Application of convolutional temporal fusion networks in spectrum optimization for UAV swarms

Given that a single UAV often experiences significant performance degradation when executing complex tasks due to limitations in endurance, payload capacity, and sensing range, UAV swarms have attracted considerable attention as a means to enhance task execution efficiency. However, as the size of UAV swarms increases, the demand for spectrum resources rises substantially. This challenge becomes especially pronounced when multiple clusters simultaneously transmit data, resulting in dynamic variations in data volume as tasks switch, which further exacerbates issues related to spectrum scarcity and interference. Consequently, optimizing spectrum resources under dynamically changing communication tasks and interference-prone environments, in order to ensure the stable and efficient operation of UAV swarms, has emerged as a critical problem that urgently needs to be addressed.

This study proposed a spectrum resource optimization algorithm for UAV swarms based on a convolutional temporal fusion network. The proposed algorithm integrated the local feature extraction capabilities of CNNs (convolutional neural networks) with the dynamic sequence modeling capabilities of LSTM (long short-term memory) networks to form a CL-Net (convolutional-temporal fusion network). Furthermore, the approach incorporated policy optimization using DDQN (Double Deep Q-Networks). Specifically, the CNN module was employed to extract useful local features from the input state space, thereby mitigating the negative impact of high-dimensional state spaces on training efficiency. The LSTM was utilized to capture long-term dependencies within temporal sequences, enhancing the retention of key interference patterns across frequency bands. These two components work synergistically and were integrated with DDQN to optimize the sequential decision-making process during Q-value estimation.

The proposed CL-Net algorithm was compared with LSTM-DDQN, DQN, and DDQN in terms of the reward curve, swarm throughput curve, spectrum collision rate curve, and power consumption per unit throughput. Across all four evaluation metrics, CL-Net demonstrates clear advantages. Specifically, CL-Net achieves the reward peak in the shortest time and maintained stability after reaching the peak, indicating both high efficiency and stability in learning the optimal policy. In terms of throughput, the proposed algorithm achieves the highest throughput values, demonstrating its effectiveness in optimizing spectrum resources and improving system throughput. With respect to spectrum collision rate, CL-Net remained stably around 0.2, which is significantly better than the other baseline algorithms. This result indicates that the proposed method can proactively avoid interfered channels and substantially reduce the probability of encountering interference, thus minimizing transmission disruption. Regarding power consumption, CL-Net achieved the lowest energy consumption among the four algorithms, showing that it can maintain higher throughput while ensuring better energy efficiency.

The proposed algorithm exhibits enhanced adaptability and flexibility for UAV swarms operating under dynamically varying communication tasks and interference. Compared with traditional approaches, it achieves superior performance in optimal policy discovery, spectrum resource utilization efficiency, and system stability.