Learning Based CSI Feedback Prediction for 5G NR

Acquisition of accurate channel state information (CSI) is critical in multiple-input and multiple-output (MIMO) systems to achieve efficient link adaptation. As CSI is typically estimated at the receiver, effective and efficient acquisition at the transmitter is challenging. The primary concerns are estimation accuracy, reporting overhead, and channel aging effects caused due to delayed CSI usage. In this paper, we design a deep learning based CSI prediction framework (DCP) to address these challenges. The DCP consists of a channel prediction network to compensate the aging effects, followed by a deep learning based CSI estimator (DCE) for accurate estimation. Through simulations, we show that the DCE can yield up to 15% higher spectral efficiency (SE) due to the improved estimation accuracy while significantly lowering the complexity compared to conventional approaches. Moreover, for the same CSI reporting overhead, DCP can improve the SE by up to 20% over conventional techniques. Further, even at 50% of the CSI reporting overhead, DCP improves the SE by up to 6.5% over the conventional methods.