Handover Management in 5G Vehicular Networks

: Fifth-Generation (5G) vehicular networks support novel services with increased Quality of Service (QoS) requirements. Vehicular users need to be continuously connected to networks that fulﬁl the constraints of their services. Thus, the implementation of optimal Handover (HO) mechanisms for 5G vehicular architectures is deemed necessary. This work describes a scheme for performing HOs in 5G vehicular networks using the functionalities of the Media-Independent Handover (MIH) and Fast Proxy Mobile IPv6 (FPMIP) standards. The scheme supports both predictive and reactive HO scenarios. A velocity and alternative network monitoring process prepares each vehicle for both HO cases. In the case of predictive HO, each time the satisfaction grade of the vehicular user drops below a predeﬁned threshold, the HO is initiated. On the other hand, in the case of reactive HO, the vehicle loses the connectivity with its serving network and connects to the available network that has obtained the higher ranking from the network selection process. Furthermore, the HO implementation is based on an improved version of the FPMIPv6 protocol. For the evaluation of the described methodology, a 5G vehicular network architecture was simulated. In this architecture, multiple network access technologies coexist, while the experimental results showed that the proposed scheme outperformed existing HO methods.


Introduction
Fifth-generation (5G) vehicular networks have emerged rapidly.In a typical 5G vehicular network architecture, vehicular equipment includes On-Board Units (OBUs) [1] and mobile User Equipment (UEs).Vehicles communicate with each other, as well as with access network infrastructures.Access networks usually construct a Fog [2] infrastructure, which interacts with a Cloud environment, offering vehicular services.Many vehicular users can be served from each vehicle, while each user can simultaneously use multiple services with different constraints with each other.
In this network environment, increased communication requirements are arising for the vehicular users and, thus, an increased number of network access resources is needed.To address this issue, the technology of Ultra-Dense Networking (UDN) [3,4] indicates that dense 5G access network infrastructures can be implemented, where an increased number of Points of Access (PoAs) can provide plenty of communication resources to users [5,6].Regarding the operational principles of UDN, several different technologies can coexist in the network access environment.Indicatively, the Macrocells and Femtocells that construct the network access environment can implement technologies such as the IEEE 802.11pWireless Access for Vehicular Environment (WAVE) [7], the 3GPP Long-Term Evolution Advanced (LTE-A) [8], and the IEEE 802.16 Worldwide Interoperability for Microwave Access (WiMAX) [9].Furthermore, regarding the design principles of the 5G Core (5GC) [10], the core network of a 5G vehicular architecture applies virtualization techniques for resource manipulation, including network slicing [11] and mobility management [12].Communication among virtual network functionalities is achieved using RESTful application programming interfaces (RESTful APIs) [13] over HTTPS-/JSON-based protocols [14].As a result, the implementation of novel network manipulation algorithms becomes independent of the specific underlying hardware optimizing the support for heterogeneous network technologies.
The co-existence of heterogeneous wireless access technologies indicates the need for efficient manipulation of the vehicles' mobility.This need has involved the employment of the IEEE Media Independent Handover (MIH) and the Fast Proxy Mobile IPv6 (FPMIPv6) standards [15] for the mobility transfer.Specifically, the MIH protocol defines HO operations across different access networks.Its architecture provides generic interfaces and primitives independent of the type of vehicular equipment and access networks, supporting link-layer events and network information retrieval.On the other hand, two reference HO scenarios are introduced in FPMIPv6, the predictive and the reactive scenario, depending on the vehicle's speed.In the predictive scenario, the vehicle before disconnecting from its previous network sends a handover indication report to its Mobility Access Gateway (MAG) [16] declaring the new network, which is selected as the target for performing the HO.In case the vehicle connects to the new network without informing of its MAG in the previous network, the reactive operation of FPMIPv6 is performed.
Vehicles should always be connected to the most appropriate access network to fulfil the QoS constraints of their services.To address this issue, Multi-Attribute Decision-Making (MADM) algorithms can be applied to select the best alternative among candidate networks by taking into consideration a set of evaluation criteria.Well-known methodologies include the Analytic Hierarchy Process (AHP) [17], the Analytic Network Process (ANP) [18], Simple Additive Weighting (SAW) [17,19], Fuzzy SAW (FSAW) [20], Multiplicative Exponential Weighting (MEW) [17] and the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) [21].
This paper describes a methodology for supporting Handovers (HOs) on 5G vehicular networks.Both predictive and reactive HO modes are supported.During the vehicles movement, a velocity and alternative network monitoring process executes.This process applies the Mobility State Estimation (MSE) technique of the LTE [22], which was improved in [23], in order to facilitate the decision process by taking into consideration that both Macrocells and Femtocells can serve the territory of each vehicle.Thus, the vehicle's velocity is continuously monitored and characterized as Normal, Medium, or High, considering the number of HOs performed by the vehicle during a sliding time window [23].Furthermore, it ranks the candidate networks by taking into consideration the vehicle's velocity, network characteristic criteria, provider policy criteria, and energy-aware criteria.Specifically, if the vehicle's velocity is Normal, all the networks, including both Macrocells and Femtocells, are considered as network alternatives.On the other hand, if the vehicle's velocity is Medium, some Femtocells along the vehicle's trajectory are skipped, in order for the HO rate to be reduced.Finally, if the vehicle's velocity is High, only the available Macrocells are considered as network alternatives.Moreover, in the predictive HO scenario, each time that the grade of the user satisfaction drops below a predefined threshold, the HO is initiated and the process of the network selection is performed.Correspondingly, in the reactive HO operation, whenever the vehicular user loses the connectivity with his/her serving network, it connects to the available network that obtained the higher ranking during the network selection process.
The discussed scheme implements several characteristics, including: • The HO management services, namely the velocity and alternative network monitoring, the HO initiation, the network selection, and the HO execution are performed at the Fog and the Cloud infrastructures to reduce the workload at the vehicle; • Both predictive and reactive HOs modes are supported for the mobility transfer; • The vehicle's velocity is considered in order for unnecessary HOs to short-range cells (such as Femtocells) to be avoided, since a vehicle that moves with high velocity will remain for a short time period inside their coverage area; • HO initiation is performed by taking into consideration the user satisfaction, which is estimated using both the Signal-to-Noise-plus-Interference (SINR) and the perceived Quality of Service (QoS) parameters, due to the fact that users that perceive a good SINR may not be satisfied by the QoS that they also perceive from their current networks; • Network selection takes into consideration contradictory criteria for different application types and users' SLAs to satisfy the requirements of demanding services; • The HO scheme applies an improved version of the MIH framework and the FP-MIPv6 protocol.
The remainder of the paper is organized as follows: In Section 2, the related research literature is revised.Section 3 describes the proposed HO scheme, and Section 4 presents the implemented signaling process.Subsequently, Section 5 presents the simulation setup and the experimental results.Finally, Section 6 concludes the discussed work.

Related Work
In general, the HO process usually consists of the HO initiation, the network selection, and the HO execution subprocesses.Several HO schemes have been proposed in the research literature supporting the above network management functions.
In [24], the authors introduced the necessity of Quality of Service (QoS) and Quality of Experience (QoE) parameters in mobility management algorithms.In this approach, a QoE algorithm for video streaming in 5G vehicular networks (HoVe) is proposed in order to improve the overall handover performance.The Analytic Hierarchy Process (AHP) technique is used to identify the importance of the current metrics including radio resources, vehicle's mobility, QoE.etc.The Structural Similarity (SSIM) is evaluated as a QoE metric estimating the corruption between the original and the transmitted video.The proposed algorithm outperforms the standard LTE RSSI-based algorithm [25] for video services.
In [26], the authors proposed an HO triggering scheme based on dynamic criteria.An important aspect of the handover process is the momentum of the trigger event, which initiates the HO.An imprecise trigger could produce service interruptions, data loss, latency, and radio-link failure.The proposed scheme benefits from the IEEE 802.21 Media Independent Handover (MIH) and its services, introducing the link going down event.A dynamic value is calculated to adjust the trigger point during the handover delivery time according to the mobile node's current speed.
In [27], the authors presented a machine learning technique to predict the necessity of a handover so as to maintain the users' QoE.The proposed method makes use of a support vector regression algorithm (SVR) to evaluate an HO score periodically per mobile node.The algorithm obtains the signal strength, the throughput, and the previous handover score of the node as the input criteria.The mobile node is triggered to perform the HO to a new network in case its handover cost is above a threshold specified by the system policy.
In [28], an MIH-enhanced FPMIPv6 methodology was proposed.In particular, the functionalities of the Evolved Packet Core (EPC) [29] were considered, while a scheme that implements the IEEE 802.11MIH protocol in the EPC architecture for performing seamless HO was described.Additionally, the use of a transient flag was proposed for eliminating both the ping-pong effect and false HOs by retaining the initial and transient bindings at the LMA.
In [30], the Predictive Handover Mechanism for Video streaming in cloud-based urban VANET (PHMVV) method was proposed considering a list of alternative Points of Access (PoAs).The PHMVV takes into account parameters such as the velocity of the vehicle, its distance from each PoA, its moving direction, and the link quality of each PoA.Accordingly, when the vehicle approaches an alternative PoA with better link quality than that of its current PoA, the vehicle performs an HO to it.
In [31], the Mobility-Aware Handover scheme for Smart Cities Environment (MAH-SCE) was described.In particular, road traffic conditions were gathered from a Smart City [32] environment using sensors.An HO initiation mechanism initiates an HO considering the road traffic, as well as the SNR that the vehicle perceives from its current PoA.Additionally, in order to reduce the HO rate, Femtocells are not considered as network alternatives when the vehicle's velocity is above 16 m/s.However, it has to be noted that the authors did not provide an adequate explanation about the selection of this specific threshold value.However, an adequate explanation about the selection of this specific threshold value was provided in [31].Therefore, in our work, the vehicle's velocity was also monitored and a different approach to this task was implemented by considering the operation principles of the LTE MSE technique [22].
In [33], the Mobility-Aware Handover in Ultra-Dense Vehicular Environment (MAH-UDVE) method was proposed.The direction to which each vehicle moves is considered for the most appropriate PoA to be selected.Specifically, the vehicle's mobility is analyzed in order for a set of network alternatives to be created.Regarding the target network to perform the HO, three options were assessed including the network cell, the next cell in the same or opposite direction of the vehicle, and the cell with the lowest load.
In [34], the Context-Aware Load Balancing (HO-CALB) mobility management scheme was proposed.HO-CALB is applied to a network architecture where both WiFi and WiMAX networks coexist.When the QoS that each user perceives drops below a predefined threshold, an HO will be initiated.Thereafter, during the network selection, the data traffic load is distributed to WiMAX and WiFi PoAs by taking into consideration the available bandwidth of each one.
In [35], the Velocity-Aware Handover (VAH) technique was described.Two network tiers are defined.The first tier consists of Macrocells, while the second tier contains Femtocells.Furthermore, four HO mechanisms are employed.The first one is called the Best Connected (BC)mechanism, and it is applied to users with Low mobility or static users.In this case, all the network cells available in the area of each user, including both Macrocells and Femtocells, are considered as candidate networks.The user selects the nearest Macrocell if otherwise the nearest Femtocell is selected as the target network for the HO.The parameter P represents the transmission power.Furthermore, the parameter B represents the bias factor.R represents the user distance, and η is the path loss exponent for each tier.The second tier is called the Femto Skipping (FS) mechanism, and it is used by vehicles with Medium mobility.In this case, some Femtocells are skipped to reduce the HO rate, when Furthermore, the third tier is called the Femto Disregard (FD) mechanism, and it is applied to vehicular users with High mobility.In this case, all the Femtocells are skipped.Finally, the fourth tier is called the Macro Skipping (MS) mechanism and is applied to users with Extremely High mobility, while in this case, all the Femtocell,s as well as some of the Macrocells are not considered as alternative networks.

Preliminaries
The proposed HO management scheme uses Interval-Valued Icosagonal Fuzzy Numbers (IVIFNs) for the representation of the criteria values considered in both the HO initiation and network selection processes.For the creation of the IVIFNs, the Equalized Universe Method (EUM) is applied.Thus, the created fuzzy numbers are equally distributed inside the predefined domain [0, 1].Additionally, the criteria weights for both HO initiation and network selection are estimated using the Icosagonal Fuzzy Analytic Network Process (IFANP) method.
During the HO initiation, the Mamdani Icosagonal Fuzzy Inference System (MIFIS) is used to decide when an HO must be initiated.Furthermore, during the network selection, the most appropriate network is selected by applying the Dynamic Icosagonal Fuzzy TOPSIS with Adaptive Criteria Weights (DIFT-ACW) method.In this work, criteria such as the vehicle's velocity, the QoS constraints of vehicular user services, the user SLA, and the remaining energy of the user equipment are considered for the network selection.Furthermore, in special cases (e.g., emergency situations), the severity level of a natural disaster is also considered.
The following subsections describe the aforementioned methods, while the design of the discussed mobility management scheme is described in Section 4.

The Interval-Valued Icosagonal Fuzzy Numbers
Fuzzy logic can be used to solve problems by taking into consideration approximate information, while a fuzzy number is represented by a set of real values indicating an uncertain quantity.In this work, icosagonal fuzzy numbers were used, providing increased computational accuracy in comparison with the aforementioned ones.An icosagonal fuzzy number can be defined as a vector x = (x 1 , x 2 , x 3 , x 4 , x 5 , x 6 , x 7 , x 8 , x 9 , x 10 , x 11 , x 12 , x 13 , x 14 , x 15 , x 16 , where where the A L parameter denotes the lower fuzzy number and the parameter A U denotes the upper one [36].IVFNs replace the crisp numbers by intervals in [0, 1].In general, they are useful for solving Multiple Attribute Decision-Making (MADM) problems due to the fact that fuzzy information is better expressed using intervals than using crisp values.Specifically, an IVFS A is defined as follows [37]: The Interval-Valued Icosagonal Fuzzy Number (IVIFN) [38] is the most general IVFN case (Figure 1), defined as: The form of an interval-valued icosagonal fuzzy number.

The Equalized Universe Method
In this work, the Equalized Universe Method (EUM) [39] was applied for the creation of IVIFNs by distributing their centroids equally along a predefined domain of values.In particular, the values of the ith IVIFN are calculated using Formula (6), where d i represents the centroid of the ith IVIFN, calculated using Formula (7).Furthermore, p U is the distance between consecutive intervals of the upper icosagon calculated using Formula (8) where z U is the width of the upper icosagon.Specifically, z U is calculated using Formula (9), where c represents the number of IVFNs and θ ∈ [0, c] is a constant factor for configuring the width of the icosagon.Similarly, p L is the distance between consecutive intervals of the lower icosagon calculated using Formula (10) where z L is the width of the lower icosagon calculated using Formula (11).Finally, the , and v U A8 parameters are defined by the user.

The Icosagonal Fuzzy Analytic Network Process
The Icosagonal Fuzzy Analytic Network Process (IFANP) method implements the Analytic Network Process (ANP) using IVIFNs, which usually is applied for the estimation of the weights of criteria considered in MADM problems.In particular, for the estimation of criteria weights, the considered criteria are organized into clusters.The relations between criteria that belong to the same cluster are called "inner dependences", while the relations between criteria that belong to different clusters are called "outer dependences".Indicatively, in this work, three clusters of criteria were considered.The first one contains the network characteristics' criteria such as throughput, latency, jitter, and packet loss.Correspondingly, the second cluster contains operator policy criteria such as service reliability, security, and monetary cost, while the third cluster contains energy-aware criteria such as the energy consumption that occurs for a user equipment that interacts with the network.
Regarding the implementation of the IFANP method, a fuzzy pairwise comparison matrix Ã is constructed for each cluster of criteria.The form of this matrix is presented in Formula (12), where n represents the number of criteria that belong to the corresponding cluster.Furthermore, Figure 2 presents the linguistic terms that can be used for the representation of the relative importance of each criterion in comparison with each one of the other criteria.It has to be noted that the aforementioned linguistic terms were created using the EUM method.Ã = Furthermore, Formula ( 13) is applied for the calculation of the geometric mean r Ãi of each row i in the pairwise comparison matrix Ã.In this formula, the ⊗ operator represents the multiplication of two IVIFNs.
Following this, Formula ( 14) is applied for the construction of the priority vector Ωi of each criterion of the cluster.
In this priority vector, i8 )] is calculated using Formula (15).In this formula, the ⊕ operator represents the addition operator of two IVIFNS.
Subsequently, the fuzzy supermatrix W is created.It represents both the inner and outer dependencies of the considered clusters of criteria.In particular, the supermatrix W is a partitioned matrix, where each matrix segment represents the relationship between two clusters of criteria.For the construction of the supermatrix, the local priority vectors Ω are grouped and placed in the appropriate positions in the supermatrix by taking into consideration the relative importance of each cluster in comparison with each one of the other clusters.Indicatively, by assuming that there are q clusters available, with each cluster C k , k = [1, 2, . . ., q] containing n k criteria, denoted as e k1 , e k2 , . . ., e kn k , then the supermatrix is expressed as presented in Formula (16) Thereafter, the Weighted Supermatrix W is created by applying Formula (17).
Then, the Weighted Average method is applied to defuzzify the Weighted Supermatrix.This method applies Formula (18), where v represents the height and d represents the centroid of each W k,j icosagon, respectively.In particular, W is raised to limiting powers to produce the Limited Supermatrix, where all the columns have the same values indicating the global priority of each criterion.Thus, the overall criteria weights are estimated.

The Mamdani Icosagonal Fuzzy Inference System
The Mamdani Icosagonal Fuzzy Inference System (MIFIS) takes into consideration two inputs (Input 1 and Input 2 ) to calculate the value of the Output using fuzzy logic.The values of both Input 1 and Input 2 are normalized within the range [0, 1].Additionally, the MF Input 1 , MF Input 2 , MF Output Membership Functions (MFs) are created, indicating the linguistic terms and the corresponding IVIFNs for the fuzzy representation of Input 1 , Input 2 , and Output, respectively.As a result, for each crisp value, two membership degrees are determined in the respective MF, one for the upper icosagon and one for the lower icosagon.
Regarding the implementation of the MIFIS method, initially, the Membership Functions (MFs) for the Input 1 , Input 2 , and Output parameters are determined.Furthermore, a knowledge base that contains R fuzzy rules is created.In this base, each rule r ∈ R is a simple if-then statement with a condition and a conclusion.The rule's condition consists of MF Input 1 and MF Input 2 membership functions, while its conclusion indicates an MF Output membership function.Subsequently, the Input 1 and Input 2 crisp values are converted to degrees of membership indicated as Input 1 and Input 2 by a lookup in the MF Input 1 and MF Input 2 membership functions, respectively.Thereafter, fuzzy operations are performed in order for the fuzzified inputs to be combined.In particular, the rule's condition is considered a Z r strength of the r rule to be calculated.In addition, the fuzzy operators "AND" and "OR" can be used in multiple conditions to be combined.In this case, the "algebraic product" is applied for the "AND" operator, while the "algebraic sum" is applied for the "OR" operator.In this work, Formula (19) was used for the estimation of the "algebraic product", while Formula (20) was used for the estimations of the "algebraic sum".
Then, the consequence MF Output c r of the rule conclusion is estimated by applying an implication method.Specifically, the implication method trims the MF Output r height by taking into consideration the Z r strength, by applying Formula (21), which implements the Min method.
MF Output c r Height Subsequently, the R rules' consequences are combined by applying an aggregation method in the Output A fuzzy set to be calculated.In particular, the aggregation method uses Formula (22), which implements the Max method.
Finally, defuzzification is performed.Specifically, the Output A fuzzy set is transformed to the crisp value Output.In this case, the Weighted Average method is applied using Formula (23).In this formula, the µ r represents the height of each rule and h r represents the centroid of each rule, as they are obtained from Output A .Moreover, the parameters U and L indicate the upper and the lower icosagon of each rule, respectively.

The Dynamic Icosagonal Fuzzy TOPSIS with Adaptive Criteria Weights
The Dynamic Icosagonal Fuzzy TOPSIS with Adaptive Criteria Weights (DIFT-ACW) algorithm improves the Icosagonal Fuzzy TOPSIS (IFT) methodology proposed in [40] by embedding the characteristics of the Dynamic Trapezoidal Fuzzy TOPSIS with Adaptive Criteria Weights (DTFT-ACW) algorithm described in [41].Thus, the DIFT-ACW adapts its decisions according to specific factors such as the battery level of the user equipment.Regarding the implementations of the DIFT-ACW algorithm, suppose C = {C 1 , C 2 , . . ., C m } is the set of criteria and w 1 , w2, . . ., w m are the weights of each criterion.The criteria weights are adapted by taking into consideration the fact that some services could be more important than the others.In particular, the relative importance of each service q is determined using a priority vector Ω, which is calculated using Formula (24) where each Furthermore, at each time t, a list In particular, initially, each x ij (t) element of the n × m decision matrix D(t) is an IVIFN, which expresses the performance of alternative i for the criterion j, as presented to Formula (25). where: In case there are Q services, the decision matrix includes the average of the performance values by applying Formula (26).Correspondingly, in this case, the criteria weights are obtained as the average of the individual weights of the considered services by applying Formula (27).
Furthermore, for each network selection performed in time t, if D(t) = D(t prev ), then the ranks of the candidate networks obtained at time t prev are considered.Thus, the proposed algorithm achieves O(n) complexity in this case.Otherwise, the methods defined by the DIFT-ACW algorithm are performed.In this case, the algorithm achieves O(n 2 ) complexity, which is also introduced by the most MADM methods proposed in the literature.Thereafter, the decision matrix is normalized.To accomplish this task, the algorithm considers that Ω b is the set of benefit attributes and Ω c is the set of cost attributes.Subsequently, Formula ( 28) is applied for the estimation of the elements of the normalized decision matrix for the case of benefit attributes where b j (t) = max i x U ij (t) for each j ∈ Ω b .Correspondingly, Formula ( 29) is applied for the estimation of the elements of the normalized decision matrix for the case of cost attributes where c j (t) Subsequently, each element of the normalized decision matrix r ij (t) is multiplied with the respective weight w j (t) according to Formula (30) in order for the weighted normalized decision matrix to be created.
Then, the positive ideal solution is determined using Formula (31), where i Similarly, the negative ideal solutions is calculated using Formula (32), where i Following this, Formulas (33) and (34) are applied in order for the distances of each candidate network from the positive ideal solution to be calculated.Similarly, (35) and (36) are applied in order for the distances of each candidate network from the negative ideal solution to be calculated.
Thereafter, in a way similar to [37], the distances of the candidate networks from the positive and negative ideal solutions are expressed using intervals such as Finally, the relative closeness of the distances from the ideal solutions are computed using Formulas (37) and (38). and: The compound relative closeness is obtained by calculating the average of the RC i1 (t) and RC i2 (t) values using Formula (39).
The candidate networks are ranked by taking into consideration their RC i values, and the one with the higher RC i value is considered as the best alternative.

The Proposed Mobility Management Scheme
During vehicles' movement, the velocity and network monitoring process is executed, enabling the effective transmission of each vehicle to the target access network in the case of an HO event.The HO process includes the HO initiation, the network selection, and the HO execution procedures, while both predictive and reactive HO operations are supported.In the case of predictive HO, HO initiation, network selection, and HO execution are performed, while in the case of reactive HO, the HO initiation is not performed since the link between the vehicle and its previous serving network goes down unexpectedly without an indication of a potential connection loss.Figure 3 presents the workflow of the proposed methodology.The design of the discussed scheme was optimized to be applied in 5G network architectures where both Fog and Cloud infrastructures are available, while the implemented signaling combines messages of both the IEEE 802.21 Media Independent Handover (MIH) [42] and the IETF Proxy Mobile IPv6 (PMIPv6) [43] standards.Furthermore, it should be noted that the proposed scheme can be used for the mobility management of both pedestrian and vehicular users.More specifically, handling the mobility issues in dynamic vehicular environments is of paramount importance, since the vehicles may move with increased velocities and provide heterogeneous services to multiple onboard users.Thus, this paper proposes and evaluates an HO scheme that can be adopted in a challenging vehicular network.

Velocity and Alternative Network Monitoring
The Cloud/Fog infrastructure continuously monitors the velocity of each vehicle by applying the methodology described in [23].Specifically, an improved version of the Mobility State Estimation (MSE) technique defined in 3GPP LTE Release 11 [22] was applied.This methodology categorizes the moving velocity of the vehicle as Normal, Medium, or High, by taking into consideration the number of HOs N CR u that the vehicle has performed during a sliding time window T CRmax .
Additionally, each vehicle maintains a set A = {A 1 , A 2 , . . ., A n } of candidate networks based on the SINR perceived from each one of them as specified by the Cloud/Fog environment.The set A is continuously updated, since the SINR of each network is constantly changing as the vehicle moves inside the wireless network access environment.Thereafter, in the case of predictive HO, the HO initiation will be performed, while in the case of reactive HO, the network selection is immediately executed.

HO Initiation
HO initiation is performed only in the case of predictive HO.During the HO initiation, the methodology described in [23] is applied.Specifically, the S u,i indicator is defined, determining the satisfaction grade of user u from his/her current network i.The value of the S u,i indicator is calculated using the MIFIS algorithm where the normalized values of the SINR and the QoS the vehicle perceives from its current network are used as the inputs.In this work, the MIFIS was implemented using the IVIFNs presented in Figures 4-6.
During the instantiation of the system, the Satisfaction Indicators Chart (SIC) presented in Figure 7 is created at the Fog by applying the knowledge base described in [44].The SIC chart presents the S u,i values obtained for each possible SINR and QoS combination.When either the SINR or the QoS drops below a predefined threshold, the vehicle sends the obtained values to the MAG of its current network using an MI H_Link_Going_down [45] message.Subsequently, the Fog infrastructure where the current MAG belongs uses the SIC chart, in order for the S u,i satisfaction grade of the user to be determined.Subsequently, the network selection process will be performed if the estimated satisfaction grade is less than a predefined S th threshold value.

Network Selection
During the network selection (Figure 8), the vehicle sends a 1.MI H_Get_In f ormation_-Request to the MIIS.This message embeds information about the services and the remaining energy of the vehicle.In the case of predictive HO, the message is transmitted by the current network of the vehicle.However, in the case of reactive HO, the vehicle chooses the network alternative with the higher SINR to transmit the message, since the connection with the current network has been lost.Subsequently, the velocity and network monitoring process considers the mobility state of each vehicle to create a subset A ⊂ A of the available networks.Specifically, if the vehicle's velocity has been characterized as Normal, then all the available networks are considered as alternatives, including both Macrocells and Femtocells in the A subset.On the other hand, if the vehicle's velocity has been characterized as Medium, then some Femtocells are skipped along the vehicle's trajectory and are not included in A .In particular, the Femtocell i is considered as an alternative if the estimated residence time of the vehicle u (t residence u,i ) is greater than or equal to the average residence time of the vehicle considering all the available Femtocells in its current location.It has to be noted that the t residence u,i parameter of vehicle u for the ith cell is estimated using Formula (40) [46], where r f is the cell radius and v u is the vehicle velocity obtained by Formula (41) [47].The λ u parameter represents the cell's density in the location of the vehicle, as it is obtained by the Software-Defined Networking (SDN) controller.Finally, if vehicle's velocity has been characterized as High, then only the available Macrocells are included in the A as alternatives.
Following this, the availability of the resources offered from each network of the subset A is considered, so that a subset A ⊂ A is created.Specifically, the MIIS checks the availability of the resources offered from each network alternative by exchanging 2.MI H_N2N_HO_Query_Resource_Request and 3.MI H_N2N_HO_Query_Resource_Response messages.Thereafter, networks included in the subset A" are ranked using the DIFT-ACW algorithm.The ranking is performed considering both network characteristics (throughput, latency, jitter, packet loss, and energy consumption) and the operator's policy criteria (service reliability, security, and monetary cost) that each network offers for the services of the vehicle.Since the energy consumption parameter is considered during the ranking process, energy-intensive networks such as LTE-A Pro FD-MIMO, which operates with multiple antennas, obtain a low score for vehicles with low remaining energy of the user equipment.Furthermore, since the decision is made considering the vehicle's velocity, as well as QoS and policy-related parameters, the ping-pong HO effect is eliminated.Following this, the network that obtains the higher score is selected, while the vehicle is informed about the target access network through a 4.MI H_Get_In f ormation_Response message.

HO Execution
An improved scheme for the HO execution is proposed based on the FPMIPv6 protocol.Furthermore, if the selected and the current network belong to different MAGs, inter-MAG HO execution is performed.On the contrary, if the two networks belong to the same MAG, intra-MAG HO execution is performed.

Proposed Predictive HO Scheme
In the case of predictive inter-MAG HO execution (Figure 9), the current network initially sends a 1.MI H_N2N_HO_Commit_request to the selected MAG to reserve the necessary resources for the upcoming handover.Following this, the selected MAG responds with a 2.MI H_N2N_HO_Commit_response message indicating that the required resources have been reserved.In this case, the current MAG informs the vehicle about the successful resource allocation at the selected network using a 3.MI H_Net_HO_Commit_request message, while the vehicle acknowledges the reception of the information using a 4.MI H_Net_-HO_Commit_response message.
Subsequently, the MAG of the current network sends a 5.Handover_Initiate message to the MAG of the selected one, which includes the MN Identifier (MN-ID), the Home Network Prefix (HNP), and the MN Logical Link Identifier (MN-LLID).Furthermore, the Proxy (P) and the Buffering (B) flags of the message are set equal to 1 to inform the selected MAG that the buffering of data has been performed at the current MAG.By receiving this message, the MAG of the selected network replies with a 6.Handover_Ack message, which acknowledges that the aforementioned information has been successfully received, while a bi-directional tunnel is established between the two MAGs.Thereafter, a copy of the buffered packets from the current MAG is transmitted through the tunnel to the selected MAG to be delivered to the vehicle when it connects to the selected network.Subsequently, the new MAG sends a 7.PBU message to the LMA to create a transient record in the binding cache.Once the temporal binding is established at the LMA, it sends a 8.PBA message to the MAG of the selected network, and a bi-directional tunnel is established between the two entities.Thus, downlink packets are forwarded from the LMA to the MAG of the selected network where they are buffered to be delivered to the vehicle when it completes its HO to the selected network.Furthermore, the bi-directional tunnel between the current MAG and the LMA remains active until the transient binding becomes permanent.Then, the transient binding introduces a smooth way for performing the handover by gradually changing the current MAG to avoid packet losses and false HOs.
Subsequently, the vehicle sends a 9.MI H_Link_up message to the selected MAG to enable the communication channel.Furthermore, the vehicle sends a 10.UN A message to the selected network, to complete the establishment of both the Link and Network layers.Thereafter, the MAG of the selected network exchanges the 11.PBU and 12.PBA messages with the LMA to finalize the binding.Finally, the selected network sends a 13.MI H_N2N_Complete_request message to the previous network.By receiving this message, the previous network releases the resources that have been reserved for the vehicle and sends back a 14.MI H_N2N_Complete_response message to finalize the HO execution.Similar to the predictive inter-MAG HO execution, in the case of the predictive intra-MAG HO execution (Figure 10), the current network sends a 1.MI H_N2N_HO_Commit_request to the selected network to reserve the necessary resources for the upcoming handover.Following this, the selected network responds with a 2.MI H_N2N_HO_Commit_response message indicating that the required resources have been reserved.In this case, the current network informs the vehicle about the successful resource allocation at the selected network using a 3.MI H_Net_HO_Commit_request message, while the vehicle acknowledges the reception of the information using a 4.MI H_Net_HO_Commit_response message.Subsequently, the vehicle sends a 5.MI H_Link_up message to the selected network to enable the communication channel.Furthermore, the vehicle sends a 6.UN A message to the selected network, to complete the establishment of both the Link and Network layers.Finally, the selected network replies back with a 7.MI H_N2N_Complete_request to the current one.By receiving this message, the MAG releases the resources that have been reserved for the current network and replies back with a 8.MI H_N2N_Complete_response message to successfully finish the HO execution.It has to be noted that in the case of predictive intra-MAG HO execution (Figure 10), since the involved networks are served by the same MAG, the use of 5.Handover_Initiate and 6.Handover_Ack messages is not necessary [48], as the MN-ID, the MN-LLID, and the Local Mobility Anchor Address (LMAA) are already known by their common MAG.Furthermore, the MAG does not exchange 7.PBU and 8.PBA messages with the LMA since it retains a Proxy Binding Update List (PBUL).Specifically, the MAG's PB-UL stores information about the vehicle including the MN-ID, the Mobile Network Prefix (MNP), and the vehicle's current network.Thus, when the vehicle performs an intra-MAG handover, the MNP is obtained from the PBUL.Furthermore, due to the common serving MAG, false handovers do not cause any malfunction.

Proposed Reactive HO Scheme
In the case of reactive inter-MAG HO execution (Figure 11), the vehicle sends a 1.MI H_Li-nk_up_React to the selected network, which includes information about the MN-ID and the MN-LLID and enables the communication channel.Subsequently, the vehicle sends a 2.UN A message to the selected network, to complete the establishment of both the Link and Network layers, as well as a 3.MI H_N2N_HO_Commit_request message to reserve the required resources.By receiving this message, the selected network reserves the required resources and sends back a 4.MI H_N2N_HO_Commit_response to the vehicle.
Thereafter, the MAG of the selected network identifies that the vehicle was not previously connected to any of the networks that it manipulates.Thus, the selected network sends a 5.Handover_Initiate message to the previous one through the LMA.In this message, the Proxy (P) and the Buffering (B) flags of the message are set equal to 1.By receiving this message, the previous network responds with a 6.Handover_Ack message, and a bidirectional tunnel is established between the two networks.Thereafter, the buffered packets from the previous network are transmitted through the tunnel to the selected network to be delivered to the vehicle when its HO completes.Furthermore, the MAG of the selected network exchanges the 7.PBU and 8.PBA messages with the LMA to update the Binding Cache Entry (BCE) of the LMA.Consequently, the vehicle becomes ready to transmit or receive packets through the selected network.For the intra-MAG reactive VHO execution (Figure 12), the messages after the 4.MI H_-N2N_HO_Commit_response of the above case are not exchanged between the vehicle and the selected network.Specifically, in this case, the MAG of the selected network identifies that the vehicle was previously connected to a network that it manipulates.Thus, the exchange of signaling messages between the MAG and the LMA is not required, and the vehicle is ready to transmit or receive packets through the selected network.Finally, it has to be noted that after the successful execution of each of the above HO schemes, the number of handovers for the vehicle N CR u is increased for the estimation of the vehicle's velocity.

The Computational Complexity of the Proposed Scheme
As far as the computational complexity of the proposed methodology is concerned, the worst-case scenario is observed during the network selection process.In particular, during the first run of the DIFT-ACW algorithm, an O(n 2 ) complexity is introduced due to the weighting and the normalization of the two-dimensional decision matrices.Nevertheless, a constant time duration is required for the algorithm execution provided that D(t) = D(t prev ).In this case, the algorithm checks only the recent results obtained during the time period t prev , and an O(n) complexity is observed.

Simulation Setup and Results
In our simulation, the software-defined vehicular cloud topology presented in Figure 13 was evaluated.A mobility trace designating the map of the city of Athens along with transportation data was generated using the Open Street Map (OSM) software.Afterwards, the mobility trace was used as the input in the Simulator of Urban Mobility (SUMO), establishing a realistic mobility pattern including 77,797 vehicles driving inside Athens within a 24 h period of time.The network topology, as well as a Cloud/Fog infrastructure were created on the map using Network Simulator 3 (NS3) [49].The Fog infrastructure consists of 28 802.11pWAVE RSUs, 5 LTE-A Pro FD-MIMO Macrocell e-Node Bs (eNBs), 8 LTE-A Macrocell eNBs, 70 LTE-A Femtocells eNBs, 8 WiMAX Macrocell Base Stations (BSs), and 41 WiMAX Femtocell BSs, equipped with additional computational and storage resources.Furthermore, the access networks are located in geographical positions based on the data available from the Hellenic Telecommunications and Post Commission (EETT) [50].The positions and the spectrum of the base stations are presented in Tables A2-A4 (Appendix A) for the LTE-A Pro, the LTE-A, as well as the WiMAX and the WAVE technologies, respectively.Furthermore, the Fog infrastructure includes an SDN controller that constantly monitors the entire topology by implementing the Mobility Management as a Service (MMaS) model.Additional components including the Fog Manager, Fog Monitor, Network Manager, Network Monitor, Topology Discovery, Joint Resource Provisioner, and Power Consumption Estimator are incorporated into the Fog and are administered by the SDN controller.Specifically, the Fog Manager, as well as the Fog Monitor control the available resources of the topology.At the same time, the Network Manager manipulates OpenFlow forwarding tables, default path settings, and QoS-aware configurations.The Network Monitor provides the status of the switches of the entire infrastructure.Furthermore, the Topology Discovery component stores information about the generated topologies and the available UEs.Moreover, the Joint Resource Provisioner is utilized as a load coordinator for the VMs, the nodes, and the network traffic.Finally, the Power Consumption Estimator identifies the energy-demanding network access elements.
The Cloud infrastructure deploys a number of Virtual Machines (VMs) providing both Non-Disaster Management (nDM) and Disaster Management (DM) services.In particular, the nDM services include Navigation Assistance (NAV), Voice over IP (VoIP), Conversa-tional Video (CV), Buffered Streaming (BS), and Web Browsing (Web) applications.On the other hand, DM services include Disaster-Aware Navigation Assistance (DA-NAV), Disaster-Aware Voice over IP (DA-VoIP), Live Video Streaming for Emergency Manipulation (LVS-EM), and Image Transmission for Emergency Manipulation (IT-EM) systems.Furthermore, the Cloud infrastructure includes a Software-Defined Cloud (SDC) controller.The SDC controller consists of a Cloud Monitor and a Cloud Manager, which monitor and manipulate the Cloud resources, respectively.
Each access network supports at least one of the aforementioned services.Furthermore, three Service Level Agreements (SLAs) are determined.The SLA1 supports the entire service types, while at the same time, it provides the best values for both the QoS and the operator's policy decision criteria.SLA2 does not support CV, BS, and LVS-EM services.In addition, it provides slightly worse values about the criteria in comparison with the ones provided by SLA1.Finally, SLA3 supports only the NAV, web, and DA-NAV services with satisfactory values for the considered criteria.
The specifications of the networks are expressed using linguistic terms corresponding to specific value ranges per service type according to [23], whereas Table 1 summarizes the parameters of the simulation.Besides, the simulated network architecture uses real-world data, including the positions of the access networks and the mobility pattern of the vehicles.Hence, the proposed scheme can be easily adopted by telecommunication providers that leverage 5G key enabling technologies, such as Cloud/Fog computing and SDN.

Study of a Simulation Snapshot
In this subsection, a simulation snapshot is studied.This snapshot was taken at a simulation time equal to 43,200 s (namely a 12 h interval), while 10 of the vehicles were monitored.Each vehicle needs to be connected to a PoA that fulfills the constraints of its services and at the same time complies with its corresponding SLA agreement.Table 2 presents the status of each monitored vehicle.Furthermore, Table 3 presents the PoAs that are available in the location of each vehicle.Initially, during the HO initiation, the estimated Q u,i and SI NR u,i values are used to obtain the user satisfaction S u,i with a lookup to the Satisfaction Indicators Chart (SIC).Figure 14 presents the service weights used for the estimation of the Q u,i as they are obtained from the IFANP method.Furthermore, Table 4 includes the satisfaction grade, the SLA threshold, and the HO decision per vehicle connected to a network with specific QoS and SINR rates.In addition, the minimum acceptable Q SLA and SI NR SLA values per SLA are considered in order for the S th,SLA thresholds to be estimated.Consequently, if the satisfaction grade S u,i drops below the corresponding S th,SLA threshold, then the vehicle should perform an HO.Thus, according to the HO initiation results, Vehicle 6 will remain connected to its current PoA, while the rest of the vehicles will perform an HO to a new PoA.During the network selection, the weights of the corresponding criteria per service type and SLA are estimated using the IFANP method.In particular, the considered criteria are classified into three clusters.The first cluster is called "Network QoS characteristics" and contains criteria such as the throughput, the latency, the jitter, and the packet loss.Correspondingly, the second cluster is called "network policy characteristics" and contains criteria such as monetary cost, security, and service reliability.Finally, the third cluster is called "network energy-aware characteristics" and includes the energy consumption criterion, which mainly is the weight adaptation factor.Figure 15 presents the importance of each service with respect to the disaster severity level, while the estimated weights per SLA are presented in Figures 16-18.As can be observed, these weights are proportional to the requirements of each service, as well as to the agreements of each SLA.Specifically, in the case of SLA1, the weight of the monetary cost criterion is low, while in this SLA, the most important factors are the network QoS characteristics and the service reliability.On the contrary, in SLA3, the monetary cost is the most important criterion resulting in a high estimated weight value.Furthermore, intermediate values were estimated for the SLA2.Additionally, Figure 19 presents the importance of the energy consumption factor in conjunction with the remaining energy in each UE, while Figure 20 presents the final weight for each vehicle considered for the network selection process.Furthermore, Table 2 presents the SLA of each one of the monitored vehicles, as well as the services that it uses, its position, its velocity, the disaster severity level in its location, and the remaining energy of its equipment.The appropriate network for each vehicle is selected using the DFT-ACW algorithm described in [41].The algorithm receives as the input the specifications of the networks of the 10 vehicles outlined in Tables A5 and A6 (Appendix B).As follows, the network that each vehicle selects is given in Table 5.The always best connected principle is ensured since the required satisfaction grade per vehicle based on its mobility pattern and services is maintained.
In addition, the signaling transmission cost of the discussed methodology was compared with the one obtained from the FMIPv6 [51] and the eFPMIPv6 [52] schemes.For the latter two schemes, the algorithms considered for the HO initiation and the network selection processes were the Velocity-Aware Handover (VAH) [35], the Fuzzy AHP TOPSIS (FAT) [53], the Fuzzy AHP SAW (FAS) [53], the Fuzzy AHP MEW (FAM) [53], and the Fuzzy AHP ELECTRE (FAE) [54].The number of hops that each signaling message performs (Table 6), as well as the size in bytes of each message (Table 7) according to [55] were considered for the evaluation.
In particular, the signaling cost of the discussed methodology was estimated using Formula (42).In this formula, the P f parameter indicates the connection failure probability, which is supposed to be equal to 0.5 [55].Furthermore, the n parameter indicates the number of available PoAs.
Similarly, the signaling cost of the FMIPv6 was estimated using Formula (43), while the corresponding cost of the eFPMIPv6 was estimated using Formula (44).The signaling costs obtained from each scheme are presented in Figure 21.In particular, the proposed scheme achieved a lower signaling cost than the one achieved from the FMIPv6 and eFPMIPv6 schemes, since these schemes exchange additional signaling messages during the HO execution.

Twenty-Four-Hour Evaluation Results
In this scenario, the efficiency of the proposed scheme was evaluated during a 24 h simulation.Specifically, the number of HOs and the respective signaling cost were studied.According to the obtained results, the discussed scheme achieved better performance than the other schemes, since it performed fewer HOs, while at the same time, its signaling cost per HO was lower than the ones observed from the other schemes for both inter-MAG and intra-MAG predictive and reactive HOs.Specifically, the proposed scheme performed 372,170 HOs in total.Furthermore, the VAH scheme accomplished acceptable HO counts, which were equal to 533,590, by considering the vehicle's velocity for the handover process.However, it performed the network selection considering only the strength of the perceived signal, providing worse results than the proposed scheme.The rest of the schemes accomplished HO counts equal to 805,751, since they were based on the perceived RSS per user for the HO initiation process.
Finally, Figure 22 presents the total signaling cost observed for each scheme during the 24 h simulation by taking into consideration the number of HOs that each scheme performed, as well as the respective signaling cost.As shown, the proposed scheme achieved a lower signaling cost for both inter-MAG and intra-MAG predictive and reactive HOs.
MAG and intra-MAG predictive and reactive HOs.Specifically, the proposed scheme performed 372,170 HOs in total.Furthermore, the VAH scheme accomplished acceptable HO counts, which were equal to 533,590, by considering the vehicle's velocity for the handover process.However, it performed the network selection considering only the strength of the perceived signal, providing worse results than the proposed scheme.The rest of the schemes accomplished HO counts equal to 805,751, since they were based on the perceived RSS per user for the HO initiation process.
Finally, Figure 22 presents the total signaling cost observed for each scheme during the 24 h simulation by taking into consideration the number of HOs that each scheme performed, as well as the respective signaling cost.As shown, the proposed scheme achieved a lower signaling cost for both inter-MAG and intra-MAG predictive and reactive HOs.

Conclusions
In this paper, an HO management methodology for 5G vehicular networks was described supporting both predictive and reactive HO scenarios.Four HO subprocesses complemented the system's functionality, namely the velocity and network monitoring, the HO initiation, the network selection, and the HO execution.The velocity and network monitoring module estimates the velocity of each vehicle and provides the alternative access networks to the HO operation.If the user satisfaction grade becomes less than a predefined threshold, then predictive HO is performed.In such a case, the HO is initiated and network selection is performed.Reactive HO is performed whenever the vehicle unexpectedly loses the connectivity with its serving network.In such a case, network selection is immediately executed.Furthermore, the HO execution phase includes several improvements in the operation of the FPMIPv6 protocol.The proposed scheme was evaluated in a 5G vehicular network architecture, while the performance evaluation showed that it outperformed existing HO schemes, maintaining the user satisfaction at acceptable levels and reducing the HO signaling cost.

Figure 3 .
Figure 3.The workflow of the proposed methodology.

Figure 4 .Figure 5 .Figure 6 .
Figure 4.The membership functions used for the representation of the SINR indicator.

Figure 7 .
Figure 7.The S values range as obtained using the FIS.

Figure 8 .
Figure 8.The signaling performed during the network selection.

Figure 15 .
Figure 15.The importance of each service with respect to the disaster severity level.

Figure 20 .
Figure 20.The final network selection weights for each vehicle.

Figure 21 .
Figure 21.The signaling cost of each HO scheme.

Figure 22 .
Figure 22.The total signaling cost of each HO scheme during the 24 h simulation.

Linguistic Terms: ______________________________________________________________________________________ Figure 2. The
membership functions used for the pairwise comparisons of criteria.

Table 1 .
The simulation parameters.

Table 2 .
The monitored vehicles' status.

Table 3 .
The networks that are available in the area of each monitored vehicle.

Table 4 .
The results of the HO initiation process.

Table 5 .
The network that each vehicle selects.

Table 6 .
The hops between the components of the considered topology.

Table 7 .
The cost of each message exchanged during the signaling process.

Table A3 .
The positions and the frequencies of the simulated LTE-A networks.

Table A4 .
The positions and the frequencies of the simulated WiMAX networks.

Table A6 .
The networks of SLA2 and SLA3 considered in the study of the simulation snapshot.