Frequency Extrapolation for Carrier Aggregation as a Super-Resolution Problem: Rethinking Conventional Forecasting Methods

Daoud Burghal, Yan Xin, Jianzhong Charlie Zhang

Keywords: Time Series Forecasting, Frequency Extrapolation, Carrier Aggregation, Vision-Based Transformers

TL;DR: Formulating the challenging problem of frequency extrapolation as a super-resolution problem in the inverse domain shows promising gains. We highlight the differences from DL-based series forecasting and the challenges in the wireless domain.

Abstract:

Frequency extrapolation plays a pivotal role in modern wireless communication systems, particularly in Carrier Aggregation (CA), where service providers combine multiple Carrier Components (CCs) to enhance system throughput (Tput). This technique enables the efficient utilization of available spectrum, ensuring robust and high-performance connectivity. However, the effectiveness of CA heavily relies on accurate Channel State Information (CSI) for each CC, which is essential for tasks such as precoding and User Equipment (UE) coordination. In practice, obtaining CSI beyond the primary CC (Primary Cell, PCell) is challenging, necessitating advanced prediction methods to estimate CSI for secondary CCs (SCell).

Frequency extrapolation can be framed as a series forecasting problem, where the goal is to predict future CSI values based on historical data. Traditional time-series forecasting methods, while effective in many domains, fall short in the context of wireless communication. This is due to the unique challenges posed by the dynamic and multi-user nature of wireless environments, where predictions must account for varying signal characteristics across multiple users and scenarios. To address these limitations, this work redefines frequency extrapolation as a super-resolution (SR) problem, leveraging advanced deep learning techniques to enhance the accuracy and robustness of CSI predictions.

In this study, we propose a novel solution that applies SR to combined CCs in the inverse domain (delay domain). By utilizing state-of-the-art deep learning algorithms, including vision-based transformers, we demonstrate that our approach outperforms existing AI-based solutions. Our simulation results highlight the superior prediction robustness and significant throughput gains achieved over a wide frequency range, reaffirming the potential of super-resolution techniques in revolutionizing frequency extrapolation for Carrier Aggregation.

Submission Number: 74