Abstract
Channel state information (CSI) is crucial for down link beamnforming in massive multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) systems. However, obtaining timely and accurate CSI remains a challenge in frequency division duplex (FDD) systems due to the time-varying channel and feedback delay. This paper proposes a deep learning (DL)-based joint prediction and feedback network for eigenvector-based CSI, termed as PCRANet, which could leverage history CSI to predict and feedback future CSI. On one hand, the PCRANet could fit for different subband numbers of input CSI. On the other hand, to enable flexible input lengths of temporal-domain CSI sequences for various UE speeds, padding and masking operations are employed to preprocess input CSI sequences. Furthermore, a feedback bit control unit (FBCU) module is introduced to realize the variable feedback overhead. Experiments indicate that the proposed PCRANet outperforms separated prediction-feedback schemes, and maintains robust performance across various temporal-frequency sizes of input CSI and adjustable feedback overhead.