A Cascaded BiLSTM-Based Cross-Dimensional Prediction Model for Non-Stationary Channels

Non-stationary channel prediction remains a critical challenge in 5G/6G high-frequency communications. Traditional linear time-series models with fixed parameters fail to capture multi-scale interactions (e.g., multipath fading, Doppler shifts) and abrupt signal variations. To address these issues, we propose a cascaded bidirectional Long Short-Term Memory (Bi-LSTM) framework for joint channel state prediction. The model uses bidirectional feature fusion. The forward layer models channel gain decay and delay spread pulses. The backward layer analyzes phase offset variations and signal-to-noise ratio (SNR) dynamics. Dropout regularization is added to prevent overfitting. Experiments in 2.4 GHz dynamic channels show significant improvements. The bidirectional architecture demonstrates superior Doppler shift tracking capability through backward-layer phase dynamics analysis, while achieving multi-noise robustness with 89.8% SNR error reduction in dense scattering environments. Experimental validation shows 0.9751 R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> for channel gain prediction and 93% MAE reduction in delay spread estimation, establishing a cross-dimensional solution for high-mobility wireless communications.