Adaptive Application Deployment Methods in Mobility Prediction-Based Edge Computing Networks

Abstract

This invention discloses an adaptive application deployment method in an edge computing network based on mobility prediction. The deployment framework implemented by this method is based on a general adaptive mobility prediction framework, achieving good accuracy for different movement trajectories while saving on model design and training costs. An adaptive application deployment method is proposed, using mobile user mobility predictions over multiple future time periods as input. The goal is to optimize and maximize user service quality while minimizing deployment costs at edge nodes, dynamically deploying mobile user services to the most suitable edge nodes. This invention constructs a joint optimization mechanism to determine the deployment location and duration of user applications.

Description

Technical Field

[0001] This invention pertains to network resource management technology, specifically relating to the deployment of applications in edge computing networks, and more particularly to a method for deploying applications in edge computing networks based on mobility prediction. Background Technology

[0002] Mobile devices, due to their portability requirements, are typically limited in size, weight, and battery capacity, resulting in a scarcity of computing and storage resources. This severely hinders the execution of many latency-sensitive and computationally intensive tasks. While edge computing offers more real-time, convenient, and even flexible computing services compared to traditional cloud computing, it also has inherent limitations. For example, edge servers typically have limited computing power and storage resources, and their access coverage is generally limited. Furthermore, the mobility of end users leads to frequent switching across gateways, making it difficult to guarantee service continuity.

[0003] In recent years, with the rapid development of big data technology and artificial intelligence algorithms, and the diverse user mobile data generated by various mobile devices, and with the increasing demand for personalized applications, there is an urgent need to study the dynamic deployment optimization of edge computing application services for ubiquitous mobile scenarios. Simultaneously, leveraging the excellent capabilities of deep learning on time-series datasets, we are prompted to consider combining user mobility prediction and proactive application deployment as a solution to ensure service continuity in edge computing. Jointly considering research on mobility prediction and application deployment in the edge computing field is necessary and crucial for the following reasons:

[0004] 1) In the future, more portable mobile devices will connect to base stations, leading to greater demand for computational offloading. However, due to varying user mobility patterns, current mobility prediction algorithms and frameworks are only applicable to specific users or those with similar mobility patterns. Edge computing scenarios involve a large number of mobile devices, resulting in low prediction efficiency. Furthermore, general-purpose predictive frameworks suffer from low prediction accuracy, leading to inefficient deployment algorithms for some mobility prediction-based applications. Therefore, researching prediction models that can adaptively adjust to different mobility trajectories to ensure prediction accuracy while significantly reducing model design and training costs is crucial.

[0005] 2) User mobility and computing load are time-varying, and application placement and migration should also be able to adaptively adjust to the movement of mobile users, thereby providing users with higher service quality. At the same time, the computing power and storage resources of edge nodes are limited, and the cost of deploying applications on edge nodes also needs to be considered.

[0006] While research on mobility prediction and application deployment is valuable, it also faces certain challenges, for the following reasons:

[0007] a) When the complexity of the predictive model matches the complexity of the data, its generalization performance is robust. Overfitting may occur when the complexity of the data exceeds the complexity of the model, and underfitting may occur if the data complexity is less than the model complexity. If different predictive models are designed for different mobile users, the model design cost and training time increase dramatically with the number of users. However, if a single model is designed to predict all mobile trajectories, there is a significant number of users with poor mobility prediction accuracy.

[0008] (b) To improve user quality for heterogeneous mobile users in dynamic edge computing environments, it is necessary to simultaneously optimize the deployment cycle and location of applications. However, this raises another optimization problem: minimizing the deployment cost of edge nodes involves the allocation of computing resources to those nodes. These decision variables are highly dependent on the accuracy of mobility predictions over multiple future time periods. Furthermore, there is an inherent conflict between prediction accuracy and the prediction timeframe, further complicating this mobility prediction-based application deployment problem. Summary of the Invention

[0009] Purpose of the invention: To address the problems of insufficient resource optimization for edge server deployment and difficulty in guaranteeing user service quality in mobile scenarios in the prior art, this invention provides an adaptive application deployment method in an edge computing network based on mobility prediction.

[0010] Technical Solution: An adaptive application deployment method in an edge computing network based on mobility prediction. The method is applied to an application deployment system that provides dynamic allocation of applications for mobile users based on edge computing servers. The method is designed for mobile users with different mobility modes, aiming to maximize the quality of service for users and minimize the deployment cost of edge nodes. Based on a mobility prediction model, the method dynamically deploys the service programs of mobile users to the corresponding edge nodes.

[0011] The method for establishing a mobile user application deployment model in an edge computing scenario includes the following steps:

[0012] (1) Determine the decision variables for the application deployment system, including adjusting the location μ of the service application deployment for each user. i,j And the timeframe for advance deployment of service applications for each user. i,j ;

[0013] (2) Collect user mobility data and design an adaptive mobility prediction framework within a period W.i,j Within this range, output the location of the mobile user for one or more future time periods. and future edge nodes

[0014] (3) Calculate the deployment overhead F of the edge computing network, including the application deployment cost costi on the edge nodes. ,j and service delays for mobile users

[0015] (4) Based on the mobility prediction results for several future time periods, optimize the application deployment location and cycle for mobile users, and construct a function to minimize service latency and deployment cost in the edge computing network. The function is based on the application deployment optimization problem in the edge computing network based on mobility prediction, and its expression is as follows:

[0016]

[0017] Constraints:

[0018]

[0019]

[0020]

[0021] In the formula, μ i,j W indicates the location of the edge nodes where the user's application needs to be deployed. i Indicates the deployment cycle required for the application; p i p j These represent the computing resources consumed by the task and the processor resources owned by the edge node, respectively. This indicates the effective deployment cycle for achieving the required accuracy.

[0022] (5) Solve the application deployment optimization problem for mobile users through an adaptive deployment algorithm, and complete the dynamic deployment of mobile users' computing tasks to the edge computing server.

[0023] Furthermore, in step (1), each mobile user i∈N generates a single type of computational task, which is offloaded to an edge node j∈M to receive related services; during the pre-deployed period W i,j The longest future time is divided into several time slots, and the creation and unloading of tasks occur within each time slot.

[0024] Furthermore, the user mobility data in step (2) includes the user's mobility rate and mobility path, taking into account possible handover scenarios for access to base stations along the user's mobility path, including routing overhead caused by access handover. and deployment cost i,j The calculation, and in terms of service latency and deployment cost i,j Minimum is the target.

[0025] In step (2), the edge nodes for future access are obtained through the mobility prediction framework. Due to the computing power p of edge servers j It is restricted, and the specific mathematical expression is as follows:

[0026]

[0027] Since it is not possible to deploy the application services of all users who have switched access to the corresponding edge nodes, the deployment location of the application services of certain users is dynamically adjusted and deployed to the edge nodes of the access network. In nearby low-load edge nodes.

[0028] Furthermore, the mobility prediction framework described in step (2) is implemented based on the LSTM network model with residual structure. It takes the user's location coordinates and mobility rate as features, outputs the user's future multi-time slot location coordinates and access base station number, and consists of a basic block and three residual blocks. The basic block and the residual blocks both contain a normalization layer and a dropout layer.

[0029] Basic blocks are used to enable at least one stacked LSTM network to extract temporal information from long-term positional relationships;

[0030] Residual blocks are used to achieve the performance of this mobility prediction framework, which outperforms a single stacked LSTM network framework.

[0031] The mobility prediction framework ultimately obtains future location information and access base station information through a fully connected layer.

[0032] Furthermore, in step (3), the mathematical expression for the deployment cost F of the entire system in the edge computing network is as follows:

[0033]

[0034] Among them, service delay The expression is:

[0035]

[0036] Service delay The time it takes for each mobile user task to be unloaded Task calculation time And routing time caused by the lack of applications on the access nodes. composition;

[0037] Among them, uninstallation time The expression is:

[0038]

[0039] Calculation time The expression is:

[0040]

[0041] Routing time The expression is:

[0042]

[0043] Additionally, F includes deployment costs. i,j Its expression is:

[0044] cost i,j =μ i,j app i / W i

[0045] Where β is a scaling factor used to balance deployment costs and service latency, and θ represents the propagation speed of light; This indicates the size of the task, measured in bits per second (b). i p represents the rate at which users uninstall tasks. i This indicates the processor speed allocated to the task. Indicates the number of processor revolutions required for the task, app i This indicates the storage capacity of the service program required to process the user's corresponding task; The distance between the user and the edge node is represented using Euclidean distance:

[0046]

[0047] Furthermore, in step (5), the deployment of applications and computing tasks for mobile users includes the following process:

[0048] (51) Feed the historical data of mobile users into the RELS mobility prediction framework, train the model, and obtain the location coordinates and access base station numbers for multiple time slots in the future.

[0049] (52) Based on the prediction accuracy requirements, the range of values ​​for the application deployment cycle for each mobile user is selected and obtained.

[0050] (53) Calculate the pre-deployment cost of each user's application in different edge nodes and deployment cycles. The pre-deployment cost consists of the user's service latency and the deployment cost of the edge nodes.

[0051] (54) The problem is transformed into a resource-constrained generalized assignment problem, and the optimal deployment period and deployment location for each user are obtained through greedy algorithms or dynamic programming.

[0052] Furthermore, the above method, during deployment, first determines the range WWiitth of the application deployment cycle for each user based on the required prediction accuracy. Secondly, it calculates the pre-deployment cost for each application across different edge nodes and deployment cycles. Finally, a greedy algorithm is used to obtain the optimal deployment strategy.

[0053] Beneficial effects: Compared with the prior art, the significant features and substantial progress of this invention include the following three points:

[0054] First, this invention comprehensively considers the factors affecting the deployment of mobile user applications on corresponding edge nodes, including user mobility, limited storage space and computing power of edge nodes, the trade-off between providing high-quality services and deployment costs, and the accuracy of predictions for multiple future time slots.

[0055] Secondly, this invention designs a mobility prediction framework that takes into account the individual differences in different mobility data. It can achieve good accuracy in predicting different movement trajectories while saving on model design and training costs. Furthermore, it considers the mobile user's movement rate and path, and comprehensively considers the handover characteristics of base stations along the route to improve prediction accuracy and minimize system overhead.

[0056] Third, the present invention adopts adaptive deployment, taking the mobility prediction of mobile users in multiple time periods in the future as input, optimizing the deployment cycle and location of user applications, which can ensure the service quality of mobile users in multi-access edge environments and obtain cost-effective edge computing services. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the system model structure constructed by the framework described in this invention;

[0058] Figure 2 This is a system flowchart of the framework described in this invention;

[0059] Figure 3 This is a diagram of the model structure for mobility prediction in this invention;

[0060] Figure 4 The bar chart shows the user mobility accuracy under different prediction periods in the example.

[0061] Figure 5 This is a line graph showing the service latency at different transmission rates for edge nodes in the example.

[0062] Figure 6 The example shows a line graph illustrating the deployment cost of edge nodes at different transmission rates. Detailed Implementation

[0063] To illustrate the technical solutions disclosed in this invention in detail, the invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0064] First, the framework described in this invention primarily addresses the problem of minimizing user service latency and edge node deployment costs in dynamic edge computing scenarios by adjusting the deployment cycle and location. A specific system model is as follows: Figure 1 As shown.

[0065] The main idea of ​​this invention is to collect past mobility data from mobile users, design a general mobility prediction framework to train on the mobility data, and thus obtain future mobility information for users. Using this prediction information, future time periods that meet deployment conditions are selected, and the pre-deployment cost of all applications on each edge node is estimated over all feasible deployment cycles. Based on the pre-deployment cost, a corresponding deployment algorithm is executed to obtain the optimal deployment strategy, such as... Figure 2 As shown.

[0066] Specifically, an adaptive application deployment method in an edge computing network based on mobility prediction includes the following steps:

[0067] Step 1: Establish a mobile user application deployment model for edge computing scenarios.

[0068] This step will mathematically analyze the practical problems of adaptive application deployment frameworks in edge computing networks based on mobility prediction.

[0069] Based on the actual application scenario and the deployment of edge nodes, each edge node contains an access point (i.e., a base station) where mobile users offload their tasks, and an edge server that provides computing services for tasks with installed service programs. This is represented by a set M, with cardinality |M| = m. Simultaneously, there exists a group of mobile users with heterogeneous mobility patterns, such as bicycles, cars, and trains, represented by a set N, with cardinality |N| = n. Each mobile user i ∈ N generates a single type of computing task, such as transportation system computing services, web browsing, or navigation route planning, which needs to be offloaded to an edge node j ∈ M to receive the relevant services. Time is divided into several time slots, within which tasks are generated and offloaded.

[0070] Each user i has the following specifications: app i p represents the amount of storage space required by the service application for user i; i b represents the processor computing resources required for user i's computational task. i This represents the uplink data rate for user i. Furthermore, within each time slot t, the location coordinates of each user are represented as follows: This depends on its movement pattern. Furthermore, the data size of user i's computational task is defined as... The CPU cycles required for each computing task are defined as

[0071] Each edge node j has the following specifications: (x j ,y j This indicates the position coordinates of edge node j. j This indicates the computing resources of the processors owned by edge node j, while b j This represents the transmission rate of edge node j.

[0072] The primary objective of this invention is to minimize overall average service latency and deployment cost based on user movement trajectories. Therefore, with the goal of minimizing the overall deployment cost F, the entire edge computing system can make corresponding decisions, including the following:

[0073] a) Adjust the deployment location of the service application for each user. i,j

[0074] b) Determine the pre-deployment cycle μ for each user's service application. i,j

[0075] The following constraints must be met:

[0076]

[0077]

[0078]

[0079] Step 2: Collect user mobility data and design an adaptive mobility prediction framework to achieve mobility prediction.

[0080] Different users exhibit different mobility patterns, manifested in variations in mobility speed and movement regularity, resulting in varying access handover frequencies. Due to these varying handover frequencies, a uniform deployment cycle presents the following problems: a longer deployment cycle increases service latency, while a shorter cycle incurs additional deployment costs. Therefore, the deployment cycle of mobile user services should be flexibly adjusted to reduce service latency and deployment costs. Furthermore, since edge node computing resources are limited, it is impossible to deploy corresponding services for all mobile users who will undergo future access handovers. Therefore, dynamic decision-making regarding the deployment location of mobile users is necessary.

[0081] Furthermore, considering that mobility trajectory data is considered to be time-series data collected at different points in time, it is readily available. Therefore, this invention can utilize this data for mobility prediction to guide the deployment of proactive applications.

[0082] To address the problem of mobile trajectory prediction, this invention combines neural networks with machine learning, using RNNs for time series prediction tasks. However, RNNs suffer from several issues, such as vanishing and exploding gradients, and limited ability to handle long-term dependencies. Therefore, training struggles to converge, resulting in an unstable model. Instead, a variant of RNN, the so-called LSTM, is utilized by introducing storage units and three gates (input gate, forget gate, and output gate) to capture long-term information. Considering that deep network structures can learn more temporal dependencies in user movement trajectories, a stacked LSTM network structure is considered. Due to the varying amounts and complexities of mobile datasets, the model structure and its parameters need to be dynamically adjusted for each user. However, for small or simple datasets (i.e., regular movement patterns), prediction errors increase with the number of network layers (i.e., "degradation"). Residual networks allow deeper networks to address degradation by providing shortcuts across layers, thus achieving at least the performance of shallower networks. Therefore, a residual LSTM framework is designed to predict the future movement patterns of mobile users. This framework can be trained to obtain the most suitable model parameters for each user. This prediction framework uses the user's latitude and longitude location coordinates as training features. Then, it calculates the base station the user will connect to in the future based on the predicted location coordinates.

[0083] The mobility prediction framework consists of a base block and three residual blocks, as follows: Figure 3The former guarantees that at least one stacked LSTM network can extract temporal information from long-term positional relationships, while the latter guarantees that its performance is always better than a framework with only one stacked LSTM network. Both the basic block and the residual block consist of a stacked LSTM, a layer normalization layer, and a dropout layer. The layer normalization layer adjusts the mean and variance of the computed intermediate data, making the transmission of hidden states more stable and the training process faster. The dropout layer placed at the end of each block reduces overfitting, making the prediction model more robust. Finally, the extracted high-level semantic features are flattened and fed into a fully connected layer to obtain future positional information.

[0084] Step 3: Calculate the deployment cost F of the edge computing network and solve it using an adaptive deployment algorithm.

[0085] In the aforementioned edge computing scenario, to construct a mobile user application deployment system model, with the goal of minimizing overall average service latency and deployment cost, the mathematical expression for the deployment overhead F of the entire system in the edge computing network is as follows:

[0086]

[0087] Among them, service delay The expression is:

[0088]

[0089] Service delay The time it takes for each mobile user task to be unloaded Task calculation time And routing time caused by the lack of applications on the access nodes. composition.

[0090] Among them, uninstallation time The expression is:

[0091]

[0092] Calculation time The expression is:

[0093]

[0094] Routing time The expression is:

[0095]

[0096] Additionally, F includes deployment costs. i,j Its expression is:

[0097] costi,j =μ i,j app i / W i

[0098] Here, β is a scaling factor used to balance deployment costs and service latency. θ represents the speed of light, numerically 3 * 10⁻⁶. 8 m / s. This indicates the size of the task, measured in bits per second (b). i p represents the rate at which the user uninstalls tasks. i This indicates the processor speed allocated to the task. Indicates the number of processor revolutions required for the task, app i This indicates the storage capacity of the service programs required to process the user's corresponding tasks. The distance between the user and the edge node is represented using Euclidean distance:

[0099]

[0100] Here, β is a scaling factor used to balance deployment costs and service latency.

[0101] Step 4: Based on the different mobile user mobility patterns, and considering the prediction effect of the mobile prediction framework for different prediction durations, design a low-complexity adaptive application deployment method that balances user service quality and server deployment costs, and uses mobility prediction results from multiple future time periods to make the best deployment decision.

[0102] The application deployment optimization problem in this invention is solved, thereby enabling the offloading of computing tasks for mobile users. The implementation process is as follows:

[0103] 1) Feed historical data of mobile users into the RELS mobility prediction framework, train the model, and obtain the location coordinates and access base station numbers for multiple future time slots.

[0104] 2) Based on the required prediction accuracy, filter and obtain the range of values ​​for the application deployment cycle for each mobile user.

[0105] 3) Calculate the pre-deployment cost of each user's application across different edge nodes and deployment cycles. The pre-deployment cost consists of the user's service latency and the deployment cost of the edge nodes.

[0106] 4) Transform the problem into a resource-constrained generalized assignment problem, and obtain the optimal deployment cycle and deployment location for each user through a greedy algorithm or dynamic programming.

[0107] For the corresponding algorithm implementation, the algorithm receives three input parameters, which are:

[0108] a) Minimum acceptable prediction accuracy th , used to control the accuracy of prediction;

[0109] b) User's historical mobility data U;

[0110] c) The proportionality factor β that balances deployment costs and service delays.

[0111] Finally, it should be noted that this invention first obtains multiple future time-period mobility prediction tables through the residual LSTM framework in step 2. These tables contain the accuracy of the prediction results for each future time period for all users. (According to acc...) th The deployment cycle range M for all user applications is obtained by querying this table. Next, all potential pre-deployment costs are estimated, which is the expected weighted sum of service latency and deployment costs for each user's application deployed to each edge node over all reasonable deployment cycles. To obtain these potential deployment costs, the method mentioned in step 1 will be used. and cost i,j The calculation formula is as follows. It's important to note that additional routing costs are only calculated when the deployed edge nodes differ from the predicted access edge nodes. Finally, based on these estimated potential pre-deployment costs, a greedy algorithm is used to obtain the optimal application deployment strategy.

[0112] To fully illustrate the adaptive application deployment framework in the edge computing network based on mobility prediction proposed in this invention, performance is evaluated using the following three metrics:

[0113] (1) The accuracy of user mobility prediction;

[0114] (2) Overall service delay for all users;

[0115] (3) Deployment cost of all edge nodes.

[0116] In this embodiment, a neural network model is built using the PyTorch deep learning architecture. All algorithms and simulation experiments written in Python 3.9 are completed on a PC with a 2.9GHz CPU and 16GB of memory. It is assumed that there are 40 users and 9 edge nodes, each with a different mobility pattern. The average value is selected after numerous experiments. The adaptive application deployment framework in the mobility prediction-based edge computing network proposed in this invention is called AADA; the fixed-period deployment framework is called FADA; the pure migration scheme following the mobility pattern is called Mig; and the strategy of always assigning the task to the initial edge node for computation is called NoMig. Simultaneously, the residual LSTM mobility prediction architecture proposed in this invention is called RELS; the residual LSTM mobility prediction architecture based on base station ID features is called RELS_BS; the prediction model using reinforcement learning for adaptive adjustment is called TL_RL_LSTM; and the prediction model using a comprehensive search neural network is called GS_LSTM.

[0117] like Figure 4 As shown, the RELS mobility prediction architecture proposed in this invention outperforms other models in prediction accuracy, especially in predicting longer timeframes. Figure 5 As shown, the AADA architecture proposed in this invention significantly outperforms other architectures in terms of user service latency. While AADA's service latency increases with the increase in edge node transmission rate, this is because AADA considers the trade-off between deployment cost and user service latency, sacrificing some service latency to ensure lower deployment costs. Figure 6 As shown, the AADA algorithm not only has very low service latency, but also relatively low deployment cost (NoMig has a deployment cost of 0 because it does not require deploying a new application). In summary... Figure 4 , Figure 5 and Figure 6 This demonstrated the superior performance of the application deployment framework in the mobility prediction-based edge computing network described in this invention.

Claims

1. An adaptive application deployment method in an edge computing network based on mobility prediction, the method being applied to an application deployment system that provides dynamic allocation of applications for mobile users based on edge computing server architecture, characterized in that: The method targets mobile users with different mobility modes, aiming to maximize the quality of service for users and minimize the deployment cost of edge nodes. Based on a mobility prediction model, it dynamically deploys the service programs of mobile users to the corresponding edge nodes. The method for establishing a mobile user application deployment model in an edge computing scenario includes the following steps: (1) Determine the decision variables for the application deployment system, including adjusting the location of service application deployment for each user. And the decision-making process for the pre-deployment cycle of service applications for each user. ; Where i represents the user and j represents the edge node; (2) Collect user mobility data and design an adaptive mobility prediction framework in a periodic manner. Within this range, output the location of the mobile user for one or more future time periods. and future edge nodes ; Location( () represents the predicted position coordinates of user i in time slot t, and the superscript pre indicates the prediction marker; (3) Calculate the deployment overhead of the edge computing network This includes application deployment costs on edge nodes. and service delays for mobile users ; (4) Based on the mobility prediction results for several future time periods, optimize the application deployment location and cycle for mobile users, and construct a function to minimize service latency and deployment cost in the edge computing network. The function is based on the application deployment optimization problem in the edge computing network based on mobility prediction, and its expression is as follows: ， Constraints: , In the formula, This indicates the location of the edge nodes where the user's application needs to be deployed. Indicates the deployment cycle required for the application; , These represent the computing resources consumed by the task and the processor resources owned by the edge node, respectively. This indicates the effective deployment period to achieve the required accuracy; the superscript "th" indicates the upper limit. (5) Solve the deployment optimization problem of the application corresponding to the mobile user through the adaptive deployment algorithm, and complete the dynamic deployment of the mobile user's computing tasks to the edge computing server.

2. The adaptive application deployment method in an edge computing network based on mobility prediction according to claim 1, characterized in that: In step (1), each mobile user Generate a single type of computational task and offload it to the edge node. To receive relevant services; within the pre-deployment period. The longest future time is divided into several time slots, and the creation and unloading of tasks occur within each time slot.

3. The adaptive application deployment method in an edge computing network based on mobility prediction according to claim 1, characterized in that: The user mobility data in step (2) includes the user's movement rate and movement path. The handover scenarios for possible access to base stations along the user's movement path are considered, including the routing overhead caused by the access handover. and deployment costs The calculation, and in terms of service latency and deployment costs Minimum is the target; k represents the edge node different from j, and the routing cost. This represents the routing delay for user i's computation task to travel from edge node j to edge node k in time slot t.

4. The adaptive application deployment method in an edge computing network based on mobility prediction according to claim 3, characterized in that: In step (2), the edge nodes for future access are obtained through the mobility prediction framework. Due to the computing power of edge servers It is restricted, and the specific mathematical expression is as follows: ， Since it is not possible to deploy the application services of all users who have switched access to the corresponding edge nodes, the deployment location of the application services of certain users is dynamically adjusted and deployed to the edge nodes of the access network. In nearby low-load edge nodes.

5. The adaptive application deployment method in an edge computing network based on mobility prediction according to claim 1 or 4, characterized in that: The mobility prediction framework described in step (2) is implemented based on the LSTM network model with residual structure. It takes the user's location coordinates and mobility rate as features and outputs the user's future multi-time slot location coordinates and access base station number. It consists of a basic block and three residual blocks. The basic block and the residual blocks both contain a normalization layer and a dropout layer. Basic blocks are used to enable at least one stacked LSTM network to extract temporal information from long-term positional relationships; Residual blocks are used to achieve the performance of this mobility prediction framework, which outperforms a single stacked LSTM network framework. The mobility prediction framework ultimately obtains future location information and access base station information through a fully connected layer.

6. The adaptive application deployment method in an edge computing network based on mobility prediction according to claim 1, characterized in that: In step (3), the deployment overhead of the entire system in the edge computing network The mathematical expression for is as follows: ， Among them, service delay The expression is: ， Service delay The time it takes for each mobile user task to be unloaded Task calculation time And routing time caused by the lack of applications on the access nodes. composition; Among them, uninstallation time The expression is: ， Calculation time The expression is: , Routing time The expression is: ， Additionally, F includes deployment costs. Its expression is: ， in, It is a scaling factor used to balance deployment costs and service latency. Indicates the speed at which light travels; This indicates the size of the task, measured in bits per second. This indicates the rate at which the user uninstalls tasks. This indicates the processor speed allocated to the task. This indicates the number of processor revolutions required for the task. This indicates the storage capacity of the service program required to process the user's corresponding task; The distance between the user and the edge node is represented using Euclidean distance: 。 7. The adaptive application deployment method in an edge computing network based on mobility prediction according to claim 1, characterized in that: In step (5), the deployment of applications and computing tasks for mobile users includes the following process: (51) Feed the historical data of mobile users into the RELS mobility prediction framework, train the model, and obtain the location coordinates and access base station numbers for multiple time slots in the future; (52) Based on the required prediction accuracy, filter and obtain the range of values ​​for the application deployment cycle for each mobile user. ; (53) Calculate the pre-deployment cost of each user's application in different edge nodes and deployment cycles. The pre-deployment cost consists of the user's service latency and the deployment cost of the edge nodes. (54) The problem is transformed into a resource-constrained generalized assignment problem, and the optimal deployment period and deployment location for each user are obtained through greedy algorithm or dynamic programming.

Images