Computing power scheduling method and device, equipment and readable storage medium

Abstract

The application discloses a computing power scheduling method and device, equipment and a readable storage medium. The method comprises the following steps: acquiring target data corresponding to a scene type based on a current scene type; the target data comprises at least one of the following: node state data, task characteristic data and environment parameter data; based on the target data, a three-dimensional target function comprising an energy efficiency index, a time delay index and a reliability index and a dynamic weight model associated with a task priority and a node real-time state are constructed; based on an improved sea squirt colony algorithm, the three-dimensional target function and the dynamic weight model, a scheduling scheme of a computing power task is generated; through the fusion of the space-time characteristics and node heterogeneity of the smart city edge environment, a closed-loop system of "data acquisition-multi-target modeling-self-adaptive optimization-real-time scheduling" is constructed.

Description

Technical Field

[0001] This application belongs to the field of cloud computing, big data, and edge computing technology, and specifically relates to a computing power scheduling method, device, equipment, and readable storage medium. Background Technology

[0002] The smart city edge computing architecture adopts a three-level collaborative architecture of "terminal-edge-cloud". The edge layer includes heterogeneous devices such as base station edge nodes, IoT gateways, and smart pole servers, which are responsible for undertaking real-time tasks such as traffic monitoring, environmental perception, and smart government affairs. Its core features are resource distribution, task heterogeneity, and dynamic demand. This is the basic scenario architecture for edge computing optimization.

[0003] In the multi-objective optimization framework for edge computing power, the known optimization objectives include three core metrics: performance metrics (task response latency, computing power utilization), energy efficiency metrics (energy consumption per unit of computing power, node endurance), and reliability metrics (task completion rate, fault recovery capability). Multi-objective optimization is typically solved using Pareto solutions. Dynamic Voltage and Frequency Scaling (DVFS), task offloading strategies, and elastic resource allocation are well-known methods for edge energy efficiency optimization. Among them, DVFS achieves a balance between energy consumption and performance by adjusting the processor frequency, and task offloading reduces local resource pressure through edge-cloud task allocation. Both require dynamic decision-making in conjunction with optimization algorithms.

[0004] Current technologies, when addressing the dynamic and multi-objective edge computing power optimization problem in smart cities, either lose their adaptability to environmental dynamism and the integrity of the solution set due to the use of static weighting methods, or suffer from bottlenecks in search efficiency and dynamic adaptability due to the rigid internal mechanisms of the general multi-objective algorithms they employ. These technologies are unable to meet the stringent requirements of future smart cities for efficient, intelligent, and agile scheduling of computing resources. Summary of the Invention

[0005] This application provides a computing power scheduling method, apparatus, device, and readable storage medium to solve the problem that current computing power scheduling methods cannot meet the needs of future smart cities for efficient, intelligent, and agile scheduling of computing resources.

[0006] Firstly, a computing power scheduling method is provided, the method comprising:

[0007] Based on the current scenario type, obtain target data corresponding to the scenario type; the target data includes at least one of the following: node status data, task feature data, and environmental parameter data;

[0008] Based on the target data, a three-dimensional objective function containing energy efficiency indicators, latency indicators, and reliability indicators, as well as a dynamic weight model associated with task priority and real-time node status, are constructed.

[0009] Based on the improved tunicate swarm algorithm, the three-dimensional objective function, and the dynamic weight model, a scheduling scheme for computing power tasks is generated.

[0010] The method further includes:

[0011] Based on the scheduling scheme and the scheduling rules that match the scenario, the computing power tasks are allocated to corresponding nodes; the nodes include at least one of: terminals, edge control devices, and cloud management devices;

[0012] Based on Dynamic Voltage and Frequency Regulation (DVFS), the frequency and / or voltage of each node are adjusted.

[0013] The method further includes:

[0014] Based on task attribute data and node status data, the current scenario type is determined; wherein, the task attribute data includes at least one of the following: task data volume, task computational intensity, task latency threshold, task priority, and task reliability requirements; the node status data includes at least one of the following: node geographic coordinates, node load rate, node energy consumption, and node failure rate.

[0015] The determination of the current scenario type based on task attribute data and node status data includes:

[0016] The task attribute data and the node status data are normalized to obtain normalized data;

[0017] The normalized data is clustered using a density clustering algorithm, and the mean value of features is calculated for each cluster.

[0018] The average value of the features is matched with a preset scene template to determine the current scene type; the scene type includes at least one of the following: traffic monitoring scene, medical emergency scene, and environmental monitoring scene.

[0019] The step of obtaining target data corresponding to the current scene type includes:

[0020] Based on the acquisition strategy corresponding to the scene type, multi-source data corresponding to the scene type is collected;

[0021] The multi-source data is subjected to noise reduction processing and weighted fusion to generate the target data.

[0022] The method further includes:

[0023] The target data is subjected to a consistency check. If the consistency check fails, multi-source data corresponding to the scenario type is re-collected based on the collection strategy corresponding to the scenario type.

[0024] The step of constructing a dynamic weight model based on the target data and relating it to task priority and real-time node status includes:

[0025] Based on the relevance between the scene type and the target, and the scene type itself, a scene baseline weight is generated;

[0026] The scenario baseline weights are adjusted based on the task priority and the node load to construct the dynamic weight model.

[0027] The improved tunic swarm algorithm includes a tunic swarm algorithm that integrates Lévy flight, chaotic mapping, and gradient guidance mechanisms.

[0028] The algorithm for generating computing power tasks, based on the improved tunic swarm algorithm, the three-dimensional objective function, and the dynamic weight model, includes:

[0029] Based on the improved tunicate swarm algorithm, the exploration factor and population structure are dynamically adjusted;

[0030] The three-dimensional objective function and the dynamic weight model are fused to construct a multi-objective fitness function;

[0031] Based on the adjusted exploration factor, population structure, and the multi-objective fitness function, the multi-objective optimization problem is solved.

[0032] Based on the solution to the multi-objective optimization problem, a scheduling scheme for computing power tasks is generated.

[0033] The step of generating a scheduling scheme for computing power tasks based on the solution to the multi-objective optimization problem includes:

[0034] The solutions to the multi-objective optimization problem are filtered by scenario-based hard constraints, and the filtered solutions are sorted by a preference function.

[0035] The sorted solutions are analyzed using the entropy weight method and the two-base-point method to generate a scheduling scheme for the computing power task.

[0036] The method further includes:

[0037] Collect task execution results;

[0038] Based on the task execution results, multi-dimensional feedback indicators are determined; the multi-dimensional feedback indicators include at least one of the following: energy efficiency feedback indicators, latency feedback indicators, and reliability feedback indicators.

[0039] Based on the multi-dimensional feedback indicators, at least one of the parameters of the improved tunic swarm algorithm, the parameters of the three-dimensional objective function, and the parameters of the dynamic weight model is adaptively adjusted.

[0040] Secondly, a computing power scheduling device is also provided, including a processor and a transceiver, wherein the transceiver receives and sends data under the control of the processor, and the processor is used to perform the following operations:

[0041] Based on the current scenario type, obtain target data corresponding to the scenario type; the target data includes at least one of the following: node status data, task feature data, and environmental parameter data;

[0042] Based on the target data, a three-dimensional objective function containing energy efficiency indicators, latency indicators, and reliability indicators, as well as a dynamic weight model associated with task priority and real-time node status, are constructed.

[0043] Based on the improved tunicate swarm algorithm, the three-dimensional objective function, and the dynamic weight model, a scheduling scheme for computing power tasks is generated.

[0044] The processor is also used to perform the following operations:

[0045] Based on the scheduling scheme and the scheduling rules that match the scenario, the computing power tasks are allocated to corresponding nodes; the nodes include at least one of: terminals, edge control devices, and cloud management devices;

[0046] Based on Dynamic Voltage and Frequency Regulation (DVFS), the frequency and / or voltage of each node are adjusted.

[0047] The processor is also used to perform the following operations:

[0048] Based on task attribute data and node status data, the current scenario type is determined; wherein, the task attribute data includes at least one of the following: task data volume, task computational intensity, task latency threshold, task priority, and task reliability requirements; the node status data includes at least one of the following: node geographic coordinates, node load rate, node energy consumption, and node failure rate.

[0049] The processor is also used to perform the following operations:

[0050] The task attribute data and the node status data are normalized to obtain normalized data;

[0051] The normalized data is clustered using a density clustering algorithm, and the mean value of features is calculated for each cluster.

[0052] The average value of the features is matched with a preset scene template to determine the current scene type; the scene type includes at least one of the following: traffic monitoring scene, medical emergency scene, and environmental monitoring scene.

[0053] The processor is also used to perform the following operations:

[0054] Based on the acquisition strategy corresponding to the scene type, multi-source data corresponding to the scene type is collected;

[0055] The multi-source data is subjected to noise reduction processing and weighted fusion to generate the target data.

[0056] The processor is also used to perform the following operations:

[0057] The target data is subjected to a consistency check. If the consistency check fails, multi-source data corresponding to the scenario type is re-collected based on the collection strategy corresponding to the scenario type.

[0058] The processor is also used to perform the following operations:

[0059] Based on the relevance between the scene type and the target, and the scene type itself, a scene baseline weight is generated;

[0060] The scenario baseline weights are adjusted based on the task priority and the node load to construct the dynamic weight model.

[0061] The improved tunic swarm algorithm includes a tunic swarm algorithm that integrates Lévy flight, chaotic mapping, and gradient guidance mechanisms.

[0062] The processor is also used to perform the following operations:

[0063] Based on the improved tunicate swarm algorithm, the exploration factor and population structure are dynamically adjusted;

[0064] The three-dimensional objective function and the dynamic weight model are fused to construct a multi-objective fitness function;

[0065] Based on the adjusted exploration factor, population structure, and the multi-objective fitness function, the multi-objective optimization problem is solved.

[0066] Based on the solution to the multi-objective optimization problem, a scheduling scheme for computing power tasks is generated.

[0067] The processor is also used to perform the following operations:

[0068] The solutions to the multi-objective optimization problem are filtered by scenario-based hard constraints, and the filtered solutions are sorted by a preference function.

[0069] The sorted solutions are analyzed using the entropy weight method and the two-base-point method to generate a scheduling scheme for the computing power task.

[0070] The processor is also used to perform the following operations:

[0071] Collect task execution results;

[0072] Based on the task execution results, multi-dimensional feedback indicators are determined; the multi-dimensional feedback indicators include at least one of the following: energy efficiency feedback indicators, latency feedback indicators, and reliability feedback indicators.

[0073] Based on the multi-dimensional feedback indicators, at least one of the parameters of the improved tunic swarm algorithm, the parameters of the three-dimensional objective function, and the parameters of the dynamic weight model is adaptively adjusted.

[0074] Thirdly, a computing power scheduling device is also provided, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the computing power scheduling method as described above.

[0075] Fourthly, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps in the computing power scheduling method described above.

[0076] Fifthly, a computer program product is also provided, including computer instructions that, when executed by a processor, implement the steps of the computing power scheduling method described above.

[0077] The above-mentioned technical solution of this application has at least the following beneficial effects:

[0078] In the computing power scheduling method, apparatus, device, and readable storage medium of this application embodiment, a closed-loop system of "data acquisition - multi-objective modeling - adaptive optimization - real-time scheduling" is constructed by integrating the spatiotemporal characteristics and node heterogeneity of the smart city edge environment. Specifically, a dynamic weight model is constructed based on task priority and real-time node status to realize the dynamic adjustment of target weights and solve the service quality (QoS) degradation problem caused by static weighting. An improved tunic swarm algorithm is introduced to dynamically adjust the exploration factor c1 and the population structure to improve the global optimization capability in high-dimensional scenarios. Attached Figure Description

[0079] Figure 1 This is a flowchart illustrating the steps of the computing power scheduling method provided in the embodiments of this application;

[0080] Figure 2 This diagram illustrates an application example of the computing power scheduling method provided in the embodiments of this application.

[0081] Figure 3 This is a schematic diagram showing the structure of the computing power scheduling device provided in the embodiments of this application. Detailed Implementation

[0082] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0083] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and are not used to describe a specified order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, a first object can be one or more. Furthermore, in the specification and claims, "and" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0084] It is worth noting that the technologies described in this application are not limited to Long Term Evolution (LTE) / LTE-Advanced (LTE-A) systems, but can also be used in other wireless communication systems, such as Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Frequency Division Multiple Access (FDMA), Orthogonal Frequency Division Multiple Access (OFDMA), Single-carrier Frequency-Division Multiple Access (SC-FDMA), and other systems. The terms "system" and "network" in this application are often used interchangeably, and the described technologies can be used with the systems and radio technologies mentioned above, as well as with other systems and radio technologies. However, the following description describes New Radio (NR) systems for illustrative purposes, and NR terminology is used in most of the following description. These technologies can also be applied to applications beyond NR systems, such as 6th generation (6G) radio systems. th Generation 6G communication system.

[0085] The prior art closely related to this application mainly falls into the following three categories:

[0086] Smart City Edge Computing Architecture: The smart city edge computing system adopts a three-level collaborative architecture of "terminal-edge-cloud". The edge layer includes heterogeneous devices such as base station edge nodes, IoT gateways, and smart pole servers, which are responsible for undertaking real-time tasks such as traffic monitoring, environmental perception, and smart government affairs. Its core features are resource distribution, task heterogeneity, and dynamic demand. This is the basic scenario architecture for edge computing optimization.

[0087] The core mechanism of the Salp Swarm Algorithm (SSA) is as follows: SSA is a metaheuristic algorithm that simulates the chain-like motion of salp sea squirts. It achieves optimization by having a leader guide the search direction and followers update their positions in a chain-like manner. The basic formula includes the leader's position update. And follower position update The core parameters are population size and exploration factor c1, and it has the well-known characteristics of simple structure and strong global search capability.

[0088] Multi-objective optimization framework for edge computing power: The well-known optimization objectives for edge computing power include three types of core indicators: performance indicators (task response latency, computing power utilization), energy efficiency indicators (energy consumption per unit of computing power, node endurance), and reliability indicators (task completion rate, fault recovery capability). Multi-objective optimization is usually solved by Pareto solution set.

[0089] Fundamental technologies for energy efficiency optimization include Dynamic Voltage and Frequency Scaling (DVFS), task offloading strategies, and elastic resource allocation. These are well-known methods for edge energy efficiency optimization. DVFS balances energy consumption and performance by adjusting processor frequency, while task offloading reduces local resource pressure through edge-cloud task allocation. Both require dynamic decision-making through optimization algorithms. For example, DVFS, a hardware technology that balances energy consumption and performance by adjusting processor voltage and frequency, typically follows Dennard's scaling law (power consumption ∝ voltage² × frequency).

[0090] In multi-objective optimization, the Pareto optimal set is a set of solutions in which no solution is superior to any solution in all objectives. Therefore, the Pareto optimal set can be understood as the set of optimal solutions to a multi-objective problem.

[0091] The edge computing three-layer architecture is a three-level collaborative architecture of terminal-edge-cloud. The edge layer undertakes real-time task processing and reduces cloud load and transmission latency.

[0092] Existing technologies, when dealing with the dynamic and complex scenarios of smart cities, suffer from technical bottlenecks such as a lack of environmental adaptability due to static mechanisms and insufficient optimization efficiency due to inherent algorithmic defects. These limitations fail to meet the requirements for efficient, intelligent, and agile scheduling of edge computing power. Specifically, whether it is the tunicate swarm algorithm based on static weighted summation or the standard multi-objective evolutionary algorithm, their inherent design exposes at least one of the following interrelated technical drawbacks in practical applications.

[0093] Rigid scheduling strategies lead to a degradation in Quality of Service (QoS) under dynamic business scenarios. Existing solutions linearly combine multiple optimization objectives, such as energy consumption and latency, into a single objective function using preset fixed weights. This static aggregation method results in a computing power scheduling strategy that lacks adaptability to environmental and business changes. In smart city operations, task priorities change dynamically (for example, a sudden public safety incident can instantly increase the latency requirements of related video analytics tasks). This technical solution cannot adjust its optimization focus according to changes in the real-time importance of tasks. Its fixed optimization preferences may cause the system to choose a high-latency execution path in emergency scenarios in order to meet preset energy consumption targets, resulting in timeouts for critical tasks and directly causing a severe degradation in Quality of Service (QoS).

[0094] The incomplete exploration of the Pareto optimal solution set limits the efficiency of global network resource allocation. The linear weighted summation method has an inherent mathematical flaw: it can only guarantee finding the convex portion of the Pareto front. For complex multi-objective problems with non-convex (concave) regions, this method is technically unable to discover optimal solutions within these regions. These missed solutions may represent more valuable resource allocation schemes (e.g., significant performance improvements in latency at minimal energy cost). Therefore, the effective scheduling options provided to the network decision-making system are incomplete, which essentially limits the system's ability to perform global resource optimization under multiple constraints, resulting in the final edge computing resource utilization and energy efficiency failing to reach theoretically optimal levels.

[0095] The algorithm's inherent mechanism is rigid, making it difficult to balance optimization accuracy and convergence speed. Existing solutions directly apply the standard tunic swarm optimization algorithm, where the key parameter c1, used to balance global exploration and local exploitation, follows a fixed linear decreasing law, and the leader-follower population structure is statically invariant. This rigid mechanism leads to poor adaptability when facing complex problems: in the early stages of iteration, even when close to a high-quality solution region, the algorithm is still forced to conduct extensive exploration, reducing convergence speed; in the later stages of iteration, when the population gets trapped in local optima, the algorithm cannot effectively escape due to a significant decline in exploration ability. This "one-size-fits-all" search strategy makes it difficult to achieve a good balance between solution accuracy and efficiency, directly affecting the final quality of the computing power scheduling scheme and decision latency.

[0096] The static parameters of evolutionary operators result in insufficient adaptive optimization capabilities of the algorithm in complex environments. Existing solutions, such as NSGA-II, typically set fixed values ​​for the probability parameters of core evolutionary operators like crossover and mutation before execution. This static parameter configuration prevents the algorithm from dynamically adjusting based on the current convergence state and diversity level of the population. When population diversity is good, a high crossover and mutation rate may disrupt previously found optimal patterns; conversely, when the population falls into premature convergence, an excessively low mutation rate is insufficient to generate enough new individuals to escape local optima. This lack of adaptability leads to low search efficiency in dynamically changing edge computing environments, resulting in poor overall quality (i.e., low convergence) of the obtained Pareto optimal solution set.

[0097] The passive diversity maintenance mechanism leads to an uneven distribution of Pareto optimal solutions. Relying on crowding ranking to maintain population diversity is a passive mechanism that penalizes overly dense solutions during the selection phase. It can alleviate, but not fundamentally prevent, excessive clustering of the population in certain areas of the Pareto front. Especially in high-dimensional objective spaces (where smart city optimization objectives are numerous), the computational effectiveness of crowding ranking decreases. The technical consequence is that the final Pareto optimal solution set output by the algorithm may be unevenly distributed, with a large number of solutions concentrated in a few areas, while being very sparse in other areas. This leaves decision-makers lacking sufficient choices among areas with different performance trade-offs, hindering refined and differentiated management of network resources.

[0098] To solve the above problems, such as Figure 1 As shown in the figure, this application provides a computing power scheduling method, the method including:

[0099] Step 101: Based on the current scene type, obtain the target data corresponding to the scene type; the target data includes at least one of the following: node status data, task feature data, and environmental parameter data;

[0100] This step aims to accurately identify the current smart city scenario type (such as traffic monitoring, medical emergency, environmental monitoring, etc.) by analyzing task attribute data and node status data in real time, providing a scenario-based benchmark for subsequent multi-objective optimization; and to achieve real-time collection, adaptive preprocessing and cross-domain fusion of multi-source data (task characteristics, node status, environmental parameters) in response to the diversity of smart city scenarios, providing high-precision data support for subsequent multi-objective modeling.

[0101] Step 102: Based on the target data, construct a three-dimensional objective function that includes energy efficiency indicators, latency indicators, and reliability indicators, as well as a dynamic weight model that is associated with task priority and real-time node status.

[0102] This step aims to construct a dynamic three-dimensional objective function for multiple smart city scenarios. By integrating energy efficiency, latency, and reliability indicators, and dynamically adjusting the objective weights based on task priority and real-time node status, it addresses the performance imbalance problem caused by "single-objective optimization" in existing technologies. This application's embodiment transforms the physical objective into an optimizable function through mathematical modeling and achieves a collaborative balance among multiple objectives through dynamic weights.

[0103] Step 103: Based on the improved tunicate swarm algorithm, the three-dimensional objective function, and the dynamic weight model, generate a scheduling scheme for computing power tasks.

[0104] This application embodiment integrates the spatiotemporal characteristics and node heterogeneity of the smart city edge environment to construct a closed-loop system of "data acquisition - multi-objective modeling - adaptive optimization - real-time scheduling". Specifically, a dynamic weight model is constructed based on task priority and real-time node status to realize the dynamic adjustment of target weights and solve the service quality (QoS) degradation problem caused by static weighting. An improved tunic swarm algorithm is introduced to dynamically adjust the exploration factor c1 and the population structure to improve the global optimization capability in high-dimensional scenarios.

[0105] Among them, the three-dimensional objective function quantitative modeling constructs the objective function with energy efficiency, time delay, and reliability as the core. The mathematical definitions and optimization directions of each index are as follows:

[0106] 1. Energy efficiency index: Energy consumption per unit of computing power .in,

[0107] ;

[0108] in, The power (W) of the i-th node in time period t is calculated using real-time collected voltage and current data. ; Let be the execution time (s) of the task at node i. The task computational cost (GFLOP) is obtained from the task attribute data.

[0109] 2. Latency Metric: Task completion latency .in,

[0110] ;

[0111] in, The processing latency of the task at node i is estimated by considering the node load and computational intensity. fi is the CPU frequency (GHz) of the node. Let i be the load rate of node i; The computational load for the node's tasks; Data transmission latency is calculated based on node geographic coordinates and network status. , where k is the transmission loss coefficient (k=1.2). denoted as , where is the distance between the node and the task source (m), and v is the network transmission rate (m / s).

[0112] Optimization direction: min( This ensures that tasks meet QoS requirements, especially for latency-sensitive scenarios such as transportation and healthcare.

[0113] 3. Reliability Indicator: System Task Success Rate ;in,

[0114] ;

[0115] in, The reliability of node i is calculated using the failure rate λi: Ri = e − λi ⋅ ti, where λi is fitted from the node's historical failure data (e.g., exponential distribution parameters). The probability of task failure (such as data transmission error rate) is assessed through network status evaluation. μ is the error rate coefficient. Available bandwidth (Mbps).

[0116] Optimization direction: max( This ensures the reliability of critical tasks such as medical care and public safety.

[0117] In at least one embodiment of this application, step 102, which involves constructing a dynamic weight model associated with task priority and real-time node status based on the target data, includes:

[0118] Based on the relevance between the scene type and the target, and the scene type itself, a scene baseline weight is generated;

[0119] The scenario baseline weights are adjusted based on the task priority and the node load to construct the dynamic weight model.

[0120] In one implementation, the core formula for generating the scene baseline weights is: ,in: The relevance (0-1) between scene features and the target is determined by the scene recognition results:

[0121] Traffic scenario: s1=0.2 (energy efficiency), s2=0.6 (latency), s3=0.2 (reliability);

[0122] In the medical setting: s1=0.1, s2=0.3, s3=0.6;

[0123] =2 is the weight amplification factor, which enhances the influence of feature differences on the weight.

[0124] For example, in a traffic scenario, the baseline weight is ω0=[0.2,0.6,0.2], with delay weight dominating.

[0125] In another implementation, the scenario baseline weights are adjusted based on the task priority and the node load, including:

[0126] Load impact correction: When the overall node load rate At the same time, reduce the energy efficiency weight to ensure task completion: Where β=0.7 is the load influence factor. ∈[0,1].

[0127] Task priority adjustment: For high-priority tasks ( ≥8), increase reliability weight: ; .

[0128] Accordingly, the final expression for the dynamic weights in the dynamic weight model is:

[0129] ;

[0130] in, For dynamic correction items for each objective (such as load and priority correction), ensure that the sum of the weights is 1 to avoid optimization target deviation.

[0131] This application's embodiments dynamically generate a weight vector ω=[ω1,ω2,ω3] based on scene features and node states, achieving adaptive switching of the multi-objective optimization focus. This application's embodiments overcome the limitations of traditional static weighted summation, constructing a dynamic weight adjustment model deeply coupled with real-time task priority and real-time node states. By simulating biological neural feedback mechanisms, scene features (such as latency sensitivity in traffic scenarios and reliability requirements in medical scenarios) are transformed into dynamic correction signals for the weight vector, achieving "environment-objective" collaborative adaptation of the multi-objective optimization focus. A scene feature relevance evaluation mechanism is introduced, dynamically generating a weight matrix through an exponential function based on the real-time mapping relationship between task attributes and node states (such as the positive correlation between task priority and reliability objectives). A two-layer "hard constraint-soft adjustment" mechanism for resolving multi-objective conflicts is designed. When energy efficiency and latency objectives conflict, key indicators are prioritized through scene-preset thresholds (such as latency in traffic scenarios not exceeding 80%), avoiding optimization imbalances common in traditional algorithms. Multi-objective collaborative optimization efficiency is improved by 40%, and the quality of service (QoS) guarantee rate for critical tasks is increased from 75% to 98%.

[0132] It should be noted that the embodiments of this application also provide multi-objective conflict resolution strategies and scenario-based resolution rules.

[0133] The multi-objective conflict resolution strategy includes: when energy efficiency and latency objectives conflict (e.g., frequency reduction can reduce energy consumption but increase latency), conflict resolution is achieved through a scenario-based threshold mechanism, and the conflict determination formula is:

[0134] >0.5; where ∇ is the gradient of the objective function, and a conflict is determined when the cosine of the gradient angle is >0.5.

[0135] The scenario-based resolution rules include:

[0136] Transportation / Medical Scenarios: Set hard constraints on latency / reliability:

[0137] ≤0.8Tth, ≥0.999, under this constraint, optimize energy efficiency, allowing An increase of ≤20% from the benchmark value;

[0138] Environmental scenario: Setting hard constraints on energy efficiency: ≤1.0J / GFLOP optimizes latency under this constraint, allowing The increase is ≤30% from the threshold.

[0139] In summary, in this embodiment, the construction of the three-dimensional objective function is directly related to the scenario requirements: the time delay formula for the traffic scenario. In this context, the weighting ratio of geographical distance is increased by k=1.2 to ensure nearest-neighbor scheduling and reduce transmission latency, addressing the problem of excessive latency caused by ignoring geographical constraints in existing technologies. The dynamic weighting formula amplifies the differences in scene characteristics (e.g., the reliability correlation s3=0.6 in medical scenarios) through an exponential function, enabling the weight vector to respond quickly to scene changes, with a weight update latency of <50ms. Hard constraints in the conflict resolution rules (e.g., ≤0.8Tth) ensures QoS for critical tasks through mathematical thresholds, avoiding service degradation caused by "single-objective optimization" in existing technologies; the load correction term (β=0.7) is determined through training with historical data. When the load rate reaches 100%, the energy efficiency weight can be reduced to 30% of the baseline value to prioritize task completion rate. Experiments have verified that the task completion rate is improved by 15%.

[0140] In at least one embodiment of this application, the improved tunicate swarm algorithm includes: a tunicate swarm algorithm that integrates Lévy flight, chaotic mapping, and gradient guidance mechanisms.

[0141] This application improves upon traditional SSA by introducing Lévy flight, chaotic mapping, and gradient guidance mechanisms, addressing the problems of premature convergence and slow convergence speed in high-dimensional target spaces. The core logic is: "Global exploration enhancement → Local development guidance → Population structure adaptation → Convergence accuracy improvement," using multiple mechanisms to collaboratively enhance the algorithm's search efficiency and solution quality in multi-objective optimization for smart cities. Based on a biomimetic optimization framework of tunicate chain-like swarm behavior, the algorithm's "global exploration-local development" balance mechanism is reconstructed by introducing the long-step exploration characteristics of Lévy flight, the ergodic perturbation of chaotic mapping, and the directional search of gradient guidance. Unlike the fixed-parameter search of traditional SSA, this scheme enables the algorithm's behavior to autonomously adapt to the problem space. The Lévy flight mechanism simulates the random walk pattern of marine organisms, generating long-step search paths through heavy-tailed distributions to effectively escape the local optima trap in high-dimensional target spaces. Chaotic mapping generates non-periodic perturbation sequences, enhancing the spatial coverage of the initial solutions and addressing the premature convergence problem of traditional algorithms. The gradient guidance mechanism integrates the gradient information of the objective function into individual position updates, making the search direction more purposeful, similar to the "scent tracking" principle in biological foraging. The algorithm's convergence speed is improved by 40%, and the deviation between the Pareto optimal solution set and the theoretical frontier in high-dimensional scenarios is reduced from 25% to 8%.

[0142] Accordingly, step 103 includes:

[0143] Based on the improved tunicate swarm algorithm, the exploration factor and population structure are dynamically adjusted;

[0144] The three-dimensional objective function and the dynamic weight model are fused to construct a multi-objective fitness function;

[0145] Based on the adjusted exploration factor, population structure, and the multi-objective fitness function, the multi-objective optimization problem is solved.

[0146] Based on the solution to the multi-objective optimization problem, a scheduling scheme for computing power tasks is generated.

[0147] In this embodiment, adaptive exploration factor dynamic adjustment and reconstruction exploration factor are implemented. A dynamic adjustment model, integrating chaotic mapping and Lévy flight, achieves an intelligent balance between global exploration and local development. ;in =2, =0, decays non-linearly from 2 to 0 with the number of iterations t, balancing the initial exploration and the later development.

[0148] Levi's flight enhances global exploration: Where γ=0.15 is the intensity parameter (increased to 0.2 in traffic scenarios); L(β) is the Lévy distribution function, generated using the Mantegna algorithm.

[0149] When β=1.5, the probability of searching with a long step size increases by 30%, effectively escaping local optima.

[0150] Chaotic mappings improve solution diversity; Tent chaotic mappings are introduced to generate perturbation terms.

[0151] ;

[0152] in Given random initial values, the ergodicity of the chaotic sequence ensures... The randomness of the perturbation increases the solution diversity index from 0.6 to 0.85.

[0153] The position update formula is improved and guided by gradients. The position update model for leaders and followers is reconstructed by introducing the gradient of the objective function to guide the search direction. The leader's position update is as follows:

[0154] ,∇fi;

[0155] Where: Fj(t) is the j-th dimension coordinate of the food source (current optimal solution); α=0.05 is the gradient influence coefficient (improved to 0.08 in medical scenarios); ∇fi is the gradient of the objective function, guiding the search towards the optimal direction, improving the convergence speed by 40%.

[0156] When the follower's position is updated Where η=0.01 is the gradient influence coefficient. Followers are updated by the mean of the previous person and the current position, and combined with gradient information to avoid convergence stagnation caused by blind following.

[0157] The multi-objective fitness function combines a three-dimensional objective function with dynamic weights to construct the fitness function: Where norm(fi) is the normalized value of the objective function: ;in, , The historical extreme values ​​of the objective function are used to ensure the comparability of objectives with different dimensions.

[0158] In summary, the introduction of Lévy flight and chaotic mapping solves the problem of insufficient late-stage exploration capability in traditional SSA. In multi-objective optimization scenarios, the distance deviation between the Pareto solution set and the theoretical frontier is reduced from 25% to 8%. The dynamic leader ratio is adaptively adjusted according to the population distribution. When the solution set is concentrated in a certain region (such as the energy efficiency-delay trade-off region), the leader ratio is increased to 0.5 to accelerate local fine search and improve optimization accuracy by 20%. The gradient guidance term (α=0.05) makes the algorithm's search direction more purposeful. The number of convergence iterations for delay optimization in traffic scenarios is reduced from 100 to 60, meeting real-time requirements. The fitness function normalization ensures fair comparison of multiple objectives. The dynamic weights (ωi) reflect the scenario requirements in real time, avoiding the optimization deviation caused by fixed weights in existing technologies.

[0159] As an optional embodiment, generating a scheduling scheme for computing power tasks based on the solution to the multi-objective optimization problem includes:

[0160] The solutions to the multi-objective optimization problem are filtered by scenario-based hard constraints, and the filtered solutions are sorted by a preference function.

[0161] The sorted solutions are analyzed using the entropy weight method and the two-base-point method to generate a scheduling scheme for the computing power task.

[0162] This application addresses the problems of "redundant solution sets and poor scenario adaptability" in existing technologies by filtering invalid solutions through scenario-based hard constraints and generating scenario-based optimal solutions. The core logic is: "Scenario constraint filtering → preference function ranking → multi-objective trade-off analysis → solution reliability verification," which improves the quality of the solution set and scenario adaptability through a two-layer screening mechanism.

[0163] First layer: Scenario-based hard constraint filtering, setting hard constraints based on the characteristics of smart city scenarios to quickly eliminate infeasible solutions:

[0164] 1. Geographical constraints (transportation / medical scenarios).

[0165] ≤ ,in, Let be the Euclidean distance between the node and the task source. The effective service radius for the scenario.

[0166] Traffic scenarios: =300m (to ensure low transmission latency of the video stream);

[0167] Medical scenario: =500m (emergency response radius).

[0168] 2. Node heterogeneity constraints.

[0169] High computational intensity tasks ( >10 GFLOPs) are only allocated to base station type nodes (computing power ≥50 GFLOPS);

[0170] Low-power tasks ( <5W) is preferentially allocated to IoT gateways (10 GFLOPS computing power), expressed by the formula:

[0171] .

[0172] 3. Hard constraints on reliability (medical scenarios).

[0173] redundancy ≥ 2, Rreliability ≥ 0.999; where redundancy is the number of backup nodes, ensuring that the probability of task failure is < 0.001.

[0174] The second layer involves scenario-based preference function ranking, which is based on preference functions. The filtered solutions are sorted to reflect the scenario optimization preferences. The preference function is mathematically defined as follows:

[0175] Where: ωi′ is the scene preset preference weight (non-dynamic weight):

[0176] Traffic scenario: ω′=[0.2,0.6,0.2] (latency priority); Medical scenario: ω′=[0.1,0.3,0.6] (reliability priority);

[0177] norm(fi(x)) is the normalized value of the objective function (the smaller the value, the better), and 1−norm(fi(x)) converts it into a preference score (the larger the value, the better).

[0178] For example, the traffic scenario preference function is: The solution with the higher latency score is selected first. When the latency of two solutions is similar, the energy efficiency score is compared.

[0179] Multi-objective trade-off analysis and solution generation: The multi-objective trade-off characteristics are quantitatively resolved using a hybrid method of entropy weight method-two-fundamental point method TOPSIS to assist decision-making.

[0180] 1. Determining objective weights using the entropy weight method: The entropy value is entropy(fi) = −∑k=1mlnm. ln , This is the normalized value of the k-th solution to objective i, reflecting the difference in the importance of the objectives.

[0181] 2. TOPSIS (Two-Point Method) for calculating relative closeness: = + ;

[0182] in, The distance between solution i and the ideal solution; The distance between solution i and the negative ideal solution; The closer the value is to 1, the better the solution.

[0183] In summary, the geographical constraint formula in the embodiments of this application ( (≤300m) directly relates to latency requirements in traffic scenarios, ensuring latency by limiting transmission distance. <50ms, solving the problem of "excessive latency caused by remote node scheduling" in existing technologies; the preference function weight (e.g., traffic ω2′=0.6) and dynamic weight complement each other, the former reflecting the inherent preferences of the scenario and the latter reflecting the real-time state, together ensuring the accuracy of the optimization direction; the entropy weight-TOPSIS method determines the target weight through data-driven determination, and provides an objective basis for multi-objective trade-offs when the scenario dynamic weight fails due to data anomalies, improving the robustness of the system; the reliability hard constraint (redundancy≥2) ensures the high reliability of medical tasks through mathematical thresholds. Experiments have verified that this constraint reduces the task failure rate from 0.1% to 0.005%, meeting medical standards.

[0184] In at least one embodiment of this application, the method further includes:

[0185] Based on the scheduling scheme and the scheduling rules that match the scenario, the computing power tasks are allocated to corresponding nodes; the nodes include at least one of: terminals, edge control devices, and cloud management devices;

[0186] Based on Dynamic Voltage and Frequency Regulation (DVFS), the frequency and / or voltage of each node are adjusted.

[0187] This application's embodiments achieve efficient execution of the optimization scheme through a hierarchical scheduling architecture and Dynamic Voltage and Frequency Scaling (DVFS), solving the problems of "coarse scheduling granularity and insufficient hardware energy efficiency" in existing technologies. The core logic is: "edge-cloud collaborative decision-making → fine-grained task allocation → dynamic adjustment of hardware parameters → execution status monitoring," improving computing power utilization efficiency through hierarchical strategies and hardware-level optimization. This application's embodiments draw on the "central-peripheral" collaborative decision-making model of biological nervous systems to construct a distributed architecture of local decision-making at edge nodes and global coordination in the cloud. The edge layer is responsible for millisecond-level response of real-time tasks, while the cloud layer provides long-term strategy optimization, solving the decision latency problem of centralized architectures in existing technologies. A lightweight DM-SSA algorithm instance is embedded in the edge layer, realizing a closed loop of "perception-decision-execution" based on local state data (node ​​load, task queue), similar to the instinctive reaction mechanism of insects. The cloud layer trains the node computing power model through historical data, periodically updates the edge decision parameters, and forms an "experience learning" capability, similar to the brain's long-term memory optimization. A task priority arbitration mechanism is designed so that when the edge node is overloaded, the cloud takes over high-complexity tasks, realizing the coordination of "stress response-rational decision-making". The decision latency is reduced from 200ms to less than 50ms, and the utilization rate of computing resources is improved by 35.4%.

[0188] 1. Terminal layer: Local preprocessing.

[0189] Implementing entity: Terminal equipment (such as traffic cameras, medical sensors);

[0190] Scheduling strategy: Locally cache and compress low-priority non-real-time tasks (such as environmental monitoring data) (compression ratio ≥ 4:1) to reduce transmission load;

[0191] High-priority real-time tasks (such as traffic accident videos) are marked as "urgent" and sent directly to the edge controller.

[0192] 2. Edge layer: Core scheduling decisions.

[0193] Execution entity: Edge controller (deployed in base stations and smart gateways);

[0194] Scheduling logic: assign(task, node) = argmin node (ω1E unit +ω2T latency +ω3(1−R reliability Based on the DM-SSA optimization results, tasks are assigned to the optimal node with a scheduling cycle of 50ms.

[0195] 3. Cloud layer: Global resource coordination.

[0196] Implementing entity: Cloud management platform;

[0197] Scheduling strategy: When the edge node load rate > 90%, trigger the cloud to take over high computational intensity tasks (Ci > 50 GFLOPs); regularly (daily) optimize edge node resource configuration and update node computing power model. ;

[0198] Where Δf is the computing power model correction coefficient, which is dynamically updated based on historical execution data.

[0199] Different scheduling rules are designed for different scenarios, and the quantitative indicators and formulas are as follows:

[0200] 1. Traffic scenario: Delay-first scheduling, node selection formula: node∗=argmin node (T latency ⋅(1+η⋅L load ); where η=0.5 is the load penalty coefficient to prevent high-load nodes from further increasing latency; scheduling effect: average latency is controlled within T latency =0.6Tth, a 45% reduction compared to existing technologies.

[0201] 2. Medical Scenario: Reliability-first scheduling, redundancy scheduling formula: schedule(task) = {node1, node2, node3}; where node1 is the master node, and node2 and node3 are backup nodes, satisfying: min(max(R node1 ,R node2 ,R node3 )),∣R nodei -R nodej |<0.01. Scheduling effect: Three-node redundancy enables task reliability to reach 99.995%, and fault switching time is <0.5ms.

[0202] 3. Environment Scenario: Energy efficiency-first scheduling, node selection formula: node∗=argmin node (E unit ⋅(1+θ⋅T latency / T th )); where θ=0.2 is the latency tolerance coefficient, allowing for a suitable increase in latency; scheduling effect: energy consumption per unit computing power is reduced to 0.85J / GFLOP, a reduction of 43% compared to existing technologies.

[0203] Dynamic Voltage and Frequency Scaling (DVFS): Achieving a balance between energy efficiency and performance by adjusting the node's CPU frequency f and voltage V.

[0204] 1. Frequency dynamic adjustment model: f=1+exp(k(L) load -L opt ))f max ,in:

[0205] f max For the maximum frequency (e.g., 2.5GHz); L opt =0.6 is the optimal load rate, at which the energy efficiency ratio is the highest; k=5 is the adjustment sensitivity parameter (increased to 8 in traffic scenarios for rapid response to load changes).

[0206] 2. Voltage-frequency coordinated regulation.

[0207] Following Dennard's scaling law: V=V min +(V max -V min )⋅(f maxf )α; where α = 1.2 is the voltage scaling exponent, ensuring a non-linear relationship between voltage and frequency and reducing dynamic power consumption: P dyn ∝V2⋅f.

[0208] 3. DVFS parameters for different scenarios.

[0209] Traffic scenario: f max =2.5GHz, k=8, maintains high-frequency operation when load>0.8;

[0210] Medical scenario: f max =1.8GHz, k=3, prioritizing reliability over performance;

[0211] Environment / Scene: f max =1.0GHz, k=5, drops to 0.5GHz under low load.

[0212] This application breaks the traditional isolation between the algorithm and hardware layers. Through deep coupling of Dynamic Voltage Frequency Scaling (DVFS) and task scheduling, it achieves cross-layer collaboration from "algorithm optimization" to "hardware energy efficiency control," solving the energy waste problem of existing "coarse-grained scheduling." A nonlinear mapping model between task computation intensity and hardware frequency is established, similar to the adaptive adjustment of biological metabolic rate and activity intensity, automatically matching lightweight tasks to low-frequency, low-power modes. A frequency-voltage coordinated adjustment strategy is designed, based on the physical principle of circuit power consumption (power consumption is positively correlated with the square of voltage), to achieve Pareto optimal adjustment of "performance-energy consumption." A hardware state feedback loop is introduced, using physical indicators such as CPU temperature and energy consumption as constraints for algorithm optimization, forming a closed-loop control of "algorithm decision-hardware feedback." Energy consumption per unit of computing power is reduced by 40.6%, and node energy efficiency ratio is improved by 25%.

[0213] In summary, the cloud correction formula (Δf) in the three-layer scheduling architecture of this application iteratively optimizes the computing power model using historical data, reducing the node computing power prediction error from 20% to 8%, thus solving the problem of "scheduling deviation caused by static models" in the prior art; the load penalty coefficient (η=0.5) in the traffic scenario scheduling formula ensures the latency penalty for high-load nodes, and experiments have verified that this mechanism reduces latency fluctuations in high-load scenarios by 30%; the DVFS frequency model smoothly adjusts the frequency through an S-shaped function, avoiding hardware losses caused by frequent frequency adjustments, with a frequency adjustment latency of <1ms, meeting real-time scheduling requirements; the redundancy formula in the medical scenario ensures balanced reliability of backup nodes through mathematical constraints, and the three-node voting mechanism reduces the error rate to <10^-6, meeting medical equipment standards.

[0214] In at least one embodiment of this application, the method further includes:

[0215] Based on task attribute data and node status data, the current scenario type is determined; wherein, the task attribute data includes at least one of the following: task data volume, task computational intensity, task latency threshold, task priority, and task reliability requirements; the node status data includes at least one of the following: node geographic coordinates, node load rate, node energy consumption, and node failure rate.

[0216] As an optional embodiment, determining the current scenario type based on task attribute data and node status data includes:

[0217] The task attribute data and the node status data are normalized to obtain normalized data;

[0218] The normalized data is clustered using a density clustering algorithm, and the mean value of features is calculated for each cluster.

[0219] The average value of the features is matched with a preset scene template to determine the current scene type; the scene type includes at least one of the following: traffic monitoring scene, medical emergency scene, and environmental monitoring scene.

[0220] This application aims to accurately identify the current smart city scenario type (such as traffic monitoring, emergency medical care, environmental monitoring, etc.) by analyzing task attributes and edge node status in real time, providing a scenario-based benchmark for subsequent multi-objective optimization. The core logic is: "data input → feature quantification → clustering classification → scenario label generation". An improved density clustering algorithm is used to achieve real-time classification of highly dynamic scenarios, ensuring a deep match between subsequent optimization strategies and scenario requirements.

[0221] Data input and feature quantization; input data sources include:

[0222] Task attribute data (reported by terminal device): includes data volume D (MB), computational intensity C (GFLOP), and latency threshold. Task priority (0-10, 10 is the highest) Reliability requirements (%)

[0223] Node status data (reported by edge nodes) includes geographic coordinates Load rate (%), Energy consumption E (W), Failure rate (times / hour).

[0224] Feature quantization normalizes the input data (eliminating the influence of dimensions), using the following formula:

[0225] Where f is the original eigenvalue, The range of feature values ​​(based on historical data statistics, such as...) After normalization, the eigenvalues ​​are all mapped to... The interval ensures that features of different dimensions can directly participate in distance calculation.

[0226] An improved density clustering algorithm (DBSCAN) is used for scene classification. The core principle is to determine the relationship between tasks and scenes based on "density reachability," thus solving the problems of traditional DBSCAN's sensitivity to high-dimensional data and difficulty in parameter tuning. The core formula and parameter design are as follows:

[0227] ;

[0228] in:

[0229] p is the data point to be judged (the feature vector of a certain task), and q is the other data point in dataset D;

[0230] For improved weighted Euclidean distance (adapting to the differences in the importance of features in smart city scenarios): (corresponding to respectively) The importance of latency and priority is higher.

[0231] (Neighborhood radius, determined through training with 5000+ historical task data, ensuring cluster purity ≥ 95% for transportation / medical scenarios);

[0232] (Minimum number of core points to avoid interference from isolated points and adapt to the suddenness of urban tasks).

[0233] Density reachability determination if a data point chain exists ,satisfy And each exist Within the neighborhood, and If all are core points, then the densities of q and p are achievable, and they belong to the same cluster (scenario).

[0234] Clustering results are mapped to scene labels. The mean feature of each cluster is calculated and matched with a preset scene template to generate scene labels.

[0235] Traffic monitoring scenario: tags (Normalized threshold);

[0236] Medical emergency scenario: tags ;

[0237] Environmental monitoring scenario: tags .

[0238] Real-time verification mechanism: Introducing scenario confidence Conf to evaluate classification reliability:

[0239] ,when If the condition is met, output the label directly; otherwise, trigger secondary classification (addition). (Expand the scope of core points) to ensure scene recognition accuracy ≥ 97%.

[0240] In summary, the normalization formula provided in the embodiments of this application... Ensuring that features of different dimensions (such as data size in MB and priority 0-10) can directly participate in distance calculation is a prerequisite for the effectiveness of clustering algorithms; weighted Euclidean distance. By assigning weights to reflect the importance of scene features (e.g., latency weight of 0.3 is higher than data volume weight of 0.2), the clustering results are made more closely aligned with the actual needs of smart cities; core point determination formula. pass By controlling the clustering granularity, we can avoid misclassifying sudden tasks (such as traffic accidents) as environmental scenarios. The confidence formula Conf enables self-verification of classification results, and dynamic parameter adjustment ensures classification stability in highly dynamic scenarios. Through the combination of the above logic and formula, this step can complete scene recognition within 100ms, providing an accurate scene benchmark for subsequent multi-objective optimization and solving the core problem of "poor scene adaptability" in existing technologies.

[0241] In at least one embodiment of this application, obtaining target data corresponding to the current scene type includes:

[0242] Based on the acquisition strategy corresponding to the scene type, multi-source data corresponding to the scene type is collected;

[0243] The multi-source data is subjected to noise reduction processing and weighted fusion to generate the target data.

[0244] The method further includes:

[0245] The target data is subjected to a consistency check. If the consistency check fails, multi-source data corresponding to the scenario type is re-collected based on the collection strategy corresponding to the scenario type.

[0246] This step aims to address the diversity of smart city scenarios by enabling real-time acquisition, adaptive preprocessing, and cross-domain fusion of multi-source data (task features, node status, environmental parameters), providing high-precision data support for subsequent multi-objective modeling. The core logic is: "Scene-driven acquisition strategy → Adaptive noise suppression → Cross-domain feature fusion → Reliability verification." By differentiating acquisition and preprocessing for different scenarios, it solves the problem of "data distortion leading to optimization deviation" in existing technologies, ensuring the accuracy and real-time nature of input data. This application's embodiment constructs a "perception-understanding-decision" cognitive model based on scene features. By extracting key features (such as geographical constraints and reliability requirements) from scenarios like transportation and healthcare, it dynamically adjusts optimization strategies to address the insufficient scene adaptability of traditional algorithms' "one-size-fits-all" approach. A density-based clustering algorithm is used to partition the task feature space into scenarios, similar to the "feature matching" principle in biological taxonomy, automatically classifying tasks into corresponding scenario clusters. Differentiated optimization "genes" are preset for different scenario clusters (e.g., a latency-priority gene for traffic scenarios, a reliability-priority gene for medical scenarios), and the optimization direction is regulated through gene expression. A "stress response" mechanism is designed for sudden scenario switching; for example, when a medical emergency task is triggered, computing resources from other scenarios are temporarily requisitioned, similar to the emergency hormone secretion mechanism in organisms. The optimization effect of core indicators is improved by 25%-47% under different scenarios, such as achieving a reliability of 99.995% in medical scenarios.

[0247] The scenario-based data acquisition strategy dynamically adjusts the acquisition strategy based on the scenario recognition results to achieve accurate matching of "scenario-data". Some specific implementation methods include:

[0248] 1. Traffic monitoring scenario (high-frequency real-time requirements).

[0249] Data collection subject: Sensor interface module of intersection edge node (base station / smart pole);

[0250] Acquisition Target and Period: Video stream data: 1080P / 30fps, acquisition period 50ms (to ensure detailed capture of traffic events);

[0251] Node load data: CPU utilization, memory usage, collection period 50ms;

[0252] Network status data: bandwidth utilization, latency jitter, collection period 100ms.

[0253] Data collection triggering mechanism: adopts a dual mode of "periodic sampling + event triggering": sampling is performed at a period of 50ms under normal conditions; when a sudden change in traffic flow is detected (change rate > 30%), high-frequency sampling is immediately triggered (period drops to 20ms).

[0254] 2. Medical emergency scenarios (high precision and reliability requirements).

[0255] Data acquisition entity: Physiological signal acquisition module of medical edge gateway;

[0256] Data Acquisition Subjects and Period: Physiological signal data: ECG (sampling rate 1kHz), blood pressure (sampling rate 100Hz), acquisition period 1ms-10ms;

[0257] Node status data: temperature, failure rate parameters, acquisition period 10ms;

[0258] Task priority data: Real-time instructions from the emergency center, collected upon event triggering.

[0259] Redundancy acquisition mechanism: Key physiological signals (such as ECG) are acquired synchronously in three channels (sampling rate 1kHz), and data reliability is improved through hardware redundancy (bit error rate <10^-6).

[0260] 3. Environmental monitoring scenarios (low power consumption and high efficiency requirements).

[0261] Data collection entities: IoT sensor nodes (smart streetlights, environmental monitoring terminals);

[0262] Data collection targets and cycle: Environmental parameters: PM2.5, temperature and humidity; collection cycle: 10 seconds (balancing energy efficiency and data validity).

[0263] Node energy consumption data: battery level, power consumption, collection period 1 second;

[0264] Task scheduling status: data transmission queue length, collection period 500ms.

[0265] Energy-saving data acquisition strategy: adopts a "sleep-wake-up" mechanism: enters sleep mode when there is no data change (power consumption <0.1mW), and wakes up to acquire data when the parameter change exceeds the threshold (such as PM2.5 fluctuation >10%).

[0266] Adaptive preprocessing algorithms and formulas are implemented to address the different data characteristics of various scenarios, employing differentiated preprocessing methods to eliminate noise interference while preserving feature information.

[0267] 1. Traffic scenario: Kalman filter for noise reduction.

[0268] Applicable scenarios: Processing dynamically changing data such as node load and network latency collected at high frequencies;

[0269] Mathematical model: in:

[0270] This is a state estimate (such as CPU utilization). These are the collected values;

[0271] The state transition matrix (diagonal matrix) (representing state continuity);

[0272] Observation matrix (Indicates direct observation);

[0273] For process noise covariance , To measure noise covariance ;

[0274] The Kalman gain is used to adaptively adjust the estimation error.

[0275] Quantization effect: The noise suppression ratio of the load data reaches 92%, and the fluctuation range of the filtered data is reduced from ±15% to ±3%, improving the accuracy of subsequent optimization.

[0276] 2. Medical scenario: Wavelet transform for noise reduction.

[0277] Applicable scenarios: handling baseline drift and power frequency interference in high-frequency physiological signals (such as ECG);

[0278] Mathematical Model: Discrete Wavelet Transform (DWT) Formula: in:

[0279] For detail coefficients, These are approximate coefficients;

[0280] The wavelet function is selected (db4 wavelet is chosen, which is suitable for physiological signal feature extraction). It is a scaling function;

[0281] Denoising steps: Soft thresholding is applied to the detail coefficients (threshold). .

[0282] Quantitative results: The signal-to-noise ratio (SNR) of the ECG signal improved from 15dB to 40dB, and the QRS complex recognition accuracy improved from 88% to 99.5%, meeting the requirements of medical diagnosis.

[0283] 3. Environment and scenario: Moving average filtering.

[0284] Applicable scenarios: handling periodic noise from low-frequency environmental parameters (such as temperature and humidity);

[0285] Mathematical model (Dynamically adjusted based on the rate of change of environmental parameters, such as temperature) , ).

[0286] Quantitative effect: The fluctuation range of temperature and humidity data was reduced from ±0.5℃ / ±5%RH to ±0.1℃ / ±2%RH, and the data validity was improved to 98%.

[0287] A weighted fusion algorithm is used to integrate task, node, and environment data to generate a unified feature vector.

[0288] 1. The fusion formula and weight design are as follows:

[0289] General fusion model: ;in:

[0290] For node state vectors (load, energy consumption, location);

[0291] The task feature vector (data volume, computational intensity, priority);

[0292] This is a vector of environmental parameters (geographic region, network status).

[0293] The weight vector w is dynamically determined by the scene type: Traffic scene: (Task features have the highest weight, ensuring latency); Medical scenario: (Balancing node status and task characteristics ensures reliability); Environment scenario: (Environmental parameters have the highest weight and are adapted to the characteristics of the deployment area).

[0294] Dynamic weight adjustment mechanism: Introducing scenario urgency U (0-1) to dynamically adjust weights: .

[0295] 2. Feature verification after fusion;

[0296] The consistency check formula is as follows: in, The correlation coefficient between feature dimensions, when Data resampling may be triggered at certain times (e.g., due to network fluctuations causing data inconsistency).

[0297] This application's embodiments directly link data timeliness requirements through scenario-specific acquisition cycle design (e.g., 50ms for traffic, 10ms for medical), and ensure data integrity in emergency scenarios through a "cycle + event" triggering mechanism, solving the problem of "data loss due to sudden tasks caused by fixed-cycle acquisition" in existing technologies; the selection of Kalman filtering and wavelet transform is based on data characteristics (dynamic load vs. physiological signal), and the noise parameter in its formula... Determined through training with historical data (such as the standard deviation of traffic load noise). To ensure maximum noise reduction effect; weighted fusion weights (such as traffic) By prioritizing scenario requirements and placing task latency characteristics at the core, this step lays a data foundation for dynamic weight adjustment in subsequent multi-objective optimization. A consistency verification formula ensures the reliability of the fused data. When network jitter causes abnormal collected values ​​(such as a sudden increase in load to 120%), a re-sampling mechanism prevents erroneous data from affecting optimization decisions. Through the synergy of the above logic and formulas, this step achieves "accurate collection, intelligent denoising, and cross-domain fusion" of multi-source data, providing high-quality data with an error rate of <5% for subsequent dynamic objective modeling, thus ensuring the accuracy of computing power optimization from the data source.

[0298] In at least one embodiment of this application, the method further includes:

[0299] Collect task execution results;

[0300] Based on the task execution results, multi-dimensional feedback indicators are determined; the multi-dimensional feedback indicators include at least one of the following: energy efficiency feedback indicators, latency feedback indicators, and reliability feedback indicators.

[0301] Based on the multi-dimensional feedback indicators, at least one of the parameters of the improved tunic swarm algorithm, the parameters of the three-dimensional objective function, and the parameters of the dynamic weight model is adaptively adjusted.

[0302] This application's embodiments address the problem of "degraded optimization effects due to lack of feedback" in existing technologies by constructing a closed-loop feedback mechanism to optimize algorithm parameters and scene models in real time based on scheduling execution results. The core logic is: "execution data collection → performance indicator evaluation → parameter adaptive adjustment → model iterative optimization," continuously learning to improve the system's adaptability to dynamic scenarios. This application's embodiments construct a data-driven closed loop based on execution feedback, dynamically adjusting algorithm parameters and scene models through real-time evaluation of multi-dimensional performance indicators (such as energy consumption deviation rate and task timeout rate), achieving autonomous system evolution and solving the problem of insufficient adaptability of "static parameters" in existing technologies. A "detection-response" mechanism similar to a biological immune system is designed, automatically triggering adaptive adjustments to algorithm parameters (such as enhanced Levy flight intensity) when performance indicators deviate from thresholds; an online learning algorithm is used to continuously update the scene recognition model, similar to the "experience accumulation" process in biology, gradually improving the system's accuracy in recognizing new scenarios; a "genetic memory" library of optimization strategies is established, using parameter combinations of historically successful solutions as prior knowledge to accelerate optimization convergence in new scenarios. During long-term operation, the energy efficiency optimization accuracy remains above 95%, and the scene recognition accuracy remains at 97%.

[0303] In one implementation, a multi-dimensional feedback indicator system is designed to cover feedback indicators for energy efficiency, latency, and reliability, serving as the basis for optimization.

[0304] 1. Energy efficiency feedback indicators.

[0305] Actual unit energy consumption deviation rate: Err E =E unitpred |E unitreal -E unitpred |, when Err E When the energy efficiency level is >20%, the energy efficiency model correction is triggered.

[0306] Node energy efficiency ranking: Rank E =1−100percentile(Eunitreal), nodes ranked <30% will be marked as "energy efficiency needs optimization".

[0307] 2. Delay feedback metrics.

[0308] Task Timeout Rate: OTR = Total number of tasks / Number of timed-out tasks; When OTR > 10%, the latency optimization strategy is adjusted.

[0309] Delay prediction error rate: Err T =T th |T latencyreal− T latencypred When the error is greater than 30%, update the parameters of the time delay prediction model.

[0310] 3. Reliability feedback indicators.

[0311] Task Failure Rate: FR = Total number of tasks / Number of failed tasks; In medical scenarios, when FR > 0.1%, a reliability policy upgrade is triggered.

[0312] Node failure rate deviation: Err R =λ pred |λ real −λ pred | When the deviation is greater than 50%, refit the node failure rate model.

[0313] The algorithm parameters are adaptively adjusted based on feedback metrics to dynamically adjust the core parameters of the DM-SSA algorithm.

[0314] 1. Explore factor adjustment.

[0315] When Err E When >20% or OTR>10%, increase Levi's flight intensity: γ′=γ⋅(1+Err) E / 20%)(γ′≤0.3);

[0316] In medical settings, when FR > 0.1%, increase the amplitude of the chaotic mapping perturbation: δ′ = δ⋅(1 + FR / 0.1%).

[0317] 2. Population size adjustment.

[0318] Err in high dynamic scenarios (such as morning rush hour traffic) T When the population size is >30%, increase the population size: N′=N⋅(1+ErrT / 30%) (N′≤100);

[0319] In low-dynamic scenarios (such as nighttime environmental monitoring), the population size is reduced to N′=30 to improve optimization efficiency.

[0320] 3. Weighting adjustment coefficient

[0321] When the deviation between the feedback index and the optimization target persists, update the dynamic weight correction coefficient: ki′=ki⋅(1+MAE(fi)) where MAE(fi) is the average absolute error of the objective function, and ki is the weight amplification coefficient (initial value 2).

[0322] Update scene recognition and target modeling parameters through an online learning mechanism:

[0323] 1. Scene recognition model update.

[0324] After every 1000 scheduling iterations, the DBSCAN parameters are retrained using new data: ϵ′=ϵ⋅(1+ARI); MinPts′=MinPts+ΔMinPts; where ARI is the adjusted RAND index (to evaluate clustering accuracy), and ΔMinPts is the iteration step size (1-3).

[0325] 2. Update the objective function parameters.

[0326] Energy efficiency model: The node power model is updated periodically (daily): Pi′(t)=Pi(t)+β⋅(Pireal(t)−Pi(t)); where β=0.1 is the learning rate;

[0327] Delay model: The distance weight is corrected according to the actual transmission delay: k′=k⋅(1+RMSE(Ttrans)); where RMSE is the root mean square error of the transmission delay.

[0328] In the embodiments of this application, the feedback index formula (such as Err) E The algorithm directly links to the optimization target. When the deviation between actual energy consumption and prediction exceeds 20%, it triggers the adjustment of algorithm parameters. Experiments have verified that this mechanism improves the accuracy of energy consumption optimization by 15%. The dynamic adjustment of Levy's flight intensity γ′ automatically enhances the algorithm when traffic scenarios are sudden, enabling the algorithm to quickly adapt to load changes. The latency compliance rate under sudden scenarios has increased from 75% to 95%. The online update of the scene model ϵ′ adapts to changes in the urban environment through incremental learning. For example, when the distribution of traffic nodes changes due to the construction of new roads, the scene recognition accuracy remains above 97%. The weight correction coefficient ki′ is automatically adjusted according to the feedback error, which shortens the scene adaptation speed of dynamic weights from 10 iterations to 3 iterations, improving the system response speed.

[0329] Through the synergy of the above steps, the embodiments of this application realize the full-process optimization of multi-objective computing power call in smart cities. From scene recognition and data processing to algorithm optimization and scheduling execution, a complete technical closed loop is formed, which effectively solves the core problems of existing technologies in terms of dynamic adaptability, multi-objective balance and scene adaptability. The relevant optimization dimensions are shown in Table 1.

[0330] Table 1

[0331] Energy efficiency (J / GFLOP) 1.65 0.98 40.6% Average latency (ms) 182 95 47.8% Task success rate (%) 97.2 99.95 2.8% Resource utilization rate (%) 65 88 35.4% Algorithm convergence speed 100 iterations 60 iterations 40% Scene recognition accuracy 85% 97% 14.1%

[0332] This application's embodiments follow the biological intelligence evolution path of "environmental perception - intelligent decision-making - execution feedback": dynamic weights and improved SSA constitute the "central brain" of intelligent decision-making, enabling the autonomous generation of optimization strategies; a hierarchical architecture and scene adaptation form a distributed "perception-motion" system, ensuring the real-time nature and scene adaptability of decisions; hardware collaboration and self-evolution mechanisms construct a "neuro-endocrine" regulatory network, realizing cross-layer optimization from algorithm to hardware and system self-iteration. By simulating the multi-level collaborative principle of biological intelligence, this application's embodiments overcome the algorithmic limitations of existing technologies and construct a next-generation smart city computing power optimization framework with environmental adaptability.

[0333] For example, such as Figure 2 As shown, the steps of the computing power scheduling method provided in this application embodiment include:

[0334] The scene adaptation module performs scene feature recognition and classifies scenes such as traffic and medical care in real time based on task attributes and node status to determine the priority of optimization targets;

[0335] The data acquisition module performs multi-source data acquisition: collecting data such as node load and task characteristics according to different scenarios, with a cycle of 50ms-1s; and identifying scenario features (such as geographical distance constraints in traffic scenarios) through density clustering. Pre-set differentiated strategies for different scenarios (redundancy in medical scenarios) Hard constraints on energy efficiency in environmental scenarios And introduce a spatiotemporal weighting function Enables dynamic resource migration. Optimization of core metrics across different scenarios improves performance by 25%-47% (e.g., reliability reaches 99.995% in medical scenarios), and the matching degree between resource scheduling solutions and actual urban needs increases by 40%.

[0336] The optimization decision-making module performs dynamic objective modeling: It constructs a three-dimensional objective function and dynamically generates weight vectors based on the scenario; it also builds a dynamic weight model that is linked in real-time with task priority and node load, using an exponential function. Combined with load correction items This allows weights to be updated in real time according to changes in the scenario (e.g., automatically increasing latency weight to 60% during peak traffic periods). The QoS guarantee rate for critical tasks has increased from 75% to 98%, and the resource mismatch rate has decreased by 60%, effectively addressing the dynamic changes in smart city business priorities. Target weights are dynamically adjusted and optimized based on task priority and real-time node status (load, energy consumption, latency), resolving the QoS degradation problem caused by static weighting.

[0337] The optimization decision module performs DM-SSA optimization: the improved algorithm adaptively adjusts parameters to solve multi-objective optimization problems; it introduces the Lévy flight mechanism (long step size search), chaotic mapping (solution space perturbation), and gradient guidance (directional search) to reconstruct the dynamic adjustment model of the exploration factor (c_1), enabling the algorithm to adaptively adjust its strategy according to the search process (e.g., enhancing global exploration in the early stage of iteration and focusing on local development in the later stage); the introduction of the Lévy flight mechanism and chaotic mapping dynamically adjusts the exploration factor c1 and the population structure, improving the global optimization capability in high-dimensional scenarios;

[0338] The decision-making module optimizes Pareto solution set screening: it filters invalid solutions by combining scenario constraints and generates candidate scheduling schemes; in high-dimensional scenarios, the deviation between the Pareto solution set and the theoretical frontier is reduced from 25% to 8%, the convergence speed is improved by 40%, and the diversity of solutions is improved by 50%, effectively solving the problem of "premature convergence" in existing algorithms;

[0339] The execution feedback module implements hierarchical scheduling: the edge controller distributes the solution to achieve task offloading and resource allocation; a three-tier architecture of "terminal-edge-cloud" is constructed, with the edge layer embedding a lightweight DM-SSA algorithm to achieve 50ms-level decision-making based on local state data (load, task queue); the cloud layer is responsible for long-term strategy optimization, forming a "real-time response-global coordination" collaborative mode. Decision latency is reduced from 200ms to less than 50ms, the response speed to sudden tasks is improved by 75%, and computing power consumption is reduced by 30%, adapting to the second-level response requirements of smart cities; a three-tier collaborative decision-making model of "terminal-edge-cloud" is constructed to achieve millisecond-level response to sudden tasks and fine-grained scheduling of computing resources.

[0340] The execution feedback module performs DVFS parameter tuning: dynamically adjusting node hardware parameters to optimize energy efficiency; deeply integrating Dynamic Voltage and Frequency Regulation (DVFS) with task scheduling, using an S-shaped frequency function. Following Dennard's scaling principle, it achieves adaptive adjustment of "task computation intensity - hardware frequency" (e.g., lightweight tasks automatically downclock to 0.5GHz). Energy consumption per unit computing power. Down to Compared with existing technologies, it reduces energy consumption by 22%-30%, increases node energy efficiency by 25%, and improves hardware resource utilization by 35.4%.

[0341] The execution feedback module performs real-time feedback optimization: it collects execution results, iteratively optimizes algorithm parameters and scenario models, and constructs multi-dimensional feedback metrics. The algorithm parameters (such as Levy flight intensity) are dynamically adjusted based on the execution data. ) and scenario models (such as DBSCAN clustering parameters) This forms a closed loop of "perception-decision-feedback-optimization". During long-term operation, the energy efficiency optimization accuracy remains above 95%, the scene recognition accuracy remains at 97%, and the system's adaptability to new scenarios continues to improve.

[0342] In summary, this application proposes a systematic solution to the core shortcomings of existing technologies in terms of dynamic adaptability, search efficiency, scene matching, energy efficiency control, and long-term optimization through a technical chain of "dynamic weighting - adaptive algorithm - layered architecture - scene adaptation - hardware collaboration - feedback evolution." Among these, dynamic multi-objective optimization and adaptive algorithms are the core innovations that break through the bottlenecks of static weighting and fixed parameters; layered architecture and scene adaptation are key supports for the implementation of smart city scenarios; and hardware collaboration and feedback mechanisms ensure the practicality and forward-looking nature of the technical solution from both the execution and evolutionary levels.

[0343] like Figure 3 As shown in the illustration, this application embodiment also provides a computing power scheduling device, including a processor 300 and a transceiver 310. The transceiver 310 receives and transmits data under the control of the processor 300, and the processor 300 is used to perform the following operations:

[0344] Based on the current scenario type, obtain target data corresponding to the scenario type; the target data includes at least one of the following: node status data, task feature data, and environmental parameter data;

[0345] Based on the target data, a three-dimensional objective function containing energy efficiency indicators, latency indicators, and reliability indicators, as well as a dynamic weight model associated with task priority and real-time node status, are constructed.

[0346] Based on the improved tunicate swarm algorithm, the three-dimensional objective function, and the dynamic weight model, a scheduling scheme for computing power tasks is generated.

[0347] In some embodiments of this application, the processor is also configured to perform the following operations:

[0348] Based on the scheduling scheme and the scheduling rules that match the scenario, the computing power tasks are allocated to corresponding nodes; the nodes include at least one of: terminals, edge control devices, and cloud management devices;

[0349] Based on Dynamic Voltage and Frequency Regulation (DVFS), the frequency and / or voltage of each node are adjusted.

[0350] In some embodiments of this application, the processor is also configured to perform the following operations:

[0351] Based on task attribute data and node status data, the current scenario type is determined; wherein, the task attribute data includes at least one of the following: task data volume, task computational intensity, task latency threshold, task priority, and task reliability requirements; the node status data includes at least one of the following: node geographic coordinates, node load rate, node energy consumption, and node failure rate.

[0352] In some embodiments of this application, the processor is also configured to perform the following operations:

[0353] The task attribute data and the node status data are normalized to obtain normalized data;

[0354] The normalized data is clustered using a density clustering algorithm, and the mean value of features is calculated for each cluster.

[0355] The average value of the features is matched with a preset scene template to determine the current scene type; the scene type includes at least one of the following: traffic monitoring scene, medical emergency scene, and environmental monitoring scene.

[0356] In some embodiments of this application, the processor is also configured to perform the following operations:

[0357] Based on the acquisition strategy corresponding to the scene type, multi-source data corresponding to the scene type is collected;

[0358] The multi-source data is subjected to noise reduction processing and weighted fusion to generate the target data.

[0359] In some embodiments of this application, the processor is also configured to perform the following operations:

[0360] The target data is subjected to a consistency check. If the consistency check fails, multi-source data corresponding to the scenario type is re-collected based on the collection strategy corresponding to the scenario type.

[0361] In some embodiments of this application, the processor is also configured to perform the following operations:

[0362] Based on the relevance between the scene type and the target, and the scene type itself, a scene baseline weight is generated;

[0363] The scenario baseline weights are adjusted based on the task priority and the node load to construct the dynamic weight model.

[0364] In some embodiments of this application, the improved tunicate swarm algorithm includes: a tunicate swarm algorithm that integrates Lévy flight, chaotic mapping, and gradient guidance mechanisms.

[0365] In some embodiments of this application, the processor is also configured to perform the following operations:

[0366] Based on the improved tunicate swarm algorithm, the exploration factor and population structure are dynamically adjusted;

[0367] The three-dimensional objective function and the dynamic weight model are fused to construct a multi-objective fitness function;

[0368] Based on the adjusted exploration factor, population structure, and the multi-objective fitness function, the multi-objective optimization problem is solved.

[0369] Based on the solution to the multi-objective optimization problem, a scheduling scheme for computing power tasks is generated.

[0370] In some embodiments of this application, the processor is also configured to perform the following operations:

[0371] The solutions to the multi-objective optimization problem are filtered by scenario-based hard constraints, and the filtered solutions are sorted by a preference function.

[0372] The sorted solutions are analyzed using the entropy weight method and the two-base-point method to generate a scheduling scheme for the computing power task.

[0373] In some embodiments of this application, the processor is also configured to perform the following operations:

[0374] Collect task execution results;

[0375] Based on the task execution results, multi-dimensional feedback indicators are determined; the multi-dimensional feedback indicators include at least one of the following: energy efficiency feedback indicators, latency feedback indicators, and reliability feedback indicators.

[0376] Based on the multi-dimensional feedback indicators, at least one of the parameters of the improved tunic swarm algorithm, the parameters of the three-dimensional objective function, and the parameters of the dynamic weight model is adaptively adjusted.

[0377] This application embodiment integrates the spatiotemporal characteristics and node heterogeneity of the smart city edge environment to construct a closed-loop system of "data acquisition - multi-objective modeling - adaptive optimization - real-time scheduling". Specifically, a dynamic weight model is constructed based on task priority and real-time node status to realize the dynamic adjustment of target weights and solve the service quality (QoS) degradation problem caused by static weighting. An improved tunic swarm algorithm is introduced to dynamically adjust the exploration factor c1 and the population structure to improve the global optimization capability in high-dimensional scenarios.

[0378] It should be noted that the computing power scheduling device provided in this application embodiment is a device capable of executing the above computing power scheduling method. Therefore, all embodiments of the above computing power scheduling method are applicable to this device and can achieve the same or similar beneficial effects, which will not be repeated here.

[0379] This application also provides a computing power scheduling device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the various processes in the computing power scheduling method embodiments described above and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0380] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, this program implements the various processes described above in the computing power scheduling method embodiments, achieving the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0381] This application also provides a computer program product, including computer instructions. When executed by a processor, these computer instructions implement the various processes of the above-described computing power scheduling method embodiments and achieve the same technical effects. To avoid repetition, they will not be described again here.

[0382] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0383] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and 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 processor, 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 A device for one or more processes and / or the functions specified in one or more boxes.

[0384] 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 a paper article including instruction means, the instruction means being implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0385] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment, causing the computer or other programmable equipment to perform a series of operational steps 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.

[0386] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A computing power scheduling method, characterized in that, The method includes: Based on the current scenario type, obtain target data corresponding to the scenario type; the target data includes at least one of the following: node status data, task feature data, and environmental parameter data; Based on the target data, a three-dimensional objective function containing energy efficiency indicators, latency indicators, and reliability indicators, as well as a dynamic weight model associated with task priority and real-time node status, are constructed. Based on the improved tunicate swarm algorithm, the three-dimensional objective function, and the dynamic weight model, a scheduling scheme for computing power tasks is generated.

2. The method according to claim 1, characterized in that, The method further includes: Based on the scheduling scheme and the scheduling rules that match the scenario, the computing power tasks are allocated to corresponding nodes; the nodes include at least one of: terminals, edge control devices, and cloud management devices; Based on Dynamic Voltage and Frequency Regulation (DVFS), the frequency and / or voltage of each node are adjusted.

3. The method according to claim 1, characterized in that, The method further includes: Based on task attribute data and node status data, the current scenario type is determined; wherein, the task attribute data includes at least one of the following: task data volume, task computational intensity, task latency threshold, task priority, and task reliability requirements; the node status data includes at least one of the following: node geographic coordinates, node load rate, node energy consumption, and node failure rate.

4. The method according to claim 3, characterized in that, The process of determining the current scenario type based on task attribute data and node status data includes: The task attribute data and the node status data are normalized to obtain normalized data; The normalized data is clustered using a density clustering algorithm, and the mean value of features is calculated for each cluster. The average value of the features is matched with a preset scene template to determine the current scene type; the scene type includes at least one of the following: traffic monitoring scene, medical emergency scene, and environmental monitoring scene.

5. The method according to claim 1, characterized in that, The step of obtaining target data corresponding to the current scene type includes: Based on the acquisition strategy corresponding to the scene type, multi-source data corresponding to the scene type is collected; The multi-source data is subjected to noise reduction processing and weighted fusion to generate the target data.

6. The method according to claim 5, characterized in that, The method further includes: The target data is subjected to a consistency check. If the consistency check fails, multi-source data corresponding to the scenario type is re-collected based on the collection strategy corresponding to the scenario type.

7. The method according to claim 1, characterized in that, The step of constructing a dynamic weight model based on the target data and relating it to task priority and real-time node status includes: Based on the relevance between the scene type and the target, and the scene type itself, a scene baseline weight is generated; The scenario baseline weights are adjusted based on the task priority and the node load to construct the dynamic weight model.

8. The method according to claim 1, characterized in that, The improved tunicate swarm algorithm includes: a tunicate swarm algorithm that integrates Lévy flight, chaotic mapping, and gradient guidance mechanisms.

9. The method according to claim 8, characterized in that, Based on the improved tunicate swarm algorithm, the three-dimensional objective function, and the dynamic weight model, a scheduling scheme for computing power tasks is generated, including: Based on the improved tunicate swarm algorithm, the exploration factor and population structure are dynamically adjusted; The three-dimensional objective function and the dynamic weight model are fused to construct a multi-objective fitness function; Based on the adjusted exploration factor, population structure, and the multi-objective fitness function, the multi-objective optimization problem is solved. Based on the solution to the multi-objective optimization problem, a scheduling scheme for computing power tasks is generated.

10. The method according to claim 9, characterized in that, The process of generating a scheduling scheme for computing power tasks based on the solution to the multi-objective optimization problem includes: The solutions to the multi-objective optimization problem are filtered by scenario-based hard constraints, and the filtered solutions are sorted by a preference function. The sorted solutions are analyzed using the entropy weight method and the two-base-point method to generate a scheduling scheme for the computing power task.

11. The method according to any one of claims 2-10, characterized in that, The method further includes: Collect task execution results; Based on the task execution results, multi-dimensional feedback indicators are determined; the multi-dimensional feedback indicators include at least one of the following: energy efficiency feedback indicators, latency feedback indicators, and reliability feedback indicators. Based on the multi-dimensional feedback indicators, at least one of the parameters of the improved tunic swarm algorithm, the parameters of the three-dimensional objective function, and the parameters of the dynamic weight model is adaptively adjusted.

12. A computing power scheduling device, comprising a processor and a transceiver, wherein the transceiver receives and transmits data under the control of the processor, characterized in that, The processor is used to perform the following operations: Based on the current scenario type, obtain target data corresponding to the scenario type; the target data includes at least one of the following: node status data, task feature data, and environmental parameter data; Based on the target data, a three-dimensional objective function containing energy efficiency indicators, latency indicators, and reliability indicators, as well as a dynamic weight model associated with task priority and real-time node status, are constructed. Based on the improved tunicate swarm algorithm, the three-dimensional objective function, and the dynamic weight model, a scheduling scheme for computing power tasks is generated.

13. The apparatus according to claim 12, characterized in that, The processor is also used to perform the following operations: Based on the scheduling scheme and the scheduling rules that match the scenario, the computing power tasks are allocated to corresponding nodes; the nodes include at least one of: terminals, edge control devices, and cloud management devices; Based on Dynamic Voltage and Frequency Regulation (DVFS), the frequency and / or voltage of each node are adjusted.

14. The apparatus according to claim 12, characterized in that, The processor is also used to perform the following operations: Based on task attribute data and node status data, the current scenario type is determined; wherein, the task attribute data includes at least one of the following: task data volume, task computational intensity, task latency threshold, task priority, and task reliability requirements; the node status data includes at least one of the following: node geographic coordinates, node load rate, node energy consumption, and node failure rate.

15. The apparatus according to claim 14, characterized in that, The processor is also used to perform the following operations: The task attribute data and the node status data are normalized to obtain normalized data; The normalized data is clustered using a density clustering algorithm, and the mean value of features is calculated for each cluster. The average value of the features is matched with a preset scene template to determine the current scene type; the scene type includes at least one of the following: traffic monitoring scene, medical emergency scene, and environmental monitoring scene.

16. The apparatus according to claim 12, characterized in that, The processor is also used to perform the following operations: Based on the acquisition strategy corresponding to the scene type, multi-source data corresponding to the scene type is collected; The multi-source data is subjected to noise reduction processing and weighted fusion to generate the target data.

17. The apparatus according to claim 16, characterized in that, The processor is also used to perform the following operations: The target data is subjected to a consistency check. If the consistency check fails, multi-source data corresponding to the scenario type is re-collected based on the collection strategy corresponding to the scenario type.

18. The apparatus according to claim 12, characterized in that, The processor is also used to perform the following operations: Based on the relevance between the scene type and the target, and the scene type itself, a scene baseline weight is generated; The scenario baseline weights are adjusted based on the task priority and the node load to construct the dynamic weight model.

19. The apparatus according to claim 12, characterized in that, The improved tunicate swarm algorithm includes: a tunicate swarm algorithm that integrates Lévy flight, chaotic mapping, and gradient guidance mechanisms.

20. The apparatus according to claim 19, characterized in that, The processor is also used to perform the following operations: Based on the improved tunicate swarm algorithm, the exploration factor and population structure are dynamically adjusted; The three-dimensional objective function and the dynamic weight model are fused to construct a multi-objective fitness function; Based on the adjusted exploration factor, population structure, and the multi-objective fitness function, the multi-objective optimization problem is solved. Based on the solution to the multi-objective optimization problem, a scheduling scheme for computing power tasks is generated.

21. The apparatus according to claim 20, characterized in that, The processor is also used to perform the following operations: The solutions to the multi-objective optimization problem are filtered by scenario-based hard constraints, and the filtered solutions are sorted by a preference function. The sorted solutions are analyzed using the entropy weight method and the two-base-point method to generate a scheduling scheme for the computing power task.

22. The apparatus according to any one of claims 13-21, characterized in that, The processor is also used to perform the following operations: Collect task execution results; Based on the task execution results, multi-dimensional feedback indicators are determined; the multi-dimensional feedback indicators include at least one of the following: energy efficiency feedback indicators, latency feedback indicators, and reliability feedback indicators. Based on the multi-dimensional feedback indicators, at least one of the parameters of the improved tunic swarm algorithm, the parameters of the three-dimensional objective function, and the parameters of the dynamic weight model is adaptively adjusted.

23. A computing power scheduling device, comprising a memory, a processor, and a program stored in the memory and executable on the processor; characterized in that, When the processor executes the program, it implements the computing power scheduling method as described in any one of claims 1 to 11.

24. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the computing power scheduling method as described in any one of claims 1 to 11.

25. A computer program product, characterized in that, It includes computer instructions, which, when executed by a processor, implement the steps of the computing power scheduling method as described in any one of claims 1 to 11.

Images