Mobility-Aware Cell Clustering Mechanism for Self-Organizing Networks

Self-Organizing Networks (SON) which automate the mobile networks in a cost-efficient way can provide extensive benefits for mobile network operators that are facing several challenges, such as providing maximum coverage and balanced/efficient usage of spectrum and energy. However, a major challenge in future wireless cellular systems is the design of self-organizing architecture that can enable re-configurable, scalable, flexible, low-cost, and efficient solutions for supporting increasingly diverse applications, products, and services of user and network requirements. In this paper, we discuss the benefits and opportunities of mobility-aware clustering capability on different conventional SON functions and advanced Long-Term Evolution features. We further propose a novel cell clustering methodology, which utilizes handover attempt data and aims to collect the cells that have higher mobility activity with each other in a single cluster. One of the main benefits of the clustering mechanism is to reduce the complexity of network optimization processes via checking smaller number of cells instead of evaluating entire network, as their number is gradually increasing. We consider real-world data of evolved Node-Bs (eNodeBs) in one of the major cities while evaluating our novel clustering technique that considers the mobility activities and location information of eNodeBs. In our evaluations, k-means clustering is used as a benchmark. The clustering technique provides the detection of the cells whose serving regions are overlapped. Our results indicate that up to %6 increment is achieved in number of the ratio of intra-cluster handover attempts to total handover attempts. Additionally, the results further show a reduction in the standard deviation of the number of cells within a single cluster with a ratio of %8, which yields more uniform distribution of the cells across different clusters.

SECTION I. Introduction

International Telecommunication Union (ITU) is expected to approve Fifth Generation (5G) standards around 2020 and initial commercial deployments are expected to start in early 2020s. The 3rd Generation Partnership Project (3GPP) is currently standardizing 5G in Release 15 and will advance the support for new features in Release 16 [1]. Existing commercial systems on the other hand, (both 3GPP and non-3GPP based) heavily rely on static and detailed network planning. However, depending on traffic load and mobility characteristics, the network infrastructure needs to be flexible and elastic, i.e., shrink or enlarge depending on demand as required. Moreover, the timescale of this scaling should be much shorter than current timescales of days. As a result of this elasticity demands, traditional network services (e.g. providing mobility support, session management) need to be adapted and self-organize themselves to the current needs without strict and conservative network planning.

Advanced physical and link layer solutions (including orthogonal frequency division multiplexing (OFDM), multiple-input multiple-output (MIMO), adaptive modulation and coding scheme (MCS) etc.) included by specifications of Global System for Mobile Communications (GSM), Universal Mobile Telecommunications Service (UMTS), and Long Term Evolution (LTE) with the aim of improving the link efficiency have reached their theoretical limits [2]. Therefore, increasing node deployment density is the only possible solution to improve the Radio Access Network (RAN) performance [3]. Heterogeneous networks (HetNets) that include picocells, femtocells, relay nodes etc. have been increasingly gaining momentum to meet the requirements of the data explosion. On the other hand, HetNet developments have been bringing with both capital expenditure (CapEx) and operating expenditure (OpEx) increments due to the deployment of large number of the nodes, planning of their locations, their construction and management etc. For today’s complex networks, various futuristic network solutions are being considered especially in wireless cellular network domain [4]. Self-Organizing Network (SON) concept introduced within 3GPP has emerged which can provide advanced features including self-configuration, self-healing and self-optimization for next generation cellular networks [5]. SON has the potential for enabling easier management, configuration, planning and optimization of the cellular networks while reducing both CapEx and OpEx and providing better quality-of-experience (QoE) and overall improved network performance. Network densification with the introduction of HetNet has brought rapid evolution and integration of SON functions with the purposes of minimizing human intervention in the networking processes and automating mobile infrastructure operation, administration and management (OAM). Therefore, SON is considered to be the driving technology that leads the next stage in the current cellular network evolution towards 5G. Due to different requirements such as energy saving, throughput enhancements, latency reduction, fair load distributions; resource optimization, interference mitigation, mobility management (e.g. for reducing number of handover), caching, fault detection, cell sleep management, cell cooperation etc. are required in the next generation networks. For this reason, clustering, an unsupervised learning approach, that generates isolated group of cells based on interaction among them is considered to be a common solution for challenges of 5G networks [6]. It reduces the complexity of network optimization processes via checking smaller number of cells instead of evaluating entire network.

Main SON functionality can be divided into three categories: (i) self-configuration, (ii) self-optimization, and (iii) self-healing. SON functions in the category of self-configuration enable automatic parameter adjustments such as transmit power, electrical antenna angle, handover management parameters including measurement report triggering event offsets, thresholds, system updates. Additionally, automated network integration of new evolved Node-B (eNodeB) by auto connection and auto configuration, core connectivity (via S1 interface) and automated neighbor site configuration (via X2 interface) can also be considered as some additional actions inside this category. Self-optimization targets achieving the main network level goals by optimizing various aspects such as link quality, coverage and capacity; auto-tune the network with the help of user equipment (UE) and eNodeB measurements on local eNodeB level and/or network management level. Finally, self-healing functions enable network to recover from network related failures such as cell outage, sick cell, sleeping cell, and removal of failures for resiliency purposes. For example, when a cell or eNodeB goes down during disaster scenarios, self-healing feature may enable the traffic to use alternative routes for service continuity.

Together with the expected exponential growth of wireless devices and data rates of 5G networks, the amount of data traffic generated inside telecommunication systems is also expected to increase spectacularly fast. Hence, data analytic techniques such as machine learning will be promising solutions to combat performance inefficiency problems. Some examples can be to reduce unnecessary handovers and increase energy efficiency of Base Stations (BSs) while making wireless cellular network more sustainable. Therefore, the proliferation of new technologies such as machine learning and advanced cellular techniques (e.g. SON) calls for innovative ideas that can jointly analyze, optimize and integrate Mobile Network Operators (MNOs)’ infrastructure data. As such, it is time to develop innovative solutions that consider emerging network intelligence, transform this data surge into knowledge and perform smart self organizing network management.

SECTION II. Related Works and Main Contributions

SON has recently evolved into a very active research field as well as a topic of rapid technological process in both Standard Developing Organizations (SDOs) and telecommunication industries [7]–​[9]. SON for LTE has been proposed by 3GPP with the aim of interference reduction, Coverage and Capacity Optimization (CCO), random access channel (RACH) optimization, Automatic Neighbor Relation (ANR) optimization, Physical Cell Identity (PCI) optimization, Mobility Load Balancing (MLB), Cell Outage Detection and Compensation (CODC), Mobility Robustness Optimizer (MRO), Area Code Optimization (ACO) and Energy Saving (ES) etc. through optimizing configuration management (CM) parameters of the cells and/or eNodeB in the mobile networks (SON functionality and principles and SON use cases in [10] and [11]). Prior SON research addresses the problem of self-healing, self-optimization and self-configuration, i.e. automation of network functionality.

Machine learning approaches, which have evolved as a sub-field of Artifical Intelligence (AI), have also caused tremendous advances in many areas including health, transportation, automotive and telecommunication. In telecommunication domain, machine learning algorithms can substitute the traditional configuration based network operations where analytics on accumulated data can yield better operational efficiency [12]. Current works on SON are focusing on automatizing networks with machine learning techniques such as supervised, unsupervised and reinforcement learning [13]. The success of machine learning enabled different applications of learning models where learning process can be done with labeled or unlabeled data-sets (unsupervised and supervised learning) and trial-and-error based actions with rewards and punishment (reinforcement learning). However, developing mobility-aware machine learning techniques have not received much research focus. Additionally, not so many telecommunication infrastructure are utilizing learning models in MNO’s infrastructure especially for executing SON functionality.

Relevant literature on unsupervised learning approach for SON can be classified into different categories including: operational parameter configuration [14], handover management [15], spectrum learning [16], caching [17], [18], cell outage management [19], [20]. A recent survey of application of machine learning solution in the context of SON is given in [21]. Application of clustering techniques on BSs have shown many benefits in previous works [18], [22]–​[28]. Zhao and Lei [22] have investigated clustering techniques for the purpose of cooperation between BSs in Coordinated Multi-point (CoMP) transmission/reception. Zhang et al. [29] are grouping BSs according to their heterogeneous traffic load along space and time. He et al. [30] have introduced clustering-enabled channel modeling where multi-path channel clustering approach using different clustering algorithms are investigated. Some of the prominent applications of clustering in the literature are: CoMP applications in clustering small cells for interference mitigation purposes [22], increasing offloading or caching gains via clustering mobile users [18], [23], [24], enhancing energy efficiency or channel quality in Device-to-Device (D2D) networks via device clustering [25], [26], increasing system capacity based on user’s content request and mobility pattern predictions in Cloud RAN (C-RAN) by clustering remote radio head (RRH) [27], radio access points clustering for fog computing applications [28], grouping of cells with similar characteristics (e.g. time) so that same configurations can be applied to them [31] and so on.

In summary, current approaches towards realizing SON infrastructure using machine learning concepts are in use in many different research studies. However, considering mobility-awareness while exploiting machine learning capabilities in SON has not attracted much research focus in cellular network design. To the best of our knowledge, none of the previous works in the literature use the relational mobility data such as handover attempts to improve the accuracy of the cell clustering mechanism. While recent advances on machine learning have provided increased opportunities for enabling SON for MNOs, the application of those algorithms on real-world data-sets still remains an open research area. Although extensive research has been established in the field of machine learning and SON, practical results closing the gap between theoretical and practical knowledge still need to be further investigated. This is where working on real-world data becomes useful where analytic algorithms and data generated from network infrastructure can be brought together. In contrast to previous application of different machine learning algorithms into SON, in this paper we propose a mobility-aware cell clustering methodology for cellular network systems. Main of contributions of the paper can be summarized as follows:

• We first discuss the benefits and opportunities of clustering capability on different conventional SON functions including PCI optimization, ANR, ACO, CODC, ES, and MLB and advanced LTE features including carrier aggregation (CA), CoMP, distributed MIMO (D-MIMO), and Single Frequency Network (SFN) after briefly describing their main functionalities.

• We then propose a novel mobility-aware cell clustering methodology which includes a pre-processing step to update cell locations based on relational handover attempts data before performing conventional distance-based clustering. It can decrease the inter-cluster handover attempts while increasing intra-cluster handovers attempts, which aims easier and faster detection of the cell sets in which overlapping serving regions exist. Hence, network optimization phase becomes more effective by evaluating a smaller number of cells (e.g., only intra-cluster or intra and adjacent clusters) rather than the entire network.

• A real world data-set is used for validation purposes of appropriately grouping cells that have both outgoing and incoming handover attempts from neighboring cells. In order to investigate the performance of the algorithm, we define an indicator which is the ratio of total intra-cluster handover attempts to overall handover attempts (including both intra-cluster and inter-cluster). The performance is later evaluated with conventional distance-based clustering algorithm, $k$ – Means, which is utilized as a benchmark. The results reveal that novel clustering algorithm with pre-processing step outperforms the conventional distance-based clustering.

The rest of this paper is organized as follows. In Section III, we present our system architecture including working principles of the clustering algorithms and SON functions & advanced LTE features. In Section IV, we discuss conventional SON functions & advanced LTE features and the benefits of our proposed clustering algorithm over them. In Section V, we explain the proposed algorithm and and we demonstrate the performance of the mobility-aware clustering algorithm in Section VI. Finally, we conclude the paper in Section VII.

Notation: The scalars are represented by regular symbols and vectors/matrices are denoted by bold face regular letters, e.g., for a vector of x, ${\textit {x}}_{i}$ denotes the $i$ -th element of x.

SECTION III. System Model and Architecture

Self-organizing RAN architecture with learning capability including multi-vendor network equipment is depicted in Fig. 1. It includes several macro and small cells which have different number of frequency layers in bottom layer of the figure. Each eNodeB has multiple sectors (generally 3 sectors) and each sector can include different number of carriers (also called as cells). In addition, there are several overlapping areas covered by more than one eNodeBs and their corresponding carriers. Information exchange is done between network management entity and domain management entity, domain management entity and eNodeB and eNodeB-eNodeB though northbound interface (Itf-N), southbound interface (Itf-S) and X2 interface, respectively. The basic idea behind self organizing framework together with proposed clustering mechanism of Fig. 1 is to help MNOs develop advanced SON functionality within their infrastructure. In this architecture, the clustering algorithms described in the following sections and other SON related functions can be deployed in network management entity. This will cater vendor-agnostic centralized SON architecture while exploiting the existing infrastructure data. The centralized information is used to tune the vendor-dependent CM parameters of different LTE features that work under eNodeBs in distributed manner.

FIGURE 1.

Self-organizing RAN architectures with domain management and network management entities.

Fig. 2 shows the aggregated handover attempts with the neighbor cells for different types of environments including outdoor (residential, highway and rural) and indoor (shopping mall) locations. The dashed lines in all sub-figures of Fig. 2 represent the neighbor cells with non-zero incoming handover attempt in Neighbor Relation Table (NRT)¹ of the considered example cell which is denoted by black color. It should be noted that the neighbor cells that do not have any incoming handover attempts from the cell under consideration and other near cells are not displayed. Fig. 2 (a) shows an example cell located in a residential district. This site is in one of the highly active region of the city. There are several handover attempts to neighborhood cells in many directions from the considered cell. However, most of the handover updates were done towards north-west region (which is the serving region) of the considered cell’s location. Note that the north-west direction is matched with azimuth direction of the cell. Fig. 2 (b) illustrates another example eNodeB that serves to outdoor location serving to highway. Compared to residential site, this site has less neighbor cells and coverage and service is more concentrated towards a certain direction, i.e. highway. The neighborhood sites with high mobility handover attempts are observed in western areas, whereas no neighborhood is observed in eastern area of the graph due to cell antenna positioning towards highway. Fig. 2 (c) shows the neighbor list of the example site located in rural outdoor region. In rural environment, there are few neighbors and with less number of handover attempts compared to residential district. Therefore, only one neighborhood site on the south-west direction has excessive handover attempts from the considered site. Finally, Fig. 2 (d) gives another example of a site with omni-directional antennas located in indoor environment, i.e. in shopping mall. In this figure, the neighbor list relation is only with two sites that have non-zero incoming handover attempt, and there are several intra-eNodeB handover attempts compared to inter-eNodeB handovers.

FIGURE 2.

Aggregated Handover Attempts and defined relations of the sectors. (a) Outdoor (Residential) (b) Outdoor (Highway) (c) Outdoor (Rural) (d) Indoor (Shopping Mall).

From Fig. 2, we can observe that the behavior of handover attempts to neighboring sites depends highly on the type of environment (indoor, outdoor) and cell serving region (that depends on different factors such as location, azimuth, antenna angle, transmit power etc.). Mobility-activity extracts the overlapping serving areas and best neighbor relations. Therefore, grouping those handover attempts is expected to be beneficial when designing a network planning strategy.

SECTION IV. Benefits and Opportunities of Mobility-Aware SON Applications

In this section, we discuss about the benefits and opportunities of applying mobility-aware clustering in different SON applications. The main functionality of SON mechanisms discussed in this context are PCI optimization, ANR, ACO, CODC, ES, and MLB.

A. Automatic Cell Identity Optimization

In LTE, mobility is based on PCI that is used to differentiate one cell from the others. It also allows UEs to uniquely identify the source of a received signal. There are total of 504 unique PCIs, which are determined by 3 different primary synchronization signal (PSS) and 168 different secondary synchronization signal (SSS). Similar to LTE, 5G New Radio (NR) extends PCIs to 1008 including 336 different SSS [32].

Motivation: PCIs should be uniquely allocated to the cells in a local area in order to avoid from PCI conflicts, which is due to same PCI usage by at least two different intra-frequency cells² serving in the same coverage area. Using the same PCI on intra-frequency cells that have overlapping serving region reduces the probability of successful synchronization for UEs that are located at this region. PCI confusions are also due to a cell that has at least two intra-frequency outgoing neighbors using the same PCI and it causes handover failures.

Benefits & Opportunities: The local areas are determined based on the overlapping or close coverage areas. The cells that have overlapping or close serving areas can be found from relational handover attempts. Then, the intra-frequency cells serving the same or close areas should not use the same PCI to decrease conflict cases. Additionally, at least two intra-frequency cells that have common neighbor with incoming relation should not use the same identity to avoid from PCI confusions. The last situation depends on directly relational information instead of overlapping or close coverage areas. Clustering mechanism based on handover attempt plays a critical role while determining the local areas. The mechanism extracts the cells that have overlapping or close serving region or common neighbors that causes handover failures due to cell identity confusion. After determining the clusters, the intra-frequency cells within the same cluster have higher priority than the others while allocating unique identity codes in order to maximize the network performance.

B. Automatic Neighbor Relations (ANR)

ANR manages NRTs that include the neighbor cell relations between LTE-LTE, LTE-UMTS and LTE-GSM. The neighbors are defined regarding to different network handover strategies in terms of source and neighbor’s operation frequencies, measurement reports that are sent by UE, and cell planning information.

Motivation: High number of defined neighbor relations leads inefficient cell identity planning and it can block the definition of necessary neighbor relations due to the limitations on NRT. On the other hand, less number of defined neighbor relations causes missing neighbor issues that results in packet drops due to handover or connection failures.

Benefits & Opportunities: It should be noted that two or multiple cells that have not any handover attempt with each other due to not defined relations even if they have overlapping coverages can be assigned to the same cluster due to an intermediate cell that has overlapping coverage area with them. To extract the best real or potential neighbors of a given site, mobility-aware clustering mechanism that uses the handover attempts can be used and neighbor definitions of the cells within same clusters can be prioritized during neighbor addition process. Hence, neighbors can be defined based on their high mobility activities. Similarly, neighbor definitions of the cells within different clusters should be prioritized during neighbor removal process when they have less mobility activities.

C. Area Code Optimization (ACO)

LTE architecture has a multi-tier hierarchical structure to support mobility activities. Cells are located at the lowest level of this hierarchy. LTE has Mobility Management Entity (MME) pool including one or more MMEs and Serving Gateway (S-GW) pool including one or more S-GWs. Those include non-overlapping tracking areas. Each tracking area has a tracking area code (TAC) for identification within a particular network area and UEs obtain a tracking area identifier (TAI) list that includes multiple TACs when they attach to an LTE network.

Motivation: All codes defined above are necessary to track the locations of mobile terminals. Moving to another area triggers a change on those codes and the UE is supposed to send tracking area update message when new TAC is not included by previous TAI list so that it can generate a signaling traffic³ to core network. ACO aims at reducing this signaling traffic from RAN to core network and minimizing mobility activity at the area code borders. There is a trade-off between number of code reuse and generated core network signaling traffic in the network. High number of reusing the same code yields reduced core network signaling traffic. This is due to assigning the same codes to large tracking areas, resulting in less UE-level updates even when UE has mobility. However, this also brings unnecessary paging loads. On the other hand, less number of reusing the same code reduces the paging load, but leads to huge core network signaling traffic and battery consumption at the UE side. Network optimization aims to share the existing codes in an appropriate manner inside the network such that reduction of paging load and core network signaling can be achieved simultaneously.

Benefits & Opportunities: 3GPP Release 16, which is the second phase of 5G, will add support for UE power consumption reduction [1]. Therefore, reducing UE-level power consumption is one of the envisioned significant requirements for 5G deployment. During optimization phase, all codes defined above should be distributed according to mobility activity of cells with their neighborhood cells. When cells that have higher mobility activity use the codes within the same TAI list, even for moving UEs that perform cell (re)selection mechanism, tracking area update messages are not required. This benefits to reduction in core network signaling traffic and UE battery consumption, and also keeps paging load low. Therefore, mobility-aware clustering mechanism plays a key role to balance the trade-off between paging load and core network signaling traffic.

D. Cell Outage Detection and Compensation (CODC)

Hardware and software failures, external failures including power supply and network connectivity, or even mis-configuration decrease the network performance significantly. CODC aims to replace manual tasks for detection of these type problems with autonomous processes including both auto-detection and auto-compensation. It mitigates the degradation on coverage, capacity, and service quality due to unexpected failures in cell or eNodeB level including cell outage, sick cell, sleeping cell etc.

Motivation: CODC minimizes the outage-induced performance effects via tuning control parameters of the cells surrounding the affected cells. In order to perform best compensation actions, the serving region of the compensator cells should be overlapped with the serving region of the problematic cells for offloading.

Benefits & Opportunities: To reduce the cost of search mechanism of finding an optimal compensator cell(s), a pre-defined list can be used. The list includes cells that have high mobility activity and partially overlapping serving areas that guarantee successful offloading process via tuning necessary configuration parameters. In this case, clustering can be used to generate these lists in which cell and potential compensator cell(s) are mapped to each other. Hence, process time of finding optimal compensator can be reduced in a failure situation.

E. Energy Saving (Es)

MNOs deploy new sites or add a new carriers into existing sectors concentrating the capacity of each on a smaller geographical area. This leads to overlapping coverage areas served by multiple sites and carriers. Mobile networks are provisioned and deployed in order to guarantee a specific quality-of-service (QoS) under the consideration of busy hours. However, these busy hours last approximately a couple of hours which the network experiences the highest amount of traffic over a 24–hour span. During non-busy hours which are in majority, the network is significantly under-utilized and power amplifiers and signal processing units still operate even though there is no user traffic. Therefore, it is possible to decrease energy consumption in mobile network infrastructure without any effect on QoS as a consequence of turning off some parts of the network equipment.

Motivation: ES aims to detect under-utilized carriers where turning-off those carriers does not cause any QoS degradation. The cells that will be turned-off should be carefully selected to avoid from QoS degradation. This will also avoid unnecessary inter-frequency handovers increments due to cell turn offs of different frequencies.

Benefits & Opportunities: To prevent inter-frequency handovers, local areas (clusters) including cells that have overlapping or close serving regions can be generated. Therefore, cells operating at same frequencies are turned off in each local area. When local areas are generated based on handover attempt, different frequencies (different number of frequencies as well) can be selected to be turned off in each area, since mobility activity is relatively less across cells in the different local areas.

F. Mobility Load Balancing (MLB)

Load imbalance between overlapping inter-frequency cells is one of the most important problems that MNOs face today due to different path loss coefficients, bandwidth etc. The congested cells cause significant deterioration of system performance due to these unbalanced load distributions. MLB provides traffic load balancing between cells which have coverage overlaps so that congestion problems can be avoided. This can improve network performance by avoiding from load imbalance situations and utilizing radio resources more efficiently. This is achieved by managing cell and relational parameters that handle unexpected overloading situations due to dynamic nature of mobile traffic.

Motivation: Due to instant traffic increments or degradation, MLB has to find the best under-utilized neighbor cell(s) where the traffic of over-utilized cells can be offloaded to.

Benefits & Opportunities: As in CODC’s case, a pre-defined list including cells that have high mobility activity and partially overlapping serving areas (that guarantee successful offloading process via tuning necessary configuration parameters such as a3offset, a5-Threshold1, a5-Threshold2, cellIndividualOffset etc. [33]) can be generated. This can also reduce the cost of search mechanism. In this case, mobility-aware clustering can be used to generate these lists in which cells and their potential neighbor cell(s) are mapped to each other. Therefore, when a cell instantly becomes over-utilized, process time of finding the optimal neighbor cell can be reduced. As a result, traffic can be offloaded to the neighbor cell(s) in a very short-time.

In addition to SON algorithms, mobility-aware clustering mechanism has different benefits and opportunities to several LTE features that are designed to handle interference issues and increase network performance in Heterogeneous network (HetNet) environments. The benefits and opportunities of applying clustering in LTE features including CA, CoMP, D-MIMO, and SFN are explained in detail in the following section.

G. Advanced LTE Features

CA has been introduced with 3GPP Release 10 to deal with 1 Gbit/s downlink peak rate under the consideration of the limited bandwidth of a single chunk of the spectrum. In 3GPP Release 15 specification, up to sixteen NR carriers can be bundled together for CA to increase the available bandwidth to approximately 1 GHz [1]. The main aim is to enhance the performance of UEs. However, the cells that are used in CA mechanism for each UE have to be carefully selected. For instance, serving UE should be located in the overlapping serving region of the cells in order to get good signal quality from each of them.

CoMP, another LTE network feature adopting intra-frequency networking to improve the spectral efficiency, increases received signal power from serving cell and reduce the interference level at UE side with the aim of increasing the throughput of cell edge UEs without decreasing the average cell throughput. The feature coordinates Physical Downlink Shared Channel (PDSCH) and Physical Downlink Control Channel (PDCCH) of multiple cells that have overlapping serving region. CoMP mechanism has multiple modes that have different benefits according to network conditions. Dynamic Point Selection (DPS) and Joint Transmission (JT) are the two of them. In DPS, PDCCH and PDSCH data of the UE are sent by serving cell to the coordinating cell, then forwarded to UE by coordinating cell. DPS increases the capacity of hot-spot cells when the network load is heavy and imbalanced. To get the full benefit from DPS mechanism, source cell should be heavy loaded and coordinating cell should be selected among under-utilized cells. In JT where network load should be light, coordinating cell first receives PDSCH data from serving cell. Cell edge UEs then receive PDCCH data from only serving cell and PDSCH data from both serving and coordinating cells simultaneously. Therefore, selection of cells and CoMP mode play a critical role for performance of CoMP mechanism.

D-MIMO using multiple antennas for data transmission in a coordinated way is benefiting from increasing spatial channel resolution as a result of different spatial locations of antennas. Similar to CoMP mechanism, D-MIMO improves especially the cell edge UE throughput as a consequence of high antenna array gains and interference suppression gains. However, it is useful for UEs that are located overlapping coverage areas of the cells.

SFN includes two or more RRUs and allows multiple physical cells to be combined in a one logical cell. SFN aims to mitigate inter-cell interference issues and reduce number of neighboring cells. Using joint scheduling, received signals from different cells that are interfering with each other in distributed schema becomes multi-path signals arriving from a single centralized node. This increases cell edge throughput in a SFN cell. To get the full benefit from SFN mechanism, the cells within an SFN should be determined under the consideration of overlapping serving regions of the cells as in CoMP and D-MIMO features.

Benefits & Opportunities: To make decision mechanism easier and faster, pre-defined separate CA, CoMP, D-MIMO, and SFN sets can be determined with respect to handover attempts between cells and relation type.⁴ Overlapping regions across the multiple cells that are used in LTE features are searched through within each cluster independently. For instance, in CA, together with many sites and the ability to aggregate up to sixteen NR component carriers (each having a serving cell) for 5G node and UEs, the selection process of component carriers and sites can be huge. Thanks to appropriate mobility based clustering between cells, this search space for finding the appropriate cells for CA can be reduced only within each cluster with performance enhancements. Another example of application of mobility-aware clustering is in the process of selection of the most feasible cells in a SFN cell. During this process, it is better to use cluster of cells that has high intra-cluster handover attempts which is triggered by coverage. This is due to increased overlapping areas between the physical cells that become part of SFN cell center which ensures efficient joint scheduling inside.

SECTION V. Mobility-Aware Clustering Algorithm for SON

In this section, we first describe the conventional distance-based ${k}$ – Means clustering mechanism and later our proposed location update algorithm using relational mobility data which are basically handover attempts among cells. The algorithm is used as pre-processing step before ${k}$ – Means to generate clusters with cells having higher mobility-activity rather than cells with closer distances.

A. Distance-Based Clustering

The goal of ${k}$ – Means algorithm is to find an assignment of given $N$ points, denoted by ${\mathbf {x}}_{n}$ , to $K$ different clusters⁵ using $D$ dimensional ${\mathbf {u}}_{k}$ centers (for $k \in \{1,2\ldots K\}$ ). Here, $n \in \{1,2\ldots N\}$ and each ${\mathbf {x}}_{n} = [x_{1,n} \; x_{2,n }\; \ldots \; x_{D,n}]^{\text {T}}$ has dimension $D$ (or number of attributes). The aim is to minimize the sum of the squares of the distances of each data point to the closest center. The analytic model, which aims to find optimum ${\mathbf {u}}_{k}$ , can be defined in terms of cost function, $J$ as \begin{equation*} J = \sum \limits _{n=1}^{N} \sum \limits _{k=1}^{K} r_{n,k} \left \|{ {\mathbf {x}}_{n} - {\mathbf {u}}_{k} }\right \|^{2}, \tag{1}\end{equation*} where $r_{n,k}$ is defined as \begin{equation*} r_{n,k} = \begin{cases} 1 & {\mathrm {if}} \; k = \underset {j}{{\textrm {argmin}}}\left \|{ {\mathbf {x}}_{n} - {\mathbf {u}}_{j} }\right \|^{2}\\ 0 & {\mathrm {otherwise \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;}}. \end{cases} \tag{2}\end{equation*} In order the minimize $J$ , we first take derivative with respect to ${\mathbf {u}}_{k}$ and set it to zero, which gives \begin{equation*} {\mathbf {u}}_{k} = \frac {\sum \limits _{n=1}^{N}r_{n,k}{\mathbf {x}}_{n}}{\sum \limits _{n=1}^{N}r_{n,k}}. \tag{3}\end{equation*} Under the consideration of (3), calculation of $r_{n,k}$ given in (2) is repeated until $\mathbf {u}_{k}$ given in (3) is not changed. For the rest of paper, we assume two-dimensional points $(D=2)$ where the first dimension denotes the latitude of the cell and the second dimension denotes the longitude of the cell. The cells under the same sector⁶ are further aggregated and the cell set is referred to sector in the following sections.

B. Location Update Based on Mobility Activity

The handover attempts between sectors (or aggregated cells) and their neighbor sectors give an idea about their location. The approximated location can be extracted from weighted average of its neighbors. The weights are the ratio of individual handover attempt between source and destination sectors to the total handover attempt associated with source sector. The analytic formula can be written as \begin{equation*} {\bar {\mathbf {x}}}_{j} = \frac {\sum \limits _{i=1}^{N} {\bar {\mathbf {x}}}_{i} \left ({{\mathrm {\Phi }}_{j \rightarrow i} + {\mathrm {\Phi }}_{i \rightarrow j} }\right)}{\sum \limits _{i=1}^{N} \left ({{\mathrm {\Phi }}_{j \rightarrow i} + {\mathrm {\Phi }}_{i \rightarrow j} }\right)}, \tag{4}\end{equation*} where ${\bar {\mathbf {x}}}_{j}$ is updated location of the source sector and rest of the sectors including neighbors and non-neighbors for $j$ -th sector and ${\mathrm {\Phi }}_{i \rightarrow j}$ ⁷ denotes the number of handover attempt from $i$ to $j$ .

In order to consider the original locations while updating the locations, we add another term (original locations denoted by ${\mathbf {x}}$ ) to (4) by introducing a weighted term, denoted by $\alpha $ , between expected locations based on handover attempts and real locations. The mathematical expression can be written as \begin{equation*} {\bar {\mathbf {x}}}_{j} = \alpha \frac {\sum \limits _{i=1}^{N} {\bar {\mathbf {x}}}_{i} \left ({{\mathrm {\Phi }}_{j \rightarrow i} + {\mathrm {\Phi }}_{i \rightarrow j} }\right)}{\sum \limits _{i=1}^{N} \left ({{\mathrm {\Phi }}_{j \rightarrow i} + {\mathrm {\Phi }}_{i \rightarrow j} }\right)} + \left ({1-\alpha }\right) {{\mathbf {x}}}_{j}. \tag{5}\end{equation*} After locations are updated for all sectors, differences between current and previous locations are calculated. The total difference at each iteration is denoted by ${\Delta }$ . The process is repeated until ${\Delta }$ values are less than a pre-defined threshold, ${\Delta }_{t}$ . A summary of the flow is given in Algorithm 1.⁸

C. Clustering Using Updated Locations

${k}-\text {Means}$ clustering algorithm is run over the updated locations, ${{\bar {\mathbf {x}}}}$ , so that we change $\mathbf {x}_{n}$ variables in (1), (2), and (3) with ${\bar {\mathbf {x}}}$ . Hence, the resulting equations are as follows:\begin{align*} J=&\sum \limits _{n=1}^{N} \sum \limits _{k=1}^{K} r_{n,k} \left \|{ {\bar {\mathbf {x}}}_{n} - {\mathbf {u}}_{k} }\right \|^{2}, \tag{6}\\[-2pt] r_{n,k}=&\begin{cases} 1 & {\mathrm {if}} \; k = \underset {j}{{\textrm {argmin}}}\left \|{ {\bar {\mathbf {x}}}_{n} - {\mathbf {u}}_{j} }\right \|^{2},\\[-2pt] 0 & {\mathrm {otherwise \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;}} \end{cases} \tag{7}\\[-2pt] {\mathbf {u}}_{k}=&\frac {\sum \limits _{n=1}^{N}r_{n,k}{\bar {\mathbf {x}}}_{n}}{\sum \limits _{n=1}^{N}r_{n,k}}. \tag{8}\end{align*} Similar to conventional ${k}-\text {Means}$ algorithm, (7) is repeated until $\mathbf {u}_{k}$ given in (8) is not changed.

D. Computation Complexity and Signalling Load

Note that handover attempt and site plan (e.g., latitude, longitude, sector information etc.) data that are used by proposed algorithm in location update phase are already available in network management entity and utilized as input in most of the centralized SON function. Therefore, collecting this information from existing infrastructure yields no additional traffic on network for clustering purpose. The analysis can be performed offline and appropriate decisions can be fed back to operational units for maximization of network efficiency. In terms of computational complexity, location update Algorithm 1 ’s overall complexity is $O(N*I)$ where $I$ is the number of iterations. It can be observed that the main computation complexity lies in the performance evaluation of (5) for $N$ times in order to compute the updated locations as per line 5 of Algorithm 1. The time complexity of Lloyds algorithm (as well as most variants) for ${k}$ – Means clustering is on the order of $O(K*N*I*D)$ [34]. Note that the time complexity is also sensitive to initial conditions. Therefore, the overall complexity is low and can be further reduced through trade-off analysis between complexity and convergence.

SECTION VI. Numerical Results

In this section, we describe our used mobility data-set and present performance improvement using mobility data which is handover attempt between cells. ${k}$ – Means clustering algorithm which is conventional distance-based clustering without location updates is used as benchmark. To evaluate the validation of the proposed mobility-aware clustering approach, we use a large-scale data-set containing one week of mobility activity for one region. Weekly data is collected from total number of 5732 LTE cells. The particular RAN under consideration includes 3 different frequency layers with the different number of cells. Numerically, first layer includes 2682 cells operating at lower band, second layer includes 2365 cells, and third layer includes 685 cells. There are totally 2848 unique sectors and 363 sectors have omni-directional antenna with relatively less transmit power than sectorial antennas.

The histogram and probability density estimation⁹ of aggregated handover attempts (natural logarithm-based) are depicted in Fig. 3 where natural logarithm of most of the number of handover attempts are between 3.96 and 4.39 with frequency around 10,800 where the highest estimated probability density is 0.18. Fig. 3 shows a Gaussian-like distribution with long right tail and the statistical information our handover attempt data-set (natural logarithm based) is depicted in Table 1.

TABLE 1 Statistical Information of Handover Attempt Data-set

FIGURE 3.

Histogram and probability density estimation of collected natural logarithm based handover attempt.

We consider 20 clusters that give %95 variance explained and 50 clusters that gives %95 variance explained (see Fig. 4). For 20 cluster, original and updated locations with respect to different $\alpha $ values are depicted in Fig. 5 where the dimensions of location matrix is reduced to one using Principal Component Analysis (PCA) for easy demonstration. The results show that increment on $\alpha $ value make the sectors with high handover attempt closer. This is due to prioritization of the first term in (5) that denotes the locations extracted from weighted average of neighbor cells’ locations.

FIGURE 4.

Number of clusters v.s. percentage of variance explained to decide on number of clusters.

FIGURE 5.

Cluster assignments with respect to (a) original locations of the sectors and (b) new locations of the sectors with $\alpha = 0.95$ .

In Figs. 6 and 7, location shifts with respect to the previous iterations and original values are depicted for different $\alpha $ values under the consideration of ${\Delta }_{t} = 10^{-10}$ . When we increase the $\alpha $ value, shifts in each iteration becomes larger (see Fig. 6). This is because of prioritized locations extracted from weighted average of neighbor cells’ locations which yields large shift with respect to original locations (see Fig. 7). However, updating process duration becomes larger as well.

FIGURE 6.

Amount of location shifts with respect to the location in previous iteration.

FIGURE 7.

Amount of location shifts with respect to the original location.

In Fig. 8, we present the ratio of number of intra-cluster handover attempts (which is the summation of the handover attempts between the sectors in the same cluster) to total number of handover attempts including both intra-cluster and inter-cluster (which is the summation of the total number of handover attempts between the sectors) in different clusters under the consideration of 20 number of clusters. $\alpha =0$ is the conventional distance-based clustering without location updates (which is ${k}$ – Means clustering) and it achieves a ratio of 91.53% with the standard deviation of 0.71%. This value is increased to 92.70% with $\alpha $ of 0.50 and increments on $\alpha $ up to 0.95 yields a ratio of 93.13%. The results show that increasing $\alpha $ achieves better intra-cluster to total handover attempts ratio. This behavior is the result of collecting sectors that have higher mobility interaction inside the same cluster. Higher $\alpha $ values gives more priority to the weighted average of neighbor cells’ locations that make the sectors including higher interaction closer. Then, running ${k}$ – Means algorithm over the sectors with updated locations increases the intra-cluster handover attempts significantly. Moreover, since extending cluster size also leads to higher ratio between intra and inter cluster handovers, the standard deviation of the total number of sectors included in clusters is also investigated. Conventional case achieves a standard deviation of 185.9 and it is decreased down to 176.2 with the use of $\alpha =0.95$ . The results reveal that proposed novel algorithm makes the distribution of sectors across different clusters more uniform. It can be concluded that location update process based on handover attempts and neighbor cells’ locations increases the uniformity of the cells over coverage areas. Hence, ${k}$ – Means clustering over updated locations provides more uniform sector distribution.

FIGURE 8.

Ratio of intra-cluster handover attempts to total handover attempts and standard deviation of the number of sectors within each cluster after location updates and clustering with 20 clusters.

In Fig. 9, we present the same ratio under the consideration of 50 clusters which achieve %99 variance explained. $\alpha =0$ , conventional distance-based clustering, achieves the ratio of 82.41% with the standard deviation of 0.60%. This value is increased to 9531 with the standard deviation of 87.51% using $\alpha $ of 0.5. The ratio becomes 87.66%, 87.75%, 87.95%, 88.17%, 88.21%, 88.27%, 88.30%, and 88.33% with the use of $\alpha =0.50$ , $\alpha =0.55$ , $\alpha =0.60$ , $\alpha =0.65$ , $\alpha =0.70$ , $\alpha =0.75$ , $\alpha =0.80$ , $\alpha =0.85$ , and $\alpha =0.90$ , respectively. Increment of $\alpha $ up to 0.95 further yields the ratio of 88.40%. The results show that increasing $\alpha $ achieves better ratio of intra-cluster handover attempt to total handover attempt. When we check the standard deviation of the total number of sectors includes by clusters, conventional case achieves the standard deviation of 68.57 and it is decreased up to 62.83 with the use of $\alpha =0.95$ . The deviations are gradually decreased which leads more homogeneous distribution while increasing $\alpha $ values. Numerically, for $\alpha $ values of 0.50, 0.55, 0.60, 0.65, 0.70, 0.75, 0.80, 0.85, and 0.90, standard deviations become 66.71, 66.15, 65.58, 65.57, 65.09, 65.08, 64.38, 64.22, 63.71, and 63.59, respectively. As in previous case with 20 clusters, proposed location update algorithm increases the intra-cluster handover attempt as a result of collecting sectors that have higher mobility interaction inside the same cluster. Additionally, based on decreasing standard deviation values of cluster sizes as we increase $\alpha $ value, uniformity is increased among number of sectors included by clusters. The results reveal that an increment with the amount of up to 6% in number of the ratio of inter-cluster handover attempts to total handover attempts and a decrement on standard deviation of the number of sectors within the clusters with the ratio of %8.

FIGURE 9.

Ratio of number of intra-cluster handover attempts to total number handover attempts and standard deviation of the number of sectors within each cluster after location updates and clustering with 50 clusters.

In Figs. 10 and 11, average distance values between the sectors within the same clusters are depicted for 20 and 50 clusters, respectively. The conventional distance-based case achieves the average values of 9.861 km for 20 clusters and 3.898 km for 50 clusters. Locations are updated based on calculating weighted average of neighbor cells’ locations and making sectors with more mobility interactions closer to each other. This causes higher distance between cells since mobility activity is included as pre-processing before ${k}$ – Means algorithm. Numerically, for $\alpha $ values of 0.50, 0.55, 0.60, 0.65, 0.70, 0.75, 0.80, 0.85, and 0.90, average distances become 4.411, 4.434, 4.503, 4.554, 4.625, 4.661, 4.690, 4.723, 4.762, and 4.873, respectively, when 50 clusters are considered. For the case of 20 clusters, 9.861 km average distance obtained in conventional distance-based clustering is increased up to 10.004 km with the use of pre-processing step. However, there is no concern about the average distance increment due to the fact that the aim of the clustering is to find the cells that have higher mobility activities rather than closeness.

FIGURE 10.

Average distance values between the sectors within the same clusters with 20 clusters.

FIGURE 11.

Average distance values between the sectors within the same clusters with 50 clusters.

SECTION VII. Conclusion

This paper is studying a novel mobility-aware cell clustering technique and its benefits and opportunities for conventional SON functions including PCI optimization, ANR, ACO, CODC, ES, and MLB and further advanced LTE features such as CA, CoMP, D-MIMO, and SFN. We used the mobility characteristics of real-world data of eNodeBs located in different regions of a selected city. After describing our particular RAN under consideration, we demonstrate that the utilized method can increase the ratio of the intra-cluster attempts to overall handover attempts. The method yields better performance in terms of collecting the nodes that have mobility-activity with each other within the same cluster as compared to traditional network planning methods which use only distance based metrics. Our new mobility-aware method shows an increment with an amount of up to 6% in number of the ratio of inter-cluster handover attempts to total handover attempts. It further yields a decrement on standard deviation of the number of sectors within the clusters with a ratio of %8.