Exploring Potential of ML-aided Mobile Traffic Prediction for Energy-efficient Optimization of Network Resources Using Real World Dataset

Received 15 June 2024, accepted 24 June 2024, date of publication 1 July 2024, date of current version 16 July 2024. Digital Object Identifier 10.1109/ACCESS.2024.3421633 Exploring Potential of ML-Aided Mobile Traffic Prediction for Energy-Efficient Optimization of Network Resources Using Real World Dataset ANETA KOLACKOVA 1, SALIH SEVGICAN2, MUHAMMET FATIH ULU 2, JALE SADREDDINI3, PAVEL MASEK 1, (Member, IEEE), JIRI HOSEK 1,4, (Senior Member, IEEE), JAN JERABEK1, AND TUNA TUGCU 2 1Department of Telecommunications, Brno University of Technology, 601 90 Brno, Czech Republic 2Department of Computer Engineering, NETLAB, Boˇ gaziçi University, 34342 Istanbul, Turkiye 3Department of Computer Engineering, Istanbul Arel University, 34537 Istanbul, Turkiye 4Unit of Electrical Engineering, Tampere University, 33100 Tampere, Finland Corresponding author: Aneta Kolackova (aneta.kolackov[email protected]) The research described in this paper was financed by the Technology Agency of the Czech Republic under the grant FW10010014. ABSTRACT To meet the extremely stringent but diverse requirements of Beyond Fifth-Generation (B5G) networks, traffic-aware adaptive utilization of network resources is becoming essential. To cope with that, a detailed traffic data analysis enables opportunities for mobile network operators to improve the Quality of Service (QoS) in the next-generation mobile communication systems. This paper presents a comprehensive analysis of the real world data collected from an operator’s 4G+and 5G infrastructure during a sevenmonth campaign. Efficient Machine Learning (ML) based network traffic predictions are presented together with a statistical model to develop optimal resource allocation strategies by using the data gathered during the pandemic, an era when the data volume, as well as the bandwidth requirements and the end users’ expectations, were significantly elevated in terms of QoS, given the huge shift to the online world. Data analysis confirmed the assumption that there are traffic changes during the day and the whole week, which helped us to find new research directions regarding resource allocation optimization of next-generation mobile networks. Furthermore, we introduce the Predictive Energy Saver for Baseband Units (PESBiU) algorithm, which utilizes traffic prediction and power consumption analysis to manage the power states (sleep or active) of BBUs in a network. The PESBiU algorithm utilizes the results from ML predictions to effectively balance energy efficiency and network performance, demonstrating its potential for practical deployment in future mobile communication networks by transitioning BBUs to sleep mode during low-traffic periods, thereby achieving significant power savings. INDEX TERMS Beyond 5G networks, data analysis, machine learning, mobile traffic, resource allocation, BBU energy saving. I. INTRODUCTION The rapid growth in cellular technologies, driven by the ever-increasing mobile data volume, is expected to spawn new communication services and play a major role in applications over the air, such as tactile Internet, holographic The associate editor coordinating the review of this manuscript and approving it for publication was Miguel López-Benítez . telepresence, and collective virtual reality (VR), which can be unified under the extended reality (XR) vision. To enable the operation of all devices and to handle their traffic requirements, telco operators have to deploy new solutions and expand their infrastructure by innovating and implementing the mechanisms of Beyond fifth-generation (B5G) such as Cloud Radio Access Network (C-RAN) [1] and Edge Computing [2]. 93606 2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 12, 2024 A. Kolackova et al.: Exploring Potential of ML-Aided Mobile Traffic Prediction With more equipment and technology all around us, carbon emissions are also increasing. Companies from various industries are diligently working on methods to reduce their carbon footprint. Within this widespread effort, the telecommunications industry is deeply involved, striving to achieve ‘‘Net Zero’’ emissions by 2050 through persistent and sustainable actions [3]. Mobile operators aim to reduce their Operating Expenditure/Capital Expenditure (OPEX/CAPEX) while maintaining the Quality of Service (QoS) at the required level. Due to the significance of QoS-aware networks in mobile communication services for both operators and customers, many researchers have conducted studies on their ideal method of QoS implementation [1]. Considering the forthcoming requirements of the B5G networks, it is also necessary to establish QoS-driven resource management schemes and identify appropriate Key Performance Indicators (KPIs) for the Radio Access Network (RAN) as well as the Core network. With the increasing dynamicity that comes with B5G networks due to new emerging applications and services, it is necessary to address this change. To develop such dynamically adopting architectures and to increase the efficiency of network resource allocation while contributing to the minimization of energy consumption, a comprehensive understanding of the network traffic patterns and users’ behavior is required. To enable this, emerging Machine Learning (ML) technologies are among promising solutions [4]. The aim of this work is to demonstrate an exploratory ML-aided analysis of a real world mobile dataset, which opens up the path for novel ideas regarding the orchestration of B5G networks in return for higher efficiency in operating costs, lower energy consumption, and increased QoS. To this end, we gathered a real world dataset of network traffic for a period of seven months (2020-08-15 to 2021-02-28) in an average-sized city from a 4G+and 5G network. The data explained in Section III-A consists of measurements from different mobile sites, where each site is located in a different area in the city and includes several evolved NodeBs (eNodeBs). Considering that the measurement campaign was conducted during the COVID-19 pandemic, we also want to study the impact of such a specific period on the user’s behavior and cellular traffic data. We investigate and discuss this real world dataset. These data served as a source for network traffic prediction results, which we have achieved with the help of Recurrent Neural Networks (RNNs), utilizing gating mechanisms of Long Short-Term Memory (LSTM) unit and Gated Recurrent Unit (GRU). We compare these results with another MLaided tool, called Facebook’s Prophet (FB’s Prophet) [5], which uses regression techniques based on the statistical and temporal features of the data. Furthermore, based on analysis and prediction results, we created Predictive Energy Saver for Baseband Units (PESBiU) Algorithm which can help to improve utilization of resource allocation, and energy efficiency in B5G networks. The emphasis on this topic is rooted in the considerable share of energy consumption in the total expenses of mobile networks. The research referenced here [6] highlights that energy costs account for 23% of all network-related expenditures. Additionally, it points out that the energy usage of the RAN constitutes 73% of the total energy consumption. The contributions of this paper are summarized as follows: •A comprehensive analysis of recent real world data from 4G+and 5G systems to extract and interpret the real cellular data traffic patterns. •Comparison of ML-aided efficient network data prediction techniques and statistical models to develop optimal resource allocation strategies and minimize energy consumption using the data gathered during the seven-month measurement campaign in a live cellular network. •Creation of the PESBiU algorithm, which leverages ML predictions to manage the power states of BBUs by transitioning them to sleep mode during lowtraffic periods, thus optimizing energy efficiency while maintaining network performance. •Discussion of proposed methodologies and algorithms as solutions for the major challenges that need to be significantly improved in the B5G networks, such as radio resource optimization and green networking. II. RELATED WORKS Much research has been conducted concerning network optimization, resource allocation, and traffic prediction. Feng et al. [7], employ deep learning (DL) techniques for mobile traffic prediction and manage to forecast the 3G/4G traffic with only 6-8% error. Predicting the cellular traffic with high accuracy allows for shutting down underutilized BBUs. However, the dataset used has a few drawbacks. It is from 2014, and the data does not contain separate information about downlink and uplink traffic volumes, which prevents training the traffic volume types and optimizing BBU usage based on downlink and uplink separately. In [8], the authors focus on in-network pooling framework in B5G era and the attention-based deep reinforcement learning algorithm. In [9], Wu et al. focus on energy-efficient base station switching techniques, while in [10], the authors propose strategies for switching off base stations in heterogeneous networks. However, most of the propositions and calculations come from theoretical assumptions and do not consider the use of real world data for 4G+and 5G systems and ML techniques. Similarly, in [11], the authors address the dearth of research on per-user traffic analysis in cellular networks and advocate for a shift from traditional cell-aggregated traffic management to a more energyand cost-efficient user-centric approach, but they artificially generate their dataset. In [12], the real world data was used. The authors manage to cluster the traffic on BBUs in the network alongside the traffic forecast in order to reduce the costs of C-RAN deployments. Their framework effectively increases VOLUME 12, 2024 93607 A. Kolackova et al.: Exploring Potential of ML-Aided Mobile Traffic Prediction the average capacity utility to 83.4% and 76.7% in two datasets and reduces the overall deployment cost to 48.4% and 51.7% of the traditional RAN architecture. However, the authors do not consider the minimization of energy consumption, and their model is trained on a dataset from 2013. For these reasons, there is an important area that requires further investigation and improvement, and that is serving as our motivation. Further, none of the related works utilize recent real world datasets or even data during the COVID-19 pandemic. On the other hand, the literature is rich in studies using synthetic data concerning power consumption generated via simulators. In [13], the authors select a Remote Radio Head (RRH) based on the traffic density using Efficient Local Search Algorithm (ELSA). Lee et al., in [14] minimize power consumption in both BBU and RRH by threshold-based RRH switching and BBU aggregation. The authors in [15] propose reducing the number of active BBUs and power consumption by switching the BBUs on/off, based on the traffic load, using combined Best Fit Decreasing and Modified Best Fit Decreasing algorithms. Having in mind the outcomes from the studies discussed above, intelligent data-oriented approaches remain untouched. When it comes to studies including data-oriented approaches, ML can be seen as one of the core approaches for reducing power consumption, optimal resource allocation in the cloud management systems, optimizing network performance and QoS with Software-defined Networking (SDN) and many other studies in the next generation mobile networks. In [16], the DL methodology called LSTM is proven to work best compared to other linear and non-linear ML algorithms for predicting network traffic patterns. A more recent study presented in [17] runs the LSTM DL algorithm on the Long Term Evolution (LTE) dataset and shows that LSTM is a promising prediction algorithm for mobile traffic. As we can see from the previous works, the main subjects we are interested in are more comprehensive and, therefore, more promising in predicting the cellular traffic pattern. Consequently, our main subjects are to predict the cellular traffic pattern using ML-based tools with a real world dataset and to minimize energy usage with emphasis on the energy efficiency of the BBU, while maintaining a considerable level of QoS. III. RESEARCH METHODOLOGY Let us focus on the measured dataset, which is thoroughly described and explained. Following that, in subsection III-B, implemented ML methodologies are explained. A. DATASET The dataset has overall records of the network quality information from 12 different eNodeBs (RRH+BBU) scattered into four cell sites. For 200 days, 4800 log records consisting of hourly observations of mobile data usage have been recorded for each eNodeB. In the coverage of sites, a total of 212,000 users are served with 1.5 terabytes of traffic volume daily on average. Table 1summarizes the cell site’s important physical communication parameters. TABLE 1. Physical communication parameters. The data is collected from four macro cell sites. Two sites, A and B, are located in the office area with specific antenna deployment on tall buildings’ roofs. The other two, sites C and D, are located in the residential area in cell site shelters between residential buildings. Site C is in a densely populated area, than site D, which is more on the edge of the city. The distance between office cell site B and residential cell site C is approximately 3 km, see Fig. 1. The specific locations of the base stations are left unspecified intentionally to protect the commercial benefits of the operator. FIGURE 1. Location of macro cell sites - Bird’s-eye view. The data has been recorded through the analytic platform, which monitors the performance of cell sites in the network and other information regarding the eNodeBs (cell site status, number of drops, blocks, power consumption etc.). Each eNodeB sends KPIs to this platform database hourly in the form of raw counters. The collected data contains information about the data volumes in the downlink and uplink directions, the number of connected UEs and packet loss, even for specific QoS Class Identifier (QCI) etc. And especially the power consumption data for each BBU and RRH measured in Watt hours. B. MACHINE LEARNING ALGORITHMS The recently introduced and standardized features of the next-generation wireless communications enable operators to gather more data from their mobile networks [18]. In [19], a component called Network Data Analytics Function (NWDAF), which is first introduced with 5G 93608 VOLUME 12, 2024 A. Kolackova et al.: Exploring Potential of ML-Aided Mobile Traffic Prediction for ML purposes, is studied, and a proof-of-concept of ML applications for a 5G network with synthetic data is shown. Thanks to real world data, NWDAF can better optimize a B5G network by predicting near-future network traffic, hence utilizing the network resources better by means of reduced power consumption as well as increased QoS via reduced latency over the SDN network. Not only does the NWDAF component contribute to nextgeneration networking, but it also creates opportunities for the management of B5G networks. For example, an operator can reduce its OPEX and CAPEX by analyzing the data and making future predictions. In [20], recent research about the usage of ML in next-generation networks is discussed, and a detailed inspection of the results that address network efficiency, resource optimization, power allocation, and virtualization is provided. In [12], a comparison of ML methodologies between Artificial Neural Networks (ANNs) and the statistical prediction called Autoregressive Integrated Moving Average (ARIMA) is discussed, showing that ANN outperforms in terms of network prediction purposes. In [16], a comparison among ANN algorithm types shows that LSTM has the minimum Mean Squared Error (MSE) of 0.042 while Feed Forward Neural Network (FNN) has the highest MSE score of 0.091. In [21], the authors show that for network traffic prediction, the accuracy of LSTM is around 90%, which is approximately 10% higher than other DL algorithms. In [16], although LSTM shows more accurate results than GRU in predicting GÉANT backbone network traffic, its running time and computational cost are higher than GRU. Considering the mentioned efficiency of GRU and LSTM, both mechanisms are comparable. Therefore, besides LSTM, a more recent gating mechanism, GRU, is proposed for this work to show how both methods perform in network traffic with our dataset. As the studies discuss, ML is one of the great solutions for the network scalability problems in the future of wireless networking. Hence, working with recent real-life data is crucial to making ML solutions more realistic and keeping them as up-to-date as possible, as network patterns may change. This study uses two types of RNN, LSTM and GRU, to achieve the best network traffic prediction results based on their proven superior performance among other DL mechanisms, as explained above. The technical reasons for LSTM and GRU’s superiority are that both are good at predicting while utilizing historical data. On the other hand, as Cho et al. suggest in the GRU’s proposition paper [22], GRU’s optimal performance can be achieved in small datasets. Therefore, LSTM tends to perform better than GRU in large datasets. This conditional performance difference can be better understood if we examine the implementation of the LSTM and GRU algorithms. Before moving into implementation details, we should define some fundamental terms, such as gates and identity connection. Gates in Gated Neural Networks (GNNs) are decision-makers for using identity connections or regular stacked layers. The significance of identity connections is that not only do they use the output of consecutive layers, but also they can use the output of much lower layers. Thus, this gate mechanism allows GNNs to learn incrementally, making neural networks safe from long-term dependencies. Two algorithms, LSTM and GRU, are types of RNNs and use a gating mechanism. Besides RNNs, Prophet is used, which is a powerful tool that works for linear or non-linear growth and can help generate accurate and reliable forecasts for time series data with multiple seasonality. A brief explanation of RNNs, the implementation details of LSTM and GRU, and a description of FB’s Prophet are as follows: 1) RNN RNNs use sequential data, which can be time-dependent, and are a type of ANN. A significant feature of RNNs is that they can utilize previous input to update the current output. The significance of the data at the very beginning can be preserved. RNN usageareas include Natural Language Processing (NLP), image captioning, speech recognition, and language translation. Therefore, using RNN mechanics to analyze sequential mobile network traffic data is appropriate. However, one drawback of Vanilla RNN is that the model’s learning can slow down or even stop. Gradients become increasingly small as they are back-propagated to earlier layers during training. This phenomenon can make it difficult for the model to learn long-range dependencies, as the gradients may become so small that they vanish, preventing the network from effectively updating the weights of earlier layers. One of the main causes of the vanishing gradient problem in RNNs is the repeated multiplication of gradients through the weight matrices during Back-Propagation Through Time (BPTT). If these weight matrices have eigenvalues close to zero, the gradients can quickly diminish as they are multiplied across multiple time steps. To mitigate the vanishing gradient problem, various techniques have been proposed, including using activation functions that alleviate the saturation of gradients (such as ReLU), using gradient clipping to prevent gradients from becoming too large or too small, and using architectures like Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU) which are designed to better capture long-range dependencies by controlling the flow of information. 2) LSTM LSTM is proposed by S. Hochreiter and J. Schmidhuber in [23], the first attempt to solve the vanishing gradient problem by designing an advanced activation function. Additionally, LSTMs are trained using the BPTT method, a variant of backpropagation that allows gradients to flow through the entire sequence, enabling the model to learn longterm dependencies. The representation of a single unit of LSTM is in Fig. 2. There are three types of gates in an LSTM unit: a forget gate, an input gate, and an output gate. The previous hidden VOLUME 12, 2024 93609 A. Kolackova et al.: Exploring Potential of ML-Aided Mobile Traffic Prediction FIGURE 2. A Representation of an LSTM Unit. state (known as short-term memory) vector (Path 2 in Fig. 2) and the input vector (Path 1 in Fig. 2) concatenate (say the input&hidden concatenation for the semantic convenience for this section), and all gates use the input&hidden concatenation. The forget gate (Path 8 in Fig. 2, abbreviated as ft) is the result of the sigmoid function (defined as σ(x)=ex/(1+ex) applied to the summation of the dot product of the forget gate weight matrix and the input&hidden concatenation (Path 3 in Fig. 2) and forget gate bias vector, i.e., ft=σ(Wf·[ht−1,xt]+bf),(1) where σ(arg) is the sigmoid function that takes argument arg and maps the argument between 0 and 1, Wfis the current forget gate weight matrix, [ht−1,xt] is the interpolation of the previous hidden state vector and input vector, and bfis the current forget gate bias vector. The semantic responsibility of the forget vector is to decide what and how much to discard or to decide what and how much to remember. Eventually, the previous cell state vector (Path 13 in Fig. 2) is multiplied by the forget vector element-wise (and results as Path 14 in Fig. 2). Applying the sigmoid function to the summation of the dot product of the input gate weight matrix and the input&hidden concatenation and input gate bias vector results as the input gate (Path 9 in Fig. 2, abbreviated as it), which will be then multiplied by the new candidate values vector (Path 10 in Fig. 2, abbreviated as ˜ Ct). it=σ(Wi·[ht−1,xt]+bi).(2) A similar approach can find the new candidate values vector, ˜ Ct. (Only the activation function is different, and it is tanh (defined as tanh(x)=(e2x−1)/(e2x+1) now). ˜ Ct=tanh(W˜ C·[ht−1,xt]+b˜ C).(3) The semantic responsibility of the input gate is to decide what to learn from the new input matrix, Xt. In the input gate-related calculations, the sigmoid function and the tanh function are used. tanh maps input value arg between −1 and 1. Multiplication of itwith ˜ Ctmaps the values between −1 and 1 (Path 11 in Fig. 2). Without calculations of the output gate, the cell state (known as long-term memory) can be calculated (Path 15 in Fig. 2). Ct=Ct−1⊙ft+˜ Ct⊙it,(4) where ⊙symbolizes the Hadamard product. This step opens the possibility that the boundaries of some cell state vector variables are not limited to −1 or 1. Therefore, the tanh function will shrink the cell state in the output gate-related calculations. Till now, the cell state vector has been updated. What remains is an update of the hidden state vector. The output gate-related calculations update the short-term memory. Similar to the previous calculations, the output gate vector (Path 12 in Fig 2) is: Ot=σ(Wo·[ht−1,xt]+bO).(5) Connecting this equation with the cell state vector, we have: Ht=Ot⊙tanh(Ct).(6) Thus, the working principle of all three gates’ and candidate values vectors is explained, and selecting LSTM for our case is justified. 3) GRU The second type of RNN used in this work is GRU, proposed by Cho et al. in [22]. GRU is similar to LSTM in methodology to solve the vanishing gradient problem. However, the key differences make one preferable depending on the machine learning applications, dataset length, etc. FIGURE 3. A Representation of a GRU. A single GRU has two types of gates: a reset gate and an update gate. The reset gate (known as short-term memory) decides the amount of negligible past information and ignores it, and it works as the unit’s short-term memory. The calculation of the GRU gates has more parameters than any gate calculation in an LSTM unit. Therefore, let us divide these calculations into steps and start with the reset gate. The first step is to find the impact of the input vector (say this impact vector Ii) on the reset gate: Ii=Wr·Xt+br,(7) 93610 VOLUME 12, 2024 A. Kolackova et al.: Exploring Potential of ML-Aided Mobile Traffic Prediction where Wris the weight matrix of the reset gate for processing the input vector (Xt, there is another weight matrix for the previous hidden state vector, making this calculation a bit longer), and bris the bias vector of the reset gate for processing the input vector (there is another bias vector too). The second step is to find the impact of the previous hidden state vector: Ih=ωr·ht−1+βr,(8) where ωris the weight matrix of the reset gate for processing the previous hidden state vector (ht−1), and βris the bias vector of the reset gate for processing the previous hidden state vector. Then, the forget gate vector (Path 10 in Fig. 3) can be calculated as: Rt=σ(Ii+Ih).(9) The update gate decides the amount of previous information that is important and provides this information survive. It behaves as long-term memory. The calculation of the update gate is similar to the reset gate. Impact of the input vector (say this impact vector 8i) on the update gate: 8i=Wu·Xt+bu,(10) where Wuis the weight matrix of the update gate for processing the input vector, and buis the bias vector of the update gate. The impact of the previous hidden state vector is: 8h=ωu·ht−1+βu,(11) where ωuis the weight matrix of the update gate, and βuis the bias vector of the update gate. Both are for processing the hidden state vector. Then, the update gate vector (Path 12 in Fig. 3) can be calculated as: Ut=σ(8i+8h).(12) The first step to calculate the candidate hidden state vector (Path 16 in Fig. 3) is to find the impact of the input vector (say this impact vector τi) on the candidate hidden state vector: τi=W˜ h·Xt+b˜ h,(13) where w˜ his the weight matrix and b˜ his the bias vector of the candidate hidden state. Both are for processing the input vector. The second step is to find the impact of the reset gate, calculated in the Eq. 9: τr=ω˜ h·(ht−1⊙Rt)+β˜ h,(14) where ω˜ his the weight matrix and β˜ his the bias vector of the candidate state. Then, the candidate hidden state vector (Path 16 in Fig. 3) can be calculated as: ˜ ht=tanh(τi+τr).(15) Lastly, we can calculate the current hidden state vector as follows: ht=ht−1⊙Ut+˜ ht⊙(1 −Ut).(16) 4) PROPHET Prophet allows forecasting sequential data depending on time. In the procedure, considerations of daily, weekly, yearly, seasonality, etc. effects are present. Thus, non-linear trends fit these considerations. It is a statistics-based method that works best when the effects are robust. To compare RNN algorithms with the statistical techniques, we used FB’s Prophet prediction. We aim to show that LSTM and GRU outperform statistical techniques when evaluated using real world, up-to-date data and demonstrate a comparison between LSTM and GRU performance-wise. C. ENERGY CONSUMPTION IN B5G To accommodate the increasing traffic demands associated with B5G networks, it’s imperative to expand capacity, which requires the deployment of more eNodeBs within the RAN segment. However, eNodeBs are not only expensive but also highly energy-intensive. According to survey [6], eNodeBs equipment is responsible for 50% of the total energy consumption in RAN, with air conditioning systems following closely at 40%. These cooling systems are essential to prevent eNodeBs from overheating. Therefore, reducing the energy consumption of eNodeBs and enhancing their energy efficiency emerges as a primary strategy for lowering overall energy usage. 1) ENERGY CONSUMPTION OF BASEBAND UNIT (BBU) As was mentioned in III-A, eNodeB has two parts - BBU + RRH. The BBU handles the baseband processing tasks, while the RRH is responsible for converting BBU signals into radio signals (and vice versa) to facilitate transmission or reception through antennas. Typically, a single BBU collaborates with several RRHs to manage the processing of radio signals. This implies that the power consumption of a BBU escalates with the number of connected RRHs and the volume of signaling required, which varies according to traffic demands [3]. From a hardware viewpoint, the energy consumption of a BBU varies based on the design of its components. The types of processors and modems incorporated into a BBU play a significant role in the overall energy efficiency. Given that BBU equipment predominantly consists of digital components, the specific architecture or design chosen for a BBU can result in varying levels of energy usage. Additionally, certain BBU models offer advanced powersaving capabilities, which are influenced by their physical form factor and the design of their chipsets. Considering it from software perspective in B5G networks, energy saving strategies for BBUs have centered around techniques like Dynamic Scaling, Sleep Modes and Cloud-Based and Virtualized BBUs. Dynamic Scaling involves adjusting the BBU’s processing capacity to match the current traffic load, potentially powering down unneeded resources during low usage periods. Sleep Modes, another tactic, entail putting BBUs into a low-power state during off-peak times, which can drastically cut energy use, provided it’s done without negatively impacting network VOLUME 12, 2024 93611 A. Kolackova et al.: Exploring Potential of ML-Aided Mobile Traffic Prediction service. Transitioning BBUs to cloud-based or virtualized settings enables more efficient resource use network-wide, minimizing the necessity for physical infrastructure and thus conserving energy. These virtualized BBUs can be adjusted dynamically to match demand, enhancing energy efficiency. Additionally, employing Artificial Intelligence (AI) and ML offers a forward-thinking method by predicting traffic flows and adjusting BBU operations in real-time for optimal energy savings. This article highlights the use of a combined approach to improve energy efficiency in B5G networks. It illustrates a strategy to reduce energy consumption in BBUs, showcasing innovative solutions for sustainable network management [3]. In real-world commercial networks, traffic generation fluctuates between full and empty loads, influenced by factors such as time, user count, and coverage area. Leveraging this dataset, we can correlate the actual traffic load with energy consumption, enabling us to develop an algorithm aimed at enhancing the energy efficiency of BBUs. 2) ENODEB’S FORMULATION OF ENERGY CONSUMPTION The ratio of energy consumption between BBUs and RRHs depends on used technology, cell load, weather conditions, temperature etc. In [3], SK Telecom reports that the amount of power consumed by BBU is 22% whereas that by RRH is 78%, which is almost 1:4. According to our measurements, from the data set, the ratio is 1:5. Generally, RRHs contribute more significantly to the overall energy consumption within a network than BBUs. On the other hand, BBUs are generally underutilized because of the dynamic behavior of network traffic, as will be shown later in this paper. As a result, it is critical to address these issues by leveraging the complex characteristic of traffic for improving the energy efficiency of BBUs at low traffic hours. To compute the weighted average of power/energy consumption, weighting factors are applied according to a daily (24-hour) traffic load distribution profile, which includes three load levels: low load (low), medium load (med), and high load (high) load. This profile is characterized by the duration, in hours, of each load level, denoted as tlow, tmed , and thigh, respectively. Operators have the flexibility to establish a load distribution profile that accurately mirrors the network’s conditions and require that this specific profile be utilized in calculations. The average power consumption [W] of central and remote parts defined as [24]: Peqp =PBBU +PRRH ,(17) in which PBBU and PRRH [W] are average power consumption of central and remote parts defined as: PBBU =(Phigh,BBU ·thigh +Pmed,BBU ·tmed +Plow,BBU ·tlow) (thigh +tmed +tlow) (18) PRRH =(Phigh,RRH ·thigh +Pmed,RRH ·tmed +Plow,RRH ·tlow) (thigh +tmed +tlow) (19) IV. ANALYSIS AND EXPERIMENTAL RESULTS ML algorithms require that the raw data is cleaned and processed in such a way that the algorithm would perform with minimal error during training and testing. In our study, we utilized Python for data cleaning, preprocessing, and the implementation of machine learning models. A. PRE-PROCESSING OF DATA The dataset used in this study and explained in Section III-A consists of traffic volume, throughput, and packet loss in the mobile network where these parameters are KPIs of B5G networks [25]. Parameters that are out of the scope of this study, such as call duration, are discarded. The data samples are taken every hour from eNodeBs. However, due to the occasional maintenance periods in the network, it was not possible to record some data points on eNodeBs. To prevent bias and misinterpretation, the daily observation periods when the eNodeBs were under maintenance are removed, which corresponds to 0.80% of all records. Secondly, the data has been split into training and testing datasets. The training part is approximately 75% of the overall dataset, while the testing dataset is about 25%. The testing dataset is required to confirm the robustness of the trained algorithm. The ML models explained in Section III-B have been trained using training data, and the results are given by calculating the difference between the actual testing data and the predictions done by the trained ML algorithm. The training and testing datasets have been scaled using a standard, common algorithm before training a DL algorithm. After pre-processing, we train the algorithms explained in Section III-B and discuss the results in the following section. B. DATA ANALYSIS The operation of a mobile network has a dynamic nature. Over time, the patterns can vary significantly, which makes them hard to predict, especially in a short-term perspective. Therefore, the data gathered in a mobile network can exhibit temporal and spatial variations. In our study, we have gathered and categorized the data under four different cell sites, namely sites A, B, C, and D, as explained in Section III-A. Sites A and B correspond to the region where offices and business centers are located. Sites C and D correspond to the residential areas (see Fig. 1). Data analysis shows that sites A and B have similar characteristics through time, and the same goes for sites C and D. Due to the similarities among the sites, we explain the figures of sites A and C as representatives of the business and the residential districts and discuss them together with sites B and D. Fig. 4shows a comparison of uplink and downlink traffic volumes across all site cells. As shown, the processed data indicate diverse characteristics. In the case of site C (green line), the traffic volume pattern surpasses the other sites by 179%, which results in a 7.5 GB difference on the average per day and leads to higher utilization of site C. Another indication is that site C (green line) and site D (red line) have 93612 VOLUME 12, 2024 A. Kolackova et al.: Exploring Potential of ML-Aided Mobile Traffic Prediction FIGURE 4. Daily avg. DL and UL traffic volume. FIGURE 5. Daily number of active users. FIGURE 6. Daily average packet loss in the network. different traffic volume trends than site A (blue line) and site B (orange line) after about 2:00 p.m. Comparing the traffic volume by the end of the day, after 4:00 p.m., the sites located in the office area (A and B) have a 52 % traffic volume drop. Meanwhile, the sites in the residential areas (C and D) have a 10 % drop in traffic volume. In Fig. 5, the average number of active users per site is shown. This figure has similarities with the daily traffic volume in Fig. 4. This similarity is expected since a high number of users registered on the cell site could generate higher traffic volume in the area. One data characteristic in sites C and D is unique to residential areas: The number of active users towards the evening does not show a steep decline as in sites A and B, but the number goes down until 5:00 a.m of the following morning. Despite a shallow declination in the number of active users, the traffic volume of sites C and D does not show a similar pattern as the number of users, which shows that at this time period of the day, after 4:00 p.m., the amount of data traffic per device is much higher than FIGURE 7. Average traffic volume on sites through a week. average. Therefore, we can conclude that the traffic volume stays almost the same for the rest of the work shift hours. We also look at the packet loss parameter’s behavior to understand the cell sites’ reliability and ease of connectivity. In our investigation, we average the packet loss values over time and remove the effects of packet loss spikes, such as data anomalies (values beyond the possible boundaries, i.e., values taken on extreme fluctuations). However, information about network reliability can be obtained by examining the packet loss inspection figure, which comprehends the whole time period. In Fig. 6, we observe the packet loss parameter follows a similar pattern as traffic volume data. However, if we consider cell sites individually, there are some conclusions to make. When sites C and D are compared, despite the low number of active users in site D, packet loss is higher than in site C, which is a cell site from a residential area as well. Site B’s packet loss reaches much higher values at the beginning of the work shift around 8:00 a.m. and continues to have higher values than the other sites when compared. Aside from daily statistics, weekly changes in traffic volume are also investigated. In Fig. 7, we compare the total traffic volume across all sites. The typical behavior is that the total traffic volume shows a decreasing trend on weekends. Site cells in residential districts have gentle decline patterns compared to the site cells located in business districts. C. RESULTS OF ML-BASED NETWORK TRAFFIC PREDICTION With the insights we obtained from Section IV-B, we applied three different algorithms for network traffic volume prediction. Implementing three different algorithms allows us VOLUME 12, 2024 93613 A. Kolackova et al.: Exploring Potential of ML-Aided Mobile Traffic Prediction to compare them. The comparison may enable the selection of the optimum method under similar circumstances. First, considering the temporal characteristics of the data, we apply FB’s Prophet algorithm that makes predictions by using temporal features such as seasonality, trend, and holiday effects. The second approach is implementing a neural network consisting of multiple layers of LSTM neurons. We have a neural network structure consisting of multiple GRU neuron layers as the third and last approach. Further, we discuss the implementation insights and pairwise comparisons of the three different algorithms. 1) FB’S PROPHET As discussed in Section III-B, the authors in [5] created a library to generalize processing implementations and predict time-series data. This library is popular because it helps researchers make predictions based on the data’s statistical features by decomposing the time series into its seasonality and trends while taking the holidays and similar time-specific features into account. The method predicts the value of a function at time tusing a trend component and an error term. A trend component can also be modeled as a linear function. Moreover, Prophet uses the Fourier series to model different types of modeling, allowing flexibility in this area. Prophet predicts the posterior distribution of parameters, where Bayesian inference is used. Also, the Markov Chain Monte Carlo (MCMC) method is utilized in this part of the algorithm. 2) LSTM In Section II, we have discussed the studies that proved LSTM is one of the best practices when it comes to making predictions on time series data. With this motivation, we apply an LSTM neural network consisting of seven layers together with input and output layers. Layers are added sequentially with different dimensionalities of output spaces. Linear and Exponential Linear Unit activation functions are used to construct the layers, and three of five hidden layers use linear activation functions. The remaining two layers use exponential linear unit activation functions. To initialize weights with statistical function, our choice as a kernel initializer is the normal distribution for every layer. We choose the MAE loss function and Adaptive Moment Estimation Algorithm (ADAM) optimizer for the LSTM network to minimize prediction errors. The parameters and configuration we have chosen for the LSTM algorithm constitute an initial prediction model, which does not heavily lean on the problem’s complexity (in our case, the traffic volume prediction). An example of LSTM predicting the traffic volume on January 26, 2021, can be seen in Fig. 10. 3) GRU We have explained the GRU’s mechanism and referred to some papers that show GRU’s significance in predicting sequential data. Therefore, we construct a GRU neural network consisting of six hidden layers. Six hidden layers TABLE 2. Network traffic volume predictions for downlink. TABLE 3. Network traffic volume predictions for uplink. are experimentally selected as optimum. Having more layers did not bring any considerable accuracy improvement in predicting the test inputs but increased overall memory consumption. Having fewer layers decreased accuracy, and despite that, it lessened overall memory consumption; losing accuracy is not preferred. Further, a batch normalization layer, an input, and an output layer. Our implementation mainly has two differences between LSTM-aided neural network and GRU-aided neural network: the first difference is the choice of GRU, and the second one is that we add a batch normalization layer. The remaining parts of the implementation are the same. Prediction results of FB’s Prophet, LSTM, and GRU are given in Tables 2and 3. The Mean Absolute Percentage Error (MAPE)1and Mean Absolute Error (MAE)2results of FB’s Prophet prediction in Tables 2and 3show that the error for downlink is lower than that for uplink. This observation follows the fact that the downlink traffic is more predictable by using statistics. Therefore, uplink traffic has a higher standard deviation and is less predictable. The results are provided by taking the average of ten trials (i.e., using different numbered epochs and/or different neural network architecture) since the performance of AI prediction changes based on those parameters. The comparison between pairs of FB’s Prophet, LSTM, and GRU is made by taking the lower one as the reference point. The calculation is made by 100 ×(Ai−Aj)/Aj, where Astands for ith or jth accuracy, provided that Aiis greater than Aj. Further, Ais can be found with 100 −MAPEi. For instance, by checking Table 2, the accuracy of LSTM for site A is 100 −7.01 = 92.99 (%), whereas the accuracy of FB’s Prophet is 100 − 29.32 =70.68 (%). Then, using the calculation formula, 1MAPE represents how much the prediction deviates from the actual data in terms of percentage. 2MAE calculates the difference between the prediction and the actual data using the sum of absolute differences in terms of megabytes. 93614 VOLUME 12, 2024 A. Kolackova et al.: Exploring Potential of ML-Aided Mobile Traffic Prediction [2] S. Math, L. Zhang, S. Kim, and I. Ryoo, ‘‘An intelligent real-time traffic control based on mobile edge computing for individual private environment,’’ Secur. Commun. Netw., vol. 2020, pp. 1–11, Oct. 2020. [3] Green Mobile Network: Energy Saving Efforts by SK Telecom and NTT DOCOMO. Accessed: Mar. 2024. [Online]. Available: https://www.docomo.ne.jp/english/corporate/technology/rd/docomo6g/ [4] A. Dogra, R. K. Jha, and S. Jain, ‘‘A survey on beyond 5G network with the advent of 6G: Architecture and emerging technologies,’’ IEEE Access, vol. 9, pp. 67512–67547, 2021. [5] S. J. Taylor and B. Letham, ‘‘Forecasting at scale,’’ Amer. Statistician, vol. 72, no. 1, pp. 37–45, 2018. [6] 5G Energy Efficiencies: Green is New Black. Accessed: Mar. 2024. [Online]. Available: https://data.gsmaintelligence.com/research/research/ research-2020/5g-energy-efficiencies-green-is-the-new-black [7] J. Feng, X. Chen, R. Gao, M. Zeng, and Y. Li, ‘‘DeepTP: An end-to-end neural network for mobile cellular traffic prediction,’’ IEEE Netw., vol. 32, no. 6, pp. 108–115, Nov. 2018. [8] Z. Di, T. Luo, C. Qiu, C. Zhang, Z. Liu, X. Wang, and J. Jiang, ‘‘In-network pooling: Contribution-aware allocation optimization for computing power network in B5G/6G era,’’ IEEE Trans. Netw. Sci. Eng., vol. 10, no. 3, pp. 1190–1202, May 2023. [9] J. Wu, Y. Zhang, M. Zukerman, and E. K. Yung, ‘‘Energy-efficient base-stations sleep-mode techniques in green cellular networks: A survey,’’ IEEE Commun. Surveys Tuts., vol. 17, no. 2, pp. 803–826, 2nd Quart., 2015. [10] F. Han, S. Zhao, L. Zhang, and J. Wu, ‘‘Survey of strategies for switching off base stations in heterogeneous networks for greener 5G systems,’’ IEEE Access, vol. 4, pp. 4959–4973, 2016. [11] A. Azari, F. Salehi, P. Papapetrou, and C. Cavdar, ‘‘Energy and resource efficiency by user traffic prediction and classification in cellular networks,’’ IEEE Trans. Green Commun. Netw., vol. 6, no. 2, pp. 1082–1095, Jun. 2022. [12] L. Chen, D. Yang, D. Zhang, C. Wang, J. Li, and T.-M.-T. Nguyen, ‘‘Deep mobile traffic forecast and complementary base station clustering for C-RAN optimization,’’ J. Netw. Comput. Appl., vol. 121, pp. 59–69, Nov. 2018. [13] W. Zhao and S. Wang, ‘‘Traffic density-based RRH selection for power saving in C-RAN,’’ IEEE J. Sel. Areas Commun., vol. 34, no. 12, pp. 3157–3167, Dec. 2016. [14] Y. Lee, K. Miyanabe, H. Nishiyama, N. Kato, and T. Yamada, ‘‘Thresholdbased RRH switching scheme considering baseband unit aggregation for power saving in a cloud radio access network,’’ IEEE Syst. J., vol. 13, no. 3, pp. 2676–2687, Sep. 2019. [15] S. R. Aldaeabool and M. F. Abbod, ‘‘Reducing power consumption by dynamic BBUs-RRHs allocation in C-RAN,’’ in Proc. 25th Telecommun. Forum (TELFOR), 2017, pp. 1–4. [16] R. Vinayakumar, K. P. Soman, and P. Poornachandran, ‘‘Applying deep learning approaches for network traffic prediction,’’ in Proc. Int. Conf. Adv. Comput., Commun. Informat. (ICACCI), Sep. 2017, pp. 2353–2358. [17] S. Jaffry and S. F. Hasan, ‘‘Cellular traffic prediction using recurrent neural networks,’’ in Proc. IEEE 5th Int. Symp. Telecommun. Technol. (ISTT), Nov. 2020, pp. 94–98. [18] I. Rahman, S. M. Razavi, O. Liberg, C. Hoymann, H. Wiemann, C. Tidestav, P. Schliwa-Bertling, P. Persson, and D. Gerstenberger, ‘‘5G evolution toward 5G advanced: An overview of 3GPP releases 17 and 18,’’ Ericsson Technol. Rev., vol. 2021, no. 14, pp. 2–12, Oct. 2021. [19] S. Sevgican, M. Turan, K. Gökarslan, H. B. Yilmaz, and T. Tugcu, ‘‘Intelligent network data analytics function in 5G cellular networks using machine learning,’’ J. Commun. Netw., vol. 22, no. 3, pp. 269–280, Jun. 2020. [20] A. Mughees, M. Tahir, M. A. Sheikh, and A. Ahad, ‘‘Towards energy efficient 5G networks using machine learning: Taxonomy, research challenges, and future research directions,’’ IEEE Access, vol. 8, pp. 187498–187522, 2020. [21] I. Alawe, A. Ksentini, Y. Hadjadj-Aoul, and P. Bertin, ‘‘Improving traffic forecasting for 5G core network scalability: A machine learning approach,’’ IEEE Netw., vol. 32, no. 6, pp. 42–49, Nov. 2018. [22] K. Cho, B. van Merrienboer, D. Bahdanau, and Y. Bengio, ‘‘On the properties of neural machine translation: Encoder–Decoder approaches,’’ in Proc. SSST-8, 8th Workshop Syntax, Semantics Struct. Stat. Transl., 2014, pp. 103–111. [23] S. Hochreiter and J. Schmidhuber, ‘‘Long short-term memory,’’ Neural Comput., vol. 9, no. 8, pp. 1735–1780, Nov. 1997. [24] Environmental Engineering (EE); Metrics and Measurement Method for Energy Efficiency of Wireless Access Network Equipment; Part 1: Power Consumption—Static Measurement Method, Standard ETSI 202 706-1, 2022. [25] T. Norp, ‘‘5G requirements and key performance indicators,’’ J. ICT Standardization, vol. 6, no. 1, pp. 15–30, 2018. [26] Product Description for Ericsson 5216. Accessed: Apr. 2024. [Online]. Available: https://www.tempestns.com/products/ericsson-baseband-5216kdu137925-31/ ANETA KOLACKOVA received the master’s degree in telecommunications and information technology from Brno University of Technology, in 2019, where she is currently pursuing the Ph.D. degree in 4G and 5G mobile technologies. During her studies, she also studied communication technologies with the Instituto Superior de Engenharia de Lisboa, in 2016 and 2017. In her research and development activities, she has been involved in projects dealing with mobile network virtualization, transport and backbone performance optimization of wireless systems, and cloud technologies. During her internship at a foreign operator, she gained hands-on experience with equipment from Ericsson and Cisco, where she is also further involved in building 5G networks. SALIH SEVGICAN received the B.Sc. and M.Sc. degrees in computer engineering from Bogazici University, Turkiye, in June 2018 and January 2022, respectively. He was a Research Assistant during his exchange term with the University of Oulu, Finland. He is currently the Head of the Mobile Software Development Team and AI Team, Harpeon, Istanbul. His research interests include next-generation cellular networks, artificial intelligence, and cloud computing. MUHAMMET FATIH ULU received the bachelor’s degree in physics from Bilkent University, in 2020. He is currently pursuing the master’s degree in computer engineering with Boˇ gaziçi University. He is part of the Software Engineering Team, Mavinci, where he focuses on developing research and development applications powered by machine learning technologies. VOLUME 12, 2024 93621 A. Kolackova et al.: Exploring Potential of ML-Aided Mobile Traffic Prediction JALE SADREDDINI received the Ph.D. degree in computer engineering from Karadeniz Technical University, in 2018. She currently holds the position of a 5G BB System Developer with Ericsson Canada Inc., focusing on carrier aggregation implementation in both downlink and uplink. Prior to her industry role, she was an Assistant Professor with Istanbul Arel University, also coordinating the Erasmus+ Program, from 2018 to 2020. In Summer 2019, she collaborated with Prof. Halim Yanikomeroglu with Carleton University. He has a history of international research, including a Visiting Researcher with Brno University of Technology, Czech Republic, in 2016, where she worked on a 3GPP long-term evolution testbed. Recognized for her teaching excellence, she was awarded ‘‘Best Teaching Assistant’’ from the Computer Engineering Department (2015–2016). PAVEL MASEK (Member, IEEE) received the M.Sc. and Ph.D. degrees in electrical engineering from the Faculty of Electrical Engineering and Communication, Brno University of Technology (BUT), Czech Republic, in 2013 and 2017, respectively. He is currently a Researcher with the Department of Telecommunications, BUT. He is also co-supervising the WISLAB Research Group, where his current interests include various aspects in the area of heterogeneous communication networks. He has co-authored more than 120 research works on a variety of networking-related topics in internationally recognized venues, including those published in IEEE Communications Magazine and several technology products. JIRI HOSEK (Senior Member, IEEE) received the M.Sc. and Ph.D. degrees in electrical engineering from Brno University of Technology (BUT), in 2007 and 2011, respectively. He is currently an Associate Professor and the Deputy Vice-Head of Research and Development (R&D) and International Relations with the Department of Telecommunications, BUT. Since 2012, he has been coordinating the WISLAB Research Group, focusing on modern wireless technologies and applications. He has co-authored over 150 research works in networking, mobile communications, quality of service, quality of experience, and the IoT services. JAN JERABEK was born in Bruntal, Czech Republic, in 1982. He received the B.Sc. and M.Sc. degrees and the Ph.D. degree in electrical engineering from Brno University of Technology, Czech Republic, in 2005, 2007, and 2011, respectively. He is currently an Associate Professor with the Department of Telecommunications, Faculty of Electrical Engineering and Communication, Brno University of Technology. His research interests include analogue signal processing including circuit design, analyses, and measurements. TUNA TUGCU received the B.S. degree in computer engineering from Boˇ gaziçi University, the M.S. degree in computer and information science from New Jersey Institute of Technology, and the Ph.D. degree in computer engineering from Boˇ gaziçi University. He was a Postdoctoral Fellow and a Visiting Professor with Georgia Institute of Technology. He is currently a Professor with the Computer Networks Research Laboratory (NETLAB) and Nanonetworking Research Group (NRG), Boˇ gaziçi University. His research interests include nanonetworking, molecular communications, and wireless networks. He has served in the NATO Science and Technology Organization IST104-RTG050 Group (Cognitive Radio in NATO II) and IST-ET-074 Group (Network Aspects of Cognitive Radio). He also serves for the IEEE P1906.1 Nanoscale and Molecular Communications Working Group, contributing to the IEEE 1906.1-2015 Recommended Practice for Nanoscale and Molecular Communication Framework and IEEE Standard Data Model for Nanoscale Communication System standard specifications. He is currently an Associate Editor of IEEE TRANSACTIONS ON MOLECULAR BIOLOGICAL AND MULTI-SCALE COMMUNICATIONS. He is a devoted and active member of the #WeDoNotAcceptWeDoNotGiveUp movement for free, autonomous, and democratic academia. 93622 VOLUME 12, 2024