (g u ) and network vocabulary embedding wu = h1 (v). To mitigate LLM hallucinations during the semantic mapping process, we

employ a pre-trained sentence-level LLM and then fine-tune After end-to-end intent-to-QoS mapping, we minimize the model. In the fine-tuning process, we consider positive (8), a binary (0-1) linear program that can be solved and negative training pairs (g u , v, ygv ), where ygv = 1 if the with modern mixed-integer programming (MIP) solvers network vocabulary term v is semantically associated with such as Gurobi or CPLEX. Although NP-hard in general, sentence g u (positive pair), and ygv = 0 otherwise (negative such solvers employ branch-and-bound and cutting-plane pair). The pre-trained LLM is fine-tuned by minimizing techniques that can efficiently handle such formulated the following cosine-similarity-based loss function: optimization problems. Furthermore, the LLMs and optiX mization approach for intent-to-QoS mapping and conflict 1 2 (cos_sim(ou , wu ) − ygv ) , (23) resolution needed to be implemented in the INO available L1 = |Ṽ| u (g ,v)∈Ṽ at the IAB donor. In other words, the IAB donor is always where the cosine similarity between embeddings is defined active. as: ou · w u cos_sim(ou , wu ) = u u . (24) B. Solution for Energy Efficiency Maximization Problem |o ||w | In network setup and configuration, we need to solve the Ṽ is the set of all sentences and network vocabulary pairs. formulated optimization problem (22) to maximize energy Once the model is fine-tuned, each sentence g u is mapped efficiency. The formulated problem (22) combined node and to the most relevant vocabulary terms v by computing link activation with traffic routing, where traffic routing cos_sim(ou , wu ) and minimizing the cosine similarity loss uses activated nodes and links. Therefore, node and link function. This approach ensures that both semantically activation should start, followed by traffic routing. This meaningful and quantitatively precise mappings are ob- motivates us to simplify and decompose the formulated tained between user intents and network vocabulary. problem. To solve (22), we first simplify the objective For b2 : V → Q, the second LLM learns a semantic function and then propose a practical two-stage Bender mapping between network vocabulary and QoS vocabulary. decomposition algorithm as a solution. In other words, we In this work, we consider a semantic mapping from ex- first reformulate the objective and then develop a two-stage tracted network-related vocabulary in intent to a predefined Benders decomposition framework that separates node and QoS vocabulary. We use a pre-trained embedding LLM link activation decisions from traffic routing. model h2 (·) with extracted network vocabulary embedding To simplify the objective function, let us assume that, wu = h2 (v) and QoS requirement embedding qu = h2 (q). for each time slot t, the per-terminal data rate demand du t Then, for LLM fine-tuning, we minimize cosine similarity can be satisfied whenever terminal u is admitted, and that loss defined as follows: the admission variable xut is obtained from solving (8). In
1 X 2 L2 = cos_sim(wu , qu ) − yvq (25) other words, if the terminal is not admitted, it will not |Q̃| be connected to the network. The achievable data rate of (v,q)∈Q̃ terminal u is therefore where Q̃ is the set of all network vocabulary and QoS Ctu,k = xut dut , ∀u ∈ U, ∀t. (27) vocabulary pairs. The yvq = 1 if the extracted network vocabulary term v is semantically associated with QoS term The total system throughput at time slot t becomes q (positive pair), and yvq = 0 otherwise (negative pair). X u,k X Rttot ≜ Ct = xut dut . (28) For example, if the extracted network vocabulary loweru∈U u∈U delay is matched with the lower-latency QoS vocabulary, yvq = 1. Suppose lower delay is matched with throughput, Under this assumption, maximizing the ratio, i.e., energy yvq = 0. In other words, yvq = 1 if similar meaning, efficiency with a known numerator, is equivalent to minii.e., high similarity. For a different meaning, i.e., lower mize the denominator, i.e., total energy consumption for similarity, yvq = 0. Similarity between embeddings v and each time slot: q is measured using the cosine similarity cos_sim(wu , qu ). X min Ei (29) For example if v is lower delay and q is latency, cosine i {a } t u i∈S∪N similarity is cos_sim(w , qu ) = 0.85 and lvq = 1. The subject to: (22a) - (22e). (29a) loss function is L2 = (0.85 − 1)2 = 0.0225. This is a good mapping because the 0.0225 loss is too small. After To minimize energy consumption, we consider activating fine-tuning the LLM, the embeddings become domainonly a subset of nodes and links that can route all traffic specific for the network, improving semantic alignment demands while satisfying link capacity, satellite visibility, for QoS terms. Finally, we formulate end-to-end intent-toand latency constraints, rather than keeping all nodes QoS mapping, where we can learn the composite mapping and links active at all times. To obtain a self-contained function b = b2 ◦ b1 : M → K by minimizing the following formulation, we introduce a multi-commodity flow variable loss function: fte,u . For each terminal u ∈ U, directed link e = (i → j), L = λ1 L1 + λ2 L2 . (26) and time slot t, let fte,u ≥ 0 denote the flow rate of Here, λ1 > 0 and λ2 > 0 are the weights to balance the terminal link e. The aggregate data frafic on link P u over e,u contributions of each loss function. e is f . For a fixed time slot t, the problem u∈U t

can be formulated as the following Mixed-Integer Linear Programming (MILP): X min Ei (30)

The objective function in (31) minimizes node activation energy consumption, while the constraints ensure that node activations depend on satellite visibility and data rate requirements. (a,s,z,f ) i∈S∪N Stage II: Subproblem for Routing Feasibility: Given s.t satellite visibility, activated nodes, and links (a = {ait }, s = X e,u e e ft ≤ Ctu,k set zte , ∀e, (30a) {st }, z = {zt }) obtained from Stage I, in the Stage II, (30) becomes linear program (LP) by only considering u∈U  constraints in (30a), (30b), (30c), and (30d) to determine u  X e,u X e,u dt , i = src(u), flows f = {fte,u } that respect link capacity and satisfy u ft − ft = −dt , i = dst(u), latency requirements. If the LP is infeasible, Benders   e:(i→j) e:(j→i) 0, otherwise, feasibility cuts are generated and added to the master (30b) problem (31). The master problem is then re-solved. If  the LP is feasible, the node and link activation and u dt , j = src(u), X e,u X e,u  routing decisions obtained constitute an optimal solution u ft − ft = −dt , j = dst(u), to the original MILP. The overall two-stage Benders   e:(j→k) e:(j→k) 0, otherwise, decomposition algorithm is summarized as follows: (30c) j,k L̃u,k + L̃j,k + L̃i,j ≤ L̃max , (30d) Algorithm 1 Two-Stage Benders Decomposition for h + L̃ Energy-Efficient IAB Activation. set ≤ ait , set ≤ ajt , (30e) 1: Preconditions: Satellite visibility is predictable, zte ≤ ait , zte ≤ ajt , ∀e = (i → j) ∈ E, (30f) LLMs and conflict resolution are performed to get e,u i e e admitted users’ requirements; at ∈ {0, 1}, zt ∈ {0, 1}, st ∈ {0, 1}, ft ≥ 0, (30g) u,k i,j j,k 2: Input: Demands {du t }, capacities {Ct , Ct , Ct }, e e where zt is link e activation decision variable at time t. visibility {st }, latency bound L̃max ; Problem (30) ensures the link is activated when the 3: Precompute candidate nodes and links by solving nodes that make the link are visible and activated. Then, master MILP (31); traffic routing on the link can occur when the link is 4: repeat activated. In other words, (30) jointly optimizes node/link Solve routing (30) LP by only considering con5: activation and traffic routing. We propose a two-stage straints in (30a), (30b), (30c), and (30d); Benders decomposition approach [15] that separates node 6: if LP infeasible then and link activation decisions (the master problem) from 7: Generate Benders feasibility cuts and add to traffic routing (the subproblem). We choose the two-stage master; Benders decomposition approach over other optimization 8: end if methods due to its decomposition capability, scalability, 9: until LP feasible; and ability to handle uncertainty [34]. 10: Output: Variables a, s, z, f and optimal energy efficiency. Stage I: Master Problem for Node and Link Activation. For each terminal u, data-rate and latency requirements are known from intent analysis, and satellite visibility is Remark 1. Computational Complexity of the Proposed predictable. Based on the terminal’s location and satellite Approach is O(n3 ). visibility, we determine candidate nodes and links to The LLM-based mapping model h1 (·) and h2 (·) are preconnect the terminals to the internet. We formulate the trained using datasets (M, V ) and (V, Q), respectively. For master problem as follows: LLM model with θ1 and θ2 parameters, the mapping has X min Ei (31) complexity O(θ1 |V|+θ2 |Q|), while the LLM model fine tun(a,s,z) i∈S∪N ing considering ζ epochs has complexity O(ζ(θ1 |V|+θ2 |Q|)). s.t. In other words, during LLM inference, complexity reduces X u,k e e u dt ≤ Ct st zt , (31a) to O(θ1 |V| + θ2 |Q|), which scales linearly with model size. After LLM-based mapping, the conflict resolution problem u∈U X is formulated as a binary linear program with xu as a binary xut dut ≤ set Ctj,k , (31b) variable. In the worst case, when the vectort x = {xu } t u∈U j,k becomes large, solving it has a computational complexity X i,j xut dut ≤ zte set Ct , (31c) of O(2n ), where n is the dimension of x. i,j u∈U The proposed two-stage Benders decomposition splits set ≤ ait , zte set ≤ ajt , (31d) (30) into sub-problems for node/link action activation and routing sub-problems. Let n denote the number of Benders zte ≤ ait , zte ≤ ajt , ∀e = (i → j) ∈ E, (31e) iterations. The master problem, defined over binary variable ait ∈ {0, 1}, zte ∈ {0, 1}, set ∈ {0, 1}. (31f) variables (a = {ai }, s = {se }, z = {z e }), has worstt t t

0.030 Loss (MSE)

0.025 0.020 0.015 0.010 0.005 0.000

MiniL

-v2 -v2 -v2 e-v2 -v1.5 M-L6 niLM-L12 MPNet Base E5-Bas Mi BGELLM Models

Energy (J)

Figure 6. LLMs comparison in terms of MSE minimization.

Figure 4. Selected region for rural areas.

FixedIABB

Figure 7. LLMs comparison in terms of energy consumption.

SatelliteB MobileIAB

SatelliteA

-v2 t-v2 e-v2 2-v2 -v1.5 M-L1 MPNe Base E5-Bas MiniL BGELLM Models

M-L6

MiniL

SatelliteC Gateway

1750 1500 1250 1000 750 500 250 0

FixedIABA Figure 5. Interconnection between satellites and terrestrial IAB stations.

3.10.12 language [35] is used for numerical analysis. For intent-to-QoS mapping, we use all-MiniLM-L6-v2 from Hugging Face [36], [37] as the sentence-transformer foundation LLM. For solving optimization problems, we use PuLP [38], a Python-based linear and mixed-integer programming modeler, with the GLPK solver [39].

case complexity O(n3 ), when vectors a, s,and z become large. The subproblem in Stage II, which is for routing A. Simulation Setup feasibility, Given the satellite visibility, activated nodes and, links (a, s, z) obtained from Stage I, in the Stage To create a network topology, we randomly choose a II, (30) becomes linear program (LP) by only considering rural region in Kenya, Africa. The selected region is shown constraints in (30a), (30b), (30c), and (30d) to determine in Fig. 4. In that region, we have two residential areas, each flows f . This has computation complexity O((n). Overall, served by one terrestrial IAB node. Area A has 15 houses the total complexity of two-stage Benders decomposition is served by fixed IAB node A, while Area B has 20 houses
O n3 + n . The proposed approach significantly improves served by fixed IAB node B. Houses are in black, while scalability by decoupling routing from node and link IAB nodes A and B are in blue. The population of these activation decisions. residential areas uses a farming area located in the region, By accounting for LLM inference, conflict resolution, where IAB functionality is mounted on an agricultural truck and the two-stage Benders decomposition, the overall as a mobile IAB to serve 10 connected Internet of Things computational complexity of the proposed approach is (IoT) devices (these devices are in orange). In the region,
given by O(θ1 |V| + θ2 |Q|) + O(2n ) + O n3 + n . Since the we have one IAB donor in green. In other words, the IAB cubic term dominates for large n, the overall computational donor/gateway and fixed IAB node locations (latitudes complexity of the proposed method is O(n3 ). This level and longitudes) are fixed. of complexity remains acceptable for rural deployment To obtain the list of satellites and the times of each scenarios, where the number of terminals is relatively low satellite pass in the region, we use the OneWeb satellite compared to urban and sub-urban areas. dataset from CelesTrak [30] [40], which includes 651 satellites. For loading the dataset and tracking satellites VI. Simulation Results and Analysis in the region, we use the Skyfield Python library [41]. In this section, we present the performance evaluation of Based on the latitudes and longitudes of terrestrial fixed the proposed LLM-driven intent-aware Satellite–IAB–based or mobile IAB stations, we continuously track and select FWA approach to connect rural areas, where Python three satellites that maximize the elevation angle θj to

LLM fine tuning loss

0.10 0.08 0.06 0.04 0.02 0.00

35 Number of admitted nodes

L with 1=0.01, 2=3 L with 1=0.1, 2=0.5 L with 1=1, 2=1 L with 1=1.5, 2=0.6 L with 1=3, 2=0.9 L1 L2

35

35

34

30 25 20 15 10 5

0

Figure 8. tuning.

10

20 Epochs

30

0

40

Loss function minimization for LLM fine

35

0.70

ry

0.65

oca Sv Qo

delay

25

Admitted nodes

Mapping score

0.75

bu la

packet loss rate

all-MiniLM-L6-v2 mapping score

throughput

0.80

Greedy

Optimal FCFS Greedy

30

0.85 0.80 0.75 0.70 0.65 0.60

FCFS

Figure 10. Summary of admitted nodes.

0.85

{'delay': 1000} {'packet loss rate': 0.55, 'bandwidth': 610.0} {'delay': 1000} {'bandwidth': 80.0} {'delay': 1000} {'bandwidth': 350.0} {'packet loss rate': 0.93} {'delay': 1000} {'bandwidth': 370.0} {'delay': 1000}

Optimal ILP

20 15 10 5 0

1.64

1.65

1.66

1.67

Time (seconds)

1.68

1.69

1.70

1e9

Figure 11. Admitted nodes over time.

0.60

inter-satellite, feeder, and satellite-to-terrestrial backhaul links. To establish the temporal network in farming areas and to coordinate temporal and FWA in residential areas A and B, we assume that users in areas A and B express their interests as intents. For the intent expression, we use the intent dataset available in [42]. To map intent to Figure 9. Mapping network vocabulary to QoS using the network vocabulary, we use the network vocabulary LLM. from [43] and the LLM all-MiniLM-L6-v2 [36]. To map the extracted network vocabulary to network QoS, we that region to establish intersatellite links connecting rural- use the 5G QoS requirements [44] and all-MiniLM-L6-v2. area nodes to the core network via an IAB donor or Therefore, among the many sentence-level LLMs available, gateway. Furthermore, each terrestrial IAB station selects the rationale for selecting all-MiniLM-L6-v2 over other two satellite links based on RSSI, with the strongest link models is discussed in the next subsection. being the strong link and the second link being the weak link. The interconnection between satellites and terrestrial B. Simulation Results IAB stations is shown in Fig. 5. The link capacities are computed based on the physical For intent-to-network vocabulary mapping and mapping characteristics of each connection in the network graph. extracted network vocabulary to network QoS requirements, For every link, the geographical distance between nodes is numerous sentence-level LLMs are available for semantic first calculated using the Haversine formula. The distance embedding tasks; selecting an appropriate model requires a is then used to estimate the signal propagation conditions. systematic comparison of both semantic representation Each link is categorized into one of three types: laser quality and computational efficiency. To this end, we intersatellite links, feeder or satellite-to-terrestrial backhaul evaluate five widely used sentence embedding models, links, and access links between fixed or mobile IAB nodes namely all-MiniLM-L6-v2 [36], [45], all-MiniLM-L12-v2 with CPEs or IoT devices. For instance, laser intersatellite [45], [46], all-mpnet-base-v2 [47], BGE-Base-v1.5 [48], links operate at 193 THz with 500 MHz bandwidth. In and E5-Base-v2 [49]. The comparison is conducted using contrast, feeder and satellite-to-terrestrial backhaul links semantic similarity learning. Specifically, each model is use Ka-band (26.5 GHz) with 250 MHz bandwidth, while fine-tuned on sentence pairs with similarity labels, where access links use sub-6 GHz (3.5 GHz) with 100 MHz semantically related pairs are assigned higher similarity bandwidth. The baseline scenario discussed in [20], [21] uses scores and unrelated pairs are assigned lower scores. Cosia 2 GHz carrier frequency and a 40 MHz bandwidth for neSimilarityLoss is adopted during training to optimize the

400 300

-33.3%

250 200

Latency (ms)

30 20 10

150 0

100 50 0

All IAB nodes active

Our approach: some IAB nodes active

Figure 12. Power consumption based on IAB activation. Our approach 3000 Baseline

Gateway-SatelliteA Gateway-SatelliteB FixedIABA-H_A1 FixedIABA-H_A2 FixedIABA-H_A3 FixedIABA-H_A4 FixedIABA-H_A5 FixedIABA-H_A6 FixedIABA-H_A7 FixedIABA-H_A8 FixedIABA-H_A9 FixedIABA-H_A10 FixedIABA-H_A11 FixedIABA-H_A12 FixedIABA-H_A13 FixedIABA-H_A14 FixedIABA-H_A15 FixedIABA-SatelliteA FixedIABA-SatelliteB FixedIABB-H_B1 FixedIABB-H_B2 FixedIABB-H_B3 FixedIABB-H_B4 FixedIABB-H_B5 FixedIABB-H_B6 FixedIABB-H_B7 FixedIABB-H_B8 FixedIABB-H_B9 FixedIABB-H_B10 FixedIABB-H_B11 FixedIABB-H_B12 FixedIABB-H_B13 FixedIABB-H_B14 FixedIABB-H_B15 FixedIABB-H_B16 FixedIABB-H_B17 FixedIABB-H_B18 FixedIABB-H_B19 FixedIABB-H_B20 FixedIABB-SatelliteB FixedIABB-SatelliteC MobileIAB-D1 MobileIAB-D2 MobileIAB-D3 MobileIAB-D4 MobileIAB-D5 MobileIAB-D6 MobileIAB-D7 MobileIAB-D8 MobileIAB-D9 MobileIAB-D10 MobileIAB-SatelliteB MobileIAB-SatelliteC SatelliteA-SatelliteB SatelliteB-SatelliteC

Power consumption (W)

350

Links

Figure 14. 1e6Latency per link.

2500

3.5

2000

3.0

EE (bit/J)

Capacity (Mbps)

Our approach Baseline

40

1500 1000

Our EE - All IAB nodes active Our EE - Some IAB nodes active (+49.8% vs All active, +160.0% vs Baseline) Baseline EE - All IAB nodes active

2.5 2.0 1.5

500

1.0

Gateway-SatelliteA Gateway-SatelliteB FixedIABA-H_A1 FixedIABA-H_A2 FixedIABA-H_A3 FixedIABA-H_A4 FixedIABA-H_A5 FixedIABA-H_A6 FixedIABA-H_A7 FixedIABA-H_A8 FixedIABA-H_A9 FixedIABA-H_A10 FixedIABA-H_A11 FixedIABA-H_A12 FixedIABA-H_A13 FixedIABA-H_A14 FixedIABA-H_A15 FixedIABA-SatelliteA FixedIABA-SatelliteB FixedIABB-H_B1 FixedIABB-H_B2 FixedIABB-H_B3 FixedIABB-H_B4 FixedIABB-H_B5 FixedIABB-H_B6 FixedIABB-H_B7 FixedIABB-H_B8 FixedIABB-H_B9 FixedIABB-H_B10 FixedIABB-H_B11 FixedIABB-H_B12 FixedIABB-H_B13 FixedIABB-H_B14 FixedIABB-H_B15 FixedIABB-H_B16 FixedIABB-H_B17 FixedIABB-H_B18 FixedIABB-H_B19 FixedIABB-H_B20 FixedIABB-SatelliteB FixedIABB-SatelliteC MobileIAB-D1 MobileIAB-D2 MobileIAB-D3 MobileIAB-D4 MobileIAB-D5 MobileIAB-D6 MobileIAB-D7 MobileIAB-D8 MobileIAB-D9 MobileIAB-D10 MobileIAB-SatelliteB MobileIAB-SatelliteC SatelliteA-SatelliteB SatelliteB-SatelliteC

0

0

10

20

Iterations

30

40

Figure 15. Energy efficiency based on IAB activation.

Links

loss function, which corresponds to the mapping from the extracted network vocabulary to network QoS. By jointly considering L1 and L2 , the figure also depicts the composite loss function L defined in (26). We chose L with λ1 = 0.1, semantic embedding space. The evaluation considers both λ2 = 0.5 over other settings. After fine-tuning the alllearning performance and computational cost. Learning MiniLM-L6-v2 model for the aforementioned tasks, Fig. 9 performance is assessed using the mean squared error presents an example of network vocabularies extracted (MSE) between the predicted cosine similarities and the from user intents, mapped to corresponding network QoS ground-truth similarity scores on benchmark sentence pairs. requirements. Following this mapping, the required network To assess computational efficiency, we measure training QoS in rural areas is determined. time, inference latency, CPU utilization, and estimated To provision such a network and estimate the number of energy consumption. Energy consumption is approximated users that can be served in rural areas under non-conflicting from CPU utilization and execution time using a CPU user intents, we solve the admission optimization problem power model. Furthermore, a combined tradeoff metric is in- defined in (8) using PuLP with the GLPK solver. We troduced by jointly considering normalized training loss and then compare the proposed solution with two baseline energy consumption, enabling the identification of models approaches: First-Come, First-Served (FCFS) and a greedy that achieve a favorable balance between semantic accuracy approach. In the greedy method, users are admitted based and computational efficiency. Fig. 6 shows a comparison of on their data rate requirements, starting with those having LLMs in terms of MSE minimization, while Fig.7 shows a the lowest demand. The results of this comparison are comparison of LLMs in terms of energy consumption. Since presented in Figs. 10 and 11. The results show that although MSE is small across all LLMs, below 0.030, we choose LLMs FCFS is computationally efficient, it may reject users even that consume less energy, as the objective of this paper is when resources are still available, resulting in sub-optimal to minimize energy consumption subject to performance performance compared to the proposed optimal solution. constraints. Based on this comparison, the most suitable Fig. 11 shows that the number of terminals varies over sentence-level LLM for the proposed semantic mapping time. During periods with fewer active users, certain IAB task is all-MiniLM-L6-v2. functionalities are deactivated to reduce energy consumpUsing the intent and network vocabulary datasets, Fig. 8 tion. By incorporating this adaptive activation mechanism, illustrates the fine-tuning loss function L1 of the all- Fig. 12 illustrates the resulting power consumption of the MiniLM-L6-v2 LLM model for mapping user intents to proposed approach, which dynamically deactivates IAB network vocabulary. In addition, the figure shows the L2 functions when the number of terminals is low, compared Figure 13. Data rate per link.

to a baseline scheme that keeps all IAB functions continuously active. The simulation results demonstrate that the proposed approach achieves up to a 33.3% reduction in power consumption compared to the baseline. Considering terminal connectivity across the network, Figs. 13 and 14 present the achievable data rate and latency, respectively. The results in these figures show that the proposed approach outperforms the baseline, achieving higher data rates and lower latency. In both approaches, feeder (satellite-to-terrestrial) backhaul links exhibit higher latency compared to access links. This high latency is primarily due to the long propagation distance between terrestrial IAB nodes and satellite IAB nodes. Based on the achievable data rate and power consumption of IAB stations, Fig. 15 compares the energy efficiency (EE) of the proposed approach with the baseline. The proposed approach considers deactivating IAB stations when the number of connected terminals is low, whereas the baseline assumes that all IAB functions remain continuously active. VII. Conclusion This work proposes an LLM-driven, intent-aware satellite–Integrated Access and Backhaul framework for rural connectivity, designed to support both temporary field and fixed household broadband access. By leveraging a large language model, the proposed framework translates users’ intents into explicit network requirements, enabling the network to adapt dynamically to highly heterogeneous, time-varying rural network demands. Based on inferred intents, we developed a satellite-IAB-enabled Fixed Wireless Access optimization framework that jointly manages node and link activation and traffic routing to improve energy efficiency while satisfying data-rate and latency requirements. We formulated the problem as a mixed-integer linear programming (MILP) model and solved it using a two-stage Benders decomposition, thereby improving scalability for large-scale rural networks. Simulation results demonstrated that the proposed intentaware framework significantly improves energy efficiency and resource utilization compared with conventional static rural networking approaches, while maintaining reliable connectivity for both household and field operations. The obtained results highlight the potential of combining LLM-based intent inference, satellite communications, and IAB technologies to enable adaptive, energy-efficient, and intelligent next-generation rural networks. Future work will investigate online learning mechanisms as replacements for traditional optimization approaches to reduce computational complexity, and will conduct experimental validation in realistic rural deployment scenarios. References [1] FAO, “Income-generating rural economic activities,” https://www.fao.org/4/t1675e/t1675e03.htm#:~:text= Rural%20dwellers%20can%20boost%20income,agricultural% 2C%20artisanal%20and%20commercial%20activities, [Online; accessed April 8, 2026].

[2] Maravedis LLC, “5G fixed wireless gigabit services today: An industry overview,” https://shop.maravedis-bwa.com/products/ 5g-fixed-wireless-gigabit-services-today-an-industry-overview, [Online; accessed April 8, 2025]. [3] A. Ndikumana, K. K. Nguyen, and M. Cheriet, “Renewable energy powered and open ran-based architecture for 5G fixed wireless access provisioning in rural areas,” IEEE Transactions on Green Communications and Networking, 2024. [4] A. Ndikumana, K. K. Nguyen, A. Larabi, and M. Cheriet, “Energy-efficient multi-radio microwave and iab-based fixed wireless access for rural areas,” IEEE Transactions on Green Communications and Networking, 2026. [5] A. Ndikumana, K. K. Nguyen, and M. Cheriet, “Digital twin assisted closed-loops for energy-efficient open ran-based fixed wireless access provisioning in rural areas,” in Proceedings of 2023 IEEE Global Communications Conference (GLOBECOM). IEEE, 2023, pp. 6285–6290. [6] Z. Yao, Y. Zhong, Q. Liao, J. Wu, H. Liu, and F. Yang, “Understanding human activity and urban mobility patterns from massive cellphone data: Platform design and applications,” IEEE Intelligent Transportation Systems Magazine, vol. 13, no. 3, pp. 206–219, 2020. [7] A. Leivadeas and M. Falkner, “A survey on intent-based networking,” IEEE Communications Surveys & Tutorials, vol. 25, no. 1, pp. 625–655, 2022. [8] 3GPP-TR, “Technical specification group radio access network; nr; study on integrated access and backhaul (release 16). (3GPP TR tr 38.874 v16.0.0 ),” 2018-12. [9] X. Luo, H.-H. Chen, and Q. Guo, “Leo/vleo satellite communications in 6g and beyond networks–technologies, applications, and challenges,” IEEE Network, vol. 38, no. 5, pp. 273–285, 2024. [10] E.-M. Oproiu, I. Gimiga, and I. Marghescu, “5g fixed wireless access-mobile operator perspective,” in 2018 International Conference on Communications (COMM). IEEE, 2018, pp. 357–360. [11] 3GPP, “3rd generation partnership project; technical specification group radio access network; NR; study on integrated access and backhaul; (release 16), 3gpp tr 38.874 v16.0.0,” 3GPPTechnical Specification, 2018-12. [12] A. Ndikumana, K. K. Nguyen, and M. Cheriet, “Energy-efficient 5g integrated access and backhaul open ran-based fixed wireless access provisioning in rural areas,” IEEE Transactions on Green Communications and Networking, vol. 10, pp. 1642–1653, 2025. [13] E. Yaacoub and M.-S. Alouini, “Efficient fronthaul and backhaul connectivity for IoT traffic in rural areas,” IEEE Internet of Things Magazine, vol. 4, no. 1, pp. 60–66, 2020. [14] J. Jia, H. Wang, X. Xia, and M. Yu, “A network architecture for 5g and 6g satellites based on integrated access and backhaul,” in 2025 International Wireless Communications and Mobile Computing (IWCMC). IEEE, 2025, pp. 144–148. [15] R. Rahmaniani, T. G. Crainic, M. Gendreau, and W. Rei, “The benders decomposition algorithm: A literature review,” European Journal of Operational Research, vol. 259, no. 3, pp. 801–817, 2017. [16] N. Anselme, K. K. Nguyen, and M. Cheriet, “Digital twin backed closed-loops for energy-aware and open ran-based fixed wireless access serving rural areas,” IEEE Transactions on Mobile Computing, 2024. [17] O. Kodheli, E. Lagunas, N. Maturo, S. K. Sharma, B. Shankar, J. F. M. Montoya, J. C. M. Duncan, D. Spano, S. Chatzinotas, S. Kisseleff et al., “Satellite communications in the new space era: A survey and future challenges,” IEEE Communications Surveys & Tutorials, vol. 23, no. 1, pp. 70–109, 2020. [18] M. Majamaa, H. Martikainen, J. Puttonen, and T. Hämäläinen, “Satellite-assisted multi-connectivity in beyond 5g,” in 2023 IEEE 24th International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM). IEEE, 2023, pp. 413– 418. [19] M. Xu, D. Niyato, Z. Xiong, J. Kang, X. Cao, X. S. Shen, and C. Miao, “Quantum-secured space-air-ground integrated networks: Concept, framework, and case study,” IEEE Wireless Communications, vol. 30, no. 6, pp. 136–143, 2022. [20] Z. Abdullah, S. Kisseleff, E. Lagunas, V. N. Ha, F. Zeppenfeldt, and S. Chatzinotas, “Integrated access and backhaul via satellites,” in 2023 IEEE 34th Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC). IEEE, 2023, pp. 1–6.

[21] Z. Abdullah, E. Lagunas, S. Kisseleff, F. Zeppenfeldt, and S. Chatzinotas, “Integrated access and backhaul via leo satellites with inter-satellite links,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC). IEEE, 2024, pp. 1–6. [22] F. Kahmann, R. Tönjes, L. Schonebeck, T. Kranz, T. Zimmermann, D. Laniewski, H. Andreesen, and A. Möller, “Nomadic 5g network with satellite based internet connectivity for agriculture,” in 2025 IEEE 21st International Conference on Factory Communication Systems (WFCS). IEEE, 2025, pp. 1–8. [23] K. Mehmood, K. Kralevska, and D. Palma, “Intent-driven autonomous network and service management in future cellular networks: A structured literature review,” Computer Networks, vol. 220, p. 109477, 2023. [24] Y. Wang, C. Yang, T. Li, Y. Ouyang, X. Mi, and Y. Song, “A survey on intent-driven end-to-end 6g mobile communication system,” IEEE Communications Surveys & Tutorials, 2025. [25] S. K. S. Sidhu and A. Sharma, “Intent-based networking for 5g, b5g, and 6g: An overview of autonomous and intelligent network management,” in 2025 3rd International Conference on Advances in Computation, Communication and Information Technology (ICAICCIT), vol. 1. IEEE, 2025, pp. 365–372. [26] H. Yang, K. Zhan, B. Bao, Q. Yao, J. Zhang, and M. Cheriet, “Automatic guarantee scheme for intent-driven network slicing and reconfiguration,” Journal of Network and Computer Applications, vol. 190, p. 103163, 2021. [27] A. Ramírez-Arroyo, M. López, I. Rodríguez, S. B. Damsgaard, and P. Mogensen, “Multi-connectivity solutions for rural areas: Integrating terrestrial 5g and satellite networks to support innovative iot use cases,” Smart Agricultural Technology, vol. 12, p. 101260, 2025. [28] M. A. Ullah, R. D. Souza, G. Pasolini, J. M. de Souza Sant’Ana, M. Höyhtyä, K. Mikhaylov, H. Alves, E. Paolini, and A. AlHouran, “Extending the lora direct-to-satellite limits: Doppler shift pre-compensation,” IEEE Open Journal of the Communications Society, 2025. [29] R. F. Vieira, M. G. G. Pantoja, C. Natalino, and D. L. Cardoso, “Virtual network function placement and routing: Formulations and solutions,” IEEE Access, vol. 14, pp. 7715–7729, 2026. [30] CelesTrak, “Norad gp element sets,” https://celestrak.org/ NORAD/elements/index.php, [Online; accessed April. 22, 2026]. [31] B. Yang, Y. Wu, X. Chu, and G. Song, “Seamless handover in software-defined satellite networking,” IEEE Communications Letters, vol. 20, no. 9, pp. 1768–1771, 2016. [32] Y. Chang, X. Wang, J. Wang, Y. Wu, L. Yang, K. Zhu, H. Chen, X. Yi, C. Wang, Y. Wang et al., “A survey on evaluation of large language models,” ACM transactions on intelligent systems and technology, vol. 15, no. 3, pp. 1–45, 2024. [33] N. Tu, S. Nam, and J. W.-K. Hong, “Intent-based network configuration using large language models,” International Journal of Network Management, vol. 35, no. 1, p. e2313, 2025. [34] N. van der Laan and W. Romeijnders, “A converging benders’ decomposition algorithm for two-stage mixed-integer recourse models,” Operations Research, vol. 72, no. 5, pp. 2190–2214, 2024. [35] M. Lutz, Programming Python: powerful object-oriented programming. " O’Reilly Media, Inc.", 2010. [36] C. Yin and Z. Zhang, “A study of sentence similarity based on the all-minilm-l6-v2 model with “same semantics, different structure” after fine tuning,” in 2024 2nd International Conference on Image, Algorithms and Artificial Intelligence (ICIAAI 2024). Atlantis Press, 2024, pp. 677–684. [37] P. Safikhani and D. Broneske, “Automl meets hugging face: Domain-aware pretrained model selection for text classification,” in Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 4: Student Research Workshop), 2025, pp. 466–473. [38] S. Mitchell et al., “An introduction to pulp for python programmers,” Python Papers Monograph, vol. 1, no. 14, p. 2009, 2009. [39] F. Gurski and J. Rethmann, “Distributed solving of mixedinteger programs with glpk and thrift,” in Operations Research Proceedings 2016: Selected Papers of the Annual International Conference of the German Operations Research Society (GOR), Helmut Schmidt University Hamburg, Germany, August 30September 2, 2016. Springer, 2017, pp. 599–605.

[40] S. Harvey, “Tracking older artificial satellites,” Yearbook of Astronomy 2024, p. 271, 2023. [41] Skyfield, “Elegant Astronomy for Python, Skyfield,” https:// rhodesmill.org/skyfield/, [Online; accessed April 20, 2026]. [42] J. Li, S. Zou, Y. Sun, H. Gao, and W. Ni, “Business intent and network slicing correlation dataset from data-driven perspective,” Scientific Data, vol. 12, no. 1, p. 419, 2025. [43] Pratyush Puri, “5g network eda analysis 2025, howpublished ="https://www.kaggle.com/code/pratyushpuri/5g-networkeda-analysis-2025", note = "[online; accessed april. 20, 2026]".” [44] J. Meredith, M. Soveri, and M. Pope, “Management and orchestration; 5g end to end key performance indicators (kpi),” 3rd Generation Partnership Project (3GPP), Technical specification (TS), vol. 28, 2021. [45] W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, and M. Zhou, “Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers,” Advances in neural information processing systems, vol. 33, pp. 5776–5788, 2020. [46] R. Aperdannier, M. Koeppel, T. Unger, S. Schacht, and S. K. Barkur, “Systematic evaluation of different approaches on embedding search,” in Future of Information and Communication Conference. Springer, 2024, pp. 526–536. [47] M. Siino, “All-mpnet at semeval-2024 task 1: Application of mpnet for evaluating semantic textual relatedness,” in Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024), 2024, pp. 379–384. [48] S. Xiao, Z. Liu, P. Zhang, and N. Muennighoff, “C-pack: Packaged resources to advance general chinese embedding,” 2023. [49] L. Wang, N. Yang, X. Huang, B. Jiao, L. Yang, D. Jiang, R. Majumder, and F. Wei, “Text embeddings by weakly-supervised contrastive pre-training,” arXiv preprint arXiv:2212.03533, 2022.