Lightweight Convolutional Neural Network for Multiscale Temporal Modeling in Ultra-Reliable Low-Latency Communication

Ultra-Reliable Low-Latency Communication (URLLC) plays a crucial role as one of the key components of 5 G communication services. Predictive tasks tailored for URLLC scenarios can provide support for latency stability. However, current time-series models fail to consider the correlation between different scales of periods in latency, making it difficult to effectively capture the statistical characteristics of real data cycles. Moreover, existing computer vision models and natural language processing lack adaptation to the physical meaning of 5G Channel State Information (CSI), resulting in poor performance. In this paper, we propose a lightweight convolutional neural network based on depthwise separable modules for multiscale temporal modeling. Specifically, we employ channels with independently learnable weights of different sizes to separately capture cycles, trends, and short-term fluctuations, thus breaking the patterns of inherent cycles captured by traditional methods. Validation in real-world URLLC scenarios demonstrates the effectiveness of the proposed approach.