A Novel Grant Prediction Method for 5G NR Terminals

Abstract: 5G NR user equipment suffers from high power consumption due to continuous PDCCH monitoring. Predictive dynamic power management (DPM) can save energy by forecasting data grants, but accurate prediction is challenging due to unobservable scheduling states and bursty grant patterns. This paper proposes IOHMM-BO, a high-order input-output hidden Markov model with Bayesian optimization. Based on real 5G NR traces, we capture long-range dependencies via a compound state and jointly optimize model order and listening window using Bayesian optimization. Experiments on real traces show that IOHMM-BO achieves 45.3% accuracy, 5.0% false negative rate, and 43% energy saving with low computational overhead. The method provides a balanced trade-off between reliability and energy efficiency.

• The paper introduces a high-order IOHMM-BO model that leverages temporal dependencies in grant sequences for precise prediction.
• It employs a two-level optimization combining EM and Bayesian methods to balance energy savings with minimal false negative rates.
• Experimental results on diverse real-world 5G traces demonstrate improved accuracy, reduced energy consumption, and low computational overhead.

Grant Prediction for Energy-Efficient 5G NR UE: IOHMM-BO Approach

Introduction

The proliferation of 5G New Radio (NR) networks has amplified the challenge of energy consumption in user equipment (UE), primarily due to continual monitoring of the physical downlink control channel (PDCCH). Existing discontinuous reception (DRX) mechanisms are insufficient for fine-grained energy savings as UE must remain active for potential grants, even when actual data arrivals are infrequent. Predictive dynamic power management (DPM), reliant on accurate forecasting of grant arrivals, offers a viable avenue for reducing energy waste. However, accurately predicting grant arrivals is difficult due to bursty, temporally correlated grant patterns and unobservable aspects of the scheduling state. This paper introduces a high-order Input-Output Hidden Markov Model with Bayesian Optimization (IOHMM-BO) for predictive DPM, designed to exploit the long-memory and input-dependent structure evident in real 5G NR traces (2606.22125).

Data-Driven Motivation and Sequence Analysis

Extensive real-world data collection across cities, applications, and scenarios reveals two salient properties of grant indicator sequences: (i) grants occur in bursty clusters, reflecting periods of scheduler activity corresponding to pending data, and (ii) scheduling requests (SR) from UE significantly increase the conditional probability of a grant in subsequent transmission time intervals (TTIs), with effects persisting for tens of TTIs. These findings contradict the assumption of memoryless or independent grant arrivals and necessitate a model architecture encapsulating both temporal dependencies and exogenous signaling.

The modeling approach leverages hidden states to abstract the gNB scheduler’s unobservable internal dynamics, while observable features such as SR events and DRX state transitions serve as input variables, enabling explicit conditioning in prediction.

IOHMM-BO: Model Architecture and Optimization

The grant prediction task is formulated as a high-order IO-HMM, where the hidden state at time $t$ depends on the compound history of preceding $K$ states and inputs, and the observation likelihood similarly incorporates high-order dependencies. State space explosion issues are mitigated via aggregation into equivalent first-order Markov models with compound states. Parameter estimation is handled through an EM algorithm, with emission and transition probabilities parameterized by weighted logistic regression over feature maps derived from stacked input vectors.

Critical model hyperparameters—the order $K$ and maximum monitoring window length $W_{dz}$—along with model parameters, are optimized by a two-level scheme. The inner loop maximizes likelihood via EM for fixed hyperparameters; the outer loop evaluates performance on validation data and applies Bayesian Optimization using a custom acquisition function that penalizes excessive false negative rate (FNR), ensuring a desired trade-off between energy savings and grant miss-detection risk.

Online grant prediction leverages the inferred state distribution and computes a per-TTI grant probability. Rather than using static thresholds, the method implements a probability-driven, adaptive listening window policy that proportionally extends monitoring durations in phases of heightened grant likelihood while reducing unnecessary listening during inactivity.

Experimental Evaluation

Experiments are conducted on 80 traces spanning prominent applications and scenario diversity, covering both indoor and outdoor environments. The grant sequence is highly imbalanced (≈10% TTIs with actual grants), exacerbating the modeling challenge.

Comparison with Baselines

Three baseline predictive DPM approaches are considered: a conventional FFNN, a deep Q-network (DQN) for RL-based prediction, and SARSA-X, a tabular RL method with eligibility traces. Results demonstrate:

• IOHMM-BO achieves the highest average overall accuracy (45.3%) while maintaining a low FNR (5.0%), outperforming both DQN (ACC: 40.7%, FNR: 23.2%) and SARSA-X (ACC: 31.1%, FNR: 24.7%). While FFNN yields a low FNR (4.6%), it does so at the cost of markedly low accuracy, indicating near-constant prediction of "no grant".
• Regarding energy efficiency, IOHMM-BO demonstrates a 43% reduction in normalized receiver energy consumption relative to reactive DPM. While DQN achieves slightly higher energy savings (46%), this comes with a much higher FNR.
• Computational overhead for IOHMM-BO is modest (about 2,180 FLOPs/TTI for $K=4$), far lower than DQN (14,520 FLOPs) and comparable to or higher than FFNN and SARSA-X. The paper rigorously accounts for both receiver operation and algorithmic computation energy in its assessment.

Sensitivity and Trade-off Analysis

Systematically varying the FNR penalty threshold in the Bayesian optimization objective reveals a controlled trade-off. Permitting higher FNRs allows for more aggressive sleeping, reducing normalized energy by up to ≈15% as threshold increases from 0.02 to 0.10, while FNR increases only moderately. Notably, accuracy slightly improves as overly tight penalties constrain model flexibility, hampering prediction quality.

Implications and Potential Extensions

The IOHMM-BO framework yields a balanced, practical predictive DPM solution for 5G NR UEs, effectively capturing grant temporal dependencies and external signaling influences assumed to characterize future wireless protocols. The joint parameter and hyperparameter optimization framework ensures operable trade-offs between energy efficiency and communication reliability, addressing a key requirement for commercial adoption.

The theoretical implications include stronger modeling of grant sequence structure and offering a path for formal performance guarantees in the context of energy-reliability optimization under real-world operating constraints. The use of Bayesian optimization for hyperparameter tuning demonstrates efficacy in operating-point selection and provides a template for further application in adaptive wireless system design.

Practical deployment requires extension to real-time closed-loop scenarios and accommodation of application-level QoS differentiation, potentially integrating further deep sequential architectures with the statistical priors captured by IO-HMMs. Broader datasets encompassing extreme and edge-case behaviors (e.g., rare channel or handover events) will be key to robustifying the approach.

Conclusion

The IOHMM-BO method constitutes a significant advancement in predictive DPM for 5G NR UE, combining high-order temporal modeling with input conditioning and principled hyperparameter selection. Compared to both deep learning and reinforcement learning baselines, it delivers superior accuracy-FNR-energy trade-offs and practical computational cost. These results point to the merit of structured probabilistic sequential models for grant prediction, highlighting opportunities for extension to differentiated, adaptive DPM in wireless terminals. Future research directions include closed-loop validation, integration of rich application context, and further model/lightweight architecture optimization for diverse real-world deployment scenarios.