Network Spectrum Resource Allocation and Optimization Based on Deep Learning and TRDM

This study proposes a Transformer-based real-time adaptive scheduling method (TRDM) to improve performance under dynamic network loads and low latency requirements. Compared with traditional scheduling methods, TRDM shows significant advantages in throughput, latency, QoS satisfaction, and computational efficiency. To verify the effectiveness of the method, we use a dataset of 10,000 samples covering a variety of network environments and load conditions. Standard deep learning benchmarks are used in the experiments, and the performance of TRDM is compared with the state-of-the-art (SOTA) and traditional methods under the same experimental settings. Experimental results show that TRDM has significant advantages in throughput and latency, especially in scenarios with high load and high realtime requirements. By adopting the Soft Actor-Critic (SAC) reinforcement learning algorithm for model training, we test TRDM in multiple network environments and verify its superiority in real-time adaptability and computational efficiency. Compared with the SOTA method, TRDM shows better realtime and adaptability, especially under low latency and dynamic load conditions. In addition, TRDM maintains high computational efficiency when dealing with complex network environments, and is suitable for practical application scenarios such as large-scale IoT or urban cellular networks. The experimental settings include hyperparameters such as a 6-layer self-attention network of the Transformer architecture, a learning rate of 0.001, and a batch size of 128. Through comparative experiments, TRDM outperforms existing methods in multiple key performance indicators, and the performance difference is statistically significant. This study provides an effective solution for real-time network scheduling and provides an important reference for future deployment in applications such as the IoT and remote urban networks.