January 17, 2024
Blare Technologies recently implemented an R&D project co-financed by European Funds:
An intelligent system for analyzing and optimizing 5G network plans using neural networks (deep machine learning) and expert algorithms.
The project aimed to develop an IT solution for professional analysis and optimization of 5G network plans. The result of the project is an intelligent system using AI methods, including neural networks (deep machine learning) and expert algorithms.
To say that planning a 5G network for FR2 (above 24 GHz) is a complex task is an understatement. Due to the specificity of millimeter waves and the challenges associated with planning 5G networks, this task requires considering a number of additional factors and a large number of possible planning scenarios on a scale incomparably larger than in the case of previous generations of networks.
While in the case of older networks, it was enough, to put it very simply, to find a tower or a tall building on which a mast with antennas could be installed, in the case of 5G NR networks, it may be necessary to install several dozen antennas (small cell) to cover the same area. Moreover, due to the characteristics of these antennas and the susceptibility of mm waves to all kinds of obstacles, it will be necessary to consider factors previously of marginal importance. Such antennas must be installed much closer to the receiver (user) and, therefore, at a much lower height than before, where their signal can be blocked by nearby trees, infrastructure elements, or even passing vehicles, e.g., trucks. In this situation, it is necessary to consider a large number of potential site candidates in the network spatial planning process. Then, choose a combination of site candidates that will provide the desired coverage parameters for a given area while maintaining the optimal number of antennas and base stations, usually as few as possible.
In the above planning scenario, we defined 321 site candidates for an urban space of 500 m by 500 m, and our AI solution designated 35 sector antennas to cover the given area (main communication routes). In this system, the number of possible combinations was, therefore, astronomical 7.55×10^46! It would take several hundred hours to consider such a massive number of potential scenarios using traditional computational methods. In fact, the task is even more complex, as discussed later in the article. However, using neural networks, AI, and ML made it possible to achieve an optimal result in less than 3 hours, which is more than a hundred times faster!
A prerequisite for the 5G network planning process, and its first stage, is the precise reconstruction of the planning area, e.g., a city, in the form of a 3D digital spatial model. Such modeling can be done using available open-source data for a given area or specific data for a given project. In case of investments planned for implementation, it is necessary to use design data, usually CAD or BIM. Our solution makes both of these scenarios possible.
The digital 3D model should reflect the planning area as precisely as possible, considering not only buildings and the digital terrain model (DTM) but also vegetation, infrastructure elements, and (optionally) dynamic elements such as vehicles.
The next key step is defining the coverage parameters required from the 5G radio network in a given planning area and designing potential site candidates for antennas and base stations. Defining the planning area in our solution can be done in several ways:
1. by indicating a set of specific points for which defined radio parameters are to be provided
2. drawing a linear set of requirements, e.g., for communication routes, roads, etc.
3. drawing a surface area on which radio parameters are to be ensured, e.g., for typical urban or industrial scenarios
4. it is also possible to indicate no-go zones that are to be excluded from signal coverage due to specific local regulations or requirements
Then, the requirements for the desired parameters of radio signal coverage will vary depending on the planning scenario and the technology being implemented. For example, they will be different in the design of the network for autonomous driving (C-V2X) and the university campus network.
Site candidates can also be appointed by providing a ready-made list of mounting points (e.g., poles dedicated to radio equipment), or the application can automatically, with AI support, indicate potential mounting points on roofs or facades of buildings, lighting poles, etc.
With the planning environment prepared in this way, the application proceeds to determine the optimal spatial configuration of the 5G network using a number of AI / ML methods. In our case, the development of those methods proceeded in successive phases.
In the first phase, only a genetic algorithm was used, which uses a number of metrics, such as signal range, signal emission angle, and location of base stations, to optimize the distribution of base stations in a defined urban area. Based on those results, subsequent site candidates from all possible points are accepted or rejected. The figure below shows the initial situation, including a set of potential mounting points and mounting directions (azimuth) of antennas:
A genetic algorithm evaluates whether a given point and azimuth of the antenna meet the predefined requirements. The next picture shows the setting that is qualified as correct:
On the other hand, the following setting was discarded because too many beams were hitting adjacent buildings:
The operation of the genetic algorithm in the first phase made it possible to select a set of settings (site candidate + antenna azimuth) with the greatest potential to contain the optimal set among them. However, it took many hours, which we considered a result that did not meet our expectations.
At the same time, it should be noted that in the case of conventional computation methods, the above task would be practically impossible to solve due to the increasing computational difficulty as the number of parameters increases. This problem belongs to the NP-hard class, which means that even a small increase in the number of initial parameters can cause an exponential increase in computational time, with the number of potential combinations being ~2.91×10^290.
The result of the operation of the genetic algorithm in the first phase of development is shown in the below picture:
In the second development phase, we divided the planning process into modules. The first module was designed to filter out suboptimal antenna orientations by applying linear programming algorithms. The second used a genetic algorithm to initially limit the set of possible base station locations by simultaneously analyzing the range for all antennas. An example of the operation of these modules is shown in the picture:
In the third phase, we implemented a hybrid model that combines a genetic algorithm with neural networks that act as penalty functions. These networks were previously trained on datasets manually classified as correct base station and sector configurations in various planning scenarios.
The results obtained in the subsequent phases of the development are presented in the table below:
As indicated by the above data, using the hybrid model significantly increased the speed and precision of the optimization process, and the results confirmed that this approach effectively solves previous computational challenges. Although the time it takes to perform a single assessment in the genetic algorithm has increased, the overall precision and efficiency of finding the optimal solution have increased, ultimately increasing the accuracy and efficiency of 5G network planning.
Over the course of the R&D project, several commercial implementations were carried out, which provided valuable data for real-world validation. The most important implementations concerned the development of 5G network plans for selected sections of highways in the US. During these implementations, detailed plans were created that defined the parameters and locations of base stations, mainly on existing lighting poles and other dedicated structures. The entire planning was conducted in a detailed simulation environment, including multi-level road junctions. The AI-supported planning process was much faster, and the generated 5G network plans proved to be more efficient regarding the number of base stations and antennas planned (especially around road junctions) than if they had been developed by traditional methods.
The picture below shows the 5G network plan generated for the 38 km section of the highway:
You can learn more about our R&D project, An intelligent system for analyzing and optimizing 5G network plans using neural networks (deep machine learning) and expert algorithms, in this website section.
© 2022 Blare Technologies. All rights reserved. Cookie Policy