Multi-dimensional dynamic modeling edge computing power network multi-domain resource optimization method

Abstract

The invention discloses an edge computing power network multi-domain resource optimization method for multi-dimensional dynamic modeling, and the method comprises the steps: constructing a computing power task demand dynamic model and a router health state dynamic model according to the historical data information of a computing power task demand and a router health state in an edge computing power network; constructing a computing power task demand composite feature function and a router health state composite feature function; predicting the computing power task demand and the health state of the router; and constructing a multi-domain resource optimization model, and solving to realize optimal communication resource allocation and edge computing power node matching of the computing power task. According to the method, the dynamic uncertain condition of the computing power task requirement in the edge computing power network and the health state of the router is fully considered, and under the condition that network resources are limited, the optimal communication resource allocation and edge computing power node matching decision are carried out on the computing power task by combining available idle resources in the network and user requirements.

Description

Technical Field

[0001] The present invention relates to the field of computing power tasks, and in particular to a method for optimizing multi-domain resources in an edge computing power network using multi-dimensional dynamic modeling. Background Art

[0002] With the continuous development of communication technology, user demand for various services continues to increase, and the demand for computing resources has shown exponential growth. The edge computing network achieves the widespread availability of computing resources by organically integrating distributed edge computing nodes and routing architecture. The edge computing network obtains computing task requirements and available computing power, storage, and other resource information of edge computing nodes through the resource orchestration hub. Based on network conditions and resource requirements, the edge computing network provides corresponding computing power and communication resources, and simultaneously allocates and schedules computing tasks, thereby achieving efficient orchestration and management of the entire edge computing network resources.

[0003] The dynamic uncertainty of computing task demands and router health within edge computing networks leads to problems such as unmet task demands, inefficient processing, and insufficient resource utilization. Therefore, given the limited resources of edge computing nodes, how to efficiently manage and utilize the computing resources of edge computing networks has become a pressing challenge. Therefore, it is necessary to conduct a comprehensive analysis based on user-submitted computing task information and the available idle resources in the network to achieve optimal resource allocation. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the deficiencies in the prior art, the present invention provides a multi-domain resource optimization method for edge computing network based on multi-dimensional dynamic modeling. By dynamically modeling the computing task requirements and the router health status and constructing a composite characteristic function, the computing task requirements and the router health status are predicted. This is used to solve the problems of the dynamic uncertainty of the computing task requirements and the router health status, which lead to the inability to meet the task requirements, low processing efficiency, and insufficient resource utilization.

[0005] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0006] A multi-domain resource optimization method for edge computing network based on multi-dimensional dynamic modeling, comprising the following steps:

[0007] Step 1: Based on the historical data information of computing task demand and router health status in the edge computing network, a dynamic model of computing task demand and a dynamic model of router health status are constructed.

[0008] Step 2: Construct a composite characteristic function of computing power task requirements.

[0009] Step 3: Construct a composite feature function of the router health status.

[0010] Step 4: Predict the computing power task requirements and router health status.

[0011] Step 5: Establish a comprehensive optimization model for multi-domain resources.

[0012] Step 6: Improve the proximal strategy optimization method based on the hybrid clipping strategy update mechanism.

[0013] In step 7, the improved proximal strategy optimization method based on the hybrid tailoring strategy update mechanism is used to solve the multi-domain resource comprehensive optimization model and output the optimal communication resource allocation and edge computing node matching strategy.

[0014] Preferably, the multi-domain resource comprehensive optimization model established in step 5 is:

[0015]

[0016] Among them, net (t) is the overall efficiency of the network, ρ t and ρ c is the weight factor. represents the optimal solution vector, They represent the optimal strategies for transmission power, bandwidth resource allocation, and edge computing node matching. Constraints C1 and C2 indicate that each task can only be fully matched to a certain edge computing node for execution. C3 and C4 indicate that the resource orchestration hub is the computing task. The total allocated transmit power cannot exceed the maximum available transmit power of base station m. C5 and C6 are expressed as The total allocated bandwidth resources cannot exceed the maximum available bandwidth of base station m. C7 indicates matching to edge computing node e n The total computing power required by all computing tasks at a location cannot exceed e n The amount of idle computing resources. C8 indicates that all computing tasks passing through a link cannot exceed the link capacity of the link. C9 indicates that the transmission rate of the computing task in the upload link should be greater than or equal to the rate requirement of the computing task. 10 It means that at time t, if router r n If it is not available, the edge computing nodes are constrained to match the decision variables

[0017] Preferably, the establishment of a multi-domain resource comprehensive optimization model in step 5 includes the following steps:

[0018] By predicting the characteristic values of computing power task requirements, the multi-dimensional requirements of computing power tasks at time t can be further expressed as follows:

[0019]

[0020] Among them, the computing power task demand generated by the kth user accessing base station m at time t can be expressed as a tuple l∈[1,L]. The total number of computing resource types required for the task is L, L=5. Indicates a task The size of computing power resources of type l∈[1,L] required at time t. Indicates the amount of data, and They represent the ideal processing delay and the maximum tolerable delay respectively. Indicates the data transmission rate requirement. Indicates the execution mode. Indicates that the execution mode is parallel processing. Indicates serial processing.

[0021] The original link state matrix is It can be represented by the following Z×Z dimensional matrix:

[0022]

[0023] in, Represents the link connection state matrix between routers. There are Z routers in the network. If router r m ,m∈[1,Z] and router r n , there is a connection relationship between n∈[1,Z], It can be further expressed as the data transmission rate of the link between two routers. If there is no connection between the two routers, then Prediction results based on the health status of router i∈[1,Z] at time t The corrected transmission rate between router links can be expressed as:

[0024]

[0025] in, Indicates the revised routing link transmission rate, Indicates that router i is in an available state at time t. This means the router is unavailable.

[0026] The upload link transmission delay is calculated as follows:

[0027]

[0028] in, Indicates computing power tasks The upload link transmission rate from user k to access base station m is, and Respectively represent the resource orchestration center allocated to computing tasks transmit power and bandwidth resources, represents the background Gaussian white noise power. The channel power gain between user k and access base station m is defined as follows:

[0029]

[0030] Where g0 represents the signal power gain at a reference distance of 1 meter and a transmission power of 1W, using (x k (t),y k (t),0) and (x m (t),y m (t),H) represent the positions of user k and base station m respectively, and the height of the base station is H.

[0031] The transmission delay of the fronthaul link can be calculated by the following formula:

[0032]

[0033] in, Indicates a task The amount of data, is the forward link transmission rate.

[0034] The routing link transmission delay is calculated as follows:

[0035]

[0036] Among them, the edge computing power node matching strategy is used To express, Indicates that the computing task Match to edge computing node e n The shortest path Dijkstra method is used to find the optimal path between the edge entry route and the edge computing node, where l m,n represents the set of links that the task passes through during the optimal path transmission process, l p,q Indicates a specific link in the set. Indicates the predicted value of the router health status Corrected link l p,q There are N edge computing nodes and M access base stations in the network, and there are K users within the coverage area of each access base station.

[0037] The formula for calculating delay is as follows:

[0038]

[0039] The total delay is calculated as follows:

[0040]

[0041] Among them, α up , α fr , α r and α com is the weighted coefficient of the delay in different stages.

[0042] The resource utilization estimation model is expressed as follows:

[0043]

[0044] in, Represents the edge computing node e n The size of the idle computing power resources of the l∈[1,L]th type.

[0045] Then a comprehensive optimization model of multi-domain resources is obtained.

[0046] Preferably, in step 6, based on the hybrid clipping strategy update mechanism, the improved proximal strategy optimization method includes the following steps:

[0047] The decision process of communication resource allocation and edge computing node matching is modeled as a Markov decision process, clarifying the state space, action space, and reward function.

[0048] Based on the state space, action space and reward function, an improved policy update mechanism based on a hybrid clipping objective function that integrates the standard clipping objective function, policy gradient and KL divergence penalty term is proposed. The specific implementation is as follows:

[0049]

[0050] Among them, L(θ) is the hybrid clipping objective function, is the weight factor. The standard clipping objective function is as follows:

[0051]

[0052] Among them, r t (θ) is the probability ratio between the current policy and the old policy is the advantage function Is the clipping factor. Use the clipping function clip(r t (θ),1-∈,1+·) to limit the change of strategy ratio. V(s t ) is the value network’s estimate of the current state value. The policy gradient objective function is calculated as follows:

[0053]

[0054] Among them, π θ (a t ∣s t ) is the probability of the policy network for the action. The KL divergence penalty is calculated as follows:

[0055]

[0056] Among them, π old (a t ∣s t ) is the old strategy, π θ (a t ∣s t ) is the current policy.

[0057] Preferably, the state space S is expressed as S = [θ, Q, ξ]:

[0058]

[0059] Among them, θ represents the state of each access base station in the network, θ m Indicates access to base station BS m Status, including the maximum available bandwidth for accessing the base station and status information of the maximum available transmit power Q represents the computing task status within the range of each access base station. ξ represents the status of each type of idle computing resource of each edge computing node in the network. Indicates access to base station BS m The kth task within .

[0060] The action space is represented as A = [P, B, Ω], Matching strategy for edge computing nodes, and They are transmit power and bandwidth allocation strategies respectively.

[0061] The reward function is as follows:

[0062]

[0063] in, and is the reward function weight factor, η pn is the weight of the penalty term.

[0064] Optimally, the composite characteristic function of computing power task requirements in step 2 is:

[0065]

[0066] Among them, f i q (·) is the composite characteristic function of computing power task requirements, is the mean of the random variable of computing power task demand, is the variance. is the weight coefficient.

[0067] Preferably, the composite characteristic function of the router health status in step 3 is:

[0068]

[0069] Among them, f i r (·) is the composite characteristic function of the router health status, is the average health status of the router, is the variance, and is the weight coefficient.

[0070] Preferably, the method for predicting computing power task requirements and router health status in step 4 includes the following steps:

[0071] The historical data sequence of the characteristic value of the computing power task demand within the historical sliding window time is obtained based on the statistics of the composite characteristic function of the computing power task demand Among them, the window length is λ, is the characteristic value of the computing power task demand at time t-1, which is specifically expressed as follows:

[0072]

[0073] in, and Computing tasks The mean and variance of the i-th demand at time j are used to calculate the fluctuation weight of the computing power task demand characteristic value based on the standard deviation within the sliding window. The specific calculation formula is as follows:

[0074]

[0075] in, is the fluctuation weight of the characteristic value of computing power task demand, is the standard deviation within the sliding window.

[0076] Weight of trend change of characteristic value of computing power task demand The calculation formula is as follows:

[0077]

[0078] The final time point weight of the computing power task demand characteristic value is expressed as follows:

[0079]

[0080] in, is the final time point weight of the computing power task demand characteristic value, and is an adjustable weight coefficient.

[0081] Based on the above content, a weighted autoregressive integrated moving average model is constructed. The specific formula is as follows:

[0082]

[0083] in, It represents the predicted value of the computing power task demand characteristic value at time t-1, and denote the autoregressive coefficient and the moving average coefficient, respectively. and are the orders of the autoregressive term and the moving average term, respectively.

[0084] The model parameters are optimized by minimizing the weighted residual sum of squares:

[0085]

[0086] in, Represents the optimized autoregressive coefficient and moving average coefficient.

[0087] Computing tasks The forecast value of the i-th demand at time t is calculated as follows:

[0088]

[0089] The historical data sequence of the router health status characteristic value within the historical sliding window time is obtained based on the composite characteristic function of the router health status [Θ i (t-λ),Θ i (t-λ+1),…,Θ i (t-1)], where Θ i (t-1) is the characteristic value of the router health status at time t-1, which is specifically expressed as follows:

[0090]

[0091] in, and The mean and variance of the health status of the i-th router at time j are respectively used to calculate the fluctuation weight of the router health status eigenvalue, which is specifically expressed as follows:

[0092]

[0093] in, is the router health status characteristic value fluctuation weight, rsd i(t-1) is the standard deviation within the sliding window.

[0094] The weight of the router health state characteristic value trend change is calculated based on the change amplitude of the router health state characteristic value at adjacent time points. i (t) is calculated as follows:

[0095]

[0096] The weight of the router's health status at the final time point is as follows:

[0097]

[0098] Among them, wre i (t) is the weight of the router health status characteristic value at the final time point, and is an adjustable weight coefficient.

[0099] Based on the above content, a weighted autoregressive integrated moving average model is constructed. The specific formula is as follows:

[0100]

[0101] in, represents the predicted value of the router health status characteristic value at time t-1, and Represent the autoregressive coefficient and the moving average coefficient, p reg and p sli are the orders of the autoregressive term and the moving average term, respectively.

[0102] The model parameters are optimized by minimizing the weighted residual sum of squares:

[0103]

[0104] in, Represents the predicted value of the router health status characteristic value at time t.

[0105] Preferably, step 1 comprises the following steps:

[0106] The computing power resource requirements, data volume, ideal processing delay, maximum tolerable delay, data transmission rate and execution mode of the computing power tasks submitted by users have significant random dynamics. Through the analysis of historical data, it can be seen that the computing power resource requirements, data volume, ideal processing delay, maximum tolerable delay and data transmission rate requirements are all random variables that obey Gaussian distribution. The demand is a multidimensional vector To express:

[0107]

[0108] Among them, the multidimensional vector represents the multidimensional demand vector of the kth task in the access base station m, Respectively Various computing resource requirements, Respectively The amount of data, ideal delay requirements, maximum tolerable delay, data transmission rate requirements and task execution method requirements, therefore, the random variable The distribution of is expressed as follows:

[0109]

[0110] in, Indicates computing power tasks The expected demand value of the i-th demand, is the variance of the demand, and then the probability density function of the Gaussian distribution model of the i-th demand of the computing task It is expressed as follows:

[0111]

[0112] Task execution method Different from other computing tasks that require random variables, the execution mode of computing tasks is Assume that it is a binary random variable, Representative computing power tasks Using parallel processing, The representative computing task adopts serial processing mode, and the task execution mode is the result of conditional probability, which is expressed as follows:

[0113]

[0114] Among them, f p (·) is a mapping function, Represents a multi-dimensional demand vector consisting of various types of computing resource requirements, data volume, ideal delay requirements, maximum tolerable delay threshold, and data transmission rate requirements of the task. Logistic regression is used to model this conditional probability. are the parameters of the regression model.

[0115] Set the router health status to y i ,i∈[1,Z] is set as a binary random variable of 0 or 1. Through historical data analysis, it can be obtained that the health status of the router follows the Bernoulli distribution. There are Z routers in total, y i =1 means router i is working normally, y i= 0 means the router is unavailable. The router health state transition process is modeled using a non-homogeneous continuous-time Markov chain, and the router health state space is defined as S device ={0,1}, state transition intensity matrix Λ i (t) are as follows:

[0116]

[0117] in, Indicates the instantaneous transition rate of the router's health status from normal to faulty. represents the instantaneous transition rate from failure to repair. The failure rate function follows the Weibull distribution, and the repair rate function follows the lognormal distribution:

[0118]

[0119] in, represents the shape parameter, is the scale parameter, and are the maintenance time data of router i in time period m Logarithmic mean and standard deviation of .

[0120]

[0121] The transient probability can be solved by the Kolmogorov forward equation in the above formula.

[0122] Another object of the present invention is to provide a multi-domain resource optimization system for edge computing network using multi-dimensional dynamic modeling, which is used to implement the multi-domain resource optimization method for edge computing network using multi-dimensional dynamic modeling, including an information input unit, a dynamic model construction unit, a computing power task demand composite feature function unit, a router health status composite feature function unit, a prediction unit, a multi-domain resource comprehensive optimization model unit, and a solution unit, wherein:

[0123] The information input unit is used to input historical data information on computing task requirements and router health status in the edge computing network.

[0124] The dynamic model construction unit is used to construct a computing power task demand dynamic model and a router health status dynamic model based on historical data information of computing power task demand and router health status in the edge computing power network.

[0125] The computing power task requirement composite characteristic function unit is used to construct a computing power task requirement composite characteristic function.

[0126] The router health state composite characteristic function unit is used to construct a router health state composite characteristic function.

[0127] The prediction unit is used to predict the computing task requirements and the health status of the router.

[0128] The multi-domain resource comprehensive optimization model unit is used to establish a multi-domain resource comprehensive optimization model.

[0129] The solving unit uses an improved proximal strategy optimization method based on a hybrid pruning strategy update mechanism to solve the multi-domain resource comprehensive optimization model, and outputs the optimal communication resource allocation and edge computing power node matching strategy.

[0130] Compared with the prior art, the present invention has the following beneficial effects:

[0131] (1) The present invention fully considers the dynamic uncertainty of the computing task requirements and router health status submitted by users in the edge computing network, and constructs a dynamic model of computing task requirements and a dynamic model of router health status.

[0132] (2) The present invention constructs a composite characteristic function of computing power task requirements based on historical data analysis of multi-dimensional computing power task requirements.

[0133] (3) The present invention constructs a composite characteristic function of the router health status based on the historical data analysis of the router health status.

[0134] (4) The present invention performs data analysis based on the historical characteristic values of computing power task requirements and the historical characteristic values of router health status, and predicts the computing power task requirements and router health status.

[0135] (5) The present invention combines the computing task delay model and the resource utilization model, and establishes a multi-domain resource comprehensive optimization model with computing task resource requirements and communication requirements as constraints.

[0136] (6) The present invention uses a hybrid clipping strategy update mechanism to improve the proximal strategy optimization method.

[0137] (7) The present invention uses an improved proximal strategy optimization method based on a hybrid clipping strategy update mechanism to solve the multi-domain resource comprehensive optimization model and output the optimal communication resource allocation and edge computing power node matching strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0138] Figure 1 is a flow chart of the present invention;

[0139] Figure 2 This is the edge computing network system model diagram;

[0140] Figure 3 This is a flow chart of the proximal strategy optimization method based on the improved hybrid clipping strategy update mechanism;

[0141] Figure 4This is a comparison chart of the network comprehensive efficiency of the present invention and other methods.

[0142] Figure 5 This is a comparison chart of the task response rates of the present invention and other methods. DETAILED DESCRIPTION

[0143] The present invention is further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0144] This embodiment proposes a multi-domain resource optimization method for edge computing network based on multi-dimensional dynamic modeling. According to the historical data information of computing task demand and router health status in the edge computing network, a dynamic model of computing task demand and a dynamic model of router health status are constructed; a composite characteristic function of computing task demand and a composite characteristic function of router health status are constructed; computing task demand and router health status are predicted; a multi-domain resource optimization model is constructed and solved to achieve the best communication resource allocation of computing task and matching of edge computing node, such as Figure 1-3 The specific steps are as follows:

[0145] Step S101: Based on the historical data of computing task requirements and router health status in the edge computing network, a dynamic model of computing task requirements and a dynamic model of router health status are constructed. The specific representation method is as follows:

[0146] The computing power resource requirements, data volume, ideal processing delay, maximum tolerable delay, data transmission rate and execution mode of the computing power tasks submitted by users have significant random dynamics. Through the analysis of historical data, it can be seen that the computing power resource requirements, data volume, ideal processing delay, maximum tolerable delay and data transmission rate requirements are all random variables that obey Gaussian distribution. The demand can be expressed as a multidimensional vector To express:

[0147]

[0148] Among them, the multidimensional vector represents the multidimensional demand vector of the kth task in the access base station m, Respectively Various types of computing resources such as CPU, GPU, NPU, FPGA and ASIC, Respectively The amount of data, ideal delay requirements, maximum tolerable delay, data transmission rate requirements and task execution method requirements, therefore, the random variable The distribution of can be expressed as:

[0149]

[0150] in, Indicates computing power tasks The expected demand value of the i-th demand, is the variance of the demand, and the probability density function of the Gaussian distribution model of the i-th demand of computing power task can be expressed as follows:

[0151]

[0152] Task execution method Different from other computing tasks that require random variables, the execution mode of computing tasks is Assume that it is a binary random variable, Representative computing power tasks Using parallel processing, The representative computing task adopts serial processing mode, and the task execution mode is the result of conditional probability, which is expressed as follows:

[0153]

[0154] Among them, f p (·) is a mapping function, Represents a multidimensional demand vector consisting of various types of computing resource requirements, data volume, ideal delay requirements, maximum tolerable delay threshold, and data transmission rate requirements of the task; logistic regression is used to model this conditional probability. are the parameters of the regression model;

[0155] Set the router health status to y i ,i∈[1,Z] is set as a binary random variable of 0 or 1. Through historical data analysis, it can be obtained that the health status of the router obeys the Bernoulli distribution. There are Z routers in the network, y i =1 means router i can work normally, y i = 0 means that the router is faulty and unavailable; the router health state transition process is modeled using a non-homogeneous continuous-time Markov chain, and the router health state space is defined as S device ={0,1}, the state transition intensity matrix is as follows:

[0156]

[0157] in, Indicates the instantaneous transition rate of the router's health status from normal to faulty. represents the instantaneous transition rate from failure to repair; the failure rate function follows the Weibull distribution, and the repair rate function follows the lognormal distribution:

[0158]

[0159] in, represents the shape parameter, is the scale parameter, and are the maintenance time data of router i in time period m Logarithmic mean and standard deviation of ;

[0160]

[0161] The transient probability can be solved by the Kolmogorov forward equation in the above formula.

[0162] Step S102: Analyze historical data of computing power task requirements and construct a composite characteristic function of computing power task requirements. The specific implementation method is as follows:

[0163]

[0164] Among them, f i q (·) is the composite characteristic function of computing power task requirements, is the mean value of computing power task requirements, is the variance; Represents the weight coefficient. Based on the above formula, the characteristic value of computing power task demand can be obtained.

[0165] Step S103: Based on the historical data analysis of the router health status, a composite characteristic function of the router health status is constructed. The specific implementation method is as follows:

[0166]

[0167] Among them, f i r (·) is the composite characteristic function of the router health status, is the average health status of the router, is the variance, and is the weight coefficient. Based on the above formula, the router health status characteristic value can be obtained.

[0168] Step S104: Perform data analysis based on historical characteristic values of computing task demand and historical characteristic values of router health status to predict computing task demand and router health status. The specific steps are as follows:

[0169] The historical data sequence of the characteristic value of the computing power task demand within the historical sliding window time is obtained based on the statistics of the composite characteristic function of the computing power task demand Among them, the window length is λ, is the characteristic value of the computing power task demand at time t-1, which is specifically expressed as follows:

[0170]

[0171] in, and Computing tasks The mean and variance of the i-th demand at time j are used to calculate the fluctuation weight of the computing power task demand characteristic value based on the standard deviation within the sliding window. The specific calculation formula is as follows:

[0172]

[0173] in, is the fluctuation weight of the characteristic value of computing power task demand, is the standard deviation within the sliding window;

[0174] Weight of trend change of characteristic value of computing power task demand The calculation formula is as follows:

[0175]

[0176] The final time point weight of the computing power task demand characteristic value is expressed as follows:

[0177]

[0178] in, is the final time point weight of the computing power task demand characteristic value, and is the adjustable weight coefficient;

[0179] Based on the above content, a weighted autoregressive integrated moving average model is constructed. The specific formula is as follows:

[0180]

[0181] in, It represents the predicted value of the computing power task demand characteristic value at time t-1, and denote the autoregressive coefficient and the moving average coefficient, respectively. and are the order of autoregressive term and moving average term respectively;

[0182] The model parameters are optimized by minimizing the weighted residual sum of squares:

[0183]

[0184] Based on the above content, computing power tasks The forecast value of the i-th demand at time t is calculated as follows:

[0185]

[0186] The historical data sequence of the router health status characteristic value within the historical sliding window time is obtained based on the composite characteristic function of the router health status [Θ i (t-λ),Θ i (t-λ+1),…,Θ i (t-1)], where Θ i (t-1) is the characteristic value of the router health status at time t-1, which is specifically expressed as follows:

[0187]

[0188] in, and The mean and variance of the health status of the i-th router at time j are respectively used to calculate the fluctuation weight of the router health status eigenvalue, which is specifically expressed as follows:

[0189]

[0190] in, is the router health status characteristic value fluctuation weight, rsd i (t-1) is the standard deviation;

[0191] The weight of the router health state characteristic value trend change is calculated based on the change amplitude of the router health state characteristic value at adjacent time points. i (t) is calculated as follows:

[0192]

[0193] The weight of the router's health status at the final time point is as follows:

[0194]

[0195] Among them, wre i (t) is the weight of the router health status characteristic value at the final time point, and is the adjustable weight coefficient;

[0196] Based on the above content, a weighted autoregressive integrated moving average model is constructed. The specific formula is as follows:

[0197]

[0198] in, represents the predicted value of the router health status characteristic value at time t-1, and Represent the autoregressive coefficient and the moving average coefficient, p reg and p sli are the order of autoregressive term and moving average term respectively;

[0199] The model parameters are optimized by minimizing the weighted residual sum of squares:

[0200]

[0201] Therefore, the predicted value of the health status of the i-th router at time t can be calculated using the above formula.

[0202] Step S105: Establish a comprehensive optimization model for multi-domain resources. The specific method is as follows:

[0203] By predicting the characteristic value of computing power task demand, the computing power task demand at time t can be further expressed as follows:

[0204]

[0205] Among them, the computing power task demand generated by the kth user accessing base station m at time t can be expressed as a tuple l∈[1,L] represents the total number of computing resource types required for the task, L, L=5; where, Indicates a task The required computing power resource size of type l∈[1,L] at time t; Indicates the amount of data, and They represent the ideal processing delay and the maximum tolerable delay respectively; Indicates data transmission rate requirements; Indicates the execution mode. Indicates that the execution mode is parallel processing; Indicates serial processing;

[0206] The original link state matrix is It can be represented by the following Z×Z dimensional matrix:

[0207]

[0208] in, Represents the link connection state matrix between Z routers, there are Z routers in total; if router r m ,m∈[1,Z] and router r n, there is a connection relationship between n∈[1,Z], It can be expressed as the data transmission rate of the link between two routers. If there is no connection between the two routers, then Prediction results based on the health status of router i∈[1,Z] at time t The corrected transmission rate between router links can be expressed as:

[0209]

[0210] in, Indicates the modified router link transmission rate. If Indicates that router i is in an available state at time t; Indicates that the router is in an unavailable state;

[0211] The upload link transmission delay is calculated as follows:

[0212]

[0213] in, Indicates computing power tasks The upload link transmission rate from user k to access base station m is, and Respectively represent the resource orchestration center allocated to computing tasks transmit power and bandwidth resources, represents the background Gaussian white noise power. The channel power gain between user k and access base station m is defined as follows:

[0214]

[0215] Where g0 represents the signal power gain at a reference distance of 1 meter and a transmission power of 1W; using (x k (t),y k (t),0) and (x m (t),y m (t),H) represent the positions of user k and base station m respectively. All base stations are fixed and the height of the base station is H;

[0216] The transmission delay of the fronthaul link can be calculated by the following formula:

[0217]

[0218] in, Indicates a task The amount of data, For the task The transmission rate of the fronthaul link between access base station m and edge ingress router m;

[0219] The routing link transmission delay is calculated as follows:

[0220]

[0221] Among them, the edge computing node matching strategy is used To express, Indicates that the computing task Match to edge computing node e n The shortest path Dijkstra method is used to find the optimal path between the edge entry route and the edge computing node, where l m,n represents the set of links that the task passes through during the optimal path transmission process, l p,q Indicates a specific link in the set. Indicates the predicted value of the router health status Corrected link l p,q The link capacity of the network has N edge computing nodes and M access base stations. Each access base station has K users within its coverage area.

[0222] The formula for calculating delay is as follows:

[0223]

[0224] The total delay is calculated as follows:

[0225]

[0226] Among them, α up , α fr , α r and α com is the weighting coefficient of the delay at different stages;

[0227] The resource utilization estimation model is expressed as follows:

[0228]

[0229] in, Represents the edge computing node e n The size of idle computing resources of type l∈[1,L];

[0230] The multi-domain resource comprehensive optimization model can be expressed as follows:

[0231]

[0232] Among them, net (t) is the overall efficiency of the network, ρ t and ρ c is the weight factor; represents the optimal solution vector, They represent the optimal strategies for transmission power, bandwidth resource allocation, and edge computing node matching respectively; constraints C1 and C2 indicate that each task can only be fully matched to a certain edge computing node for execution; C3 and C4 indicate that the resource orchestration hub is the computing task The total allocated transmit power cannot exceed the maximum available transmit power of base station m; C5 and C6 are expressed as The total allocated bandwidth resources cannot exceed the maximum available bandwidth of base station m; C7 indicates matching to edge computing node e n The total computing power required by all computing tasks at a location cannot exceed e n The amount of idle computing resources; C8 indicates that all computing tasks passing through a certain link cannot exceed the link capacity of the link; C9 indicates that the transmission rate of the computing task in the upload link should be greater than or equal to the rate requirement of the computing task; C 10 It means that at time t, if router r n If it is not available, the edge computing nodes are constrained to match the decision variables

[0233] Step S106: improving the proximal strategy optimization method based on the hybrid clipping strategy update mechanism, specifically implemented as follows:

[0234] In the decision-making process of communication resource allocation and edge computing node matching, the current state, i.e., computing task requirements and computing resources, determines the next decision. The decision only depends on the current state and is not directly affected by the past state. This conforms to the Markov property. Therefore, the decision-making process of communication resource allocation and edge computing node matching is modeled as a Markov decision process, and the state space, action space, and reward function are defined as follows:

[0235] The state space S can be expressed as S = [θ, Q, ξ]:

[0236]

[0237]

[0238] Among them, θ represents the state of each access base station in the network, θ m Indicates access to base station BS m Status, including the maximum available bandwidth for accessing the base station and status information of the maximum available transmit power Q represents the computing task status within the range of each access base station; ξ represents the status of each type of idle computing resource of each edge computing node in the network. Indicates access to base station BS m The kth task in ;

[0239] The action can be expressed as A=[P,B,Ω], Matching strategy for edge computing nodes, and They are transmit power and bandwidth allocation strategies respectively;

[0240] The reward function is as follows:

[0241]

[0242] in, and is the reward function weight factor, η pn is the weight of the penalty term;

[0243] Based on the above definition, a policy update mechanism based on the hybrid clipping objective function that integrates the standard clipping objective function, policy gradient and KL divergence penalty term is proposed. The specific implementation is as follows:

[0244]

[0245] Among them, L(θ) is the hybrid clipping objective function, is the weight factor; the standard clipping objective function is as follows:

[0246]

[0247] Among them, r t (θ) is the probability ratio between the current policy and the old policy is the advantage function is the clipping factor; use the clipping function clip(r t (θ),1-∈,1+∈) to limit the change of strategy ratio; V(s t ) is the value network's estimate of the current state value; the policy gradient objective function is calculated as follows:

[0248]

[0249] Among them, π θ (a t ∣s t ) is the probability of the policy network for the action; the KL divergence penalty term is calculated as follows:

[0250]

[0251] Among them, π old (a t ∣s t ) is the old strategy, π θ (a t ∣s t ) is the current policy.

[0252] Step S107, as Figure 3 As shown in the figure, the improved proximal strategy optimization method based on the hybrid clipping strategy update mechanism is used to solve the multi-domain resource comprehensive optimization model. The specific steps are as follows:

[0253] S1: Initialize the parameters θ of the policy network and value network A ,θ V , iter=1;

[0254] S2: Interact with the environment and use the current strategy Collect data such as state, action, reward and next state;

[0255] S3: Calculate the advantage function A based on each time step t =R t +γV(s t+1 )-V(s t );

[0256] S4: Calculate the improved hybrid loss function

[0257] S5: Update policy network parameters

[0258] S6: Update the value network parameters using mean squared error

[0259] S7: Determine whether the maximum number of iterations has been reached. If iter <N iter , jump to step S2, if iter>N iter , output the optimal communication resource allocation and edge computing node matching strategy.

[0260] In another embodiment, a multi-dimensional dynamic modeling edge computing network multi-domain resource optimization system is provided, which is used to implement the multi-dimensional dynamic modeling edge computing network multi-domain resource optimization method, including an information input unit, a dynamic model construction unit, a computing power task demand composite feature function unit, a router health status composite feature function unit, a prediction unit, a multi-domain resource comprehensive optimization model unit, and a solution unit, wherein:

[0261] The information input unit is used to input historical data information on computing task requirements and router health status in the edge computing network.

[0262] The dynamic model construction unit is used to construct a computing power task demand dynamic model and a router health status dynamic model based on historical data information of computing power task demand and router health status in the edge computing power network.

[0263] The computing power task requirement composite characteristic function unit is used to construct a computing power task requirement composite characteristic function.

[0264] The router health state composite characteristic function unit is used to construct a router health state composite characteristic function.

[0265] The prediction unit is used to predict the computing task requirements and the health status of the router.

[0266] The multi-domain resource comprehensive optimization model unit is used to establish a multi-domain resource comprehensive optimization model.

[0267] The solving unit uses an improved proximal strategy optimization method based on a hybrid pruning strategy update mechanism to solve the multi-domain resource comprehensive optimization model, and outputs the optimal communication resource allocation and edge computing power node matching strategy.

[0268] The network comprehensive efficiency of the present invention is compared with that of other methods. Figure 4 As shown in the figure, the task response rate of the present invention is compared with that of other methods. Figure 5 shown.

[0269] This invention fully considers the dynamic uncertainty of computing task requirements and router health status in the edge computing network. Based on the computing task requirements proposed by users and the historical information of router health status, it dynamically models and predicts the computing task requirements and router health status. When network resources are limited, it combines the available idle resources in the network and user needs to make the optimal communication resource allocation and edge computing node matching decisions for computing tasks.

[0270] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-domain resource optimization method for edge computing network based on multi-dimensional dynamic modeling, characterized in that: The following steps are involved: Step 1: Based on the historical data of computing task demand and router health status in the edge computing network, a dynamic model of computing task demand and a dynamic model of router health status are constructed; Step 2: Construct a composite feature function of computing power task requirements; Step 3: Construct a composite characteristic function of the router health status; Step 4: Predict computing power requirements and router health status. Step 5: Establish a comprehensive optimization model for multi-domain resources; Step 6: Improve the proximal strategy optimization method based on the hybrid clipping strategy update mechanism; In step 7, the improved proximal strategy optimization method based on the hybrid tailoring strategy update mechanism is used to solve the multi-domain resource comprehensive optimization model and output the optimal communication resource allocation and edge computing node matching strategy.

2. The method for optimizing multi-domain resources of edge computing network based on multi-dimensional dynamic modeling according to claim 1 is characterized by: In step 5, the multi-domain resource comprehensive optimization model is established as follows: Among them, net (t) is the overall efficiency of the network, ρ t and ρ c is the weight factor; represents the optimal solution vector, They represent the optimal strategies for transmission power, bandwidth resource allocation, and edge computing node matching respectively; constraints C1 and C2 indicate that each task can only be fully matched to a certain edge computing node for execution; C3 and C4 indicate that the resource orchestration hub is the computing task The total allocated transmit power cannot exceed the maximum available transmit power of base station m; C5 and C6 are expressed as The total allocated bandwidth resources cannot exceed the maximum available bandwidth of base station m; C7 indicates matching to edge computing node e n The total computing power required by all computing tasks at a location cannot exceed e n The amount of idle computing resources; C8 indicates that all computing tasks passing through a certain link cannot exceed the link capacity of the link; C9 indicates that the transmission rate of the computing task in the upload link should be greater than or equal to the rate requirement of the computing task; C 10 It means that at time t, if router r n If it is not available, the edge computing nodes are constrained to match the decision variables 3. The method for optimizing multi-domain resources of edge computing network based on multi-dimensional dynamic modeling according to claim 1 is characterized by: Establishing a comprehensive optimization model for multi-domain resources in step 5 includes the following steps: By predicting the characteristic values of computing power task requirements, the multi-dimensional requirements of computing power tasks at time t can be further expressed as follows: Among them, the computing power task demand generated by the kth user accessing base station m at time t can be expressed as a tuple l∈[1,L] represents the total number of computing resource types required for the task, L, L=5; where, Indicates a task The required computing power resource size of type l∈[1,L] at time t; Indicates the amount of data, and They represent the ideal processing delay and the maximum tolerable delay respectively; Indicates data transmission rate requirements; Indicates the execution mode. Indicates that the execution mode is parallel processing; Indicates serial processing; The original link state matrix is It can be represented by the following Z×Z dimensional matrix: in, Represents the link connection state matrix between routers. There are Z routers in the network. If router r m ,m∈[1,Z] and router r n , there is a connection relationship between n∈[1,Z], It can be further expressed as the data transmission rate of the link between two routers. If there is no connection between the two routers, then Prediction results based on the health status of router i∈[1,Z] at time t The corrected transmission rate between router links can be expressed as: in, Indicates the revised routing link transmission rate, Indicates that router i is in an available state at time t; This means that the router is unavailable. The upload link transmission delay is calculated as follows: in, Indicates computing power tasks The upload link transmission rate from user k to access base station m is, and Respectively represent the resource orchestration center allocated to computing tasks transmit power and bandwidth resources, represents the background Gaussian white noise power. The channel power gain between user k and access base station m is defined as follows: Where g0 represents the signal power gain at a reference distance of 1 meter and a transmission power of 1W, using (x k (t),y k (t),0) and (x m (t),y m (t),H) represent the positions of user k and base station m respectively, and the height of the base station is H; The transmission delay of the fronthaul link can be calculated by the following formula: in, Indicates a task The amount of data, is the transmission rate of the fronthaul link; The routing link transmission delay is calculated as follows: Among them, the edge computing power node matching strategy is used m∈[1,M],n∈[1,N],k∈[1,K], Indicates that the computing task Match to edge computing node e n The shortest path Dijkstra method is used to find the optimal path between the edge entry route and the edge computing node, where l m,n represents the set of links that the task passes through during the optimal path transmission process, l p,q Indicates a specific link in the set. Indicates the predicted value of the router health status i∈[1,Z] The corrected link l p,q The link capacity of the network consists of N edge computing nodes and M access base stations. Each access base station has K users within its coverage area. The formula for calculating delay is as follows: The total delay is calculated as follows: Among them, α up , α fr , α r and α com is the weighting coefficient of the delay at different stages; The resource utilization estimation model is expressed as follows: in, Represents the edge computing node e n The size of idle computing resources of type l∈[1,L]; Then a comprehensive optimization model of multi-domain resources is obtained.

4. The method for optimizing multi-domain resources of edge computing network based on multi-dimensional dynamic modeling according to claim 1 is characterized by: In step 6, based on the hybrid clipping strategy update mechanism, the improved proximal strategy optimization method includes the following steps: The communication resource allocation and edge computing node matching decision process is modeled as a Markov decision process, clarifying the state space, action space, and reward function; Based on the state space, action space and reward function, an improved policy update mechanism based on a hybrid clipping objective function that integrates the standard clipping objective function, policy gradient and KL divergence penalty term is proposed. The specific implementation is as follows: Among them, L(θ) is the hybrid clipping objective function, is the weight factor; the standard clipping objective function is as follows: Among them, r t (θ) is the probability ratio between the current policy and the old policy is the advantage function is the clipping factor; use the clipping function clip(r t (θ),1-∈,1+∈) to limit the change of strategy ratio; V(s t ) is the value network's estimate of the current state value; the policy gradient objective function is calculated as follows: Among them, π θ (a t ∣s t ) is the probability of the policy network for the action; the KL divergence penalty term is calculated as follows: Among them, π old (a t ∣s t ) is the old strategy, π θ (a t ∣s t ) is the current policy.

5. The method for optimizing multi-domain resources of edge computing network based on multi-dimensional dynamic modeling according to claim 1 is characterized by: The state space S is represented as S = [θ, Q, ξ]: Among them, θ represents the state of each access base station in the network, θ m Indicates access to base station BS m Status, including the maximum available bandwidth for accessing the base station and status information of the maximum available transmit power Q represents the computing task status within the range of each access base station; ξ represents the status of each type of idle computing resource of each edge computing node in the network. Indicates access to base station BS m The kth task in ; The action space is represented as A = [P, B, Ω], Matching strategy for edge computing nodes, and They are transmit power and bandwidth allocation strategies respectively; The reward function is as follows: in, and is the reward function weight factor, η pn is the weight of the penalty term.

6. The method for optimizing multi-domain resources of edge computing network based on multi-dimensional dynamic modeling according to claim 1, characterized in that: The composite characteristic function of computing power task requirements in step 2 is: Among them, f i q (·) is the composite characteristic function of computing power task requirements, is the mean of the random variable of computing power task demand, is the variance; is the weight coefficient.

7. The method for optimizing multi-domain resources of edge computing network based on multi-dimensional dynamic modeling according to claim 1, characterized in that: The composite characteristic function of the router health status in step 3 is: Among them, f i r (·) is the composite characteristic function of the router health status, is the average health status of the router, is the variance, and is the weight coefficient.

8. The method for optimizing multi-domain resources of edge computing network based on multi-dimensional dynamic modeling according to claim 1, characterized in that: The method for predicting computing power task requirements and router health status in step 4 includes the following steps: The historical data sequence of the characteristic value of the computing power task demand within the historical sliding window time is obtained based on the statistics of the composite characteristic function of the computing power task demand Among them, the window length is λ, is the characteristic value of the computing power task demand at time t-1, which is specifically expressed as follows: in, and Computing tasks The mean and variance of the i-th demand at time j are used to calculate the fluctuation weight of the computing power task demand characteristic value based on the standard deviation within the sliding window. The specific calculation formula is as follows: in, is the fluctuation weight of the characteristic value of computing power task demand, is the standard deviation within the sliding window; Weight of trend change of characteristic value of computing power task demand The calculation formula is as follows: The final time point weight of the computing power task demand characteristic value is expressed as follows: in, is the final time point weight of the computing power task demand characteristic value, and is the adjustable weight coefficient; Based on the above content, a weighted autoregressive integrated moving average model is constructed. The specific formula is as follows: in, It represents the predicted value of the computing power task demand characteristic value at time t-1, and denote the autoregressive coefficient and the moving average coefficient, respectively. and are the order of autoregressive term and moving average term respectively; The model parameters are optimized by minimizing the weighted residual sum of squares: in, Represents the optimized autoregressive coefficient and moving average coefficient; Computing tasks The forecast value of the i-th demand at time t is calculated as follows: The historical data sequence of the router health status characteristic value within the historical sliding window time is obtained based on the composite characteristic function of the router health status [Θ i (t-λ),Θ i (t-λ+1),…,Θ i (t-1)], where Θ i (t-1) is the characteristic value of the router health status at time t-1, which is specifically expressed as follows: in, and The mean and variance of the health status of the i-th router at time j are respectively used to calculate the fluctuation weight of the router health status eigenvalue, which is specifically expressed as follows: in, is the router health status characteristic value fluctuation weight, rsd i (t-1) is the standard deviation within the sliding window; The weight of the router health state characteristic value trend change is calculated based on the change amplitude of the router health state characteristic value at adjacent time points. i (t) is calculated as follows: The weight of the router's health status at the final time point is as follows: Among them, wre i (t) is the weight of the router health status characteristic value at the final time point, and is the adjustable weight coefficient; Based on the above content, a weighted autoregressive integrated moving average model is constructed. The specific formula is as follows: in, represents the predicted value of the router health status characteristic value at time t-1, and Represent the autoregressive coefficient and the moving average coefficient, p reg and p sli are the order of autoregressive term and moving average term respectively; The model parameters are optimized by minimizing the weighted residual sum of squares: in, Represents the predicted value of the router health status characteristic value at time t.

9. The method for optimizing multi-domain resources of edge computing network based on multi-dimensional dynamic modeling according to claim 1, characterized in that: Step 1 includes the following steps: The computing power resource requirements, data volume, ideal processing delay, maximum tolerable delay, data transmission rate and execution mode of the computing power tasks submitted by users have significant random dynamics. Through the analysis of historical data, it can be seen that the computing power resource requirements, data volume, ideal processing delay, maximum tolerable delay and data transmission rate requirements are all random variables that obey Gaussian distribution. The demand is a multidimensional vector To express: Among them, the multidimensional vector represents the multidimensional demand vector of the kth task in the access base station m, Respectively Various computing resource requirements, Respectively The amount of data, ideal delay requirements, maximum tolerable delay, data transmission rate requirements and task execution method requirements, therefore, the random variable The distribution of is expressed as follows: in, Indicates computing power tasks The expected demand value of the i-th demand, is the variance of the demand, and then the probability density function of the Gaussian distribution model of the i-th demand of the computing task It is expressed as follows: Task execution method Different from other computing tasks that require random variables, the execution mode of computing tasks is Assume that it is a binary random variable, Representative computing power tasks Using parallel processing, The representative computing task adopts serial processing mode, and the task execution mode is the result of conditional probability, which is expressed as follows: Among them, f p (·) is a mapping function, Represents a multidimensional demand vector consisting of various types of computing resource requirements, data volume, ideal delay requirements, maximum tolerable delay threshold, and data transmission rate requirements of the task; logistic regression is used to model this conditional probability. are the parameters of the regression model; Set the router health status to y i ,i∈[1,Z] is set as a binary random variable of 0 or 1. Through historical data analysis, it can be obtained that the health status of the router follows the Bernoulli distribution. There are Z routers in total, y i =1 means router i is working normally, y i = 0 means the router is unavailable; a non-homogeneous continuous-time Markov chain is used to model the router health state transition process, and the router health state space is defined as S device ={0,1}, state transition intensity matrix Λ i (t) are as follows: in, Indicates the instantaneous transition rate of the router's health status from normal to faulty. represents the instantaneous transition rate from failure to repair; the failure rate function follows the Weibull distribution, and the repair rate function follows the lognormal distribution: in, represents the shape parameter, is the scale parameter, and are the maintenance time data of router i in time period m Logarithmic mean and standard deviation of ; The transient probability can be solved by the Kolmogorov forward equation in the above formula.

10. A system for implementing the edge computing network multi-domain resource optimization method of multi-dimensional dynamic modeling according to claim 1, characterized in that: It includes information input unit, dynamic model construction unit, computing power task demand composite feature function unit, router health status composite feature function unit, prediction unit, multi-domain resource comprehensive optimization model unit, and solution unit, among which: The information input unit is used to input historical data information of computing task requirements and router health status in the edge computing network; The dynamic model construction unit is used to construct a dynamic model of computing task demand and a dynamic model of router health status based on historical data information of computing task demand and router health status in the edge computing network; The computing power task demand composite characteristic function unit is used to construct a computing power task demand composite characteristic function; The router health state composite characteristic function unit is used to construct a router health state composite characteristic function; The prediction unit is used to predict the computing task demand and the router health status; The multi-domain resource comprehensive optimization model unit is used to establish a multi-domain resource comprehensive optimization model; The solving unit uses an improved proximal strategy optimization method based on a hybrid pruning strategy update mechanism to solve the multi-domain resource comprehensive optimization model, and outputs the optimal communication resource allocation and edge computing power node matching strategy.

Images