Isomeric data center multi-micronet computer collaboration double-layer optimization scheduling method

Abstract

The application discloses a heterogeneous data center multi-micronet electric algorithm collaborative double-layer optimization scheduling method, designs a double-layer optimization architecture of upper-layer micronet energy management and lower-layer algorithm collaborative scheduling, adopts a second-order cone relaxation technology to convexly process quadratic non-convex terms of edge side local calculation delay in an optimization model of lower-layer algorithm collaborative scheduling, converts the lower-layer algorithm collaborative scheduling sub-problem into a standard second-order cone programming problem, solves local optimization problems independently through each subject, alternately updates global variables and Lagrange multipliers until convergence conditions are met, and outputs optimal scheduling schemes of each micronet and the data center. The performance saturation effect of the edge server caused by limited calculation resources is more accurately reflected, and the service quality is obviously improved. Meanwhile, long-term stability of a task queue and energy storage capacity is ensured.

Description

Technical Field

[0001] This invention relates to technologies that consider the heterogeneous characteristics of cloud and edge computing in the context of power system optimization scheduling and cloud computing collaboration, and particularly to a two-layer optimization scheduling method for multi-microgrid power computing collaboration in heterogeneous data centers. Background Technology

[0002] In recent decades, the booming development of the digital economy has driven the rapid popularization of cloud computing, big data, artificial intelligence, and other businesses. Data centers, as core infrastructure, have seen continuous expansion in scale and number, leading to explosive growth in electricity load. Driven by both the "dual carbon" goals and the "East Data West Computing" project, the large-scale integration of renewable energy (wind, solar, etc.) has become a key direction for the green transformation of data centers. Microgrid technology, which integrates distributed power sources, energy storage devices, and controllable loads, has gradually become a core support for ensuring the reliability of power supply and the efficiency of renewable energy absorption in data centers, thanks to its autonomous operation and flexible adjustment capabilities. However, the operation of multi-data center microgrids faces the challenge of strong uncertainty on both the source and load sides. Random fluctuations in wind and solar power output and dynamic changes in computing load lead to an imbalance between supply and demand in the computing system, severely restricting the overall energy utilization efficiency and service quality stability. There is an urgent need to build a collaborative mechanism that takes into account both energy optimization and computing power scheduling.

[0003] With the development of data center microgrid technology, energy optimization of single data centers has become a research hotspot. Existing work largely focuses on independent scheduling on the energy side of the microgrid, using methods such as two-stage robust optimization and Wasserstein partial robust optimization to address the uncertainties of renewable energy output and electricity price fluctuations, achieving optimized configuration of energy storage charging and discharging and distributed power output. Some studies model the energy consumption characteristics of power supply equipment such as UPS systems, combining genetic algorithms and gradient descent methods to optimize virtual machine placement strategies, reducing energy consumption losses in individual data centers. Other studies employ hardware and control methods such as liquid cooling technology and dynamic cooling strategies to reduce the data center PUE to below 1.1, significantly improving energy utilization efficiency. While these methods alleviate the energy consumption pressure of single data centers to some extent, they lack a global perspective across data centers, fail to consider the potential for spatiotemporal migration of computing loads, and struggle to achieve resource complementarity and collaborative optimization among multiple data center microgrids.

[0004] Exploring the coupling potential of computing power and energy through the collaborative scheduling of multiple data center clusters has become an important direction for improving the overall system efficiency. In the field of multi-data center collaborative optimization, existing research has achieved certain results: some studies construct multi-agent cooperation models based on game theory, achieving distributed solutions through the alternating direction multiplier method, optimizing energy interaction and profit distribution across data centers; other studies adopt a combination of day-ahead scheduling and real-time scheduling to maximize renewable energy utilization and cope with load fluctuations; scheduling methods based on Lyapunov optimization techniques achieve overall cost optimization and carbon emission reduction of multi-data center systems without requiring future prediction information. In addition, some studies have stimulated the demand response potential of multiple data centers by constructing a business flow and energy flow coupling model, achieving dynamic matching of computing load and clean energy. However, the above studies mostly focus on the balanced allocation of computing resources or cross-domain energy scheduling, failing to deeply integrate queue optimization and load migration, and not fully considering the fine energy management of energy storage systems and the two-way interaction of computing systems, making it difficult to achieve efficient collaboration of "electric computing power".

[0005] In the application of computational load migration and queue optimization techniques, existing solutions have accumulated certain technical expertise. Traditional load migration often triggers virtual machine migration based on CPU and memory utilization thresholds, achieving load balancing through cold or hot migration techniques to avoid server overload or idleness. Some research introduces queuing theory to optimize scheduling logic, balancing task waiting costs and resource idle costs by dynamically adjusting task queuing rules and resource allocation weights. However, existing load migration strategies have significant limitations: threshold settings rely on experience, which can lead to biased decision-making; frequent migrations can cause system instability; load classification lacks refined design, making it difficult to distinguish between latency-tolerant and real-time load service requirements; and queue optimization techniques are mostly applied within a single data center and have not been extended to multi-data center micronet scenarios, failing to achieve orderly scheduling and collaborative matching of tasks across nodes, thus limiting the actual effectiveness of load migration.

[0006] In the application of optimization theory to address uncertainty, existing methods are mainly divided into two categories: stochastic optimization and robust optimization. Stochastic optimization relies on precise probability distribution functions, but in practice, it is difficult to obtain complete data. Traditional robust optimization covers the worst-case scenario through uncertainty sets, which can easily lead to overly conservative solutions. Although online scheduling methods based on Lyapunov optimization do not rely on future predictions and can effectively cope with random fluctuations, their application in the collaborative computing scenarios of multi-data center microgrids is insufficient. Some studies adopt historical data-driven optimization strategies, reducing the difficulty of real-time decision-making through reference sequences, but their robustness and adaptability need to be improved when facing dynamically changing loads and energy states. How to achieve multi-objective balance while mitigating the impact of uncertainty through the deep integration of optimization theory and scheduling technology has become an urgent problem to be solved.

[0007] Chinese patent (publication number: CN116231655A) discloses a two-layer optimization scheduling method for virtual power plants that considers multiple types of source and load reserves. It optimizes shared energy storage and demand response resources through a two-stage model of day-ahead and intraday optimization. However, this application ignores the spatiotemporal migration characteristics of data center computing loads and fails to achieve deep integration of computing tasks and power dispatch. Another patent proposes a microgrid group request migration scheduling method based on federated learning. Although it achieves dynamic load migration, it does not incorporate queue optimization and stochastic optimization theories, making it difficult to cope with the complex scenarios of multi-data center microgrids. Therefore, the core development direction of multi-data center microgrid scheduling technology lies in how to achieve accurate load classification and orderly scheduling through queue optimization, combine load spatiotemporal migration to establish a collaborative power supply link, utilize stochastic optimization theory to mitigate the impact of uncertainty, and balance the reduction of overall energy costs, the improvement of renewable energy absorption rate, and the guarantee of service level agreements (SLAs).

[0008] However, existing computer-based collaborative scheduling technologies still have the following key shortcomings: 1. Insufficient characterization of heterogeneous loads: Most data centers are treated as homogeneous load nodes, failing to fully distinguish the essential differences in service quality constraints between delay-sensitive tasks (DSTs) and delay-tolerant tasks (DTTs), resulting in insufficient exploration of the design space for load spatiotemporal migration strategies. 2. Reliance on predictive information: Traditional stochastic optimization and robust optimization methods both rely on predicted data of future wind power output and computational load. Prediction errors can significantly affect scheduling performance and make it difficult to achieve real-time online decision-making. 3. Lack of queue stability: Existing studies mostly use hard constraints with a fixed upper limit for queue backlog, which cannot dynamically balance system economy and queue stability, and are prone to causing task delays to exceed the limit or resources to be wasted. 4. Disadvantages of centralized solution: Centralized scheduling requires each micronet and data center to upload all operating parameters to the central scheduling center, which poses a serious risk of privacy leakage and communication burden, and the system has poor scalability; 5. Insufficient depth of computer-computer collaboration: A two-way coupling mechanism between the queue status of the computing power side and the power supply capacity of the energy side has not been established. Most of the time, the energy side meets the needs of the computing power side in a one-way manner, and it is impossible to realize the computing power load actively responding to the energy side signals. Summary of the Invention

[0009] This invention addresses the problems of existing heterogeneous data center multi-micronet collaborative scheduling, such as reliance on predictive information, poor queue stability, insufficient collaboration depth, and high privacy risks of centralized solution, by providing a two-layer optimization scheduling method for heterogeneous data center multi-micronet collaborative scheduling that takes into account queue stability.

[0010] The objective of this invention is achieved as follows: A two-layer optimized scheduling method for multi-micronet computer-aided collaborative scheduling in heterogeneous data centers, comprising the following steps: S1. Construct a heterogeneous data center multi-microgrid system architecture including wind power generation units, energy storage units, centralized cloud data centers and geographically distributed edge base stations. Based on queuing theory, the computing tasks are divided into latency-sensitive tasks and latency-tolerant tasks. A cloud-edge task migration mechanism is established, and cloud data center operation model, edge base station operation model, microgrid energy unit model and microgrid power balance constraints are constructed respectively. S2. Introducing Lyapunov optimization theory, virtual queues are constructed for the cloud-side delay-tolerant task queue and the energy storage SOC state, respectively. Lyapunov functions and drift penalty functions are defined to decouple the long-term stochastic optimization problem into a single-time-slot deterministic optimization problem, enabling real-time decision-making based solely on the current time-slot system state. S3. Design a two-layer optimization architecture for upper-layer microgrid energy management and lower-layer computing power collaborative scheduling: The upper-layer microgrid energy management aims to minimize the total operating cost and carbon emission cost of the entire system within the scheduling cycle, optimizes the energy storage charging and discharging power of each microgrid and the power purchase and sale power of the distribution network, and transmits the marginal cost of electricity consumption and renewable energy consumption margin signals to the lower-layer computing power collaborative scheduling; The lower-layer computing power collaborative scheduling aims to minimize the total delay cost and task migration cost of the entire system, optimizes the processing volume of latency-tolerant tasks and the local processing and migration allocation of latency-sensitive tasks, and feeds back the computing power load electricity demand to the upper-layer microgrid energy management, forming a two-way coupled closed loop of computer and computing; The optimization time step of the upper-layer microgrid energy management is 1 hour, and the optimization sub-period step of the lower-layer computing power collaborative scheduling is 5 seconds, to match the fine scheduling requirements of latency-sensitive tasks; S4. The second-order cone relaxation technique is used to convexize the quadratic non-convex term of the local computation delay on the edge side in the optimization model of lower-level computing power collaborative scheduling, and the lower-level computing power collaborative scheduling subproblem is transformed into a standard second-order cone programming problem. S5. Based on the Alternating Direction Multiplier Method (ADMM), a distributed solution framework is constructed. The power demand of each microgrid computing load is defined as a global coupled variable and a global consistency constraint is constructed. Each subject independently solves the local optimization problem and alternately updates the global variables and Lagrange multipliers until the convergence condition is met, and the optimal scheduling scheme of each microgrid and data center is output.

[0011] Furthermore, the multi-microgrid system architecture specifically includes: The multi-micronet system architecture is constructed as follows: a distributed architecture consisting of one cloud data center micronet and three edge base station micronets. The cloud data center serves as the global computing power hub, responsible for centralized processing of latency-tolerant tasks and migration and assignment of latency-sensitive tasks. The edge base stations serve as local computing power nodes, responsible for providing low-latency, latency-sensitive task services to nearby users. Each micronet is independently configured with wind power and energy storage units, and weak interconnection is achieved through the power distribution network. Task interaction between the cloud data center and the edge base stations is achieved through a high-speed fiber optic private network. Heterogeneous task partitioning and migration mechanism establishment are performed, and computational tasks are divided into two categories based on queuing theory: One type is latency-tolerant tasks: these tasks are characterized by large computational demands and are used to process offline data analysis and AI model training; at the same time, due to the large computational demands and flexible completion time limits, they are directly submitted to the cloud data center for queuing and processing, and have hourly time transfer flexibility. Another type is latency-sensitive tasks: these tasks are characterized by low computational load but strict requirements on processing latency, and are used to handle real-time inference and autonomous driving edge computing. Due to the low computational load and latency in milliseconds or seconds, they are initially processed locally by the edge base station. When the local computing power utilization exceeds 90% or the queue length exceeds the threshold, some latency-sensitive tasks are migrated to the cloud data center for real-time processing. The cloud-edge task migration mechanism is as follows: Latency-tolerant tasks involve large computational loads and have lenient completion time limits, so they are directly submitted to the cloud data center for queuing and processing. Latency-sensitive tasks involve small computational loads and have strict latency requirements, so they are prioritized for local processing by nearby edge base stations. When the local computing power of an edge base station is overloaded or the queue backlog exceeds a threshold, some latency-sensitive tasks are migrated to the cloud data center for real-time processing via high-speed fiber optic channels. When the total number of latency-sensitive tasks received by the cloud data center is equal to the total number of latency-sensitive tasks migrated from all edge base stations, the task traffic conservation constraint is satisfied.

[0012] Furthermore, the cloud data center operation model includes a task queue model, an energy consumption model, and service quality constraints: The task queue model states that the cumulative processing volume of delay-tolerant tasks satisfies the deadline constraint, i.e.: (1) in: h From the initial time period to the time period t Time index; For time period h Latency-tolerant workloads handled by the inner cloud data center; lower bound Indicates the end time t The cumulative amount of tasks that must be completed; upper bound The total number of tasks reached cumulatively is expressed as follows: (2) (3) in: I h This is a binary variable used to determine the time period. h Will the arriving task be during the time period? t Deadline reached; Time period t The arrival delay-tolerant task attributes include { , }, and These represent the amount of task arrival and the maximum tolerable delay, respectively. The total task processing volume of the cloud data center is the sum of the local latency-tolerant task processing volume and the edge migration latency-sensitive task reception volume: (4) in: D t For cloud data centers during time periods t Total task processing volume, For latency-sensitive task reception during the migration of edge base stations to cloud data centers; The energy consumption model, based on the Power Usage Effectiveness (PUE) index, states that the total energy consumption of a cloud data center is: (5) (6) in: for t Total energy consumption of the Shike Cloud Data Center for t Static power of cloud data center servers for t TimeCloud Data Center Cooling Energy Consumption for t Energy consumption of auxiliary lighting in the Shike Cloud data center; η dc For electrical energy utilization efficiency; N dc This represents the total number of servers in the cloud data center. P dc,s This refers to the static power of a single server. k dc This refers to the dynamic power consumption coefficient of the cloud server. The service quality constraint is as follows: based on the M / M / 1 queuing model, the average service time of cloud data center servers shall not exceed a preset upper limit. T dc,max : (7) Where: μ dcFor cloud server service speed, λ dc This represents the task completion rate.

[0013] The edge base station operation model includes a task migration model, a computation latency model, an energy consumption model, and quality of service constraints. The task migration model: a single edge base station during a time period t The arrival of delay-sensitive tasks is satisfied. (8) in: A ec ( t (For edge base stations during time periods) t The amount of latency-sensitive tasks arriving, x ec ( t Select the local processing load for latency-sensitive tasks at the edge base station. x mig ( t Select the workload for migrating edge base stations to cloud data centers; The computational delay model uses a quadratic function to characterize the local computational delay at the edge. (9) in: d ec ( t (Time period) t Local computation latency, α , b , c The delay factor is used to calculate the migration and transmission delay for delay-sensitive tasks, as shown in the following formula: (10) in: d mig ( t (Time period) t Task migration transmission latency, k The interference coefficient is... For the transmission distance between edge base stations and cloud data centers, V tr For the transmission speed, the total processing delay must satisfy the following constraint: (11) in: T ec,max Maximum preset delay limit for edge base station task processing; The energy consumption model, i.e., the energy consumption of edge base stations, includes the computing energy consumption of edge base station servers and the energy consumption per unit task migration, as shown in the following formula: (12) in: For all edge base stations at time t Energy consumption power, M This represents the total number of edge base stations. N ec The total number of servers configured for a single edge base station. Calculate the energy consumption factor for edge servers. x ec,n ( t This refers to the latency-sensitive task volume processed locally by a single server on a single edge base station. x mig,n ( t For latency-sensitive tasks that require migration from a single server at a single edge base station to a cloud data center, β mig The unit migration energy consumption coefficient.

[0014] The microgrid energy unit model and microgrid power balance constraints are expressed as follows: Wind power generation model: The wind turbine output power is as follows, and the specific model is as follows. (13)

[0015] in: w t For time period t The actual wind speed inside; These are the inlet velocity, rated velocity, and outlet velocity of the fan, respectively. This refers to the rated output power of the fan; The formula for dynamically updating the energy storage model is as follows: (14) (15) (16) (17) in: Time periods t End of period t- The remaining energy stored at the end; This refers to the energy storage self-loss rate; η c , η d For charging efficiency and discharging efficiency; They are respectively t The charging / discharging power at any given moment; The duration is the unit time period; to ensure the service life and operational safety of energy storage equipment, it is necessary to meet equations (13)-(15). These are the lower and upper limits of the energy storage state of charge, respectively; These are the maximum charging power and maximum discharging power of the energy storage, respectively. This is a binary variable representing the charging / discharging state. u t =1 indicates that the energy storage is in a charging state. u t =0 indicates that the energy storage is in a discharging state, to avoid simultaneous charging and discharging of the energy storage; The wind power output, energy storage charging and discharging power, and interaction power with the distribution network within the microgrid must meet power balance constraints. Taking a cloud data center microgrid as an example, the expression is: (18) (19) In the formula: Time periods t The power purchased and sold by the internal microgrid to the upper-level power grid cannot both be non-zero at the same time. For time period t The sum of the power consumption of all edge base stations, For the first m Edge base stations during time periods t Energy consumption.

[0016] The Lyapunov function and drift penalty function include: 1) Construct a latency-tolerant virtual task queue on the cloud data center side, with the following dynamic update equation: (20) in: For time period t Initial task queue backlog x ( t (Time period) t The workload of handling latency-tolerant tasks, A ( t (Time period) t The amount of delayed-tolerant tasks arriving; 2) Construct a virtual queue for energy storage power. B ( t The dynamic update equation is: (twenty one) Where: θ is the energy storage charging and discharging threshold parameter; 3) Define the Lyapunov function L(Θ( t ))for: (twenty two) in: This is the system queue state vector; 4) Define single-slot Lyapunov drift as: (twenty three) Introducing a drift penalty function ,in V These are weighting coefficients used to balance the economic efficiency of system operation with queue stability. C ( t The total system cost is denoted as ) within a single time slot. 5) By minimizing the upper bound of the drift penalty function, the long-running stochastic optimization problem is transformed into a single-slot deterministic optimization problem, with the objective function being: (twenty four) The design incorporates a two-layer optimized architecture for upper-layer microgrid energy management and lower-layer computing power collaborative scheduling, including: The goal of upper-level optimization is to minimize the total operating cost and carbon tax cost of all microgrids in the entire system within the scheduling cycle. The upper-level optimization objective function is... f upper for: (25) (26) (27) (28) Where: S(·) is the mathematical expectation operator; For time period t The cost of purchasing and selling electricity in the entire microgrid and distribution network within the system; For time period t Internal equipment operation and maintenance costs; For time period t Internal carbon emission costs; the parameter definitions for each cost item are as follows: For micro-network indexing, For cloud data center micronets, For each edge base station micronet; These are the purchase and sale prices of electricity during the respective time periods; k WT k ES These are the unit power operation and maintenance cost coefficients for wind power and energy storage, respectively. μ For carbon emissions trading prices; δ The carbon emission factor per unit of electricity generated by the upper-level power grid; The constraints that the upper-level optimization needs to satisfy include: the wind power and energy storage operation constraints and microgrid balance constraints shown in equations (13)-(17), as well as the non-negative constraints of power purchase and sale and the mutual exclusion constraints of ES charging and discharging. The goal of lower-level optimization is to minimize the total latency and task migration costs of task processing between cloud data centers and edge base stations within the scheduling cycle, while ensuring queue stability and service quality. The objective function for lower-level optimization is... for: (29) (30) (31) in: For time period t Latency costs of task processing on both the inner cloud and the edge. For time period t Intra-micronet task migration costs, parameter definitions for each cost item, These are the unit task latency penalty cost coefficients for the cloud side and the edge side, respectively. The first term is the latency cost of the backlog of latency-tolerant task queues in the cloud data center, and the second term is the latency cost of processing latency-sensitive tasks at the edge base station. For the first The migration cost coefficient per unit task for migrating an edge base station to a cloud data center. For time period t No. The latency-sensitive workload of migrating edge base stations to cloud data centers; The constraints that lower-level optimization needs to meet include: task queue constraints, traffic conservation constraints, service quality constraints, and server processing capacity constraints.

[0017] Furthermore, the steps of the standard second-order cone programming problem are as follows: The quadratic function of the local computation delay of the edge base station, i.e., equation (9), is relaxed into an inequality constraint as shown below: (32) And transform equation (32) into the standard second-order cone constraint form: (33) Since the optimization objective includes a delay penalty term, the inequality constraints at the optimal solution take the equality sign, and the problems before and after relaxation are equivalent.

[0018] Furthermore, the steps for constructing global consistency constraints specifically include: 1) Define global coupling variables For the first received by the upper layer Local computing power load electricity demand of microgrids, local variables The first output of the lower layer For each microgrid's local computing power load power demand, establish global consistency constraints: (34) 2) Constructing the augmented Lagrangian function L p : (35) in: f i ( x i ) is the first The local optimization objective function of each subject, For Lagrange multipliers, ρ As a penalty factor; 3) Initialize global variables Lagrange multipliers With penalty factor ρ 0 ; 4) Each entity independently solves the local optimization problem and updates its local decision variables. With local computing load ; 5) Update the global coupling variables as shown in equation (36): (36) 6) Update the Lagrange multipliers: (37) 7) Determine the original residuals r k With dual residual s k If all values ​​are less than the preset threshold ε, then convergence is achieved, and the optimal scheduling scheme is output; otherwise, return to step 4) to continue iteration. The definitions of the original residual and the dual residual are as follows: (38) (39) The present invention has the following beneficial effects: (1) In cloud data centers, task queuing quantifies the time flexibility of cloud servers in handling large-scale computing tasks, and the service quality constraints are accurately characterized based on the M / M / 1 queuing model. In edge base stations, a quadratic function is used to characterize the nonlinear relationship between edge computing performance and the amount of tasks processed, which more accurately reflects the performance saturation effect caused by limited computing resources of edge servers. The average task latency is controlled within 0.72 seconds, and the service quality is significantly improved.

[0019] (2) By constructing a delay-tolerant virtual task queue and a virtual energy storage queue, the long-term stochastic optimization problem is decoupled into a single-time-slot deterministic optimization problem, which completely eliminates the dependence on future prediction information such as wind power output and task load, avoids the negative impact of prediction errors on scheduling effect, and ensures the long-term stability of task queues and energy storage.

[0020] (3) The design adopts a multi-time scale of "upper-level 1-hour energy dispatch + lower-level 5-second computing power dispatch", which takes into account the slow dynamic characteristics of the energy system and the fast response requirements of computing tasks. Through the bidirectional signal interaction of electricity marginal cost, renewable energy absorption margin and computing power load demand, the accurate spatiotemporal matching of computing power load with wind power output and time-of-use electricity price is realized. Compared with the independent dispatch scheme of computing power and electricity, the total cost of system operation is reduced by 41.8%, and the wind power absorption rate is increased to over 98%.

[0021] (4) The edge-side secondary calculation delay constraint is relaxed to a standard second-order cone constraint. The relaxation equivalence is rigorously proven, which transforms the lower-level non-convex optimization into a second-order cone programming problem that can be solved efficiently. The global optimal solution can be quickly obtained through a mature commercial convex optimization solver. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the modeling process of the present invention; Figure 2 This is a diagram of the heterogeneous data center microgrid system architecture of the present invention; Figure 3 This is a schematic diagram of the cloud-edge collaborative task queue interaction of the present invention; Figure 4 This is a diagram of the double-layer optimized structure of the present invention; Figure 5 This is a flowchart of the ADMM algorithm in this invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. See also: Figures 1 to 5 A heterogeneous data center multi-micronet collaborative two-layer optimization scheduling method, with specific implementation steps and procedures (e.g.) Figure 1 (as shown) S1. Construct a heterogeneous data center multi-microgrid system architecture including wind power generation unit, energy storage unit, 1 centralized cloud data center and 3 geographically distributed edge base stations. Based on queuing theory, the computing tasks are divided into latency-sensitive tasks and latency-tolerant tasks. Establish a task migration mechanism for cloud data center edge base stations. Construct cloud data center operation model, edge base station operation model, microgrid energy unit model and microgrid power balance constraints respectively. S2. Introducing Lyapunov optimization theory, virtual queues are constructed for the cloud-side delay-tolerant task queue and the energy storage SOC state, respectively. Lyapunov functions and drift penalty functions are defined to decouple the long-term stochastic optimization problem into a single-time-slot deterministic optimization problem, enabling real-time decision-making based solely on the current time-slot system state. S3. Design a two-layer optimization architecture for upper-layer microgrid energy management and lower-layer computing power collaborative scheduling: The upper layer aims to minimize the total operating cost and carbon emission cost of the entire system within the scheduling cycle, optimizes the energy storage charging and discharging power of each microgrid and the power purchase and sale power of the distribution network, and transmits the marginal cost of electricity consumption and renewable energy consumption margin signals to the lower layer; The lower layer aims to minimize the total delay cost and task migration cost of the entire system task processing, optimizes the processing volume of latency-tolerant tasks and the local processing and migration allocation of latency-sensitive tasks, and feeds back the computing power load electricity demand to the upper layer, forming a two-way coupled closed loop of computer computing. S4. The second-order cone relaxation technique is used to make the quadratic non-convex term of the local computation delay of the edge base station in the lower-level optimization model convex, and the lower-level subproblem is transformed into a standard second-order cone programming problem. S5. Based on the Alternating Direction Multiplier Method (ADMM), a distributed solution framework is constructed. The power demand of each microgrid computing load is defined as a global coupled variable and a global consistency constraint is constructed. Each subject independently solves the local optimization problem and alternately updates the global variables and Lagrange multipliers until the convergence condition is met, and the optimal scheduling scheme of each microgrid and data center is output.

[0024] The heterogeneous data center multi-micronet system targeted by this invention adopts a typical cloud-edge collaborative architecture of "1 cloud + 3 edges". Each micronet is independent and autonomous, and is weakly interconnected through the power distribution network. The cloud data center and edge base stations interact through a high-speed fiber optic private network. The overall system topology (e.g.) Figure 2 As shown in the diagram, the system is divided into four main modules: energy side, task side, cloud side, and edge side. The energy side includes the upstream power grid, wind turbines, and batteries, responsible for supplying power to the entire system. The task side includes data users, task queues, and a task scheduling center, which generates and coordinates the scheduling of computing tasks. Tasks are distributed to latency-sensitive and latency-tolerant tasks via the scheduling center. The cloud side includes power distribution and lighting, electric cooling, a cloud data center, and servers. The cloud data center relies on electric cooling to obtain cooling capacity for operation, and the servers provide centralized computing power. The edge side deploys 1-3 edge base stations to receive the distributed computing tasks. In the diagram, dashed lines represent task flows, indicating the transmission path of computing tasks; dotted lines represent power flows, transmitting system power; and long dashed lines represent cooling flows, delivering the cooling capacity required for cooling. The system relies on these three types of energy and task flows to achieve coordinated and optimized scheduling of computing and energy resources.

[0025] The hardware configuration settings for cloud data center micronets and edge base station micronets are as follows: The cloud data center's computing unit comprises 25,000 Intel Xeon Platinum 8375C servers, each with 32 cores and 64 threads, a base frequency of 2.9GHz, a static power consumption of 53W, a dynamic power consumption coefficient of 6.5W / GHz, and a service rate of 5.4GHz / server. The energy output unit is equipped with 10MW wind turbines and a 20MW energy storage system, with a charge / discharge efficiency of 95%, a self-loss rate of 0.1% / h, and a SOC safety range of 20%~90%. The auxiliary facilities have a PUE value of 1.4, and the cooling system adopts a hybrid air-cooling + water-cooling mode, with a supply air temperature of 18℃ and a return air temperature limit of 25℃.

[0026] Three edge base stations are configured, each equipped with 400 NVIDIA Jeston AGX Orin edge servers, with a single server service rate of 5GHz and a computing energy consumption coefficient of 0.08W / GB. The energy unit of the edge base station microgrid is equipped with a 2MW wind turbine and a 2MWh energy storage system for each base station, with parameters identical to those on the cloud side. The transmission distances between the three edge base stations and the cloud data center are 50km, 100km, and 200km, respectively, with a transmission rate of 800Mbps and a unit task migration energy consumption of 0.02W / GB / km.

[0027] The upper layer of this invention is a microgrid energy management layer with a time step of 1 hour and a scheduling cycle of 24 hours, which matches the slow dynamic characteristics of wind power output and energy storage charging and discharging; the lower layer is a computing power collaborative scheduling layer, which divides each hour into 720 5-second sub-time periods to match the millisecond-level response requirements of latency-sensitive tasks.

[0028] The cloud data center operation model is derived and modeled in detail. It is mainly divided into task queue model, energy consumption model, and service quality constraint model.

[0029] Task queue model: Let the total number of time periods in the scheduling cycle be... T =24, Time Period Index t =1,2,....,24, Time period t The arrival delay-tolerant task attribute is ( A ( t ), D ( t )), where A(t) is the workload (unit: GB). D ( t ) represents the maximum tolerable delay (unit: hours).

[0030] Define binary indicator variables δ ( s,t ): (1) The above formula represents the time period.s Is the arriving task within the specified time period? t Deadline has been reached.

[0031] The cumulative processing volume of the delay-tolerant task queue must meet the deadline constraint: (2) in: x ( k (Time period) k The latency-tolerant workload processed by cloud data centers.

[0032] The total task processing volume of the cloud data center is the sum of the local latency-tolerant task processing volume and the edge migration latency-sensitive task reception volume: (3) in: x mig,i (t) For the first Each edge base station during the time period t The migration is a latency-sensitive task.

[0033] Energy consumption model: The Power Usage Effectiveness (PUE) metric used for total energy consumption in cloud data centers is defined as the ratio of total data center energy consumption to IT equipment energy consumption. Total energy consumption is broken down into IT equipment energy consumption, cooling energy consumption, and power distribution and lighting energy consumption. (4) in: PUE =1.4 represents the energy utilization efficiency.

[0034] Server computing power consumption includes static idle power consumption and dynamic load power consumption: (5) in: N c,act ( t (Time period) t The number of running cloud servers, P c,sta =53w is the static power of a single server. This is the dynamic power consumption coefficient.

[0035] The modeling of cooling energy consumption is based on the air enthalpy difference method: (6) in: ρ =1.2kg / m 3 air density, v =2 m / s represents the air velocity, and c p =1005 J / The specific heat capacity of air at constant pressure. T in ( t The supply air temperature is 18℃. T out ( t ≤25℃ is the upper limit for return air temperature. COP =3.5 is the coefficient of performance (COP).

[0036] Energy consumption for power distribution and lighting is directly proportional to the amount of service provided. (7) in: β lig =12W / unit is the unit server power distribution lighting energy consumption coefficient.

[0037] Service quality constraints: Based on the M / M / 1 queuing model, the average service time of the cloud server does not exceed a preset limit. : (8) in: Service speed for a single server; This represents the task arrival rate, measured in cycles / second. A 1MB task requires 60M cycles of CPU time.

[0038] The edge base station operation model includes a task migration model, a computation latency model, and an energy consumption model, the specific derivation and establishment of which are shown below: In the task migration model, edge base stations during time periods t The arrival time of the delay-sensitive task is A ec ( t It is divided into two parts: local processing and cloud migration, which satisfies traffic conservation. The cloud-edge collaborative task queue interaction diagram is shown in the figure. Figure 3As shown in the diagram, based on the above modeling, this invention constructs a cloud-edge collaborative task queue interaction logic. Data users issue user requests, which are sent to a heterogeneous load unit containing latency-tolerant tasks (DTTs) and latency-sensitive tasks (DSTs). The request information is transmitted to the task scheduling center, which completes task allocation. The allocated tasks are distributed to the cloud data center and edge base stations 1, 2, and 3 for communication transmission. Each edge base station interacts with the cloud data center. The cloud data center connects to a server cluster, which distributes tasks to computing nodes 1, 2, and 3. The three types of computing nodes carry different types of tasks; the different shapes in the diagram represent differentiated task data flows. This architecture enables collaborative allocation of heterogeneous computing tasks between computing nodes on the cloud data center side and the edge base station side, accommodating different load requirements and achieving optimized scheduling of computing resources. This mechanism ensures low-latency service for latency-sensitive tasks while releasing the time-space scheduling potential of latency-tolerant tasks, achieving complementary cloud-edge computing power. The modeling is as follows: (9) Where: x ec ( t ) represents the processing capacity of latency-sensitive tasks handled locally, x mig ( t This refers to the latency-tolerant workload of migrating edge base stations to the cloud.

[0039] The computational latency model provides a complete characterization of nonlinear characteristics. Due to the limited resources of servers at edge base stations, significant performance saturation effects occur under high loads. A quadratic function is used to accurately characterize the nonlinear relationship between local computational latency and task processing volume. (10) Wherein: the delay coefficient is obtained through actual measurement fitting: Task migration transmission latency is directly proportional to transmission distance and task volume. (11) Where: k = 0.001s / The interference coefficient is... For the first The transmission distance between an edge base station and a cloud data center, v tr =100MB / s is the effective transfer rate.

[0040] The total processing latency of latency-sensitive tasks must meet the maximum tolerable latency constraint. T ec,max =1s: (12) The total energy consumption of edge base stations includes server computing energy consumption and task migration and transmission energy consumption: (13) in: Calculate the energy consumption factor for edge servers. The energy consumption coefficient for unit task migration.

[0041] The microgrid energy unit model includes a wind power generation model, an energy storage model, and microgrid power balance constraints. The relationship between the output of a single wind turbine and wind speed in the wind power generation model is as follows: (14) Where: w in =3m / s is the cut-in wind speed, w set =12m / s is the rated wind speed, w out =25m / s is the cut-out wind speed. This is the rated output of the fan.

[0042] Considering the wake effect of the wind turbine, the total power output of the wind farm is: (15) Where: N wt Number of wind turbines This is the wake effect coefficient.

[0043] The dynamic update equation for energy storage capacity is: (16) in: For energy storage self-loss rate, This refers to the charge / discharge efficiency.

[0044] Energy storage operation safety constraints: (17) (18) (19) (20) in: These are the upper and lower limits of SOC. For maximum charge and discharge power, u ch ( t ), u dis ( t ) is a binary variable representing the charging and discharging status.

[0045] Taking a cloud data center micronet as an example, the power balance constraint equation within the micronet is: (twenty one) in: They are respectively tThe power purchased and sold to the upper-level power grid must be constantly monitored to ensure mutual exclusion. (twenty two) The power balance constraints of edge base station microgrids are similar to those of cloud data centers, and will not be elaborated here.

[0046] Based on stochastic decoupling using Lyapunov optimization, the long-term stochastic optimization problem is transformed into a single-slot deterministic optimization problem, which does not rely on any future prediction information. The complete mathematical derivation is given below. Virtual queue construction: To transform long-term constraints into queue stability constraints, two types of virtual queues are constructed: a delay-tolerant task virtual queue and an energy storage power virtual queue.

[0047] A latency-tolerant virtual queue describes the dynamic changes in task backlog: (twenty three) in: For time period t The dynamic changes in the initial task queue backlog.

[0048] The construction of the virtual queue for energy storage power aims to transform the long-term security constraints of the energy storage SOC into queue stability constraints: (twenty four) in; B ( t (Time period) t The initial length of the virtual energy storage queue is θ = 0.02E. es,rated These are the energy storage charge / discharge threshold parameters used to control the SOC fluctuation range.

[0049] The derivation of the Lyapunov function and the upper bound of drift defines the Lyapunov function as a quantifier of the crowding level of a system queue: (25) in: This is a system queue state variable.

[0050] Define single-slot Lyapunov drift as the difference between adjacent Lyapunov functions: (26) Derivation of the upper bound of drift: From equation (23), we can see that Squaring both sides, we get: (27) Similarly, from equation (24), we can obtain: (28) Substituting equations (23) and (24) into equation (25), and rearranging, we get: (29) in: C 1 is a constant, guaranteed by the boundedness of the mission arrival rate and the energy storage charging and discharging power: (30) To achieve a trade-off between system economy and queue stability, a drift penalty function and a single-slot optimization objective are constructed. (31) Where: V>0 is the weighting coefficient, C( t The total system cost within a single time slot includes electricity purchase and sale costs, operation and maintenance costs, carbon tax costs, delay costs, and relocation costs.

[0051] Substituting equation (29) into equation (31), we obtain the upper bound of the drift penalty function: (32) According to Lyapunov optimization theory, by minimizing the above upper bound in each time slot, the long-term stability of the system queue can be guaranteed, while simultaneously achieving asymptotic optimization of the total system cost. Due to the constant... C 1. Without affecting the optimization results, the objective function for single-slot deterministic optimization is: (33) The objective function depends on the system state of the current time slot. It achieves true online real-time scheduling without requiring any future prediction information.

[0052] The two-layer optimization framework is divided into a computing power side and a power side. The computing power side comprises an upper-layer microgrid energy management system and a lower-layer computing power collaborative scheduling system. The upper layer is responsible for optimizing power purchase and sale plans and energy storage charging and discharging, outputting marginal electricity costs and renewable energy consumption. The lower layer optimizes data transmission duration for latency-tolerant tasks, local processing and task migration for latency-sensitive tasks, outputting the power demand of the computing power load. The power side consists of wind power generation, energy storage devices, and the upper-level power grid, providing wind power output, energy storage operating status, and time-of-use electricity price information. The upper and lower layers interact to transmit task queue status, marginal electricity costs, and DST migration amounts. The computing power side and the power side bidirectionally exchange power demand, wind power information, and electricity price parameters, achieving layered and collaborative optimization of energy scheduling and heterogeneous computing power task scheduling, constituting a typical two-layer optimization problem (such as...). Figure 4 (As shown). The complete form of the upper-level objective function is: (34) The cost items are calculated as follows: (35) in: For cloud data centers, For a microgrid with three edge base stations; time-of-use pricing c buy ( t ), c sell ( t As shown in Table 1.

[0053] ;

[0054] Equipment maintenance costs at branch offices: (36) Where: k wt =0.015 yuan / kWh is the wind power operation and maintenance cost coefficient, k es =0.02 yuan / kWh is the energy storage operation and maintenance cost coefficient.

[0055] Carbon tax costs: (37) middle: c carbon =50 yuan / ton is the carbon emission trading price, γ grid =0.58kg / kWh is the carbon emission factor of the power generation unit of the upstream power grid.

[0056] The constraints that the upper-level optimization needs to satisfy include: wind power and energy storage operation constraints and microgrid balance constraints shown in equations (14)-(22), as well as non-negative constraints on power purchase and sale, mutual exclusion constraints on ES charging and discharging, etc.

[0057] The lower layer aims to minimize the total system-wide task processing latency cost and task migration cost. The complete form of the objective function is: (38) Among them, the task processing latency cost is: (39) in: The unit latency cost factor for latency-tolerant tasks (DTTs) on the cloud side. This is the unit delay cost factor for delay-sensitive tasks (DSTs) at edge base stations.

[0058] Task migration cost: (40) Where: k mig,i =0.01×d i Yuan / GB For the first Unit task migration cost coefficient for each edge base station.

[0059] The constraints that the lower-level optimization needs to satisfy include: task queue constraints, traffic conservation constraints, service quality constraints, server processing capacity constraints, etc., as shown in equations (1) to (13).

[0060] Convexification of non-convex problems: In the lower-level optimization model, the local computation delay of the edge base station shown in Equation (10) is a quadratic non-convex term, which makes the lower-level subproblem a non-convex optimization problem. This invention uses a second-order cone relaxation technique to make it convex and rigorously proves the equivalence of the relaxation.

[0061] Second-order cone relaxation process: Relax the equality constraint in equation (10) into an inequality constraint: (41) Transform it into a standard second-order cone constraint form: (42) The relaxation equivalence is proven, that is, under the premise that the lower-level optimization objective function contains a penalty term, equation (41) is completely equivalent to the original equality constraint (10).

[0062] Proof of feasible solution: Assume there exists a feasible solution. satisfy Because the lower-level objective function contains Item, and Therefore, a new feasible solution can be constructed. ,in Obviously Therefore, the objective function value of the new solution is smaller and satisfies all the constraints. This indicates that the original feasible solution... It cannot be the optimal solution.

[0063] Therefore, the optimal solution to the lower-level optimization problem must satisfy... That is, the optimization problems before and after relaxation are completely equivalent.

[0064] After second-order cone relaxation, the lower-level subproblem is transformed into a standard second-order cone programming problem, which can be solved globally in polynomial time using mature convex optimization solvers such as Gurobi and CPLEX.

[0065] Distributed solution based on ADMM: To protect the privacy of multi-entity data and improve system scalability, this invention employs the Alternating Direction Multiplier Method (ADMM) to achieve distributed solution of the two-level optimization model. Each entity only needs to exchange the boundary variable of computing power load and electricity demand, without needing to upload detailed local operating parameters. The specific solution algorithm flowchart is shown below. Figure 5As shown, after the algorithm starts, it first initializes the global variables, Lagrange multipliers, and penalty factors, and then enters an iterative loop. Each layer independently solves its local optimization subproblems, then updates the global variables representing the microgrid's computing power load and power demand, and iteratively updates the Lagrange multipliers. Next, it checks the convergence condition by determining whether both the original residual and the dual residual are less than a set threshold. If the convergence condition is not met, it returns to the loop to continue solving; if it is met, it generates the optimal scheduling scheme for each entity, and the algorithm terminates.

[0066] First, define global consistency constraints and the augmented Lagrangian function, and then define global coupling variables. For the first received by the upper layer Local computing power load electricity demand of microgrids, local variables The first output of the lower layer For each microgrid's local computing power load power demand, establish global consistency constraints: (43) Constructing the augmented Lagrangian function L ρ : (44) in For the first The local optimization objective function of each subject, For Lagrange multipliers, ρ >0 is the penalty factor; The complete iterative steps of the ADMM algorithm are shown below: Step 1: Initialize global variables Lagrange multipliers Punishment factor ρ =1e-4, convergence threshold ε=1e-3, number of iterations k =0.

[0067] Step 2: Local optimization solution: Each entity independently solves its own local optimization problem, updating its local decision variables and local computing load. The upper microgrid energy management entity, i.e., solving equation (34), yields... The lower-level computing power scheduling entity, i.e., solving equation (38), yields... And calculate the local computing load. .

[0068] Step 3: Update global coupling variables: The task scheduling center collects the local computing load of each entity and updates the global coupling variables: (45) Step 4: Update the Lagrange multipliers: Each entity updates the Lagrange multipliers based on the updated global variables: (46) Step 5: Convergence check: Calculate the original residuals and dual residuals: (47) (48) If r k <ε and s k If ε < 0, the algorithm converges and outputs the optimal scheduling scheme; otherwise, let ε < 0. k = k +1, return to step 2 and continue iterating.

[0069] Throughout the entire iteration process, the central coordinator only receives the local computing load from each entity. It is impossible to obtain sensitive data such as the local task arrival volume, server running status, and energy storage SOC of each entity; the entities do not exchange any data with each other, but only interact with the central coordinator for boundary variables, which effectively protects the data privacy of multiple entities.

[0070] Complete example verification and result analysis: This invention is based on a simulation model built on the MATLAB 2024a platform, and uses the Gurob 10.0 solver for optimization. The simulation hardware configuration is an Intel Xeon Platinum 8375C CPU@2.9GHz and 500GB of memory.

[0071] The system simulation parameters are summarized in Table 2: ; Based on the above parameters and system modeling, the heterogeneous load and queue stability verification results of this invention were analyzed. Four sets of comparison scenarios were set up, and the specific scenario division is shown in Table 3: ; Scenario 1: Without distinguishing between the heterogeneous characteristics of latency-sensitive and latency-tolerant tasks, a unified M / M / 1 queuing model is adopted, with no queue stability constraints and the computing load is treated rigidly as a benchmark for optimization effect.

[0072] Scenario 2: Differentiate between latency-sensitive and latency-tolerant tasks and build latency models for each, without a queue-free stability optimization module, to verify the independent improvement effect of heterogeneous modeling on service quality.

[0073] Scenario 3: Differentiate heterogeneous loads, set a hard constraint on the upper limit of the queue backlog, adopt day-ahead deterministic optimization, and compare the performance difference between fixed constraint and dynamic queue optimization.

[0074] Scenario 4: The heterogeneous load fine-grained modeling method of this paper is fully adopted, and a virtual queue is constructed based on Lyapunov optimization to achieve dynamic stability control, which serves as the verification group for the method of this paper.

[0075] The total cost of the four scenarios and their operating results were analyzed, and the comparison results are shown in Table 4: ; As shown in Table 4, the decrease in total cost exhibits a strict scenario-based progression. Scenario 1, the baseline group, does not differentiate between heterogeneous loads and has no task migration, resulting in a total cost of 87,179.4 yuan. This includes electricity purchase cost of 65,542.15 yuan, carbon tax cost of 19,861.04 yuan, and latency cost of 1,776.21 yuan—all three being the highest among the four scenarios. Scenario 2, after differentiating between latency-sensitive and latency-tolerant tasks, reduces the total cost to 32,131.26 yuan, a 63.14% decrease compared to Scenario 1. Latency cost drops significantly to 230.20 yuan, and task migration reaches 39,861 GB. Scenario 3, with the addition of a fixed queue hard constraint, further reduces the total cost to 21,310.54 yuan. Electricity purchase cost decreases by 42.77% compared to Scenario 2, and the migration volume slightly increases to 39,970 GB. In Scenario 4, employing the complete method presented in this paper, the total cost was reduced to RMB 16,877.43, a cumulative reduction of 80.64% compared to Scenario 1. All component costs were the lowest among the four scenarios, with a task migration volume of 40,616.67 GB. The above analysis shows that refined heterogeneous load modeling can reduce the total system cost by 63.14% and increase the wind power absorption rate by 17.2 percentage points. Dynamic queue control based on Lyapunov optimization, compared to fixed hard constraints, reduces the total cost by 20.8%, latency costs by 15.3%, and increases the wind power absorption rate by 4.5 percentage points. The method of this invention reduces the total cost by 80.64% compared to the baseline scenario, increases the wind power absorption rate to 98.2%, and controls the average task latency to within [missing information]. Similarly, the effectiveness of the two-layer collaborative optimization of the computer system in this invention was analyzed. Three sets of comparison scenarios were set up to verify the effectiveness of the two-layer optimization architecture and the two-way collaborative strategy of the computer system. The specific scenario division is shown in Table 5: ; Scenario 4: The two-layer optimization architecture presented in this paper is fully adopted. It achieves bidirectional coupling and iterative optimization through "electricity marginal cost computing power load demand", serving as the strategy verification group of this paper.

[0076] Scenario 5: Microgrid energy management and computing power scheduling are completely independent. The computing power load is regarded as a rigid load, with no task migration and computer coupling, serving as a benchmark for collaborative optimization effects.

[0077] Scenario 6: Only perform cloud-edge task migration and computing power scheduling, without responding to energy-side electricity prices and wind and solar power output signals, and without bidirectional coupling between electricity and computing, to verify the optimization increment of pure computing power scheduling.

[0078] Based on the above scenario division, the comparison of the running results data for the three scenarios is shown in Table 6: ; Table 6 shows that the system's economic efficiency varies significantly under different collaborative modes. Scenario 5 has the highest total cost of 28,999.99 yuan, with electricity purchase and carbon tax being the highest among the three scenarios, and no migration required. After introducing task migration in Scenario 6, the total cost decreased to 21,876.54 yuan, with a migration volume of 19,876.54 GB and a latency cost of 245.17 yuan, but the optimization was limited due to the lack of energy-side collaboration. Scenario 4 has the lowest cost of 16,877.43 yuan, a reduction of 41.8% compared to Scenario 5 and 22.9% compared to Scenario 6; it also has the lowest costs for electricity purchase, carbon tax, and latency, but a migration volume as high as 40,616.67 GB, approximately twice that of Scenario 6. This fully verifies that the proposed strategy, through the bidirectional coupling of "electricity marginal cost and computing power load demand," can effectively utilize the guiding role of electricity prices and wind power signals, accurately migrating computing tasks to low-electricity-price, high-wind-power periods while ensuring service quality, achieving a dual reduction in energy costs and carbon emissions, demonstrating the significant superiority of computer-based collaborative optimization. The above analysis shows that the method of the present invention reduces the total cost by 41.8%, the electricity purchase cost by 46.2%, and the carbon tax cost by 33.4% compared with the independent scheduling mode; and reduces the total cost by 22.9% and increases the migration volume by 104.3% compared with the pure computing power scheduling mode, verifying the significant superiority of bidirectional computer-based collaborative scheduling.

Claims

1. A two-layer optimized scheduling method for multi-micronet computer collaborative operation in heterogeneous data centers, characterized in that, The steps are as follows: S1. Construct a heterogeneous data center multi-microgrid system architecture including wind power generation units, energy storage units, centralized cloud data centers and geographically distributed edge base stations. Based on queuing theory, the computing tasks are divided into latency-sensitive tasks and latency-tolerant tasks. A cloud-edge task migration mechanism is established, and cloud data center operation model, edge base station operation model, microgrid energy unit model and microgrid power balance constraints are constructed respectively. S2. Introducing Lyapunov optimization theory, virtual queues are constructed for the cloud-side delay-tolerant task queue and the energy storage SOC state, respectively. Lyapunov functions and drift penalty functions are defined to decouple the long-term stochastic optimization problem into a single-time-slot deterministic optimization problem, enabling real-time decision-making based solely on the current time-slot system state. S3. Design a two-layer optimization architecture for upper-layer microgrid energy management and lower-layer computing power collaborative scheduling: The upper-layer microgrid energy management aims to minimize the total operating cost and carbon emission cost of the entire system within the scheduling cycle, optimizes the energy storage charging and discharging power of each microgrid and the power purchase and sale power of the distribution network, and transmits the marginal cost of electricity consumption and renewable energy consumption margin signals to the lower-layer computing power collaborative scheduling; The lower-layer computing power collaborative scheduling aims to minimize the total delay cost and task migration cost of the entire system, optimizes the processing volume of latency-tolerant tasks and the local processing and migration allocation of latency-sensitive tasks, and feeds back the computing power load electricity demand to the upper-layer microgrid energy management, forming a two-way coupled closed loop of computer and computing; The optimization time step of the upper-layer microgrid energy management is 1 hour, and the optimization sub-period step of the lower-layer computing power collaborative scheduling is 5 seconds, to match the fine scheduling requirements of latency-sensitive tasks; S4. The second-order cone relaxation technique is used to convexize the quadratic non-convex term of the local computation delay on the edge side in the optimization model of lower-level computing power collaborative scheduling, and the lower-level computing power collaborative scheduling subproblem is transformed into a standard second-order cone programming problem. S5. Based on the Alternating Direction Multiplier Method (ADMM), a distributed solution framework is constructed. The power demand of each microgrid computing load is defined as a global coupled variable and a global consistency constraint is constructed. Each subject independently solves the local optimization problem and alternately updates the global variables and Lagrange multipliers until the convergence condition is met, and the optimal scheduling scheme of each microgrid and data center is output.

2. The heterogeneous data center multi-micronet computer collaborative two-layer optimization scheduling method according to claim 1, characterized in that, The multi-microgrid system architecture specifically includes: The multi-micronet system architecture is constructed as follows: a distributed architecture consisting of one cloud data center micronet and three edge base station micronets. The cloud data center serves as the global computing power hub, responsible for centralized processing of latency-tolerant tasks and migration and assignment of latency-sensitive tasks. The edge base stations serve as local computing power nodes, responsible for providing low-latency, latency-sensitive task services to nearby users. Each micronet is independently configured with wind power and energy storage units, and weak interconnection is achieved through the power distribution network. Task interaction between the cloud data center and the edge base stations is achieved through a high-speed fiber optic private network. Heterogeneous task partitioning and migration mechanism establishment are performed, and computational tasks are divided into two categories based on queuing theory: One type is latency-tolerant tasks: these tasks are characterized by large computational demands and are used to process offline data analysis and AI model training; at the same time, due to the large computational demands and flexible completion time limits, they are directly submitted to the cloud data center for queuing and processing, and have hourly time transfer flexibility. Another type is latency-sensitive tasks: these tasks are characterized by low computational load but strict requirements on processing latency, and are used to handle real-time inference and autonomous driving edge computing. Due to the low computational load and latency in milliseconds or seconds, they are initially processed locally by the edge base station. When the local computing power utilization exceeds 90% or the queue length exceeds the threshold, some latency-sensitive tasks are migrated to the cloud data center for real-time processing. The cloud-edge task migration mechanism is specifically as follows: Latency-tolerant tasks involve large computational loads and have lenient completion time limits, so they are directly submitted to the cloud data center for queuing and processing. Latency-sensitive tasks involve small computational loads and have strict latency requirements, so they are prioritized for local processing by nearby edge base stations. When the local computing power of an edge base station is overloaded or the queue backlog exceeds a threshold, some latency-sensitive tasks are migrated to the cloud data center for real-time processing via high-speed fiber optic channels. When the total number of latency-sensitive tasks received by the cloud data center is equal to the total number of latency-sensitive tasks migrated from all edge base stations, the task traffic conservation constraint is satisfied.

3. The heterogeneous data center multi-micronet computer collaborative two-layer optimization scheduling method according to claim 1, characterized in that, The cloud data center operation model includes a task queue model, an energy consumption model, and service quality constraints. The task queue model states that the cumulative processing volume of delay-tolerant tasks satisfies the deadline constraint, i.e.: （1） in: h From the initial time period to the time period t Time index; For time period h Latency-tolerant workloads handled by the inner cloud data center; lower bound Indicates the end time t The cumulative amount of tasks that must be completed; upper bound The total number of tasks reached cumulatively is expressed as follows: (2) (3) in: I h This is a binary variable used to determine the time period. h Will the arriving task be within the time period? t Deadline reached; Time period t The arrival delay-tolerant task attributes include { , }, and These represent the amount of task arrival and the maximum tolerable delay, respectively. The total task processing volume of the cloud data center is the sum of the local latency-tolerant task processing volume and the edge migration latency-sensitive task reception volume: （4） in: D t For cloud data centers during time periods t Total task processing volume, For latency-sensitive task reception during the migration of edge base stations to cloud data centers; The energy consumption model, based on the Power Usage Effectiveness (PUE) index, states that the total energy consumption of a cloud data center is: （5） （6） in: for t Total energy consumption of the Shike Cloud data center for t Static power of cloud data center servers for t TimeCloud Data Center Cooling Energy Consumption for t Energy consumption of auxiliary lighting in the Shikeyun data center; η dc For electrical energy utilization efficiency; N dc This represents the total number of servers in the cloud data center. P dc,s This refers to the static power of a single server. k dc This refers to the dynamic power consumption coefficient of the cloud server. The service quality constraint is as follows: based on the M / M / 1 queuing model, the average service time of cloud data center servers shall not exceed a preset upper limit. T dc,max : （7） Where: μ dc For cloud server service speed, λ dc This represents the task completion rate.

4. The heterogeneous data center multi-micronet computer collaborative two-layer optimization scheduling method according to claim 1, characterized in that, The edge base station operation model includes a task migration model, a computation latency model, an energy consumption model, and quality of service constraints. The task migration model: a single edge base station during a time period t The arrival of delay-sensitive tasks is satisfied. （8） in: A ec ( t (For edge base stations during time periods) t The amount of latency-sensitive tasks arriving, x ec ( t Select the local processing load for latency-sensitive tasks at the edge base station. x mig ( t Select the workload for migrating edge base stations to cloud data centers; The computational delay model uses a quadratic function to characterize the local computational delay at the edge. （9） in: d ec ( t (Time period) t Local computation latency, α , b , c The delay factor is used to calculate the migration and transmission delay for delay-sensitive tasks, as shown in the following formula: （10） in: d mig ( t (Time period) t Task migration transmission latency, k The interference coefficient is... For the transmission distance between edge base stations and cloud data centers, V tr For the transmission speed, the total processing delay must satisfy the following constraint: （11） in: T ec,max Maximum preset delay limit for edge base station task processing; The energy consumption model, i.e., the energy consumption of edge base stations, includes the computing energy consumption of edge base station servers and the energy consumption per unit task migration, as shown in the following formula: （12） in: For all edge base stations at time t Energy consumption power, M This represents the total number of edge base stations. N ec The total number of servers configured for a single edge base station. Calculate the energy consumption factor for edge servers. x ec,n ( t This refers to the latency-sensitive task volume processed locally by a single server on a single edge base station. x mig,n ( t For latency-sensitive tasks that require migration from a single server at a single edge base station to a cloud data center, β mig The unit migration energy consumption coefficient.

5. The heterogeneous data center multi-micronet computer collaborative two-layer optimization scheduling method according to claim 1, characterized in that, The microgrid energy unit model and microgrid power balance constraints are expressed as follows: Wind power generation model: The wind turbine output power is The specific model is as follows: （13） in: w t For time period t The actual wind speed inside; These are the inlet velocity, rated velocity, and outlet velocity of the fan, respectively. This refers to the rated output power of the fan; The formula for dynamically updating the energy storage model is as follows: (14) (15) (16) (17) in: Time periods t End of period t- The remaining energy stored at the end; This refers to the energy storage self-loss rate; η c , η d For charging efficiency and discharging efficiency; They are respectively t The charging / discharging power at any given moment; The duration of a unit time period; to ensure the service life and operational safety of energy storage equipment, it is necessary to meet equations (13)-(15). These are the lower and upper limits of the energy storage state of charge, respectively; These are the maximum charging power and maximum discharging power of the energy storage, respectively. This is a binary variable representing the charging / discharging state. u t =1 indicates that the energy storage is in a charging state. u t =0 indicates that the energy storage is in a discharging state, to avoid simultaneous charging and discharging of the energy storage; The wind power output, energy storage charging and discharging power, and interaction power with the distribution network within the microgrid must meet power balance constraints. Taking a cloud data center microgrid as an example, the expression is: (18) (19) In the formula: Time periods t The power purchased and sold by the internal microgrid to the upper-level power grid cannot both be non-zero at the same time. For time period t The sum of the power consumption of all edge base stations, For the first m Edge base stations during time periods t Energy consumption.

6. The heterogeneous data center multi-micronet computer collaborative two-layer optimization scheduling method according to claim 1, characterized in that, The Lyapunov function and drift penalty function include: 1) Construct a latency-tolerant virtual task queue on the cloud data center side, with the following dynamic update equation: (20) in: For time period t Initial task queue backlog x ( t (Time period) t The workload of handling latency-tolerant tasks, A ( t (Time period) t The amount of delayed-tolerant tasks arriving; 2) Construct a virtual queue for energy storage power. B ( t The dynamic update equation is: （21） Where: θ is the energy storage charging and discharging threshold parameter; 3) Define the Lyapunov function L(Θ( t ))for: （22） in: This is the system queue state vector; 4) Define single-slot Lyapunov drift as: （23） Introducing a drift penalty function ,in V These are weighting coefficients used to balance the economic efficiency of system operation with queue stability. C ( t The total system cost is denoted as ) within a single time slot. 5) By minimizing the upper bound of the drift penalty function, the long-running stochastic optimization problem is transformed into a single-slot deterministic optimization problem, with the objective function being: （24）。 7. The heterogeneous data center multi-micronet computer collaborative two-layer optimization scheduling method according to claim 1, characterized in that, The design incorporates a two-layer optimized architecture for upper-layer microgrid energy management and lower-layer computing power collaborative scheduling, including: The goal of upper-level optimization is to minimize the total operating cost and carbon tax cost of all microgrids in the entire system within the scheduling cycle. The upper-level optimization objective function is... f upper for: （25） （26） （27） （28） Where: S(·) is the mathematical expectation operator; For time period t The cost of purchasing and selling electricity in the entire microgrid and distribution network within the system; For time period t Internal equipment operation and maintenance costs; For time period t Internal carbon emission costs; the parameter definitions for each cost item are as follows: For micro-network indexing, For cloud data center micronets, For each edge base station micronet; These are the purchase and sale prices of electricity during the respective time periods; k WT k ES These are the unit power operation and maintenance cost coefficients for wind power and energy storage, respectively. μ For carbon emissions trading prices; δ The carbon emission factor per unit of electricity generated by the upper-level power grid; The constraints that the upper-level optimization needs to satisfy include: the wind power and energy storage operation constraints and microgrid balance constraints shown in equations (13)-(17), as well as the non-negative constraints of power purchase and sale and the mutual exclusion constraints of ES charging and discharging. The goal of lower-level optimization is to minimize the total latency and task migration costs of task processing between cloud data centers and edge base stations within the scheduling cycle, while ensuring queue stability and service quality. The objective function for lower-level optimization is... for: （29） （30） （31） in: For time period t Latency costs of task processing on both the inner cloud and the edge. For time period t Intra-micronet task migration cost, parameter definitions for each cost item. These are the unit task latency penalty cost coefficients for the cloud side and the edge side, respectively. The first term is the latency cost of the backlog of latency-tolerant task queues in the cloud data center, and the second term is the latency cost of processing latency-sensitive tasks at the edge base station. For the first The migration cost coefficient per unit task for migrating an edge base station to a cloud data center. For time period t No. The latency-sensitive workload of migrating edge base stations to cloud data centers; The constraints that lower-level optimization needs to meet include: task queue constraints, traffic conservation constraints, service quality constraints, and server processing capacity constraints.

8. The heterogeneous data center multi-micronet computer collaborative two-layer optimization scheduling method according to claim 1, characterized in that, The steps of the standard second-order cone programming problem are as follows: The quadratic function of the local computation delay of the edge base station, i.e., equation (9), is relaxed into an inequality constraint as shown below: （32） And transform equation (32) into the standard second-order cone constraint form: （33） Since the optimization objective includes a delay penalty term, the inequality constraints at the optimal solution take the equality sign, and the problems before and after relaxation are equivalent.

9. The heterogeneous data center multi-micronet computer collaborative two-layer optimization scheduling method according to claim 1, characterized in that, The steps for constructing global consistency constraints specifically include: 1) Define global coupling variables For the first received by the upper layer Local computing power load electricity demand of microgrids, local variables The first output of the lower layer For each microgrid's local computing power load power demand, establish global consistency constraints: (34) 2) Constructing the augmented Lagrangian function L p : （35） in: f i ( x i ) is the first The local optimization objective function of each subject, For Lagrange multipliers, ρ As a penalty factor; 3) Initialize global variables Lagrange multipliers With penalty factor ρ 0 ; 4) Each entity independently solves the local optimization problem and updates its local decision variables. With local computing load ; 5) Update the global coupling variables as shown in equation (36): (36) 6) Update the Lagrange multipliers: (37) 7) Determine the original residuals r k With dual residual s k If all values ​​are less than the preset threshold ε, then convergence is achieved, and the optimal scheduling scheme is output; otherwise, return to step 4) to continue iterating. The definitions of the original residual and the dual residual are as follows: (38) (39) 。

Images