A Handover Method for Dense Heterogeneous Networks Based on User Location Prediction and Multi-Attribute Decision Making

By employing user location prediction and multi-attribute decision-making methods in dense heterogeneous networks, and utilizing the improved MS-BiTCN-BiLSTM-Attention model and SDN global perception capabilities, a multi-dimensional evaluation system is constructed. This solves the problems of redundant switching and load imbalance in traditional switching algorithms, achieving more efficient network resource utilization and a seamless communication experience.

CN123124045APending Publication Date: 2026-11-17JIAXING XINGHAN INTERNET TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611004231.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-07
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In dense heterogeneous networks, traditional handover algorithms struggle to ensure users access appropriate cells while suppressing redundant handover frequencies, leading to ping-pong effects, small cell traps, and global load imbalances, thus impacting network performance.

Method used

A user location-based prediction and multi-attribute decision-making approach is adopted. An improved MS-BiTCN-BiLSTM-Attention multi-scale model is used to predict future channel quality. Combined with the global perception capability of SDN, a multi-dimensional evaluation system is constructed that integrates prediction of RSRP, normalized load, transmission delay and expected dwell time. The optimal target cell is selected through the AHP-EWM-TOPSIS joint decision-making mechanism.

Benefits of technology

It significantly reduced the frequency of redundancy switching, optimized the load balancing status, and improved resource utilization and the user's seamless communication experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123124045A_ABST
    Figure CN123124045A_ABST
Patent Text Reader

Abstract

This invention discloses a handover method for dense heterogeneous networks based on user location prediction and multi-attribute decision-making. In the triggering phase, an improved MS-BiTCN-BiLSTM-Attention multi-scale model is introduced to enhance trajectory prediction accuracy and predict future channel quality evolution trends. In the decision phase, a multi-dimensional evaluation system is constructed, fusing predicted reference signal received power, normalized load, transmission delay, and expected dwell time. An AHP-EWM-TOPSIS joint decision-making mechanism is used to select the optimal target cell for access. Simulation results show that this invention can significantly suppress redundant handover frequency, optimize the overall network load balancing, and improve resource utilization and seamless communication experience for users in dense scenarios by utilizing prior information from trajectory prediction and the global awareness capability of SDN, while ensuring basic connectivity.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a dense heterogeneous network handover method based on user location prediction and multi-attribute decision-making, belonging to the field of computer communication technology. Background Technology

[0002] The global wireless communication business is experiencing unprecedented rapid growth, and the booming mobile internet ecosystem is driving the evolution of next-generation heterogeneous wireless networks towards high-density deployment. The overlapping coverage areas of base stations and wireless access points have led to a dramatic increase in the number of network options available to terminals during vertical handover. The inherent differences between heterogeneous network standards present even more challenging problems for vertical handover: each handover requires the sequential teardown of the existing connection and the establishment of a new one, causing a brief communication interruption during which users are completely unable to send or receive data packets, severely impacting service continuity and user experience. Simultaneously, traditional handover triggering algorithms based on received signal strength indicators are struggling to sensitively capture dynamic environmental changes in increasingly complex network scenarios, making it difficult to accurately determine the appropriate handover timing and easily leading to frequent ping-pong handovers. To address the dual challenges of the expanding candidate network scale and increased scenario complexity, next-generation vertical handover algorithms in heterogeneous wireless networks must, while maintaining moderate computational overhead and reliable execution time, focus on enhancing the accuracy of handover decisions and the rationality of strategies to avoid erroneous handovers caused by excessive similarity between different decision-making schemes.

[0003] With the rapid development of 5G and next-generation mobile communication technologies and the explosive growth of mobile data traffic, users are increasingly situated in dense heterogeneous network (HetNet) environments comprised of nested deployments of macro base stations, micro base stations, and femtocells. This invention addresses the connectivity issues in network edge communication-constrained scenarios and further investigates network connection optimization in dense network deployment scenarios. Based on the long-standing definition of multi-layer cell deployment in the 3GPP standard system, heterogeneous wireless networks encompass two typical paradigms: one is cross-standard heterogeneity, i.e., the convergence of different wireless access technologies such as 5G NR, LTE, and Wi-Fi; the other is intra-layer heterogeneity, i.e., a hybrid deployment architecture formed by multi-level nodes such as macro base stations, micro base stations, and femtocells under the same standard due to fundamental differences in coverage radius, transmit power, and access capacity.

[0004] While HetNet effectively improves network capacity and coverage, the overlapping coverage of multiple standards and multi-level cells results in extremely frequent cross-level and cross-cell handovers for users moving at high speeds. Traditional handover algorithms based on Reference Signal Received Power (RSRP) only use RSRP as the decision criterion, and users always hand over to the base station with the highest RSRP. This "greedy" access strategy easily leads to problems such as inappropriate handover timing and inaccurate decisions in dense heterogeneous networks, specifically manifesting as the ping-pong effect, small cell traps, and global load imbalance, seriously affecting overall network performance. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings and deficiencies of existing technologies by proposing a dense heterogeneous network handover method based on user location prediction and multi-attribute decision-making. In the triggering phase, this method introduces an improved MS-BiTCN-BiLSTM-Attention multi-scale model to enhance trajectory prediction accuracy and predict future channel quality evolution trends. In the decision phase, a multi-dimensional evaluation system is constructed, fusing predicted reference signal received power, normalized load, transmission delay, and expected dwell time. An AHP-EWM-TOPSIS joint decision-making mechanism is then used to select the optimal target cell for access. Simulation results show that this invention can significantly suppress redundant handover frequency, optimize the overall network load balancing, and improve resource utilization and seamless user communication experience in dense scenarios by utilizing prior information from trajectory prediction and the global awareness capability of SDN, while ensuring basic connectivity.

[0006] The technical solution adopted by this invention to solve its technical problem is: a dense heterogeneous network handover method based on user location prediction and multi-attribute decision-making, which includes the following steps:

[0007] Step 1: Define the set of base stations within the system as follows ,in The total number of base stations, each base station It has fixed geographical coordinates Transmission power Coverage radius and the maximum number of users allowed to access the system. Define the set of mobile terminals (users) that may generate handover requests in the current scenario as follows: ,in To realistically simulate resource contention and network congestion in dense heterogeneous networks, stationary background nodes are randomly distributed during system initialization, and their initial home cells are determined based on the maximum received power criterion. Simultaneously, a specific user is selected... As a candidate mobile user;

[0008] Step 2: Assumption Accessibility matrix between mobile terminals and candidate cells for:

[0009] (4.1)

[0010] Among them, matrix elements The meaning is as follows:

[0011]

[0012] When a terminal traverses an extremely narrow, tiny cell boundary at high speed, the target cell selection logic of traditional hard decision can be expressed as:

[0013]

[0014] Step 3: The network handover is transformed into a comprehensive utility optimization problem based on user location prediction and multi-attribute decision-making. The system comprehensively considers the physical layer signal trend derived from trajectory prediction and the real-time state of the network layer, and selects four key link parameters—reference signal received power, normalized load, transmission delay, and expected dwell time—based on the fusion of prediction information to quantitatively evaluate candidate cells.

[0015] Reference signal received power of fusion prediction To overcome the short-sightedness of traditional measurement mechanisms, a trajectory prediction-based extrapolation value is introduced. Construct a time-forward reference signal received power (benefit-oriented attribute):

[0016]

[0017] in, To integrate weighting factors, decision weights are assigned to the RSRP extrapolation values ​​based on trajectory prediction, while the measured RSRP weights are retained to prevent prediction errors from causing decision failures.

[0018] Step 4: Define the normalized load (cost attribute) of the candidate cell as the number of current access users. With its maximum capacity The ratio:

[0019]

[0020] Normalized load during severe network congestion The calculated value is allowed to exceed 1.0, so as to objectively quantify the extreme physical limit state of cell overload;

[0021] A quadratic penalty function model based on queuing theory is used to quantitatively model network latency (a cost attribute) (unit: ms):

[0022]

[0023] in, For unloaded foundation delay, The congestion coefficient;

[0024] Based on terminal prediction average speed With cell coverage radius The expression for estimating a user's expected dwell time (a benefit-oriented attribute) using the probability density function is as follows:

[0025]

[0026] in, This parameter is used to characterize the degree of random fluctuation in user speed, thereby avoiding decision oscillations caused by minor speed perturbations and enhancing the robustness of the judgment; This is the spatial dimension scaling correction factor, taking into account that the terminal often does not travel along the ideal longest diameter when traversing a cell, utilizing... The theoretical geometric coverage is mapped to the effective expected distance that the terminal actually traverses the cell;

[0027] Step 5: Use vector normalization to make the initial decision matrix dimensionless. Let the standardized decision matrix be... Its elements The calculation formula is as follows:

[0028]

[0029] Through this transformation, the initial attribute values ​​are mapped to the unit vector space, and the indicators of each dimension are transformed into dimensionless standardized values. This not only preserves the relative proportional relationship between the attributes of each base station, but also provides a unified data foundation for subsequent ranking of the ideal solution of TOPSIS by combining combined weights.

[0030] Furthermore, step 5 of the present invention includes, under accessible constraints... Next, we will search for a comprehensive utility evaluation function for the target community. The process of maximizing the optimal solution can be expressed as follows:

[0031] .

[0032] Beneficial effects:

[0033] 1. This invention first constructs a hybrid deep learning model of MS-BiTCN-BiLSTM-Attention to predict user movement trajectories and future channel quality. Then, it utilizes an SDN global view to build an evaluation system encompassing fused predictions of RSRP, load, latency, and expected dwell time, and uses an AHP-EWM-TOPSIS multi-attribute decision mechanism to determine the optimal target cell. Finally, the SDN controller issues flow table instructions based on the decision results to guide the terminal to complete a seamless handover. Simulation results show that this invention has high user location prediction accuracy, effectively reduces the number of frequent handovers, and achieves load balancing while ensuring users access suitable cells, providing an optimized solution for mobility management in dense heterogeneous network scenarios.

[0034] 2. This invention can utilize prior information from trajectory prediction and the global awareness capability of SDN to significantly suppress redundant switching frequency while ensuring basic connectivity, optimize the load balancing state of the entire network, and improve resource utilization and seamless communication experience for users in dense scenarios. Attached Figure Description

[0035] Figure 1 This is a diagram of the dense heterogeneous network model of the present invention.

[0036] Figure 2 This is a structural diagram of the MS-BiTCN-BiLSTM-Attention model of the present invention.

[0037] Figure 3 This is a schematic diagram illustrating the switching of a dense heterogeneous network simulation scenario according to the present invention.

[0038] Figure 4 This is a location prediction error diagram of the present invention.

[0039] Figure 5 This is a location prediction error diagram of the present invention.

[0040] Figure 6 This is a comparison chart of the predicted RSRP and the actual RSRP based on the predicted location of the present invention.

[0041] Figure 7 This is a network initialization user space distribution and cell load status diagram for the present invention.

[0042] Figure 8 This is a graph showing the change of normalized load of the serving cell over time, as presented in this invention.

[0043] Figure 9 This is a schematic diagram of the user access cell trajectory according to the present invention. Detailed Implementation

[0044] The invention will now be described in further detail with reference to the accompanying drawings.

[0045] This invention addresses mobility management in dense urban road network scenarios within 5G and next-generation heterogeneous wireless networks. In such scenarios, to cope with the explosive growth of mobile services, a dense network deployment of micro base stations and femtocells nested within the coverage area of ​​macro base stations has become a core means to improve system capacity and coverage quality. According to the 3GPP standard definition, this multi-level base station hybrid dense deployment architecture, formed under the same standard but with fundamental differences in coverage radius, transmit power, and access capacity, belongs to a typical intra-layer heterogeneous wireless network. Under the global management of the SDN architecture, the system can perceive the entire network topology and the resource status of base stations at each level in real time. However, limited by the small coverage area of ​​microcells, high-speed mobile users are prone to frequent cross-layer handovers, accompanied by serious small cell traps, ping-pong effects, and global load imbalance. This high degree of overlap among multi-level cells and intense competition for resources makes how to comprehensively utilize multi-dimensional network parameters to achieve optimal target cell selection and intelligent mobility management a core research challenge in dense heterogeneous network environments.

[0046] like Figure 1 As shown, the present invention defines the set of base stations within the system as follows: ,in This represents the total number of base stations. Each base station... It has fixed geographical coordinates Transmission power Coverage radius and the maximum number of users allowed to access the system. Define the set of mobile terminals (users) that may generate handover requests in the current scenario as follows: ,in The number of terminals. To realistically simulate resource contention and network congestion in a dense heterogeneous network, stationary background nodes are randomly distributed during system initialization, and their initial home cell is determined based on the maximum received power criterion; simultaneously, a specific user is selected. As a candidate mobile user.

[0047] In dense, heterogeneous networks, due to the overlapping coverage of multi-level base stations and the extremely narrow coverage radius of micro base stations, high-speed mobile terminals are highly susceptible to frequent cross-level handovers and the ping-pong effect. Meanwhile, the traditional greedy access criterion based on maximum RSRP often leads to localized network congestion and global load imbalance. Assuming... Accessibility matrix between mobile terminals and candidate cells for:

[0048] (4.1)

[0049] Among them, matrix elements The meaning is as follows:

[0050]

[0051] When a terminal traverses an extremely narrow, tiny cell boundary at high speed, the target cell selection logic of traditional hard decision can be expressed as:

[0052]

[0053] To overcome the limitations of the aforementioned single-dimensional evaluation, this invention transforms network handover into a comprehensive utility optimization problem based on user location prediction and multi-attribute decision-making. The system comprehensively considers the physical layer signal trend derived from trajectory prediction and the real-time state of the network layer, selecting four key link parameters—reference signal received power, normalized load, transmission delay, and expected dwell time—based on fused prediction information, to quantitatively evaluate candidate cells.

[0054] Reference signal received power of fusion prediction To overcome the short-sightedness of traditional measurement mechanisms, a trajectory prediction-based extrapolation value is introduced. Construct a time-forward reference signal received power (benefit-oriented attribute):

[0055]

[0056] in, To integrate weighting factors, decision weights are assigned to the RSRP extrapolation values ​​based on trajectory prediction, while the measured RSRP weights are retained to prevent prediction errors from causing decision failures.

[0057] The normalized load (cost attribute) of a candidate cell is defined as the number of currently accessing users. With its maximum capacity The ratio:

[0058]

[0059] It should be noted that, in the event of severe network congestion, the normalized load... The calculated value is allowed to exceed 1.0, so as to objectively quantify the extreme physical limit state of cell overload.

[0060] A quadratic penalty function model based on queuing theory is used to quantitatively model network latency (a cost attribute) (unit: ms):

[0061]

[0062] in, For unloaded foundation delay, This represents the congestion coefficient.

[0063] Based on terminal prediction average speed With cell coverage radius The expression for estimating a user's expected dwell time (a benefit-oriented attribute) using the probability density function is as follows:

[0064]

[0065] in, The parameter is used to characterize the degree of random fluctuation in user speed (set to 1 m / s in this invention) to avoid decision oscillations caused by small speed disturbances, thereby enhancing the robustness of the decision; This is the spatial dimension scaling correction factor, taking into account that the terminal often does not travel along the ideal longest diameter when traversing a cell, utilizing... The theoretical geometric coverage is mapped to the effective expected distance of the terminal actually traversing the cell (set to 0.5 in this invention).

[0066] Because the four attribute parameters mentioned above have significant differences in dimensions and orders of magnitude, they cannot be directly calculated together. This invention employs vector normalization to render the initial decision matrix dimensionless. Let the standardized decision matrix be... Its elements The calculation formula is as follows:

[0067]

[0068] Through this transformation, the initial attribute values ​​are mapped to a unit vector space, and the indicators of each dimension are converted into dimensionless standardized values. This not only preserves the relative proportional relationships between the attributes of each base station, but also provides a unified data foundation for subsequent ranking of the ideal solution of TOPSIS by combining combined weights. The dense heterogeneous network handover problem discussed in this invention is essentially under accessibility constraints. Next, we will search for a comprehensive utility evaluation function for the target community. The process of maximizing the optimal solution can be expressed as follows:

[0069]

[0070] In summary, the dense networking of multi-layered nested cellular coverage makes mobility management challenging due to multi-dimensional resource competition and extreme network dynamics caused by overlapping coverage layers. Traditional decision-making mechanisms typically focus on a single physical layer metric (such as RSRP), which, limited by the extremely narrow coverage of tiny cells, easily leads to highly dynamic users getting trapped in small cells, causing frequent redundant handovers and localized cell congestion. Such one-sided decision-making mechanisms not only fail to guarantee handover accuracy in complex deployment environments but also easily cause global load imbalances.

[0071] To address the aforementioned multi-dimensional resource conflict problem, this invention proposes a dense heterogeneous network handover algorithm based on user location prediction and multi-attribute decision-making. This algorithm leverages the global awareness capability of SDN to acquire real-time network status. First, it utilizes an improved multi-scale MS-BiTCN-BiLSTM-Attention hybrid deep learning model to improve prediction accuracy by extracting multi-scale features from the mobile trajectory, thereby predicting future channel trends. Subsequently, it combines AHP-EWM subjective and objective joint weighting with the TOPSIS multi-attribute decision-making mechanism to construct a four-dimensional evaluation system that integrates predicted RSRP, normalized load, transmission delay, and expected dwell time, quantitatively evaluating the comprehensive utility of candidate cells. This algorithm aims to provide better target cell access selection for highly dynamic users, effectively suppressing redundant handover while achieving balanced optimization of the entire network load.

[0072] Dense Heterogeneous Network Handover Algorithm Based on User Location Prediction and Multi-Attribute Decision Making

[0073] MS-BiTCN-BiLSTM-Attention-based trajectory prediction

[0074] In densely populated urban areas, mobile terminal trajectories exhibit a multi-timescale characteristic, intertwining long-term stable straight-line motion with short-term abrupt changes in direction. Traditional TCNs, limited by a single fixed receptive field, often struggle to simultaneously capture low-frequency macroscopic motion trends and high-frequency microscopic local fluctuations, leading to significant cumulative errors when processing complex non-stationary sequences. This multi-scale physical characteristic of trajectory data places stringent demands on mobility management in dense heterogeneous networks (HetNet). Due to the large span of base station coverage radii in HetNet, macro base station handover decisions rely on long-term trajectory evolution trends to ensure basic connectivity; while precise access and congestion avoidance for micro base stations heavily depend on the keen capture of short-term microscopic trajectory changes. To meet these multi-level decision accuracy requirements, this invention proposes a hybrid trajectory prediction model that integrates a multi-scale bidirectional temporal convolutional network (MS-BiTCN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism. This model, by configuring dilated convolution kernels of different sizes, utilizes differentiated receptive fields to extract physical features of trajectories at different temporal granularities in parallel, thereby providing a reliable prior space decision-making basis for the accurate and seamless switching of SDN controllers in complex heterogeneous environments. The model structure is as follows: Figure 2 As shown.

[0075] This model, by configuring dilated convolution kernels of different sizes, utilizes differentiated receptive fields to extract multi-scale physical features of trajectories in parallel. Combined with a collaborative design of bidirectional temporal modeling and an attention mechanism, it achieves multi-granularity analysis and robust prediction of mobile terminal trajectory data. Furthermore, to better adapt to the dynamic switching requirements in heterogeneous network environments, the model adopts a hierarchical architecture, transforming the prediction target from absolute coordinates into relative displacement increments, thus endowing the model with stronger translation invariance and spatial generalization capabilities. The model's data processing flow can be formally described as follows:

[0076] (1) Input Layer and Feature Preprocessing: After feature engineering, the original trajectory data is used to construct an input sequence containing multi-dimensional motion states such as position, velocity, and acceleration. After data standardization, the data is input into the multi-scale bidirectional temporal convolutional network (MS-BiTCN). This module consists of several sets of parallel bidirectional temporal convolutional networks (BiTCN) configured with different kernel sizes (3, 5, 7). Through differentiated receptive field design, the model can simultaneously capture multi-temporal granular features such as short-term local fluctuations and long-term motion trends, solving the problem of incomplete feature representation of single-scale models in cross-spatial scale scenarios.

[0077] (2) Gain of the bidirectional temporal convolution module: By constructing two temporal convolutional streams, BiTCN enables the output at each moment to simultaneously perceive the historical motion inertia and future trend priors, under the premise of strictly ensuring causal constraints (the reverse path is achieved by flipping the time axis), which greatly enhances the modeling stability and prediction accuracy of non-stationary abrupt behaviors such as turning, acceleration and deceleration, and provides key bidirectional structural support for closely tracking the trajectory evolution across the tiny cell boundary.

[0078] (3) Bidirectional temporal modeling and attention mechanism: The BiLSTM module receives features after multi-scale BiTCN abstraction and cross-scale fusion, and uses its gating mechanism and bidirectional loop structure to further refine the encoding of long-range motion patterns and deep dependencies in the sequence. Thus, the local fine features of BiTCN and the global trend features of BiLSTM complement each other in the model, enabling the prediction results to respond sensitively to short-term changes without deviating from long-range motion trends, significantly improving the spatiotemporal consistency of continuous prediction under complex paths. By dynamically calculating the attention weights of each time step, the model can adaptively focus on the historical trajectory segments (such as sudden turns or abrupt acceleration points) that have the greatest influence on the current prediction, and aggregate them into a highly condensed context feature vector.

[0079] (4) Incremental Regression Prediction and Output Optimization: The context feature vector is input to the regression output module. After nonlinear transformation and feature compression by the fully connected layer, the linear regression layer outputs the relative displacement increment at future time steps. The entire training process combines Dropout layer, Huber loss function, and regularization strategy, which effectively suppresses overfitting and improves the model's robustness to abnormal jump trajectory points. This architecture significantly improves the ability to represent non-stationary motion sequences and provides methodological support for trajectory prediction and multi-parameter network switching algorithms in dense heterogeneous network environments.

[0080] The trajectory sequence is first input into the MS-BiTCN module to extract deep multi-scale spatial features. These features, after cross-scale interaction, are incorporated with target temporal prior information and then fed into the BiLSTM module to deeply capture forward and reverse motion dependencies. An Attention mechanism is used to dynamically focus on key time steps, generating a context vector. Finally, a fully connected network performs a regression task on the relative displacement increment, which is then superimposed onto the baseline coordinates to obtain the final prediction result. In summary, the MS-BiTCN-BiLSTM-Attention model is an adaptive improvement to meet the requirements of dense heterogeneous networks for multi-scale coverage features and higher switching accuracy. This allows the improved prediction model to provide more accurate and robust spatiotemporal look-ahead information for subsequent network switching algorithms based on multi-attribute decision-making.

[0081] Switching decision strategy based on multi-attribute decision

[0082] After predicting the future trajectory of mobile users, a comprehensive performance evaluation of each candidate cell needs to be conducted. This invention proposes to use a multi-attribute decision algorithm that combines subjective and objective weighting (AHP-EWM) with the Top-Approximation Ideal Solution Ranking (TOPSIS) method to select the optimal cell. When evaluating the comprehensive performance of candidate cells, this invention selects four key parameters: reference signal received power, normalized load, transmission delay, and expected dwell time, based on the fused prediction information.

[0083] (1) Subjective and objective joint empowerment based on AHP-EWM

[0084] To balance the guidance of expert experience with the objectivity of real-time data, this invention employs a combined subjective and objective weighting method based on AHP-EWM. This invention constructs a system consisting of a target layer (optimal target cell selection) and an attribute layer (fusion prediction). This invention employs a multi-dimensional evaluation system comprising normalized load, transmission delay, estimated dwell time, and a scheme layer (each candidate access cell). When evaluating the overall performance of candidate cells, considering that inter-layer heterogeneous handover involves not only physical layer signal quality but also network layer resource status and user layer mobility characteristics, this invention selects the aforementioned four key parameters, among which the fused prediction... The expected dwell time and the normalized load represent coverage quality and connection stability, respectively, while the normalized load and transmission latency reflect the network resource status and service response capability.

[0085] (a) Subjective weight calculation

[0086] Saaty's 1-9 scale (see Table 4.1) was used to quantify the relative importance of each attribute. Based on the strategy of "prioritizing mobility support and coverage quality," a 4th-order positive-reciprocal judgment matrix was constructed. .

[0087] Table 4.1 Definition of Saaty scale

[0088]

[0089] In highly dynamic mobile scenarios with dense, multi-layered heterogeneous networks, ensuring basic connectivity and suppressing the ping-pong effect are primary objectives. Therefore, fusion prediction, which represents signal coverage quality, is crucial. The expected dwell time, representing connection stability, should be considered more important than normalized load and transmission latency, representing resource allocation. Based on a "mobility support and coverage quality priority" strategy, we determine the importance of dwell time... It has a "slightly more important" advantage compared to load and latency, so the corresponding position in the judgment matrix is ​​assigned a value of 2, constructing a 4th order judgment matrix. :

[0090]

[0091] The largest eigenvalue of the matrix is ​​found using the root method. The corresponding normalized feature vector is the subjective weight vector. To perform a consistency check, for a 4th order matrix, the random consistency index is... Calculate the consistency index. Then the consistency ratio This indicates that the judgment matrix is ​​logically sound and passes the test.

[0092] (b) Calculation of objective weights

[0093] Entropy weight method (EWM) is introduced to mine the dispersion of candidate base station attributes in the current decision period, correcting subjective bias. First, for The decision matrix of each candidate base station is obtained by range standardization. To eliminate dimensional differences. Calculate the first... The characteristic proportion of each base station under the item indicator :

[0094]

[0095] Then calculate the first Information entropy of the indicator Finally, the objective weights of each attribute are determined based on information entropy:

[0096]

[0097]

[0098] The greater the difference in the measured data of a certain indicator, the higher its information entropy. The smaller the value, the more effective information the indicator contains in the current period, and thus it is given a higher objective weight.

[0099] (c) Combined weights

[0100] The Analytic Hierarchy Process (AHP) assigns weights by having experts compare indicators pairwise based on experience. This process is inevitably influenced by individual expert preferences and cognitive differences, resulting in a significant subjective bias in weight allocation. The Entropy Weighting Method, on the other hand, uses information entropy to measure the variability of each indicator's data, determining weights entirely from objective data distribution and eliminating reliance on expert experience. However, this method suffers from over-idealization, easily overlooking non-data-driven special factors of significant importance in specific real-world scenarios. To overcome the limitations of single weighting methods, this invention employs a comprehensive weighting strategy that integrates subjective and objective information. This involves linearly weighting and combining the weights obtained from the AHP and Entropy Weighting Methods. By rationally configuring the weight coefficients of both methods, the dual advantages of subjective experience and objective data mining are fully integrated, ultimately generating a more comprehensive, balanced weighting result with practical explanatory power, thus compensating to some extent for the shortcomings of single methods in information coverage. Let the weight coefficients be... Simple weighting yields the combined weights. for:

[0101]

[0102] Among them, coefficient The degree of reliance on expert experience rules can be adaptively set according to different scenarios. Considering that the heterogeneous urban dense road network scenario addressed in this invention may experience local congestion, but the overall network topology and communication infrastructure are in a stable operating state, under simulation conditions without extreme network failures, the following parameters are taken: The subjective and objective factors are equally weighted.

[0103] (2) Comprehensive utility evaluation of candidate base stations based on TOPSIS

[0104] After obtaining the combined weights, the access utility of each cell is quantified using the Top-Optimal Solution Ranking Method (TOPSIS).

[0105] (a) Decision matrix normalization: construction The original decision matrix of the candidate cells Vector normalization is used to eliminate the influence of dimensions, resulting in a normalized matrix. :

[0106]

[0107] Then combine AHP weights Construct a weighted normalized matrix ,in .

[0108] (b) Determine the positive and negative ideal solutions: Define the positive ideal solution The set of minimum values ​​for cost-related indicators and maximum values ​​for benefit-related indicators, the negative ideal solution. Conversely:

[0109]

[0110]

[0111] (c) Calculate proximity and ranking: Calculate the proximity of each candidate cell Euclidean distance to the ideal solution Distance to the negative ideal solution To obtain the relative closeness :

[0112]

[0113] The larger the value, the better the overall performance of the network, and it can be used as the initial target network.

[0114] (3) Physical constraints and hysteresis threshold decision

[0115] To further enhance the algorithm's robustness in environments with fluctuating signals, a physical layer veto mechanism is introduced. A minimum signal reception threshold is set before proceeding with the sorting process. If the candidate cell The system will assess its overall relevance. The forced zeroing constraint mechanism fundamentally eliminates the risk of invalid handovers to communication dead zones, ensuring the reliability of the basic data link. Furthermore, in dense heterogeneous networks, the high-density stacked deployment of micro / femto cells, due to the significant differences in coverage radii between base stations at different layers, is highly susceptible to severe inter-layer ping-pong effects caused by the combined effects of fast fading in the wireless channel and user micro-trajectory jitter. Therefore, this algorithm incorporates a hysteresis threshold decision logic at the execution level. Let the TOPSIS comprehensive utility score of the current serving cell be... The highest score among the candidate communities is To avoid invalid switching due to minor fluctuations in ratings, a hysteresis threshold is defined. The switching trigger condition is defined as follows:

[0116]

[0117] This mechanism ensures that the SDN controller will only issue a switching command when the overall performance of the target network is significantly better than that of the current network, thereby effectively reducing redundant switching.

[0118] Algorithm overall process

[0119] This invention elaborates on the overall execution logic of the proposed algorithm. In the SDN architecture, each level of base station acts as a data forwarding network element, responsible for reporting the local network status to the SDN controller; the SDN controller, as the logical hub, maintains a global view containing the entire network topology and load distribution. Based on the above architecture, the handover algorithm based on trajectory prediction and multi-attribute decision-making proposed in this invention mainly includes the following steps:

[0120] (1) Obtain network state and prediction parameters at the controller end.

[0121] The underlying base stations periodically report real-time network performance parameters such as load and number of connected users to the SDN controller via the southbound interface. To achieve proactive awareness, this is implemented for mobile users. Utilizing an improved MS-BiTCN-BiLSTM-Attention multi-scale hybrid model, based on past... Historical trajectory sequences at each time step, predicting future forecast windows. Movement trends within the system. The SDN controller acquires future movement trends in real time. Predicted coordinates at time This allows for the early estimation of the predicted RSRP (Receiving Reference Signal Power) for users reaching neighboring base stations. The collection of network attribute parameters mainly includes four network performance parameters for fusion prediction: Effective Reference Signal Power, Normalized Load, Transmission Delay, and Estimated Dwell Time. Among these, Effective Reference Signal Power and Estimated Dwell Time are benefit-related attributes, while Normalized Load and Transmission Delay are cost-related attributes.

[0122] (2) Implement subjective and objective combination weight calculation at the controller end.

[0123] After network status information is collected, the controller uses the analytic hierarchy process (AHP) to assess the importance of each network attribute and generate subjective weights accordingly. Simultaneously, it uses the entropy weighting method to derive objective weights, ultimately merging the two into a comprehensive weight for the network attribute. This combined subjective and objective weighting approach more comprehensively measures the importance of network attributes, thereby ensuring the scientific rigor and reliability of the multi-attribute evaluation process.

[0124] (3) Implement comprehensive ranking of candidate cells at the controller end.

[0125] After weighting, the SDN controller substitutes the combined weights into the Top-Ideal Solution Ranking (TOPSIS) model to comprehensively calculate the overall relative proximity (i.e., the final utility score) of each candidate cell, and initially selects the base station with the highest score as the optimal target cell. To further suppress the ping-pong effect that is easily caused by dense overlap of heterogeneous boundaries, the SDN controller will perform a secondary verification in conjunction with a hysteresis threshold mechanism before finally issuing the command, thereby triggering the handover conditions.

[0126] (4) The controller triggers and performs network switching.

[0127] After completing the selection and matching of the optimal target cell, the SDN controller will rely on the OpenFlow protocol to synchronously issue new flow table entries to the data forwarding layer, thereby controlling the underlying network devices to disconnect the terminal from the original serving cell and synchronously build the transmission link of the target cell, thus realizing seamless heterogeneous handover in the mobile communication process.

[0128] At the specific algorithm execution level, the detailed process of the dense heterogeneous network handover algorithm based on user location prediction and multi-attribute decision-making is as follows:

[0129]

[0130] Simulation Experiments and Result Analysis

[0131] Simulation platform and parameter settings

[0132] (1) Constructing a dense heterogeneous network simulation scenario

[0133] This invention, based on the MATLAB platform, models and simulates the handover algorithm based on trajectory prediction and multi-attribute decision-making under the proposed SDN architecture. The simulation scenario constructs a typical three-layer heterogeneous wireless network, with the coverage area consisting of nested macrocells, microcells, and femtocells. Macrocells provide basic wide-area coverage, microcells enhance capacity in hotspot areas, and femtocells fill coverage gaps and provide deep indoor coverage. To simulate the load and interference environment in a real network, stationary background users are generated in a uniformly random distribution within the simulation area. During the initialization phase, all background users complete initial access according to the maximum reference signal received power (RSRP) criterion, thus establishing the baseline load state of each cell in the entire network. Background users remain in their positions throughout the simulation period to maintain a stable interference field and baseline load; candidate mobile users are driven by external trajectory data, continuously moving within the scenario at 1-second time steps, simulating maneuvering behavior during emergency rescue or inspection operations. At each location update, the system makes a multi-attribute handover decision based on the predicted trajectory information and integrates multiple attributes, triggering the handover logic of the optimal target cell in real time. Figure 3 A schematic diagram of the switching process in the studied dense heterogeneous network scenario is given, and the detailed simulation parameter configuration is shown in Table 4.2.

[0134] Table 4.2 Simulation Parameter Configuration

[0135]

[0136] (2) Constructing a dataset and a user trajectory prediction model

[0137] In heterogeneous network environments, end-user mobility leads to fluctuations in the current access network status, raising the question of when to switch and which network to switch to. To endow the SDN controller with forward-looking switching decision-making capabilities, this invention utilizes an improved multi-scale hybrid neural network for trajectory prediction. To comprehensively evaluate algorithm performance, both the Manhattan grid model and the random walk model are introduced in the trajectory prediction evaluation phase to verify the model's generalization ability and robustness. In the system network switching evaluation phase, considering that dense heterogeneous networks are often deployed in urban building clusters, the focus is on the Manhattan grid model for analysis. The specific construction process and parameter settings are as follows:

[0138] (a) Motion Model and Raw Data Generation: This invention utilizes the motion scene generation tool BonnMotion to construct two candidate user motion models: Manhattan and Random Walk. A trajectory sequence lasting 50,000 seconds (approximately 14 hours) was generated for model training. The Manhattan grid model is used to characterize the physical paths and sudden changes in direction of vehicles in dense urban blocks, serving as the core driving scenario for subsequent handover simulations. The Random Walk model simulates pedestrian movement without fixed constraints, used to test the robustness and generalization ability of the prediction algorithm under highly uncertain motion conditions.

[0139] (b) Configure the SDN controller with a sampling period (Set to 1 second) User movement trajectory sampling. Given the extremely small coverage radius of micro-base stations in dense heterogeneous networks, excessively long forward prediction times can easily lead to accumulated errors and are not beneficial for handover decisions at the microscale. Therefore, this experiment sets the target prediction time span for the Manhattan grid model and the random walk model to 5 seconds and 2 seconds, respectively. After normalization, the data is divided into an 80% training set and a 20% validation set.

[0140] (c) Model Training and Network Handover Simulation: The model is trained offline using the trajectory dataset generated by BonnMotion. During the network handover simulation phase, the system predicts the user's future location based on the pre-trained model and estimates the predicted RSRP of the target base station. Simultaneously, it combines the real-time load and number of access users reported from the data plane to calculate the normalized load, transmission delay, and expected dwell time of candidate cells. Finally, the AHP-EWM-TOPSIS multi-attribute decision-making mechanism is used to complete the selection and handover of the globally optimal target cell.

[0141] Evaluation indicators

[0142] This invention constructs an evaluation index system from two aspects: trajectory prediction accuracy and network handover performance. The former is used to quantify the reliability of predicting the future location of mobile nodes; the latter is used to verify the improvement effect of introducing the prediction mechanism on handover decisions and the overall network service quality.

[0143] (1) Trajectory prediction evaluation index

[0144] To quantify the accuracy of the improved MS-BiTCN-BiLSTM-Attention model in predicting user mobility, root mean square error (RMSE) and average displacement error (ADE) were used as core evaluation metrics.

[0145] (a) Root Mean Square Error (RMSE)

[0146] Because RMSE has a magnifying penalty effect on large local coordinate deviations, this metric is used to quantify the dispersion of the predicted trajectory, focusing on verifying the model's temporal robustness and prediction stability in extreme scenarios such as sudden user turns. Its calculation formula is:

[0147]

[0148] in, This represents the total number of test samples. and They represent the first The test sample in the future The predicted trajectory coordinates and actual trajectory coordinates of the prediction step. The smaller the RMSE value, the higher the model's fit to highly dynamic movement features, the more stable the time series prediction results, and the less likely abnormal trajectory jumps will occur.

[0149] (b) Average displacement error (ADE)

[0150] ADE intuitively quantifies the average Euclidean distance (in meters) between the predicted and actual locations. In the dense heterogeneous network of this invention, the coverage radius of the small base stations is relatively small, placing more stringent requirements on spatial location resolution. Its definition is as follows:

[0151]

[0152] A lower ADE indicates that the predicted trajectory closely matches the actual physical path at the microscale, providing a reliable spatial decision basis for the SDN controller to perform accurate switching at tiny cell overlap edges.

[0153] (2) Network handover evaluation indicators

[0154] After fusing prior prediction information in the multi-attribute decision model, it is necessary to further evaluate the global performance of the algorithm in real-world dense heterogeneous networks. This invention selects the following core indicators for comparative analysis from two dimensions: handover behavior characteristics and network service performance:

[0155] (a) Cumulative Handover Count: The total frequency of cross-base station connection re-establishment for mobile users within the simulation lifecycle. An excessive number of handovers typically indicates unstable handover decisions, easily triggering the ping-pong effect and increasing signaling overhead. By comparing the handover counts under different algorithms, the improvement effect of trajectory prediction information on the rationality of handover triggering timing can be evaluated.

[0156] (b) System average throughput: System average throughput measures the amount of data successfully transmitted by the network per unit time and is an important indicator reflecting the overall service capacity of the network. By introducing trajectory prediction and multi-parameter collaborative handover mechanisms, better serving base stations can be selected for users in advance, thereby improving link quality and enhancing system throughput performance.

[0157] (c) Normalized Serving Cell Load: Normalized serving cell load reflects the real-time resource occupancy of the base station currently accessing the mobile user. It is defined as the ratio of the number of currently accessing users to the maximum allowed access capacity of the base station. By tracking the evolution of serving cell load during the movement of highly dynamic users, it can be verified whether the algorithm of this invention can effectively guide users to avoid local congestion hotspots and achieve balanced traffic distribution among multi-level base stations across the entire network.

[0158] (d) Base station dwell rate at each level: This metric represents the percentage of time a mobile user spends connected to macro base stations, micro base stations, and femtocells throughout the entire simulation period. This indicator quantifies and verifies whether the algorithm can enable high-speed users to remain more frequently at the macro base station layer with stable coverage, thereby prioritizing the connection stability of the underlying links.

[0159] Comparison and Analysis of Experimental Results

[0160] This invention aims to comprehensively evaluate the trajectory prediction accuracy and network switching performance of the proposed algorithm. First, in the trajectory prediction stage, a Manhattan grid and random walk model are used to verify the generalization ability and temporal robustness of the improved MS-BiTCN-BiLSTM-Attention model when handling non-stationary spatiotemporal sequences. Then, in the network switching stage, focusing on a typical Manhattan urban road network scenario, multi-dimensional evaluation metrics are used to quantitatively compare the performance advantages of this algorithm and traditional strategies in switching between layers of multi-layer heterogeneous networks.

[0161] (1) Trajectory prediction error assessment

[0162] (a) Manhattan grid model

[0163] To evaluate the model's predictive performance in regularized urban road scenarios, this invention uses user trajectory data generated by the Manhattan grid movement model for experimental evaluation. The dataset is divided into an 80% training set and a 20% validation set. All experimental frameworks use the Huber loss function (…). The system uses the AdamW optimizer. The initial learning rate starts at 0.001, and a dynamic learning rate decay strategy is introduced. If the validation set loss does not show a decreasing trend for three consecutive training epochs, the learning rate will be adaptively adjusted with a decay factor of 0.5. The time step for historical observations is set. The model batch size was set to 64, and the maximum training epochs were set to 100. To verify the actual performance of the prediction algorithm in a dense heterogeneous network environment, this invention selected a 600-second test trajectory rich in high-dynamic features such as frequent crossings and sudden turns during the network handover evaluation phase.

[0164] Figure 4 The model's performance and error distribution characteristics in predicting the user's position over the next 5 seconds are demonstrated when the user's movement speed dynamically varies between 10-20 m / s (mean 15 m / s). The 5-second prediction window was chosen to mitigate the ping-pong effect by considering the physical critical time for the user to cross the microcell edge. Figure (a) compares the actual and predicted trajectories of the end-user. Despite the complex nonlinear characteristics and random turns in the user's movement trajectory, the predicted trajectory still maintains a high degree of overlap with the actual trajectory. Figures (b) and (d) show the prediction results of the trajectory prediction model in the x and y coordinates, respectively. Figure (c) quantifies the position prediction deviation in a single axis direction in detail. Statistical results show that the model's average displacement error (ADE) is 32.24 m, and the root mean square error (RMSE) is 41.85 m. The error peaks at certain times are mainly attributed to the random abrupt changes in the user's movement direction in the Manhattan model, leading to cumulative bias in multi-step predictions. Overall, the trajectory prediction model can effectively learn the user's movement behavior and provide reliable information support for the SDN controller to perform proactive handover and network resource reservation.

[0165] (b) Random walk model

[0166] This invention further employs a random walk model to evaluate the scene generalization ability of the trajectory prediction network. The movement direction and speed exhibit high randomness and uncertainty, and lack fixed road constraints. To ensure the rigor of the model evaluation, the random walk model uses the same network hyperparameters as the Manhattan model. This invention reduces the prediction window to 2 seconds to avoid the cumulative error caused by long-term prediction.

[0167] Figure 5The study demonstrates the model's performance and error distribution characteristics in predicting the user's location 2 seconds in advance, with a user movement speed dynamically varying between 2-5 m / s and an average speed of 3 m / s. Statistical results show that the model's root mean square error (RMSE) is only 7.25 m, and the average displacement error (ADE) is 6.46 m. ​​Combined with heterogeneous network topology analysis, this prediction accuracy meets the spatiotemporal constraints of system mobility management and can support the SDN controller in making forward-looking handover decisions.

[0168] (2) Network handover performance analysis

[0169] This invention primarily evaluates the performance of the proposed handover algorithm in heterogeneous network environments. Considering that the Manhattan model is more realistic in simulating urban road topology and node mobility characteristics, this invention uses it as the main simulation scenario for detailed performance analysis. Combining the global view of SDN, the performance advantages of the proposed algorithm compared to the traditional single RSRP strategy are analyzed through multi-dimensional core indicators such as serving cell load state evolution, cumulative handover count, and base station retention ratio at different levels.

[0170] Figure 6 The results demonstrate a comparison between the actual RSRP of the serving cell and the model-predicted RSRP. Experimental results show that the spatiotemporal trajectory features captured by the MS-BiTCN-BiLSTM-Attention model can accurately predict signal fluctuations caused by distance changes and topological dynamics in dense heterogeneous networks. In the region of deep signal fading and the violent oscillation range at the boundary of heterogeneous networks, the predicted curve can closely track the evolution trend and abrupt inflection points of the actual signal strength; although there are slight instantaneous amplitude deviations at a very few sampling points, they do not deviate from the overall temporal trend. This provides robust prior data support for subsequent multi-attribute decision matrices based on TOPSIS, which is a key prerequisite for achieving forward-looking seamless handover.

[0171] like Figure 7 As shown, the network exhibits significant heterogeneous hierarchical characteristics in the initial simulation phase. A large number of background users are diverted to hotspot coverage areas such as microcells and femtocells, effectively alleviating the access pressure on macrocells. However, limited by the finite system capacity of the underlying microcells, their normalized load fluctuates dramatically, with some cells exceeding the overload threshold (normalized load > 1.0) at the initial moment. This non-uniform traffic distribution and localized congestion phenomenon realistically reflect the hotspot effect under dense networking, providing a stringent benchmark scenario for subsequent verification of the SDN handover algorithm's performance in load balancing and congestion avoidance.

[0172] Figure 8This paper demonstrates the evolution of the normalized load of the serving cell accessed by a candidate mobile user over time during mobility. The RSRP-based handover algorithm, due to its over-reliance on a single signal strength indicator, leads users to blindly switch to congested hotspot base stations, causing frequent and severe load fluctuations and congestion in their serving cells. In contrast, the algorithm of this invention proactively guides users to a lighter load level through multi-attribute decision-making, significantly smoothing the user's perceived load experience curve. It is worth noting that the instantaneous high-load access observed in the SDN scheme (e.g., at t≈140 seconds) reflects the algorithm's strategic trade-off: under strict topology constraints or high mobility conditions, the decision logic prioritizes signal continuity and handover stability rather than absolutely minimizing the load, thus avoiding radio link failures at the cost of temporarily deviating from the optimal load.

[0173] Figure 9 The access cell trajectories of candidate mobile users during the simulation period were compared. The RSRP-based handover algorithm, due to its over-reliance on instantaneous RSRP, is prone to frequent handovers at the overlapping coverage edges of heterogeneous networks. This not only disrupts the continuity of data flow but also leads to wasted signaling overhead. In contrast, the handover algorithm proposed in this invention, based on user location prediction and multi-attribute decision-making, effectively avoids invalid handovers caused by local channel fast fading or edge oscillations by extrapolating future network states and combining multi-attribute joint decisions under the coordination of the global SDN perspective. As can be seen from the figure, the access trajectory of the algorithm in this invention has a wider step span and a lower jump frequency, resulting in significant continuity and stability of the user's network connection.

[0174] The statistical results are shown in Table 4.3. In the Manhattan grid scenario, handover algorithms based on single signal strength lack forward-looking awareness of network state evolution, making them prone to falling into local optima and small cell traps, leading to frequent ping-pong handovers. In contrast, the SDN architecture-based handover algorithm proposed in this invention, which is based on user location prediction and multi-attribute decision-making, can fully consider the high mobility characteristics of users and the potential congestion risk of underlying small base stations, actively suppressing the blind deployment to femtocells.

[0175] Table 4.3 Performance Comparison of Different Algorithm Switching

[0176]

[0177] Statistical results show that the algorithm of this invention guides highly dynamic users to stay stably at the macro base station layer with wider coverage for a higher proportion of time (51.7%), while the handover algorithm based on signal strength only achieves a 0.5% retention rate at the macro base station layer. This significant difference is due to the fact that traditional algorithms prioritize the base station with the strongest instantaneous signal in densely heterogeneous scenarios. Because micro base stations and femtocells are densely deployed and their signal strength is often higher than that of macro base stations in local areas, users are frequently attracted to these small cells with extremely narrow coverage, resulting in almost no stable stay at the macro base station layer. In contrast, the algorithm of this invention, through trajectory prediction and multi-attribute decision-making, can identify and effectively avoid such small cell traps in advance, prioritizing keeping users at the macro base station layer with wider coverage and more stable connections. Simultaneously, the cumulative number of handovers using the algorithm of this invention is reduced from 38 times for the traditional strategy to 32 times, a reduction of approximately 15.8%. Under the centralized control architecture of SDN, each cross-base station handover involves signaling interaction and path reconfiguration between the controller and data plane devices. Reducing the number of handovers directly translates to a simultaneous decrease in handover signaling overhead and service interruption frequency, effectively alleviating the signaling processing load on the core network. More importantly, thanks to precise congestion avoidance and better resource scheduling, the algorithm of this invention reduces the average serving cell load for users from 0.72 in traditional handover strategies to 0.46, significantly alleviating network congestion in highly dynamic scenarios. In summary, the algorithm of this invention, by introducing a prediction mechanism and combining multi-attribute decision-making, effectively suppresses frequent handovers and localized congestion, thereby comprehensively improving the seamless handover service experience for highly dynamic users in dense heterogeneous networks.

Claims

1. A method for handover in dense heterogeneous networks based on user location prediction and multi-attribute decision-making, characterized in that, The method includes the following steps: Step 1: Define the set of base stations within the system as follows ,in The total number of base stations, each base station It has fixed geographical coordinates Transmission power Coverage radius and the maximum number of users allowed to access the system. Define the set of mobile terminals (users) that may generate handover requests in the current scenario as follows: ,in To realistically simulate resource contention and network congestion in dense heterogeneous networks, stationary background nodes are randomly distributed during system initialization, and their initial home cells are determined based on the maximum received power criterion. Simultaneously, a specific user is selected... As a candidate mobile user; Step 2: Assumption Accessibility matrix between mobile terminals and candidate cells for: (4.1) Among them, matrix elements The meaning is as follows: When a terminal traverses an extremely narrow, tiny cell boundary at high speed, the target cell selection logic of traditional hard decision can be expressed as: Step 3: The network handover is transformed into a comprehensive utility optimization problem based on user location prediction and multi-attribute decision-making. The system comprehensively considers the physical layer signal trend derived from trajectory prediction and the real-time state of the network layer, and selects four key link parameters—reference signal received power, normalized load, transmission delay, and expected dwell time—based on the fusion of prediction information to quantitatively evaluate candidate cells. Reference signal received power of fusion prediction To overcome the short-sightedness of traditional measurement mechanisms, a trajectory prediction-based extrapolation value is introduced. Construct a time-forward reference signal received power (benefit-oriented attribute): in, To integrate weighting factors, decision weights are assigned to the RSRP extrapolation values ​​based on trajectory prediction, while the measured RSRP weights are retained to prevent prediction errors from causing decision failures. Step 4: Define the normalized load (cost attribute) of the candidate cell as the number of current access users. With its maximum capacity The ratio: Normalized load during severe network congestion The calculated value is allowed to exceed 1.0, so as to objectively quantify the extreme physical limit state of cell overload; A quadratic penalty function model based on queuing theory is used to quantitatively model network latency (a cost attribute) (unit: ms): in, For unloaded foundation delay, The congestion coefficient; Based on terminal prediction average speed With cell coverage radius The expression for estimating a user's expected dwell time (a benefit-oriented attribute) using the probability density function is as follows: in, This parameter is used to characterize the degree of random fluctuation in user speed, thereby avoiding decision oscillations caused by minor speed perturbations and enhancing the robustness of the judgment; This is the spatial dimension scaling correction factor, taking into account that the terminal often does not travel along the ideal longest diameter when traversing a cell, utilizing... The theoretical geometric coverage is mapped to the effective expected distance that the terminal actually traverses the cell; Step 5: Use vector normalization to make the initial decision matrix dimensionless. Let the standardized decision matrix be... Its elements The calculation formula is as follows: Through this transformation, the initial attribute values ​​are mapped to the unit vector space, and the indicators of each dimension are transformed into dimensionless standardized values. This not only preserves the relative proportional relationship between the attributes of each base station, but also provides a unified data foundation for subsequent ranking of the ideal solution of TOPSIS by combining combined weights.

2. The dense heterogeneous network handover method based on user location prediction and multi-attribute decision-making according to claim 1, characterized in that, Step 5 includes, under accessible constraints Next, we will search for a comprehensive utility evaluation function for the target community. The process of maximizing the optimal solution can be expressed as follows: 。