High-Accuracy Predictive Channel Modeling for 6G Wireless Communications With an Improved Diffusion-Driven Learning Framework

To address sparse channel measurement data and inadequate predictive capabilities in conventional channel models, predictive channel modeling employs joint generative and predictive architectures to enhance robustness. In this paper, we propose an enhanced diffusion-driven predictive framework that integrates generative augmentation and prior-aware prediction into a unified learning pipeline. We first introduce a space-time-frequency (STF) coupled diffusion network based on transformers that generates synthetic channel data preserving critical channel statistical properties. Additionally, we compress measured channel state information into a low-dimensional manifold via a latent encoder and introduce an innovative composite training scheme that couples diffusion-driven prior generation with prediction, equipping the predictive module with rich latent features that lift its performance ceiling and markedly improve generalization across diverse scenarios. Extensive experiments confirm the superiority of our algorithm, and its performance is further validated using channel measurement data, thereby demonstrating its robustness for advanced wireless communications in real-world deployment scenarios.