Virtual power plant-oriented data center resource scheduling method, equipment and medium

Abstract

The invention relates to a virtual power plant-oriented data center resource scheduling method, equipment and medium, and the method comprises the steps: carrying out the multi-time scale division and dynamic characteristic decoupling of IT equipment, energy storage system and refrigeration system resources, and forming a decoupling resource model; dividing a high-collaboration resource cluster, and calibrating an elastic capacity boundary of the resource cluster; solving a multi-market collaborative optimization model in a global optimization layer at a first preset time scale according to the electric power information, the carbon information and the computing power demand information, and generating a day-ahead baseline strategy; carrying out distributed decision making by adopting a federal reinforcement learning framework at a local decision making layer based on the day-ahead baseline strategy and the real-time operation data at a second preset time scale, and generating a real-time regulation and control instruction; and when a trigger event is monitored, triggering the global optimization layer to re-plan the day-ahead baseline strategy, issuing a re-planning result to a local decision-making layer, and updating a real-time regulation and control instruction. Compared with the prior art, the method has the advantages of high reliability, collaboration, robustness and the like.

Description

Technical Field

[0001] This invention relates to the field of power resource scheduling, and in particular to a data center resource scheduling method, equipment and medium for virtual power plants. Background Technology

[0002] New power systems are rapidly transitioning to a higher proportion of renewable energy. Virtual power plants, as key carriers for aggregating distributed resources, urgently need to expand their flexibility and regulation capabilities to cope with the increasing uncertainties on both the source and load sides. Data centers, as a new type of entity combining high energy consumption and flexible regulation potential, possess significant power regulation potential due to their spatiotemporal migration capabilities of computing loads and the thermal inertia effect of their cooling systems. They are considered important "virtual energy storage" resources supporting the dynamic balance of virtual power plants.

[0003] However, existing virtual power plant resource aggregation technologies primarily focus on traditional distributed energy sources, such as energy storage, controllable loads, and distributed power sources, paying insufficient attention to new entities like data centers with spatiotemporal elasticity, thus failing to fully unleash their regulation potential. Specifically, existing technologies suffer from the following main problems: First, existing virtual power plant resource aggregation models lack the ability to spatiotemporally decouple and model the dynamic characteristics of various heterogeneous resources within data centers. The time constants of resources within data centers vary significantly, covering multiple scales from seconds to hours, while traditional centralized optimization models typically use a single time scale, making it difficult to accurately characterize the dynamic coupling relationships between different resources, leading to significant biases in regulation margin assessments. Second, data centers participating in virtual power plant regulation involve multi-objective collaborative problems such as electricity market revenue, computing power service quality, and carbon footprint transfer, constituting a high-dimensional non-convex optimization problem. Existing research often employs linear weighted or simplified game models, which struggle to effectively handle objective conflicts under multiple markets and constraints, resulting in high solution complexity and difficulty in meeting real-time decision-making requirements. Furthermore, existing regulation methods have significant bottlenecks in real-time response efficiency. The rapid fluctuations in renewable energy output and computing load require control systems with second-level decision-making capabilities. However, existing methods are limited by communication latency and computational complexity, making efficient coordination difficult and hindering the practical effectiveness of data center resources in frequency regulation and power balancing. Finally, the value quantification mechanism for data center elastic resources participating in the electricity market is not yet perfect, lacking trading rules and cost-sharing systems that balance fairness and incentive compatibility, affecting the feasibility and economics of its large-scale application.

[0004] Therefore, how to effectively aggregate and coordinate the spatiotemporal elastic resources of data centers to fully unleash their flexibility potential in virtual power plants and support the safe, low-carbon, and economical operation of high-proportion renewable energy power systems is a technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a data center resource scheduling method, device and medium for virtual power plants. By constructing a "global-local" hierarchical distributed collaborative optimization architecture and adopting a federated reinforcement learning framework, it not only achieves efficient collaborative scheduling of data center resources, but also protects the data privacy and security of all participating parties.

[0006] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a data center resource scheduling method for virtual power plants is provided, comprising the following steps: S1, dividing IT equipment, energy storage systems, and cooling system resources with different response time constants within the data center into multiple time scales and decoupling their dynamic characteristics to form a decoupled resource model; S2, based on the decoupled resource model, dividing highly collaborative resource clusters and defining the elastic capacity boundaries of the resource clusters; S3, based on a hierarchical distributed collaborative optimization architecture, solving a multi-market collaborative optimization model at the global optimization layer at a first preset time scale, based on power information, carbon information, and computing power demand information, to generate a day-ahead baseline strategy including an energy storage charging and discharging plan and an IT load allocation plan; S4, at the local decision-making layer at a second preset time scale, based on the day-ahead baseline strategy and real-time operating data, using a federated reinforcement learning framework for distributed decision-making, generating real-time control instructions; S5, monitoring the system operating status through an event triggering mechanism, and when a triggering event is detected, triggering the global optimization layer to replan the day-ahead baseline strategy, and sending the replanning result to the local decision-making layer to update the real-time control instructions.

[0007] Furthermore, the specific steps of multi-timescale division and dynamic characteristic decoupling in S1 include: IT equipment is divided into fast time scales of seconds, and its dynamic model is built based on CPU utilization and chip temperature; energy storage systems are divided into medium time scales of minutes, and their dynamic models are built based on state of charge (SOC) and charging / discharging power; the thermal inertia dynamics of refrigeration systems are divided into slow time scales of hours, and their dynamic models are built based on thermodynamic differential equations. Temporal decoupling coding technique is used to separate cross-scale coupling terms between models at different time scales. The expression is: , in, This is the feature extraction matrix for this scale. This is a cross-scale interference suppression matrix. For inter-scale propagation delay, These are the state variables for this scale; For other scales of state variables; For time indexing.

[0008] Furthermore, in step S2, the resource cluster is obtained through dynamic resource clustering using state potential game theory, specifically including the following steps: Define a cluster potential function based on the maximum adjustment capacity of the nodes and the communication latency. The expression is: , in, For nodes i Maximum adjustment capacity For communication delay, , These are the weighting coefficients. C For resource clusters, For nodes i The virtual power plant control capacity that is currently occupied within the dispatch period; Based on the cluster potential function, the optimal resource cluster partitioning scheme is obtained by solving the Nash equilibrium, so that the total adjustment capability of the nodes in each cluster reaches the optimal under the communication delay constraint. A Bayesian update method is used to dynamically adjust the margin parameter based on the deviation between the actual adjustment capacity and the predicted capacity of the resource cluster. The expression for the Bayesian update method is as follows: , in, For the updated cluster C margin parameters, For the cluster before the update C margin parameters, γ For learning rate, and Clusters C The actual regulating capacity and the predicted regulating capacity; By introducing a sub-Brønsted bar chance constraint, an elastic capacity boundary considering the impact of uncertainty is constructed. The sub-Brønsted bar chance constraint measures the deviation of the uncertainty distribution through Wasserstein distance and transforms the constraint conditions into a solvable form.

[0009] Furthermore, in S3, the hierarchical distributed collaborative optimization architecture includes a global optimization layer, a local decision-making layer, and an elastic feedback channel connecting the global optimization layer and the local decision-making layer; the global optimization layer operates at a first preset time scale, and the local decision-making layer operates at a second preset time scale that is shorter than the first preset time scale.

[0010] Furthermore, the operations performed by the global optimization layer include: Based on electricity market information, carbon market information, and computing power demand information, a multi-market joint optimization model is constructed. The objective function of the multi-market joint optimization model is to minimize the total cost, which includes electricity purchase cost, carbon cost, and computing power service quality default cost. The constraints of the multi-market joint optimization model include power balance constraints and carbon quota constraints, with the carbon quota constraints dynamically allocating green electricity quotas based on a carbon flow tracking model. The adaptive alternating direction multiplier method (AT-ADMM) is used to solve the problem and generate a day-ahead baseline strategy that includes energy storage charging and discharging plans, IT load allocation, and backup power start-up and shutdown timing.

[0011] Furthermore, the operations performed by the local decision-making layer include: A federated reinforcement learning controller group is deployed, and each edge device generates control instructions based on real-time operational data. The update expression for the policy network parameters of the federated reinforcement learning controller group is as follows: , in, θ i For nodes i The strategy network parameters; η The learning rate controls the step size for updating the network parameters of the policy. R i The local reward function includes energy cost and QoS metrics; γ is the discount factor. For nodes i The current state of the system environment as observed at the present moment; For nodes i The policy network is based on the state The decision-making instructions made; For nodes i The observed updated system environment status; In the new state Next, based on the current policy network, calculate the next action to be executed; For iterative indexing; The gradient information of the local policy network is uploaded to the cloud for aggregation using gradient encryption to update the global model weights; The decision-making strategy is dynamically adjusted based on the system disturbance level, which is based on the real-time power deviation classification.

[0012] Furthermore, the elastic feedback channel operates based on an event-triggered mechanism. When the wind and solar power output prediction deviation exceeds the first preset threshold or the computing power service quality default risk exceeds the second preset threshold, the elastic feedback channel is triggered. After triggering, the local decision-making layer sends an interruption request to the global optimization layer, and the global optimization layer initiates replanning and sends lightweight correction instructions downward.

[0013] Furthermore, the updated real-time control instructions obtained in S5 are executed to obtain the final resource scheduling result of the virtual power plant. The real-time control instructions include the output charging and discharging power instructions of the energy storage system, the load adjustment instructions of the IT equipment, the start and stop instructions of the backup power supply, and the cooling power adjustment instructions.

[0014] According to a second aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.

[0015] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) Improve the accuracy of resource modeling and the reliability of scheduling: By dividing the IT equipment, energy storage system and cooling system resources with different response characteristics in the data center into multiple time scales and decoupling their dynamic characteristics, and using time-domain decoupling coding technology to separate cross-scale coupling terms, the dynamic response characteristics of different types of resources are accurately characterized, overcoming the limitations of traditional single-time-scale modeling, thereby improving the accuracy of resource regulation margin assessment, enhancing the reliability of virtual power plant resource aggregation and scheduling, making full use of the data center's ability to regulate energy storage, providing more reliable and flexible resources for virtual power plants, and effectively supporting the stable operation of high proportion of renewable energy access to the grid.

[0017] (2) Achieving efficient collaborative optimization and privacy protection: By constructing a hierarchical distributed collaborative optimization architecture that includes a global optimization layer, a local decision-making layer and an elastic feedback channel, and adopting a federated reinforcement learning framework for distributed decision-making, this scheme achieves the collaboration between global policy and local regulation. On the one hand, it reduces computational complexity and improves optimization efficiency by solving in parallel. On the other hand, by making decisions locally and uploading only encrypted gradient information, it effectively protects the privacy of operational data of each data center while ensuring training results. This solves the computing power problem of large-scale resource collaborative optimization, eliminates concerns about data sharing between different entities, and is conducive to multiple data centers participating in virtual power plants.

[0018] (3) Enhance the robustness of the system in response to emergencies: By establishing an event-triggered elastic feedback channel, when the wind and solar power output prediction deviation or the risk of default on computing power service quality exceeds the preset threshold, the system can trigger online replanning of the global strategy. This enables the system to dynamically adjust the scheduling strategy and correct operational deviations in a timely manner when facing uncertainties such as renewable energy output fluctuations and load mutations. This improves the adaptability and overall robustness of the virtual power plant in complex operating environments. The rapid response mechanism effectively prevents power shortages caused by sudden drops in wind and solar power, while ensuring the continuity of key computing power supply. This enhances the adaptability and overall robustness of the virtual power plant in complex operating environments and ensures the safe and stable operation of the power system. Attached Figure Description

[0019] Figure 1 A flowchart of a data center resource scheduling method for virtual power plants; Figure 2 A framework diagram for the interaction between data centers and virtual power plants; Figure 3 This is a diagram of a hierarchical, distributed, collaborative optimization architecture. Figure 4 A flowchart of the solution method; Figure 5 These are the predicted values ​​for wind and solar power output and load in the VPP; Figure 6 Wiring diagram for IEEE 33-node; Figure 7 The resource power allocation and system power balance results for scenario 4 in Example 1 are optimized by time period. Figure 8 The results show the charge state and cost optimized for scenario 4 in Example 1. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] Example 1 Existing virtual power plant resource aggregation technologies primarily focus on traditional distributed energy resources, employing centralized optimization models for unified scheduling of resources on a single time scale. This approach struggles to accurately characterize the significant differences in dynamic response characteristics among heterogeneous resources such as IT equipment and energy storage systems within data centers, leading to inaccurate assessments of resource regulation potential. Furthermore, its computational complexity is high when addressing multi-market collaborative optimization, making it difficult to meet the real-time decision-making needs under renewable energy fluctuations, and it also poses a risk of data center privacy data leakage.

[0022] like Figure 1 As shown, this embodiment provides a data center resource scheduling method for virtual power plants, and the specific steps include: S1. Divide IT equipment, energy storage systems and cooling system resources with different response time constants in the data center into multiple time scales and decouple their dynamic characteristics to form a decoupled resource model; S2. Based on the decoupled resource model, highly collaborative resource clusters are divided, and the elastic capacity boundary of the resource clusters is defined. S3. Based on the hierarchical distributed collaborative optimization architecture, at the global optimization layer, using the first preset time scale, the multi-market collaborative optimization model is solved according to power information, carbon information and computing power demand information to generate day-ahead baseline strategies including energy storage charging and discharging plans and IT load allocation plans. S4. At the local decision-making level, based on the previous day baseline strategy and real-time operation data, a federated reinforcement learning framework is used to make distributed decisions and generate real-time control instructions at the second preset time scale. S5. Monitor the system's operating status through an event triggering mechanism. When a triggering event is detected, trigger the global optimization layer to replan the day-ahead baseline strategy and send the replanning results to the local decision-making layer to update real-time control instructions.

[0023] like Figure 2 The diagram illustrates the interaction framework between data centers and virtual power plants. Data centers interact with virtual power plants through spatiotemporal elastic resources, enhancing the flexibility and resilience of the virtual power plants.

[0024] In a virtual power plant, data center resources exhibit significant dynamic differences. Therefore, resource tier management based on response speed is implemented, dividing IT equipment, energy storage systems, and cooling systems into fast, medium, and slow time scales. Dynamic characteristic decoupling technology separates the coupling effects of different time scales, enabling independent modeling and collaborative optimization of resource characteristics. Highly collaborative resource clusters are dynamically partitioned using state potential game theory, and cluster capacity is updated in real time through an online learning mechanism, enhancing the flexibility and adaptability of resource scheduling. By quantifying the impact of uncertainty, elastic capacity boundaries are constructed, fully releasing the regulatory potential of the data center as "virtual energy storage."

[0025] The specific steps for multi-timescale partitioning and dynamic characteristic decoupling in S1 include: IT equipment is divided into fast time scales of seconds, and its dynamic model is built based on CPU utilization and chip temperature; energy storage systems are divided into medium time scales of minutes, and their dynamic models are built based on state of charge (SOC) and charge / discharge power; the thermal inertia dynamics of refrigeration systems are divided into slow time scales of hours, and their dynamic models are built based on thermodynamic differential equations.

[0026] Fast timescales are in the seconds, dynamic power consumption response of IT devices The state equation is: , in, For CPU utilization, For chip temperature, It is Gaussian white noise. and These are the impact coefficients of CPU utilization and chip temperature on the power consumption of IT equipment, respectively, measuring the contribution of each to power consumption.

[0027] For medium timescales on the order of minutes, the expression for the charging and discharging dynamics of an energy storage system is: , in, for t The state of stored charge at all times, , The charging power and discharging power at time t are respectively. and These are charging efficiency and discharging efficiency, respectively. C ESS For energy storage capacity, and They represent t The charging and discharging power of the energy storage system at any given time. Constraints on the energy storage charge state include energy storage state boundaries and power limits.

[0028] The thermodynamic differential equation for the thermal inertia dynamics of the refrigeration system, on a slow timescale of hours, is expressed as follows: , in, Q IT To generate heat for IT equipment For cooling capacity, For heat capacity, This refers to the indoor temperature.

[0029] Temporal decoupling coding technique is used to separate cross-scale coupling terms between models at different time scales. The expression is: , in, This is the feature extraction matrix for this scale. This is a cross-scale interference suppression matrix. For inter-scale propagation delay, These are the state variables for this scale; For other scales of state variables; For time indexing.

[0030] In S2, resource clusters are obtained through dynamic clustering of resources using state potential game theory. The specific steps include: Define a cluster potential function based on the maximum adjustment capacity of a node and the communication latency. The expression is: , in, For nodes i Maximum adjustment capacity For communication delay, , These are the weighting coefficients. C For resource clusters, For nodes i The capacity of virtual power plants that are currently in use within the current scheduling.

[0031] Based on the cluster potential function, the optimal resource cluster partitioning scheme is obtained by solving the Nash equilibrium, which maximizes the overall adjustment capability of nodes within each cluster under communication delay constraints. The expression for solving the optimal cluster partitioning using Nash equilibrium is: , in, This represents the maximum aggregate capacity of the cluster. For nodes i Used capacity.

[0032] The Bayesian update method is used to dynamically adjust the margin parameter based on the deviation between the actual adjustment capacity and the predicted capacity of the resource cluster. The expression for the Bayesian update method is: , in, For the updated cluster C margin parameters, For the cluster before the update C margin parameters, γ For learning rate, and Clusters C The actual regulating capacity and the predicted regulating capacity; By introducing the bibru bar chance constraint, an elastic capacity boundary considering the influence of uncertainty is constructed. The bibru bar chance constraint uses the Wasserstein distance to measure the deviation of the uncertainty distribution and transforms the constraint conditions into a solvable form. The expression for the elastic capacity boundary is: , in, ϵ The confidence level of the robustness constraint. This is the required capacity.

[0033] The expression for the transformed constraint condition is: , in, , Clusters C Capacity mean and standard deviation.

[0034] like Figure 3 The diagram shows a hierarchical distributed collaborative optimization architecture. This architecture includes a global optimization layer, a local decision-making layer, and an elastic feedback channel connecting the global optimization layer and the local decision-making layer. The global optimization layer operates on a first preset time scale, while the local decision-making layer operates on a second preset time scale, shorter than the first. The global optimization layer coordinates electricity market, carbon market, and computing power demand on an hourly time scale, generating a day-ahead baseline strategy through an improved Lagrange dual relaxation method and dynamically allocating green electricity quotas using a carbon flow tracking model. The local decision-making layer performs real-time control at a minute-to-second time granularity, employing a federated reinforcement learning framework to achieve privacy-preserving distributed optimization and reducing communication data volume through an event-triggered mechanism. The elastic feedback channel, acting as a cross-layer collaborative link, triggers global strategy correction and local parameter updates when wind and solar power output suddenly drops or computing power default risk exceeds limits, ensuring robustness in extreme scenarios. The core objective of this hierarchical distributed collaborative optimization architecture is to break through the efficiency bottleneck of traditional centralized optimization, balancing market returns, low-carbon goals, and service quality (SLA) constraints, providing dynamic resource support for high-proportion renewable energy grid integration.

[0035] The global optimization layer constructs a multi-market joint optimization model based on Mixed Integer Programming (MIP). Inputs include electricity market information such as spot price forecasts, carbon market information such as wind and solar power output, and load forecasts. The optimization objective is to minimize the total cost, which includes electricity purchase cost, carbon cost, and computing power service quality default cost. The model prioritizes green energy sources, namely wind and solar power and energy storage systems, with diesel generators serving only as backups. An improved ADMM algorithm decomposes the problem into grid interaction and carbon quota issues at the transmission layer, node voltage balancing at the distribution layer, and computing power migration at the data center layer. Combined with sparse matrix compression and parallel computing, node optimization time is reduced. The output is the day-ahead baseline strategy, including BESS charging and discharging plans, IDC load allocation, and diesel generator start-up and shutdown timings.

[0036] Specifically, the operations performed by the global optimization layer include: Based on electricity market information, carbon market information, and computing power demand information, a multi-market joint optimization model is constructed. The objective function of the multi-market joint optimization model is to minimize the total cost, which includes electricity purchase cost, carbon cost, and computing power service quality default cost; the expression of the objective function is: , in, for t Time-of-use electricity price To purchase external power, for t Carbon emission costs per period Carbon emissions, calculated using a carbon flow tracking model. for t Time-based computing power service quality breach rate per unit SLA violation The number of service quality breaches for computing power is the number of tasks.

[0037] The constraints of the multi-market joint optimization model include electricity balance constraints and carbon quota constraints. The expression for the electricity balance constraint is as follows: , in, P IT For IT equipment power, P cool For the power of the refrigeration equipment, P PV Contribute to photovoltaic power P ESS For energy storage charging and discharging power, This refers to the net electrical power purchased or sold by the virtual power plant from the external main power grid.

[0038] carbon quotas The constraint is based on a carbon flow tracking model to dynamically allocate green electricity quotas, expressed as follows: , in, Green electricity quotas allocated to the government.

[0039] The adaptive alternating direction multiplier method (AT-ADMM) is used to solve the problem and generate a day-ahead baseline strategy that includes energy storage charging and discharging plans, IT load allocation, and backup power start-up and shutdown timing.

[0040] The lower layer is the local decision-making layer, deploying a federated reinforcement learning controller group. Edge-based IDC, BESS, and diesel generators generate second-level control commands based on real-time data such as SOC, electricity price, and SLA. Specifically, the BESS power adjustment ΔP is uploaded to the cloud via gradient encryption to aggregate and update model weights, reducing single-decision latency. The dynamic elastic resource adaptation algorithm adjusts its strategy according to the disturbance level: normal fluctuations (ΔP≤5%) execute the baseline; moderate disturbances (5%<ΔP≤15%) initiate online federated learning correction; extreme events (ΔP>15%) trigger the elastic feedback channel, calling up diesel for backup across layers to reduce fuel costs.

[0041] Specifically, the operations performed by the local decision-making level include: A federated reinforcement learning controller group is deployed, and each edge device generates control commands based on real-time operational data. The update expression for the policy network parameters of the federated reinforcement learning controller group is as follows: , in, θ i For nodes i The strategy network parameters; η The learning rate controls the step size for updating the network parameters of the policy. R i The local reward function includes energy cost and QoS metrics; γ is the discount factor. For nodes i The current state of the system environment as observed at the present moment; For nodes i The policy network is based on the state The decision-making instructions made; For nodes i The observed updated system environment status; In the new state Next, based on the current policy network, calculate the next action to be executed; For iterative indexing; The gradient information of the local policy network is uploaded to the cloud for aggregation using gradient encryption to update the global model weights; The decision-making strategy is dynamically adjusted based on the system disturbance level, which is determined by real-time power deviation.

[0042] The elastic feedback channel operates based on an event-triggered mechanism. When the predicted deviation of wind and solar power output exceeds the first preset threshold or the risk of default on computing power service quality exceeds the second preset threshold, the elastic feedback channel is triggered. After triggering, the local decision-making layer sends an interruption request to the global optimization layer, which then initiates replanning and sends lightweight correction instructions downwards.

[0043] The event trigger expression for the elastic feedback channel is: , in, This represents the predicted value of locally distributed renewable energy within the virtual power plant. Δ represents the actual measured total output of locally distributed renewable energy within the virtual power plant. P th The power deviation threshold, i.e. the first preset threshold, is used to dynamically adjust the communication frequency to reduce redundant data by 87%. λ e ( t The volatility is the risk of default on computing power service quality, used to determine whether the system state has changed significantly.

[0044] In this embodiment, when the wind and solar power prediction deviation is ≥15% or the SLA default risk is ≥10%, a local interruption request is sent to the global layer, triggering global model replanning, increasing the BESS discharge priority, and issuing a lightweight correction command (completed within 50ms). For example, when the wind and solar power suddenly drops by 40% at 14:00, the edge BESS response latency is shortened from 8 seconds to 2 seconds, the power deviation is reduced from 12% to 3%, and the forced start of the diesel generator is avoided for 30 minutes, significantly improving economy and reliability.

[0045] The layered distributed collaborative optimization architecture achieves breakthroughs in multi-dimensional performance through a three-layer collaborative mechanism of "global coordination, local agility, and elastic fault tolerance." In terms of efficiency and real-time performance, the spatiotemporal decoupling strategy significantly improves the optimization speed of large-scale nodes, the federated learning framework supports second-level dynamic decision-making, and the event-triggered mechanism greatly reduces redundant communication. A balance between privacy and efficiency is achieved through gradient encryption and sparse matrix technology, ensuring data security while reducing memory resource consumption. Regarding robustness and economy, the system maintains stable operation under extreme fluctuations, effectively controls energy storage state deviations, significantly improves green energy utilization, and achieves outstanding carbon emission reduction. The rapid collaborative response of energy storage and backup power can promptly fill power gaps, and the computing power task migration strategy simultaneously ensures service quality and low-carbon goals. This constructs a new paradigm for the collaborative optimization of market revenue, low-carbon constraints, and computing power guarantees for virtual power plants, enhancing the system's economy and operational resilience while improving the capacity for renewable energy absorption.

[0046] When solving the problem using the method in this embodiment, the following method is used: Figure 4 The solution is obtained through the illustrated process. A spatiotemporal decoupling acceleration strategy is employed to decouple the long-term day-ahead planning based on mixed-integer programming from the short-term real-time control using model predictive control, avoiding the dimensionality explosion problem caused by cross-scale iteration. Simultaneously, a sharpness-aware minimization algorithm is combined to prioritize the optimization of high-sensitivity parameters during reinforcement learning policy updates, significantly improving training efficiency. Secondly, a distributed parallel computing architecture is constructed, relying on containerized cluster technology to decompose and elastically scale the global optimization problem, and sparse matrix storage technology is used to compress the state transition matrix, greatly reducing computational resource overhead. Furthermore, a lightweight iteration strategy uses feasible region approximation theory to approximate the non-convex constraint space with a convex envelope, reducing the number of iterations while ensuring optimization accuracy; simultaneously, a lightweight edge-side inference engine is deployed, combining operator fusion and quantization compression techniques to achieve millisecond-level local decision response. Through the synergy of these technologies, the system balances high-precision control and real-time computing capabilities in complex scenarios, providing an efficient solution for multi-scale resource optimization.

[0047] Specifically, in the time-scale decoupling, the long-term timescale (day-ahead planning) and the short-term timescale (implementation of control) are decoupled, employing mixed-integer programming and model predictive control methods respectively. In the day-ahead planning, the timescale is 24 hours, and a mixed-integer programming model is constructed with economic efficiency as the objective: , in, C grid For grid electricity price, P grid For the power purchase capacity, C start For equipment start-up and shutdown costs, x ( t () is a binary variable representing the device status.

[0048] In real-time control, based on MPC rolling optimization, the optimization objective is: , in, The target power in the day-ahead baseline strategy issued by the global optimization layer. For the aggregated power of controllable resources, constraints include the power balance equation ∑ P res = P demand , Q This is the weight matrix. H To predict the time domain, P demandThis refers to the local total load demand that the virtual power plant needs to meet within the current rolling optimization window.

[0049] In spatial scale decoupling, for multi-region virtual power plants, the distributed alternating direction multiplier method (ADMM) is used to decompose the global optimization problem: , in, To alternately update local variables, This is a preset threshold.

[0050] Based on resource geographic distribution and communication topology, the global optimization problem is divided into M independent subtasks and deployed on a Kubernetes cluster to achieve parallel processing and efficient resource utilization. Parallel solutions to each subtask are obtained by synchronously updating intermediate variables through lightweight containers, reducing cross-node communication latency. The state transition matrix is ​​stored in a compressed sparse row format.

[0051] By using the feasible region approximation theory, a convex envelope is constructed to approximate the original feasible region for the non-convex constrained space. Through sampling and convex combination generation, 90% optimization accuracy is guaranteed while reducing the number of iterations by 75%.

[0052] To improve the real-time performance and efficiency of local decision-making, this embodiment deploys a lightweight inference mechanism at the edge. Specifically, a quantized and compressed inference engine is integrated into the local controller. Through operator fusion technology, convolutional layers and activation functions are integrated into a single computational unit, effectively reducing memory accesses and thus lowering computational resource consumption. Simultaneously, low-bit quantization technology is used to compress the original 32-bit floating-point parameters into 8-bit fixed-point numbers, shortening inference latency. This lightweight edge-side inference design not only improves data processing speed but also ensures fast and accurate decision-making even in resource-constrained environments, providing strong technical support for data centers to participate in the real-time control of virtual power plants.

[0053] Execute the updated real-time control commands obtained from S5 to obtain the final resource scheduling results of the virtual power plant. The real-time control commands include outputting the charging and discharging power commands of the energy storage system, the load adjustment commands of the IT equipment, the start and stop commands of the backup power supply, and the cooling power adjustment commands.

[0054] This embodiment is based on a simulation of a virtual power plant. Table 1 shows the internal resource parameters of the virtual power plant after data expansion under a spot electricity market environment, including resource parameters, forecast data, and market parameters. Table 2 shows the key operating parameters and market rules related to the State of Charge (SOC) of the Battery Energy Storage System (BESS) in this embodiment, including charge / discharge efficiency, initial SOC, intraday market penalty coefficient, and SOC operating range. Table 3 shows the specific equipment parameters and cost coefficients of the Internet Data Center (IDC), BESS, and diesel generator set in this embodiment. Figure 5 The wind and solar power output and load forecast curves for VPP during a typical day are used as input data for the global optimization layer to formulate day-ahead baseline scheduling strategies. Figure 6 This is the wiring diagram of the IEEE 33-node distribution system in this embodiment. This test network is used to verify the effectiveness and adaptability of the method in this embodiment under multi-node, complex power grid topology.

[0055] Table 1. Internal Resource Parameters of the Virtual Power Plant Table 2 Operating parameters and market rules for the state of charge of energy storage systems Table 3 Equipment parameters and cost coefficients for Internet data centers, energy storage systems, and diesel generator sets. Depending on the charging and discharging state, the BESS can function as either a load or a power source at different times. When the BESS is in charging mode, it absorbs power from the external power grid, similar to a user role; while when the BESS is in discharging mode, the power stored within it is transmitted to the external power grid, equivalent to a generator role. This embodiment simulates and analyzes the following four scenarios: Scenario 1, VPP operation without considering the BESS, IDC, and diesel generator; Scenario 2, considering the BESS participating in VPP operation; Scenario 3, considering the IDC and internal diesel generator participating in VPP operation; Scenario 4, simultaneously considering the VPP operation involving the BESS, IDC, and diesel generator. The key optimization indicators for each scenario in the simulation optimization are shown in Table 4. Figure 7 The resource power allocation and system power balance results for scenario 4 time-segment optimization specifically characterize the power output of each distributed resource (including energy storage system, adjustable load of data center and diesel generator) and the relationship with system power balance. Figure 8 The state of charge and cost results for scenario 4 time-segment optimization specifically characterize the dynamic changes of key state parameters (including the state of charge (SOC) of the energy storage system, node voltage, and market electricity price) and system operation indicators.

[0056] Table 4 Comparison of Key Optimization Indicators for Each Scenario Under the collaborative optimization scenario, the system achieved a comprehensive improvement in economy, low carbon emissions, and resource efficiency. In terms of economy, the total cost was reduced by 21.4% compared to the baseline scenario (saving RMB 267,700), mainly through the following measures: The energy storage system (BESS) adopted a peak-valley charging and peak-discharging strategy, charging 12.4MW during off-peak hours (0:00-6:00) and discharging 18.2MW during peak hours (10:00-14:00), effectively reducing the demand for high-priced electricity; the IDC further reduced peak-hour electricity purchases by dynamically adjusting the 4.5MW computing power load to match peak wind and solar output; the diesel generator was only used as a backup power source, only activated briefly when wind and solar output dropped sharply (5.1MW was activated at 14:00), reducing routine operating costs and carbon emissions. In terms of low-carbon benefits, total carbon emissions were reduced by 28.9% (107.6 tons), mainly due to the replacement of diesel power generation with BESS at 26.7 MWh, and the improvement in wind and solar power utilization (wind and solar curtailment rate decreased from 8.7% to 3.8%, and green electricity usage increased by 35.6 MWh). Resource synergy efficiency was significantly optimized: BESS charging and discharging efficiency remained stable at 90%, with the state of charge (SOC) dynamically adjusted between 20% and 90%, and the daily market plan deviation consistently below 5%; the adjustable load of the IDC (accounting for 20% of the total load) could flexibly adjust 4.5 MW within 15 minutes, accurately tracking wind and solar fluctuations; diesel generators started when wind and solar output was 30% lower than the predicted value, strictly adhering to minimum operating time constraints (≥1 hour). These results validate the key role of multi-resource synergy control in cost reduction, carbon reduction, and system resilience improvement in this embodiment.

[0057] Table 5 illustrates the performance improvement effect of the spatiotemporal decoupling mechanism during peak power system periods. It shows that traditional centralized optimization methods suffer from high computation time and large memory consumption due to single-threaded dense matrix calculations, stemming from the matrix dimension explosion problem caused by global variable coupling. The spatiotemporal decoupling mechanism proposed in this embodiment uses an improved ADMM algorithm to decouple and parallelize the optimization tasks of the transmission layer, distribution network layer, and data center layer, and utilizes CSR format sparse matrix compression technology to significantly reduce computational resource consumption. Simultaneously, addressing the issues of large prediction deviations and lag in strategy updates for wind and solar power output, an MPC rolling compensation mechanism and event triggering rules are introduced, suppressing prediction deviations to 2.3% and increasing response frequency by 300%. For example, when photovoltaic output suddenly drops by 12% due to cloud cover, the system completes three rounds of strategy iteration updates within 30 seconds, avoiding a 4.5MW power shortfall caused by traditional fixed-cycle updates. The spatiotemporal decoupling mechanism is implemented in a layered and collaborative manner: in the time dimension, a day-ahead baseline strategy (power output of thermal power units and electric vehicle charging baseline) is generated through the upper-level MIP model, while in the spatial dimension, the load of distribution network nodes is dynamically scheduled based on the distributed MPC framework. For example, during peak hours, by adjusting the priority of electric vehicle charging piles, the load rate is dynamically transferred from overloaded nodes with a load of 98% to low-load nodes, achieving efficient resource utilization and safe system operation.

[0058] Table 5 Comparison of optimization efficiency under the spatiotemporal decoupling mechanism (taking the 10:00 peak period as an example) In the extreme scenario of a 40% drop in wind and solar power output, as shown in Table 6, the dynamic elastic resource adaptation mechanism of the hierarchical collaborative architecture in this implementation demonstrates high-efficiency response capabilities. When the wind and solar power output prediction deviation exceeds the 15% threshold (corresponding to a power gap of 8.2MW), the local decision loop detects the anomaly in real time through the federated reinforcement learning controller and activates the cross-layer collaborative channel: the local layer sends a correction request to the global optimization layer (delay <50ms), triggering the rolling replanning of the global model. The global layer updates the carbon-electricity joint optimization strategy based on the latest measured data, prioritizing the use of highly elastic resources—the battery energy storage system (BESS) discharge power is increased from 12MW to 20MW (SOC decreases from 65% to 35%), while the IDC releases 4.3MW of adjustable load by delaying non-urgent computing tasks. The above collaborative mechanism achieves dynamic compensation for an 8.5MW power gap in the extreme event at 14:00, verifying the elastic resource adaptation capability of the "local fast response - global optimization correction" under the hierarchical architecture of this embodiment.

[0059] Table 6. Dynamic Elastic Resource Adaptation Effect (Scenario of Sudden Drop in Wind and Light at 14:00) Table 7 shows a comparison of resource adaptation effects. This embodiment achieves efficient resource adaptation and multi-objective optimization in the scenario of sudden wind and solar power downtime through technological innovation. BESS adopts sparse matrix compression technology (CSR format) to achieve second-level issuance of charging and discharging commands, increasing the SOC release rate to 4.5% / minute and eliminating power tracking lag caused by traditional communication delays; IDC migrates non-real-time tasks based on dynamic resource allocation technology, releasing 4.3MW of adjustable load and completing cross-node scheduling under privacy protection through federated learning controller; diesel generator backup optimization requires only a single 15MW unit through precise start-stop strategy, saving fuel costs of RMB 27,000 / hour. In terms of economics, diesel generator fuel costs are reduced by 57.1%, and combined with the collaborative scheduling of IDC and BESS, the cost per kilowatt-hour is reduced from RMB 0.48 / kWh to RMB 0.39 / kWh; in terms of reliability, node voltage fluctuations are strictly controlled within ±2%, and the SLA default rate is reduced to 1.2%, ensuring the continuity of high-priority computing power. In terms of technical support, federated reinforcement learning combined with 8-bit quantization and gradient encryption achieves privacy protection, the dynamic elasticity coefficient algorithm optimizes resource allocation weights in real time, and the cross-layer collaborative protocol achieves 50ms-level policy synchronization through lightweight communication (data packets <1.2KB). Addressing issues such as SOC alarm (20%) caused by high-power discharge of BESS and the IDC adjustment limit (10MW), improvements are proposed, including introducing hydrogen energy storage to supplement short-term gaps and deploying edge nodes to enhance IDC elasticity to 15MW. The conclusions show that this mechanism, through cross-layer collaboration and rapid response technology, simultaneously achieves power balance, cost reduction, and reliability assurance in extreme scenarios, providing key technical support for high-proportion renewable energy power systems.

[0060] Table 7 Comparison of Resource Adaptation Effects Table 8 shows the effect analysis of green computing power scheduling through carbon-electricity-computing power coupling. Coordinated optimization of carbon-electricity-computing power significantly improves energy efficiency and sustainability. In terms of energy efficiency, the supercomputing center uses liquid cooling technology and dynamic voltage regulation algorithms to optimize PUE from 1.5 to 1.15, increasing the power density per rack to 30kW and reducing energy consumption per unit of computing power by 42%. Combined with model compression and in-memory computing technologies, the energy consumption of AI training tasks decreased from 1.2 kWh / TOPS to 0.68 kWh / TOPS, and redundant computing energy consumption decreased by 30%. Regarding carbon emission control, the IDC has increased its carbon emissions to 58% through cross-regional green electricity trading, driving a reduction of 16.34 million tons of annual carbon emissions from data centers (equivalent to the emission reduction of 15 coal-fired power plants). Specifically, the carbon emission intensity per unit of computing power at nodes decreased to 0.35 kgCO2 / kWh, and the carbon emission intensity of the IDC decreased by 31%. In terms of economics, by leveraging peak-valley electricity price arbitrage (nighttime electricity price of 0.28 yuan / kWh) and direct green electricity supply, the operating cost of computing power decreased by 19%. The IDC achieved annual revenue of 120 million yuan through green certificate carbon trading, increasing the profit margin by 8%. In terms of resource synergy, liquid cooling waste heat recovery improved the energy efficiency of park heating by 25%, saving 18,000 tons of coal annually. During wind and solar power fluctuations, BESS energy storage and dynamic adjustment of task priorities kept the interruption rate of high-value tasks (medical image processing) below 0.1%, achieving a multi-dimensional win-win situation in energy efficiency, environmental protection, and economic benefits.

[0061] Table 8 Analysis of Green Computing Power Scheduling Effect Example 2 This embodiment is a data center resource scheduling system for virtual power plants constructed using the method of Embodiment 1, including a resource modeling and aggregation unit, a hierarchical optimization and decision-making unit, an elastic feedback and replanning module, and a scheduling execution unit.

[0062] The resource modeling and aggregation unit is used to divide IT equipment, energy storage systems and cooling system resources with different response time constants in the data center into multiple time scales and decouple their dynamic characteristics to form a decoupled resource model; and based on the decoupled resource model, to divide out highly collaborative resource clusters and define the elastic capacity boundary of the resource clusters.

[0063] The hierarchical optimization and decision-making unit is used to solve the multi-market collaborative optimization model based on the hierarchical distributed collaborative optimization architecture at the global optimization layer with a first preset time scale, based on power information, carbon information and computing power demand information, and generate a day-ahead baseline strategy including energy storage charging and discharging plan and IT load allocation plan; at the local decision-making layer with a second preset time scale, based on the day-ahead baseline strategy and real-time operation data, a federated reinforcement learning framework is used to make distributed decisions and generate real-time control instructions.

[0064] The elastic feedback and replanning unit is used to monitor the system's operating status through an event-triggered mechanism. When a triggering event is detected, the global optimization layer is triggered to replan the day-ahead baseline strategy, and the replanning result is sent to the local decision-making layer to update the real-time control instructions.

[0065] The scheduling execution unit is used to execute updated real-time control instructions and output the final resource scheduling results of the virtual power plant.

[0066] This embodiment proposes a spatiotemporal elastic resource dynamic aggregation model for data centers oriented towards virtual power plants. Through a three-layer modeling approach of "characteristic decoupling - dynamic clustering - elastic calibration," this model successfully addresses the significant differences in resource time constants within the data center, achieving dynamic characteristic decoupling and efficient aggregation across time scales. This innovation not only reduces cross-scale coupling errors but also quantifies the heterogeneous impact of photovoltaic power output fluctuations and sudden changes in task load by introducing bibloc bar chance constraints, significantly enhancing the regulation potential of data centers as "virtual energy storage."

[0067] Furthermore, a layered distributed collaborative optimization architecture, characterized by "double-layer nesting and event triggering," is constructed. This architecture comprises three parts: a global optimization layer, a local decision-making layer, and an elastic feedback channel. The global optimization layer coordinates the electricity market, carbon market, and computing power demand on an hourly time scale, achieving multi-market collaborative optimization. The local decision-making layer executes real-time control at a minute-to-second time granularity, employing a federated reinforcement learning framework to achieve distributed optimization with privacy protection. The elastic feedback channel, acting as a cross-layer collaborative link, triggers global policy correction and local parameter updates in extreme scenarios, ensuring the system's robustness and response speed. This effectively breaks through the efficiency bottleneck of traditional centralized optimization, achieving significant performance improvements across multiple dimensions.

[0068] To further improve optimization efficiency and real-time performance, this embodiment proposes a spatiotemporal decoupling optimization mechanism and a lightweight iteration strategy. By decoupling time scales and space scales, the long-term day-ahead planning and short-term real-time control are decoupled in a hierarchical manner, avoiding the dimensionality explosion problem caused by cross-scale iteration. A distributed alternating direction multiplier method is used to decompose the global optimization problem, enabling parallel solution across multiple regions. Through the application of feasible region approximation theory and a lightweight edge-side inference engine, the number of iterations and local decision latency are significantly reduced. This mechanism ensures optimization accuracy while improving the system's computational efficiency and real-time response capability, providing strong support for the efficient scheduling of large-scale data center resources.

[0069] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0070] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0071] Multiple components in the device are connected to an I / O interface, including: input units such as a keyboard, mouse, etc.; output units such as various types of displays, speakers, etc.; storage units such as disks, optical disks, etc.; and communication units such as network interface cards, modems, wireless transceivers, etc. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks. The processing unit performs the various methods and processes described above, such as the method of the present invention. For example, in some embodiments, the method of the present invention may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or the communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the method of the present invention described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute the method of the present invention by any other suitable means (e.g., by means of firmware).

[0072] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0073] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0074] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0075] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data center resource scheduling method for a virtual power plant, characterized in that, The specific steps include: S1, multi-time scale division and dynamic characteristic decoupling of IT equipment, energy storage system and refrigeration system resources with different response time constants in the data center are performed to form a decoupled resource model; S2, based on the decoupled resource model, a resource cluster with high synergy is divided, and the elastic capacity boundary of the resource cluster is calibrated; S3, based on a hierarchical distributed collaborative optimization architecture, a multi-market collaborative optimization model is solved at a first preset time scale in the global optimization layer according to power information, carbon information and computing power demand information, and a day-ahead baseline strategy including energy storage charging and discharging plan and IT load allocation plan is generated; S4, in the local decision layer, a real-time regulation instruction is generated by using a federal reinforcement learning framework for distributed decision-making based on the day-ahead baseline strategy and real-time operation data at a second preset time scale; and S5, the system operation state is monitored through an event triggering mechanism, when a triggering event is monitored, the global optimization layer is triggered to re-plan the day-ahead baseline strategy, and the re-planned result is issued to the local decision layer to update the real-time regulation instruction. 2.The data center resource scheduling method for virtual power plant according to claim 1, wherein, The specific steps of the multi-time scale division and dynamic characteristic decoupling in S1 include: The IT equipment is divided into a fast time scale of seconds, and the dynamic model of the IT equipment is constructed based on CPU utilization and chip temperature; the energy storage system is divided into a medium time scale of minutes, and the dynamic model of the energy storage system is constructed based on state of charge SOC and charging and discharging power; the thermal inertia dynamics of the refrigeration system is divided into a slow time scale of hours, and the dynamic model of the refrigeration system is constructed based on thermodynamic differential equations; The time-domain decoupling coding technique is adopted to separate the cross-scale coupling terms between different time-scale models The expression is: ， wherein, is a local scale feature extraction matrix, is a cross scale interference suppression matrix, is an inter-scale propagation delay, is a local scale state variable; is a cross scale state variable; is a time index.

3. The data center resource scheduling method for virtual power plant according to claim 1, wherein, In S2, the resource cluster is obtained by resource dynamic clustering through state potential game theory, and the specific steps include: A cluster potential function is defined based on the maximum regulation capacity of a node and communication time delay, and the cluster potential function The expression is: ， wherein, is a node i maximum regulation capacity, is a communication delay, , is a weight coefficient, C is a resource cluster, is a node i capacity of the virtual power plant regulation that has been occupied within the current schedule; According to the cluster potential function, the optimal resource cluster division scheme is obtained by solving the Nash equilibrium, so that the total adjustment capacity of the nodes in each cluster reaches the optimum under the communication time delay constraint; A Bayesian updating method is used to dynamically correct the margin parameter according to the deviation between the actual adjustment capacity and the predicted capacity of the resource cluster, and the expression of the Bayesian updating method is: ， wherein, is the margin parameter of the updated cluster C , is the margin parameter of the cluster C before the update, γ is the learning rate, and are the actual and predicted regulation capacities of the cluster C , respectively. A distribution robust chance constraint is introduced to construct an elastic capacity boundary considering the influence of uncertainty, and the distribution robust chance constraint measures the deviation of the uncertainty distribution through the Wasserstein distance and converts the constraint condition into a solvable form.

4. The data center resource scheduling method for a virtual power plant according to claim 1, wherein, In S3, the hierarchical distributed collaborative optimization architecture includes a global optimization layer, a local decision layer and an elastic feedback channel connecting the global optimization layer and the local decision layer. The global optimization layer operates at a first preset time scale, and the local decision layer operates at a second preset time scale shorter than the first preset time scale.

5. The data center resource scheduling method for virtual power plant according to claim 4, wherein, The operations performed by the global optimization layer include: Based on power market information, carbon market information and computing power demand information, a multi-market joint optimization model is constructed; The objective function of the multi-market joint optimization model is to minimize the total cost, and the total cost includes power purchase cost, carbon cost and computing power service quality default cost; the constraint conditions of the multi-market joint optimization model include power balance constraint and carbon quota constraint, and the carbon quota constraint dynamically allocates green power quota based on a carbon flow tracking model; The adaptive alternating direction multiplier method (AT-ADMM) is used for solving, to generate a day-ahead baseline strategy including energy storage charging and discharging plans, IT load distribution, and standby power start-stop timing.

6. The data center resource scheduling method for virtual power plant according to claim 4, wherein, The operations performed by the local decision layer include: A federal reinforcement learning controller group is deployed, each edge end generates regulation and control instructions based on real-time operation data, and an update expression of a policy network parameter of the federal reinforcement learning controller group is: ， wherein, Theta i is a policy network parameter of the node i ; Eta is a learning rate controlling the step size of the policy network parameter update; R i is a local reward function including energy cost and QoS indicator; γ is a discount factor; is a system environment state observed at the current time instant by the node i ; is a decision instruction made by the policy network of the node i according to the state ; is an updated system environment state observed by the node i ; is a next action to be executed according to the current policy network at the new state ; is an iteration index; Gradient information of the local policy network is uploaded to the cloud end for aggregation in a gradient encryption mode, and the global model weight is updated; The decision strategy is dynamically adjusted according to the system disturbance level, and the disturbance level is divided based on real-time power deviation.

7. The data center resource scheduling method for virtual power plant according to claim 4, wherein, The elastic feedback channel is operated based on an event triggering mechanism, when it is monitored that the wind and light output prediction deviation exceeds a first preset threshold or the computing power service quality default risk exceeds a second preset threshold, the elastic feedback channel is triggered; after being triggered, the local decision layer sends an interruption request to the global optimization layer, the global optimization layer starts re-planning and sends lightweight correction instructions downward.

8. The data center resource scheduling method for virtual power plant according to claim 1, wherein, The updated real-time regulation and control instructions obtained by executing the S5 are executed to obtain the final resource scheduling result of the virtual power plant, and the real-time regulation and control instructions include charging and discharging power instructions of the output energy storage system, load adjustment instructions of the IT equipment, start-stop instructions of the standby power, and refrigeration power adjustment instructions.

9. An electronic device comprising a memory and a processor, said memory having stored thereon a computer program, characterized in that, The processor executes the program to implement the method of any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method of any one of claims 1-8.

Images