Data center virtual power plant coordination optimization system and method based on load aggregation

Abstract

The application relates to a data center virtual power plant coordination optimization system and method based on load aggregation, and the system comprises a load aggregation evaluation module, which is used for quantifying time-varying adjustable capacities of data center heterogeneous loads and energy storage devices by means of unsupervised learning and supervised regression methods by utilizing historical operation data; a real-time collaborative decision module, which is used for modeling a real-time coordination problem as a Markov decision process and outputting continuous control instructions by means of a deep reinforcement learning intelligent agent of a deep deterministic policy gradient; and a planning optimization module, which is used for establishing a double-target model of operation income and carbon emission and solving a Pareto optimal planning strategy by means of a non-dominated sorting genetic algorithm with an elite strategy. The application realizes end-to-end optimization of fine-grained load aggregation, real-time adaptive continuous decision and multi-target coordination, effectively improves operation economy of the data center virtual power plant and reduces carbon emission.

Description

Technical Field

[0001] This invention relates to the field of power resource dispatching, and in particular to a data center virtual power plant coordination and optimization system and method based on load aggregation. Background Technology

[0002] With the rapid growth of energy consumption in data centers, power system operators face the dual challenges of increased peak grid pressure and continuously rising carbon emissions. Virtual power plants (VPPs) offer a clever solution by integrating distributed energy resources into dispatchable entities. Due to the inherent flexibility of data centers, such as flexible IT load scheduling and energy storage, they become the ideal platform for VPP integration, thus forming the concept of data center virtual power plants (DC-VPPs).

[0003] However, coordinating DC-VPP faces a series of clear and difficult technical challenges. Data center loads are highly heterogeneous and uncertain, there is a conflict between operational efficiency and carbon reduction targets, and the system urgently needs fast, adaptive, real-time decision-making. Traditional optimization methods, such as mixed-integer linear programming (MILP) and model predictive control (MPC), suffer from large computational latency, high dependence on accurate predictions, and are easily affected by prediction errors. Most existing reinforcement learning (RL) applications are only geared towards a single optimization objective or discrete action space, making it difficult to effectively address the aforementioned complex needs.

[0004] Existing research can be summarized into three main areas. In data center energy management, technology has evolved from internal efficiency optimization to demand response, but mainstream MILP and MPC methods are sensitive to prediction errors. In VPP research, while various resource integration schemes have been explored, they generally oversimplify the flexibility of data centers, failing to fully reflect their fine-grained adjustment capabilities. In artificial intelligence applications, deep reinforcement learning (DRL) has shown good application potential in sequential decision-making in energy systems; however, existing research on DC-VPP often focuses on a single economic objective, neglecting explicit carbon emission coordination and failing to achieve joint optimization of economic efficiency and carbon reduction.

[0005] Therefore, there is an urgent need to provide a solution that can integrate multiple AI technologies and achieve end-to-end coordination of data center VPPs in order to overcome the aforementioned limitations of existing technologies. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a data center virtual power plant coordination and optimization system and method based on load aggregation. This solves or partially solves the problem that existing data center virtual power plant optimization methods are unable to simultaneously achieve fine-grained adjustable capacity assessment of heterogeneous loads, real-time collaborative decision-making of continuous action space, and multi-objective coordinated optimization of economic benefits and carbon emissions under real-time requirements.

[0007] The objective of this invention can be achieved through the following technical solutions: A first aspect of the present invention provides a data center virtual power plant coordination and optimization system based on load aggregation, comprising: The load aggregation assessment module is used to quantify the time-varying adjustable capacity of heterogeneous loads and energy storage devices in the data center by using historical operating data, through unsupervised learning and supervised regression, so as to obtain the total adjustable capacity of the data center virtual power plant. The real-time collaborative decision-making module is used to model the real-time coordination problem as a Markov decision process, interact with the simulated environment in the form of a deep reinforcement learning agent, learn the optimal real-time control strategy under dynamic power grid conditions, and output continuous control commands. The planning and optimization module is used to establish a multi-objective optimization model and use an evolutionary algorithm to solve the Pareto optimal planning strategy to obtain a scheduling plan that takes into account both operating costs and carbon emissions.

[0008] As a preferred technical solution, the load aggregation assessment module is used for: Divide the data center load into rigid loads, elastic loads, and energy storage devices; K-means clustering algorithm was used to perform cluster analysis on historical elastic load data to obtain several typical load patterns; For each cluster, an XGBoost regression model is used, trained based on feature parameters including time, job queues, environmental indicators and market signals, to predict the maximum schedulable power corresponding to that cluster. The total adjustable capacity of the data center virtual power plant is obtained by summing up the adjustable power predictions of all clusters.

[0009] As a preferred technical solution, in the Markov decision-making process constructed by the real-time collaborative decision-making module: the state space includes power grid signals and the internal state of the data center, the power grid signals include real-time electricity price, power grid carbon intensity and regional load, and the internal state of the data center includes job queue length and time information; The action space is a continuous control vector, including power purchase and sale from the grid and adjustment commands for flexible loads; A reward function is constructed that integrates revenue, energy cost, carbon emissions, and constraint violation penalties, and the agent is trained using a deep deterministic policy gradient algorithm to maximize the cumulative reward.

[0010] As a preferred technical solution, the reward function is: in, For the reward function, For electricity sales revenue, For operating costs, For carbon emission costs, Penalties for violating service quality or equipment constraints, to These are the weighting coefficients for each component.

[0011] As a preferred technical solution, the planning optimization module is used for: An optimization model with the dual objectives of maximizing total operating revenue and minimizing total carbon emissions was established, with an optimization period of 24 hours. The constraints included workload cutoff time, dynamic state of charge of the energy storage system, power limit of the cooling system, and power balance. The non-dominated sorting genetic algorithm NSGA-III with an elitist strategy is used to solve the bi-objective optimization model, and a Pareto optimal solution set reflecting the trade-off between operating revenue and carbon emissions is obtained.

[0012] Another aspect of the present invention provides a data center virtual power plant coordination and optimization method based on load aggregation, characterized in that it is applied to the aforementioned data center virtual power plant coordination and optimization system, and the method includes the following steps: S1 collects historical and real-time operational data from data centers, power grid market signals, and carbon intensity signals; S2, Quantify the real-time adjustability potential of the data center virtual power plant using the load aggregation assessment module; S3, obtain the real-time optimal control command using the real-time collaborative decision-making module; S4, use the planning optimization module to obtain the Pareto optimal scheduling plan for the future period; S5, execute the control commands and monitor system performance indicators.

[0013] As a preferred technical solution, S2 includes: K-means clustering algorithm is used to cluster historical elastic load data to obtain typical load patterns; For each cluster, the XGBoost regression model is used to predict the maximum adjustable power using time, job queue, environmental indicators and market signals as features. By summing the capacities of each cluster, the total adjustable capacity of the data center virtual power plant is obtained.

[0014] As a preferred technical solution, S3 includes: Construct a Markov decision process, with states including real-time electricity price, grid carbon intensity, regional load, data center job queue length and time, and actions including grid power purchase and sale and flexible load adjustment instructions; A reward function is constructed, and the agent is trained using a deep deterministic policy gradient algorithm to calculate the optimal control command in real time.

[0015] As a preferred technical solution, S4 includes: A dual-objective optimization model is constructed for a 24-hour cycle, with the objectives of maximizing total operating revenue and minimizing total carbon emissions. Constraints include workload cutoff time, dynamic changes in the state of charge of the energy storage system, power limits of the cooling system, and power balance. The NSGA-III algorithm is used to solve the bi-objective optimization model to obtain the Pareto optimal scheduling plan.

[0016] As a preferred technical solution, in step S5, the performance indicators include: average daily carbon emissions, total operating cost, average delay and 95th percentile delay of batch processing jobs, scheduling instruction calculation time, and target tracking error.

[0017] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Achieving accurate quantification of time-varying adjustable capacity: Existing technologies are coarse in quantifying the adjustable capacity of heterogeneous loads and are insufficient in characterizing time-varying characteristics. This invention adopts a load aggregation assessment module based on the combination of K-means clustering and XGBoost regression, which subdivides data center resources into rigid loads, elastic loads and energy storage devices. By clustering historical data to identify typical load patterns, and using multi-feature regression to predict the schedulable power of each pattern, the time-varying adjustable capacity is quantified in a fine-grained and high-precision manner, providing a reliable data foundation for subsequent optimization and overcoming the shortcomings of existing VPP research in simplifying the flexibility of data centers.

[0018] (2) Reduce computational latency: Traditional optimization methods have large computational latency, rely on accurate prediction and are susceptible to errors under real-time decision-making requirements. This invention models the real-time coordination problem as a Markov decision process and uses a deep reinforcement learning agent based on the deep deterministic policy gradient algorithm to solve it. The state space covers the power grid signal and internal state, and the action space is a continuous control vector. It can quickly output adaptive real-time control commands without relying on accurate prediction, thus reducing computational latency and improving the system's response to dynamic power grid conditions and uncertainties.

[0019] (3) Collaborative optimization of operating costs and carbon emissions: Existing DC-VPP optimizations are mostly based on a single economic objective and ignore carbon emission coordination. This invention designs a multi-weighted reward function that integrates electricity sales revenue, operating costs, carbon emission costs and constraint penalties in the real-time collaborative decision-making module, and establishes an optimization model with the dual objectives of maximizing total operating revenue and minimizing total carbon emissions in the planning optimization module. The NSGA-III algorithm is used to solve the Pareto optimal solution set, and the revenue and carbon emission reduction are collaboratively optimized at the two levels of real-time control and day-ahead planning, providing operators with a strategy for balancing efficiency and emission reduction.

[0020] (4) Forming an end-to-end AI integrated optimization framework: This invention breaks through the limitations of single AI technologies, organically integrating unsupervised learning, supervised regression, deep reinforcement learning, and multi-objective evolutionary optimization into a unified framework. From the adjustable capacity quantification of the load aggregation module to the real-time adaptive control of the real-time collaborative decision-making module, and then to the long-term strategic planning of the planning optimization module, a full-chain, end-to-end decision-making process is formed, significantly improving the intelligence level and overall operational efficiency of the data center virtual power plant.

[0021] (5) Effectively balancing Quality of Service (QoS) and power grid regulation requirements: This invention embeds considerations of QoS indicators such as job latency and equipment lifespan into the optimization objective by using the penalty term in the reward function and the workload deadline constraint in the planning model. Experimental results show that, compared with the single-objective DRL that simply pursues economy, the solution of this invention effectively controls the average latency and tail latency of batch processing jobs while achieving good economy and low carbon emissions, avoiding the impact on the normal operation performance of the data center due to excessive pursuit of regulation capabilities. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the data center virtual power plant coordination and optimization system based on load aggregation in the embodiment; Figure 2 This is a schematic diagram of the AI-enhanced DC-VPP optimization framework in the embodiment; Figure 3 This is a schematic diagram of the 24-hour running trajectory obtained in the example; Figure 4 This is a flowchart of the data center virtual power plant coordination optimization method based on load aggregation in the embodiment. Detailed Implementation

[0023] 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.

[0024] Example 1 To address the problems of the aforementioned existing technologies, this embodiment provides a data center virtual power plant coordination and optimization system based on load aggregation. It constructs a fine-grained load aggregation model, designs a DRL agent for real-time continuous control tasks, and establishes a dual-objective optimization model considering time-varying carbon emissions and operating costs. It utilizes a combination of unsupervised learning, supervised regression, DRL, and multi-objective evolutionary optimization to achieve end-to-end DC-VPP coordination. Its core components include: a fine-grained load aggregation model (corresponding to the load aggregation evaluation module), a DRL agent for real-time continuous control tasks (corresponding to the real-time collaborative decision-making module), and a dual-objective optimization model considering time-varying carbon emissions (corresponding to the planning optimization module).

[0025] See Figure 2 The AI-enhanced DC-VPP optimization framework provided in this embodiment organically and rigorously integrates data-driven flexibility models, continuous action DRL, and multi-objective carbon emission optimization into a unified framework, thus advancing the cutting-edge development in this field.

[0026] This embodiment naturally and rationally divides data center resources into three categories: rigid loads that are non-adjustable and subject to SLA constraints, flexible loads that are delay-tolerant and adjustable, and energy storage devices such as UPS batteries. Therefore, the method for handling the characteristics of flexible loads is very clear. First, K-means cluster analysis is performed on historical flexible load data to obtain the typical pattern Lk. Then, for each cluster, an XGBoost regression model is used to predict the maximum adjustable power. The selected features include time, job queues, environmental indicators, and market signals. Finally, the total adjustable capacity is the sum of the capacities of each cluster.

[0027] The real-time coordination problem is modeled as a Markov decision process. The selected state variables are grid signals and internal states, while the action variables are continuous vectors controlling power inflows and outflows and adjusting flexible loads. The reward function carefully considers factors such as economic benefits, energy costs, carbon emissions, and constraints / penalties. Grid signals can include electricity prices, carbon intensity, and load, while internal states can include queue length and time. Therefore, when using a Deep Deterministic Policy Gradient (DDPG) agent to solve for the optimal policy, the cumulative reward can be directly maximized.

[0028] The planning establishes a deterministic multi-objective optimization model with a 24-hour cycle, using maximizing total revenue and minimizing total carbon emissions as the two optimization objectives, while workload deadlines, dynamic changes in the energy storage system state (SOC), and cooling system limitations are all explicitly defined constraints. Therefore, this embodiment uses the Non-Dominated Sorting Genetic Algorithm III (NSGA-III) to solve the problem and obtain the Pareto front of the optimal trade-off strategy in the operational planning.

[0029] In summary, see Figure 1 The system includes the following components: (1) Load aggregation assessment module.

[0030] By combining historical operational data with unsupervised learning and supervised regression methods, the time-varying adjustable capacity of heterogeneous loads in data centers and energy storage devices can be objectively and reliably quantified.

[0031] In unsupervised learning clustering, it plays the role of an explorer. It requires no manual labeling and can automatically categorize various complex elastic load operation states of a data center in the past, such as peak and off-peak periods of batch processing tasks and daily fluctuations, into several typical categories. This solves the problem of not knowing which typical historical scenarios existed.

[0032] In supervised regression based on XGBoost, it acts as a predictor. After identifying the current typical mode, it utilizes more refined features, such as specific time and number of tasks, to accurately predict the maximum adjustable power within that mode. This solves the problem of accurately quantifying the regulating capacity within a known mode. The combination of the two enables refined and automated evaluation of data, patterns, and quantitative values, overcoming the shortcomings of traditional methods that rely on fixed formulas or crude assumptions.

[0033] This module specifically divides the data center load into rigid loads, elastic loads, and energy storage devices. It uses the K-means clustering algorithm to perform cluster analysis on historical elastic load data to obtain typical load patterns. For each cluster, an XGBoost regression model is trained based on characteristic parameters such as time, job queues, environment, and market signals to obtain the maximum schedulable power. Then, the schedulable power predictions of all clusters are summarized to obtain the total schedulable capacity of the data center virtual power plant.

[0034] (2) Real-time collaborative decision-making module.

[0035] The system uses a deep reinforcement learning agent to interact with a simulated environment and learn the optimal real-time control strategy under dynamic power grid conditions. The real-time coordination problem in the real-time collaborative decision-making module is modeled as a Markov decision process. First, the state space is defined as the power grid signal and the internal state of the data center, and the action space is a continuous control vector. Then, a reward function that takes into account comprehensive economic benefits, energy costs, carbon emissions, and constraint violation penalties is designed. Finally, the Deep Deterministic Policy Gradient (DDPG) algorithm is used to train the agent.

[0036] Specifically, the state variables in the Markov decision process include: real-time electricity price, grid carbon intensity, regional load, data center job queue length, and time information; the action variables include: grid power purchase and sale and adjustment commands for flexible loads.

[0037] Specifically, the reward function The specific form is as follows: in, For the reward function, For electricity sales revenue, For operating costs, For carbon emission costs, Penalties for violating service quality or equipment constraints, to These are the weighting coefficients for each component.

[0038] (3) Planning optimization module.

[0039] The Pareto optimal planning strategy is solved using a multi-objective optimization model and evolutionary algorithm.

[0040] The implementation process of the planning optimization module is as follows: First, a deterministic optimization model with the dual objectives of maximizing total operating revenue and minimizing total carbon emissions is established. The optimization period is 24 hours. The constraints of the selected model are the workload cutoff time, the dynamic state of charge of the energy storage system, the power limit and power balance of the cooling system. Then, the non-dominated sorting genetic algorithm (NSGA-III) with elitist strategy is used to solve the problem. Finally, the Pareto optimal solution set reflecting the trade-off between efficiency and emission reduction is obtained.

[0041] To verify the effectiveness of this method, a very clear and rigorous analysis was conducted on real-time minute-level data from 20 data center clusters over a year. This allowed for the natural and reasonable introduction of variables such as IT / cooling energy consumption, workload arrival, and UPS specifications, while also appropriately utilizing the regional load and time-varying carbon intensity contained in the grid data. The simulation system used OpenAIGym, PyTorch (DDPG), XGBoost, and pymoo (NSGA-III).

[0042] This embodiment provides a reasonable comparison between the improved artificial intelligence method and two benchmark methods: Model Predictive Control (MPC) with a prediction duration of 4 hours and Single Objective DRL (SO-DRL) trained solely from a profitability perspective, evaluating it in terms of daily carbon emissions, operating costs, average batch job latency / 95th percentile latency, and computational decision time.

[0043] Table 1 shows the results of each experiment during the 30-day test.

[0044] Table 1. Comparison of Overall Performance The AI-enhanced strategy in this embodiment strikes a balance between flexibility and latency performance, resulting in an average batch job latency of 3.1 hours, slightly higher than MPC's 2.8 hours but significantly better than SO-DRL's 5.2 hours. More importantly, its 95th percentile latency is 8.5 hours, slightly higher than MPC's 7.1 hours but far lower than SO-DRL's 14.3 hours. MPC is overly conservative in latency control, thus lacking long-term flexibility, while SO-DRL sacrifices carbon reduction targets to lower costs, leading to increased emissions. Therefore, the AI-enhanced strategy in this embodiment achieves a good balance between latency control, carbon reduction, and cost savings.

[0045] See Figure 3 The data represents the 24-hour operating trajectory, where (a) indicates the changes in electricity price and grid carbon intensity, (b) shows the battery system's state of charge (SOC), and (c) indicates net power input / output and flexible load adjustment. In this embodiment, the AI-enhanced agent operates during the early morning hours, approximately 2:00-6:00 AM, when electricity prices and carbon intensity are low, actively pre-charging the energy storage system. Towards evening, around 6:00-9:00 PM, when carbon-intensive peak-shaving units are used, causing both electricity prices and carbon intensity to rise simultaneously, coordinated regulation is implemented using two methods. One method is to reduce flexible load consumption by delaying the execution time of some batch operations; the other is to reduce its own net power consumption to the grid using intelligent discharge strategies. This allows coordinated regulation to be implemented precisely during the period when grid electricity prices and carbon intensity are at their highest, which is a key reason for the excellent cost-effectiveness and energy-saving and emission-reduction effects seen in Table 1.

[0046] In addition to daily operations, the framework naturally and logically uses modules optimized for dual objectives for strategic planning, thus naturally providing a representative daily Pareto front obtained using the NSGA-III method, i.e. Figure 3Each point represents a feasible operating strategy under both total cost and total carbon emissions indicators. The appropriate operating point is selected based on current policy priorities: point B is chosen when a moderate cost increase is acceptable in exchange for significant emissions reductions; points A or C are chosen when economic or environmental objectives are dominant.

[0047] In summary, addressing the limitations of existing reinforcement learning applications, which are often confined to discrete action spaces and unable to finely adjust continuous power commands, this embodiment employs the DDPG algorithm to directly output a continuous control vector including power purchase and sale commands and flexible load adjustment commands, thus overcoming the limitations of discrete actions on control precision. This achieves smoothing and adjustment of data center power inflows and outflows and load scheduling, better adapting to the continuous adjustment characteristics of actual physical equipment. Overall, through an integrated AI framework comprising load aggregation assessment, real-time collaborative decision-making, and planning optimization modules, unsupervised learning, supervised regression, deep reinforcement learning, and multi-objective evolutionary optimization are organically combined, forming a complete end-to-end solution from adjustable capacity quantification to real-time control and multi-objective planning. Simultaneously, it satisfies rapid adaptive decision-making, fine-grained load aggregation, and coordinated carbon emission optimization, effectively alleviating peak grid pressure and reducing the carbon footprint of data centers.

[0048] Example 2 Based on Example 1, this example provides a data center virtual power plant coordination optimization method based on load aggregation. See [link to example]. Figure 4 The method includes the following steps: Step S1: Collect historical and real-time operational data from the data center, as well as signals from the power grid market and carbon intensity.

[0049] Step S2: Use the load aggregation assessment module to quantify the real-time adjustability potential of the data center virtual power plant.

[0050] Specifically, a data-driven approach combining K-means clustering and XGBoost regression is used for adjustable capacity prediction.

[0051] Specifically, performance metrics include, but are not limited to: average daily carbon emissions, total operating costs, average latency and 95th percentile latency of batch processing jobs, scheduling instruction calculation time, and target tracking error.

[0052] Step S3: Use the real-time collaborative decision-making module to obtain the real-time optimal control command.

[0053] Specifically, a deep reinforcement learning agent based on the DDPG algorithm is used for online decision-making.

[0054] Step S4: Use the planning optimization module to obtain the Pareto optimal scheduling plan for the future period.

[0055] Specifically, the NSGA-III algorithm is used to solve the bi-objective optimization problem.

[0056] Step S5: Execute control commands and monitor system performance indicators.

[0057] 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 load-aggregation-based coordinated optimization system for data center virtual power plant, characterized in that, include: The load aggregation assessment module is used to quantify the time-varying adjustable capacity of heterogeneous loads and energy storage devices in the data center by using historical operating data, through unsupervised learning and supervised regression, so as to obtain the total adjustable capacity of the data center virtual power plant. The real-time collaborative decision-making module is used to model the real-time coordination problem as a Markov decision process, interact with the simulated environment in the form of a deep reinforcement learning agent, learn the optimal real-time control strategy under dynamic power grid conditions, and output continuous control commands. The planning and optimization module is used to establish a multi-objective optimization model and use an evolutionary algorithm to solve the Pareto optimal planning strategy to obtain a scheduling plan that takes into account both operating costs and carbon emissions. 2.The load aggregation based data center virtual power plant coordinated optimization system of claim 1, wherein, The load aggregation assessment module is used for: Divide the data center load into rigid loads, elastic loads, and energy storage devices; K-means clustering algorithm was used to perform cluster analysis on historical elastic load data to obtain several typical load patterns; For each cluster, an XGBoost regression model is used, trained based on feature parameters including time, job queues, environmental indicators and market signals, to predict the maximum schedulable power corresponding to that cluster. The total adjustable capacity of the data center virtual power plant is obtained by summing up the adjustable power predictions of all clusters. 3.The load aggregation based data center virtual power plant coordinated optimization system of claim 1, wherein, In the Markov decision-making process constructed by the real-time collaborative decision-making module, the state space includes power grid signals and the internal state of the data center. The power grid signals include real-time electricity price, power grid carbon intensity, and regional load. The internal state of the data center includes job queue length and time information. The action space is a continuous control vector, including power purchase and sale from the grid and adjustment commands for flexible loads; A reward function is constructed that integrates revenue, energy cost, carbon emissions, and constraint violation penalties, and the agent is trained using a deep deterministic policy gradient algorithm to maximize the cumulative reward.

4. The load aggregation based coordinated optimization system for data center virtual power plant according to claim 3, wherein, The reward function is: wherein, is a reward function, is a power selling revenue, is an operation cost, is a carbon emission cost, is a penalty term for violating quality of service or equipment constraints, to is a weight coefficient of each component.

5. The load aggregation based coordinated optimization system for data center virtual power plant according to claim 1, wherein, The planning optimization module is used for: An optimization model with the dual objectives of maximizing total operating revenue and minimizing total carbon emissions was established, with an optimization period of 24 hours. The constraints included workload cutoff time, dynamic state of charge of the energy storage system, power limit of the cooling system, and power balance. The non-dominated sorting genetic algorithm NSGA-III with an elitist strategy is used to solve the bi-objective optimization model, and a Pareto optimal solution set reflecting the trade-off between operating revenue and carbon emissions is obtained.

6. A load aggregation based coordination optimization method for data center virtual power plant, characterized in that, Applied to the data center virtual power plant coordination and optimization system as described in any one of claims 1-5, the method includes the following steps: S1 collects historical and real-time operational data from data centers, power grid market signals, and carbon intensity signals; S2, Quantify the real-time adjustability potential of the data center virtual power plant using the load aggregation assessment module; S3, obtain the real-time optimal control command using the real-time collaborative decision-making module; S4, use the planning optimization module to obtain the Pareto optimal scheduling plan for the future period; S5, execute the control commands and monitor system performance indicators.

7. The method of claim 6, wherein, S2 includes: K-means clustering algorithm is used to cluster historical elastic load data to obtain typical load patterns; For each cluster, the XGBoost regression model is used to predict the maximum adjustable power using time, job queue, environmental indicators and market signals as features. By summing the capacities of each cluster, the total adjustable capacity of the data center virtual power plant is obtained. 8.The load aggregation based coordination optimization method of data center virtual power plant according to claim 6, wherein, S3 includes: Construct a Markov decision process, with states including real-time electricity price, grid carbon intensity, regional load, data center job queue length and time, and actions including grid power purchase and sale and flexible load adjustment instructions; A reward function is constructed, and the agent is trained using a deep deterministic policy gradient algorithm to calculate the optimal control command in real time. 9.The load aggregation based coordination optimization method of data center virtual power plant according to claim 6, wherein, S4 includes: A dual-objective optimization model is constructed for a 24-hour cycle, with the objectives of maximizing total operating revenue and minimizing total carbon emissions. Constraints include workload cutoff time, dynamic changes in the state of charge of the energy storage system, power limits of the cooling system, and power balance. The NSGA-III algorithm is used to solve the bi-objective optimization model to obtain the Pareto optimal scheduling plan.

10. The method of claim 6, wherein, In S5, the performance indicators include: average daily carbon emissions, total operating cost, average delay and 95th percentile delay of batch processing jobs, scheduling instruction calculation time, and target tracking error.

Images