A Service Collaborative Caching Method in Vehicle Edge Computing

Abstract

This invention discloses a service collaborative caching method in vehicle-mounted edge computing, comprising the following steps: (i) predicting the service request volume based on historical access records using an ARIMA model; (ii) establishing a service latency model, including: 1) a latency model from the vehicle terminal to the edge server; 2) a latency model generated by transmission between edge nodes; 3) a computation latency model generated by the edge server processing requests; and 4) a total service request latency model; (iii) establishing a service caching and resource allocation model: 1) establishing a state space model using deep reinforcement learning; 2) establishing an action space model using deep reinforcement learning; 3) setting a reward function; and 4) obtaining the maximum reward value. This invention optimizes the service processing flow, improves service processing efficiency, and effectively reduces service response latency by accurately predicting service types, establishing a service processing latency model, and rationally applying caching and resource allocation.

Description

Technical Field

[0001] This invention relates to the field of edge computing, and more particularly to a service collaborative caching method in vehicle-mounted edge computing. Background Technology

[0002] In automotive edge computing scenarios, due to the limited computing power and resources of vehicles, service requests from vehicles need to be executed by edge servers before being returned to the vehicle terminal. To ensure service quality, response latency should be minimized. Therefore, in edge computing, we need to pre-determine which services should be deployed on which edge servers and how much computing resources to allocate, in order to achieve service caching (also known as dynamic service deployment). Existing research largely focuses on the vertical architecture between the user end and the edge end, considering service caching from a vertical perspective. The focus is on edge server energy consumption and independent service response latency, which, while improving user experience to some extent, does not adequately consider the relationships between services, i.e., insufficient consideration at the horizontal service level. In reality, services requested by users often involve data interaction. If related services are not cached on the same edge server, additional data transmission latency will occur. This latency is not negligible and can be eliminated through reasonable caching strategies. Conversely, if related services are cached on the same edge server, data transmission latency will not occur, thus improving user experience. Therefore, the importance of a service caching and resource allocation strategy that considers service relationships is self-evident.

[0003] Therefore, how to optimize the models, strategies, and resource utilization in vehicle edge computing to obtain the shortest service response latency has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a service collaborative caching method in vehicle edge computing. By accurately predicting service types, establishing a service processing latency model, and rationally applying caching and resource allocation, the method optimizes the service processing flow, improves service processing efficiency, and effectively reduces service response latency.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: a service collaborative caching method in vehicle edge computing, characterized by comprising the following steps:

[0006] (i) Based on historical access records, predict the number of service requests using the ARIMA model;

[0007] (ii) Establishing a service latency model, which includes:

[0008] 1) Latency model from vehicle terminal to edge server:

[0009]

[0010] in: Let c(k,i,t) be the total number of service requests k received by all nodes in time slot t, and let c(k,i,t) be the number of service requests k received by edge node i in time slot t. edge Let d(k) represent the transmission rate between the vehicle terminal and the edge server, and d(k) represent the storage space required to place service k, which satisfies... D(i) represents the storage capacity of edge server i;

[0011] 2) The latency model for transmission between edge nodes is as follows:

[0012]

[0013] Where: the transmission rate between edge nodes is v bet C(k,t)-c(k,i,t) represents the number of services that need to be transmitted to edge server i for processing; b(k,i,t)∈{0,1} is a binary variable. If the service is placed on edge node i, then b(k,i,t) = 1, otherwise b(i,k,t) = 0. Therefore, b(k,i,t)∈{0,1},k∈K,i∈S.

[0014] 3) The computational latency model generated by the edge server processing requests is as follows:

[0015]

[0016] Where m(k) represents the computing resources required for service request k, l(k,i,t)M(i) represents the size of computing resources allocated by server i to service k, M(i) represents the computing power of edge server i, and l(k,i,t) should satisfy l(k,i,t)∈[0,1],k∈K,i∈S. When the service is placed on the edge server,

[0017] 4) The total latency model for service requests is as follows:

[0018]

[0019] Among them, service k and service k * These are service pairs with a relationship; the optimization objective, aimed at reducing the average response latency of tasks, is expressed as:

[0020]

[0021] (III) Establishing a service caching and resource allocation model:

[0022] 1) Establish the state-space model S for deep reinforcement learning * (i,t):

[0023] S * (i,t)={c(k,i,t),M(i),D i ,l(k,i,t)};

[0024] 2) Establish an action space model A for deep reinforcement learning. i (t):

[0025] A i (t)={b(k,i,t),Δl(k,i,t),k∈K};

[0026] 3) Set the reward function R(t), let make but

[0027] Among them, state Take action A t Get the state Δw represents the response time difference between the two states, and α is a constant;

[0028] 4) Maximize reward value: First define Q(s,a) as the action value, then use an a * Let Q(s,a) be used to represent maximizing the action value:

[0029] a * =argmax a∈A Q(s,a);

[0030] Here, to calculate the difference between the forward computation result and the true value in each iteration of the neural network, and thus guide the next step of training in the correct direction, a loss function is used: The weight values ​​in the forward calculation formula are corrected; the parameter θ is updated using gradient descent: θ t+1 =θ t -η▽L(θ t ); where η is the learning rate.

[0031] Furthermore, in step (i), the specific prediction process for the service request volume is as follows:

[0032] 1) Collect the actual historical records of user service calls and use them as the raw dataset;

[0033] 2) Examine the service call time series data to determine if it is a white noise sequence. If so, use the ARIMA model for modeling and prediction; otherwise, use the arithmetic mean method to predict the service request volume using the following formula:

[0034]

[0035] 3) ARIMA modeling, denoted by ARMA(p,q), satisfies Where, φ o φ is a constant term. i and θ j These are the parameters for the AR model and the MA model, respectively; a t It is random error, i.e., white noise; the non-negative integers p, q and d represent the order of AR, MA and the order of difference, respectively;

[0036] 4) Use time series forecasting methods to predict future values, and perform a stationarity test on the original time series. If it is not stationary, then modify the original time series {x}. t The time series is transformed into a stationary sequence by performing d interpolation operations on the sequence |t=1,2,...,n}; then an ARMA model is constructed as a candidate model for the time series.

[0037] 5) Calculate the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the constructed ARMA model, and select appropriate values ​​for model parameters p and q based on ACF and PAC; where p is the AR model parameter and q is the MA model parameter; and perform residual tests on the model to verify its effectiveness, and select significantly effective models as candidate models.

[0038] 6) Calculate the Information Criterion (AIC) value among the candidate models, and select the model with the smallest AIC value as the prediction model to predict the number of service requests.

[0039] Furthermore, in step (ii), it is assumed that there are several edge nodes capable of deploying services and mobile users requesting services in a certain area. The set of edge nodes is represented as S = {1, 2, 3, ..., s}, D(i) represents the storage space size of edge node i, and M(i) represents the computing resources of edge node i. Edge nodes consume computing resources to process service requests. It is assumed that there is a set of service requests K = {1, 2, 3, ..., k}, d(k) represents the storage space required to deploy service k, and m(k) represents the computing resources required for service request k.

[0040] First, define a binary variable b(k,i,t)∈{0,1} to determine whether service k is placed on edge node i in time slot t. If the service is placed on edge node i, set b(k,i,t)=1; otherwise, set b(i,k,t)=0. Therefore, the service placement decision is defined as B(t)={b(k,i,t)|i∈S,k∈K}.

[0041] Furthermore, in step (ii) 3), such as service k and service k * These are service pairs with a relationship, where the number of preceding services k processed by edge node i in time slot t is c. comp Given (k,i,t), the number of services k completed by all nodes in time slot t is:

[0042]

[0043] The transmission delay caused by data interaction is:

[0044]

[0045] Where d(kk*) represents the associated service pair k and k * The amount of data that needs to be interacted with; v bet Calculate the back-end service k for the data transfer rate between edge nodes. * The resulting delay is:

[0046]

[0047] in, Indicates the processing of service request k * The required computing resources; where m(k) represents the computing resources required for service request k, and l(k,i,t)M(i) represents the size of computing resources allocated by server i to service k;

[0048] For handling association relationships, k and k * The total delay is:

[0049]

[0050] When processing a service k that has no relation to another service, the total latency is:

[0051]

[0052] Compared with existing technologies, the present invention has the following advantages: it predicts service request volume based on historical access records, while also considering the correlation between services, and performs unified caching and computing resource allocation for services in the edge server network; it establishes a relatively complete latency model and trains it using deep reinforcement learning, and performs service caching and computing resource allocation according to different edge server storage capacities and computing resource sizes; and provides users with low-latency, high-quality services by optimizing the latency of user service calls and the data interaction latency between related services when terminal node resources are limited.

[0053] The purpose of service request volume prediction is to identify patterns in service requests from historical data. Having this data makes service caching and resource allocation based on deep reinforcement learning more reliable. Establishing a service execution latency model facilitates the phased analysis of latency generation, identifying areas for optimization and fully leveraging the advantages of deep reinforcement learning. Both of these points lay the foundation for better utilization of deep reinforcement learning algorithms: with predicted service request volume data, deep reinforcement learning algorithms can implement caching and resource allocation strategies tailored to specific situations; with a service latency model, the algorithm can optimize the network from iterative results, ultimately converging to the optimal solution. In other words, the former provides reliability, and the latter provides feasibility. Attached Figure Description

[0054] Figure 1 Example plot of ACF for time series;

[0055] Figure 2 Example diagram of PACF;

[0056] Figure 3 Example of service access prediction. Detailed Implementation

[0057] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0058] Example: A service collaborative caching method in vehicle edge computing, comprising the following steps:

[0059] (i) Based on historical access records, predict the number of service requests using the ARIMA model.

[0060] Specifically,

[0061] 1) Collect the actual historical records of user service calls and use them as the raw dataset.

[0062] 2) Examine the service call time series data to determine if it is a white noise sequence. If it is, use the ARIMA model for modeling and prediction; otherwise, use the arithmetic mean method to predict the service request volume using the following formula:

[0063]

[0064] 3) ARIMA modeling, denoted by ARMA(p,q), satisfies Where φ o φ is a constant term. i and θ j These are the parameters for the AR model and the MA model, respectively; a t It is random error, i.e., white noise; the non-negative integers p, q and d represent the order of AR, MA and the order of difference, respectively.

[0065] 4) Use time series forecasting methods to predict future values, and perform a stationarity test on the original time series. If it is not stationary, then modify the original time series {x}. t The time series is transformed into a stationary sequence by performing d interpolation operations on the sequence |t=1,2,...,n}; then an ARMA model is constructed as a candidate model for the time series.

[0066] 5) Calculate the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the constructed ARMA model. Select appropriate values ​​for the model parameters p and q based on ACF and PAC; where p is the AR model parameter and q is the MA model parameter. Test different combinations of p and q, and perform residual tests on the model to verify its effectiveness. Select significantly effective models as candidate models, such as... Figure 1 , Figure 2 As shown, there are example diagrams of ACF and PACF.

[0067] By observing the characteristics of ACF and PACF, the following models can be constructed as candidate models for time series data. For example, if we select 80 hours from the service request time series data and use the historical data of the first 72 hours to predict the service request volume for the next 8 hours, the following ARIMA models can be constructed using the method described in this patent: ARIMA(1,1,2), ARIMA(2,1,3), ARIMA(3,1,3), ARIMA(2,1,4), and ARIMA(4,1,4). After determining the candidate ARIMA models in the model estimation stage, the estimates of these models are shown in Table 1.

[0068] Table 1. Estimation of candidate models. AIC: Akaike Information Criterion;

[0069]

[0070]

[0071] 6) Calculate the Information Criterion (AIC) value among the candidate models and select the model with the smallest AIC value as the prediction model to predict the service request volume. Table 1 shows that the minimum AIC value is 1695.502; therefore, the ARIMA(2,1,3) model is selected as the optimal model. Model testing confirms the model's significance; therefore, ARIMA(2,1,3) can be chosen to predict future values. The prediction graph is shown below. Figure 3 As shown, the red values ​​represent the predicted values ​​of the method. The root mean square error (RMSE) is used to evaluate the accuracy of the prediction method, and it is calculated as follows:

[0072]

[0073] Where n is the number of predictions, and y i These represent the predicted value and the actual value, respectively. By calculating the RMSE value, we can obtain an RMSE of 10, which provides a highly accurate estimate of the future service request volume.

[0074] (ii) Establishing a service latency model, which includes:

[0075] We assume there are several edge nodes in this area where services can be deployed, as well as mobile users requesting services. The set of edge nodes can be represented as S = {1, 2, 3, ..., s}, where D(i) represents the storage space size of edge node i, and M(i) represents the computing resources of edge node i. Edge nodes consume computing resources to process service requests. Assume there is a set of service requests K = {1, 2, 3, ..., k}, where d(k) represents the storage space required to deploy service k, and m(k) represents the computing resources required for service request k. First, we define a binary variable b(k, i, t) ∈ {0, 1} to determine whether service k is placed on edge node i in time slot t. If the service is placed on edge node i, we set b(k, i, t) = 1; otherwise, we set b(i, k, t) = 0. Therefore, the service placement decision can be defined as B(t) = {b(k, i, t) | i ∈ S, k ∈ K}.

[0076] 1) Latency from the vehicle terminal to the edge server. Assume that the number of service requests k received by edge node i in time slot t is c(k,i,t), then the total number of service requests k received by all nodes in time slot t is... The transmission latency between the terminal and the edge server can be calculated as follows:

[0077]

[0078] in, Let c(k,i,t) be the total number of service requests k received by all nodes in time slot t, and let c(k,i,t) be the number of service requests k received by edge node i in time slot t. edge Let d(k) represent the transmission rate between the vehicle terminal and the edge server, and d(k) represent the storage space required to place service k, which satisfies... D(i) represents the storage capacity of edge server i.

[0079] 2) When service requests need to be forwarded between edge nodes, assume the transmission rate between nodes is v. bet Then, the delay caused by service k transmitting between nodes in time slot t is:

[0080]

[0081] The transmission rate between edge nodes is v. bet C(k,t)-c(k,i,t) represents the number of services that need to be transmitted to edge server i for processing; b(k,i,t)∈{0,1} is a binary variable indicating whether service k is placed on edge node i in time slot t. If the service is placed on edge node i, b(k,i,t) = 1; otherwise, b(i,k,t) = 0. Therefore, b(k,i,t)∈{0,1}, k∈K, i∈S.

[0082] 3) The computational latency incurred by the edge server in processing requests is:

[0083]

[0084] Where m(k) represents the computing resources required for service request k, l(k,i,t)M(i) represents the size of computing resources allocated by server i to service k, M(i) represents the computing power of edge server i, and l(k,i,t) should satisfy l(k,i,t)∈[0,1],k∈K,i∈S, indicating that the allocated computing resources will not exceed the edge server's own capacity. Furthermore, once a service is placed on an edge server, a certain amount of computing resources will always be allocated, i.e.

[0085] If service k and service k * These are service pairs with a relationship, where the number of preceding services k processed by edge node i in time slot t is c. comp Given (k,i,t), the number of services k completed by all nodes in time slot t is:

[0086]

[0087] Therefore, the transmission latency caused by data interaction between related services is:

[0088]

[0089] Where d(kk*) represents the associated service pair k and k * The amount of data that needs to be interacted with; v bet This represents the data transfer rate between edge nodes.

[0090] Compute post-service k * The resulting delay is:

[0091]

[0092] in, Indicates the processing of service request k * The required computing resources; where m(k) represents the computing resources required for service request k, and l(k,i,t)M(i) represents the size of computing resources allocated by server i to service k.

[0093] Since the amount of data returned after the service execution is much smaller than the amount of data that needs to be calculated, the service return latency is negligible. Therefore, for processing the relationship pair k and k * The total delay is:

[0094]

[0095] When processing a service k that has no relation to another service, the total latency is:

[0096]

[0097] 4) In summary, the processing latency for a service request k is:

[0098]

[0099] Among them, service k and service k * These are service pairs that have a relationship;

[0100] With the goal of reducing the average response latency of a task, the optimization objective is expressed as follows:

[0101]

[0102] (III) Establishing a service caching and resource allocation model:

[0103] 1) Establish the state-space model S for deep reinforcement learning *(i,t), including the number of services k received by edge server i in time slot t, c(k,i,t), the storage space D(i) of edge server i, the computing resources M(i) of edge server i, and the proportion of computing resources l(k,i,t) allocated by edge server i to service request k in time slot t, specifically expressed as:

[0104] S * (i,t)={c(k,i,t),M(i),D i ,l(k,i,t)}.

[0105] 2) Establish the action space A for deep reinforcement learning i (t), including deployment variables b(k,i)∈{0,1}, and the smallest allocation unit of the calculated resources Δl(k,i,t)=0.1, therefore, the action space is represented as:

[0106] A i (t)={b(k,i,t),Δl(k,i,t),k∈K}.

[0107] 3) Setting the reward function R(t): Our goal is to minimize service response latency through service placement and resource allocation, making... make but:

[0108]

[0109] Among them, state Take action A t Get the state Δw represents the response time difference between the two states, and α is a constant. The reward function can be understood as follows: if the average response delay does not decrease after taking an action, the reward value is negative; if the average response delay decreases after taking an action, the reward value is set according to the magnitude of the decrease, and the greater the decrease, the greater the reward value.

[0110] 4) To maximize the reward value, we first define Q(s,a) as the action value; then, we can show the expected sum of future rewards over time step T: Where γ∈[0,1] is the discount factor, and Ε represents the expectation for the changing environment. Therefore, we can express the original optimization problem as finding the optimal service placement and resource allocation strategy, using an a * To maximize the action value Q(s,a):

[0111] a * =argmax a∈A Q(s,a)

[0112] Here, to calculate the difference between the forward computation result and the true value in each iteration of the neural network, and thus guide the next step of training in the correct direction, a loss function is used: The weight values ​​in the forward calculation formula are corrected. The parameter θ is updated using gradient descent: θ t+1 =θ t -η▽L(θ t ); where η is the learning rate.

[0113] To address the issues of caching and resource allocation for in-vehicle edge services using deep reinforcement learning, it is essential to first obtain reliable service request volumes in order to formulate more reasonable service deployment strategies. The ARIMA model predicts future service request volumes based on historical access data, and the prediction results undergo RMSE (Reliability, Response, and Execution) to determine their accuracy, thus ensuring the reliability of the prediction data.

[0114] Then, establish a data transmission latency and service computation latency model involved in the service execution process, and analyze various possible situations, including the execution latency of related services and the execution latency of unrelated services. A complete latency model is crucial for evaluating the quality of the algorithm.

[0115] Finally, a service caching and resource allocation algorithm is established. Since this scenario involves integer variables for service deployment and continuous variables for resource allocation, constituting a mixed-integer nonlinear programming problem, deep reinforcement learning is employed. This algorithm combines a target network and an estimation network for deployment and resource allocation. It takes actions at each step according to a greedy policy action function, using average response latency as the criterion for evaluating the quality of each action and simultaneously obtaining a reward value. After each action, the previous state, action, subsequent state, and reward value are stored in an experience pool. The experience pool stores the training set used to train the neural network. A loss function is used to continuously narrow the gap between the target network and the estimation network. After a certain number of iterations, the weights in the network are updated using gradient descent, enabling the network to eventually converge to the optimal result.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A service collaborative caching method in vehicle-mounted edge computing, characterized in that: Includes the following steps: (i) Based on historical access records, predict the number of service requests using the ARIMA model; (ii) Establishing a service latency model, which includes: 1) Latency model from vehicle terminal to edge server: in: Let c(k,i,t) be the total number of service requests k received by all nodes in time slot t, and let c(k,i,t) be the number of service requests k received by edge node i in time slot t. edge Let S represent the transmission rate between the vehicle terminal and the edge server, S be the set of edge nodes, and d(k) represent the storage space required to place service k, which satisfies... D(i) represents the storage capacity of edge server i, and K is the set of service requests; 2) The latency model for transmission between edge nodes is as follows: Where: the transmission rate between edge nodes is v bet C(k,t)-c(k,i,t) represents the number of services that need to be transmitted to edge server i for processing; b(k,i,t)∈{0,1} is a binary variable. If the service is placed on edge node i, then b(k,i,t) = 1, otherwise b(i,k,t) = 0. Therefore, b(k,i,t)∈{0,1},k∈K,i∈S. S is the set of edge nodes, and K is the set of service requests. 3) The computational latency model generated by the edge server processing requests is as follows: Where m(k) represents the computing resources required for service request k, l(k,i,t)M(i) represents the size of computing resources allocated by server i to service k, M(i) represents the computing power of edge server i, and l(k,i,t) should satisfy l(k,i,t)∈[0,1],k∈K,i∈S. When the service is placed on an edge server, 4) The total latency model for service requests is as follows: Among them, service k and service k * These are service pairs with a relationship, where K represents the set of service requests in this scenario, i.e., K = {1, 2, 3, ..., k}; the optimization objective, with the goal of reducing the average response latency of the task, is expressed as: (III) Establishing a service caching and resource allocation model: 1) Establish the state-space model S for deep reinforcement learning * (i,t): S * (i,t)={c(k,i,t),M(i),D i ,l(k,i,t)}; 2) Establish an action space model A for deep reinforcement learning. i (t): A i (t)={b(k,i,t),Δl(k,i,t),k∈K}; Where Δl(k,i,t) represents the smallest unit of computing resource allocation, and K represents the set of service requests in this scenario, i.e., K = {1,2,3...k}; 3) Set the reward function R(t), let make but Among them, state Take action A t Get the state Δw represents the response time difference between the two states, α is a constant, and K represents the set of service requests in the scenario, i.e., K = {1, 2, 3, ..., k}. 4) Maximize reward value: First define Q(s,a) as the action value, then use an a * Let Q(s,a) be used to represent maximizing the action value: a * =argmax a∈A Q(s,a); Here, to calculate the difference between the forward computation result and the true value in each iteration of the neural network, and thus guide the next step of training in the correct direction, a loss function is used: The weight values ​​in the forward calculation formula are corrected; the parameter θ is updated using gradient descent. Where η is the learning rate.

2. The service collaborative caching method in vehicle edge computing according to claim 1, characterized in that: In step (one), the specific prediction process for the service request volume is as follows: 1) Collect the actual historical records of user service calls and use them as the raw dataset; 2) Examine the service call time series data to determine if it is a white noise sequence. If so, use the ARIMA model for modeling and prediction; otherwise, use the arithmetic mean method to predict the service request volume using the following formula: 3) ARIMA modeling, denoted by ARMA(p,q), satisfies Where, φ o φ is a constant term. i and θ j These are the parameters for the AR model and the MA model, respectively; a t It is random error, i.e., white noise; the non-negative integers p, q and d represent the order of AR, MA and the order of difference, respectively; 4) Use time series forecasting methods to predict future values, and perform a stationarity test on the original time series. If it is not stationary, then modify the original time series {x}. t The time series is transformed into a stationary sequence by performing d interpolation operations on the sequence |t=1,2,...,n}; then an ARMA model is constructed as a candidate model for the time series. 5) Calculate the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the constructed ARMA model, and select appropriate values ​​for model parameters p and q based on ACF and PAC; where p is the AR model parameter and q is the MA model parameter; and perform residual tests on the model to verify its effectiveness, and select significantly effective models as candidate models. 6) Calculate the Information Criterion (AIC) value among the candidate models, and select the model with the smallest AIC value as the prediction model to predict the number of service requests.

3. The service collaborative caching method in vehicle edge computing according to claim 1, characterized in that: In step (ii), assume there are several edge nodes capable of deploying services and mobile users requesting services in a certain area. The set of edge nodes is represented as S = {1, 2, 3, ..., s}, D(i) represents the storage space size of edge node i, and M(i) represents the computing resources of edge node i. Edge nodes consume computing resources to process service requests. Assume there is a set of service requests K = {1, 2, 3, ..., k}, d(k) represents the storage space required to deploy service k, and m(k) represents the computing resources required for service request k. First, define a binary variable b(k,i,t)∈{0,1} to determine whether service k is placed on edge node i in time slot t. If the service is placed on edge node i, set b(k,i,t)=1; otherwise, set b(i,k,t)=0. Therefore, the service placement decision is defined as B(t)={b(k,i,t)|i∈S,k∈K}.

4. The service collaborative caching method in vehicle edge computing according to claim 1, characterized in that: In step (ii) 3), such as service k and service k * These are service pairs with a relationship, where the number of preceding services k processed by edge node i in time slot t is c. comp Given (k,i,t), the number of services k completed by all nodes in time slot t is: The transmission delay caused by data interaction is: Where d(kk*) represents the associated service pair k and k * The amount of data that needs to be interacted with; v bet Calculate the back-end service k for the data transfer rate between edge nodes. * The resulting delay is: Where m(k) * ) indicates the processing of service request k * Required computing resources; C comp (k*,t) represents the backend service k * The number of requests, l(k) * M(i) represents the amount of computing resources allocated from server i to service k; For handling association relationships, k and k * The total delay is: When processing a service k that has no relation to another service, the total latency is:

Images