An LLM-Enhanced Conditional Diffusion Model for Mobile Traffic Prediction

Accurate traffic prediction is critical for ensuring the efficient operation of mobile networks and maintaining a seamless user experience. However, existing approaches face several significant challenges: high costs and incomplete environmental perception, difficulty capturing spatial correlations between base stations, and limited adaptability to irregular traffic fluctuations driven by user activities. This article proposes an LLM-enhanced Conditional Diffusion (LEC-Diff) model for cellular traffic prediction to address the abovementioned challenges. First, we utilize easily accessible aerial images to describe the static environmental context surrounding base stations. We further enhance this representation by leveraging comprehensive textual data from a Large Language Model (LLM) for these images, extracting the abundant inherent knowledge embedded within the LLM. Second, we employ a Graph Neural Network (GNN) to automatically model the spatial dependencies between base stations and enhance spatiotemporal information through mapping. Finally, we introduce a conditional diffusion model to capture complex traffic distributions by conditioning predictions on static environmental features and dynamic historical traffic features. Extensive experiments demonstrate that our proposed model surpasses state-of-the-art methods by over 5% in mobile traffic prediction.