CSI Prediction Using Diffusion Models

TL;DR

This work tackles the problem of accurate CSI prediction under pilot overhead and channel aging by proposing a probabilistic diffusion-based framework that forecasts future CSI from historical observations. It decomposes the task into a temporal encoder and a diffusion-based generator, and supports autoregressive and sequence-to-sequence inference with multiple backbones (e.g., U-Net, DiT). The approach yields probabilistic predictions, captures the stochastic, multimodal nature of wireless channels, and demonstrates substantial NMSE gains over state-of-the-art baselines across CDL-based mmWave MIMO scenarios, with fast sampling via DDIM and strong generalization to unseen environments. Overall, diffusion-based CSI predictors offer robust, scalable, and latency-efficient improvements for real-world wireless systems.

Abstract

Acquiring accurate channel state information (CSI) is critical for reliable and efficient wireless communication, but challenges such as high pilot overhead and channel aging hinder timely and accurate CSI acquisition. CSI prediction, which forecasts future CSI from historical observations, offers a promising solution. Recent deep learning approaches, including recurrent neural networks and Transformers, have achieved notable success but typically learn deterministic mappings, limiting their ability to capture the stochastic and multimodal nature of wireless channels. In this paper, we introduce a novel probabilistic framework for CSI prediction based on diffusion models, offering a flexible design that supports integration of diverse prediction schemes. We decompose the CSI prediction task into two components: a temporal encoder, which extracts channel dynamics, and a diffusion-based generator, which produces future CSI samples. We investigate two inference schemes-autoregressive and sequence-to-sequence- and explore multiple diffusion backbones, including U-Net and Transformer-based architectures. Furthermore, we examine a diffusion-based approach without an explicit temporal encoder and utilize the DDIM scheduling to reduce model complexity. Extensive simulations demonstrate that our diffusion-based models significantly outperform state-of-the-art baselines.

Figures (7)

• Figure 1: CSI prediction with a diffusion model conditioned on the temporal encoder’s latent representation, using an AR inference strategy.
• Figure 2: NMSE of different CSI prediction models over varying prediction steps at inference SNRs of $0$, $10$, and $20$ dB.
• Figure 3: NMSE performance of CSI prediction models versus inference SNR. Results are reported for the first prediction step, the last prediction step, and the average NMSE across all prediction steps.
• Figure 4: NMSE performance of CSI prediction models as a function of prediction step, evaluated on CSI samples with user mobilities of $30$ km/h, $60$ km/h, and $120$ km/h.
• Figure 5: NMSE performance versus prediction step at $\mathrm{SNR}=10$ dB. The left panel compares DiU, DiT, and DiU3; the middle panel shows different seq2seq variants; and the right panel presents DiU with varying context lengths.