Power policy pre-forecast post-execution system based on ai large model
Patent Information
- Application Number
- CN202511834441.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-09-11
AI Technical Summary
现有超算中心的计算任务调度系统通常只考虑计算任务本身消耗的计算资源，而完全忽略了外部的电力供应状态，特别是电力的来源结构
[0034]1. The AI-based large-scale model-based power strategy prediction and execution system provided in this application fundamentally solves the problem that existing supercomputing centers can only passively receive power supply and cannot actively respond to green electricity. Compared with existing technologies, this application creatively combines weather forecasting with power generation prediction for wind and solar power, and links the electricity price consumed by the computing task with wind and solar power generation through power generation prediction, allowing users to choose for themselves. This constructs a price incentive mechanism of "peak power generation equals price trough," transforming unstable new energy power generation into an economic signal that can guide user behavior.
Smart Images
Figure CN122735992A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of power system automation control, and in particular to power system automation control in supercomputing centers. Background Technology
[0002] Supercomputing centers, or supercomputing hubs for short, are institutions that provide infrastructure and services for high-performance computing. They house a large number of computing nodes and storage systems, making them energy-intensive facilities. Existing supercomputing center task scheduling systems typically only consider the computing resources consumed by the task itself, completely ignoring the external power supply status, especially the power source structure. This prevents supercomputing centers from identifying peak periods for renewable energy generation, such as solar and wind power, and thus from scheduling computing tasks to run during these periods. Consequently, this results in the ineffective use of green electricity, hindering the reduction of carbon emissions and the promotion of green electricity consumption.
[0003] In fact, new energy sources such as wind and solar power are intermittent and fluctuating. Their peak power generation often does not naturally coincide with the peak execution of computing tasks. This means that during periods of high new energy power generation, supercomputing centers will miss the opportunity to use cheap and green electricity, and instead exacerbate the burden on the power grid during periods of power shortage. Summary of the Invention
[0004] To address the shortcomings of existing technologies, one of the objectives of this invention is to provide a power strategy prediction and execution system based on a large AI model.
[0005] The power strategy prediction and execution system based on AI large model provided in this application adopts the following technical solution:
[0006] The AI-based large-scale model-based power strategy prediction and execution system includes a green power generation system, which includes a wind power generation system, a solar power generation system, and an energy storage system. The green power generation system is connected to the power supply station of the supercomputing center. The power supply station has a power supply station data center, which is connected to the weather forecast system. The power supply station data center is connected to the computer management platform of the supercomputing center, which has platform management software.
[0007] The power station's data center has an AI power prediction model. This model associates historical power generation data with meteorological data and power generation to train the AI power prediction model to predict the power generation of the green power generation system.
[0008] The power station's data center predicts power generation for each period based on meteorological data from the weather forecasting system, and identifies peak and off-peak periods for power generation.
[0009] It is connected to the virtual grid system, which uses the valley price standard corresponding to the peak power generation period and the peak price standard corresponding to the low power generation period in the green power generation system.
[0010] The platform management software includes a user interaction module, a pricing module, an AI large model module, and a scheduling module.
[0011] The user interaction module receives computing tasks uploaded by users;
[0012] The AI large model module associates historical computing task data with computing power usage to train the AI large model to predict computing power usage.
[0013] The AI large model module calls the uploaded computing tasks, extracts related data, and uses the AI large model to predict computing power usage.
[0014] The pricing module connects to the virtual power grid system, obtains the off-peak price pricing standard and the peak price pricing standard, and combines the time period determined by the power generation forecast and the predicted computing power usage to generate the estimated power price for peak and off-peak power generation periods. The estimated power price is then fed back to the user through the user interaction module, allowing the user to select the estimated power price.
[0015] The scheduling module obtains the estimated power price and corresponding time period selected by the user, and schedules the calculation task to be executed in the time period selected by the user.
[0016] This application provides an AI-based large-scale model-based power strategy prediction and execution system, which fundamentally solves the problem that existing supercomputing centers can only passively receive power and cannot actively respond to green electricity. Compared with existing technologies, this application creatively combines weather forecasting with power generation prediction for wind and solar power, and links the electricity price consumed by computing tasks to wind and solar power generation through power generation prediction, allowing users to choose for themselves. This constructs a price incentive mechanism where "peak power generation equals price trough," transforming unstable renewable energy power generation into an economic signal that can guide user behavior.
[0017] This application utilizes a large AI model to accurately predict computing power demand, enabling users to know the prices for different time periods when submitting computing tasks. This allows them to proactively adjust their computing tasks to times when green electricity is abundant. This transforms supercomputing centers from a passive electricity consumption model to a proactive green electricity consumption model. By guiding user behavior through economic signals, it significantly increases the utilization rate of intermittent renewable energy sources such as wind and solar power, directly reducing carbon emissions.
[0018] Furthermore, by guiding users to actively migrate their computing load to green periods with abundant power, this application effectively alleviates the power consumption pressure on supercomputing centers during periods of grid shortage, enabling supercomputing centers to fully utilize their computing capabilities during periods of abundant power supply. To a certain extent, this achieves the effect of "small supercomputing centers" undertaking the computing tasks of "large supercomputing centers" through intelligent scheduling, thus achieving a performance leapfrog effect.
[0019] Preferably, the AI large model module adopts an AI large model based on the Transformer architecture.
[0020] By adopting the above technical solutions, the AI large model based on the Transformer architecture can process time-series data more effectively, capturing the complex nonlinear relationship between computing power demand and task characteristics and its long-term dependence. It can also stably predict computing power usage, providing a stable data foundation for subsequent price forecasting and scheduling decisions.
[0021] Preferably, the AI large model module continuously monitors the error between the computing power usage predicted by the computing task and the actual computing power usage consumed by the computing task;
[0022] If the error exceeds the preset threshold for at least 5 consecutive times, the AI large model will be retrained.
[0023] By adopting the above technical solutions and setting up error monitoring and passive triggering retraining mechanisms, the AI large model has the ability to self-perceive performance degradation and autonomously optimize itself. This mechanism ensures that the model can adapt to changes in computing power demand patterns in a timely manner by only initiating retraining when the prediction error continues to deviate.
[0024] Preferably, the AI large model module collects newly added historical computing task-related data at a preset period, performs incremental training on the AI large model, and continuously optimizes the prediction accuracy.
[0025] By adopting the above technical solutions, new data is collected regularly to incrementally train the AI model, enabling the AI model to proactively adapt to the changing trends of task modes and achieve continuous evolution of predictive capabilities. This update method ensures that the system can maintain a high level of computing power prediction over a long period of time.
[0026] Preferably, the scheduling module obtains the maximum computing capacity of the supercomputing center from the supercomputing center's computer management platform; when performing scheduling, the scheduling module ensures that the total computing power requirement of all scheduled computing tasks within the same time period does not exceed the maximum computing capacity of the supercomputing center.
[0027] By adopting the above technical solution and introducing the maximum computing capacity constraint of the supercomputing center, it is ensured that the scheduling module always prioritizes system stability when executing user selections, thereby preventing the risk of supercomputing center overload caused by multiple computing tasks being executed at the same time.
[0028] Preferably, when the scheduling module performs scheduling, if the total computing power demand of all scheduled computing tasks in the same time period is greater than 90% of the maximum computing capacity of the supercomputing center, the scheduling module will trigger the pricing module to increase the estimated computing power price generated by the pricing module.
[0029] By adopting the above technical solution, an intelligent control mechanism is introduced. The scheduling module does not simply refuse scheduling when the supercomputing center is overloaded. Instead, it creatively constructs a preventative control system based on price signals by monitoring resource utilization in real time and dynamically adjusting the estimated resource price. When resource scarcity is detected during a certain period, the pricing module raises the price for that period, using economic means to proactively distribute user choices, significantly improving the overall system's intelligence level.
[0030] Preferably, the energy storage system is charged during the peak power generation period corresponding to the off-peak price pricing standard of the virtual power grid, and the power station data center monitors the energy storage capacity of the energy storage system and obtains the load information of the supercomputing center from the scheduling module.
[0031] The power station data center calculates the duration for which the stored energy sustains the system's operation based on the stored energy and the load information, and uses the duration for which the stored energy sustains the system as the delay period for the off-peak price pricing standard.
[0032] By adopting the above technical solutions, the energy storage system can be charged during peak green electricity periods (corresponding to off-peak pricing), and the time extension of the value of stored electricity is established. The power station data center can accurately calculate the duration for which the stored electricity can sustain the system operation by using the energy storage system's own stored electricity and the load information from the supercomputing center of the scheduling module, and use this period as the delay period, so that the energy storage system can extend the low-priced electricity.
[0033] In summary, this application includes at least one of the following beneficial technical effects:
[0034] 1. The AI-based large-scale model-based power strategy prediction and execution system provided in this application fundamentally solves the problem that existing supercomputing centers can only passively receive power supply and cannot actively respond to green electricity. Compared with existing technologies, this application creatively combines weather forecasting with power generation prediction for wind and solar power, and links the electricity price consumed by the computing task with wind and solar power generation through power generation prediction, allowing users to choose for themselves. This constructs a price incentive mechanism of "peak power generation equals price trough," transforming unstable new energy power generation into an economic signal that can guide user behavior.
[0035] 2. This application utilizes a large AI model to accurately predict computing power demand, enabling users to know the prices for different time periods when submitting computing tasks, thus proactively adjusting their computing tasks to times with abundant green electricity. This transforms supercomputing centers from a passive electricity consumption model to a proactive green electricity consumption model, guiding user behavior through economic signals, significantly increasing the utilization rate of intermittent renewable energy sources such as wind and solar power, and directly reducing carbon emissions;
[0036] 3. Furthermore, by guiding users to actively migrate their computing load to green periods with abundant power, this application effectively alleviates the power consumption pressure on supercomputing centers during periods of grid shortage, enabling supercomputing centers to fully utilize their computing capabilities during periods of abundant power supply. To a certain extent, this achieves the effect of "small supercomputing centers" undertaking the computing tasks of "large supercomputing centers" through intelligent scheduling, thus achieving a performance leapfrog effect. Attached Figure Description
[0037] Figure 1 This embodiment of the application is a data transmission diagram illustrating a power strategy prediction and execution system based on an AI large model;
[0038] Figure 2(ae) is a code framework diagram of a power strategy prediction and execution system based on an AI large model. Detailed Implementation
[0039] The following is in conjunction with the appendix Figure 1 Figure 2(ae) provides a further detailed description of this application.
[0040] This application discloses a power strategy prediction and execution system based on an AI large model.
[0041] Reference Figure 1The AI-based large-scale power strategy prediction and execution system includes a green power generation system, which comprises a wind power generation system, a solar power generation system, and an energy storage system. The wind power generation system and the solar power generation system are connected to the energy storage system, and the green power generation system is connected to the power supply station of the supercomputing center. The power supply station has a power supply station data hub, which interacts with the green power generation system, the computer management platform of the supercomputing center, and the weather forecasting system.
[0042] Reference Figure 1 Figure 2(ae) shows that the power station's data center is connected to the weather forecasting system and obtains historical and real-time data from the system. The data includes publicly available meteorological data released by the meteorological information center and real-time data collected by distributed meteorological sensors deployed in the green power generation area, thereby obtaining meteorological parameters, specifically wind force, sunshine, and temperature for each time period. Each time period is divided into 1-hour intervals. These data are updated according to a preset cycle. In this embodiment, the preset cycle is 1 hour, which can be adjusted to 30 minutes or 2 hours according to the actual frequency of meteorological changes.
[0043] The power station data center has an AI power prediction model. In this embodiment, the AI power prediction model adopts an LSTM architecture. During the training phase, the AI power prediction model associates historical power generation data, including information such as wind speed, sunshine duration, temperature, and time period, with the power generation and inputs this data into the AI power prediction model to train it to predict the power generation of wind power and solar power systems. In the usage phase after training, when the power station data center obtains the latest weather forecast data, the AI power prediction model predicts the power generation for each time period based on the wind speed, sunshine duration, temperature, and time period from the weather forecast system. Based on this predicted value sequence and a preset threshold, the power station data center identifies peak power generation periods with relatively high predicted power generation and off-peak power generation periods with relatively low predicted power generation.
[0044] Furthermore, the power station's data center is connected to the virtual grid system, which uses the valley price standard corresponding to the peak power generation period in the green power generation system and the peak price standard corresponding to the valley power generation period.
[0045] Furthermore, the energy storage system is charged during the peak power generation period corresponding to the off-peak pricing standard of the virtual grid. The power station data center monitors the energy storage capacity of the energy storage system and obtains the load information of the supercomputing center from the scheduling module. Based on the energy storage capacity and load information, the power station data center calculates the duration for which the energy storage capacity sustains the system operation and uses the duration for which the energy storage capacity sustains the operation as the delay period of the off-peak pricing standard. The price during this period is still calculated according to the off-peak pricing standard.
[0046] The supercomputing center's computer management platform runs platform management software, which includes a user interaction module, a pricing module, an AI large model module, and a scheduling module. Data is interconnected within each module to ensure smooth process flow.
[0047] The user interaction module receives computing tasks uploaded by users, transmits the computing tasks to the AI large model module for computing power usage prediction, and after receiving the estimated computing power price from the pricing module, displays different estimated computing power prices for the computing task during price troughs and price peaks for the user to choose from.
[0048] The AI large-scale model module is the core of computing power usage prediction. In this embodiment, an AI large-scale model based on the Transformer architecture is used. The AI large-scale model module includes two stages: model training and computing power prediction. In the AI large-scale model training stage, the AI large-scale model module associates the associated data of historical computing tasks stored in the platform management software, including information such as computing task type, computing task size, and user type, with computing power usage and inputs it into the AI large-scale model. The AI large-scale model is trained to predict computing power usage when there is sufficient associated data. In the computing power prediction stage, the AI large-scale model module calls the computing tasks uploaded by users through the user interaction module, extracts the associated data, and inputs it into the trained AI large-scale model. The AI large-scale model then predicts the computing power usage of the computing task.
[0049] Furthermore, to ensure the continuous stability of the AI model's prediction accuracy, the AI model module also incorporates an error monitoring and model update mechanism. The AI model module continuously monitors the error between the predicted computing power usage and the actual computing power usage for each computational task. When this error exceeds a preset threshold for at least five consecutive times, the AI model module automatically triggers retraining of the AI model. At this time, the data from the five most recent and newly added historical computational tasks are added to the training set, and the model is retrained using the original training parameters. Secondly, the AI model module actively collects newly added historical computational task-related data at a preset cycle, correlates it with computing power usage, and performs incremental training on the AI model. In this embodiment, the preset cycle is once a week. This incremental training mechanism enables the AI model to continuously adapt to the changing trends of computational task characteristics.
[0050] The pricing module is connected to the virtual power grid system. It obtains the off-peak price standard corresponding to the peak power generation period and the peak price standard corresponding to the off-peak power generation period through the virtual power grid system. It also calls the AI power prediction model from the power station data center to predict the power generation of the wind power generation system and the time period division information, and retrieves the computing power usage of the predicted calculation task from the AI large model module. It generates the estimated power price for the peak and off-peak power generation periods, and feeds it back to the user through the user interaction module so that the user can select the estimated power price.
[0051] The scheduling module obtains the supercomputing center's maximum computing capacity from its computer management platform and calculates the computing power demand of scheduled tasks in real time for each time period. When a user selects a time period and its corresponding estimated computing power price through the user interaction module, the scheduling module first determines whether the sum of the total computing power demand of scheduled tasks in that time period and the current user's computing power usage exceeds the supercomputing center's maximum computing capacity. If it does not exceed the maximum capacity, the user is directly assigned to the corresponding time period for execution. If it does exceed the maximum capacity, the user is prompted with the message "Computing power resources are scarce in this time period; please select another time period."
[0052] Furthermore, the scheduling module is equipped with a load threshold-triggered price control mechanism. When the total computing power demand of scheduled tasks exceeds 90% of the supercomputing center's maximum computing capacity during a certain period, the scheduling module immediately sends a load warning signal to the pricing module, triggering an upward adjustment of the estimated computing power price for that period. The adjusted price is updated in real time through the user interaction module, thereby guiding subsequent users to choose other periods with lower loads through economic signals. This prevents computing tasks within the supercomputing center from approaching or reaching their computing capacity limits, ensuring the stable execution of all scheduled computing tasks.
[0053] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.
Claims
1. A power strategy prediction and execution system based on an AI large-scale model, comprising a green power generation system, including a wind power generation system, a solar power generation system, and an energy storage system, wherein the green power generation system is connected to a power supply station of a supercomputing center, and the power supply station has a power supply station data hub, characterized in that... The power station's data center is connected to the weather forecast system, and the power station's data center is connected to the computer management platform of the supercomputing center. The computer management platform has platform management software. The power station's data center has an AI power prediction model. This model associates historical power generation data with meteorological data and power generation to train the AI power prediction model to predict the power generation of the green power generation system. The power station's data center predicts power generation for each period based on meteorological data from the weather forecasting system, and identifies peak and off-peak periods for power generation. It is connected to the virtual grid system, which uses the valley price standard corresponding to the peak power generation period and the peak price standard corresponding to the low power generation period in the green power generation system. The platform management software includes a user interaction module, a pricing module, an AI large model module, and a scheduling module. The user interaction module receives computing tasks uploaded by users; The AI large model module associates historical computing task data with computing power usage to train the AI large model to predict computing power usage. The AI large model module calls the uploaded computing tasks, extracts related data, and uses the AI large model to predict computing power usage. The pricing module connects to the virtual power grid system, obtains the off-peak price pricing standard and the peak price pricing standard, and combines the time period determined by the power generation forecast and the predicted computing power usage to generate the estimated power price for peak and off-peak power generation periods. The estimated power price is then fed back to the user through the user interaction module, allowing the user to select the estimated power price. The scheduling module obtains the estimated power price and corresponding time period selected by the user, and schedules the calculation task to be executed in the time period selected by the user.
2. The power strategy prediction and execution system based on AI large model according to claim 1, characterized in that, The AI large model module adopts an AI large model based on the Transformer architecture.
3. The power strategy prediction and execution system based on an AI large model according to claim 1, characterized in that, The AI large model module continuously monitors the error between the computing power usage predicted by the computing task and the actual computing power usage consumed by the computing task. If the error exceeds the preset threshold for at least 5 consecutive times, the AI large model will be retrained.
4. The power strategy prediction and execution system based on AI large model according to claim 1, characterized in that, The AI large model module collects newly added historical computing task-related data at a preset cycle, performs incremental training on the AI large model, and continuously optimizes the prediction accuracy.
5. The power strategy prediction and execution system based on AI large model according to claim 1, characterized in that, The scheduling module obtains the maximum computing capacity of the supercomputing center from the computer management platform of the supercomputing center; when performing scheduling, the scheduling module ensures that the total computing power requirement of all scheduled computing tasks in the same time period does not exceed the maximum computing capacity of the supercomputing center.
6. The power strategy prediction and execution system based on an AI large model according to claim 5, characterized in that, When the scheduling module performs scheduling, if the total computing power demand of all scheduled computing tasks in the same time period exceeds 90% of the maximum computing capacity of the supercomputing center, the scheduling module will trigger the pricing module to increase the estimated computing power price generated by the pricing module.
7. The power strategy prediction and execution system based on AI large model according to claim 1, characterized in that, The energy storage system is charged during the peak power generation period corresponding to the off-peak price pricing standard of the virtual power grid. The power station data center monitors the energy storage capacity of the energy storage system and obtains the load information of the supercomputing center from the scheduling module. The power station data center calculates the duration for which the stored energy sustains the system's operation based on the stored energy and the load information, and uses the duration for which the stored energy sustains the system as the delay period for the off-peak price pricing standard.