Building Low-Carbon Networks with AI-Based Power Saving

As mobile networks enter the 5G era, new technologies, services and applications continue to emerge. With revolutionary advantages over 4G networks in key performance indicators such as transmission rate, delay and connection, 5G networks can support more abundant services and applications, but also bring the challenges of growing Capex and Opex to mobile operators. A report from GSMA forecasts that deploying 5G networks in a given scenario will consume at least 140% more electricity than 4G networks. In a typical network operation, wireless sites consume about 45% of the total network power consumption, among which BBUs and RRUs/AAUs consume about half or more of the power consumption. Therefore, reducing power consumption of base stations is crucial to network energy saving.
Traditional power-saving methods require manual analysis of massive data, including common parameter data, network inventory, feature adaptation, site co-coverage, and multi-band multi-mode network identification. During the implementation, strategies of power saving in a specified area will be a unified manual operation. Because there is no differential setting of the parameters, it is impossible to automatically match different network scenarios or site traffic. In some sites during busy hours, this may cause damage to services and affect network performance, while other sites cannot maximize power saving in idle hours.
With the help of AI and big data technologies, ZTE's AI-based power saving solution implements coordinated power saving in different scenarios, different sites, different time, and different systems of networks. While ensuring network KPIs, the solution can maximize power saving and achieve the best balance between power consumption and network performance.
In the AI-based power saving solution, traffic load prediction, strategy adjustment and optimization, real-time KPI and performance monitoring will form a closed loop, which includes initial parameter self-configuration, time window self-adjustment, and threshold self-optimization (Fig. 1). Through the precise insight into network and user behaviors, the solution can intelligently identify real user usage, thus reducing power consumption.

Initial Parameter Self-Configuration
The initial power saving strategy can be self-configured according to characteristics of base stations and the relationship between power saving functions. Through big data analysis, the system automatically sorts out mainstream network scenarios, and analyzes the power-saving scenarios according to the historical traffic model and base station configuration. According to user behavior habits, site hardware equipment and power-saving function constraints, the power-saving effect is estimated and the initial power saving strategy is made to realize self-configuration. The suitable power saving strategy including a relatively proper initial threshold and a time period for executable energy saving will be enabled for the sites that are expected to have power saving effects.

Time Window Self-Adjustment
Based on historical traffic data, three types of cells are distinguished: positive cells, negative cells and invalid cells. The intra-week sub-sequence split prediction is combined with the impact of holiday factors and network traffic load trend on prediction. After testing the well-known time series prediction algorithm like linear regressive, ARIMA and long-short term memory, the second-order exponential smoothing algorithm is chosen to get the prediction model with the best computational performance and optimization effect.  
The result of commercial use case shows that the prediction accuracy of uplink/downlink PRB ratio and RRC connected users exceeds 90%. The prediction value matches well with the actual value in normal scenarios, which effectively increases the efficiency of power saving in the specified period.

Threshold Self-Optimization
To find the balance between power saving and network performance and maximize the power saving effect, the precise cell-level scenario-based triggering and parameter setting replaces the traditional inflexible network-level parameter setting. Clustering algorithm will be used to find out the best step of power saving threshold adjustment through 96 groups of KPIs including setup success rate, call drop rate, handover and user experience per cell every day. After the AI self-learning, the baseline of KPI will be updated daily, and the threshold will be adjusted according to the step. In case the KPI baseline is exceeded, the threshold will roll back in real-time. 
The convergence between power saving effect and network performance will achieved within one week after commercial application is testified. 
Since mid-2019, ZTE's AI-based power-saving solution has been widely deployed for commercial use by operators in many places of China, including Shandong, Chongqing, Sichuan, Fujian, Hunan and Liaoning, with a cumulative application scale of more than 100,000 cells. Its commercial networks in overseas countries such as Malaysia, South Africa and Italy are also deployed and verified. As testified in commercial use, the AI-based power saving solution can reduce power consumption of the base station by 15 to 20%. In 5G networks, 6 hours of deep sleep combined with 18 hours of symbol shutdown can save 30% of power. In other words, every 1000 sites can save 1.5 to 2 million kilowatt-hours a year, equivalent to about 1.5 to 2 million yuan or 1100 to 1500 tons of carbon emissions.
The intelligence of the whole network is difficult to be achieved overnight and needs a long-term development. However, with the continuous accumulation of commercial network data, the machine-learning algorithm will gradually improve the AI-based power-saving solution. The AI algorithm itself will evolve iteratively to achieve higher efficiency and more accurate strategies to adapt to changes in network topology and traffic models. 
The following suggestions can be considered for improvement:
—Feed back to vendors to design more intelligent devices and make them more adaptive.
—Make consumer products such as mobile phones and mobile phone software more intelligent in the use of power and signal strength and work in synergy with the network.
—Make the design of vendor equipment more integrated with AI to reduce power usage.

The development of IoT and enterprise networks is a big issue at present. More carrier-grade solutions are deployed in enterprises, and the demand for services in different vertical industries is growing. The power-saving requirements in these scenarios will be unique and more complex than those in cellular networks for individuals.