Collaborative optimization method and system for edge nodes of computing power network

Abstract

The invention relates to the technical field of edge computing, in particular to an edge node collaborative optimization method and system of a computing power network. Comprising 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 computing power network topological graph based on the resource state data set, and extracting node feature representation of each node through a graph neural network; calculating the comprehensive credibility of each edge node based on the historical behavior data, and integrating the comprehensive credibility into the node feature representation; collecting energy consumption state data of each edge node, and calculating an energy consumption efficiency index of each edge node; and a coupling linkage mechanism of credibility and energy consumption perception is established. Through a coupling linkage mechanism of credibility and energy consumption perception, high resource utilization rate, low task delay, strong system robustness and green energy conservation of computing power network edge node scheduling are realized.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to a method and system for collaborative optimization of edge nodes in a computing power network. Background Technology

[0002] Existing edge node scheduling in computing networks largely relies on static scheduling strategies based on preset rules or distributed scheduling methods where each node makes independent decisions, and most employ a single scheduling strategy model. The former struggles to adapt to real-time changes in network conditions and cannot achieve dynamic optimal resource allocation; the latter suffers from a lack of global perspective and effective inter-node collaboration mechanisms, and a single strategy is insufficient to cover diverse scheduling needs, exhibiting inadequate robustness in dynamic heterogeneous computing environments. Furthermore, existing scheduling methods lack the ability to predict future task loads, making proactive resource allocation difficult; the reward function weights remain fixed, failing to adaptively adjust the priority of optimization objectives based on actual operating conditions.

[0003] More importantly, existing technologies have the following three key shortcomings: First, they lack a systematic evaluation mechanism for the historical behavior of edge nodes, making it impossible to effectively identify and prioritize nodes with high reliability; second, existing scheduling methods rarely consider energy consumption factors, or only treat energy consumption as an independent optimization objective, making it difficult to achieve the goal of green energy saving while ensuring system performance; third, node reliability assessment and energy consumption optimization are independent of each other, failing to achieve synergy between the two and failing to form a closed-loop optimization mechanism to incentivize nodes to improve their reliability and energy efficiency. Summary of the Invention

[0004] This invention addresses the problems in existing technologies, such as the inability to achieve optimal allocation of network resources due to independent decision-making by each edge node, the lack of effective cooperation mechanisms between edge nodes, the lack of predictive ability for future task loads, the fixed and unadaptable weights of the reward function, the lack of a mechanism for evaluating historical node behavior, and the independence of energy consumption optimization and credibility evaluation. It provides a collaborative optimization method and system for edge nodes in a computing power network to solve these problems.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a collaborative optimization method for edge nodes of a computing power network, comprising: collecting resource status information and historical behavior data of multiple edge nodes to form a resource status dataset; A computing power network topology graph is constructed based on the resource status dataset, and node feature representations of each node are extracted through a graph neural network. The comprehensive reputation score of each edge node is calculated based on the historical behavior data, and the comprehensive reputation score is incorporated into the node feature representation. Collect energy consumption status data of each edge node and calculate the energy efficiency index of each edge node; Establish a linkage mechanism between reputation and energy consumption perception. Dynamically adjust the energy consumption penalty weight of a node based on its overall reputation. When the overall reputation of a node is lower than a preset reputation threshold, increase the energy consumption penalty weight and reduce the task allocation priority. Trigger a reputation-linked scheduling strategy when the system energy consumption is abnormal. A collaborative optimization model is established based on Markov decision process, and the collaborative optimization model is trained using multi-agent deep reinforcement learning algorithm. The collaborative optimization model includes a reward function, which contains a weighted combination of five factors: positive reward for system throughput, negative penalty for average task delay, positive reward for resource utilization balance, negative penalty for total system energy consumption, and positive reward for average reputation. The node feature representations are input into the collaborative optimization model to generate task scheduling decisions; Based on the task scheduling decision, tasks are distributed to the corresponding edge nodes for execution, and the parameters of the collaborative optimization model are updated based on the execution results.

[0006] Optionally, node feature representations of each node can be extracted using a graph neural network, including: Construct the computing power network topology graph, which includes a node set and a network connection set, wherein the node set includes all edge nodes and the network connection set represents the network connection relationship between edge nodes; A graph convolutional network structure is used to perform multi-layer iterative updates on node features. During each layer update, the feature information of neighboring nodes is aggregated and normalized and weighted according to the degree of the node. A nonlinear activation function is applied to the transformed features to obtain the updated node feature representation.

[0007] Optionally, the collaborative optimization model is trained using a multi-agent deep reinforcement learning algorithm, including: A centralized training and distributed execution mechanism is adopted, and the state information of all edge nodes is collected during the training phase to construct a global state vector; A value function is computed using a deep Q-network, which is obtained by mixing the local value functions of each edge node; Based on the comprehensive reputation score, the local value function is modulated, and nodes with higher reputation scores obtain higher value function estimates, thereby increasing their probability of being selected to execute tasks. During the execution phase, each edge node independently performs actions based on its local state, achieving distributed decision-making.

[0008] Optionally, the comprehensive reputation score of each edge node is calculated based on the historical behavior data, including: Within a preset historical observation period, the task execution success rate, failure frequency, response latency score, response latency fluctuation rate, and energy efficiency score of each edge node are collected. The task execution success rate is the ratio of the number of successfully completed tasks to the total number of assigned tasks; The fault occurrence frequency is the ratio of the number of faults to the observation duration within the observation period; The response delay score is calculated based on the normalized values ​​of the node's average response delay relative to the system's minimum and maximum response delays; The response delay volatility is the ratio of the standard deviation of the response delay to the average response delay; The energy efficiency score is calculated based on the normalized values ​​of node energy efficiency relative to the minimum and maximum energy efficiency of the system. The comprehensive credit rating is obtained by weighting and summing the above five indicators, with the frequency of failures and the volatility of response delay being scored in reverse. The overall reputation score is used as an additional dimension of the input feature vector of the graph neural network.

[0009] Optionally, energy consumption status data of each edge node is collected and energy efficiency indicators are calculated, including: Collect the instantaneous power consumption, energy consumption per unit computing power, and cumulative energy consumption of each edge node; Calculate the energy efficiency index of each edge node, where the energy efficiency index is the ratio of effective computing power output to instantaneous power consumption; The energy efficiency index is normalized and then used as the basis for calculating the energy efficiency score in the overall credit rating.

[0010] Optionally, the coupling mechanism between reputation and energy consumption perception includes: Establish a dynamic adjustment relationship between reputation and energy consumption weight, where nodes with high reputation receive lower energy consumption penalty weights and nodes with low reputation receive higher energy consumption penalty weights. A reputation threshold is set. When the overall reputation of a node is lower than the reputation threshold, a low-reputation node penalty strategy is implemented, including increasing the energy consumption penalty weight of the node and proportionally reducing the task allocation probability of the node. In the adaptive adjustment of reward function weights, the trends in reputation and energy efficiency are used as input factors for weight adjustment. An energy consumption anomaly detection mechanism is set up. When the total energy consumption of the system exceeds the sum of the historical average and the standard deviation by a preset multiple, it is judged as an energy consumption anomaly. When an energy consumption anomaly is triggered, a reputation-linked scheduling strategy is executed, including proportionally reducing the task allocation to low-reputation nodes, prioritizing the allocation of tasks to nodes that meet both the reputation threshold and the energy consumption efficiency threshold, and temporarily increasing the energy consumption penalty coefficient.

[0011] Optionally, the method further includes a load balancing step: Calculate the load metrics for each edge node, where the load metrics are a weighted combination of computing resource utilization, memory utilization, and task queue length; When the load metric of an edge node exceeds a preset load threshold, new tasks are redirected to other edge nodes with the lowest load metric. For tasks with computational demands exceeding a preset scale, the task is decomposed into multiple subtasks and distributed to multiple edge nodes for parallel processing.

[0012] Optionally, the method also includes a fault-tolerance mechanism: Set a heartbeat detection cycle, and each edge node periodically sends a heartbeat signal to the central node; If a heartbeat signal is not received from a certain edge node for a preset number of consecutive detection cycles, the edge node is determined to be faulty. Tasks on the faulty edge node are migrated to the backup node, which is determined based on a combination of the candidate node's load index and network distance.

[0013] Optionally, the method further includes resource pre-allocation and adaptive weight adjustment: Collect task arrival data within the historical time window of each edge node, and predict the task arrival rate of each edge node within the future time window based on the long short-term memory network. The expected load of each edge node is calculated based on the predicted task arrival rate and the current load. When the expected load exceeds the pre-allocated trigger threshold, computing resources are migrated from low-load nodes to high-expected-load nodes in advance. Calculate performance evaluation metrics, and calculate a weight adjustment vector based on the deviation between the performance evaluation metrics and the preset target value. The performance evaluation metrics include average task completion time, system throughput, resource utilization, and load balancing. The weight coefficients of the reward function are updated according to the weight adjustment vector, and the updated weight coefficients are normalized and subject to boundary constraints.

[0014] Secondly, the present invention provides an edge node collaborative optimization system for a computing power network, comprising: The resource status acquisition module is used to collect resource status information and historical behavior data from multiple edge nodes to form a resource status dataset. The feature extraction module is used to construct a computing power network topology graph based on the resource status dataset and extract node feature representations of each node through a graph neural network. The reputation assessment module is used to calculate the comprehensive reputation score of each edge node based on historical behavior data, and to integrate the comprehensive reputation score into the node feature representation; The energy consumption monitoring module is used to collect energy consumption status data of each edge node and calculate the energy efficiency index of each edge node. The coupling and linkage module is used to dynamically adjust the energy consumption penalty weight based on the node's comprehensive reputation. When energy consumption is abnormal, it triggers the reputation-linked scheduling strategy to achieve closed-loop optimization of reputation and energy consumption. The collaborative optimization module is used to establish a collaborative optimization model based on Markov decision processes and to train the collaborative optimization model using a multi-agent deep reinforcement learning algorithm. The scheduling decision module is used to input the node feature representation into the collaborative optimization model to generate task scheduling decisions; The task execution module is used to distribute tasks to corresponding edge nodes for execution according to the task scheduling decision, and update the parameters of the collaborative optimization model based on the execution results.

[0015] By implementing this invention, the following beneficial effects can be achieved: collecting resource status information and historical behavior data provides a complete data foundation for collaborative optimization; extracting node features through graph neural networks enables the model to perceive the global network state; identifying and prioritizing nodes with excellent historical performance based on a dynamic evaluation mechanism of node reputation; reducing energy consumption while ensuring performance through an energy-aware green scheduling strategy; and establishing a virtuous cycle where reputation and energy consumption are coupled and linked, allowing the two optimization objectives to mutually promote each other, with high-reputation nodes receiving more flexible energy consumption constraints, thus incentivizing nodes to improve reliability and energy efficiency. In summary, this invention can achieve high resource utilization, low task latency, strong system robustness, rapid environmental adaptability, and green energy saving in edge node scheduling of computing power networks. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a collaborative optimization method for edge nodes in a computing network provided by the present invention; Figure 2 A schematic diagram of the structure of an edge node collaborative optimization system for a computing power network provided by the present invention; In the attached diagram, the components represented by each number are as follows: Resource status acquisition module 11, Feature extraction module 12, Reputation assessment module 13, Energy consumption monitoring module 14, Coupling and linkage module 15, Collaborative optimization module 16, Scheduling decision module 17, and Task execution module 18. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0020] Example 1, as Figure 1 As shown, this embodiment of the invention provides a collaborative optimization method for edge nodes of a computing power network, including the following steps: S100: Collects resource status information and historical behavior data from multiple edge nodes to form a resource status dataset; S200: Construct a computing power network topology graph based on the resource status dataset, and extract node feature representations of each node through a graph neural network; S300: Calculate the comprehensive reputation score of each edge node based on the historical behavior data, and integrate the comprehensive reputation score into the node feature representation; S400: Collects energy consumption status data of each edge node and calculates the energy efficiency index of each edge node; S500: Establish a coupling and linkage mechanism between credit rating and energy consumption perception to achieve closed-loop optimization of credit rating and energy consumption; S600: A collaborative optimization model is established based on Markov decision process, and the collaborative optimization model is trained using a multi-agent deep reinforcement learning algorithm; S700: Input the node feature representation into the collaborative optimization model to generate task scheduling decisions; S800: Distribute tasks to corresponding edge nodes for execution according to the task scheduling decision, and update the parameters of the collaborative optimization model based on the execution results.

[0021] In step S100 of this application embodiment, it is necessary to collect resource status information and historical behavior data of multiple edge nodes to form a resource status dataset.

[0022] The resource status information includes processor utilization, memory usage, task queue length, network bandwidth, and network latency. Specifically, processor utilization is collected from the CPU and GPU utilization of each edge node; memory usage is collected from the memory usage of each node; task queue length is collected from the number of tasks currently waiting to be executed on each node; and network bandwidth and network latency are collected from the network connection status of each node.

[0023] The historical behavior data includes task execution success rate, failure frequency, failure recovery time, response latency fluctuation rate, and energy efficiency. The collection period for historical behavior data is set to T. history The recommended value is 24 hours, collecting statistical values ​​of various indicators within the collection window.

[0024] The data collection method can be either active polling or passive reporting. The collection period for resource status information can be set according to the actual application scenario, for example, it can be set to any value between 100 milliseconds and 10 seconds.

[0025] In step S200 of this application embodiment, a computing power network topology graph is constructed based on the resource status dataset, and node feature representations of each node are extracted through a graph neural network.

[0026] First, construct the computing power network topology graph G=(V,E), where V represents the set of nodes, including all edge nodes, and E represents the set of network connections, representing the network connection relationships between edge nodes. For example, a computing power network includes 5 edge nodes, where node 1 is connected to nodes 2 and 3, node 2 is connected to node 4, and node 3 is connected to node 5. Then the node set V={1,2,3,4,5} and the network connection set E={(1,2),(1,3),(2,4),(3,5)}.

[0027] A graph convolutional network structure is used to perform multi-layer iterative updates on node features. The node feature update formula is: hᵢ ( ˡ +1) =σ(Σⱼ∈N(i)(1 / √(dᵢ·dⱼ))·W ( ˡ ) ·hⱼ ( ˡ ) ) Among them, h i (l+1) h represents the feature representation of node i at layer l+1 (i.e., after the update). j (l)Let N(i) represent the feature representation of node j at layer l, and let N(i) represent the set of neighboring nodes of node i. j∈N(i) This represents summing and aggregating all neighboring nodes j of node i, d i d represents the degree of node i (i.e., the number of neighbors of node i). j Denotes the degree of node j, 1 / √(d i ·d j W is a normalization coefficient used to eliminate the influence of differences in node degree. (l) Let represent the learnable weight matrix of the l-th layer, and σ represent the non-linear activation function. The physical meaning of this formula is: when each node updates, it aggregates the feature information of all its neighboring nodes, and after weight transformation and activation function, it obtains a new feature representation. Through multiple iterations, the node can perceive neighbor information from a wider range.

[0028] The specific parameter settings for the graph convolutional network are as follows: The network has 3 layers (L); the weight matrix of each layer has a dimension of d. in ×d out , where the input layer d in The original feature dimension is 6, which includes two parts: the first part is the 5-dimensional resource status information (processor utilization, memory usage, task queue length, network bandwidth, network latency) collected in step S100, and the second part is the 1-dimensional graph structure feature (node ​​degree) calculated based on the computing power network topology; the dimensions of the hidden layer and the output layer are both set to 128; the activation function σ adopts the ReLU function, that is, σ(x)=max(0,x); the weight matrix is ​​initialized using the Xavier initialization method, and the weights W are randomly sampled from the uniform distribution U(-a,a), where a=√(6 / (dᵢ)). n +d o ᵤ t )), dᵢ n For the input dimension, d o ᵤ t For the output dimension, this initialization method maintains the stability of the signal variance during forward and backward propagation; the learning rate is set to 0.001, and the Adam optimizer is used for parameter updates; the dropout rate is set to 0.5 to prevent overfitting.

[0029] In step S300 of this application embodiment, the comprehensive reputation of each edge node is calculated based on the historical behavior data, and the comprehensive reputation is integrated into the node feature representation.

[0030] The core objective of this step is to establish a node trust measurement mechanism, identify and prioritize edge nodes with good historical performance, reduce task failures and rescheduling overhead caused by node failures, and improve the overall reliability and stability of the system.

[0031] Task execution success rate S i The calculation formula is: Sᵢ=N success / N total Among them, S i This represents the task execution success rate of node i, with a value range of [0,1], N. success N represents the number of tasks successfully completed. total This represents the total number of tasks assigned to node i. For example, if a node receives 100 tasks in 24 hours and successfully completes 95, then S represents the total number of tasks assigned to node i. i =0.95.

[0032] Fault occurrence frequency F i The calculation formula is: Fᵢ=N fault / T observation Among them, F i N represents the frequency of failures at node i, expressed in times per hour. fault T represents the number of faults that occur within the observation period. observation This indicates the duration of the observation period (in hours). For example, if a node experiences two failures within 24 hours, then F... i =2 / 24≈0.083. For ease of normalization, F... i It is limited to the range [0,1].

[0033] Response latency score T i The calculation formula is: Tᵢ=1-(τ avgi -τ min ) / (τ max -τ min ) Among them, T i This represents the response delay score of node i, with a value ranging from [0,1]. A larger value indicates a faster response speed. τ avgi τ represents the average response delay of node i. min and τ max Let T represent the minimum and maximum average response times across all nodes, respectively. For example, if the average response time of a node is 50ms, the minimum response time in the system is 20ms, and the maximum response time is 200ms, then T... i =1-(50-20) / (200-20)≈0.833.

[0034] Response delay volatility V i The calculation formula is: Vᵢ=σ τi / τ avgi Among them, V i σ represents the response delay volatility of node i, also known as the coefficient of variation. A smaller value indicates better node stability. τi τ represents the standard deviation of the response delay of node i. avgi This represents the average response delay. For example, if the standard deviation of a node's response delay is 15ms and the average response delay is 50ms, then V... i =15 / 50=0.3. V i The smaller the value, the better the node's stability.

[0035] Overall Credit R i The calculation formula is: Rᵢ=ω s ·Sᵢ+ω f ·(1-Fᵢ)+ω t ·Tᵢ+ωᵥ·(1-Vᵢ)+ω e ·Eᵢ Among them, R i S represents the overall credibility of node i, with a value ranging from [0,1]. A larger value indicates a more trustworthy node. i Indicates the task execution success rate; F i To indicate the frequency of failures, the formula uses (1-F) i Reverse scoring is used; fewer faults result in a higher score. i V represents the response delay score; i This represents the response delay volatility, expressed in the formula as (1-V) i Reverse scoring is applied; the smaller the fluctuation, the higher the score. i This represents the energy efficiency score, ω. s ω f ω t ω v ω e ω is the weighting coefficient. s The weighting coefficient representing the task execution success rate; ω f The weighting coefficient representing the frequency of failure occurrence; ω t The weighting coefficient ω represents the response delay score. v The weighting coefficient representing the response delay volatility; ω e The weighting coefficients representing the energy efficiency score satisfy ω s +ω f +ω t +ω v +ω e =1. The calculation of the energy efficiency score Eᵢ relies on the synchronous acquisition of energy status data. Recommended values ​​for each weighting coefficient are: ω s =0.25、ω f =0.20、ω t=0.20、ω v =0.15、ω e =0.20.

[0036] Calculation example: The indicators of a certain node are S i =0.95, F i =0.083, T i =0.833, V i =0.3、E i =0.75, then R i =0.25×0.95+0.20×(1-0.083)+0.20×0.833+0.15×(1-0.3)+0.20×0.75=0.2375+0.1834+0.1666+0.105+0.15=0.8425.

[0037] The method of integrating comprehensive reputation into node feature representation is as follows: based on the original 128-dimensional node feature vector, a reputation dimension is added to form a 129-dimensional enhanced feature vector h. i_enhanced =[h i ,R i ].

[0038] In step S400 of this embodiment, energy consumption status data of each edge node is collected, and energy efficiency index of each edge node is calculated.

[0039] The core objective of this step is to reduce overall energy consumption while ensuring system performance, which aligns with the development trends of green computing and dual-carbon goals. At the same time, it links energy efficiency with node reputation, laying the foundation for subsequent coupling and linkage mechanisms.

[0040] The energy consumption status data collection period is set to T. power The recommended value is 1 second, which can be collected through the power consumption monitoring module built into the edge node or an external smart meter.

[0041] Instantaneous power consumption P i The unit of measurement is watts (W), representing the real-time power consumption of node i at the time of measurement. For example, the instantaneous power consumption of an edge server is 350W.

[0042] Power Usage Effectiveness (PUE) per Unit of Computing Power i The calculation formula is: PUEᵢ=Pᵢ / Cᵢ Among them, PUE i This represents the energy consumption per unit of computing power for node i, expressed in watts per task. A smaller value indicates higher energy efficiency. (P) i C represents the instantaneous power consumption of node i, in watts. iThe effective computing power output of node i can be measured by the number of tasks processed per second. For example, if a node has an instantaneous power consumption of 350W and can process 500 standard tasks per second, then the PUE is... i =350 / 500=0.7W / task.

[0043] Cumulative energy consumption E cumulative_i The calculation formula is: E cumulative_i =Σ k ₌1 N Pᵢ(k)·Δt Where N is the number of sampling points and Δt is the sampling interval. For example, if the sampling interval is 1 second and there are a total of 86,400 sampling points in 24 hours, and the average power consumption is 350W, then the cumulative energy consumption is 350 × 86,400 = 30,240,000Ws = 8.4kWh.

[0044] Energy efficiency index η i The calculation formula is: ηᵢ=Cᵢ / Pᵢ Where, η i This represents the energy efficiency index of node i, measured in tasks per watt; a higher value indicates higher energy efficiency. i P represents the effective computing power output of node i, in tasks per second; i η represents the instantaneous power consumption of node i, in watts. For example, if a node has an effective computing power of 500 tasks / second and a power consumption of 350W, then η i =500 / 350≈1.43 tasks / watt.

[0045] Energy efficiency score E i The normalization formula is: Eᵢ=(ηᵢ-η min ) / (η max -η min ) Among them, E i η represents the energy efficiency score of node i, with a value range of [0,1], and is used for comprehensive reputation calculation. i η represents the energy efficiency index of node i. min and η max These represent the minimum and maximum energy efficiency values ​​of all edge nodes in the current system, respectively. For example, η in the system min =0.8 tasks / watt, η max =2.0 tasks / watt, for a node η i =1.43 tasks / watt, then E i =(1.43-0.8) / (2.0-0.8)≈0.525.

[0046] In step S500 of this embodiment, a coupled linkage mechanism between reputation and energy consumption perception is established to achieve closed-loop optimization of reputation and energy consumption. This is the core innovation of this invention.

[0047] The core objective of this step is to elevate credit rating assessment and energy consumption optimization from parallel and independent processes to a collaborative and linked process, so that the two optimization objectives can promote each other. At the same time, the linkage scheduling can be automatically triggered when energy consumption is abnormal, thereby improving the system's adaptive optimization capability.

[0048] Reputation and Energy Consumption Weighting Dynamic Adjustment Function Based on the node's overall reputation R i The energy consumption penalty weight of this node is dynamically adjusted, and the calculation formula is as follows: δᵢ=δ base ·(1-μ·Rᵢ) Where, δ i δ represents the personalized energy consumption penalty weight for node i, used to apply a weighted penalty to the energy consumption of that node in the reward function; base represents the base energy consumption penalty coefficient, which is the benchmark value for the energy consumption penalty weight of all nodes, and a recommended value of 0.15; μ represents the reputation adjustment factor, which controls the degree of influence of reputation on energy consumption weight, with a value range of [0, 0.5], and a recommended value of 0.3; R i δ represents the overall reputation score of node i, with a value ranging from [0,1]. base The recommended value is 0.15; μ is the reputation adjustment factor, with a value range of [0, 0.5], and a recommended value of 0.3.

[0049] The physical meaning of this formula is: high-reputation nodes receive lower energy consumption penalty weights, meaning they have more flexible energy consumption constraints in scheduling decisions. Calculation example: A high-reputation node R... i =0.9, then δ i =0.15×(1-0.3×0.9)=0.15×0.73=0.1095; A low-reputation node R i =0.4, then δ i =0.15×(1-0.3×0.4)=0.15×0.88=0.132. This mechanism incentivizes nodes to improve their reliability in order to obtain better scheduling treatment.

[0050] Low-reputation node penalty strategy Set the reputation threshold R threshold A value of 0.5 is recommended. When the node's overall reputation R... i Below R threshold When necessary, implement the following adjustment strategy: First, increase the energy consumption penalty weight of this node, and adjust the formula as follows: δᵢ=δ base·(1+ρ·(R threshold -Rᵢ)) Where ρ is the low reputation penalty coefficient, and a value of 1.0 is recommended. For example, for a node R... i =0.3, then δ i =0.15×(1+1.0×(0.5-0.3))=0.15×1.2=0.18.

[0051] Second, reduce the task allocation priority of this node, and adjust the formula as follows: P iadjusted =P ioriginal ·(Rᵢ / R threshold ) Among them, P ioriginal Let P be the original choice probability. iadjusted This is the adjusted selection probability. For example, if the original selection probability of a node is 0.4, R... i =0.3, R threshold =0.5, then P iadjusted =0.4×(0.3 / 0.5)=0.24.

[0052] Creditworthiness and energy consumption trends are incorporated into the weighted adaptive adjustment. In the adaptive adjustment step of the reward function weights, the reputation change trend ΔR is used. avg and the trend of energy consumption efficiency change Δη avg As an input factor to the weight adjustment vector, the formula for calculating the trend of change is: ΔR avg =R avg (t)-R avg (t-1) Δη avg =η avg (t)-η avg (t-1) The formula for updating the weight adjustment vector is expanded to: Δw=η w ·[∂R / ∂α,∂R / ∂β,∂R / ∂γ,∂R / ∂δ,∂R / ∂ε] T +λ r ·ΔR avg· e ε +λ e ·Δη avg· e δ Where, η w The learning rate is adjusted for weights, and a value of 0.01 is recommended; λ r The credibility trend influence coefficient is recommended to be 0.05; λ eThe energy efficiency trend influence coefficient is recommended to be 0.05. This mechanism enables the system to automatically adjust the optimization weights based on the changing trends of creditworthiness and energy efficiency. ε =[0,0,0,0,1] T Let e ​​be the fifth-dimensional unit vector. δ =[0,0,0,1,0] T As a fourth-dimensional unit vector, the trend of reputation change only affects the adjustment of reputation weight ε, and the trend of energy efficiency change only affects the adjustment of energy consumption weight δ.

[0053] The partial derivative is calculated analytically based on the linear structure of the reward function, specifically: ∂R / ∂α=T throughput , ∂R / ∂β=-L delay , ∂R / ∂γ=U balance , ∂R / ∂δ=-E total , ∂R / ∂ε=R avg This partial derivative vector reflects the marginal contribution of each performance indicator to the reward value and is used to guide the adjustment direction of the weight coefficients.

[0054] Energy consumption anomaly detection and credit-based scheduling mechanism When the system detects an abnormal increase in energy consumption, it triggers an emergency energy consumption scheduling strategy. The criteria for determining abnormal energy consumption are: E total >E avg +k e ·σ E Among them, E avg σ represents the historical average total energy consumption of the system. E Let k be the standard deviation of the system's total energy consumption. e The anomaly detection coefficient is recommended to be 2.0.

[0055] When an energy consumption anomaly is triggered, the following reputation-based scheduling strategy adjustment will be implemented: First, for those with a credit rating lower than R threshold For nodes, reduce their task allocation ratio to ξ times the original ratio, with a recommended value of ξ=0.5.

[0056] Second, prioritize assigning tasks to those that meet condition R. i >R threshold And E i >E threshold_efficiency High-reputation, high-energy-efficiency nodes, including E threshold_efficiency The recommended value for the energy efficiency threshold is 0.6.

[0057] Third, adjust the energy consumption penalty coefficient δ in the reward function to δ emergency =δ base ×k emergency , where kemergency The emergency amplification factor is recommended to be 1.5.

[0058] Through the above three-level coupling and linkage mechanism, a deep integration of credit assessment and energy consumption optimization can be achieved, forming a virtuous cycle.

[0059] When the total energy consumption of the system is ≤ E for N consecutive cycles avg +k e ·σ E When the energy consumption is restored to normal, the scheduling parameters are restored to their normal state.

[0060] In step S600 of this embodiment, a collaborative optimization model is established based on a Markov decision process, and the collaborative optimization model is trained using a multi-agent deep reinforcement learning algorithm.

[0061] First, a collaborative optimization model is established based on the Markov decision process. The state space S is defined as the set of resource states of each edge node, and the action space A is defined as the set of task allocation schemes.

[0062] During the training phase, the central node collects the state information of all edge nodes and constructs a global state vector s. t =[s1ᵗ,s2ᵗ,...,s n ᵗ], where s i ᵗ represents the state of the i-th edge node at time t, and n is the total number of edge nodes.

[0063] Multi-agent reinforcement learning can employ the QMIX algorithm, whose total value function is expressed as: Q tot (s,a)=f mixer (Q1(s1,a1),...,Q n (s n ,a n ),s) Where f mixer This is a hybrid network function.

[0064] Based on the comprehensive reputation score, the value function is modulated by introducing a reputation score modulation factor into the local value function calculation of the QMIX algorithm: Q iadjusted (sᵢ,aᵢ)=Qᵢ(sᵢ,aᵢ)·(1+φ·Rᵢ) Where φ is the reputation modulation coefficient, and the recommended value range is [0.1, 0.5], for example, φ=0.3. This modulation mechanism enables nodes with high reputation to obtain a higher value function estimate when selecting actions, thereby increasing the probability that they are selected to execute tasks.

[0065] Set the reward function as follows: R=α·T throughput-β·L delay +γ·U balance -δ·E total +ε·R avg Where R represents the instantaneous reward value at the current moment; α represents the weighting coefficient of the system throughput, with a recommended value of 0.30; T throughput β represents system throughput, i.e., the number of tasks completed per unit time, expressed as tasks / second; β represents the weighting coefficient for average task latency, with a recommended value of 0.25; L delay γ represents the average task latency, i.e., the average time from task submission to completion, in milliseconds; γ represents the weighting coefficient for resource utilization balance, with a recommended value of 0.20; U balance This represents the resource utilization balance, with a value ranging from [0,1]. A larger value indicates a more balanced load across nodes; δ represents the weighting coefficient for the total system energy consumption, with a recommended value of 0.15; E total R represents the total system energy consumption, which is the sum of the power consumption of all edge nodes, in watts; ε represents the weighting coefficient of average reputation, with a recommended value of 0.10; avg This represents the average reputation of all edge nodes. Recommended values ​​for the weighting coefficients are: α=0.30, β=0.25, γ=0.20, δ=0.15, ε=0.10. The energy consumption penalty term in the reward function can be further refined into a weighted sum of the energy consumption of each node, i.e., E total =Σᵢδᵢ·Pᵢ, where δᵢ is the personalized energy consumption penalty weight calculated in step S500.

[0066] The specific parameter settings for the QMIX algorithm are as follows: Discount factor γ discount Set to 0.99; Target network update period T target The number of steps is set to 200. The exploration strategy adopts an ε-greedy strategy, which means randomly selecting an action for exploration with probability ε and selecting the current best action for utilization with probability 1-ε. The initial value of ε is set to 1.0, and the final value is set to 0.05. The decay step count is set to 50,000 steps, and a linear decay method is used to gradually shift from exploration to utilization. The hybrid network adopts a two-layer fully connected network with a hidden layer dimension of 32 and an activation function of ELU. The local Q network of each agent adopts a three-layer fully connected network with hidden layer dimensions of 64 and 64 respectively and an activation function of ReLU.

[0067] In step S700 of this embodiment, the node feature representation is input into the collaborative optimization model to generate a task scheduling decision. The 129-dimensional enhanced node feature vector (including the reputation dimension) extracted by the graph neural network is input into each sub-policy network of the collaborative optimization model. Each sub-policy network outputs the probability distribution of scheduling actions in parallel, and finally selects the scheduling action with the highest probability as the output.

[0068] In step S800 of this embodiment, the task is distributed to the corresponding edge node for execution according to the task scheduling decision, and the parameters of the collaborative optimization model are updated based on the execution results.

[0069] First, a scheduling decision is made, distributing the task to the target edge nodes for execution and recording the execution results. Then, the state transition tuple (s) is generated. t ,a t ,r t ,s t+1 Store in experience pool D.

[0070] The specific parameter settings for the experience pool are as follows: Experience pool capacity D size Set to 100,000 state transition records; batch size is set to 64 records; priority is given to experience replay with weight α. priority Set to 0.6, importance sampling weight β importance The initial value is set to 0.4, and the final value is set to 1.0; the experience pool filling threshold is set to 10,000 entries.

[0071] Randomly sample batches of data from the experience pool for model training, and update the network parameters using gradient descent: θ t+1 =θ t -η·∇ θ L(θ) Where η represents the learning rate, and L(θ) represents the loss function. The learning rate can be set to 0.001.

[0072] In this embodiment of the application, the method further includes a load balancing step.

[0073] The formula for calculating the load index is: L i =ω1·C i +ω2·M i +ω3·Q i C i M represents the computational resource utilization rate. i Q represents memory usage. i This represents the normalized value of the task queue length. Recommended values ​​for the weight parameters are: ω1=0.4, ω2=0.3, ω3=0.3. Load threshold L threshold The recommended value is 0.8.

[0074] When the load metric L of node i is detected i Exceeding threshold L threshold When a new task is assigned, it is redirected to the node j with the lowest load metric, where j = argmin. k∈V,k≠i L k .

[0075] In this embodiment of the application, the method further includes a fault tolerance mechanism.

[0076] Set the heartbeat detection cycle T heartbeat The recommended value is 1 second. When n consecutive seconds... miss If no heartbeat signal is received from node i in n cycles, node i is determined to have failed. miss It is recommended to take 3 values.

[0077] The strategy for selecting a backup node is as follows: node backup =argmin j ∈V available (Lⱼ+λ·Dᵢⱼ) Where V available D represents the set of available nodes. ij This represents the network distance from node i to node j, and λ is the balancing coefficient, with a recommended value of 0.5. The task migration timeout is set to 30 seconds.

[0078] In this embodiment of the application, the method further includes resource pre-allocation and weight adaptive adjustment steps.

[0079] Collect task arrival sequences within historical time windows of each edge node to construct a task arrival rate time series dataset. The recommended historical window length W is 60 seconds, and the recommended prediction time window Δt is 10 seconds. The LSTM network consists of two hidden layers, each with 128 hidden units. A pre-assigned trigger threshold L... prealloc The recommended value is 0.7.

[0080] Weight-adjusted learning rate η w A value of 0.01 is recommended, with a weight adjustment cycle of 1000 steps. The weight coefficient should be limited to the range of [0.1, 0.6] to ensure that each optimization objective receives appropriate attention. Performance targets can be set as follows: average task completion time target of 100ms, system throughput target of 1000 tasks / second, resource utilization target of 0.7, and load balancing target of 0.9.

[0081] like Figure 2 As shown, based on the same inventive concept as the edge node collaborative optimization method for a computing power network provided in Embodiment 1, this embodiment of the invention also provides an edge node collaborative optimization system for a computing power network, comprising: Resource status acquisition module 11 is used to collect resource status information and historical behavior data of multiple edge nodes to form a resource status dataset; Feature extraction module 12 is used to construct a computing power network topology graph based on the resource status dataset and extract node feature representations of each node through a graph neural network; The reputation assessment module 13 is used to calculate the comprehensive reputation of each edge node based on historical behavior data, and to integrate the comprehensive reputation into the node feature representation; The energy consumption monitoring module 14 is used to collect energy consumption status data of each edge node and calculate the energy consumption efficiency index of each edge node. The coupling and linkage module 15 is used to dynamically adjust the energy consumption penalty weight according to the node's comprehensive reputation, and trigger the reputation linkage scheduling strategy when energy consumption is abnormal, so as to realize the closed-loop optimization of reputation and energy consumption. Collaborative optimization module 16 is used to establish a collaborative optimization model based on Markov decision process and to train the collaborative optimization model using a multi-agent deep reinforcement learning algorithm. The scheduling decision module 17 is used to input the node feature representation into the collaborative optimization model to generate task scheduling decisions; The task execution module 18 is used to distribute tasks to corresponding edge nodes for execution according to the task scheduling decision, and update the parameters of the collaborative optimization model based on the execution results.

[0082] To verify the effectiveness of the method proposed in this invention, a comparative experiment was designed in this embodiment. The performance of the method of this invention and the existing typical scheduling methods were compared and analyzed through a simulation platform.

[0083] The experiment uses Python 3.8 and PyTorch 1.12 to build a simulation platform to simulate the scheduling scenario of edge nodes in a computing power network. The specific parameter settings for the simulation environment are as follows: Number of edge nodes: 20 heterogeneous edge nodes, with computing power following a uniform distribution U(100,500) GFLOPS; Network topology: A randomly generated small-world network with an average node degree of 4 and a network latency following a uniform distribution U(5,50)ms; Task characteristics: Task arrival follows a Poisson distribution with an average arrival rate of λ = 100 tasks / second; task computation follows an exponential distribution with a mean of 50 MFLOPs. Node power consumption: Idle power consumption P idle The power consumption is U(50,100)W, and the full-load power consumption is P. max It follows the U(200,400)W rule; Node reliability: 20% of the nodes are set as low reliability nodes, whose failure rate is 3 times that of normal nodes; Simulation duration: Each round of the experiment runs 10,000 time steps, with each time step corresponding to 100ms of actual time; Number of experiments repeated: Each experiment was repeated 10 times independently, and the average value was taken as the final result.

[0084] To comprehensively evaluate the performance advantages of the method of this invention, the following four methods were selected as comparative benchmarks: Method 1: Static rule scheduling. This method uses a fixed round-robin strategy for task allocation, without considering node status or historical behavior, and represents the traditional static scheduling method.

[0085] Method 2: Load Balancing Scheduling. This method allocates tasks to the node with the lowest current load based on real-time load information, representing a dynamic scheduling method based on a single metric.

[0086] Method 3: Reputation-Aware Scheduling Only. This method employs the reputation assessment mechanism of the present invention, but does not introduce energy consumption sensing and coupling linkage mechanisms. It is used to verify the gain of the coupling linkage mechanism.

[0087] Method 4: Energy-Aware Scheduling Only. An energy-aware scheduling strategy is adopted, but without introducing reputation assessment and coupling linkage mechanisms, to verify the gain of the reputation mechanism.

[0088] Method 5: The method of this invention. It employs a complete optimization method that couples and links reputation and energy consumption.

[0089] The experiment used the following five metrics to evaluate the performance of each method: Metric 1: System throughput, defined as the number of tasks successfully completed per unit of time, in tasks per second; Metric 2: Average task latency, defined as the average time from task submission to completion, in milliseconds; Indicator 3: Task success rate, defined as the ratio of the number of successfully completed tasks to the total number of submitted tasks, expressed as a percentage; Indicator 4: Total system energy consumption, defined as the cumulative energy consumption of all nodes during the experimental period, in kilowatt-hours (kWh). Indicator 5: Resource utilization balance, defined as 1 minus the standard deviation of the load rate of each node, with a value range of [0,1]. The larger the value, the more balanced the load distribution.

[0090] Experimental Results and Analysis Table 1 shows the experimental results of each method on five evaluation metrics under normal load scenarios.

[0091] Table 1 Performance Comparison of Methods under Normal Load Scenarios As shown in Table 1, under normal load scenarios, the method of this invention outperforms the comparative methods in all evaluation metrics. Specifically, compared with static rule scheduling, the method of this invention improves system throughput by 18.6%, reduces average task latency by 37.3%, improves task success rate by 9.6 percentage points, reduces total system energy consumption by 20.6%, and improves resource utilization balance by 36.1%. Compared with reputation-aware scheduling and energy-aware scheduling, the method of this invention achieves synergistic optimization of performance and energy efficiency through a coupling and linkage mechanism, with significant improvements in all metrics, verifying the effectiveness of the coupling and linkage mechanism.

[0092] Table 2 shows the performance of each method under the abnormal energy consumption scenario. The abnormal energy consumption scenario was simulated by artificially injecting a sudden increase in energy consumption in the middle of the experiment (time steps 5000-6000).

[0093] Table 2 Performance Comparison of Various Methods under Abnormal Energy Consumption Scenarios As shown in Table 2, under abnormal energy consumption scenarios, the performance of each method decreased to varying degrees, but the method of this invention exhibited a significant robustness advantage. Compared to normal scenarios, the throughput of the method of this invention decreased by only 5.3%, while methods one through four decreased by 13.1%, 12.6%, 9.2%, and 10.4%, respectively. This is because after detecting abnormal energy consumption, the method of this invention can respond quickly through a reputation-based scheduling strategy, prioritizing the allocation of tasks to high-reputation, high-energy-efficiency nodes, thereby effectively suppressing the impact of abnormal energy consumption on system performance.

[0094] Table 3 shows the performance improvement ratio of different methods compared to static rule scheduling.

[0095] Table 3 Performance improvement ratio of each method compared to static rule scheduling As shown in Table 3, the method of this invention outperforms other comparative methods in all evaluation metrics. It is particularly noteworthy that the performance improvement of the method of this invention is not a simple summation of the effects of individual mechanisms, but rather a synergistic gain effect generated through a coupled linkage mechanism. For example, Method 3 reduces energy consumption by 8.3%, Method 4 by 15.6%, while the method of this invention achieves a reduction of 20.6%, exceeding the arithmetic mean of the two, demonstrating the synergistic promoting effect of the coupled linkage mechanism on energy consumption optimization.

[0096] Through the above comparative experiments, the following conclusions can be drawn: First, the reputation-energy consumption coupled optimization method proposed in this invention significantly outperforms existing methods in five dimensions: system throughput, task latency, task success rate, system energy consumption, and resource utilization balance. Second, the coupled mechanism can generate a synergistic gain effect, making the overall system performance superior to using reputation-aware or energy-aware scheduling strategies alone. Third, in the event of sudden events such as energy consumption anomalies, the method of this invention exhibits stronger robustness and adaptability, and can quickly respond to and suppress the impact of anomalies on system performance. The experimental results fully verify the effectiveness and superiority of the technical solution of this invention.

[0097] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0098] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0099] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0100] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0102] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Clearly, those skilled in the art can make various alterations and variations to the invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the invention and its equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A collaborative optimization method for edge nodes of a computing power network, characterized in that, include: Collect resource status information and historical behavior data from multiple edge nodes to form a resource status dataset; A computing power network topology graph is constructed based on the resource status dataset, and node feature representations of each node are extracted through a graph neural network. The comprehensive reputation score of each edge node is calculated based on the historical behavior data, and the comprehensive reputation score is incorporated into the node feature representation. Collect energy consumption status data of each edge node and calculate the energy efficiency index of each edge node; Establish a linkage mechanism between reputation and energy consumption perception. Dynamically adjust the energy consumption penalty weight of a node based on its overall reputation. When the overall reputation of a node is lower than a preset reputation threshold, increase the energy consumption penalty weight and reduce the task allocation priority. Trigger a reputation-linked scheduling strategy when the system energy consumption is abnormal. A collaborative optimization model is established based on Markov decision process, and the collaborative optimization model is trained using multi-agent deep reinforcement learning algorithm. The collaborative optimization model includes a reward function. The node feature representations are input into the collaborative optimization model to generate task scheduling decisions; Based on the task scheduling decision, tasks are distributed to the corresponding edge nodes for execution, and the parameters of the collaborative optimization model are updated based on the execution results.

2. The edge node collaborative optimization method for a computing power network according to claim 1, characterized in that, Node feature representations are extracted from each node using a graph neural network, including: Construct the computing power network topology graph, which includes a node set and a network connection set, wherein the node set includes all edge nodes and the network connection set represents the network connection relationship between edge nodes; A graph convolutional network structure is used to perform multi-layer iterative updates on node features. During each layer update, the feature information of neighboring nodes is aggregated and normalized and weighted according to the degree of the node. A nonlinear activation function is applied to the transformed features to obtain the updated node feature representation.

3. The edge node collaborative optimization method for a computing power network according to claim 1, characterized in that, The collaborative optimization model is trained using a multi-agent deep reinforcement learning algorithm, including: A centralized training and distributed execution mechanism is adopted, and the state information of all edge nodes is collected during the training phase to construct a global state vector; A value function is computed using a deep Q-network, which is obtained by mixing the local value functions of each edge node; Based on the comprehensive reputation score, the local value function is modulated, and nodes with higher reputation scores obtain higher value function estimates, thereby increasing their probability of being selected to execute tasks. During the execution phase, each edge node independently performs actions based on its local state, achieving distributed decision-making.

4. The edge node collaborative optimization method for a computing power network according to claim 1, characterized in that, The comprehensive reputation score of each edge node is calculated based on the historical behavior data, including: Within a preset historical observation period, the task execution success rate, failure frequency, response latency score, response latency fluctuation rate, and energy efficiency score of each edge node are collected. The task execution success rate is the ratio of the number of successfully completed tasks to the total number of assigned tasks; The fault occurrence frequency is the ratio of the number of faults to the observation duration within the observation period; The response delay score is calculated based on the normalized values ​​of the node's average response delay relative to the system's minimum and maximum response delays; The response delay volatility is the ratio of the standard deviation of the response delay to the average response delay; The energy efficiency score is calculated based on the normalized values ​​of node energy efficiency relative to the minimum and maximum energy efficiency of the system. The comprehensive credit rating is obtained by weighting and summing the above five indicators, with the frequency of failures and the volatility of response delay being scored in reverse. The overall reputation score is used as an additional dimension of the input feature vector of the graph neural network.

5. The edge node collaborative optimization method for a computing power network according to claim 1, characterized in that, Collect energy consumption status data of each edge node and calculate energy efficiency indicators, including: Collect the instantaneous power consumption, energy consumption per unit computing power, and cumulative energy consumption of each edge node; Calculate the energy efficiency index of each edge node, where the energy efficiency index is the ratio of effective computing power output to instantaneous power consumption; The energy efficiency index is normalized and then used as the basis for calculating the energy efficiency score in the overall credit rating.

6. The edge node collaborative optimization method for a computing power network according to claim 1, characterized in that, The coupling mechanism between reputation and energy consumption perception includes: Establish a dynamic adjustment relationship between reputation and energy consumption weight, where nodes with high reputation receive lower energy consumption penalty weights and nodes with low reputation receive higher energy consumption penalty weights. A reputation threshold is set. When the overall reputation of a node is lower than the reputation threshold, a low-reputation node penalty strategy is implemented, including increasing the energy consumption penalty weight of the node and proportionally reducing the task allocation probability of the node. In the adaptive adjustment of reward function weights, the trends in reputation and energy efficiency are used as input factors for weight adjustment. An energy consumption anomaly detection mechanism is set up. When the total energy consumption of the system exceeds the sum of the historical average and the standard deviation by a preset multiple, it is judged as an energy consumption anomaly. When an energy consumption anomaly is triggered, a reputation-linked scheduling strategy is executed, including proportionally reducing the task allocation to low-reputation nodes, prioritizing the allocation of tasks to nodes that meet both the reputation threshold and the energy consumption efficiency threshold, and temporarily increasing the energy consumption penalty coefficient.

7. The edge node collaborative optimization method for a computing power network according to claim 1, characterized in that, The method also includes a load balancing step: Calculate the load metrics for each edge node, where the load metrics are a weighted combination of computing resource utilization, memory utilization, and task queue length; When the load metric of an edge node exceeds a preset load threshold, new tasks are redirected to other edge nodes with the lowest load metric. For tasks with computational demands exceeding a preset scale, the task is decomposed into multiple subtasks and distributed to multiple edge nodes for parallel processing.

8. The edge node collaborative optimization method for a computing power network according to claim 1, characterized in that, The method also includes a fault tolerance mechanism: Set a heartbeat detection cycle, and each edge node periodically sends a heartbeat signal to the central node; If a heartbeat signal is not received from a certain edge node for a preset number of consecutive detection cycles, the edge node is determined to be faulty. Tasks on the faulty edge node are migrated to the backup node, which is determined based on a combination of the candidate node's load index and network distance.

9. The edge node collaborative optimization method for a computing power network according to claim 1, characterized in that, The method also includes resource pre-allocation and adaptive weight adjustment: Collect task arrival data within the historical time window of each edge node, and predict the task arrival rate of each edge node within the future time window based on the long short-term memory network. The expected load of each edge node is calculated based on the predicted task arrival rate and the current load. When the expected load exceeds the pre-allocated trigger threshold, computing resources are migrated from low-load nodes to high-expected-load nodes in advance. Calculate performance evaluation metrics, and calculate a weight adjustment vector based on the deviation between the performance evaluation metrics and the preset target value. The performance evaluation metrics include average task completion time, system throughput, resource utilization, and load balancing. The weight coefficients of the reward function are updated according to the weight adjustment vector, and the updated weight coefficients are normalized and subject to boundary constraints.

10. An edge node collaborative optimization system for a computing power network, characterized in that, The system is used to implement the method as described in any one of claims 1 to 9, comprising: The resource status acquisition module is used to collect resource status information and historical behavior data from multiple edge nodes to form a resource status dataset. The feature extraction module is used to construct a computing power network topology graph based on the resource status dataset and extract node feature representations of each node through a graph neural network. The reputation assessment module is used to calculate the comprehensive reputation score of each edge node based on historical behavior data, and to integrate the comprehensive reputation score into the node feature representation; The energy consumption monitoring module is used to collect energy consumption status data of each edge node and calculate the energy efficiency index of each edge node. The coupling and linkage module is used to dynamically adjust the energy consumption penalty weight based on the node's comprehensive reputation. When energy consumption is abnormal, it triggers the reputation-linked scheduling strategy to achieve closed-loop optimization of reputation and energy consumption. The collaborative optimization module is used to establish a collaborative optimization model based on Markov decision processes and to train the collaborative optimization model using a multi-agent deep reinforcement learning algorithm. The scheduling decision module is used to input the node feature representation into the collaborative optimization model to generate task scheduling decisions; The task execution module is used to distribute tasks to corresponding edge nodes for execution according to the task scheduling decision, and update the parameters of the collaborative optimization model based on the execution results.

Images