Graph neural network-based cell switching for energy optimization in ultra-dense heterogeneous networks

Kang Tan, Duncan Bremner, Julien Le Kernec, Yusuf Sambo, Lei Zhang, Muhammad Ali ImranScientific 2022 · DOI

TL;DR

This paper introduces a Graph Neural Network (GNN)-based solution for intelligently switching small cell base stations on and off to save energy in ultra-dense 5G/6G networks. By modeling the network as a graph and using unsupervised learning, the proposed system achieves significant energy efficiency gains with linear computational complexity, overcoming the scalability bottlenecks of traditional heuristic and reinforcement learning approaches.

What Problem Does This Paper Address?

As we move toward 6G, cellular networks are becoming "ultra-dense," packed with a mix of Macro Cells (for coverage) and Small Cells (for high-speed data). While this improves connectivity, it creates a massive energy footprint; base stations (BSs) account for 60% to 80% of a network's total power consumption.

The standard approach is to keep all BSs "always on," even when traffic is low, leading to immense wastage. While "cell switching" (turning off idle BSs) is a known solution, finding the optimal combination of BSs to turn off is an NP-hard problem. Traditional Exhaustive Search (ES) algorithms work for tiny clusters but explode in complexity as the number of BSs grows. Meanwhile, existing Machine Learning methods like Reinforcement Learning (RL) often struggle to generalize—if you add a single new base station, the model usually needs to be completely retrained.

Methods

The authors propose GBCSS (GNN-Based Cell Switching Solution). The core innovation lies in how they represent and process the network state:

  1. Graph Representation: The network is modeled as a directed graph where each base station is a node. Node features represent current traffic loads ($\lambda$), and edge attributes represent power consumption profiles.
  2. Control Data Separated Architecture (CDSA): The model assumes a "Macro Cell" acts as a central controller, managing the traffic offloading from "Small Cells" that are put into sleep mode.
  3. GNN Architecture: They utilize a Local Extremum Convolution (LEConv) operator. This allows the model to learn the importance of specific nodes relative to their neighbors, capturing the underlying topology of the network.
  4. Unsupervised Training: Because generating "optimal" labels via Exhaustive Search is too slow for large datasets, the authors designed a custom loss function. This function directly minimizes system-wide power consumption per unit of traffic load, allowing the model to learn efficient switching policies without needing "correct" answers upfront.
  5. Complexity: Unlike the $O(2^N)$ complexity of Exhaustive Search, GBCSS operates at $O(N)$, making it viable for real-time deployment in dense urban environments.

Datasets

The study avoids synthetic data in favor of the Milan Telecommunication Dataset. This dataset is based on real-world Call Detail Records (CDRs) provided by Telecom Italia. It includes:

  • Two months of data (Nov 2013 – Jan 2014) from the city of Milan.
  • Activity types: SMS, calls, and Internet usage, aggregated into 235m x 235m grids.
  • The authors focused on 81 grids around the city center (near the Duomo di Milano) to simulate a high-density environment.

Key Insights and Ideas

  • Node Size Invariance: This is the "aha moment" for GNNs in networking. Because GNNs learn on graph structures rather than fixed-size vectors, a model trained on a network with 4 small cells can be deployed on a network with 32 small cells without retraining.
  • The QoS/Energy Balancing Act: The model doesn't just save energy; it learns the "Knapsack-style" constraint of not overloading the Macro Cell. It maintains 99.63% of original traffic loads, meaning energy savings don't come at the cost of dropped calls or slow data for the user.
  • Spatial-Temporal Awareness: By using real Milan data, the model demonstrates it can handle the "tidal effect" of urban traffic—high demand during work hours and low demand at night or on holidays—more gracefully than linear RL benchmarks.

Key Results

  • Energy Efficiency: GBCSS achieved a 10.41% energy efficiency gain compared to the "All-on" baseline.
  • Optimality: It reached 75.76% of the theoretical optimal performance (defined by the computationally expensive Exhaustive Search).
  • Superiority over RL: GBCSS outperformed a state-of-the-art RL benchmark (SARSA with linear function approximation) by 11.90% in average energy efficiency.
  • Generalization: In "node size generalization tests," models trained on small clusters (4 nodes) performed effectively when suddenly tasked with managing 12 or 32 nodes, proving the scalability of the GNN approach.

Limitations and Future Work

While promising, the authors note several areas for improvement:

  • Convergence: The GNN sometimes struggles to reach the absolute mathematical optimum when gradients become too small during training.
  • Hybrid Models: Future research could combine GNNs with Reinforcement Learning to further refine the switching policy's convergence.
  • Feature Enrichment: The current model primarily looks at traffic load. Incorporating more granular data, such as specific dates, times, or even weather/event data into the graph representation, could improve robustness.
  • Protocol Integration: The paper focuses on the algorithm; the next step is investigating how GBCSS would actually sit within the complex 3GPP protocol stacks of a real-world cellular network.

Summary updated

Abstract

The development of ultra-dense heterogeneous networks (HetNets) will cause a significant rise in energy consumption with large-scale base station (BS) deployments, requiring cellular networks to be more energy efficient to reduce operational expense and promote sustainability. Cell switching is an effective method to achieve the energy efficiency goals, but traditional heuristic cell switching algorithms are computationally demanding with limited generalization abilities for ultra-dense HetNet applications, motivating the usage of machine learning techniques for adaptive cell switching. Graph neural networks (GNNs) are powerful deep learning models with strong generalization abilities but receive little attention for cell switching. This paper proposes a GNN-based cell switching solution (GBCSS) that has a smaller computational complexity than existing heuristic algorithms. The presented performance evaluation uses the Milan telecommunication dataset based on real-world call detail records, comparing GBCSS with a traditional exhaustive search (ES) algorithm, a state-of-the-art learning-based algorithm, and the baseline without cell switching. Results indicate that GBCSS achieves a 10.41% energy efficiency gain when compared with the baseline and achieves 75.76% of the optimal performance obtained with ES algorithm. The results also demonstrate GBCSS' significant scalability and generalization abilities to differing load conditions and the number of BSs, suggesting this approach is well-suited to ultra-dense HetNet deployment.
BibTeX
@inproceedings{tan2022graph,
  author = {Kang Tan and Duncan Bremner and Julien Le Kernec and Yusuf Sambo and Lei Zhang and Muhammad Ali Imran},
  title = {{Graph neural network-based cell switching for energy optimization in ultra-dense heterogeneous networks}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-25800-3},
}
Figure