CSI prediction for enhanced feedback in emerging cellular systems

Abstract

Accurate and timely Channel State Information (CSI) feedback is a critical requirement in cellular networks, as it enables enhanced Downlink (DL) performance. However, obtaining accurate and timely CSI reports becomes more challenging in dynamic Fifth-Generation (5G) deployments, especially at high User Equipment (UE) speeds, due to rapid channel variations. CSI prediction has been identified by the 3rd Generation Partnership Project (3GPP) as a solution to combat the channel aging issue in current cellular systems. The evaluation results generated using the simulation framework proposed in this thesis allow for the verification of crucial claims made by 3GPP in its TR 38.843 and provide novel insights about CSI prediction. Quantitatively illustrating the out-dated CSI issue, using the results obtained from the proposed simulation framework, the thesis provides a timely review of 3GPP standards and literature works on CSI prediction. This thesis investigates the performance of Machine Learning (ML) and non-ML models for predicting CSI in cellular communication networks. The study compares ML techniques, namely Long Short-Term Memory (LSTM) and convolutional LSTM (convLSTM), and statistical prediction methods, namely Auto-Regressive (AR), to the Zero-Order Hold (ZOH) approach used in conventional cellular systems. The models are evaluated using metrics such as Normalized Mean Squared Error (NMSE) and Squared Generalized Cosine Similarity (SGCS) to assess their prediction accuracy and ability to capture channel variations. The generalization capability of the considered ML models is evaluated over different UE speeds and channel models.