In the rapidly evolving landscape of intelligent transportation, the integration of drone technology with vehicular networks has emerged as a transformative paradigm. I focus on addressing the critical challenges of computation-intensive and latency-sensitive services in highly dynamic internet of vehicles (IoV) scenarios. The core of this work lies in proposing a novel collaborative content caching scheme that leverages non-orthogonal multiple access (NOMA) and a swarm of drones as edge nodes. By deploying drone clusters, I aim to enhance spectrum efficiency, reduce content retrieval latency, and improve overall network robustness.
The exponential growth of connected vehicles and their applications—such as autonomous driving, real-time navigation, and multimedia streaming—generates massive data traffic. Traditional orthogonal multiple access (OMA) schemes suffer from limited spectrum utilization, especially under high vehicle density. To overcome this, I adopt NOMA, which enables multiple vehicles to share the same frequency resource through power-domain multiplexing and successive interference cancellation (SIC). This drone technology-enabled framework dynamically partitions vehicles into clusters using a K-Means++ algorithm, optimizing the deployment positions of drone cluster heads. The clustering process reduces inter-cluster interference and minimizes communication distances, directly contributing to lower latency and higher throughput.
A key contribution of my work is the formulation of the collaborative caching problem as a decentralized partially observable Markov decision process (Dec-POMDP). I design a graph convolutional network (GCN) to capture the topological relationships among drones, their caching status, and content popularity features. The dynamic adjacency matrix in the GCN allows for effective cross-node information sharing, enabling each drone to make informed caching decisions based on its neighboring drones’ states. To solve the Dec-POMDP efficiently, I introduce an attention-enhanced QMIX deep reinforcement learning algorithm. The attention mechanism weightedly fuses the states of neighboring drones, ensuring that the long-term cache hit rate is maximized while minimizing redundant communication overhead.
Before diving into the technical details, let me establish the system model. I consider a urban traffic scenario where a high-altitude platform (HAP) centrally manages multiple drones. Both drones and roadside base stations possess caching and computing capabilities. When base stations are overloaded or unavailable, drones serve as supplementary nodes for content delivery and task offloading. Vehicles are dynamically clustered based on spatial distribution and associated with corresponding drones. The drones maintain inter-cluster distance to suppress interference, and NOMA with SIC mitigates intra-cluster interference. The primary objective is to minimize the total content retrieval delay while satisfying various constraints on power, cache capacity, and interference.
System Model and Problem Formulation
I define the set of drones as \(U = \{1, 2, \dots, u_{\text{max}}\}\) and the set of vehicles as \(V = \{1, 2, \dots, v_{\text{max}}\}\). The vehicles are partitioned into \(C\) clusters. The system operates on a dual-timescale: a small timescale for cache updates and a large timescale for drone repositioning. Each vehicle has a position \(\phi_v(t_c) = (x_v(t_c), y_v(t_c), 0)\). Drones fly at a fixed altitude \(h\) and follow a Gauss-Markov mobility model:
$$ v_u(t_{c+1}) = \alpha \cdot v_u(t_c) + (1-\alpha) \cdot \mu_{v_u} + \epsilon(t_c) $$
$$ \theta_u(t_{c+1}) = \alpha \cdot \theta_u(t_c) + (1-\alpha) \cdot \mu_{\theta_u} + \epsilon(t_c) $$
where \(\alpha\) is the inertial coefficient, \(\mu_{v_u}\) and \(\mu_{\theta_u}\) are target speed and heading, and \(\epsilon\) is Gaussian noise. The drone’s position at time \(t_{c+1}\) is updated accordingly.
Each drone has a caching capacity \(\psi_u\) and decides which content items to cache. Content popularity follows a Zipf distribution with parameter \(\alpha=1\). The retrieval delay depends on three delivery modes: direct edge retrieval (from the serving drone), cooperative retrieval (from a neighboring drone via U2U link), and HAP retrieval (via H2U link). I model the communication links as follows.
Communication Models
H2U Link: The path loss follows a Weibull channel model with gain \(H_H^u = \frac{G_{TH} G_{RU} \lambda^2}{(4\pi H_{\text{HAP}})^2} 10^{-F_{\text{at}}/10}\). The transmission rate is:
$$ R_{\text{HAP}\to u}^t = \frac{B_H}{\rho_{t_c}(H_{\text{HAP}})} \log_2\left(1 + \frac{P_H H_H^u}{B_H \sigma_H^2}\right) $$
U2U Link: Free-space propagation model with path loss \(H_{u’}^u = \frac{G_{TU} G_{RU} \lambda^2}{(4\pi d_{u,u’}(t_c))^2}\). Rate:
$$ R_{u’\to u}^{t_c} = \frac{B_u}{\rho_{t_c}(u’)} \log_2\left(1 + \frac{P_S H_{u’}^u}{B_u \epsilon_u^2}\right) $$
U2V Link: Based on 3GPP Release-15, the path loss depends on LoS/NLoS conditions. The average path loss is \(L_v^u(t_c) = P_{\text{LoS}} L_{\text{LoS}} + P_{\text{NLoS}} L_{\text{NLoS}}\). The channel gain is \(g_v^u(t_c) = H_v^u(t_c) \cdot 10^{-L_v^u(t_c)/10}\). I use NOMA for U2V transmissions. The superimposed signal at drone \(u\) is:
$$ d_u(t_c) = \sum_{v=1}^{m} \alpha_{u,v}(t_c) \sqrt{P_v^u(t_c)} d_v^u(t_c) $$
where \(\alpha_{u,v}(t_c)\) is the service indicator. The SINR for the \(v\)-th decoded user is:
$$ \gamma_{\epsilon(v)}^u(t_c) = \frac{\alpha_{u,\epsilon(v)}(t_c) P_{\epsilon(v)}^u(t_c) |g_{\epsilon(v)}^u(t_c)|^2}{I_{\text{intra}}^{\epsilon(v)}(t_c) + I_{\text{inter}}^{u,\epsilon(v)}(t_c) + \sigma(t_c)^2} $$
The data rate for vehicle \(v\) associated with drone \(u\) is:
$$ R_{u,v}(t_c) = B \log_2\left(1 + (\gamma_{\epsilon(v)}^u(t_c))^2\right) $$
The total data rate of drone \(u\) is \(R_u(t_c) = \sum_{v=1}^{V_u} R_{u,v}(t_c)\). The retrieval delays for the three modes are:
$$ \tau_{u,v,w}^1(t_c) = \frac{\lambda_w}{R_{u\to v}^{t_c}}, \quad \tau_{u,v,w}^2(t_c) = \frac{\lambda_w}{R_{u’\to u}^{t_c}} + \frac{\lambda_w}{R_{u\to v}^{t_c}}, \quad \tau_{u,v,w}^3(t_c) = \frac{\lambda_w}{R_{\text{HAP}\to u}^t} + \frac{\lambda_w}{R_{u\to v}^{t_c}} $$
Optimization Problem
The goal is to minimize the total content retrieval delay over all cache cycles, subject to:
- Each vehicle is served by exactly one drone: \(\sum_{v=1}^{v_{\text{max}}} \alpha_{u,v} = 1\)
- Power constraint per drone: \(\sum_{v \in C_u^{t_c}} \alpha_{u,v}(t_c) P_v^u \leq P_u\)
- Decoding order constraint for NOMA: \(G_{\epsilon(v)}^u(t_c) \leq G_{\epsilon(o)}^u(t_c), o > v\)
- Cache capacity constraint: \(\sum_{w \in W} r_w^u(t_c) \lambda_w \leq \psi_u\)
- Content retrieval mode selection: \(\sum_{i=1}^3 \chi_i = 1, \chi_i \in \{0,1\}\)
- Delay ordering: \(0 < \tau^1 < \tau^2 < \tau^3 < T_c\)
- Rate constraint: \(R_u(t_c) < B\)
Proposed Solution
I design a three-stage solution that integrates clustering, graph learning, and multi-agent reinforcement learning. This framework fully exploits drone technology to achieve efficient collaborative caching.
K-Means++ Dynamic Clustering and Drone Deployment
I use K-Means++ to cluster vehicles based on their positions. The algorithm initializes centroids probabilistically, selecting the first centroid randomly and subsequent centroids with probability proportional to the squared distance from the nearest existing centroid. Then iterative assignment and update steps minimize the within-cluster sum of squares. The resulting centroids correspond to optimal drone deployment positions. The silhouette coefficient for various vehicle counts (20 to 100) remains above 0.54, demonstrating the algorithm’s adaptability to high-density IoV scenarios.
| Number of Vehicles | Silhouette Coefficient |
|---|---|
| 20 | 0.68 |
| 40 | 0.64 |
| 60 | 0.63 |
| 80 | 0.55 |
| 100 | 0.54 |
GCQM: Graph Convolutional QMix for Collaborative Caching
I model the caching decision as a Dec-POMDP. Each drone agent observes its local cache state, content popularity matrix over historical windows, remaining cache capacity, and neighboring drones’ cache states. The observation for drone \(u\) at time \(t_c\):
$$ o_u^{t_c} = [E_W^u(t_c), P_W^u(t_c), \psi_u^{\text{cap}}(t_c), N_u(t_c)] $$
The action is the binary caching decision for each content: \(a_u^{t_c} = [r_1^u(t_c), \dots, r_{w_{\text{max}}}^u(t_c)]\). The reward combines local hit rate and neighbor collaborative hit rate:
$$ r_u^{t_c} = \omega_1 \cdot r_{\text{self},u}^{t_c} + \omega_2 \cdot r_{\text{neig},u}^{t_c} $$
where \(\omega_1 = 0.67, \omega_2 = 0.33\). The global reward is the sum of all drone rewards.
To capture the dynamic topology of the drone swarm, I construct a graph \(\mathcal{G} = \{\mathcal{N}, \mathcal{E}\}\) where nodes are drones. The adjacency matrix \(A_u^{t_c}\) is normalized and updated per cache cycle. I combine physical proximity (Euclidean distance \(d_{u,j}(t_c)\)) and content complementarity (Jaccard dissimilarity \(D(u,j,t_c) = 1 – \frac{|E_W^u(t_c) \cap E_W^j(t_c)|}{|E_W^u(t_c) \cup E_W^j(t_c)|}\)) to compute a composite collaboration score:
$$ S_{u,j}(t_c) = \omega_d \cdot \frac{1}{d_{u,j}(t_c)} + \omega_c \cdot D(u,j,t_c) $$
where \(\omega_d = \omega_c = 0.5\). The GCN aggregates features from neighbors through two convolutional layers:
$$ H^{(l+1)} = \sigma( \tilde{A}_u^{t_c} H^{(l)} W^{(l)} ), \quad l \in \{0,1\} $$
Then multi-head attention is applied to produce final node features \(h_u^{t_c”’}\). The attention weights:
$$ \alpha_{i,j,t_c}^h = \frac{\exp\left( \tau \cdot (W_Q^h h_i^{t_c”})^T (W_K^h h_j^{t_c”}) \right)}{\sum_{k \in A_u^{t_c}} \exp\left( \tau \cdot (W_Q^h h_i^{t_c”})^T (W_K^h h_k^{t_c”}) \right)} $$
The final feature is:
$$ h_i^{t_c”’} = \sigma\left( \text{con}\left[ \sum_{j \in A_u^{t_c}} \alpha_{i,j,t_c}^h W_V^h h_j^{t_c”} \right] \right) $$
These features are fed into the local Q-network of the QMix algorithm. The global Q-value is computed by a hypernetwork that enforces monotonicity:
$$ Q_{\text{tot}}(s_{t_c}, a_{t_c} | \theta_{\text{mix}}) = f_{\text{mix}}(Q^1, Q^2, \dots, Q^{u_{\text{max}}}) $$
Training minimizes the TD error:
$$ L = \frac{1}{N_b} \sum_{j=1}^{N_b} (Q_{\text{tot}}(s_j, a_j | \theta_{\text{mix}}) – y_j)^2 $$
with target \(y_j = y_j + \gamma \max_{a’} Q’_{\text{tot}}(s_{j+1}, a’ | \theta_{\text{mix}}^-)\). The algorithm alternates between experience collection and parameter updates, optimizing GCN weights, local Q-network, and global Q-network.
Simulation Results and Analysis
I implement the simulation environment using Python 3.8 and TensorFlow 2.6, following 3GPP TR 36.885 Manhattan scenario. The key parameters are listed below.
| Parameter | Value |
|---|---|
| HAP altitude | 20 km |
| HAP transmit power | 40 dBm |
| HAP bandwidth | 100 MHz |
| H2U link frequency | 3.5 GHz |
| U2V carrier frequency | 2 GHz |
| UAV bandwidth | 20 MHz |
| UAV transmit power | 30 dBm |
| Cache capacity per drone | 200 MB (variable) |
| Number of drones | 5 |
| Vehicle speed (max) | 60 km/h |
| Number of content items | 30 |
| Zipf parameter \(\alpha\) | 1 |
| Cache cycle duration | 10 s |
I compare my proposed GCQM algorithm against several baselines: Random, LFU, FIFO, DDQN, GT-SSA, and MADDPG. I evaluate performance under varying vehicle densities and cache capacities.
NOMA vs. OMA Performance
I first quantify the benefit of NOMA over OMA. The following table shows throughput and delay for different vehicle counts.
| Number of Vehicles | NOMA Throughput (Gbps) | OMA Throughput (Gbps) | NOMA Delay (ms) | OMA Delay (ms) |
|---|---|---|---|---|
| 20 | 8.1 | 6.2 | 120 | 181 |
| 60 | 12.3 | 7.5 | 145 | 382 |
| 100 | 15.6 | 8.9 | 195 | 620 |
NOMA achieves nearly 60% improvement in latency and throughput, validating its suitability for high-density IoV environments. This improvement directly stems from the ability of drone technology to exploit NOMA for simultaneous transmissions, reducing contention and enhancing spectral efficiency.
Impact of Vehicle Density
| Vehicles | Random | LFU | FIFO | DDQN | MADDPG | GT-SSA | GCQM (Proposed) |
|---|---|---|---|---|---|---|---|
| 20 | 10.2 | 28.5 | 22.1 | 38.4 | 42.7 | 45.3 | 58.6 |
| 60 | 8.4 | 22.3 | 17.6 | 31.2 | 35.1 | 37.8 | 47.3 |
| 100 | 5.6 | 15.4 | 11.2 | 23.5 | 26.8 | 28.9 | 35.2 |
My GCQM consistently outperforms all baselines. At 100 vehicles, it maintains a 35.2% hit rate, while Random falls to 5.6%. The GCN-based feature aggregation and attention-enhanced QMix enable effective cooperation among drones, mitigating the adverse effects of high density.
| Vehicles | Random | LFU | FIFO | DDQN | MADDPG | GT-SSA | GCQM (Proposed) |
|---|---|---|---|---|---|---|---|
| 20 | 520 | 380 | 410 | 310 | 280 | 260 | 190 |
| 60 | 610 | 460 | 500 | 380 | 340 | 320 | 240 |
| 100 | 720 | 550 | 590 | 460 | 410 | 390 | 275 |
The delay reduction of GCQM is attributable to higher hit rates reducing reliance on high-latency HAP links, combined with NOMA’s lower transmission delays.
| Vehicles | Random | LFU | FIFO | DDQN | MADDPG | GT-SSA | GCQM (Proposed) |
|---|---|---|---|---|---|---|---|
| 20 | 2.85 | 1.92 | 2.21 | 1.62 | 1.45 | 1.38 | 1.13 |
| 60 | 3.86 | 2.73 | 3.04 | 2.31 | 2.05 | 1.94 | 1.78 |
| 100 | 4.52 | 3.41 | 3.72 | 2.83 | 2.51 | 2.38 | 2.26 |
GCQM reduces backhaul load significantly, as more content is served locally or via cooperative U2U links.
| Vehicles | Random | LFU | FIFO | DDQN | MADDPG | GT-SSA | GCQM (Proposed) |
|---|---|---|---|---|---|---|---|
| 20 | 88 | 70 | 75 | 58 | 52 | 48 | 38 |
| 60 | 95 | 78 | 82 | 64 | 58 | 54 | 42 |
| 100 | 102 | 85 | 89 | 71 | 64 | 60 | 46.6 |
Energy efficiency improves because higher cache hit rates reduce repeated transmissions and long-haul communication.
Impact of Cache Capacity
| Capacity | Random | LFU | FIFO | DDQN | MADDPG | GT-SSA | GCQM (Proposed) |
|---|---|---|---|---|---|---|---|
| 100 | 4.8 | 12.3 | 9.5 | 18.7 | 21.4 | 23.1 | 28.4 |
| 200 | 5.6 | 15.4 | 11.2 | 23.5 | 26.8 | 28.9 | 35.2 |
| 300 | 6.1 | 17.8 | 12.6 | 27.0 | 31.2 | 33.5 | 45.5 |
GCQM achieves a 60.2% increase in hit rate when capacity doubles from 100 to 300 MB, while DDQN only improves 44.3%. This demonstrates the GCN’s ability to predict popular content and allocate cache efficiently.
| Capacity | Random | LFU | FIFO | DDQN | MADDPG | GT-SSA | GCQM (Proposed) |
|---|---|---|---|---|---|---|---|
| 100 | 680 | 520 | 560 | 440 | 400 | 380 | 310 |
| 200 | 720 | 550 | 590 | 460 | 410 | 390 | 275 |
| 300 | 750 | 570 | 610 | 480 | 430 | 405 | 240 |
Increasing cache capacity reduces delay across all methods, but GCQM’s delay drops more significantly due to coordinated caching decisions.
| Capacity | Random | LFU | FIFO | DDQN | MADDPG | GT-SSA | GCQM (Proposed) |
|---|---|---|---|---|---|---|---|
| 100 | 4.21 | 3.12 | 3.45 | 2.65 | 2.38 | 2.25 | 1.86 |
| 200 | 4.52 | 3.41 | 3.72 | 2.83 | 2.51 | 2.38 | 2.26 |
| 300 | 4.78 | 3.65 | 3.95 | 3.01 | 2.65 | 2.52 | 2.10 |
GCQM’s backhaul load is only 58.8% of DDQN’s at 200 MB capacity, highlighting the effectiveness of U2U cooperative sharing.
| Capacity | Random | LFU | FIFO | DDQN | MADDPG | GT-SSA | GCQM (Proposed) |
|---|---|---|---|---|---|---|---|
| 100 | 95 | 78 | 83 | 66 | 60 | 56 | 48 |
| 200 | 102 | 85 | 89 | 71 | 64 | 60 | 46.6 |
| 300 | 108 | 90 | 94 | 75 | 68 | 63 | 48.5 |
The energy consumption of GCQM decreases by 58.2% from 100 to 300 MB, while others show less improvement. The attention mechanism reduces redundant communication among drones, enhancing energy efficiency.
Ablation Study: Role of GCN
| Training Episodes | GCQM (full) | GCQM-MeanPool | GCQM-NoGCN |
|---|---|---|---|
| 0 | 0.2 | 0.2 | 0.2 |
| 1000 | 0.65 | 0.48 | 0.40 |
| 2000 | 0.82 | 0.62 | 0.52 |
| 3000 | 0.90 | 0.70 | 0.58 |
| 4000 | 0.95 | 0.75 | 0.62 |
| 5000 | 0.98 | 0.78 | 0.65 |
The full GCQM achieves a final average reward of 0.98, significantly higher than GCQM-MeanPool (0.78) and GCQM-NoGCN (0.65). This validates that GCN with attention captures complex interdependencies among drones, while mean pooling flattens information and removing GCN leads to local optima.
Conclusion
I have presented a comprehensive collaborative caching framework for drone-assisted Internet of Vehicles, integrating NOMA, dynamic K-Means++ clustering, GCN feature aggregation, and attention-enhanced QMix reinforcement learning. The extensive simulation results demonstrate that my proposed GCQM scheme significantly outperforms state-of-the-art methods in terms of cache hit rate, average delay, backhaul load, and energy consumption across various vehicle densities and cache capacities. The synergy between NOMA and drone technology enables high spectral efficiency and low-latency content delivery, while the graph-based learning ensures robust cooperation among drones even in highly dynamic environments. This work provides a solid foundation for future research on joint trajectory planning, heterogeneous network integration, and adaptive drone swarm management, further pushing the boundaries of drone technology in intelligent transportation systems.