1 Introduction

Fig. 1
figure 1

Heterogeneous network architecture

Full size image

Heterogeneous Networks (HetNets) are now considered a significant architecture of fifth generation (5G) and future generation networks since they have been proven to be a promising paradigm for increasing data rates and user capacity. In HetNets, small cells, also referred to as small base stations (SBSs), are installed within macro base station (MBS) to enhance the coverage areas which are inadequately served (See Fig. 1 for an illustration of a heterogeneous network) [1]. The major differences between 4G HetNets and 5G HetNets are higher data rates, lower latency, increased capacity, enhanced coverage and intelligent network management, among many others. 5G HetNets enable a wide range of new use cases and applications that were not feasible or practical with 4G networks. These include Ultra-Reliable Low-Latency Communications (URLLC), massive Machine-Type Communications (mMTC), and mission-critical applications. Addressing the challenges of enhancing the Quality of Service (QoS), load balance, throughput, and interference mitigation in 5G HetNets are thus more critical. Therefore, many studies have been ongoing to address the problems and meet the standards of 5G and future generation networks [2].

Heterogeneous networks have been a topic of interest because of their role and benefits in establishing Internet of Things (IoT) and smart cities. Researchers explored this field mainly from a resource allocation perspective and introduced novel resource allocation schemes for heterogeneous networks with an overarching aim to help multimedia applications [3]. On the other hand, some other researchers delved into the technical challenges from a load-balancing perspective [4]. This research direction has turned out to be increasingly significant because user association, which is a key load-balancing mechanism, is generally executed before resource allocation. Hence, it significantly affects the network performance. Therefore, inefficient cell association causes load imbalance and wasteful resource distribution among User Equipment (UE), limiting QoS provisioning [5] [6].

One key strategy for advancing 5G success involves implementing HetNets, incorporating Base Stations (BSs) with high transmission power like MBS alongside those with lower transmission power, such as PBSs and Femto Cells. The primary benefits of deploying HetNets include boosting network capacity and enhancing link quality for users who are in closer proximity to BSs [7]. Nevertheless, conventional cellular user association methods tend to result in the majority of users associating with MBS, drawn by their higher transmission power, even when smaller cells like pico or Femto Cells are closer. This conventional association approach leads to a load imbalance, with MBS experiencing overload while SBSs have lighter loads [8]. Along with this, the throughput of cell-edge users of the MBS drops, which often leads to call-drop and cell-edge user being out of service. In addition, lot of other users associated with the MBS also suffer from low data rate leading to lower average network throughput. Thus, it is a struggle not only for users to get the best service but also for the operators maintaining the claim of 5G standards [9]. Despite the existing research on load-balancing for heterogeneous networks, there have not been adequate robust solutions to meet the 5G and beyond user demands, which motivates this study.

This paper is structured as follows. The next section outlines the literature review, followed by the system model in Section III. Next, in Section IV, the proposed algorithm illustrates the proposed load-aware-based biased cell association scheme. Section V outlines the results and discussion. Finally, a conclusion is presented in Section VI.

2 Literature review

In existing research works such as [10], the load carried by each BS is stated as the number of UEs served, or the quantity of resources consumed. Although, backhaul capacity, which is generally the constriction for the load that each BS can handle, is not taken into account. Therefore, the conventional load-balancing technique in [11] is not directly applicable to the practical load-balancing of base stations. Load-balancing of base stations is co-related to the backhaul capacity of the base station. Each base station has its own backhaul capacity; hence, the load-balancing problem arises because of the disparity of backhaul capacity of SBSs and MBS in a heterogeneous network. This leads to a disproportionate distribution of users among base stations. Although [12] addressed load-balancing, the different QoS user requirements have not been analyzed.

Previous research on load-balancing techniques relates mainly to networks consisting of only MBS [13]. However, HetNets comprising of SBSs (such as PBSs and Femto Cells) and MBS are more sensitive to the cell association strategy because of the significant variance in cell sizes. These disparities result in much more unbalanced loads in a signal to interference and noise ratio (SINR) or reference signal received power (RSRP) based cell association, where the mobile user distribution has been assumed to be uniform. So, if the users simply link up with the strongest base station, the difference of load in MBS networks is constrained since all these base stations have equal coverage areas. Therefore, recently some advanced RSRP methods, such as RSRP+biased association and RSRP+power control association have been introduced [14]. RSRP+power control is the method of blanking MBS power at certain instances to allow its associated users to be shifted to the neighboring small base station. This power control allows a better load balance than the general RSRP method. Meanwhile, RSRP+biased association method introduces an additional bias to the RSRP. This bias is generally added to the small base station power, to combat higher MBS transmission power disparity. This method is also named Cell Range Extension (CRE).

The authors in [15] address the crucial role of the handoff process in WiFi networks, particularly in scenarios with numerous Access Points (APs), where even a few seconds of delay can lead to information loss in time-sensitive applications. The paper, therefore, claims to achieve better throughput with reduced packet loss. In [16], an overview of load-balancing along with cell association techniques is mentioned, which proclaims most state of art techniques revolve around RSRP and SINR. In [17], the authors conducted a comparative examination of various user association strategies within HetNets, incorporating biased user association. The paper presents AURA-5G and demonstrates that the optimal solutions enhance the performance of diverse network scenarios in terms of total network throughput and system fairness when compared to the conventional cell association scenario.

Table 1 Qualitative comparison of cell association techniques specifically for load-balancing
Full size table

Cell range extension [18], enabled through appropriate cell biasing, is an effective method to balance the load among macro and small base stations. It is obtained by employing adaptive cell association based on the biased signal strength. This aids in better load-balancing; however, this often degrades the bit rate (throughput) that certain users receive. Hence, one of the critical open problems is how to design a suitable biasing factor [19]. Yasin Aghazadeh et al. proposed a method in [20] that uses estimated received signal strength from different cells and adjusted pilot signals. However, the authors did not include the least visited Femto Cell access points, which probably limits the efficacy of the technique by getting unnecessary hand-offs. They also did not discuss about the complexity and scalability of the proposed technique.

Researchers are exploring the potential advantages of incorporating Cell Range Expansion (CRE) to improve the performance of heterogeneous networks. Numerous studies, including [24] and [21], propose adjusting the CRE area based on the SINR, while others advocate expanding the region through the analysis and updates of the co-ordinated scheduler operation, as outlined in [21]. To enhance HetNet performance, Cell-Selection Offsets (CSOs) can be employed to modify the CRE’s coverage area, as suggested in [25]. An innovative concept presented in [4] involves dynamically adjusting both the CRE region and Almost Blank Sub-frame (ABS) based on the traffic conditions of each cell in the network [9]. Furthermore, researchers have explored various techniques to enhance the user experience in 5G systems, such as interference coordination in massive MIMO (Multiple-Input Multiple-Output) networks during the transition from 4G to 5G [26]. Table 1 illustrates a qualitative comparison of cell association techniques with our proposed technique.

The current study investigates load-balancing of base stations by considering the load of MBS and Small Base Stations (SBSs) and therefore also realising the QoS requirements of users. This paper emphasizes on downlink load-balancing among base stations, along with throughput maximization for 5G and beyond. Therefore, in this paper, a load-aware-based cell association method based on the distance from heterogeneous downlink networks is proposed. The proposed algorithm will allow UEs to associate with the base stations by distance rather than conventional RSRP, which will improve the load balance of the network along with the QoS (in terms of user throughput or bit rate). The study is novel because we can see from the literature that the state-of-the-art techniques are mainly based on RSRP and SINR and rarely consider load-awareness. But as HetNets are proceeding towards ultra-dense networks, a new proposition such as load-aware distance based cell association could be a solution. The simulation results from our experiments in this work add merit to our claim.

Table 2 Parameters used in the simulation
Full size table

3 System model

The simulation has been performed in MATLAB software following the specifications of Release 16, 3GPP TS 24.426 version 16.5.0. We have considered a downlink network containing one MBS overlaid by 3 SBSs. Table 2 depicts the summary of the main parameters used in the simulation. The network can be conceptualized as a multi-tier cellular network, where base stations belonging to the same tier share identical transmission power, density, and coverage area. Furthermore, all tiers are autonomously distributed using the Binomial Point Process (BPP) with a density \(\lambda _k\) for tier k. The BPP model assumes that BSs are deployed randomly and independently of each other, meaning that the location of each BS is chosen independently of the locations of other BSs. This assumption closely resembles real-world deployment scenarios where BS locations are determined without consideration for the locations of neighboring BSs. By assuming random and independent BS deployment, the BPP model provides a simplified yet realistic representation of the spatial distribution of BSs in heterogeneous networks. This allows researchers and engineers to analyze network performance, coverage, and interference in a mathematically tractable manner while capturing key characteristics of real-world deployment scenarios. The reason most researchers use BPP during system modelling is its scalability feature, meaning that it can be applied to networks of varying sizes without significantly increasing computational complexity. For the sake of simplicity in notation, we designate the BS subset as \(S = {1, 2, \ldots , S}\), with the first element representing the MBS. Each BS j is characterized by the tuple \({P_j, B_j, \chi _j}\), wherein \(P_j\) signifies transmit power in equation (1). Assuming independent and identically distributed Rayleigh fading between BS j and user location x, denoted as \(h_j(x)\), the received power at user location x is given by \(P_jh_j(x)(d_j)^{-\alpha _j}(x)\), where \(h_j(x) \sim \text {exp}(1)\), \(\alpha _j\) is the path loss exponent, and \(d_j(x)\) is the distance from user location x to the BS j. Consequently, the received SINR of BS j at user location x can be articulated as,

$$\begin{aligned} \gamma _j(x) = \frac{P_jh_j(x)(d_j)^{-\alpha _j}(x)}{I(x) + N_0} \end{aligned}$$
(1)

Here, I(x) represents the interference emanating from both the identical tier and diverse tier BSs at the user location x, while \(N_0\) stands for the noise power spectral density. Utilizing the provided \(\gamma _j(x)\), we employ the Shannon capacity to characterize the transmit rate in the following manner:

$$\begin{aligned} R_j(x) = W \log _2(1 + \gamma _j(x)) \end{aligned}$$
(2)

For the above, W denotes the bandwidth of base station j.

4 Proposed load-aware biased cell association scheme

Conventionally, the users are associated with the base stations by the RSRP. The more the transmission power of a cell, the more the probability that the user gets associated with it. However, there are path losses, fading and interference that hamper the RSRP signal received by a user. The MBS, being the bigger base station, generally has more users connected to it, while the small base stations, such as PBSs do not get that many users associated with them. This is because the users are associated with the base station with higher signal strength. The load-balancing problem becomes more challenging when each base station in the heterogeneous networks has a different backhaul capacity. For example, MBS has a higher backhaul capacity than PBS. This disparity of BS capacities underutilizes the small base stations by a disproportionate distribution of users among the base stations, creating an imbalance among the load shared by the base stations. Moreover, the MBS’s overloaded users do not get enough resource blocks or bandwidth required for 5G and beyond standards.

Algorithm 1
figure e

Proposed algorithm

Full size image

In this paper, a load-aware based cell association is proposed, as shown in Algorithm 1, where the users connect with the base stations with regards to distance from users to base stations, rather than conventional RSRP. In this scheme, the closer a user is to a BS, the higher the probability of that user being connected to the base station. Therefore, the disparity of power capacities of MBS and SBSs is not a factor anymore. Also, if the percentage of the total number of users connected to a BS reaches a threshold, the furthest cell edge users of that BS is handed over to the nearest BS. The percentage can be set by the operator (in this case by us), and it is understood that at peak hours and non-peak hours, the rush of users can vary. A bias is added to the nearest SBS so that the handover occurs accordingly from the MBS to the SBS, whenever the threshold percentage is crossed for the MBS. The bias value is also decided based on the distance of user being handed over to the nearest BS, the further the user is, the bigger the bias value is needed. The first step of this scheme is to get the location of the user equipment. The BSs are fixed, and so their location coordinates are known. In effect, the more precisely we can locate the users, the more precisely we can calculate the distance of the users from these base stations.

4.1 Location management methodology

Location Areas (LAs) are groups of network cells that users move across, updating the network with their location as per set standards. When a user receives a call, the network pages the cells in the LA to locate the user, a process known as Location Management (LM). The network can reduce paging costs by requiring frequent Location Updates (LUs), but this increases time and energy costs. Conversely, infrequent LUs lower these costs but increase paging costs. LAs can be optimized to reduce hand-offs and speed up location updates. The goal of LM is to balance these factors. Most current LM systems are static, with LUs happening periodically or with each cell change. However, frequent updates occur when users repeatedly switch between two or more LAs, known as the ping-pong effect. Static LAs are the same for all users, uniform in size and fixed. The current static LM standards, GSM (Global System for Mobile Communications) and IS-41 (also known as ANSI-41) have a hierarchical database structure [27]. LUs are executed via one of these criteria:

  1. 1.

    Update location with each BS transition: A location update is initiated whenever a UE relocates to a different BS.

  2. 2.

    Page all cells within the network: When an incoming call needs to be directed to a UE, a page is broadcasted to all cells in the network to determine the cell associated with the UE.

  3. 3.

    Divide the network into paging sub-regions: As a UE moves to a new paging sub-region, it informs the network about the identity of that sub-region. When an incoming call is received for that UE, only the cells within the current sub-region are searched to identify the cell linked to the UE. This entails additional criteria such as optimizing the location update process with reduced traffic load and refining the paging procedure to minimize traffic congestion.

The distance between users and BS is calculated from their respective geographic coordinates, enabling the association of UE with the nearest BS. This methodology incorporates a load-aware mechanism, ensuring that if a specific BS becomes overloaded-defined by surpassing a predetermined threshold of users-the system dynamically reassigns the furthest user to the next nearest BS. This intelligent handover mechanism alleviates congestion and optimizes network performance by distributing the user load more evenly across available BSs. Following the user association phase, resource allocation is executed using conventional techniques such as round-robin scheduling, wherein resource blocks are allocated uniformly among users. This uniformity in resource characteristics ensures the allocation process is consistent across all scenarios examined in this study. By leveraging these strategies, the proposed approach not only aligns with established network management practices but also incorporates advanced load-balancing techniques to enhance efficiency and service quality. This comprehensive method of user association and resource allocation highlights the robustness and scalability of the proposed network management framework in 5G and beyond HetNets. The resource allocation has been carried out similarly for all the schemes discussed in this paper.

4.2 Methodology of adding bias

Primarily, we assume unrestricted access for all base stations, allowing users to connect to any cell within the network coverage area. To balance the load among BSs, we implement a scheme based on distance plus the bias in case of hand off required. In this scheme, users selectively connect to the BS with the highest biased distance. Consequently, the biased distance can be expressed as follows:

$$\begin{aligned} \gamma _j(x) = \gamma _j(x)+B_j \end{aligned}$$
(3)

where \(B_j\) is the biasing factor of BS j. Note that the biasing factor of the MBS is always set to be \(B_1 = 1\), which is the same as the traditional max-SINR cell association. Generally, set \(B_j > 1\) for \(j \ge 2\) to make it more attractive for users. Unlike the static biased scheme, we design our proposed load-aware based scheme to achieve global load-balancing. Considering the non-uniformly distributed traffic, the biasing factor changes along with the load of BS and the traffic arrival rate. Intuitively, the SBSs located in high traffic density areas may reduce the biasing factor to shrink coverage, while the SBSs located in low traffic density areas may increase the biasing factor to attract more users.

Next, we present the definition of the Biasing Association Region (BAR) of BS j. Users located within the coverage will associate with that particular BS j based on the maximum biased SINR. Thus, the covering radius with biasing can be expressed as

$$\begin{aligned} r_{b,j} = B_j^{1/\alpha _j} \cdot r_j \end{aligned}$$
(4)

where \(r_j\) the covering radius without biasing can be obtained from the equation below,

$$\begin{aligned} r_j = \left( \frac{P_jg_j(x)}{\chi _j(I(x)+N_0)}\right) ^{1/\alpha _j} \end{aligned}$$
(5)

Here, \(P_j\) is the transmit power for the BS j, \(g_j(x)\) is the channel gain of BS j from user x, \(I(x)+N_0\) is the addition of interference and noise power respectively. To further clarify, \(\chi _j\) represents the received SINR threshold and \(\alpha _j\) is the path-loss exponent of BS j.

Fig. 2
figure 2

Simulation of system model

Full size image

5 Results and discussion

An MBS of 500 m radius overlaid with 3 PBSs located is considered, as shown in Fig. 2. The MBS is positioned at the center, and the PBSs are placed at random locations, as indicated in Fig. 2. The power transmitted by the PBSs and MBS has been set to 30 dBm and 43dBm, respectively. 500 UEs are considered, which are randomly distributed within the MBS and SBSs, which is again BPP distribution, and their required data rate is randomly set. The key factors to calculate the data rate are the path loss models, transmit power, bandwidth, noise power, and Rayleigh fading. Table 2 shows the values of these parameters used in this study.

To model the overall probability distribution of data rates, the following steps are followed:

  1. 1.

    Simulate user placement: Randomly place 500 users within the MBS and SBS coverage areas.

  2. 2.

    Assign data rates: Calculate the data rate for each user based on their distance to the nearest base station and the path loss model.

  3. 3.

    Fit distribution: Fit a suitable distribution (e.g., log-normal) to the calculated data rates.

The log-normal distribution is used to model data rates in this context, as follows:

$$\begin{aligned} f(x) = \frac{1}{x\sigma \sqrt{2\pi }} \exp \left( -\frac{(\ln x - \mu )^2}{2\sigma ^2}\right) \end{aligned}$$
(6)

where \(\mu \) and \(\sigma \) are parameters estimated from the simulation data. This equation represents the Probability Density Function (PDF) of a log-normal distribution. The variables are defined as follows:

  • f(x): The probability density function of the log-normal distribution evaluated at x.

  • x: The data rate (in this context), which is a positive real number. It represents the variable for which we are calculating the probability density.

  • \(\mu \): The mean of the natural logarithm of the data rates. It is a parameter of the log-normal distribution estimated from the simulation data.

  • \(\sigma \): The standard deviation of the natural logarithm of the data rates. It is another parameter of the log-normal distribution estimated from the simulation data.

  • \(\ln x\): The natural logarithm of x.

  • \(\exp \): The exponential function.

  • \(\sqrt{2\pi }\): The square root of \(2\pi \), a normalizing factor in the PDF of the log-normal distribution.

From the calculation, the data rate range for MBS is 95 Mbps-1 Gbps, and SBS is 51 Mbps-491 Mbps. BPP assumes that users are distributed randomly and independently of each other, which closely resembles real-world deployment scenarios in wireless networks. The required data for each user is also random, but each channel the user uses has limited bandwidth. The channel contains 100 Physical Resource Blocks (PRBs), with each having 180 kHz bandwidth. The path loss models used for PBSs are 140.7 + 36.7 log(d) (dB) and for MBS is 128.1 + 36.7 log(d) (dB), where d is the distance measured in kilometers (kms) between the BS and the UE. The noise spectral density and noise figure are set to -174 dBm/Hz and 9dB, respectively. Zero-mean log-normal shadowing with 10dB standard deviation and a channel with zero-mean unit-variance Rayleigh fading is considered. In addition, the proposed algorithm (distance based load-aware biased cell association) is compared with the maximum RSRP, RSRP+biased, RSRP+power control, max-SINR and EELUA cell association schemes with the same resource allocation mechanism. The simulation results for each scheme are run for 1000 instances, and then the average reading is considered to combat the random nature of the simulation parameters like interference, path-loss and fading. This is important because, in practical scenarios, these parameters are similarly random in nature, and this consideration in our simulation makes the comparison of the results stronger. In each instance, users have been stationary for all the compared schemes as well as the proposed one but located at various random positions to consider the mobility and experiencing different channel conditions.

Fig. 3
figure 3

Performance comparison of proposed scheme against the state-of-the-art techniques over variable number of users

Full size image
Fig. 4
figure 4

Performance comparison of the proposed scheme with state-of-the-art techniques against varying hand-off numbers

Full size image

In Fig. 3, the network’s average throughput [13], is evaluated against different number of users. The proposed load-aware-based association scheme achieves significantly higher throughput because the power disparity is avoided during user association. The power disparity in the RSRP method pushes more users to be associated with MBS rather than SBSs, which results in the throughput being shared among a large number of users in the congested MBS, leading to a relatively low throughput per user. As most users are associated with MBS rather than SBSs in the RSRP scheme, the average throughput of the network drops significantly, unlike the proposed distance-based load-aware scheme. Even the other methods presented in Fig. 3, (e.g., RSRP+biased and RSRP+power [14] control) could not sufficiently enhance the throughput in all circumstances (different number of users). The max-SINR based association method as well as EELUA method performs poorly as the number of users increased.

Load-balancing has been depicted by Jain’s fairness Index (JFI) [13] in Fig. 3, which is implemented to evaluate the fairness of distribution of users among BSs or the load-balancing [18] of the heterogeneous network. It is generally the best way to compare and indicate the efficiency of load distribution per cell versus different numbers of users. The proposed technique is found to outperform the other methods though it follows a similar downward trend as the user number increases, like the other compared schemes. However, it can be observed from Fig. 3 that when the number of users is 300, the throughput performance drops while JFI peaked. This might be due to a sub-optimal bias value. However, the overall throughput and JFI of the proposed technique outperform the other techniques in most instances. The proposed distance-based load-aware scheme provides 10–12% more load balance compared to the other network schemes. The downward trend for a higher number of users might be because of the UEs being randomly distributed (random walk) in the simulations, the number of users associated with some of the base stations increases, leading to a lower throughput per user.

Furthermore, the throughput and JFI is evaluated against variable hand-off numbers in Fig. 4. These findings provide valuable insights into how both the proposed methods and current techniques perform under conditions where users are moving at high speeds, which is an important aspect to consider in real-world scenarios. The trend of load balancing and throughput among BSs versus hand-off number in a HetNet can vary depending on various factors such as network topology, traffic patterns, and hand-off management strategies. It can be observed from Fig. 4 that the proposed technique outperforms the other state-of-the-art techniques in almost all hand-off conditions. While techniques like EELUA or RSRP+biased or RSRP+power control lost load balance with higher hand-off number, the proposed technique has shown stability. Although in the case of throughput performance, the rate dropped for the proposed technique with the increase of hand-offs, but still sufficiently better than the other techniques. Factors such as interference from neighboring cells, signaling overhead associated with hand-offs, and variations in network conditions can influence the throughput-hand-off relationship. Therefore, while there may be variations in the specific trend depending on network configuration and user mobility patterns, the overall trend often reflects an initial improvement followed by stabilization or decline in throughput as hand-off frequency increases.

Fig. 5
figure 5

Complexity analysis

Full size image

It is imperative to assess whether the advancements in throughput and load-balancing come at the expense of heightened computational complexity. Implementing a sophisticated algorithm with potential delays, particularly in the context of cell association, poses practical challenges. Thus, we meticulously compared and analyzed the computational complexity, as illustrated in Fig. 5. Notably, the proposed algorithm demonstrates a significantly shorter duration for user association and handovers between cells compared to the RSRP+biased, RSRP+power control, max-SINR and EELUA algorithms. While the load-aware algorithm requires slightly more time than the conventional RSRP technique, this trade-off is justified by the substantial improvements in throughput and load-balancing. It is easily understandable that a flat RSRP cell association does not consume as much computational time compared to the other techniques which are manipulating various parameters to obtain the necessary gain. We contend that this level of computational complexity is justified, considering the substantial enhancements achieved in the targeted parameters.

6 Conclusion

This study introduces an innovative cell association approach founded on load-awareness, where the association is determined based on proximity rather than the traditional RSRP and its variants. From the state of art, it could be observed that, there is a study gap of load-aware cell association, which simultaneously solves the load-balancing and throughput challenges of 5G and future networks. Therefore, the primary objective of this proposed technique is to enhance the load distribution among base stations while concurrently ensuring QoS provisioning, specifically in terms of throughput, for 5G and future networks. The simulation outcomes demonstrate the superior performance of the load-aware-based technique over existing cell association methodologies, showcasing notable improvements in both throughput and load-balancing.

Nevertheless, it is crucial to acknowledge certain limitations, such as the use of a random user deployment model to simulate mobility, which warrants further investigation for more accurate representations of real-world scenarios. We also believe that in future work, the complexity or the computational time can be reduced by the introduction of proper optimization techniques. Future research endeavors could delve into refining user mobility modeling, and additional considerations might encompass interference mitigation, radio resource management, and bandwidth optimization within heterogeneous network environments.