Dynamic clustering-based adaptive frequency allocation for interference mitigation and device-to-device communication in wireless cellular networks

Abstract

Efficient frequency allocation in device-to-device (D2D) communication remains a critical challenge due to the need to mitigate interference while maintaining high quality of service (QoS). Traditional static allocation methods fail to adapt to dynamic network conditions, leading to inefficient spectrum utilization and degraded performance. This paper proposes a dynamic clustering-based adaptive frequency allocation (DCB-AFA) framework to address these limitations in wireless networks operating under continuously changing environments. The proposed approach leverages user location information and communication patterns to form adaptive clusters that minimize intra-network interference while enabling efficient D2D connectivity. A machine learning-based prediction mechanism is incorporated to anticipate user behavior and dynamically adjust cluster boundaries and resource allocation strategies. Furthermore, a QoS-aware feedback system continuously monitors network conditions and refines allocation decisions to improve performance metrics such as throughput, latency, and energy efficiency. Experimental evaluation demonstrates that the proposed DCB-AFA scheme significantly enhances spectrum utilization, reduces interference, and lowers power consumption compared to conventional approaches, making it a robust solution for next-generation wireless networks.

Keywords

device-to-device (D2D) communication frequency allocation clustering interference mitigation wireless networks

1. Introduction

Users can access wireless services through reliable cellular network infrastructures, which have become essential due to the rapid growth in data-intensive applications such as video streaming and real-time communication. This increasing demand places significant pressure on network resource management, particularly in the efficient allocation of frequency spectrum and transmission power. Traditionally, cellular networks rely on fixed or static resource allocation strategies, which often lead to suboptimal utilization of available resources and degraded performance under dynamic traffic conditions.

To address these limitations, this work proposes a dynamic clustering-based adaptive frequency allocation (DCB-AFA) scheme that dynamically adjusts resource allocation based on real-time network conditions, user behavior, and data consumption patterns. The proposed system aims to enhance overall network performance by optimizing key metrics such as latency, bandwidth utilization, and signal strength. Extensive evaluation demonstrates that DCB-AFA significantly improves resource efficiency and mitigates the limitations of conventional allocation approaches.

The need for adaptive mechanisms in dynamic environments is further supported by studies in related domains. Hameurlaine¹ analyzes the performance of the Optimized Link State Routing protocol in highly dynamic flying ad hoc networks (FANETs), showing that the packet delivery ratio is highly sensitive to mobility-induced topology changes. This highlights the importance of incorporating adaptability in network design. Similarly, Hameurlaine et al.² propose a predictive approach for selecting stable multi-point relays, demonstrating that link stability can significantly enhance routing reliability and reduce network fluctuations. These insights reinforce the motivation for adopting adaptive and stability-aware strategies such as the proposed DCB-AFA in cellular networks.

1.1. Key contributions

The major technical contributions of this work are summarized as follows:

A mobility-aware clustering framework integrating prediction and proximity. The DCB-AFA scheme introduces a new approach which combines link stability prediction through current spatial relationships and $M (u_{i}, u_{j})$ to create clusters. The system operates with permanent clusters, which it maintains through strategic planning instead of forming temporary groups based on present circumstances.

A lightweight neural prediction model for forecasting stable device-to-device (D2D) links. The system employs a two-layer neural predictor to predict user pair distance maintenance probability for a defined prediction duration. The system tracks user activities by processing mobility data, which includes speed and direction measurements and received signal strength indicator (RSSI) and signal-to-interference-plus-noise ratio (SINR) values to detect wireless channel changes. The system maintains reliable operation through D2D operations because it eliminates the need for reclustering procedures.

An interference-aware frequency allocation mechanism for multi-cluster environments. The proposed method solves both internal cluster interference and external cluster interference through the SINR formulation, which distinguishes it from current clustering algorithms. The proposed method reduces interference by using graph-coloring optimization for frequency assignment, which results in maximum spectral utilization for dense networks.

A realistic fifth generation (5G)/sixth generation (6G)-grade channel and D2D discovery model. The system depends on 3GPP TR 38.901 propagation models to determine pathloss, shadowing, and Rayleigh fading. The system operates through a D2D discovery protocol that uses 3GPP standardized sidelink beaconing and RSSI thresholding to achieve operational realism and maintain consistent results.

Comprehensive performance evaluation demonstrating significant gains. The system achieved better results than baseline methods in dense 5G/6G environments according to simulation results, which evaluated 1000 users under various New Ratio (NR) numerology conditions. The system achieves better performance than baseline clustering and frequency allocation methods because it provides superior latency results, better bandwidth management, and superior signal-to-noise ratio (SNR) outcomes. The ablation study demonstrates that all four modules, which include mobility prediction, clustering, frequency assignment, and D2D pairing must operate together for the system to function properly.

The paper organization follows this sequence: Section 2 examines existing research about resource allocation strategies. Section 3 presents both the system design and mathematical framework. The DCB-AFA algorithm appears in Section 4. Section 6 evaluates the simulation framework and evaluation approach developed in Section 5. The paper presents a comparison of obtained results in Section 6. The paper concludes in Section 7 and presents potential future research directions. Resource allocation in dynamic networks requires organizations to develop both strategic planning methods and flexible response systems. The DCB-AFA system improves wireless cellular network resource management through better service quality delivery, which results in enhanced customer satisfaction.

2. Related work

Extensive research has been conducted on managing radio resources and optimizing communication protocols in cellular networks. In this section, we delve into the existing literature on resource allocation for D2D communication and strategies, for optimizing networks. Optimizing how resources are distributed is essential for improving the efficiency of cell networks.³ They suggested an algorithm for managing radio resources in D2D-based vehicle-to-vehicle (V2V) communication using reinforcement learning with a focus on adjusting resources.⁴ Implemented an aware resource sharing strategy, in D2D communications, within diverse networks (HetNet).⁵ A new approach was introduced for assigning spectrum, in D2D-enabled tier HetNets using distributed multi-agent reinforcement learning. These studies emphasize the significance of dynamic and smart resource allocation tactics. D2D communication has been getting a lot of focus lately as a way to enhance efficiency and network capacity.⁶ A new approach suggests a plan for sharing resources in D2D communication, within networks. This method utilizes repeated game theory to enhance the efficiency of spectrum usage.⁷ Delving into the idea of bidirectional multi-level cognitive swarm drone 5G networks, with D2D communication serving as a factor, in boosting network functionalities.⁸ I analyzed how well different resource allocation algorithms work, in D2D communication systems with a focus on strategies that can adapt over time. Ensuring communication and maximizing network performance are goals. Bhardwaj and Gurjar⁹ explored how resources are distributed in D2D communications with a focus on enhancing the effectiveness of the network.¹⁰ A new method for controlling access based on clusters was presented for D2D communication networks, in devices focusing on challenges linked to the density of devices. Farrag et al.¹¹ analyzed how using drones for communications can improve network connections, with technologies.¹² Explored managing interference, among cells in band D2D communication, within LTE-A networks, focusing on strategies to reduce interference.

3. Problem formulation

In a cell network where there are $N$ users scattered randomly across the cell area, the goal is to assign $M$ frequencies ( $M < N$ ) to these users in a way that reduces interference and promotes communication. This issue can be presented in the manner;

3.1. Notation and definitions

$U = {u_{1}, u_{2}, \dots, u_{N}}$ : Set of $N$ users.

$F = {f_{1}, f_{2}, \dots, f_{M}}$ : Set of $M$ available frequencies ( $M < N$ ).

$C$ : Set of dynamically formed clusters within the network.

$D_{i}$ : Communication demand of user $u_{i} \in U$ .

$P_{i}$ : Transmission power of user $u_{i} \in U$ .

$I_{i j}$ : Interference between user $u_{i}$ and user $u_{j}$ when assigned the same frequency.

$Q_{i}$ : Quality of service (QoS) metric for user $u_{i} \in U$ .

3.2. Problem definition

The objective of the DCB-AFA framework is to determine an optimal assignment of frequencies $f_{i} \in F$ to users $u_{i} \in U$ , together with a stable cluster configuration $C = {C_{1}, C_{2}, \dots, C_{K}}$ , such that both intra-cluster and inter-cluster interference are minimized while satisfying the required QoS constraints. Let $I_{i j}$ denote the interference indicator between users $u_{i}$ and $u_{j}$ , and let

δ (f_{i}, f_{j}) = {\begin{cases} 1, & f_{i} = f_{j}, \\ 0, & f_{i} \neq f_{j}, \end{cases}

represent the frequency collision function.

The frequency assignment problem is formulated as the following optimization:

\begin{aligned} min_{{f_{i}}, C} & \sum_{i = 1}^{N} \sum_{\begin{matrix} j = 1 \\ j \neq i \end{matrix}}^{N} I_{i j} δ (f_{i}, f_{j}), \end{aligned}

(1)

subject to the constraints

\begin{aligned} \sum_{m = 1}^{| F |} δ (f_{i}, f_{m}) = 1, \forall i \in {1, \dots, N}, \end{aligned}

(2)

\begin{aligned} Q_{i} \geq Q_{min}, \forall i \in {1, \dots, N}, \end{aligned}

(3)

\begin{aligned} C_{i} = ClusterAssignment (u_{i}), \forall i \in {1, \dots, N} . \end{aligned}

(4)

The objective (1) penalizes any two interfering users that are assigned the same frequency. Constraint (2) ensures that each user receives exactly one frequency from the available pool. Constraint (3) enforces a minimum QoS requirement in terms of SINR or throughput. Constraint (4) binds each user to exactly one cluster based on the clustering mechanism described next (Figure 1).

Figure 1.

System model.

4. System model

Consider a single-cell network consisting of a set of users $U = {u_{1}, u_{2}, \dots, u_{N}}$ . Users are partitioned into dynamically evolving clusters $C = {C_{1}, C_{2}, \dots, C_{K}}$ , where each cluster represents a group of geographically close and communication-relevant users. Cluster boundaries are continuously adjusted based on predicted mobility patterns to preserve long-term stability and avoid frequent reclustering events.

Each user $u_{i}$ communicates either via the base station (cellular mode) or directly with another user (D2D mode). Let $P_{i}$ denote the transmit power of user $u_{i}$ and $g_{i j}$ the channel gain between users $u_{i}$ and $u_{j}$ . The channel gain incorporates pathloss, shadowing, and Rayleigh fading:

g_{i j} = | h_{i j} |^{2} \cdot 10^{- {PL}_{i j} / 10},

where

h_{i j} \sim C N (0, 1)

represents small-scale fading and

{PL}_{i j}

follows the 3GPP TR 38.901 model.

4.1. Interference and SINR model

For a user $u_{i}$ assigned frequency $f_{i}$ , the total interference consists of (i) intra-cluster interference from users within the same cluster using the same frequency, and (ii) inter-cluster interference from all neighboring clusters sharing frequency $f_{i}$ .

Thus, the received SINR at user $u_{i}$ is

{SINR}_{i} = \frac{P_{i} g_{i}}{\underset{Intra-cluster interference}{\underset{⏟}{\sum_{u_{j} \in C_{k}, j \neq i, f_{j} = f_{i}} P_{j} g_{i j}}} + \underset{Inter-cluster interference}{\underset{⏟}{\sum_{ℓ \neq k} \sum_{u_{m} \in C_{ℓ}, f_{m} = f_{i}} P_{m} g_{i m}}} + N_{0}},

where

C_{k}

is the cluster containing

u_{i}

, and

N_{0}

is the noise power. This SINR model ensures that frequency allocation and D2D pairing decisions fully account for cross-cluster interference, which is dominant in dense 5G/6G deployments.

QoS for each user requires:

{SINR}_{i} \geq γ_{\min},

where

γ_{\min}

is the minimum SINR threshold.

4.2. D2D feasibility model

Two users $u_{i}$ and $u_{j}$ can establish a D2D link if both their instantaneous channel quality and predicted mobility stability are sufficient. A D2D pair is formed only when:

{RSSI}_{i j} \geq γ_{disc} and M (u_{i}, u_{j}) \geq η,

where

γ_{disc}

is the discovery threshold and

M (u_{i}, u_{j})

is the predicted probability that the pair remains within clusterable distance during the next

T

seconds.

The objective of the system is to maximize valid D2D connections within each cluster while ensuring interference-aware frequency reuse. Formally, the number of D2D links in cluster $C_{k}$ is

max_{E_{C_{k}}} \sum_{(u_{i}, u_{j}) \in E_{C_{k}}} D_{i j},

subject to SINR and cluster-compatibility constraints, where

D_{i j}

is the D2D compatibility score.

Thus, the system model integrates mobility-aware clustering, interference-constrained frequency allocation, and channel-driven D2D pairing to achieve a stable and spectrum-efficient 5G/6G communication environment.

5. Proposed model

The proposed DCB-AFA framework enhances resource distribution in cellular networks by jointly optimizing clustering, interference-aware frequency allocation, and D2D connectivity. Unlike conventional approaches that rely on instantaneous user positions or static rules, DCB–AFA incorporates a predictive mobility-aware model that anticipates future user interactions, enabling proactive cluster formation and stable D2D communication. This section describes the formal structure of the model, the underlying prediction mechanism, and the integrated algorithmic workflow.

5.1. Dynamic clustering

Let $U = {u_{1}, u_{2}, \dots, u_{N}}$ denote the set of users in a cell. The objective of the clustering module is to partition $U$ into clusters that remain stable under mobility and maximize the probability of sustainable D2D links. Two components drive cluster formation: a proximity metric $P (u_{i}, u_{j})$ and a mobility-aware prediction model $M (u_{i}, u_{j})$ .

The proximity metric evaluates the spatial closeness between users. For any pair $(u_{i}, u_{j})$ , the Euclidean distance

d_{i j} (t) = \sqrt{(x_{i} (t) - x_{j} (t))^{2} + (y_{i} (t) - y_{j} (t))^{2}}

is computed, and two users are considered eligible for clustering if

P (u_{i}, u_{j}) = d_{i j} (t) \leq τ,

where

τ

denotes the cluster radius threshold.

While proximity captures instantaneous spatial correlation, it is insufficient in mobile environments. To ensure long-term cluster stability, DCB-AFA employs a predictive model that forecasts the likelihood of a stable link in the near future. The model output is defined as

M (u_{i}, u_{j}) = Pr (\begin{aligned} Users u_{i} and u_{j} remain cluster-compatible \\ during [t, t + T] \end{aligned}),

where

T

is the prediction horizon.

The prediction model is driven by a feature vector capturing mobility and channel dynamics:

x_{i j} (t) = \begin{aligned} [d_{i j} (t), d_{i j} (t - Δ), v_{i} (t), v_{j} (t), θ_{i} (t), θ_{j} (t), \\ {RSSI}_{i j} (t), {SINR}_{i j} (t)] . \end{aligned}

where

v_{i} (t)

and

θ_{i} (t)

denote the speed and direction of user

u_{i}

, respectively. The model is implemented as a lightweight neural predictor,

M (u_{i}, u_{j}) = σ (W_{2} ReLU (W_{1} x_{i j} (t) + b_{1}) + b_{2}),

trained offline using mobility traces and channel-quality histories. The ground-truth labels used during training are

y_{i j} = {\begin{cases} 1, & if d_{i j} (t^{'}) \leq τ, \forall t^{'} \in [t, t + T], \\ 0, & otherwise, \end{cases}

and the loss function is the binary cross-entropy

L = - y_{i j} \log M (u_{i}, u_{j}) - (1 - y_{i j}) \log (1 - M (u_{i}, u_{j})) .

During operation, users

u_{i}

and

u_{j}

are assigned to the same cluster if both the proximity requirement and the predictive stability condition hold:

P (u_{i}, u_{j}) \leq τ and M (u_{i}, u_{j}) \geq η,

where

η

is the stability threshold. The final cluster is obtained by maximizing the expected stability,

C_{k}^{⋆} = \arg max_{C_{k}} \sum_{u_{i}, u_{j} \in C_{k}} M (u_{i}, u_{j}),

ensuring proactive and mobility-resilient grouping.

5.2. Mobility prediction model

The DCB-AFA framework maintains stable clusters through two factors, which include users’ current location proximity and their predicted future movement patterns. The proposed model includes a mobility-aware prediction function $M (u_{i}, u_{j})$ , which forecasts link stability between users throughout future time periods. The predictor enables proactive clustering through its ability to forecast user relationship changes that proximity measurements cannot detect.

For any two users $u_{i}$ and $u_{j}$ , let $d_{i j} (t)$ denote their instantaneous distance at time $t$ . The model evaluates the probability that these users will remain within a clusterable distance threshold $τ$ for the subsequent $T$ seconds:

M (u_{i}, u_{j}) = Pr (d_{i j} (t^{'}) \leq τ, \forall t^{'} \in [t, t + T]) .

The predicted strength of the D2D wireless connection between two users becomes stronger when their forecasted distance and channel conditions show positive and strong values.

The prediction model uses a feature vector that combines mobility indicators with channel-quality measurements to make its predictions.

\begin{aligned} x_{i j} (t) = [d_{i j} (t), d_{i j} (t - Δ), v_{i} (t), v_{j} (t), θ_{i} (t), θ_{j} (t), \end{aligned}

\begin{aligned} {RSSI}_{i j} (t), {SINR}_{i j} (t)] . \end{aligned}

The model uses $v_{i} (t)$ and $θ_{i} (t)$ to represent user $u_{i}$ speed and direction at time $t$ and RSSI and SINR to represent channel evolution during the previous time period. The model uses these features to understand both user movement patterns and wireless link performance under mobile conditions.

The neural predictor contains a simple design that transforms input feature vectors into stability probability outputs.

M (u_{i}, u_{j}) = σ (W_{2} ReLU (W_{1} x_{i j} (t) + b_{1}) + b_{2}),

The model contains four learnable parameters

W_{1}

W_{2}

b_{1}

, and

b_{2}

which learn during training and

σ (\cdot)

serves as the sigmoid activation function to produce outputs between 0 and 1. The two-layer architecture provides both high prediction accuracy and quick processing speed because it runs in real-time at network edges.

The model receives offline training data from both artificial mobility data and actual mobility records. The model receives its ground-truth label from

y_{i j} = {\begin{cases} 1, & if d_{i j} (t^{'}) \leq τ, \forall t^{'} \in [t, t + T], \\ 0, & otherwise . \end{cases}

Training minimizes the binary cross-entropy loss

L = - y_{i j} \log M (u_{i}, u_{j}) - (1 - y_{i j}) \log (1 - M (u_{i}, u_{j})),

The network learns to identify patterns between movement paths and wireless link development through this method.

The system uses the mobility predictor to make clustering decisions by applying the following joint condition:

P (u_{i}, u_{j}) \leq τ and M (u_{i}, u_{j}) \geq η,

The system uses

η

as a stability confidence threshold to determine cluster formation. The system uses predicted long-term compatibility to form clusters instead of depending on current geometry, which results in lower reclustering costs and better D2D link stability. The mobility prediction component functions as a core element that enables the DCB-AFA framework to develop stable clusters that fight against interference while actively forming clusters.

5.3. Interference modeling across and within clusters

The DCB-AFA framework models both user behavior and cluster behavior to achieve realistic behavior in multi-user and multi-cluster cellular environments: intra-cluster and inter-cluster interference. The frequency allocation stage reduces user interference inside one cluster, but the system assesses neighboring cluster interference to determine SINR values. The received SINR value for user $u_{i}$ in cluster $C_{k}$ is defined as

{SINR}_{i} = \frac{P_{i} g_{i}}{\sum_{u_{j} \in C_{k}, j \neq i} P_{j} g_{i j} + \sum_{ℓ \neq k} \sum_{u_{m} \in C_{ℓ}} P_{m} g_{i m} + N_{0}},

where the first summation captures intra-cluster interference and the second captures inter-cluster interference from all neighboring clusters

C_{ℓ}

ℓ \neq k

. The channel gain terms

g_{i}

g_{i j}

, and

g_{i m}

in this model derive from 3GPP pathloss, shadowing, and Rayleigh fading models. The optimization method at the cluster level prevents it from ignoring interference that occurs between different clusters because dense networks experience this type of interference as their main problem. The frequency assignment process selects specific frequency sets

F_{C_{k}}

for each cluster, which meet the required distance rules to prevent strong interference between clusters.

5.4. Adaptive frequency allocation

The system creates clusters before DCB-AFA determines frequency assignments through intra-cluster interference minimization. Let $F$ denote the global available frequency set and $F_{C_{k}} \subseteq F$ the subset allocated to cluster $C_{k}$ . For each cluster, an interference adjacency matrix

A_{i j} = {\begin{cases} 1, & if users u_{i}, u_{j} \in C_{k} interfere, \\ 0, & otherwise, \end{cases}

is constructed. Frequency assignment solves

min_{{f_{i}}} \sum_{u_{i}, u_{j} \in C_{k}} A_{i j} δ (f_{i} = f_{j}),

The problem requires a solution that matches a graph-coloring problem that demands nodes with interference to operate on different frequency channels. The optimization process continues to run until the assignment reaches a point where no further changes occur.

5.5. D2D communication pairing

The base station load decreases when users in each cluster establish D2D links with compatible users. The D2D compatibility metric $D_{i j}$ represents the combination of SINR values, RSSI measurements, and predicted stability performance. The system determines pairings by running an optimization algorithm that selects the D2D connections that will produce the highest number of operational links.

max_{E_{C_{k}}} \sum_{(u_{i}, u_{j}) \in E_{C_{k}}} D_{i j} .

The system creates high-quality direct communication channels through its operation under hardware restrictions and interference environments.

5.6. Algorithm: DCB-AFA

5.7. Discussion of assumptions

The enhanced DCB-AFA model removes simplifications through its implementation of complete inter-cluster interference and its use of actual propagation models and link prediction based on mobility patterns and a discovery system that follows D2D standards. The model becomes more practical for use because of its additional features, which demonstrate its ability to operate effectively in 5G/6G networks with their complex dense structures that experience channel interference and wireless environment changes.

The ablation results presented in Table 1 provide a quantitative comparison of different configurations of the proposed DCB-AFA framework. Configuration A1, which includes all modules, achieves the best performance, indicating that the integration of dynamic clustering, adaptive frequency allocation, and D2D communication yields complementary benefits.

Table 1.
Ablation study of the proposed DCB-AFA framework.

Removing D2D communication (A2) results in increased latency and reduced bandwidth utilization, as traffic must be routed through the base station instead of leveraging direct links. Disabling adaptive frequency allocation (A3) causes a noticeable degradation due to increased interference and inefficient spectrum usage under static allocation. Similarly, removing dynamic clustering (A4) leads to poor spatial grouping, which increases transmission distance and interference, thereby degrading signal quality and latency.

The baseline configuration (A5), where all adaptive mechanisms are disabled, shows the worst performance across all metrics. This confirms that static and non-adaptive resource allocation strategies are insufficient for dynamic wireless environments. Overall, the results demonstrate that each module contributes significantly, and their combined operation is necessary to achieve optimal performance in next-generation networks.

6. Simulation results and comparisons

The following section shows simulation results for the DCB-AFA algorithm and its performance comparison with current methods. The simulation took place in a wireless cellular network environment that mimicked real-world conditions to measure latency, bandwidth utilization, and signal quality.

6.1. Realistic wireless channel and propagation models

The simulation implements 3GPP TR 38.901 channel models to simulate urban macro, urban micro, and indoor hotspot environments for real-world deployment testing. The pathloss between transmitter $i$ and receiver $j$ is computed as

{PL}_{i j} (dB) = 32.4 + 21 \log_{10} (d_{i j}) + 20 \log_{10} (f_{c}),

supplemented by log-normal shadow fading with standard deviation

σ_{SF}

and Rayleigh fast fading

h_{i j} \sim C N (0, 1)

. The resulting channel gain is expressed as

g_{i j} = | h_{i j} |^{2} \cdot 10^{- {PL}_{i j} / 10} .

The stochastic channel components enable user mobility and frequency assignment decisions to handle urban obstructions, multipath effects, and high-frequency propagation losses, which are typical for 5G/6G systems.

6.2. Scalability to dense 5G/6G scenarios

The simulation system now enables dense user deployment scenarios which support more than 500–1000 users per cell above the standard 100-user configuration. The adaptive clustering and frequency allocation stages function at a specific operational scale that

\begin{aligned} O (N \log N) for clustering, \\ O (E) for interference graph construction . \end{aligned}

(5)

The model represents interference edges through the variable $E$ , which directly depends on neighborhood density. The model uses this scaling method to achieve fast computation in dense network environments. The frequency resource pool receives an increase to support 5G NR numerologies, which include 15–120 kHz subcarrier spacing bands for up to hundreds of orthogonal resource blocks.

6.3. Realistic D2D discovery and control signaling

The discovery protocol of D2D pairing uses a functional approach to enhance its operational efficiency. Users send discovery beacons through the sidelink channel at power level $P_{disc}$ during their assigned discovery time slots. The system enables D2D communication when both devices meet two particular requirements.

{RSSI}_{i j} \geq γ_{disc} and M (u_{i}, u_{j}) \geq η,

The system initiates D2D links through equation (1) when the physical channel conditions and mobility patterns indicate stable communication. The system initiates D2D links through equation (1) when the physical channel conditions and mobility patterns indicate stable communication. The sidelink establishment procedure follows 3GPP ProSe control signaling, which includes resource request and grant cycles to prevent conflicts.

6.4. Latency performance

Figure 2 shows that the proposed DCB-AFA model consistently attains lower latency across a wide range of operating scenarios compared with the baseline methods. This gain emerges from the simultaneous application of dynamic clustering, adaptive frequency allocation, and D2D communication. In high user-density or peak-hour traffic conditions, DCB-AFA reorganizes cluster memberships based on proximity and predicted mobility patterns, then assigns frequencies to minimize interference and congestion. Unlike conventional fixed clustering or simple roundrobin frequency assignment, which ignore real-time interference and mobility dynamics (see Alsharfa et al.¹³ and Sabella et al.¹⁴), DCB-AFA uses an interference adjacency matrix and graph-coloring optimization to reduce queuing and retransmission delays. Moreover, the option of direct D2D links within clusters bypasses base-station scheduling overhead, reducing end-to-end latency. Overall, in contrast to the static and blind allocation methods, DCB-AFA adapts to network variation and maintains a low-latency profile even under stress.

Figure 2.

Latency comparison.

The latency behavior observed in Figure 2 can be understood through a decomposition of end-to-end delay for a communication between user $u_{i}$ and user $u_{j}$ :

L_{i j} = L_{i j}^{(p)} + L_{i j}^{(q)} + L_{i j}^{(i)},

where

L_{i j}^{(p)} = d_{i j} / v

represents propagation delay over distance

d_{i j}

L_{i j}^{(q)}

is the queuing delay due to congestion and resource contention, and

L_{i j}^{(i)}

is the retransmission delay incurred from interference-induced errors. In DCB-AFA, dynamic clustering enforces a proximity condition

‖ u_{i} - u_{j} ‖_{2} \leq τ

, thereby reducing

d_{i j}

and thus

L_{i j}^{(p)}

. Unlike static clustering, which may allow

d_{i j}

to grow as users move, or roundrobin schemes which neglect proximity entirely, DCB-AFA keeps propagation delay minimal.

The queuing delay term $L_{i j}^{(q)}$ is closely related to the network interference level, and can be approximated as $L^{(q)} \propto 1 / (1 + SINR)$ . DCB-AFA constructs an interference graph with

A_{i j} = {\begin{cases} 1, & if users u_{i} and u_{j} interfere, \\ 0, & otherwise, \end{cases}

and solves the frequency assignment problem

min_{{f_{i}}} \sum_{i} \sum_{j \neq i} A_{i j} δ (f_{i} = f_{j}),

where

δ (\cdot)

is the indicator of frequency collision. Consequently, it improves SINR and reduces both queuing and retransmission delays. Prior works, such as Zafar et al.,¹⁵ adopt clustering for D2D but lack this real-time adaptation to mobility and full interference graph optimization.

Furthermore, D2D communication within clusters replaces the traditional two-hop cellular link (user $\to$ Base Station $\to$ user) with a single direct link, effectively eliminating the base-station component of queuing and scheduling delay. Therefore, the total latency in DCB-AFA becomes:

L_{DCB-AFA} < L_{Static} < L_{RoundRobin},

which aligns with the empirical trends shown in the latency comparison plot.

6.5. Bandwidth utilization

Figure 3 demonstrates that the proposed DCB-AFA consistently achieves higher bandwidth utilization than the baseline methods across all examined contexts. This enhancement arises because dynamic clustering improves spatial reuse by grouping proximate and mobility-coherent users, and adaptive frequency allocation reduces intra-cluster conflict. D2D communication further elevates resource utilization by converting conventional base-station relayed traffic into direct device exchanges, reducing system signaling overhead. Existing literature confirms that D2D clustering improves spectral efficiency (e.g. Mesbahi and Rahbar¹⁶), yet fails to integrate both clustering and real-time frequency optimization under mobility. In contrast, DCB-AFA maintains higher effective throughput and utilization even under variable load and heterogeneous mobility.

Figure 3.

Bandwidth utilization comparison.

The utilization of bandwidth for user $u_{i}$ may be expressed as

B_{i} = \frac{D_{i}}{t},

where

D_{i}

denotes the successfully transmitted data over time interval

t

. Because

D_{i}

depends on the quality of the channel and interference environment, large SINR gains produce higher

D_{i}

. By constructing an optimized interference graph and applying graph-coloring frequency allocation as above, DCB-AFA reduces the number of collision edges, thus increasing the transmission rate. Additionally, D2D links bypass backhaul and serve locally, increasing the data exchange within each cluster. Approaches proposed by Alsharfa et al.¹³ adopt clusters for fairness but do not optimize real-time frequency collision, which limits their utilization gains. In the proposed model:

B_{DCB-AFA} > B_{Static} > B_{RoundRobin},

corresponding to the bandwidth behavior shown in Figure 3.

6.6. Signal quality assessment

Figure 4 shows that the DCB-AFA model maintains consistently higher signal quality compared to the baseline methods under differing conditions. This outcome results from combining proximity-based clustering, interference- aware frequency allocation, and D2D direct links. By reducing the number of strong interfering neighbors for each user, DCB-AFA improves the SINR defined by

{SINR}_{i} = \frac{S_{i}}{\sum_{j \neq i} A_{i j} I_{i j} + N_{0}},

where

S_{i}

is the received signal strength,

I_{i j}

the interference from user

j

, and

N_{0}

the noise floor. The dynamic clustering lowers

\sum_{j \neq i} A_{i j} I_{i j}

by reducing interference edges, the frequency allocation eliminates collisions (

f_{i} = f_{j}

avoidance), and D2D links increase

S_{i}

due to reduced pathloss. Earlier methods, such as those in Sabella et al.,¹⁴ show modest SINR improvements through clustering alone, but without adaptive allocation, their gains remain limited. In our model:

{SINR}_{DCB-AFA} > {SINR}_{Static} > {SINR}_{RoundRobin},

matching the observed empirical trend in Figure 4.

Figure 4.

Signal quality comparison.

The technical justification for the superior signal quality in DCB-AFA stems from the interference graph reduction and direct link establishment within clusters. By actively managing cluster membership and frequency reuse, the effective interference seen by each user is substantially reduced, and combined with D2D link optimization, the received signal strength is strengthened. Hence, the SINR performance advantage directly translates to the signal quality improvement observed. This end-to-end treatment of proximity, mobility, resource allocation, and direct communication distinguishes DCB-AFA from prior work and yields consistently higher signal quality in dense, dynamic networks.

7. Conclusion

The DCB-AFA algorithm solves resource distribution and network communication optimization problems in wireless cellular systems through its effective solution. The network performance improves through its three main features, which include dynamic clustering, adaptive frequency allocation, and D2D communication support. The DCB-AFA system allows network operators to modify their systems according to user growth while providing top-notch service quality. The research shows that wireless networks need adaptive resource management systems, which use intelligent methods to achieve optimal performance. The user-oriented optimization approach of DCB-AFA enables network operators to obtain valuable resources while developing superior wireless communication systems for future networks.

Footnotes

ORCID iD

Noor Ahmad

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Hameurlaine

Mehallegue

Benssalah

. Packet delivery ratio assessment and analysis of OLSR protocol for FANETs. In: Proceedings of the international conference on unmanned aircraft systems (ICUAS), 2017.

Hameurlaine

Mehallegue

Benssalah

. A novel approach for selecting stable MPR in the OLSR protocol. Int J Commun Syst 2018.

Souhir

FEKI

Belghith

Zarai

. A reinforcement learning-based radio resource management algorithm for D2D-based V2V communication In: 2019 15th international wireless communications & mobile computing conference (IWCMC), 2019, pp.1367–1372. IEEE.

Ahmed

Yinxiao

, et al. Secrecy ensured socially aware resource allocation in device-to-device communications underlaying HetNet. IEEE Trans Vehic Technol 2019.

Zia

Javed

Sial

, et al. A distributed multi-agent RL-based autonomous spectrum allocation scheme in D2D enabled multi-tier HetNets. IEEE Access 2019; 7: 6733–6745.

Barik

Shukla

Datta

, et al. A resource sharing scheme for intercell D2D communication in cellular networks: A repeated game theoretic approach. IEEE Trans Vehic Technol 2020.

Jacob

Menon

Padinjappurathu

, et al. Bidirectional multi-tier cognitive swarm drone 5G network. In: Proceedings IEEE INFOCOM, 2020.

Shanbhag

Bajpai

Gupta

, et al. Performance analysis of multi-armed bandit based resource allocation algorithms for D2D communication systems. In: Proceedings of the international conference on telecommunications, 2020.

Bhardwaj

Gurjar

. Resource allocation in D2D communications. In: 5G and beyond wireless systems: PHY layer perspective, 2020, pp.317–340. Singapore: Springer Singapore.

10.

Nguyen

Dang

THG

, et al. Cluster-based access control mechanism for cellular D2D communication networks with dense device deployment. Int J Adv Comput Sci Appl 2021; 12.

11.

Farrag

Maher

El-Mahdy

, et al. Performance analysis of UAV assisted mobile communications in THz channel. IEEE Access 2021.

12.

Modak

Rahman

. Multi-cell interference management in in-band D2D communication under LTE-A network. In: 2021 international conference on computing, electronics & communications engineering (iCCECE), pp.13–18. IEEE.

13.

Alsharfa

Mohammed

Gharghan

, et al. Cellular–D2D resource allocation algorithm based on user fairness. Electronics 2020; 9: 386.

14.

Sabella

Susanto

Setyawan

FXA

. Interference management using distance-based clustering method for D2D communication underlaying multicell cellular network. J Elektron Telekomun 2023; 23: 76–84.

15.

Zafar

Khan

Alassafi

. An efficient resource optimization scheme for D2D communication with fuzzy clustering and game-theory based power control. Sustain Comput: Informat Syst 2022.

16.

Mesbahi

Rahbar

. Cluster-based architecture capable for device-to-device millimeter-wave communications in 5G cellular networks. Arab J Sci Eng 2019; 44: 9719–9733.

Dynamic clustering-based adaptive frequency allocation for interference mitigation and device-to-device communication in wireless cellular networks

Abstract

Keywords

1. Introduction

1.1. Key contributions

2. Related work

3. Problem formulation

3.1. Notation and definitions

3.2. Problem definition

4.1. Interference and SINR model

4.2. D2D feasibility model

5. Proposed model

5.1. Dynamic clustering

5.2. Mobility prediction model

5.3. Interference modeling across and within clusters

5.4. Adaptive frequency allocation

5.5. D2D communication pairing

5.6. Algorithm: DCB-AFA

5.7. Discussion of assumptions

Table 1. Ablation study of the proposed DCB-AFA framework.

6.1. Realistic wireless channel and propagation models

6.2. Scalability to dense 5G/6G scenarios

6.4. Latency performance

Footnotes

ORCID iD

Funding

Declaration of conflicting interests

References

Table 1.
Ablation study of the proposed DCB-AFA framework.