Aidons notre planète en baissant le thermostat de la 5G sans gripper le réseau

Le secteur des télécommunications est responsable d'environ 2 % des émissions mondiales de carbone, et les réseaux d'accès radio - appelés aussi RAN - représentent environ 80 % de la consommation d'énergie de l'ensemble des réseaux mobiles. Selon la GSM Association - une association internationale qui représente les intérêts des opérateurs et constructeurs d'équipements de téléphonie mobile - la consommation d'énergie représente également 20 à 40 % des dépenses d'exploitation des réseaux.

Dans ce Lightning Talk, je montrerai comment l'intelligence artificielle peut être utilisée pour relever ces défis d'optimisation des réseaux et de réduction de son impact sur notre planète. Je présenterai une solution qui prévoit la demande de trafic des utilisateurs avec une grande précision et qui contrôle le nombre de ressources RAN actives à tout moment, réduisant ainsi la consommation d'énergie de l'infrastructure jusqu'à 60 %, sans impact sur la qualité du service.

Enfin, j'expliquerai comment cette solution peut être intégrée au paradigme émergent des réseaux mobiles OpenRAN et je démontrerai ses performances en action, en considérant un déploiement urbain de plus de 200 stations de base.

Transcript

1. Aidons notre planète en baissant le thermostat de la 5G sans gripper le réseau Alexis Duque
2. Problem Growing energy consumption, spiraling costs, significant climate impact ▪ Telecoms responsible for 2% of global carbon emissions [1] ▪ Exponential growth in data consumption means deployment densification ▪ Energy prices not expected to fall below pre-2022 levels ▪ Energy spending increase outpacing sales growth by >50% ▪ Pressure to become net zero ▪ Radio access networks (RANs) account for ~80% of that [2] [1] https://www.wri.org/insights/4-charts-explain-greenhouse-gas-emissions-countries-and-sectors [2] https://www.gsmaintelligence.com/research/going-green-measuring-the-energy-efficiency-of-mobile- networks-fourth-edition
3. How the problem is addressed today? Existing solutions save energy at the price of service quality Solution #1 switches off capacity layers overnight outside typical working hours ▪ Safe approach, easy to be scheduled, preserves good service availability, BUT ▪ Enables only 15% energy savings
4. How the problem is addressed today? Existing solutions save energy at the price of service quality Solution #2 trains a ML model to learn demands on specific time of day/day of the week, which resembles statistical averaging ▪ Approach may sustain >55% reduction of energy consumption, BUT ▪ At the cost of 17% under-provisioning rate (service degradation, potential SLA breach penalties)
5. Limitations ML/AI models designed for other domains not directly applicable to networking City-scale mobile traffic forecasting ▪ Deployment with >200 cells ▪ Aggregate traffic volume sampled every 5 minutes ▪ Predicting traffic demand for the next 30 mins (6 steps) to 3 hours (36 steps)
6. Limitations ML/AI models designed for other domains not directly applicable to networking [1] C. Zhang, M. Fiore, I. Murray, and P. Patras, "CloudLSTM: A Recurrent Neural Model for Spatiotemporal Point-cloud Stream Forecasting", AAAI 2021. City-scale mobile traffic forecasting 50% reduction in average Mean Absolute Error (MAE)
7. Limitations ML/AI models designed for other domains not directly applicable to networking [2] C. Zhang, M. Fiore, and P. Patras, "Multi-Service Mobile Traffic Forecasting via Convolutional Long Short-Term Memories", IEEE M&N 2019. City-scale mobile traffic forecasting 390x fewer parameters ▪ 3 Models: • LSTM • SoothsAIer (CloudLSTM- based[2]) • ForesAIght (GNN-based, proprietary) ▪ Performance loss wrt. best model for each antenna ForesAIght 390x fewer parameters
8. Limitations ML/AI models designed for other domains not directly applicable to networking ▪ Training with typical loss functions (MAE, MSE, etc.) may not be appropriate ▪ Not all mistakes are equal - we want to reduce the number of underestimation errors ▪ That translates into fewer under-provisioning events ForesAIght: AI-assisted capacity- and uncertainty-aware forecasting
9. Closing the Loop EnergAIze: Forecast-driven energy efficiency CLOUD-NATIVE APP Telco Cloud Forecasting Engine RAN Data Collection for the RAN and the 5G Core 1 Energy Optimisation Engine Traffic demand for next 15’ 2 Compute RAN Configuration for the next 15’ per Antenna 3
10. Closing the Loop EnergAIze: Forecast-driven energy efficiency CLOUD- NATIVE APP ▪ >35% average reduction while maintaining 99.8% accessibility ▪ Integration and deployment with Red Hat OpenShift ▪ Suitable for deployment in any type of mobile/fixed network
11. Closing the Loop EnergAIze: Forecast-driven energy efficiency CLOUD- NATIVE APP It takes 43 minutes to train EnergAIze on: ▪ Dataset of 2.04 GB that contains ~3 months of data ▪ For a city scale of deployment with 729 antennas ▪ OVH VM type T1-LE-45 with characteristics: • GPU: Tesla V100 • CPU: 8 cores (3 GHz) • RAM: 45 GB Near real-time inference on CPU (sub-second) < 0.29 kWh GPT-3: 1 300 MWh
12. What is coming next? Toward an ai-native mobile network ▪ EU funded project ▪ Design a new framework for 6G ▪ Network Intelligence ▪ Sustainability ▪ Energy Efficiency ▪ Adaptability https://sns-origami.eu
13. THANKS! Questions? Alexis Duque alexis@netai.tech @alexis0duque.bsky.social https://netai.tech