Why AI-Powered RAN Is an Energy Efficiency Breakthrough
By Pushpendra Sharma, Senior Al Technical Product Manager, Manguluri Bhaskar Senior Data Scientist, Will Ng Head of AI & Analytics Portfolio, Anupama Muraleedharan Senior Manager - Data Science
Originally published by Ericsson
- The ever-increasing demand for data combined with a need to reduce energy consumption to reach Net Zero presents new challenges for network operators.
- Ericsson experts have found that efficient monitoring and tracking supported by revolutionary AI can identify optimization areas to reduce RAN energy use by up to 12 percent annually.
- Below, we walk you through our ML-based recommendation engine that generates energy-efficient configuration settings for network nodes.
At present cellular operators are seeking to reduce an ever-increasing energy bill which accounts for 5-7 percent of their operating expenditure. At the same time, in the coming years with more countries moving to 5G, like enhanced Mobile Broadband (eMBB), ultra-reliable low latency communications (URLLC) and massive machine type communications (mMTC) may require more sites, in turn negating some of the gains obtained via energy efficient parameter configuration. In addition, any measure that operators take in terms of reduced energy consumption should not impact the network performance or any of the service level agreements for these.
Energy efficiency at the RAN
Each LTE and 5G radio transmitter consumes electric power to control and transmit shared channels for user data and broadcast channels for control information. Shared channels consume slightly less electric power when user traffic is low, but broadcast channels and control circuits always consume constant power. There are means to disable or shut down cells and wake them up when required according to pre-configured scheduling parameters, but they do not consider or adapt to the subscriber experience impact of disabling these entities. AI/ML mechanisms are required to dynamically learn, adapt and act upon the best balance of power savings and subscriber experience at each cell site and radio.
Operators are already adjusting radio network requirements (such as added layers of spectrum, cell density and associated energy costs) to keep up with peak capacity demands. But much of this incremental investment is under-utilized and inefficient when demand is below peak. There is a need to reconfigure and control cells dynamically, to serve dynamic traffic patterns, not just peak traffic.
The sites and cells deployed in a network may be classified as underlaid (coverage) and overlaid (capacity). When an overlaid capacity cell is turned off, the traffic load existing in the overlaid cell is offloaded to underlaid (coverage) cells. The underlaid (coverage) cell monitors traffic conditions and key performance measures like access to the network, service quality, retainability and mobility to allow the overlaid (capacity) cell to go into sleep mode and turn on the sleeping cells when required. Once overlaid cells are turned off for energy saving reasons, the reference signal interference in the network is also reduced, improving UE throughput, and reducing operating expenditure. This phenomenon of sleeping cells with generic and static configured parameter thresholds may result in coverage loss or not using the spectrum efficiently with no optimum energy saving. This capability of cell sleep and wake up detection should be adaptive to network traffic conditions, radio resources availability, user density, service usage, user experience and overall network performance to provide best energy efficiency.
One size does not fit all
It is a known fact that the radio access network (RAN) accounts for nearly about 80- 85 percent of overall energy consumption. Depending on the geographical location and varying data traffic loads, it would be wise to put some of the capacity cells into sleep and wake them based on the traffic demand. The image below shows a cluster of cells with different energy consumption in accordance with user and network activity and depicts how not all cells need the same energy to meet the traffic demand.
This calls for having a customized approach for each capacity cell to be in a sleep or awake state. An ML/AI based approach expands the potential for such energy-saving opportunities across the network at cell level. One of the solutions is enabling dynamic thresholds configuration for cell sleep mode for coverage & capacity.
AI based dynamic thresholds for cell sleep mode
Enabling and controlling the cell sleep mode based on the physical resource block (PRB) utilization and RRC connection-based thresholds, without impacting the customer experience, needs a careful monitoring of key performance scenarios like network availability, reliability, traffic pattern, services offered and spectrum usage, while considering same of neighboring cells as well. To achieve this, it is important to determine the utilization of each cell layer for the next few days and, most importantly, to determine the impact on customer experience for the same period.
One example is a renowned CSP that was exploring avenues to optimize RAN energy consumption without degrading network performance and customer experience. In the current scenario Ericsson provides static thresholds to be set manually through the cell sleep mode feature, however, the team proposed setting thresholds dynamically. To determine the values for this threshold dynamically, it was necessary to conduct field experiments on a live network as there was no variation in data due to static nature in CSM.
The image above portrays the following:
1. High Level Solution: Energy consumption forecasting model
2. Methodology: Optimal configuration threshold grid search model
3. Model Fitting: Threshold validation on live network
4. Optimised Scenario: Impact analysis
The forecasting model predicts energy consumption levels per cell for a day in advance. It provides an idea for the optimization model for possible improvements in terms of various performance KPIs like accessibility (QCI9, QCI5 & QCI1), retainability, mobility, latency, throughput and traffic volume.
The optimization model determines the dynamic threshold at cell level at which a capacity cell should be awakened or put to sleep based on RRC Connections and PRB Utilization. In this case, it was considered as a convex optimization problem, with the objective function to maximize the sleep hours subjected to various constraints like business constraints (for example user experience or operational expenditure) or technical constraints (such as accessibility during coverage hole detection, drop call rate, average sector loading, location estimation, mobility, and lower latency handling in case of URLLC use cases).
Impact of the solution
- Target exceeded: 10-12 percent energy reduction across pilot sites
- No degradation in RRC, ERAB Success Rate and Call Drop Rate across all bands
- Stable traffic volume, mobility success rate and latency
- Stable DL/UL Throughput and other primary KPIs against historical trend
Our research shows a few more areas that operators can focus on reducing energy consumption
- MIMO Sleep: ML-enabled MIMO path and radio head control for energy savings using optimal MIMO configuration resulting in optimal power efficiency and performance balance for various traffic conditions. This learns power savings and network performance under various traffic loads, then activates and deactivates MIMO paths according to trained models.
- IoT based energy optimization: Smart Homes, Smart Factories & equipments, Smart Health monitoring, shipping & Logistics etc., are utilizing various interconnected devices to autonomously manage intent operations, which consumes a lot of energy, resulting in a need for energy optimization. A function of IoT devices is to reliably collect and share the perceived data with the physical world. The hardware element of the IoT device consists of a battery-powered sensor, an actuator, and a communication system. IoT sensors, cloud computing technologies and the telco network with the help of AI based techniques (events and data driven), improve productivity and energy efficiency.
- Smart green sourcing: Green Sourcing refers to the purchase of goods and services that cause minimal adverse environmental impact. The demand for recyclable products, energy-efficient systems, and clean technology and fuels is driving the adoption of ecologically responsible business norms. In green sourcing, concerns about environmental impact are given weight over other business decisions to reduce pollution. Sunsetting legacy systems and equipment is key in bringing sustainable smart green sourcing.
- Machine learning algorithms for HetNet traffic pattern: Energy-aware platforms analyze events through AI/ML and reinforcement learning (RL), to allow proactive energy savings coupled with reduced CO2 emissions in Heterogeneous Network (HetNet) architecture. Advanced frameworks for optimally handling the switching on or off of sleeping cells, in case of low latency services without impacting the QoS, take mobility prediction, UE location estimation and network environment into consideration.
The dynamic threshold for cell sleep mode is a unique solution that can be adapted by any operator. This solution can be used for optimal gains in energy consumption across the network. Making use of the huge volumes of data pertaining to attempts, handover and real-time demand, this ML-based approach enables highly efficient, fast, and automated decisions on the RAN components that can be put into sleep mode, thereby saving energy. The estimated energy-saving using this AI solution is up to 10-12% for some operators, in addition to any other savings that are gained from site-level efficiency measure.
How AI can reduce network energy costs
How to scale 5G to meet increasing data demands while still addressing energy concerns