CN-122027474-A - Edge node collaborative optimization method and system of computing power network

CN122027474ACN 122027474 ACN122027474 ACN 122027474ACN-122027474-A

Abstract

The invention relates to the technical field of edge computing, in particular to a method and a system for collaborative optimization of edge nodes of a computing power network. The method comprises the steps of collecting resource state information and historical behavior data of a plurality of edge nodes to form a resource state data set, constructing a computational power network topological graph based on the resource state data set, extracting node characteristic representations of all the nodes through a graph neural network, calculating the comprehensive reputation of all the edge nodes based on the historical behavior data, integrating the comprehensive reputation into the node characteristic representations, collecting energy consumption state data of all the edge nodes, calculating energy consumption efficiency indexes of all the edge nodes, and establishing a coupling linkage mechanism of the reputation and energy consumption perception. The invention realizes high resource utilization rate, low task delay, strong system robustness and green energy conservation of the edge node scheduling of the computing power network through the coupling linkage mechanism of credibility and energy consumption perception.

Inventors

HAN YUE
YANG YANG
CAO MIN
WANG LIANQING
ZHANG NA
WANG JIE
LI HAI
SHE YUHANG

Assignees

中国人民解放军信息支援部队工程大学

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. The edge node collaborative optimization method of the computing power network is characterized by comprising the following steps of: Collecting resource state information and historical behavior data of a plurality of edge nodes to form a resource state data set; Constructing a computational power network topological graph based on the resource state data set, and extracting node characteristic representation of each node through a graph neural network; calculating the comprehensive reputation of each edge node based on the historical behavior data, and integrating the comprehensive reputation into the node characteristic representation; Collecting energy consumption state data of each edge node, and calculating energy consumption efficiency indexes of each edge node; Establishing a coupling linkage mechanism of the credibility and the energy consumption perception, dynamically adjusting the energy consumption penalty weight of the node according to the comprehensive credibility of the node, increasing the energy consumption penalty weight and reducing the task allocation priority when the comprehensive credibility of the node is lower than a preset credibility threshold, and triggering a credibility linkage scheduling strategy when the energy consumption of the system is abnormal; Establishing a collaborative optimization model based on a Markov decision process, and training the collaborative optimization model by adopting a multi-agent deep reinforcement learning algorithm, wherein the collaborative optimization model comprises a reward function; inputting the node characteristic representation into the collaborative optimization model to generate a task scheduling decision; And distributing the task to the corresponding edge node for execution according to the task scheduling decision, and updating parameters of the collaborative optimization model based on an execution result.
2. The method for collaborative optimization of edge nodes of a power network according to claim 1, wherein extracting node feature representations of nodes through a graph neural network comprises: Constructing the computational power network topological graph, wherein the computational power network topological graph comprises a node set and a network connection set, the node set comprises all edge nodes, and the network connection set represents network connection relations among the edge nodes; Carrying out multi-layer iterative updating on node characteristics by adopting a graph convolution network structure, aggregating characteristic information of neighbor nodes during each layer updating, and carrying out normalization weighting according to the degree of the nodes; And applying a nonlinear activation function to the transformed features to obtain updated node feature representations.
3. The method of claim 1, wherein training the collaborative optimization model using a multi-agent deep reinforcement learning algorithm comprises: Collecting state information of all edge nodes in a training stage by adopting a centralized training and decentralized execution mechanism to construct a global state vector; calculating a value function by using a depth Q network, wherein the value function is obtained by mixing local value functions of all edge nodes; modulating the local value function based on the comprehensive credibility, and obtaining higher value function estimation by the node with high credibility, thereby increasing the probability of being selected to execute the task; In the execution stage, each edge node independently executes actions based on the local state to realize distributed decision.
4. The method of claim 1, wherein calculating the integrated reputation of each edge node based on the historical behavior data comprises: collecting task execution success rate, fault occurrence frequency, response time delay score, response time delay fluctuation rate and energy consumption efficiency score of each edge node in a preset historical observation period; the task execution success rate is the ratio of the number of successfully completed tasks to the total number of distributed tasks; The occurrence frequency of the faults is the ratio of the number of faults in the observation period to the observation duration; The response time delay score is calculated based on a normalized value of the node average response time delay relative to the minimum and maximum response time delay of the system; The response delay fluctuation rate is the ratio of the standard deviation of the response delay to the average response delay; the energy consumption efficiency score is calculated based on normalized values of node energy consumption efficiency relative to minimum and maximum energy consumption efficiency of the system; The five indexes are weighted and summed to obtain comprehensive credibility, wherein the occurrence frequency of faults and the response time delay fluctuation rate adopt reverse scoring; and taking the integrated reputation as an additional dimension of the input feature vector of the graph neural network.
5. The method for collaborative optimization of edge nodes of a power network according to claim 1, wherein collecting energy consumption status data of each edge node and calculating an energy consumption efficiency index comprises: collecting the instantaneous power consumption, the unit calculation power consumption and the accumulated energy consumption of each edge node; calculating energy consumption efficiency indexes of all edge nodes, wherein the energy consumption efficiency indexes are the ratio of effective calculation power output to instantaneous power consumption; And normalizing the energy consumption efficiency index to be used as a calculation basis of the energy consumption efficiency score in the comprehensive credibility.
6. The method for collaborative optimization of edge nodes of a power network according to claim 1, wherein the coupled linkage mechanism of reputation and energy consumption perception comprises: Establishing a dynamic regulation relation between the credibility and the energy consumption weight, wherein a high-credibility node obtains a lower energy consumption penalty weight, and a low-credibility node obtains a higher energy consumption penalty weight; Setting a reputation threshold, and executing a low reputation node punishment strategy when the comprehensive reputation of the node is lower than the reputation threshold, wherein the low reputation node punishment strategy comprises the steps of increasing the energy consumption punishment weight of the node and proportionally reducing the task allocation probability of the node; in the self-adaptive adjustment of the weight of the reward function, taking the credit degree change trend and the energy consumption efficiency change trend as input factors of weight adjustment; setting an energy consumption abnormality detection mechanism, and judging that the energy consumption is abnormal when the total energy consumption of the system exceeds the sum of the preset multiples of the historical average value and the standard deviation; When the energy consumption is abnormally triggered, executing a credibility linkage scheduling strategy, wherein the strategy comprises the steps of proportionally reducing task allocation of low-credibility nodes, preferentially allocating the tasks to nodes meeting the double conditions of credibility threshold and energy consumption efficiency threshold, and temporarily improving energy consumption penalty coefficients.
7. The method of edge node co-optimization of a power network of claim 1, further comprising the step of load balancing: calculating load indexes of all edge nodes, wherein the load indexes are weighted combinations of the calculated resource occupancy rate, the memory occupancy rate and the task queue length; When the load index of one edge node exceeds a preset load threshold, redirecting the new task to other edge nodes with the lowest load index; and decomposing the task into a plurality of subtasks for tasks with calculated amounts exceeding a preset scale, and distributing the subtasks to a plurality of edge nodes for parallel processing.
8. The method for collaborative optimization of edge nodes of a power network according to claim 1, wherein the method further comprises a fault tolerant mechanism: Setting a heartbeat detection period, and periodically sending a heartbeat signal to a central node by each edge node; when the heartbeat signal of a certain edge node is not received in a continuous preset number of detection periods, judging that the edge node fails; And migrating the task on the fault edge node to a standby node, wherein the standby node is comprehensively determined based on the load index and the network distance of the candidate node.
9. The method of claim 1, further comprising pre-allocation of resources and adaptive adjustment of weights: Collecting task arrival data in each edge node historical time window, and predicting task arrival rate of each edge node in a future time window based on a long-short-term memory network; calculating the expected load of each edge node according to the predicted task arrival rate and the current load, and migrating the calculation resource from the low-load node to the high-expected-load node in advance when the expected load exceeds the pre-allocation trigger threshold; calculating a performance evaluation index, and calculating a weight adjustment vector according to the deviation of the performance evaluation index and a preset target value, wherein the performance evaluation index comprises average task completion time, system throughput, resource utilization rate and load balancing degree; And updating the weight coefficient of the reward function according to the weight adjustment vector, and carrying out normalization processing and boundary constraint on the updated weight coefficient.
10. An edge node co-optimization system of a computing power network, characterized in that the system is adapted to implement the method of any of claims 1 to 9, comprising: The resource state acquisition module is used for acquiring resource state information and historical behavior data of a plurality of edge nodes to form a resource state data set; The feature extraction module is used for constructing a computational power network topological graph based on the resource state data set and extracting node feature representation of each node through a graph neural network; the reputation evaluation module is used for calculating the comprehensive reputation of each edge node based on the historical behavior data and integrating the comprehensive reputation into the node characteristic representation; The energy consumption monitoring module is used for collecting energy consumption state data of each edge node and calculating energy consumption efficiency indexes of each edge node; The coupling linkage module is used for dynamically adjusting the energy consumption penalty weight according to the comprehensive credibility of the nodes, triggering a credibility linkage scheduling strategy when the energy consumption is abnormal, and realizing closed-loop optimization of the credibility and the energy consumption; the collaborative optimization module is used for establishing a collaborative optimization model based on a Markov decision process, and training the collaborative optimization model by adopting a multi-agent deep reinforcement learning algorithm; The scheduling decision module is used for inputting the node characteristic representation into the collaborative optimization model to generate a task scheduling decision; And the task execution module is used for distributing the task to the corresponding edge node for execution according to the task scheduling decision, and updating the parameters of the collaborative optimization model based on the execution result.

Description

Edge node collaborative optimization method and system of computing power network Technical Field The invention relates to the technical field of edge computing, in particular to a method and a system for collaborative optimization of edge nodes of a computing power network. Background The existing computing network edge node scheduling is dependent on a static scheduling strategy based on a preset rule or a distributed scheduling method for independent decision of each node, and a single scheduling strategy model is adopted. The method has the advantages that the method is difficult to adapt to network states which change in real time, dynamic optimal configuration of resources cannot be realized, the method has the defects of global visual field loss and lack of an effective cooperation mechanism between nodes, meanwhile, a single strategy is difficult to cover diversified scheduling requirements, and robustness is insufficient in a dynamic heterogeneous computing environment. In addition, the existing scheduling method lacks the capability of prejudging future task loads, prospective configuration of resources is difficult to realize, the weight of the reward function is fixed, and the priority of the optimization target cannot be adjusted in a self-adaptive mode according to the actual running state. More importantly, the prior art has three key defects that firstly, a systematic evaluation mechanism for historical behavior of edge nodes is lacked, the nodes with high reliability cannot be effectively identified and preferentially used, secondly, the existing scheduling method takes less energy consumption factors into consideration, or only takes energy consumption as an independent optimization target, the system performance is difficult to realize, meanwhile, the green energy conservation target is difficult to realize, thirdly, the node reliability evaluation and the energy consumption optimization are mutually independent, the cooperative linkage of the node reliability evaluation and the energy consumption optimization cannot be realized, and a closed loop optimization mechanism cannot be formed to excite the nodes to promote the reliability and the energy efficiency level of the nodes. Disclosure of Invention Aiming at the problems that in the prior art, the optimal configuration of the whole network resources cannot be realized due to independent decision of each edge node, an effective cooperation mechanism is lacking among the edge nodes, the prejudging capability on future task loads is lacking, self-adaptive adjustment cannot be realized due to the fact that the weight of a reward function is fixed, a node historical behavior evaluation mechanism is lacking, energy consumption optimization and credibility evaluation are independent from each other, and the like, the invention provides an edge node cooperative optimization method and system of a computational power network for solving the problems. The technical scheme for solving the technical problems is as follows: in a first aspect, the invention provides a method for collaborative optimization of edge nodes of a computing power network, which comprises the steps of collecting resource state information and historical behavior data of a plurality of edge nodes to form a resource state data set; Constructing a computational power network topological graph based on the resource state data set, and extracting node characteristic representation of each node through a graph neural network; calculating the comprehensive reputation of each edge node based on the historical behavior data, and integrating the comprehensive reputation into the node characteristic representation; Collecting energy consumption state data of each edge node, and calculating energy consumption efficiency indexes of each edge node; Establishing a coupling linkage mechanism of the credibility and the energy consumption perception, dynamically adjusting the energy consumption penalty weight of the node according to the comprehensive credibility of the node, increasing the energy consumption penalty weight and reducing the task allocation priority when the comprehensive credibility of the node is lower than a preset credibility threshold, and triggering a credibility linkage scheduling strategy when the energy consumption of the system is abnormal; establishing a collaborative optimization model based on a Markov decision process, and training the collaborative optimization model by adopting a multi-agent deep reinforcement learning algorithm, wherein the collaborative optimization model comprises a reward function, and the reward function comprises five weighted combinations of positive rewards of system throughput, negative penalties of average task delay, positive rewards of resource utilization balance, negative penalties of system total energy consumption and positive rewards of average credibility; inputting the node characteristic representation into the collaborativ