Cross-Band Channel Prediction Model for EI Wireless Communication Systems

Abstract

A novel cross-band channel prediction model (CBCP) is proposed for embodied intelligence wireless communication systems. The basic idea of CBCP is to use CSI from a given frequency band to generate CSI predictions for multiple frequency bands. Based on the shared-encoder and independent-decoder architecture as well as the design of different loss functions, CBCP enables spatio-temporal and cross-band generative predictions. Furthermore, through the pre-training and fine-tuning paradigm, the model can capture the global correlations across multi-band channel data while preserving the key features of single-band data, thus providing synergistic support for addressing the heterogeneity and dynamic challenges in spatio-temporal and cross-band channel predictions. Experimental results demonstrate that, in the task of cross-band prediction from 2.4 GHz to 8 GHz, CBCP reduces the average normalized mean squared error by 16.8%–39.2% compared with the considered baseline methods, at the cost of a moderate degradation in spatio-temporal prediction accuracy.