CPSDNet: Cross-Power Spectral Density Guided Contrastive Learning for Channel Prediction in MISO-OFDM Systems

This letter presents CPSDNet, a novel Cross-Power Spectral Density (CPSD)-guided framework for channel state information prediction in MISO-OFDM systems. Supporting both time and frequency division duplexing, CPSDNet leverages self-supervised learning to predict future downlink CSI from historical uplink observations. Our framework introduces dual innovations: a contrastive regularization mechanism aligning frequency-domain and delay-domain representations, and a CPSD-based data augmentation strategy preserving channel coherence properties. Simulations across UMa, CDL-A, and COST2100 channel models demonstrate superior performance under various mobility scenarios and signal-to-noise ratio (SNR) conditions. CPSDNet achieves up to 19.2% average NMSE improvement while consistently outperforming state-of-the-art approaches across spectral efficiency and symbol error metrics. With 97.4% fewer parameters than recent deep learning models, CPSDNet delivers consistent performance advantages across diverse propagation environments while maintaining computational efficiency suitable for resource-constrained wireless systems.