Virtual power plant-based computer computing collaborative integrated management method and system

Abstract

The invention provides a computing collaborative integrated management method and system based on a virtual power plant, and the method comprises the steps: obtaining a to-be-scheduled batch computing power task set, and marking each task in the to-be-scheduled batch computing power task set with a time window constraint, a resource demand spectrum, a mobility identifier and a task priority; based on prospective information provided by a preset digital twinning collaborative optimization layer, by taking multi-target optimization of total economic cost and total carbon emission of the system as guidance, constructing and solving a computing power task scheduling optimization model across spatial and temporal scales, and obtaining a global scheduling blueprint including optimal execution time, execution places and resource allocation of each task; and decomposing the global scheduling blueprint into a real-time scheduling instruction sequence, and issuing the real-time scheduling instruction sequence to a corresponding physical computing power node for execution to complete computing power scheduling work. According to the method, cooperative scheduling of the electric power and the computing power is completed, and the cooperative idea that the computing power moves along with the electric power or the electric power follows the computing power is achieved.

Description

Technical Field

[0001] This invention relates to the field of energy and information technology, and more specifically, to a computer-aided collaborative integrated management method and system based on a virtual power plant. Background Technology

[0002] With the explosive growth of the digital economy, the energy consumption of computing infrastructure such as data centers and intelligent computing centers has become a key load affecting the balance of the power system. Meanwhile, the volatility and intermittency of renewable energy sources, primarily wind and solar power, pose challenges to the stable operation of the power grid when connected to the grid on a large scale. Against this backdrop, promoting deep collaboration between the power grid and the computing network ("power-computing collaboration") is considered a crucial path to improve energy efficiency, ensure a green and low-carbon computing power supply, and promote the absorption of renewable energy.

[0003] Currently, the operation and scheduling of power systems and computing power systems are basically in a state of "each doing its own thing": On the power system side, although demand-side response and virtual power plant technologies have been promoted to aggregate adjustable industrial, commercial and residential loads, their adjustment targets are mainly traditional loads such as temperature control and lighting, and their adjustment potential, response accuracy and flexibility are limited.

[0004] On the computing power system side, resource scheduling within data centers primarily focuses on computing efficiency, service level agreements (SLAs), and internal energy consumption, treating its electricity usage as an uncontrollable, rigid load. Scheduling strategies are completely decoupled from real-time electricity prices, carbon intensity, and renewable energy output. This leads to computing loads potentially concentrating in periods and regions with high electricity prices and high carbon emissions, increasing operating costs and hindering the grid's absorption of renewable energy.

[0005] Therefore, promoting the deep integration and coordinated scheduling of the "power grid" and the "computing grid" (i.e., "power grid-computing collaboration") is considered the key to breaking the deadlock. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a computer-based collaborative integrated management method and system based on a virtual power plant.

[0007] To address the aforementioned problems, in a first aspect, the present invention provides a computer-aided collaborative integrated management method based on a virtual power plant, the method comprising: Obtain a set of batch computing power tasks to be scheduled, wherein each task in the set of batch computing power tasks to be scheduled is marked with a time window constraint, resource demand spectrum, portability identifier and task priority; Based on the forward-looking information provided by the pre-set digital twin collaborative optimization layer, and guided by the multi-objective optimization of total system economic cost and total carbon emissions, a computing power task scheduling optimization model across time and space scales is constructed and solved to obtain a global scheduling blueprint that includes the optimal execution time, execution location and resource allocation of each task. The global scheduling blueprint is decomposed into a real-time scheduling instruction sequence, which is then sent to the corresponding physical computing nodes for execution to complete the computing power scheduling work.

[0008] Preferably, the method further includes: Construct and maintain a digital twin collaborative optimization layer connecting the power system and the computing center. The digital twin collaborative optimization layer synchronizes and predicts in real time the node marginal electricity price, renewable energy output curve, load curve of the power system, and the heterogeneous resource status, task queue and network topology of the computing center. The actual execution feedback data of the physical system is collected and compared with the predicted trajectory in the preset digital twin collaborative optimization layer. The model parameters of the collaborative optimization layer are calibrated online according to the deviation to form a closed-loop optimization.

[0009] Preferably, the step of constructing and solving the computing power task scheduling optimization model across spatiotemporal scales specifically includes: Set a long-term look-ahead optimization window and a short-term rolling execution window; Within the aforementioned look-ahead optimization window, with the objective of minimizing the weighted sum of total system operating cost and carbon emissions, and applying constraints such as grid security operation, task completion, computing resource capacity, task time window, and portability, a mixed-integer programming model is constructed to obtain the theoretically optimal global scheduling blueprint. After entering the rolling execution window, the global scheduling blueprint is re-optimized and fine-tuned based on the latest system status information to generate precise scheduling instructions for the current execution cycle.

[0010] Preferably, the total system operating cost includes electricity costs incurred due to the migration and scheduling of computing tasks, network transmission costs, and service quality penalty costs incurred due to task completion times deviating from their ideal time windows; the carbon emissions are quantified by multiplying the electricity consumed by the computing tasks by the average carbon emission factor of the regional power grid corresponding to the electricity source. Preferably, the method further includes: An experience sample library is formed based on historical scheduling decisions and their multi-objective benefits. Based on the aforementioned experience sample library, a deep reinforcement learning module is trained. Under different system state characteristics, the deep reinforcement learning module directly outputs the optimal scheduling strategy suggestion or the key parameter configuration of the optimization model. In each solution of the computational task scheduling optimization model across the spatiotemporal scale, the output of the deep reinforcement learning module is introduced as an initial solution or decision guide to accelerate the solution process and improve the quality of the strategy.

[0011] Preferably, before the digital twin collaborative optimization layer synchronizes and predicts in real time, the method further includes collecting real-time asynchronous data streams, performing data fusion and calibration, as follows: Obtain the characteristic parameters of the real-time asynchronous data stream, including data type, data acquisition and transmission delay, and data packet size; Based on the data type and its corresponding latency, and combined with the sensitivity of the data type to physical or business changes, the latency impact coefficient is dynamically determined. Based on the latency impact coefficient and the data acquisition and transmission latency, the deviation in state representation caused by data asynchrony is calculated. The real-time asynchronous data stream is prioritized according to its calculated deviation and dynamically divided into n data batch processing groups, where n≥1, to ensure that the total amount of data in each group is within the preset memory processing threshold. Based on historical processing logs, the calibration processing time for different types and sizes of data packets is learned and estimated. Then, a processing time window is estimated for each newly generated packet as a benchmark for subsequent process monitoring.

[0012] Preferably, when performing calibration processing on the data of the nth group, the specific steps include: The grouped data is further divided into multiple data blocks of a preset size and processed sequentially. Monitor the actual processing time of each data block and compare it with the estimated processing time in real time; If the difference between the actual time taken and the estimated time taken is within a reasonable threshold, the data processing for that group is considered normal and can continue. If the difference exceeds a reasonable threshold, an anomaly diagnosis will be triggered immediately.

[0013] Preferably, the specific process of the abnormality diagnosis is as follows: Once an anomaly is detected, the repair sub-process is immediately initiated to repair the system; Once the repair is complete, the calibration process will automatically restart from the interrupted data block until all data blocks have been processed, ensuring the final consistency of the data stream processing.

[0014] Secondly, the present invention also provides an integrated computer-aided collaborative management system based on a virtual power plant, used to implement the above-mentioned method, wherein the platform includes: The computing power task acquisition module is used to acquire a set of batch computing power tasks to be scheduled. Each task in the set of batch computing power tasks to be scheduled is marked with a time window constraint, resource demand spectrum, portability identifier and task priority. The scheduling blueprint generation module is used to construct and solve a computing power task scheduling optimization model across time and space scales based on the forward-looking information provided by the preset digital twin collaborative optimization layer, with the multi-objective optimization of total system economic cost and total carbon emissions as the guide, and obtain a global scheduling blueprint that includes the optimal execution time, execution location and resource allocation of each task. The execution module is used to decompose the global scheduling blueprint into a real-time scheduling instruction sequence, and send it to the corresponding physical computing power nodes for execution to complete the computing power scheduling work.

[0015] Thirdly, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the above-described method when executing the computer program. Beneficial effects

[0016] This invention dynamically correlates the execution time and location of computing tasks with the node marginal price (LMP) of the power grid. By optimizing scheduling, it guides migrateable computing tasks to be executed preferentially during periods and regions with low electricity prices, effectively avoiding peak electricity consumption and high-price periods, and directly reducing electricity procurement costs. At the same time, by optimizing task migration strategies, it balances network transmission costs with electricity savings, thereby minimizing the total system operating cost.

[0017] This invention transforms massive, portable computing loads into high-quality, flexible resources that can be controlled by virtual power plants. When renewable energy output is high and the grid needs to increase its load, computing tasks can be appropriately increased or executed ahead of schedule to absorb green electricity. When grid supply is tight, non-urgent computing tasks can be appropriately delayed, playing a role in "peak shaving and valley filling." This proactive and precise load regulation capability greatly enhances the grid's ability to cope with the volatility of renewable energy and improves the stability and security of system operation. Attached Figure Description

[0018] Figure 1 This is a flowchart of the computer-aided collaborative integrated management method based on a virtual power plant in an embodiment of the present invention; Figure 2 This is a block diagram of the computer-aided collaborative integrated management system based on a virtual power plant, as described in an embodiment of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] like Figure 1 As shown in the figure, this embodiment of the invention provides a computer-aided collaborative integrated management method based on a virtual power plant, which includes the following steps: Step S100: Construct and update the digital twin collaborative optimization layer connecting the power system and the computing center, wherein the digital twin collaborative optimization layer synchronizes and predicts in real time the node marginal electricity price, renewable energy output curve, load curve of the power system, and the heterogeneous resource status, task queue and network topology of the computing center. The digital twin collaborative optimization layer includes: The power module is used to map deterministic and probabilistic prediction information of power grid topology, generator combination, transmission line capacity, nodal marginal electricity price, and wind or solar power. The computing power module is used to map the dynamic utilization rate, energy efficiency indicators, cooling power consumption, and network bandwidth and latency matrix between data centers of heterogeneous computing resources in each data center. The collaborative mapping interface is used to establish the quantitative conversion relationship between standard computing power task units and electricity consumption, and to define the interaction rules for computing power load to participate in power system balancing as an adjustable flexible load. The specific process for predicting the marginal electricity price at each node is as follows: Using long short-term memory networks or temporal convolutional networks, with the historical LMP sequences of hundreds of nodes, system load, and inter-regional power transmission capacity as the main inputs, we can capture their complex spatiotemporal correlations and intraday / seasonal patterns. The model's predictions are fused with market simulation-based predictions; the market simulator simulates the market clearing process based on the latest unit bids, load forecasts, network topology, and physical constraints, and outputs the theoretical LMP; the machine learning model is responsible for correcting complex market behaviors and non-physical constraints that were not considered in the simulation.

[0021] Real-time updates and input: Data sources: Day-ahead / real-time market clearing results, transmission equipment shutdown plans, unit fault information, and weather data (affecting load and renewable energy).

[0022] Update mechanism: The model runs in a rolling manner; for example, it receives the latest system load, new energy output and network topology status every 15 minutes, re-executes the prediction, refreshes the LMP curve for the next 24-48 hours, and the uncertainty decreases over time.

[0023] Forecasting of renewable energy output curves: Deterministic prediction: Physical models based on numerical weather prediction (wind speed, irradiance, cloud cover, temperature) are combined with historical power conversion data of power plants for the same period to perform joint physical-statistical forecasts. Tree models such as XGBoost and LightGBM are commonly used for site-level ultra-short-term (0-6 hours) and short-term (6-72 hours) point forecasts.

[0024] Output: Expected power output of wind farms / photovoltaic power plants in various future time periods.

[0025] Probabilistic prediction and scene generation: Using quantile regression forest or generative adversarial network, multiple possible trajectories for future power output are generated based on the ensemble forecast and historical distribution of prediction error of NWP. A set of typical scenarios that can cover future uncertainties and their probability of occurrence are used for subsequent model optimization for risk assessment and decision-making; Real-time updates: The system receives the latest NWP data and measured power of the power plant every 5-15 minutes, updates the predictions on a rolling basis, and uses algorithms such as Kalman filtering to correct prediction errors online.

[0026] Forecasting of system and regional load curves: Traditional load forecasting: Mature multi-level forecasting. Provincial load is calculated using SARIMAX or Prophet models that consider weekday type, holidays, and weather factors (temperature, humidity). Nodal load is decomposed based on the higher-level forecast using historical load distribution coefficients.

[0027] Real-time: Executed on a rolling basis, updated every 15-60 minutes.

[0028] Computing load prediction: This is the endogenous predictive function of the digital twin collaborative optimization layer. It does not rely on external historical data, but is based on: Current scheduled computing power task blueprint; The set of tasks currently waiting to be scheduled and their resource requirements. Optimize the future scheduling decisions output by the model.

[0029] The computing power simulation engine within the digital twin collaborative optimization layer "simulates the execution" of future tasks based on predetermined scheduling strategies and task characteristics, thereby calculating the predicted power consumption curve for each computing node and each time period. This curve serves as a "flexible load" feedback to the power system twin sub-model, influencing grid power flow and LMP prediction, forming an internal coupling closed loop.

[0030] Computing power network state prediction (heterogeneous resources, task queues, network topology): Heterogeneous resource status prediction: Based on current resource utilization, task execution progress, and resource reclamation models, deterministic extrapolation is performed. For example, if a GPU pool is known to be running a task that takes 2 hours, it can be predicted that its utilization will remain high for the next 2 hours before being released. Inputs: Task execution model, server power consumption model, cooling efficiency model; Task queue prediction: By combining historical task arrival patterns (such as daily / weekly patterns) and real-time submission monitoring, time series models (such as Poisson process modified models) are used to predict the number and type distribution of new task arrivals in the next short period (such as the next hour). Network topology and state prediction: Topology: Relatively static, but requires synchronous configuration to manage database changes; Status (bandwidth, latency): Based on historical traffic data and known planned data transmission tasks, use network simulation tools (such as Mininet-based simulations) or lightweight time-series models to predict the expected utilization and latency range of critical links in future periods.

[0031] Step S200: Obtain a set of batch computing power tasks to be scheduled, wherein each task in the set of batch computing power tasks to be scheduled is marked with a time window constraint, resource demand spectrum, portability identifier and task priority. Time window constraints are generally defined by the user or inferred by the system. Specifically, user-defined constraints involve setting the earliest start time, latest finish time, or expected completion deadline when submitting a task. System inference involves estimating a committed completion time window for batch jobs based on the historical time consumption of similar tasks and the current queue length. For interactive services with service level agreements (SLAs), the response time requirement is directly converted into a strict time window. Hard windows are absolute constraints for scheduling, ensuring that critical tasks are completed on time. Soft windows provide flexibility for optimization, allowing the system to flexibly arrange task execution time within an acceptable latency range in pursuit of lower electricity costs or higher green energy consumption. Resource demand spectrum refers to the total amount of computing resources required to complete a computing task in a specific scenario. It is usually measured by indicators such as processing speed, parallel capability, memory capacity, and energy consumption. This is the fundamental basis for determining whether a server can run the task. The multi-dimensional demand spectrum ensures that the task is not assigned to a server with mismatched resources, avoiding performance bottlenecks. More importantly, it is precisely matched with the server resource status in the digital twin. It is the direct input for calculating the "resource idle rate" and judging the "resource balance rate", which directly determines the intrinsic efficiency and energy consumption of task execution. Portability tags are automatically assigned by the system or by the user based on task characteristics. They typically include: cross-regional portability: large-scale data analysis and offline rendering with no strict data locality requirements and insensitivity to network latency; portability only across data centers: tasks with some data dependencies but can be synchronized via high-speed networks, such as some distributed training tasks; and non-portable (local execution): real-time interactive applications involving sensitive data (subject to compliance constraints) or with extremely stringent latency requirements. This is the core link between computing power scheduling and power optimization. Only tasks marked as portable can become the carriers of "spatiotemporal transfer of computing load," allowing the scheduled system to migrate them from high-electricity-price, high-carbon-emission areas to low-electricity-price, high-green-energy areas. Portability tags directly define the decision variable space of the optimization problem. Task prioritization is used when system resources are scarce. Priority is the highest guiding principle for the scheduling system to make trade-offs and choices; it ensures that the needs of the enterprise's core business and key customers are given priority under any optimization scheme. Step S300: Based on the forward-looking information provided by the digital twin collaborative optimization layer, and guided by the multi-objective optimization of total system economic cost and total carbon emissions, construct and solve a computing power task scheduling optimization model across time and space scales to obtain a global scheduling blueprint that includes the optimal execution time, execution location and resource allocation of each task. Step S400: Decompose the global scheduling blueprint into a real-time scheduling instruction sequence and send it to the corresponding physical computing power nodes for execution; Step S500: Collect the actual execution feedback data of the physical system, compare it with the predicted trajectory in the digital twin collaborative optimization layer, and calibrate the model parameters of the collaborative optimization layer online according to the deviation to form a closed-loop optimization.

[0032] Furthermore, the construction and solution of the computing power task scheduling optimization model across spatiotemporal scales specifically includes: Set a long-term look-ahead optimization window and a short-term rolling execution window; Within the aforementioned forward optimization window, with the goal of minimizing the weighted sum of total system operating cost and carbon emissions, and by imposing constraints on grid safety operation, task completion, computing power capacity, task time window, and portability, a mixed integer programming model is constructed, and the theoretically optimal global scheduling blueprint is obtained by solving it. Specifically, define a three-dimensional decision variable. Where i represents a computing task to be scheduled (a set of labeled tasks), j represents a physical computing node (such as an available resource pool in a data center), and t represents a discrete time period (such as the next 24 hours, divided into 15-minute periods, for a total of 96 periods); when When, it means the decision is "task i starts executing in time period t of computing node j"; With the goal of minimizing the total system operating cost, the calculation formula is as follows: , in, The predicted electricity price for node j in time period t is provided for the digital twin collaborative optimization layer. The estimated total power consumption for task i (obtained from its resource requirement spectrum and estimated runtime). The network transmission cost for migrating the data required for task i to node j; The priority weight (labeling information) for task i; The actual completion time of task i is delayed relative to its ideal time; With the goal of minimizing total carbon emissions, the calculation formula is as follows: , in, The predicted carbon intensity factor of the power grid in the region where node j is located during time period t. This value changes in real time, approaching zero when wind / solar power generation is high, and increasing when thermal power is the main source of power. The goal is to minimize the weighted sum of total system operating costs and carbon emissions. , in, , All of these are global weights set by the system policy.

[0033] Power grid security constraints (coupled through digital twin collaborative optimization layer): Total load changes generated by dispatching must not cause any transmission line overload or node voltage exceedance. Task completion constraint: Each task must be executed one and only once; , Computing resource capacity constraint: Resource capacity constraint: At any time t on any node j, the total amount of various resources (CPU, memory, GPU, etc.) consumed by all tasks running on it must not exceed the available capacity of that node; , in, Let i be the resource requirement of the k-th type. Let be the total amount of the k-th type of resource that node j can provide during time period t; Task Time Window: Tasks must start and end within the specified time window; otherwise, ... ; Portability constraint: Based on the portability identifier of a task, restrict the set of nodes that it can be scheduled; otherwise, ; The theoretically optimal global scheduling blueprints obtained by solving include medium- to long-term blueprint planning (future 1-7 days, time granularity of 1 hour) and short-term blueprint planning (future 24 hours, time granularity of 15 minutes). Upon entering the rolling execution window, the global scheduling blueprint is re-optimized and fine-tuned based on the latest system status information to generate precise scheduling instructions for the current execution cycle. The specific process is as follows: Execution and Forecast Updates: Execute scheduling instructions for the current time period (e.g., 0-15 minutes). Simultaneously, the digital twin collaborative optimization layer receives the latest actual data (e.g., actual power generation price, actual wind power output) and continuously updates forecast information for all subsequent time periods.

[0034] Model re-solution: Starting from the current moment, the short-term optimization model is re-solved based on the updated predictions. Since most decisions are fixed, only decisions that have not been executed in the recent few hours are fine-tuned.

[0035] Output new instruction: Output the adjusted instruction for the next execution cycle.

[0036] Closed-loop learning: The actual execution results (such as real energy consumption and task completion time) are compared with the model predictions. The deviation data is used to continuously train and calibrate the prediction models in the digital twin collaborative optimization layer (such as power consumption models and task runtime models), so that the next prediction and optimization are more accurate.

[0037] Furthermore, the total operating cost of the system includes electricity costs, network transmission costs, and service quality penalty costs incurred due to the migration and scheduling of computing tasks, as well as the service quality penalty costs incurred due to the task completion time deviating from its ideal time window; the carbon emissions are quantified by multiplying the electricity consumed by the computing tasks by the average carbon emission factor of the regional power grid corresponding to the electricity source.

[0038] In some embodiments, the method further includes an adaptive policy evolution step: Historical scheduling decisions and their resulting multi-objective benefits form an empirical sample database; Based on the aforementioned experience sample library, a deep reinforcement learning agent is trained. The agent learns to directly output scheduling strategy suggestions or key parameter configurations that approximate the optimal model under different system state characteristics. In each solution of the scheduling optimization model, the output of the deep reinforcement learning agent is introduced as an initial solution or decision guide to accelerate the solution process and improve policy quality.

[0039] In some embodiments, before the digital twin collaborative optimization layer synchronizes and predicts in real time, the method further includes collecting real-time asynchronous data streams and performing data fusion and calibration, as detailed below: Obtain the characteristic parameters of the real-time asynchronous data stream, including data type, data acquisition and transmission delay, and data packet size; Based on the data type and its corresponding latency, and combined with the sensitivity of the data type to physical or business changes, the latency impact coefficient is dynamically determined. Based on the latency impact coefficient and the data acquisition and transmission latency, the deviation in state representation caused by data asynchrony is calculated. The real-time asynchronous data stream is prioritized according to its calculated deviation and dynamically divided into n data batch processing groups, where n≥1, to ensure that the total amount of data in each group is within the preset memory processing threshold. Based on historical processing logs, the calibration processing time for different types and sizes of data packets is learned and estimated. Then, a processing time window is estimated for each newly generated packet as a benchmark for subsequent process monitoring.

[0040] Furthermore, the calibration process for the data in the nth group specifically includes: The grouped data is further divided into multiple data blocks of a preset size and processed sequentially. Monitor the actual processing time of each data block and compare it with the estimated processing time in real time; If the difference between the actual time taken and the estimated time taken is within a reasonable threshold, the data processing for that group is considered normal and can continue. If the difference exceeds a reasonable threshold, an anomaly diagnosis will be triggered immediately.

[0041] Furthermore, the specific process of the abnormality diagnosis is as follows: Once an anomaly is detected, the repair sub-process is immediately initiated to repair the system; Once the repair is complete, the calibration process will automatically restart from the interrupted data block until all data blocks have been processed, ensuring the final consistency of the data stream processing.

[0042] In summary, this invention successfully constructs a collaborative management paradigm that deeply integrates power flow and information flow. It not only provides an innovative solution for cost reduction, efficiency improvement, and green development of computing infrastructure, but also provides important flexible resource support for the new power system to absorb a high proportion of renewable energy, thus having significant economic, environmental, and social value.

[0043] like Figure 2 As shown, another embodiment of the present invention provides an integrated computer-computer collaborative management system based on a virtual power plant, the platform comprising: The computing power task acquisition module is used to acquire a set of batch computing power tasks to be scheduled. Each task in the set of batch computing power tasks to be scheduled is marked with a time window constraint, resource demand spectrum, portability identifier and task priority. The scheduling blueprint generation module is used to construct and solve a computing power task scheduling optimization model across time and space scales based on the forward-looking information provided by the preset digital twin collaborative optimization layer, with the multi-objective optimization of total system economic cost and total carbon emissions as the guide, and obtain a global scheduling blueprint that includes the optimal execution time, execution location and resource allocation of each task. The execution module is used to decompose the global scheduling blueprint into a real-time scheduling instruction sequence, and send it to the corresponding physical computing power nodes for execution to complete the computing power scheduling work.

[0044] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0045] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0046] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0047] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

[0048] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A computer-aided collaborative integrated management method based on a virtual power plant, characterized in that, The method includes: Obtain a set of batch computing power tasks to be scheduled, wherein each task in the set of batch computing power tasks to be scheduled is marked with a time window constraint, resource demand spectrum, portability identifier and task priority; Based on the forward-looking information provided by the pre-set digital twin collaborative optimization layer, and guided by the multi-objective optimization of total system economic cost and total carbon emissions, a computing power task scheduling optimization model across time and space scales is constructed and solved to obtain a global scheduling blueprint that includes the optimal execution time, execution location and resource allocation of each task. The global scheduling blueprint is decomposed into a real-time scheduling instruction sequence, which is then sent to the corresponding physical computing nodes for execution to complete the computing power scheduling work.

2. The computer-aided collaborative integrated management method based on a virtual power plant according to claim 1, characterized in that, The method further includes: Construct and maintain a digital twin collaborative optimization layer connecting the power system and the computing center. The digital twin collaborative optimization layer synchronizes and predicts in real time the node marginal electricity price, renewable energy output curve, load curve of the power system, and the heterogeneous resource status, task queue and network topology of the computing center. The actual execution feedback data of the physical system is collected and compared with the predicted trajectory in the preset digital twin collaborative optimization layer. The model parameters of the collaborative optimization layer are calibrated online according to the deviation to form a closed-loop optimization.

3. The computer-aided collaborative integrated management method based on a virtual power plant according to any one of claims 1 or 2, characterized in that, The construction and solution of the computing power task scheduling optimization model across spatiotemporal scales specifically includes: Set a long-term look-ahead optimization window and a short-term rolling execution window; Within the aforementioned look-ahead optimization window, with the objective of minimizing the weighted sum of total system operating cost and carbon emissions, and applying constraints such as grid security operation, task completion, computing resource capacity, task time window, and portability, a mixed-integer programming model is constructed to obtain the theoretically optimal global scheduling blueprint. After entering the rolling execution window, the global scheduling blueprint is re-optimized and fine-tuned based on the latest system status information to generate precise scheduling instructions for the current execution cycle.

4. The computer-aided collaborative integrated management method based on a virtual power plant according to claim 3, characterized in that, The total operating cost of the system includes electricity costs, network transmission costs, and service quality penalty costs incurred due to the migration and scheduling of computing tasks, as well as the service quality penalty costs incurred due to the task completion time deviating from its ideal time window; the carbon emissions are quantified by multiplying the electricity consumed by the computing tasks by the average carbon emission factor of the regional power grid corresponding to the electricity source.

5. The computer-aided collaborative integrated management method based on a virtual power plant according to claim 1, characterized in that, The method further includes: An experience sample library is formed based on historical scheduling decisions and their multi-objective benefits. Based on the aforementioned experience sample library, a deep reinforcement learning module is trained. Under different system state characteristics, the deep reinforcement learning module directly outputs the optimal scheduling strategy suggestion or the key parameter configuration of the optimization model. In each solution of the computational task scheduling optimization model across the spatiotemporal scale, the output of the deep reinforcement learning module is introduced as an initial solution or decision guide to accelerate the solution process and improve the quality of the strategy.

6. The computer-aided collaborative integrated management method based on a virtual power plant according to claim 1, characterized in that, Before the digital twin collaborative optimization layer synchronizes and predicts in real time, the method also includes collecting real-time asynchronous data streams for data fusion and calibration, as detailed below: Obtain the characteristic parameters of the real-time asynchronous data stream, including data type, data acquisition and transmission delay, and data packet size; Based on the data type and its corresponding latency, and combined with the sensitivity of the data type to physical or business changes, the latency impact coefficient is dynamically determined. Based on the latency impact coefficient and the data acquisition and transmission latency, the deviation in state representation caused by data asynchrony is calculated. The real-time asynchronous data stream is prioritized according to its calculated deviation and dynamically divided into n data batch processing groups, where n≥1, to ensure that the total amount of data in each group is within the preset memory processing threshold. Based on historical processing logs, the calibration processing time for different types and sizes of data packets is learned and estimated. Then, a processing time window is estimated for each newly generated packet as a benchmark for subsequent process monitoring.

7. The computer-aided collaborative integrated management method based on a virtual power plant according to claim 6, characterized in that, The calibration process for the data in the nth group specifically includes: The grouped data is further divided into multiple data blocks of a preset size and processed sequentially. Monitor the actual processing time of each data block and compare it with the estimated processing time in real time; If the difference between the actual time taken and the estimated time taken is within a reasonable threshold, the data processing for that group is considered normal and can continue. If the difference exceeds a reasonable threshold, an anomaly diagnosis will be triggered immediately.

8. The computer-aided collaborative integrated management method based on a virtual power plant according to claim 7, characterized in that, The specific process for abnormal diagnosis is as follows: Once an anomaly is detected, the repair sub-process is immediately initiated to repair the system; Once the repair is complete, the calibration process will automatically restart from the interrupted data block until all data blocks have been processed, ensuring the final consistency of the data stream processing.

9. A computer-aided collaborative integrated management system based on a virtual power plant, used to implement the method as described in any one of claims 1-8, characterized in that, The platform includes: The computing power task acquisition module is used to acquire a set of batch computing power tasks to be scheduled. Each task in the set of batch computing power tasks to be scheduled is marked with a time window constraint, resource demand spectrum, portability identifier and task priority. The scheduling blueprint generation module is used to construct and solve a computing power task scheduling optimization model across time and space scales based on the forward-looking information provided by the preset digital twin collaborative optimization layer, with the multi-objective optimization of total system economic cost and total carbon emissions as the guide, and obtain a global scheduling blueprint that includes the optimal execution time, execution location and resource allocation of each task. The execution module is used to decompose the global scheduling blueprint into a real-time scheduling instruction sequence, and send it to the corresponding physical computing power nodes for execution to complete the computing power scheduling work.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 8.

Images