Abstract
The cloud radio access network (C-RAN) is a key paradigm for 5G wireless systems, where traffic prediction enables load-aware management to improve spectrum and energy efficiency. This paper introduces a scalable Gaussian process (GP) framework for large-scale wireless traffic prediction. By integrating the alternating direction method of multipliers (ADMM) for parallel training and a cross-validation-based fusion strategy for prediction, the framework achieves high accuracy while managing computational complexity. Experimental results on real-world datasets demonstrate that this scalable GP model significantly outperforms state-of-the-art approaches.
Introduction: The Scalability Challenge in 5G
As 5G networks expand, operators increasingly rely on "load-aware management" to optimize resource allocation and reduce energy consumption by dynamically activating or deactivating hardware based on demand. Accurate traffic prediction is essential to ensure service quality during these transitions.
Gaussian Processes (GPs) are highly regarded for this task due to their non-parametric nature and ability to provide uncertainty estimates. However, their $O(N^3)$ computational complexity makes them impractical for large-scale deployments. This research addresses the "scalability wall" by proposing a framework that maintains the probabilistic benefits of GPs while enabling real-time execution in a C-RAN environment.
Scalable Gaussian Process Framework
The proposed solution leverages the distributed architecture of C-RANs, where baseband units (BBUs) are pooled to handle processing for multiple remote radio heads (RRHs).
1. Tailored Kernel Design
Wireless traffic exhibits distinct temporal patterns. The authors designed a hybrid kernel composed of three parts:
- Weekly Periodic Component: Captures long-term cycles.
- Daily Periodic Component: Models daily fluctuations.
- Squared Exponential (SE) Component: Accounts for irregular, dynamic deviations from periodic norms.
2. Parallel Training via ADMM
To overcome the $O(N^3)$ complexity, the paper introduces a parallel hyper-parameter optimization method using the Alternating Direction Method of Multipliers (ADMM). This allows training to be distributed across $K$ BBUs, reducing complexity to $O(N^3/K^3)$. Each BBU optimizes parameters on a data subset, while ADMM ensures global consensus.
3. Optimal Prediction Fusion
In the prediction phase, local estimates from different BBUs are combined using a cross-validation-based weighting strategy. This approach is more robust than simple averaging, as it prioritizes the most reliable local models based on their historical performance, maintaining a closed-form GP inference.
Experimental Verification
The framework was validated using a massive dataset of hourly downlink Physical Resource Block (PRB) usage from 3,072 4G base stations in China.
Key Results
- High Accuracy: The scalable GP achieved up to 97% prediction accuracy, outperforming SARIMA, Sinusoid Superposition, and LSTMs.
- Computational Efficiency: By exploiting the Toeplitz structure of regularly sampled data, training time for 700 data points was reduced from 16.8 seconds to 0.1 seconds when scaled across 16 BBUs.
- Uncertainty Quantification: Unlike "black-box" models, the GP provides a 95% confidence interval, allowing operators to make informed decisions about resource buffering.
Conclusion
This scalable GP framework provides a cost-efficient and accurate solution for large-scale wireless traffic prediction. By balancing local estimation with global consensus, it enables the practical application of sophisticated probabilistic models in the 5G era. Future work will explore automatic kernel discovery and further optimization of data subset partitioning to enhance performance in highly volatile environments.
Summary updated