 |
|||||||||||||||||
|
|
|||||||||||||||||
        |
Regular Articles Initiatives for the Wireless Base Station Optimization AI to Balance Radio Resource Efficiency and
|
||||||||||||||||
 |
 |
1. The purpose of wireless base station optimization in mobile networks
Mobile terminals, such as smartphones, are indispensable in modern life, and mobile network services now function as critical infrastructure on par with electricity, water, and gas. Because it is not feasible to endlessly expand communication resources, such as equipment and transmission paths, in response to the ever-increasing data traffic, it is essential to maximize the efficiency of available resources. Since mobile networks are constrained by limited frequency bands available for communication, optimizing base station utilization is especially critical.
Each base station is equipped with multiple antennas that handle different frequency bands, and networks are constructed by spatially deploying multiple base stations while configuring parameters such as antenna orientation. In building these coverage areas, however, a wide range of factors must be considered. For example, to prevent the occurrence of areas where wireless communication is not possible (coverage holes), it is necessary to adjust the coverage areas of each base station so that there are no gaps. In areas where coverage overlaps, however, radio interference between base stations degrades communication quality, so overlap must be avoided as much as possible.
When constructing the coverage area, it is also necessary to take into account imbalances in wireless resource usage. If connections from mobile terminals are concentrated on a particular base station, the wireless resources of that station alone will be insufficient to handle the communication traffic. Therefore, depending on external conditions and the status of each base station, it is necessary to distribute mobile terminal connections across multiple stations and smooth out the utilization of wireless resources.
Such imbalances in wireless resource usage have generally been caused by spatial shifts in the distribution of mobile terminals, such as those resulting from events or changes in buildings. However, base station power control has emerged as a new contributing factor. To reduce power consumption, base stations use sleep control, whereby stations with low wireless resource utilization are temporarily shut down. When this occurs, mobile terminals that had been connected to the sleeping station reconnect to surrounding base stations, leading to imbalances in wireless resource usage. While mobile networks continue to advance, for example, with the transition to fifth-generation mobile communication systems (5G), the increase in power consumption has become a pressing issue. Since base stations account for more than half the total network power consumption, it is essential to consider smoothing wireless resource utilization with power control as a given condition.
Adjusting coverage areas and equalizing wireless resource usage rates are achieved by controlling multiple base station parameters, such as antenna tilt angles and sleep functions. It is thus crucial to continuously evaluate and optimize these base station parameters.
2. Challenges in base station optimization and global trends
Optimizing mobile base stations is inherently complex because changing the parameters of one station not only affects its coverage area but also impacts the coverage areas of neighboring stations. It is thus necessary to derive optimal parameters across multiple base stations simultaneously. This process is mainly carried out by network operators, who determine base station parameters for live environments through trial and error, drawing on past experience and simulator-based evaluations. Since parameter derivation through such trial and error requires considerable time, and the range of areas that can be optimized remains limited, there is growing anticipation for the automation of parameter optimization using artificial intelligence (AI) and other advanced methods.
In terms of global trends, one notable initiative is the O-RAN Alliance, a standardization body working toward the intelligentization of radio access networks (RANs), including base stations. The alliance is standardizing architectures and control interfaces to enable AI-driven operation and control. Within the RAN architecture being standardized by the O-RAN Alliance, discussions are underway regarding AI-based control procedures in key components such as Service Management and Orchestration, which handles monitoring, maintenance, and orchestration, and the near-real-time RAN Intelligent Controller, which manages fast control cycles of less than one second. These efforts suggest that AI-based optimization of base station parameters will accelerate in the coming years.
3. Initiatives in AI-based wireless base station optimization
At NTT Network Service Systems Laboratories, we are developing the Wireless Base Station Optimization AI, which aims to automate the derivation of base station parameters using AI, thus expanding the areas where optimization can be applied. In this article, we introduce our efforts focusing on antenna tilt angle and antenna sleep functionality (Fig. 1).
Fig. 1. Initiatives in the Wireless Base Station Optimization AI.
3.1 Optimization of antenna tilt angle
The antenna tilt angle refers to the downtilt angle of each antenna attached to a base station. Reducing the tilt angle points the antenna upward, expanding the coverage area, while increasing the tilt angle directs the antenna downward, shrinking the coverage area. Because of this direct impact on coverage, the tilt angle is a critical parameter in base station optimization. To calculate the optimal tilt angle, two key conditions must be satisfied. The first is to ensure a uniform level of signal strength across the area to avoid the occurrence of coverage holes while minimizing overlap between the coverage areas of neighboring base stations to reduce interference and prevent deterioration in wireless quality. The second condition is to prevent mobile terminal connections from becoming concentrated on a particular base station, thus smoothing out the utilization rate of radio resources across multiple stations. This enables communication traffic demand to be distributed and processed more evenly, which in turn leads to improvements in average throughput.
To address these requirements, NTT Network Service Systems Laboratories has developed the Base Station Tilt Angle Calculation AI [1], which automatically derives tilt angles that maximize throughput in the target area. Since the number of tilt angles to be considered is on the order of several hundred, we built a framework using particle swarm optimization (PSO)*1, which enables tilt angle values to be derived with relatively low computational cost (Fig. 2).
Fig. 2. Overview of the Base Station Tilt Angle Calculation AI.
This AI takes as input specifications such as the base station’s location, height, antenna orientation, and current tilt angle, along with traffic information collected at the base station. It then outputs the optimal combination of tilt angles that maximizes average throughput in the target area. PSO is a mathematical optimization method categorized as “black-box optimization.” Our framework incorporates a simulation environment that estimates throughput for various combinations of tilt angles across multiple base stations. To accurately derive appropriate tilt angles, it is crucial to minimize the gap between the simulation environment and real-world conditions. Therefore, the simulation environment integrates a model that estimates the distribution of mobile terminals within the target area, as well as a model that predicts throughput on the basis of the number of terminals connected to each base station.
3.1.1 Mobile terminal distribution estimation model
To derive tilt angles that avoid having mobile terminals concentrate on specific base stations and instead smooth out the utilization of radio resources, it is necessary to have information on how mobile terminals are spatially distributed. We developed a model that estimates the number of mobile terminals on a regional mesh*2 basis, using connection information observed at each base station (Fig. 3). The coverage area of each antenna is first expressed on a mesh basis using base station specification data. Among multiple predefined spatial characteristic models, the system then automatically searches for the optimal model and its parameters such that the difference between the estimated and observed number of mobile terminals falls below a specified threshold.
Fig. 3. Mobile terminal distribution estimation model.
This model is an extension of block kriging*3, a type of spatial interpolation method, and achieves high estimation accuracy even for meshes where multiple base station coverage areas overlap, an area in which conventional block kriging has difficulty.
3.1.2 Throughput estimation model
Since variations in the number of mobile terminals connected to a base station directly affect throughput, it is possible to estimate the average throughput of a base station for a given number of terminals by training a model on the relationship between terminal count and throughput, using data observable at the base station. Because each base station differs in terms of location conditions and equipment configurations such as radio frequency, it is necessary to train a separate estimation model for each station. However, the number of available observation data points also varies by base station, and for those with limited data, insufficient training leads to decreased estimation accuracy.
To address this, we established a model that learns both overall trends (including surrounding base stations) and individual base station trends and adjusts the weight given to each according to the number of observation data points. This enables suppression of accuracy degradation in throughput estimation (Fig. 4). With this model, base stations with fewer data points construct models that emphasize overall trends, while those with abundant data emphasize station-specific trends. Throughput can therefore be estimated regardless of the amount of available observation data.
Fig. 4. Throughput estimation model.
3.2 Optimization of antenna sleep function
Each base station is equipped with two types of antennas: coverage-band antennas, which ensure overall area connectivity, and capacity-band antennas, which are intended to increase available radio resources. In sleep control for base stations, it is not the entire base station that is put to sleep but rather the capacity bands, since maintaining area-wide connectivity must always be ensured. For capacity-band sleep control, the key decision factor is whether sufficient throughput can still be maintained after putting an antenna to sleep. Because the surrounding environment differs from one base station to another, however, it is difficult to establish fixed sleep-decision rules individually for each station.
To address this, NTT Network Service Systems Laboratories has developed the Base Station Band Sleep Decision AI, which determines whether capacity-band antennas can be put to sleep on the basis of the utilization rate of each antenna’s radio resources [2]. In current approaches, many AI systems use deep reinforcement learning (DRL)*4 to autonomously learn sleep-decision rules. However, to balance throughput maintenance with power consumption reduction, it is necessary to search for sleep candidates under the constraint that a certain level of throughput must always be guaranteed. Conventional AI models are not able to directly handle such constraint conditions. Our AI extends the DRL-based sleep-decision framework by introducing a safety layer, which transforms the candidate actions derived by the agent into control actions that satisfy required constraints. Thus, capacity bands can be put to sleep while still guaranteeing a specified level of communication quality, reducing power consumption (Fig. 5).
Fig. 5. Overview of the Base Station Band Sleep Decision AI.
The process works as follows. First, a reinforcement learning agent derives sleep candidates that maximize a weighted sum of target power savings and average throughput. Then, the safety layer converts those candidate actions into antenna-control decisions that are both highly similar to the agent’s output and guaranteed not to produce mobile terminals that fall below the required throughput threshold.
By applying this AI-based capacity-band sleep control, it becomes possible to reduce power consumption in the target area while maintaining mobile terminal throughput above the required level.
| *1 | Particle swarm optimization (PSO): A type of swarm intelligence optimization method inspired by the behavior of living organisms. Multiple particles are used in the search process, and each particle explores optimal solutions on the basis of the knowledge gained from its optimization attempts. |
| *2 | Regional mesh: A system that divides a geographic area into a grid of cells of nearly equal size based on latitude and longitude. |
| *3 | Block kriging: A type of spatial interpolation method. It models spatial characteristics from observed values within blocks in the target area and uses both the model and observed values to estimate values at unobserved points. |
| *4 | Deep reinforcement learning (DRL): A type of reinforcement learning that uses neural networks to represent continuous action spaces. |
4. Future outlook
We introduced our initiatives for the Wireless Base Station Optimization AI, which aims to automate the derivation of base station parameters using AI and expand the range of areas where optimization can be applied. With the advancement toward 5G-Advanced and 6G*5, base stations are expected to become increasingly sophisticated. Therefore, the number of parameters that network operators must control will grow, requiring more complex operations. NTT Network Service Systems Laboratories will continue working toward automation of such complex base station parameter control through the application of AI to meet the demands of these increasingly advanced wireless networks.
| *5 | 5G-Advanced/6G: Next-generation mobile communication standards currently being standardized by the 3rd Generation Partnership Project (3GPP). |
References
| [1] | H. Kinsho, K. Takeshita, and K. Yamagishi, “Block Kriging Based Mobile User Distribution Estimation for Antenna Tilt Optimization,” IEEE International Communications Quality and Reliability Workshop 2023 (CQR 2023), Washington, DC, USA, Oct. 2023. https://doi.org/10.1109/CQR59928.2023.10317795 |
|---|---|
| [2] | R. Tagyo, H. Kinsho, A. Shiozu, and K. Yamagishi, “Throughput-constrained Antenna Sleep Management for Saving Power,” 20th International Conference on Network and Service Management (CNSM 2024), Prague, Czech Republic, Oct. 2024. https://doi.org/10.23919/CNSM62983.2024.10814611 |
