CSI Prediction Frameworks for Enhanced 5G Link Adaptation: Performance-Complexity Trade-offs

TL;DR

This paper compares classical Wiener filtering and deep learning methods for CSI prediction in 5G systems, highlighting trade-offs in accuracy, complexity, and generalizability across different propagation scenarios.

Contribution

It introduces and evaluates two CSI prediction frameworks applicable to TDD and FDD systems, analyzing their performance and complexity trade-offs.

Findings

Wiener filter achieves near-GRU performance with lower complexity.

GRU models generalize better across diverse channel conditions.

Classical methods are preferable for low-power FDD user equipment.

Abstract

Accurate and timely channel state information (CSI) is fundamental for efficient link adaptation. However, challenges such as channel aging, user mobility, and feedback delays significantly impact the performance of adaptive modulation and coding (AMC). This paper proposes and evaluates two CSI prediction frameworks applicable to both time division duplexing (TDD) and frequency division duplexing (FDD) systems. The proposed methods operate in the effective signal to interference plus noise ratio (SINR) domain to reduce complexity while preserving predictive accuracy. A comparative analysis is conducted between a classical Wiener filter and state-of-the-art deep learning frameworks based on gated recurrent units (GRUs), long short-term memory (LSTM) networks, and a delayed deep neural network (DNN). The evaluation considers the accuracy of the prediction in terms of mean squared error (MSE), the performance of the system, and the complexity of the implementation regarding floating point operations (FLOPs). Furthermore, we investigate the generalizability of both approaches under various propagation conditions. The simulation results show that the Wiener filter performs close to GRU in terms of MSE and throughput with lower computational complexity, provided that the second-order statistics of the channel are available. However, the GRU model exhibits enhanced generalization across different channel scenarios. These findings suggest that while learningbased solutions are well-suited for TDD systems where the base station (BS) handles the computation, the lower complexity of classical methods makes them a preferable choice for FDD setups, where prediction occurs at the power-constrained user equipment (UE).