Accurate Channel Prediction Based on Spatial-Temporal Electromagnetic Kernel Learning

In wireless communication, accurate and efficient channel prediction is essential for addressing channel aging caused by user mobility. However, the actual channel variations over time are complex in high-mobility scenarios. This complexity makes it difficult for existing channel predictors to obtain future channels accurately. To overcome channel aging, we propose a channel prediction scheme based on spatial-temporal electromagnetic kernel learning (STEM-KL). Specifically, the STEM correlation function can capture the fundamental propagation characteristics of the wireless channel, making it suitable to use as a kernel function that incorporates prior information. For the channel prediction problem especially, we redesign the hyperparameters of the STEM kernel, including user velocity and concentration, which characterize the direction of the EM wave. The hyperparameters are obtained through kernel learning. Then, we use Bayesian inference to predict the future channels, employing the STEM kernel as the required covariance. To further improve the stability and model expressiveness, we propose a grid-based electromagnetic mixed kernel learning (GEM-KL) scheme. We design the mixed kernel to be a convex combination of multiple sub-kernels, where each of the sub-kernels corresponds to a grid point in the parameter space. This approach transforms the learning of concentration and speed hyperparameters into the learning of weights for different subkernels, helping the kernel learning process avoid local optima. Finally, simulation results verify that the proposed STEM-KL schemes outperform the baseline schemes.