Organizations: Ericsson Research, Stockholm, Sweden · Department of Computer and Information Science, Linköping University, Linköping, Sweden · Department of Electrical Engineering, Linköping University, Linköping, Sweden
This paper considers channel state information (CSI) prediction using neural networks for time-varying multi-antenna channels from noisy historical observations. The main challenge is to exploit temporal correlation while distinguishing channel evolution from observation noise. We propose JEPA-kMMVAE, which combines a joint-embedding predictive architecture with a k-memory Markov variational autoencoder to predict future CSI through temporally structured latent representations. The framework supports direct prediction and residual corrections relative to the latest latent state and noisy CSI observation. Simulation results show that the residual configuration outperforms the evaluated baselines at moderate and high input CSI signal-to-noise ratios, while sequence baselines remain more accurate under stronger noise.
Figures & tables
Figure 1: Causal CSI prediction from C noisy observations. The predictor maps the available history Yt to the next clean snapshot estimate xt+1 . The future target is available only for training and evaluation.
Figure 2: JEPA- k MMVAE residual prediction framework. Noisy history drives latent prediction and residual CSI decoding. The dashed branch supplies clean latent targets during training; the CSI output is supervised by the clean future snapshot.
Method
5 dB
10 dB
15 dB
20 dB
25 dB
Persistence
−4.75
−9.24
−12.94
−15.35
−16.50
Linear extrapolation
+2.01
−2.95
−7.84
−12.49
−16.53
Learned linear
−8.98±0.01
−11.32±0.01
−12.43±0.03
−12.85±0.03
−12.99±0.04
GRU
−9.69±0.01
−12.48±0.06
−14.28±0.13
−15.10±0.17
−15.41±0.19
LSTM
−9.34±0.03
−11.73±0.06
−13.14±0.14
−13.75±0.18
−13.97±0.20
Transformer
−10.02±0.06
−12.44±0.06
−13.72±0.07
−14.23±0.07
−14.41±0.07
Table 1: Prediction NMSE in dB (lower is better). Values for learned models are mean ± sample standard deviation over three runs. Both JEPA- k MMVAE variants use d=64 . Bold indicates the lowest mean in each column.
Figure 3: Direct and residual JEPA- k MMVAE prediction at d=64 . Curves show mean CSI prediction NMSE; error bars denote ± one sample standard deviation over three runs.
Ericsson Research, Stockholm, Sweden · Department of Computer and Information Science, Linköping University, Linköping, Sweden · Department of Electrical Engineering, Linköping University, Linköping, Sweden