In a mobile network, signals are transmitted via radio waves using a global network of receivers and transmitters. Communication between users may concern voice/text/data transmission. The name cellular network is closely related to its structure. It consists of overlapping geographical areas (so-called cells).
The 5G (the fifth generation) is the newest deployed technology standard for broadband cellular networks. Wireless 5G technology with higher speeds, lower latency, and higher availability (than 4G) enables new services with stringent performance requirements, e.g., enhanced mobile broadband, ultra-reliable, low-latency communications, and massive machine-type communications. Network Slicing (NS) is a solution to manage these requirements and to provide demanded Quality of Service (QoS). What exactly is QoS? It is the concept linked with characteristics of a service’s ability to satisfy telecommunication service user needs. To describe QoS quantitatively, we can consider various features, e.g., throughput, transmission delay, packet loss, bit rate, response time, echo, jitter, or availability. NS is essentially a network virtualization technique, which logically divides physical network resources into logical network layers, called slices. Resources can be dedicated to specific slices in order to separate their traffic and/or to guarantee a certain level of QoS and can be shared between slices to increase the efficiency of network utilization.
Nokia Bell Labs estimates that this addition of another level of complexity to network management without automation will increase the total cost of ownership of the network by 30% versus that of the initial physical network [2]. On the other hand, full automation of NS can lead to a 32% cost reduction.
From a mathematical perspective, traffic control in Information and Communications Technology (ICT) networks constitutes an optimization problem in resource allocation. This control problem encompasses various aspects, including flow control and congestion management. Network dimensioning can be approached from two distinct perspectives: long-term dimensioning, which is essential for investment planning, and short-term dimensioning, which focuses on the daily operations of a sliced network. The complexity of the problem arises from the variability of traffic patterns over time. Consequently, the modelling approach must be dynamic to prevent both resource underprovisioning and overestimation. The primary consequence of inefficient resource allocation is underestimation, leading to increased costs associated with violating Service-Level Agreements (SLAs) and a higher churn rate among telecommunication network users. This issue stems from the network’s inability to meet the demands of each network slice. SLAs delineate the expected quality and type of service, as well as the penalties for the operator’s failure to fulfil the contract terms. Conversely, overestimation results in the operator not possessing the necessary resources, thereby leading to inefficient resource management. Developing a digital twin that accurately mirrors the operation of a 5G network could facilitate the automation of processes related to 5G network management, thereby reducing cost, and enhancing profitability.
Following the emergent conceptions, a forecasting and dimensioning framework is proposed which is a data-driven model embodying the Network Digital Twin idea. In fact, designed framework can work independently as a forecast and “what-if” analysis module relevant for (sliced) network dimensioning and traffic engineering or can serve as one of the core components of multifunctional NDT. This framework has been developed in modules to enable easy verification, management, and scaling to specific use cases.
Here we present the results for the Stage 1, i.e., for the Scenario forecasting module. The heart of this module is the forecasting model, which is trained on real traffic and environmental data. This multivariate model forecasts throughput and delay with cell and slice granularity. Several methods have been verified and as a result the most promising have been described in the following sections.
The “what-if” scenario simulation can be performed with the use of such a pre-trained model. According to the scenario requirements, the historical model inputs are modified to estimate slice throughput and delay in the future. Because some of the input features are correlated, any modification of the inputs must consider their relation.
The dataset used in this research consists of hourly averaged time-series from the 5G Base Transceiver Stations (BTS) working in a live network deployment. This data was collected from each BTS and each configured cell within the BTS, from a period of whole month. Subscribers in this network cluster have been divided into four groups with different priority (Slice A-D).
We are focus on forecasting Delay and Throughput (for Slice A-D) using the KPIs that have been selected according to the best knowledge of the telecommunication expert. Thanks to them, it is possible to create multivariate models including traffic load and radio environment metrics that have a direct effect on throughput and delay. All the variables used in modelling with description are available in Tab. 1. Several multivariate methods for traffic forecasting have been evaluated, from which VARMAX and LSTM presented the best fit to the real network time-series. For more details, see [1].
First, we focus on developing the method of modelling Throughput and Delay for each network slice separately. The illustrative results of prediction are shown in Fig. 1 and Fig. 2.
Source: [1].
Source: [1].
However, because there is need to test the impact of changing individual variables on traffic, a general model (G) has been created. It contains information about all slices within a specific cell (for more details, see [1]).
Boxplots with normalized RMSE error corresponding to slice A are shown in Fig. 3 (the results for the remaining slices are available in [1]). It can be seen, that the differences between the general VARMAX model (G) and the unitary model (U) are low for each network slice. It is different for LSTM-based models. It is impossible to assess (based only on visual inspection) whether the differences between the prediction errors of different models are statistically significant. For verification of the differences between the pairs, the Nemenyi post hoc test is performed. The results for Slice A are presented in Tab. 2. If the p-value is less than 0.05 it means that the two algorithms differ significantly at the significance level of 0.05. Otherwise, there are no statistically significant differences. The results for the remaining slices are similar (also for delay). General and unit models created by the same methods do not differ significantly in any case. The most common differences are between CNN-BiLSTM (both U and G) and other methods. For more details, see [1].
The research described above was done with the collaboration of the Nokia Solutions and Networks and Wrocław University of Science and Technology (Faculty of Information and Communication Technology and Faculty of Pure and Applied Mathematics) . The full description of procedures, results and the explanation how this solution realizes the concept of Digital Twin-based network simulator are presented in [1].
As the continuation of the presented results the research team (Nokia and Wroclaw University of Science and Technology) submitted in March 2024 the common project under the FENG (Smart path) program (The National Centre for Research and Development, Poland) “Development of a digitized and automated human-in-the-loop class process and tools to implement it, verifying, categorizing and analyzing the performance impact of Nokia’s new 5G/5G+ product functionalities, contextually recommending optimal parameter settings by integrating synthesized domain expertise and product knowledge with a simulation model in the form of a “digital twin”.”. Within this project the research team from the Faculty of Pure and Applied Mathematics will be responsible for the methods for selecting the level of aggregation and representative sampling of data for identifying and modelling the dynamics of changes in performance and environmental indicators.
Literature
[1] Dulas, D., Witulska, J., Wyłomańska, A., Jabłoński, I., & Walkowiak, K. (2024). Data-driven model for sliced 5G network dimensioning and planning, featured with forecast and” what-if” analysis. IEEE Access.
[2] P. Subramanian, “Future X network cost economics—A network operator’s tco journey through virtualization automation and network slicing”, Bell Labs Consulting, pp. 1-19, Feb. 2019, [online]
By Dominik Dulas and Justyna Witulska (Nokia Solutions and Networks, Wrocław; Wroclaw University of Science and Technology), Agnieszka Wyłomańska (Wroclaw University of Science and Technology), Ireneusz Jabłoński (Brandenburg University of Technology, Cottbus; Fraunhofer Institute for Photonic Microsystems, Cottbus), and Krzysztof Walkowiak (Wroclaw University of Science and Technology)