Transformer-Based Rate Prediction for Multi-Band Cellular Handsets

TL;DR

This work tackles the challenge of predicting achievable rates across multiple frequency bands and antenna arrays for mobile UEs under sparse measurements. It proposes a transformer-based predictor that ingests asynchronous CQI histories, uses a temporal encoder per antenna, and a single antenna-level Transformer to forecast per-array rates, evaluated on ray-traced FR1/FR3 urban scenarios. The approach outperforms simple baselines, offering more accurate and robust rate predictions with reduced lag and fewer large errors, particularly under mobility. These predictions enable more informed band selection and handover decisions in dense, multi-band deployments, potentially reducing energy use and latency in next-generation cellular systems.

Abstract

Cellular wireless systems are witnessing the proliferation of frequency bands over a wide spectrum, particularly with the expansion of new bands in FR3. These bands must be supported in user equipment (UE) handsets with multiple antennas in a constrained form factor. Rapid variations in channel quality across the bands from motion and hand blockage, limited field-of-view of antennas, and hardware and power-constrained measurement sparsity pose significant challenges to reliable multi-band channel tracking. This paper formulates the problem of predicting achievable rates across multiple antenna arrays and bands with sparse historical measurements. We propose a transformer-based neural architecture that takes asynchronous rate histories as input and outputs per-array rate predictions. Evaluated on ray-traced simulations in a dense urban micro-cellular setting with FR1 and FR3 arrays, our method demonstrates superior performance over baseline predictors, enabling more informed band selection under realistic mobility and hardware constraints.

Figures (6)

• Figure 1: Top: Ray tracing capacity in an urban micro–cell. Top-left: difference $\Delta C=C_{15}-C_{3.5}$ (red$>$0$\rightarrow$15GHz, blue$<$0$\rightarrow$3.5GHz, Mbps). Top-right: per-band capacity at 15GHz and 3.5GHz (shared color bar, Mbps). Bottom: zoom of the dashed area with the site model and simulated pedestrian trajectories with handset rotations, for UE mobility and UE–gNB interaction, yielding time-varying band/array rates.
• Figure 2: Illustration of UE radian pattern. The red rectangular shows the UE body shape as 7cm $\times$ 15cm $\times$ 1cm. New directive 15GHz FR3 radian pattern is shown in blue while the traditional 3.5GHz dippole pattern is shown as in yellow. The figure demostrates the horizontal cut of the UE radian pattern in left with virtical cut in right.
• Figure 3: Overall rate estimation network. Each antenna’s recent history is passed through a learnable per-feature affine scaling that preserves units for rate and SNR while improving numerical conditioning. The scaled and raw features are concatenated. A lightweight temporal encoder summarizes each antenna’s $W$-sample window into one embedding. An antenna-level Transformer with one encoder layer mixes cross-antenna dependencies. The result is concatenated with the raw mean over time, which averages only along the time dimension $W$ and keeps the original antenna count $N$ and feature dimension $F$ unchanged. A shared head with Softplus produces non-negative per-antenna rate estimates. A per-antenna capacity cap conditioned on bandwidth and hardware is applied.
• Figure 4: Channel Capacity at three-level mobility. As the mobility level increases, the capacity at each RX exhibits more frequent variations. This increased temporal variability makes it challenging to maintain high transmission rates using only instantaneous feedback. Therefore, accurate rate prediction becomes essential to anticipate channel changes and enable timely adaptation of transmission strategies.
• Figure 5: Time-series performance under medium mobility. Top: comparison of the maximum predicted capacity of the proposed model, the masked-previous benchmark, and the maximum ground-truth capacity; the proposed model tracks rapid swings with less lag and fewer overshoots. Bottom: per-antenna MSE over time (gray shading marks timestamps with measurements); the proposed model (orange) is consistently below the masked-previous baseline (blue), indicating lower error both with and without measurements. Overall, per-antenna MSE is reduced by 16--38% (about 28% on average).