According to the International Energy Agency, data transmission networks consumed 260-360 terawatt-hours (TWh) in 2022, making up 1-1.5% of the world's total electricity usage. Looking ahead, there's an expected 160% increase in the energy consumption of Wireless networks up to 2030, despite 5G being more energy-efficient per unit of traffic compared to 4G. This rise is driven by the growing demands for data traffic, prompting Communications Service Providers (CSPs) to expand their networks' capacity through dense deployments, inevitably leading to higher energy consumption.
In response, the telecommunications industry is actively pursuing sustainability initiatives. These efforts include but not limited to enhancing the energy efficiency of network components, forming regulatory frameworks, and exploring renewable energy sources. Wireless backhaul is considered an environmentally friendly option among 5G backhauling technologies (e.g. fiber).
This article discusses the challenges posed by increasing energy consumption in this network segment and emphasizes the benefits of facilitating Artificial Intelligence (AI) to increase energy savings for sustainable growth.
In order to meet modern Wireless Backhaul demands including investment in infrastructure and enhance customer experience at the same time, CSPs will be struggling to balance rising OPEX and particularly energy consumption. The modern trend for Wireless Backhaul includes multi-band transport links, most often combining E-Βand with MW ones. The CSPs are trying to balance at any time three key factors:
Energy consumption optimization for Wireless Backhaul poses challenges with respect to the complexity of establishing mechanisms for intelligent management, uninterrupted deep monitoring and continuous adaptation of network resources that satisfy both traffic demand and power saving.
CSPs should facilitate tools that serve their energy efficiency strategies, with continuous data analytics, KPI forecasting, and early events detection to autonomously make sophisticated decisions based on continuous learning.
Intracom Telecom’s product portfolio has been adapted to satisfy CSPs energy efficiency strategies. The Wireless Backhaul products have been enhanced with Power Save mode functionality and uni|MS™ platform offering Network Lifecycle Management & Automation is under continuous evolution to orchestrate energy consumption optimization while peak traffic is being served.
Extending the NMS, Planning and SDN existing feature set with the brilliance of Artificial Intelligence (AI), to optimize Dual-Band radios energy consumption by up to 30% per physical link, it revolutionizes your Network Operations through a single User Interface.
The Power Save mode of the network elements allows the device to enter a low-power state, shutting down the radio transceiver and modem while still maintaining manageability through interfaces.
Similarly, link adaptation can enter an energy savings mode by limiting the selection of available modulations to the less power-intensive ones ensuring that the throughput demand is met.
The prevailing automation implementations available in the market are based on programmable criteria for activation of the Power Save mode. This rule-based approach is using as input factors the peak traffic rate and bandwidth utilization or a fixed off-peak hour time window in which Power Save mode is activated. However, this way, is not fit for every CSP and does not provide care for the dynamic traffic variations.
The uni|MS™ Energy Consumption Optimization App revolutionizes existing implementations with the introduction of an AI-powered traffic predictor used for the Power Save mode decision making. It assures prediction accuracy of future bandwidth requirements by analyzing link availability and characteristics (e.g. modulation) per direction, both downlink and uplink, in real time. This in turn reduces the probability of erroneous sleep mode decisions and triggers available fallback mechanisms. Latest LAB implementation is based on a Long Short-Term Memory (LSTM) model; a type of Recurrent Neural Network (RNN) known for its ability to effectively manage multivariate time series (i.e., simultaneously observe multiple variables), capture long-term dependencies and understand complex dynamics in sequential data, leading to the following benefits.
The uni|MS™ App exploits the network devices’ Power Save mode and orchestrates energy consumption on a network-wide scale, by intelligently controlling resources utilization and retaining network performance efficiently. Energy consumption can be decreased by up to 30% per physical link in energy consumption Power Save mode.
The typical configuration scenarios presented below are indicative but non-exhaustive in terms of available options.
In this scenario one of the two physical UltraLink™-GX80 (E-Band) links form a single logical link out of the two parallel XPIC links providing higher reliability (one physical link acts as protection to the other link in the RLA) and higher link availability for a particular capacity or double the capacity for a particular channel size.
In a network with 160 such links and for the total of off-peak hours, a CSP can save up to 10 MWh per year.
In this scenario with a fully-outdoor Dual-Band configuration including one UltraLink™-GX80 (E-Band) and two OmniBAS™-BX (MW) radios, implementing Radio Link Aggregation (RLA)-based link bonding.
Dual-Band aggregation can be deployed in greenfield installations to enable the use of E-Band in extended-range links or can be provided as an upgrade to existing microwave links to enable the multiplication of their capacity.
Like scenario 1, more than a single physical link could be opportunistically enabled or disabled based on the current and predicted link requirements. Both MW devices can be set in Power Save mode more than 95% of the time (when the E-Band link is available). In a network with 320 such links a CSP can save up to 125 MWh per year.
In deployments where Dual-Band links are used for fiber connection backup, the saving will be further increased because even the E-Band link can run, most of the time, in low modulation.
Conserving power without compromising network reliability
Reduced energy costs and improved network profitability and sustainability
Balanced energy consumption while maintaining peak network performance
Follow energy efficiency standards/regulations set by governmental authorities
Visibility into energy utilization and predictive analytics to identify opportunities for additional power savings
Future development plans will be targeting to improve the prediction accuracy and proactive capabilities of the App by considering more factors such weather forecast, energy consumption and anomaly detection mechanism.
Detecting deviations from the normal traffic conditions may indicate a sharp increase of the network traffic in the near future, ingesting upcoming weather data might reveal coming fading conditions and energy consumption analytics could indicate relevant usage patterns or any other unexpected events that could proactively end the sleep mode of a backhaul link before the scheduled time in order to effectively serve the unforeseen situations that need to be handled on time.
To support the development, deployment, execution and maintenance of related AI-powered functionalities, Intracom Telecom is working towards the implementation of an architecture for AI Lifecycle Management. Any AI-based use case is developed in our Big Data lab leveraging a High-Performance Computing (HPC) cluster and Graphical Processing Units (GPU) nodes to facilitate the parallel processing of large datasets and accelerate the training of complex neural networks. To ensure the highest possible generalization, model training and validation relies on a diverse dataset comprising of historical data collected from live networks as well as training samples from a network simulator or Digital Twin, whenever applicable. Collecting data from different sources and conditions will eventually make the model more robust to variations, corner cases, etc. and eventually perform well on unseen data when deployed in the real world.
The offline trained AI model is then deployed to the AI orchestrator which resides closer to the network and is responsible for inference, continuous monitoring and local retraining. Live data coming from the network are used as input for executing model inference and generate a respective output (e.g., prediction, classification, etc.) which in turn is mapped to an informed action that is signaled back to the network (i.e., activate or deactivate the MW device). Given that hardware constraints can be met, inference may even be placed in the device enabling faster execution and decision making. The outcome of the AI decisions is continuously monitored, and local retraining is triggered when performance degradation is detected requiring model adaptation to the varying network conditions. Another option of model adjustment would be through a software update; nonetheless, in this case the existing AI model is replaced completely by a newer (and enhanced) version developed in the Big Data lab.