Smart-CSI: Deep Learning Based Low Complexity CSI Prediction for Beyond-5G Systems

Acquiring accurate channel state information (CSI) is challenging in MIMO systems as it is quickly outdated due to fast variations in the channel. Particularly, in frequency division multiplexing (FDD) systems the problem is aggravated due to sparse frequency of CSI reporting used in practical systems; this leads to user throughput degradation. CSI prediction effectively addresses this issue, but generally at the cost of computational complexity. Traditional techniques predict using the spatial channel, hence the size of input and output scales with the number of antennas used, and the operating bandwidth, thus increasing the complexity. In this paper, we solve this problem using a DL based low-complexity approach called Smart-CSI prediction. In this approach, we predict the variation of channel capacity to predict the future CSI. Consequently, it significantly reduces the operating complexity compared to the existing prediction methods. Numerical results indicate that Smart-CSI effectively predicts the CSI in simulated and real-world channel conditions. Further, we evaluate the approach using a proof-of-concept implementation and show that Smart-CSI achieves 11.8% higher throughput, on a average, compared to the no-prediction scenario.