Intelligent computing power integration service management method and system based on cloud edge-end cooperation

Abstract

The application discloses a cloud edge-end cooperation-based intelligent computing power integration service management method and system, relates to the computing power integration management technical field, and comprises the following steps: in a cloud controller, the computing power resource state of an edge-end cooperation port and a terminal equipment is sensed in real time, and a resource topological atlas is built; a dynamic task demand is introduced, and a computing power scheduling vector is generated; a cloud training model is subjected to light-weight segmentation, and a region elastic cluster composed of adjacent edge nodes under hierarchical limitation is determined; a digital twin simulation engine is deployed to preplay a computing power distribution scene, a heterogeneous resource pooling mechanism is established, and computing power integration service management is carried out. The application solves the technical problems of low heterogeneous computing power resource management efficiency and insufficient utilization in the prior art, achieves efficient management of intelligent computing power integration service, and improves the utilization of heterogeneous computing power resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computing power integration management, in particular to an intelligent computing power integration service management method and system based on cloud edge-end collaboration. BACKGROUND

[0002] With the rapid development of cloud computing, edge computing and terminal devices, computing power demand presents diversified and dynamic characteristics, and the integration management of heterogeneous computing power resources (such as GPU, CPU, FPGA, etc.) becomes the key to improving service efficiency. However, in the prior art, there are problems such as not timely computing power resource perception, lack of flexibility in scheduling strategy, insufficient load balancing, and poor energy consumption control, resulting in low utilization of heterogeneous computing power resources and high task processing delay, which is difficult to meet the dynamic task demand in complex scenarios.

[0003] The technical problem of low efficiency and insufficient utilization of heterogeneous computing power resources in the prior art. SUMMARY

[0004] The present application provides an intelligent computing power integration service management method and system based on cloud edge-end collaboration, which is used to solve the technical problem of low efficiency and insufficient utilization of heterogeneous computing power resources in the prior art.

[0005] In view of the above problems, the present application provides an intelligent computing power integration service management method and system based on cloud edge-end collaboration.

[0006] The first aspect of the present application provides an intelligent computing power integration service management method based on cloud edge-end collaboration, which comprises:

[0007] In the cloud controller, the computing power resource state of the edge-end collaboration port and the terminal device is perceived in real time, and a resource topology map is built. Based on the resource topology map, a dynamic task demand is introduced, and a computing power scheduling vector containing computing power load balancing weight, delay sensitivity index and energy consumption constraint condition is generated. According to the computing power scheduling vector, the cloud-end training model corresponding to the cloud-end controller is segmented in a lightweight manner, the edge node inference ability is dynamically matched, the adjacent edge nodes under the level limit are determined to form a regional elastic cluster, a digital twin simulation engine is deployed to pre-play the computing power allocation scene, the resource fragmentation rate is dynamically corrected combined with reinforcement learning, a heterogeneous resource pooling mechanism is established, the heterogeneous computing power resources in the regional elastic cluster are dynamically divided on demand, and computing power integration service management is performed.

[0008] The second aspect of the present application provides an intelligent computing power integration service management system based on cloud edge-end collaboration, which comprises:

[0009] The resource topology map building module is configured to build a resource topology map by sensing the computing power resource state of the edge-end collaborative port and the terminal device in real time at the cloud controller.

[0010] The one or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0011] The resource topology map building module is configured to build a resource topology map by sensing the computing power resource state of the edge-end collaborative port and the terminal device in real time at the cloud controller. The one or more technical solutions provided in the present application have at least the following technical effects or advantages: BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0013] Figure 1 The flowchart of the intelligent computing power integrated service management method based on cloud-edge-end collaboration provided by the embodiments of the present application is shown.

[0014] Figure 2 The structural diagram of the intelligent computing power integrated service management system based on cloud-edge-end collaboration provided by the embodiments of the present application is shown.

[0015] The resource topology map building module 10, the dynamic task demand introduction module 20, the regional elastic cluster composition module 30, and the computing power integrated service management module 40. DETAILED DESCRIPTION

[0016] The present application provides an intelligent computing power integration service management method and system based on cloud edge-end collaboration, which is used to solve the technical problems of low management efficiency and insufficient utilization of heterogeneous computing power resources in the prior art.

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0018] Embodiment one, as shown in the present application provides an intelligent computing power integration service management method based on cloud edge-end collaboration, which comprises: Figure 1

[0019] Step S100: In the cloud controller, the computing power resource state of the edge-end collaborative port and the terminal device is sensed in real time, and a resource topology map is built.

[0020] Specifically, the cloud controller monitors the computing power resource state of the edge-end collaborative port and the terminal device in real time. These state information covers device load, network bandwidth, energy consumption data, etc., and is associated with related data in the user behavior portrait library and device topology data. The sensed resource information is integrated according to the device connection relationship, hierarchical structure and resource attribute, and a resource topology map that can clearly present the resource distribution, association relationship and potential interaction path between the edge-end device and the collaborative port is constructed, which provides comprehensive and real-time resource basic data support for subsequent steps of introducing dynamic task demand and generating computing power scheduling vector.

[0021] Step S200: Based on the resource topology map, introduce dynamic task demand, and generate a computing power scheduling vector containing computing power load balancing weight, time delay sensitivity index and energy consumption constraint condition.

[0022] ​Specifically, first, the load prediction weight is determined by connecting the user behavior portrait library, the partition prediction benchmark curve is formulated combined with the device topology data, and the dynamic task demand is introduced based on this; in the process of generating the computing power scheduling vector, the user power consumption behavior time series data and the device infrared thermal imaging data are integrated, the cross-dimensional features are extracted through the spatio-temporal attention mechanism, the multi-region load prediction model weight is aggregated based on the federated learning framework, the harmonic distortion rate compensation parameter is dynamically optimized combined with the cross-dimensional features, the dynamic reactive power compensation strategy is set, and the power grid voltage fluctuation trend is predicted combined with the LSTM, finally the computing power scheduling vector containing the computing power load balancing weight, the time delay sensitivity index and the energy consumption constraint condition is formed, and the deployed hardware acceleration type edge node processes the related parameters in parallel to ensure the response timeliness.

[0023] Step S300: According to the computing power scheduling vector, the cloud training model corresponding to the cloud controller is segmented, the edge node inference ability is dynamically matched, and the regional elastic cluster composed of adjacent edge nodes under the level limit is determined.

[0024] Specifically, according to the generated computing power scheduling vector, the user power consumption behavior time series data and the device infrared thermal imaging data are integrated, the cross-dimensional features are extracted through the spatio-temporal attention mechanism, the multi-region load prediction model weight is aggregated based on the federated learning framework, the harmonic distortion rate compensation parameter is dynamically optimized combined with the cross-dimensional features, and the cloud model parameter is compressed through the lightweight knowledge distillation architecture, so as to complete the lightweight segmentation of the cloud training model corresponding to the cloud controller; at the same time, the dynamic reactive power compensation strategy is set, the real-time compensation instruction is generated combined with the inductive load features and the capacitive load features, the power grid voltage fluctuation trend is predicted combined with the LSTM, and the hardware acceleration type edge node is deployed to process the harmonic suppression parameter in parallel to ensure that the dynamic compensation response time meets the timeliness limit, so as to dynamically match the edge node inference ability, and then determine the regional elastic cluster composed of adjacent edge nodes under the level limit.

[0025] Step S400: Deploy the digital twin simulation engine to pre-play the computing power distribution scene, dynamically correct the resource fragmentation rate combined with reinforcement learning, establish a heterogeneous resource pooling mechanism, dynamically cut the heterogeneous computing power resources in the regional elastic cluster on demand, and manage the computing power integration service.

[0026] Specifically, the deployment of the digital twin simulation engine preplays the computing power allocation scene, inserts the network bandwidth threshold, the upper limit of the delay tolerance, and the energy consumption increment limit under the load transfer path during the process, and performs multi-objective optimization solving with the load balancing degree, transfer cost, and service quality as the target; at the same time, the generated adversarial network simulates the sensor noise interference mode to dynamically optimize the load feature extraction threshold, associates the device fault historical data with the threshold to correct the confidence interval of the partition prediction reference curve through Monte Carlo simulation, compares the device operation data with the preset risk threshold, and if it exceeds, triggers an early warning and feeds back to the cloud controller, and on this basis, combines reinforcement learning to dynamically correct the resource fragmentation rate to establish a heterogeneous resource pooling mechanism, dynamically adjusts the allocation proportion of tasks at each edge node based on the device energy efficiency ratio and combined with the photovoltaic output and task processing progress, monitors the energy consumption after the use of the computing power of each node, dynamically splits the GPU, CPU, and FPGA computing power on demand through the fusion of hardware acceleration cards and containers, and quantitatively evaluates the energy efficiency of various computing powers according to the energy consumption data and optimizes the ratio in the regional elastic cluster, thereby dynamically splitting the heterogeneous computing power resources in the regional elastic cluster on demand, realizing the integrated service management of computing power.

[0027] In one possible implementation manner, step S300 further includes:

[0028] Step S310: integrate the user power consumption behavior time series data and the device infrared thermal imaging data, and extract cross-dimensional features through a spatiotemporal attention mechanism.

[0029] Step S320: aggregate the multi-region load prediction model weights based on a federated learning framework, dynamically optimize the harmonic distortion rate compensation parameters combined with the cross-dimensional features, and compress the cloud model parameters through a lightweight knowledge distillation architecture.

[0030] Specifically, the power consumption behavior time series data generated by the user at different time nodes (such as power consumption load fluctuation, power consumption time length change, and other data recorded in time series) and the state data such as temperature distribution and heat dissipation efficiency captured by the device through infrared thermal imaging technology are first integrated, and then a spatiotemporal attention mechanism is used to deeply process these data. This mechanism can focus on the dynamic change law of power consumption behavior in the time dimension and pay attention to the physical characteristics presented by the device thermal imaging in the space dimension. Through cross-dimensional correlation analysis and feature mining, key cross-dimensional features that can comprehensively reflect the user's power consumption habits and the device's running state are extracted, providing basic data support for the subsequent lightweight processing and parameter optimization of the cloud training model.

[0031] Based on the federated learning framework, the federated average (FedAvg) algorithm is used to aggregate the multi-region load prediction model weights. After each regional node trains the model locally using the gradient descent algorithm, only the model parameters are uploaded to the cloud. The cloud obtains the global model by weighted averaging the parameters through FedAvg. When combining cross-dimension feature dynamic optimization harmonic distortion rate compensation parameters, the random forest algorithm is introduced to mine the mapping relationship between the features and the compensation parameters. Through multiple iterations, the parameters are adjusted to minimize the harmonic distortion rate. In the process of compressing the cloud model parameters, the knowledge distillation algorithm is used, with the cloud complex model as the teacher model and the edge lightweight model as the student model. Through the KL divergence loss function, the output distribution of the two is aligned. At the same time, the L1 regularization algorithm is used to sparsify the student model parameters, realizing the effective compression of the model parameters.

[0032] In one possible implementation manner, step S300 further includes:

[0033] Step S330: Set a dynamic reactive power compensation strategy to generate real-time compensation instructions based on inductive load characteristics and capacitive load characteristics, and combine LSTM to predict power grid voltage fluctuation trends.

[0034] Step S340: At the same time, deploy hardware-accelerated edge nodes to process harmonic suppression parameters in parallel, so that the dynamic compensation response time meets the response time limit.

[0035] Specifically, by setting a dynamic reactive power compensation strategy, the inductive load characteristics (such as inductance parameters, power factor lag conditions, etc.) and capacitive load characteristics (such as capacitance parameters, power factor lead conditions, etc.) in the circuit are collected and analyzed in real time. Real-time compensation instructions corresponding to these characteristic data are generated to achieve dynamic balance of the reactive power of the power grid. At the same time, the LSTM (Long Short-Term Memory) model is introduced to take advantage of its processing of time series data. The model is trained in combination with historical power grid voltage data to predict the fluctuation trend of the power grid voltage, so that the reactive power compensation strategy can respond to voltage changes in advance, improving the accuracy and foresight of compensation.

[0036] At the same time of implementing the dynamic reactive power compensation strategy, edge nodes equipped with hardware acceleration components such as FPGA are deployed. Such nodes can perform efficient parallel calculation and real-time adjustment on harmonic suppression related parameters (such as harmonic frequency, amplitude and phase compensation amount, etc.) by virtue of the parallel processing capability at the hardware level. Through hardware acceleration technology, the operation period of parameter processing is shortened, ensuring that the dynamic compensation response time from receiving compensation demand to generating specific suppression parameters is controlled within the preset time limit, guaranteeing the timeliness of harmonic suppression action, and forming synergy with the dynamic reactive power compensation strategy to improve the overall efficiency of power grid harmonic control.

[0037] In one possible implementation manner, step S400 further includes:

[0038] Step S410: The connection user behavior portrait library determines the load prediction weight, and combines the device topology data to determine the partition prediction reference curve.

[0039] Step S420: Based on the partition prediction reference curve, a blockchain storage unit is integrated, and the blockchain storage unit is used to record the prediction model parameter update log.

[0040] Specifically, the pre-constructed user behavior portrait library is connected through the interface, the library stores user historical power consumption time, power peak period, device start-stop law and other behavior data, the weighted regression algorithm is used to evaluate the feature importance of these data, and the contribution of different user behavior characteristics in load prediction, i.e. the load prediction weight, is determined; at the same time, the device topology data (including device physical connection relationship, hierarchical distribution and regional attribution information) is called, the K-means clustering algorithm is used to partition the devices and covered areas according to the power consumption similarity, and the historical load average, fluctuation range and trend characteristics of each partition are combined to generate the load change reference line of each partition, i.e. the partition prediction reference curve, by time series smoothing method.

[0041] Based on the partition prediction reference curve, a blockchain storage unit is integrated by deploying a consortium chain architecture, the unit uses a chain data structure and a hash encryption algorithm to generate blocks for the key information of the prediction model parameter update (including update timestamp, specific values before and after parameter change, operation node identifier, etc.), and stores them in sequence according to time. At the same time, the trigger conditions and recording rules of the smart contract preset parameter update are used to ensure that after each model parameter adjustment, the relevant log information can be automatically chained and stored, realizing the non-tamperability and full traceability of the parameter update process, and providing reliable data record support for the model parameter management of the partition load prediction.

[0042] In one possible implementation manner, step S400 further includes:

[0043] Step S430: Based on the computing power allocation scenario, the network bandwidth threshold, the upper limit of delay tolerance, and the energy consumption increment limit under the load transfer path are inserted.

[0044] Step S440: Taking load balancing degree, transfer cost and service quality as the target, multi-objective optimization is solved.

[0045] Specifically, based on the computing power allocation scenario pre-played by the digital twin simulation engine, three types of key limiting parameters are accurately inserted for the load transfer path: network bandwidth threshold, i.e. the maximum bandwidth value allowed for data transmission on this path, to ensure that there is no congestion during data transmission; upper limit of delay tolerance, to clearly define the maximum delay time acceptable for the load during transfer, to ensure the timeliness of service response; and energy consumption increment limit, to set the maximum energy consumption increment allowed for load transfer compared to the original state, taking into account energy saving needs. The insertion of these parameters establishes clear constraint boundaries for the computing power allocation scenario, providing a basic condition for subsequent multi-objective optimization solution.

[0046] In the multi-objective optimization solution, NSGA-II (Non-dominated Sorting Genetic Algorithm II) is used, with load balance, transfer cost, and service quality as optimization objectives, and network bandwidth threshold, delay tolerance upper limit, and energy consumption increment limit as constraint conditions to construct the fitness function. The algorithm generates multiple possible computing power allocation schemes through initialization of the population, calculates the specific values of the three objectives for each scheme, divides the schemes into different levels using non-dominated sorting, and maintains population diversity through crowding distance. After iteration and optimization through genetic operations such as selection, crossover, and mutation, the Pareto optimal solution set is obtained, with each solution corresponding to a set of computing power allocation schemes that balance the three objectives, ensuring that load balancing, cost control, and service quality meet the requirements under the premise of meeting the constraints.

[0047] In one possible implementation, step S400 further includes:

[0048] Step S450: Simulate sensor noise interference patterns based on a generative adversarial network to dynamically optimize the load feature extraction threshold.

[0049] Step S460: Use a digital twin simulation engine to associate device fault historical data with the load feature extraction threshold for risk assessment.

[0050] Specifically, a noise simulation model is constructed based on a generative adversarial network, which simulates various noise interference patterns that may occur under different operating conditions of the sensor through a generator, while a discriminator is used to identify the authenticity of the generated noise. After multiple rounds of adversarial training, the accuracy of noise simulation is improved; the simulated noise interference patterns are combined with the actual collected load data to simulate the load data input scenario in a complex noise environment, and on this basis, the threshold for load feature extraction is dynamically adjusted to ensure that key load features can still be accurately captured in the presence of noise interference, thereby optimizing the effectiveness and anti-interference ability of load feature extraction.

[0051] A digital twin simulation engine is built, a virtual mapping model consistent with the physical device and system is constructed, a device historical fault database (containing fault type, occurrence time, associated operating parameters, and impact degree data) and a load feature extraction threshold value optimized by a generative adversarial network are imported; the engine simulates the operating state of the device under different threshold value settings, establishes an association model of fault occurrence probability and feature extraction threshold value, quantitatively analyzes the influence degree of threshold value variation on fault risk, and outputs a risk assessment report to provide data support for subsequent risk early warning and threshold adjustment.

[0052] In a possible implementation manner, step S460 further includes:

[0053] Step S461: Based on Monte Carlo simulation, dynamically correcting the confidence interval of the partition prediction reference curve.

[0054] Step S462: Comparing the device operating data with the preset risk threshold value through the corrected confidence interval, triggering the early warning mechanism if it exceeds, and feeding back the risk assessment result to the cloud controller.

[0055] Specifically, the Monte Carlo simulation method is used, a large number of simulated load data are generated by setting random variables related to load (such as user power fluctuation, device operating state change, etc.) and repeating sampling based on the probability distribution of these variables. Comparing these simulation data with the original partition prediction reference curve, calculating the deviation between the curve prediction value and the simulation value, and then dynamically adjusting the upper and lower limits of the confidence interval of the reference curve, so that the corrected confidence interval can more accurately reflect the fluctuation range that the actual load may appear, and improve the reliability and adaptability of the interval.

[0056] The confidence interval of the partition prediction reference curve corrected by Monte Carlo simulation is used as the reference range, the operating data of the device (such as real-time load value, device temperature, energy consumption index, etc.) are collected in real time, and these data are compared with the preset risk threshold value one by one. If the device operating data exceeds the boundary range of the corrected confidence interval and exceeds the preset risk threshold value, the early warning mechanism is triggered immediately, and warnings are issued through sound and light alarms, information pushing, etc.; at the same time, the detailed results of this risk assessment (including threshold value exceeding data, risk level, possible impact range, etc.) are fed back to the cloud controller in real time, so that the cloud can further schedule computing power or adjust strategies according to the feedback information.

[0057] In a possible implementation manner, step S400 further includes:

[0058] Step S470: Based on the device energy efficiency ratio, combining the photovoltaic output and the task processing progress, dynamically adjusting the allocation proportion of the task in each edge node.

[0059] Step S480: At the same time, monitor the energy consumption of each edge node after the use of computing power, and update the heterogeneous resource pooling mechanism.

[0060] Specifically, a weighted greedy algorithm is used to set weight coefficients for the device energy efficiency ratio, photovoltaic output, and task processing progress (e.g., energy efficiency ratio weight 0.5, photovoltaic output weight 0.3, and task progress weight 0.2), and then normalize the three indicators of each edge node (mapping to the 0-1 interval). The overall score of each node is calculated by the weighted formula (node score = energy efficiency ratio x 0.5 + photovoltaic output x 0.3 + task progress adaptation degree x 0.2), and the task proportion is dynamically allocated according to the score ratio, i.e., the task allocation proportion of a certain node = the score of the node ÷ the total score of all nodes, so that the task is inclined to the node with high energy efficiency, stable photovoltaic output, and suitable task progress.

[0061] Energy consumption monitoring sensors and data acquisition modules are deployed on each edge node to record real-time data such as power consumption and energy consumption fluctuations during the processing of computing power tasks, and transmit the data to the cloud controller through the edge-end collaboration port; the cloud controller analyzes the received energy consumption data, calculates energy efficiency indicators such as unit computing power energy consumption based on the node's computing power type (e.g., GPU, CPU, FPGA) and task processing volume; based on these indicators, hardware acceleration cards and containerization fusion technology support dynamic division of various types of computing power, and the ratio of different types of computing power in the regional elastic cluster is adjusted based on the energy efficiency quantitative evaluation results, thereby updating the heterogeneous resource pooling mechanism to ensure that resource allocation better meets energy efficiency requirements.

[0062] In one possible implementation, step S480 further includes:

[0063] Step S481: Through hardware acceleration cards and containerization fusion, GPU computing power, CPU computing power, and FPGA computing power are dynamically divided as needed.

[0064] Step S482: At the same time, based on energy consumption data, energy efficiency quantitative evaluation of GPU computing power, CPU computing power, and FPGA computing power is performed to optimize the ratio in the regional elastic cluster.

[0065] Specifically, hardware acceleration cards and containerization technology are fused to build a unified computing power scheduling framework. Hardware acceleration cards provide efficient hardware support for different types of computing power such as GPU, CPU, and FPGA, while containerization technology realizes the virtualization, encapsulation, and isolation of these computing power resources, allowing various types of computing power to be flexibly divided into multiple independent computing units. Through this fusion method, the system can dynamically divide GPU computing power, CPU computing power, and FPGA computing power in real time and accurately according to the actual task requirements for different types of computing power, meet the flexible requirements of various tasks in terms of computing power type and scale, and improve the utilization efficiency of heterogeneous computing power resources.

[0066] The energy efficiency quantitative evaluation model is constructed, the energy consumption data (such as power consumption, energy consumption fluctuation, etc.) of GPU, CPU and FPGA in the process of processing tasks is collected in real time, and the unit energy consumption computing power value and other quantitative indicators are calculated in combination with the computing power output data; based on these indicators, the energy efficiency of the three computing powers is scored and sorted, and then combined with the task type and computing power demand in the regional elastic cluster, the intelligent scheduling algorithm is used to dynamically adjust the proportion of the three in the cluster, and the high energy consumption task is preferentially allocated to the computing power type with high energy efficiency score, and the load proportion of the computing power with low energy efficiency is reduced, so as to realize the optimization of the proportion of heterogeneous computing power in the cluster and improve the overall energy efficiency.

[0067] Embodiment two, based on the same inventive concept as the intelligent computing power integrated service management method based on cloud edge end cooperation in the foregoing embodiments, as shown in Figure 2 The application provides an intelligent computing power integrated service management system based on cloud edge end cooperation, and the system and method embodiments in the application are based on the same inventive concept. The system comprises:

[0068] The resource topology graph building module 10 is used for real-time sensing of the computing power resource state of the edge end cooperation port and the terminal device in the cloud controller, and building a resource topology graph.

[0069] The dynamic task demand introduction module 20 is used for introducing dynamic task demand based on the resource topology graph, and generating a computing power scheduling vector containing computing power load balancing weight, time delay sensitivity index and energy consumption constraint condition.

[0070] The regional elastic cluster composition module 30 is used for performing lightweight segmentation on the cloud end training model corresponding to the cloud end controller according to the computing power scheduling vector, dynamically matching the edge node inference ability, and determining the regional elastic cluster composed of adjacent edge nodes under the level limit.

[0071] The computing power integrated service management module 40 is used for deploying a digital twin simulation engine to pre-play a computing power allocation scene, dynamically correcting the resource fragmentation rate in combination with reinforcement learning, establishing a heterogeneous resource pooling mechanism, dynamically cutting the heterogeneous computing power resources in the regional elastic cluster on demand, and performing computing power integrated service management.

[0072] Further, the system is also used to realize the following functions:

[0073] The integrated user power consumption behavior time sequence data and device infrared thermal imaging data are extracted through a space-time attention mechanism; based on a federated learning framework, the multi-region load prediction model weight is aggregated, the harmonic distortion rate compensation parameter is dynamically optimized in combination with the cross-dimension feature, and the cloud model parameter is compressed through a lightweight knowledge distillation architecture.

[0074] Further, the system is also used to realize the following functions:

[0075] The dynamic reactive power compensation strategy is set to generate real-time compensation instructions according to the inductive load characteristics and the capacitive load characteristics, and the LSTM is combined to predict the power grid voltage fluctuation trend; at the same time, the hardware acceleration type edge node is deployed, and the harmonic suppression parameters are processed in parallel, so that the dynamic compensation response time meets the response time limit.

[0076] Further, the system is also used to realize the following functions:

[0077] The connection user behavior portrait library determines the load prediction weight, and the device topology data is combined to determine the partition prediction reference curve; based on the partition prediction reference curve, a blockchain storage unit is integrated, and the blockchain storage unit is used to record the prediction model parameter update log.

[0078] Further, the system is also used to realize the following functions:

[0079] Based on the computing power allocation scene, the network bandwidth threshold, the upper limit of delay tolerance, and the energy consumption increment limit under the load transfer path are inserted; the load balancing degree, the transfer cost, and the service quality are taken as the target to solve the multi-objective optimization.

[0080] Further, the system is also used to realize the following functions:

[0081] Based on the generative adversarial network, the sensor noise interference mode is simulated, and the load feature extraction threshold is dynamically optimized; the digital twin simulation engine is used to associate the device fault historical data with the load feature extraction threshold for risk assessment.

[0082] Further, the system is also used to realize the following functions:

[0083] Based on the Monte Carlo simulation, the confidence interval of the partition prediction reference curve is dynamically corrected; through the corrected confidence interval, the device operation data and the preset risk threshold are compared, if it is exceeded, the early warning mechanism is triggered, and the risk assessment result is fed back to the cloud controller.

[0084] Further, the system is also used to realize the following functions:

[0085] Based on the device energy efficiency ratio, the photovoltaic output and the task processing progress are combined to dynamically adjust the allocation proportion of the task in each edge node; at the same time, the energy consumption of each edge node after the use of computing power is monitored, and the heterogeneous resource pooling mechanism is updated.

[0086] Further, the system is also used to realize the following functions:

[0087] By fusing hardware acceleration cards with containerization, GPU computing power, CPU computing power and FPGA computing power are supported to be dynamically divided on demand; meanwhile, according to energy consumption data, energy efficiency of the GPU computing power, the CPU computing power and the FPGA computing power is quantitatively evaluated, and the GPU computing power, the CPU computing power and the FPGA computing power are proportioned and optimized in the regional elastic cluster.

[0088] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned describes a specific embodiment of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0089] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0090] The specification and drawings are merely exemplary of the present application, and any and all modifications, variations, combinations or equivalents that fall within the scope of the present application should be covered. Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technology, the present application is intended to include these modifications and variations.

Claims

1. A method for intelligent computing power integration service management based on cloud edge-end collaboration, characterized in that, The method comprises: In the cloud controller, the computing power resource state of the edge-end collaborative port and the terminal equipment is perceived in real time, and a resource topology map is built; Based on the resource topology map, a dynamic task demand is introduced, and a computing power scheduling vector containing a computing power load balancing weight, a time delay sensitivity index and an energy consumption constraint condition is generated; According to the computing power scheduling vector, the cloud training model corresponding to the cloud controller is subjected to lightweight segmentation, the edge node inference ability is dynamically matched, and a region elastic cluster composed of adjacent edge nodes under the level limitation is determined; A digital twin simulation engine is deployed to pre-play a computing power allocation scene, a resource fragmentation rate is dynamically corrected combined with reinforcement learning, a heterogeneous resource pooling mechanism is established, and the heterogeneous computing power resources in the region elastic cluster are dynamically cut according to needs for computing power integrated service management; The cloud training model corresponding to the cloud controller is subjected to lightweight segmentation, and the method further comprises: Integrated user power consumption behavior time series data and equipment infrared thermal imaging data, cross-dimensional features are extracted through a space-time attention mechanism; Based on a federated learning framework, the weights of multi-region load prediction models are aggregated, the harmonic distortion rate compensation parameters are dynamically optimized combined with the cross-dimensional features, and the cloud model parameters are compressed through a lightweight knowledge distillation architecture; The method further comprises: Connecting a user behavior portrait library to determine a load prediction weight, and combining equipment topology data to determine a partition prediction benchmark curve; Based on the partition prediction benchmark curve, a blockchain storage unit is integrated, which is used to record a prediction model parameter update log; A digital twin simulation engine is deployed to pre-play a computing power allocation scene, a resource fragmentation rate is dynamically corrected combined with reinforcement learning, and a heterogeneous resource pooling mechanism is established, the method comprising: Based on the computing power allocation scene, a network bandwidth threshold, a time delay tolerance upper limit and an energy consumption increment limit under a load transfer path are inserted; Taking load balancing degree, transfer cost and service quality as targets, multi-objective optimization is solved; The method further comprises: Based on a generative adversarial network, a sensor noise interference mode is simulated, and a load feature extraction threshold is dynamically optimized; Using a digital twin simulation engine, risk assessment is performed by associating equipment failure historical data with the load feature extraction threshold; Risk assessment is performed by associating equipment failure historical data with the load feature extraction threshold, the method comprising: Based on Monte Carlo simulation, the confidence interval of the partition prediction benchmark curve is dynamically corrected; Through the corrected confidence interval, device operation data and a preset risk threshold are compared, and if the comparison result exceeds the preset risk threshold, a warning mechanism is triggered, and the risk assessment result is fed back to the cloud controller; A heterogeneous resource pooling mechanism is established to dynamically cut the heterogeneous computing power resources in the region elastic cluster according to needs for computing power integrated service management, and the method further comprises: Based on the equipment energy efficiency ratio, the allocation proportion of tasks in each edge node is dynamically adjusted combined with photovoltaic output and task processing progress; At the same time, the energy consumption of each edge node after use of computing power is monitored, and the heterogeneous resource pooling mechanism is updated.

2. The cloud edge-end collaboration based intelligent computing power integration service management method of claim 1, wherein The method further comprises: A dynamic reactive power compensation strategy is set to generate real-time compensation instructions according to the characteristics of inductive and capacitive loads, and the trend of power grid voltage fluctuation is predicted in combination with LSTM; At the same time, deploy hardware-accelerated edge nodes to process harmonic suppression parameters in parallel, so that the response time of dynamic compensation meets the response time limit. 3.The cloud edge-end collaboration based intelligent computing power integration service management method of claim 1, wherein, Monitor the energy consumption of each edge node after the use of computing power, and update the heterogeneous resource pooling mechanism, the method comprising: Through the fusion of hardware acceleration cards and containers, GPU computing power, CPU computing power and FPGA computing power can be dynamically divided on demand; At the same time, according to the energy consumption data, the energy efficiency of GPU computing power, CPU computing power and FPGA computing power is quantitatively evaluated, and the ratio is optimized in the regional elastic cluster.

4. The intelligent computing power integration service management system based on cloud edge-end collaboration, characterized in that, The system is used to implement the intelligent computing power integrated service management method based on cloud edge-end collaboration according to any one of claims 1-3, and the system comprises: A resource topology map building module is used to build a resource topology map by real-time sensing of the computing power resource state of the edge-end collaboration port and the terminal device in the cloud controller; A dynamic task demand introduction module is used to introduce dynamic task demand based on the resource topology map, and generate a computing power scheduling vector containing computing power load balancing weight, time delay sensitivity index and energy consumption constraint condition; A regional elastic cluster composition module is used to perform lightweight segmentation of the cloud training model corresponding to the cloud controller according to the computing power scheduling vector, dynamically match the edge node inference ability, and determine the adjacent edge node composition regional elastic cluster under the level limit; A computing power integrated service management module is used to deploy a digital twin simulation engine to pre-play the computing power allocation scene, dynamically correct the resource fragmentation rate in combination with reinforcement learning, establish a heterogeneous resource pooling mechanism, dynamically divide the heterogeneous computing power resources in the regional elastic cluster on demand, and manage the computing power integrated service.

Images