A method for coordinated regulation of power data center resources based on operational status prediction

Abstract

This invention discloses a method for coordinated control of power data center resources based on operational status prediction, belonging to the field of resource control technology. By first generating a pre-state risk image under conditions where no control action was executed in the current round after predicting operational risks, and then performing coordinated control and recording the action object, type, and intensity, the system can compare actual operational feedback data with the pre-state risk image to identify the degree to which the feedback data is affected by the control action, thereby distinguishing between "real risks suppressed by control" and "risks that do not exist but are excessively intervened in." Furthermore, by correcting the coordinated control strategy for the next control cycle based on the control masking identification results, it can avoid misinterpreting stable feedback caused by control as accurate prediction or necessary control, reducing excessive cooling, frequent task migration, and excessive power supply margin reservation, thereby improving resource utilization and the rationality of closed-loop control.

Description

Technical Field

[0001] This invention relates to the field of resource regulation technology, and specifically to a method for coordinated regulation of power data center resources based on operational status prediction. Background Technology

[0002] Existing power data center resource coordination and control methods based on operational status prediction typically predict future operational risks based on currently collected data such as server load, power supply load, cooling status, and ambient temperature. When risks of temperature rise, power increase, or resource insufficiency are predicted, control actions such as task migration, enhanced cooling, or power consumption limitation are implemented in advance. This type of approach can reduce the risks of localized overheating, power overload, and service congestion to a certain extent, enabling data centers to shift from passive response to prediction-driven control.

[0003] However, in actual closed-loop operation, once the prediction result triggers a control action, the subsequent operating state is no longer the natural evolution of the predicted object under uninterrupted conditions, but rather the result influenced by task migration, enhanced cooling, power supply limitations, or power consumption control. For example, when the system predicts a temperature rise risk in a certain area and increases the cooling output in advance, the subsequently collected stable temperature state may indicate that the predicted risk actually exists and has been suppressed by control, or it may indicate that the predicted risk itself does not exist and the control action is an over-intervention. Since both situations are reflected in the feedback data as stable temperature, controlled power, or reduced load, existing systems cannot determine whether the original predicted risk actually exists solely based on the feedback results after control.

[0004] Therefore, existing methods easily misinterpret the stable feedback caused by control actions as the accuracy of the predictive model or the necessity of the control strategy. This leads the system to continuously reinforce conservative strategies such as early cooling, early task migration, or early power reserve in subsequent operation, resulting in increased cooling energy consumption, higher task migration frequency, decreased power resource utilization, and contamination of model training samples by control results. Thus, how to identify the masking effect of control actions on subsequent feedback in the collaborative control closed loop, and how to determine the original authenticity of predicted risks under uncontrolled conditions, remains an unresolved problem in existing predictive power data center resource collaborative control methods. Summary of the Invention

[0005] The purpose of this invention is to address the problem mentioned in the background art that existing methods easily misinterpret the stable feedback caused by control actions as accurate prediction models or necessary control strategies, leading the system to continuously reinforce conservative strategies such as early cooling, early task migration, or early power supply margin in subsequent operation, resulting in increased cooling energy consumption, increased task migration frequency, decreased power supply resource utilization, and contamination of model training samples by control results. Therefore, this invention proposes a power data center resource collaborative control method based on operational status prediction.

[0006] In terms of implementation, this invention provides a method for coordinated control of power data center resources based on operational status prediction, the method comprising: S1: Obtain the operating status data of the power data center within the current control cycle, and predict the operating risks within the future prediction window based on the operating status data to obtain the prediction risk results; S2, Based on the predicted risk results, generate the original risk image results under the condition that the current round of coordinated control actions have not been implemented. The original risk image results are used to characterize the expected evolution state of the operational risk when it has not been intervened by the current round of coordinated control actions. S3. Generate the current round of coordinated regulation strategy based on the predicted risk results, and execute the corresponding coordinated regulation actions according to the current round of coordinated regulation strategy. At the same time, record the target, action type and action intensity of the coordinated regulation actions to obtain the regulation action recording results. S4. After the coordinated control action is executed, the actual operation feedback data of the power data center within the future prediction window is collected, and the actual operation feedback data is compared with the original risk image result to obtain the feedback offset result. S5. Based on the correspondence between the feedback offset results and the control action recording results, identify the degree to which the actual operation feedback data is affected by the coordinated control action, and obtain the control masking identification result. S6, based on the control masking identification results, corrects the coordinated control strategy for the next control cycle, and executes the resource coordinated control for the next control cycle according to the corrected coordinated control strategy.

[0007] Optionally, the step of generating the original risk snapshot result under the condition that the current round of coordinated control actions has not been implemented, based on the predicted risk results, is as follows: Based on the predicted risk results, determine the target risk area, risk type, and future prediction window; Extract the server load sequence, rack power sequence, power supply branch load sequence, temperature change sequence, and cooling output sequence corresponding to the target risk area within the current control cycle from the operational status data; Based on the last sampling moment before the coordinated control action is executed, the server load sequence, cabinet power sequence, power supply branch load sequence, temperature change sequence, and cooling output sequence are extended to the future prediction window to obtain the uncontrolled load extension sequence, uncontrolled power extension sequence, uncontrolled power supply load extension sequence, uncontrolled temperature extension sequence, and uncontrolled cooling output extension sequence. Based on the risk type, select the uncontrolled risk extension sequence corresponding to the risk type from the uncontrolled load extension sequence, uncontrolled power extension sequence, uncontrolled power supply load extension sequence, uncontrolled temperature extension sequence, and uncontrolled cooling output extension sequence; By binding the uncontrolled risk extension sequence with the target risk area, risk type, and future prediction window, the original risk image result is generated.

[0008] Optionally, the steps of generating a coordinated regulation strategy for this round based on the predicted risk results, executing corresponding coordinated regulation actions according to the strategy, and recording the target, type, and intensity of the coordinated regulation actions to obtain the regulation action recording results include: Based on the target risk area and risk type in the predicted risk results, determine the resource objects to be regulated corresponding to the target risk area; When the risk type is load concentration risk, generate a task migration strategy that includes the business tasks to be migrated and the hosting servers; When the risk type is power increase risk, a power limiting strategy is generated that includes the power limiting object and the power limiting range; When the risk type is that the power supply branch is close to the upper limit risk, a power supply protection strategy is generated that includes the target power supply branch and the task allocation limit range. When the risk type is temperature rise risk, generate a cooling enhancement strategy that includes the target cooling equipment and the increase in cooling output; When the risk type is insufficient cooling capacity risk, a joint heat reduction strategy is generated that includes task migration objects, power limitation objects, and cooling output enhancement objects. The task migration strategy, power limiting strategy, power supply protection strategy, cooling enhancement strategy, or combined heat reduction strategy are used as the collaborative control strategy for this round, and control commands are sent to the corresponding business scheduling system, server management system, power supply control system, or cooling control system to execute the collaborative control action. When performing coordinated control actions, the target, type, intensity, start time, and end time of each coordinated control action are recorded to obtain the control action recording results.

[0009] Optionally, the steps to compare the actual operational feedback data with the original risk snapshot results to obtain the feedback offset results are as follows: Between the start of the coordinated control action and the end of the future prediction window, the actual feedback sequence corresponding to the target risk area is collected according to the preset sampling interval. Extract the target risk area, risk type, future prediction window, and uncontrolled risk extension sequence from the original risk snapshot results; Based on the future forecast window, the actual feedback sequence and the uncontrolled risk extension sequence are aligned according to the sampling time to obtain the aligned feedback sequence and the aligned image sequence; Based on the risk type, select the actual risk feedback sequence from the aligned feedback sequence that corresponds to the uncontrolled risk extension sequence; By comparing the numerical differences between the actual risk feedback sequence and the aligned image sequence at the same sampling time, multiple single-point feedback offset values ​​are obtained. Multiple single-point feedback offset values ​​are arranged in the order of sampling time to obtain a feedback offset sequence. The feedback offset sequence, target risk area, risk type and future prediction window are then bound together to generate the feedback offset result.

[0010] Optionally, based on the correspondence between the feedback offset results and the control action recording results, the steps to identify the degree to which the actual operation feedback data is affected by the coordinated control actions and obtain the control masking identification results are as follows: Based on the correspondence between the feedback offset results and the control action recording results, the action landing point echo index and the risk trajectory reversal index are calculated. The action landing point echo index and the risk trajectory reversal index are added together to obtain the influence index. Based on the influence index and the preset influence index threshold, the degree of influence of the actual operation feedback data on the coordinated control action is identified, and the control masking identification result is obtained.

[0011] 6. The method for coordinated control of power data center resources based on operational status prediction according to claim 5, characterized in that the calculation step of the action landing echo index includes: Read the feedback offset sequence from the feedback offset result, and divide the absolute value of each feedback offset value by the absolute value of the maximum feedback offset value to obtain the feedback offset normalization sequence. When the absolute value of the maximum feedback offset value is zero, all normalized feedback offset values ​​are set to zero. Read the start time of each coordinated control action from the control action record results, and determine the first sampling time that is no earlier than the corresponding start time as the action landing point; For any coordinated control action, construct a mirror sampling group with equal distances before and after the action landing point as the center. When the normalized feedback offset value behind the action landing point is greater than the normalized feedback offset value in front, divide the difference between the two by the larger of the two values ​​to obtain the dominant value behind the mirror. Otherwise, set the dominant value behind the mirror to zero. Complementary multiplication and compression are performed on the dominance values ​​of each mirror image corresponding to the coordinated control action to obtain the echo dominance value of the action. Based on the echo dominance value behind the action, the number of consecutive mirror sampling groups that satisfy the echo dominance value behind the action not less than the echo dominance value behind the action is counted to obtain the echo length close to the landing point, and then divided by the number of mirror sampling groups to obtain the landing point proximity. The echo value of a single action is obtained by multiplying the echo dominance value of the action by the proximity of the landing point and taking the square root. The echo values ​​of the single action landing points corresponding to each coordinated control action are then multiplied and compressed to obtain the action landing point echo index.

[0012] Optionally, the calculation steps for the risk trajectory reversal index include: Read the original risk extension sequence and the actual risk feedback sequence from the feedback offset results, and determine the risk advancement direction according to the risk type; According to the direction of risk advancement, the changes of each sampling time in the two sequences relative to the first sampling time are oriented and normalized to obtain the original same-direction risk trajectory and the actual same-direction feedback trajectory. For adjacent sampling times, calculate the risk advance amount of the original same-direction risk trajectory, the feedback retreat amount of the actual same-direction feedback trajectory, and the trajectory separation amount of the two respectively. Then, compress the three by dividing the original value by one and adding the original value, multiply them by the cube root, and obtain the single-segment return deduction value. Adjacent time periods with a single-segment return deduction value greater than zero and continuous time are merged into a continuous return segment. The segment return value is calculated based on the square root of the product of all single-segment return deduction values ​​within the continuous return segment and the continuous proportion of the continuous return segment. The maximum value among all segment return values ​​is used as the risk trajectory return index. When continuous return segments cannot be formed, the risk trajectory return index is set to zero.

[0013] Optionally, the steps for identifying the degree to which actual operational feedback data is affected by coordinated control actions based on the impact index and a preset impact index threshold, and obtaining the control masking identification result, include: The preset impact index thresholds include a first impact index threshold and a second impact index threshold, and the first impact index threshold is less than the second impact index threshold; When the impact index is less than the first impact index threshold, the degree to which the actual operation feedback data is affected by the coordinated control action is determined as low occlusion, and this is used as the low occlusion identification result. When the impact index is not less than the first impact index threshold and less than the second impact index threshold, the degree to which the actual operation feedback data is affected by the coordinated control action is determined as medium occlusion, and this is used as the occlusion identification result. When the impact index is not less than the second impact index threshold, the degree to which the actual operation feedback data is affected by the coordinated control action is determined as high occlusion, and this is used as the high occlusion identification result.

[0014] Optionally, the steps of modifying the coordinated control strategy for the next control cycle based on the control masking identification results, and executing the resource coordinated control for the next control cycle according to the modified coordinated control strategy are as follows: Read the occlusion level, target risk area and risk type from the control occlusion identification results, and obtain the initial collaborative control strategy corresponding to the target risk area and risk type in the next control cycle; When the occlusion level is low, the initial coordinated control strategy is used as the revised coordinated control strategy for the next control cycle, and resource coordinated control is performed according to the revised coordinated control strategy. When the shading level is medium shading, the intensity of the coordinated control action in the initial coordinated control strategy is reduced by one level to obtain the revised coordinated control strategy for the next control cycle, and resource coordinated control is executed according to the revised coordinated control strategy. Among them, reducing the intensity of the coordinated control action by one level includes at least one of the following: reducing the number of business tasks to be migrated, reducing the power limit range, narrowing the task allocation limit range, reducing the fan speed increase range, reducing the coolant flow rate increase range, or reducing the chilled water valve opening increase range. When the occlusion level is high, the coordinated control actions in the initial coordinated control strategy are changed from direct execution to segmented execution to obtain the revised coordinated control strategy for the next control cycle, and resource coordinated control is executed according to the revised coordinated control strategy. The segmented execution method includes first executing the coordinated control action intensity of the preset proportion in the initial coordinated control strategy, collecting stage feedback data in the next control cycle, and then determining whether to continue executing the remaining coordinated control action intensity based on the stage feedback data. After the resource coordination and regulation is completed in the next control cycle, record the strategy type, target, action intensity, and execution time of the revised coordination and regulation strategy, and use the recorded results as the record results for the next round of regulation actions.

[0015] The beneficial effects of this invention are: This invention proposes a method for coordinated control of power data center resources based on operational status prediction. After predicting the operational risks within the future prediction window, it first generates a projection of the original risk under conditions where coordinated control actions were not executed in the current round. Then, it executes the corresponding coordinated control actions and records the target, type, and intensity of these actions. This allows the system to no longer directly judge the accuracy of the prediction model based solely on results such as temperature stability, power control, or load reduction after control, but rather compares the actual operational feedback data with the projection of the original risk and, combined with the recorded control actions, identifies the degree to which the actual operational feedback data is affected by the coordinated control actions. This distinguishes between "predicted risks that actually exist but have been suppressed by coordinated control actions" and "predicted risks that are not actually present." There are two situations that are similar in feedback performance but different in technical meaning: "existence and coordinated control actions constitute over-intervention". Furthermore, by modifying the coordinated control strategy for the next control cycle based on the control masking identification results, it is possible to avoid mistaking the stable feedback caused by the coordinated control actions as the accuracy of the prediction model or the necessity of the control strategy. This reduces the situation where early cooling, early task migration, and power supply margin reservation are continuously reinforced in subsequent operation, thereby reducing ineffective cooling energy consumption and unnecessary task migration frequency, improving the utilization rate of power supply resources, and reducing the interference caused by the feedback data after the control actions are rewritten to the prediction model correction process. This enables the coordinated control process of power data center resources to more accurately identify the authenticity of the original prediction risks while ensuring operational safety, and improve the rationality and long-term stability of the closed-loop control strategy. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a method for coordinated control of power data center resources based on operational status prediction, provided as an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0018] This invention provides a method for coordinated control of power data center resources based on operational status prediction. See also... Figure 1 , Figure 1 A flowchart illustrating a method for coordinated control of power data center resources based on operational status prediction, provided in an embodiment of the present invention. The method includes the following steps: S1: Obtain the operating status data of the power data center within the current control cycle, and predict the operating risks within the future prediction window based on the operating status data to obtain the prediction risk results; S2, Based on the predicted risk results, generate the original risk image results under the condition that the current round of coordinated control actions have not been implemented. The original risk image results are used to characterize the expected evolution state of the operational risk when it has not been intervened by the current round of coordinated control actions. S3. Generate the current round of coordinated regulation strategy based on the predicted risk results, and execute the corresponding coordinated regulation actions according to the current round of coordinated regulation strategy. At the same time, record the target, action type and action intensity of the coordinated regulation actions to obtain the regulation action recording results. S4. After the coordinated control action is executed, the actual operation feedback data of the power data center within the future prediction window is collected, and the actual operation feedback data is compared with the original risk image result to obtain the feedback offset result. S5. Based on the correspondence between the feedback offset results and the control action recording results, identify the degree to which the actual operation feedback data is affected by the coordinated control action, and obtain the control masking identification result. S6, based on the control masking identification results, corrects the coordinated control strategy for the next control cycle, and executes the resource coordinated control for the next control cycle according to the corrected coordinated control strategy.

[0019] Based on the operational status prediction-based collaborative control method for power data center resources provided by this invention, after predicting the operational risks within the future prediction window, the method first generates the original risk image under the condition that the current round of collaborative control actions has not been executed. Then, it executes the corresponding collaborative control actions and records the target, type, and intensity of the actions. This allows the system, after obtaining actual operational feedback data, to no longer directly judge the accuracy of the prediction model based solely on results such as temperature stability, power control, or load reduction after control. Instead, it compares the actual operational feedback data with the original risk image and, combined with the control action recording results, identifies the degree to which the actual operational feedback data is affected by the collaborative control actions. This allows for the distinction between "predicted risks that actually exist but have been suppressed by collaborative control actions" and "predicted risks that are inherently suppressed by collaborative control actions." The two situations, "the entity does not exist but the coordinated control action is an over-intervention" and "the coordinated control action is not an over-intervention," are similar in feedback performance but have different technical meanings. Furthermore, by modifying the coordinated control strategy for the next control cycle based on the control masking identification results, it is possible to avoid mistaking the stable feedback caused by the coordinated control action as the accuracy of the prediction model or the necessity of the control strategy. This reduces the situation where premature cooling, premature task migration, and power supply margin reservation are continuously reinforced in subsequent operations, thereby reducing ineffective cooling energy consumption and unnecessary task migration frequency, improving power supply resource utilization, and reducing the interference caused by the feedback data rewritten by the controlled action on the prediction model correction process. This enables the coordinated control process of power data center resources to more accurately identify the authenticity of the original prediction risks while ensuring operational safety, and improve the rationality and long-term stability of the closed-loop control strategy.

[0020] In one embodiment, S1, the operating status data of the power data center within the current control cycle is acquired, and the operating risks within the future prediction window are predicted based on the operating status data to obtain the predicted risk results. Specifically, at the beginning of the current control cycle, the operating status data of the power data center within the current moment and several consecutive sampling moments before the current control cycle are collected through the server management platform, environmental monitoring system, power distribution monitoring system, cooling control system, and business scheduling system. The operating status data includes server load data, server power consumption data, rack power data, power supply branch load data, UPS operating data, PDU current data, precision air conditioning operating data, CDU operating data, cold aisle temperature data, hot aisle temperature data, rack inlet air temperature data, rack outlet air temperature data, business task queue data, and task migration status data. Among them, server load data is used to characterize computing resource pressure, rack power data and power supply branch load data are used to characterize power supply carrying pressure, precision air conditioning operating data, CDU operating data, and temperature data are used to characterize cooling response status, and business task queue data is used to characterize the source of subsequent load growth. After acquiring the above operating status data, the data is processed according to equipment area, rack number, and sampling time. Alignment is performed to form a regional operating state sequence that reflects the correlation between "business load changes, power consumption changes, power supply pressure changes, temperature changes, and cooling output changes" within the same region. This regional operating state sequence is then input into a pre-trained operating state prediction model, or matched with similar historical operating state segments, to predict the load change trend, power change trend, temperature change trend, and power supply margin change trend of each region within the future prediction window. When the prediction results show that a certain region has a continuously rising temperature, power approaching the upper limit of the power supply branch, cooling output approaching the adjustment limit but temperature continuing to rise, and business load changes within the future prediction window, the system will take appropriate action. When a service queue is about to be concentrated in a certain rack group, the area is identified as having operational risks, and a predicted risk result is generated. For example, if the GPU utilization rate in area A continues to increase during the current control cycle, the rack power rises rapidly from the original medium level, and the corresponding cold aisle temperature and CDU return liquid temperature rise simultaneously, and the service scheduling system shows that a batch of training tasks will soon enter area A, then the operational status prediction model can predict that there is a risk of temperature rise and a risk of increased pressure on the power supply branch in area A within the next ten minutes, and output the risk area, risk type, risk occurrence time range, and risk change trend as the predicted risk result.

[0021] In one embodiment, S2, based on the predicted risk results, the original state risk image result under the condition that the current round of coordinated control actions has not been performed is generated. The original state risk image result is used to characterize the expected evolution state of the operational risk when it has not been intervened by the current round of coordinated control actions. Based on the predicted risk results, the steps to generate the original risk snapshot under the condition that the current round of coordinated control actions has not been implemented are as follows: Based on the predicted risk results, determine the target risk area, risk type, and future prediction window; Extract the server load sequence, rack power sequence, power supply branch load sequence, temperature change sequence, and cooling output sequence corresponding to the target risk area within the current control cycle from the operational status data; Based on the last sampling moment before the coordinated control action is executed, the server load sequence, cabinet power sequence, power supply branch load sequence, temperature change sequence, and cooling output sequence are extended to the future prediction window to obtain the uncontrolled load extension sequence, uncontrolled power extension sequence, uncontrolled power supply load extension sequence, uncontrolled temperature extension sequence, and uncontrolled cooling output extension sequence. Based on the risk type, select the uncontrolled risk extension sequence corresponding to the risk type from the uncontrolled load extension sequence, uncontrolled power extension sequence, uncontrolled power supply load extension sequence, uncontrolled temperature extension sequence, and uncontrolled cooling output extension sequence; By binding the uncontrolled risk extension sequence with the target risk area, risk type, and future prediction window, the original risk image result is generated.

[0022] It should be noted that, in this step, the predicted risk results should include at least the target risk area, the risk type, and the future prediction window. The target risk area refers to the data center partition, cabinet group, power supply branch coverage area, or cooling coverage area that is predicted to have operational risks. The risk type refers to the specific direction of resource anomalies corresponding to the predicted risk, including load concentration risk, power increase risk, power supply branch approaching its limit risk, temperature increase risk, and insufficient cooling capacity risk. The future prediction window refers to the prediction time range extending forward from the end of the current control cycle, such as the next five minutes, ten minutes, or fifteen minutes.After identifying the target risk area, the system extracts the server load sequence, rack power sequence, power supply branch load sequence, temperature change sequence, and cooling output sequence for that target risk area within the current control cycle from the operational status data. The server load sequence can be formed by arranging data such as CPU utilization, GPU utilization, memory usage, and task queue length over time. The rack power sequence can be formed by arranging the total rack power or server power consumption over time. The power supply branch load sequence can be formed by arranging PDU current, UPS output load rate, or distribution branch load rate over time. The temperature change sequence can be formed by arranging cold aisle temperature, rack inlet air temperature, rack outlet air temperature, or CDU return air temperature. Liquid temperatures are arranged over time, and the refrigeration output sequence can be formed by arranging precision air conditioner fan speed, supply air temperature, chilled water valve opening, CDU coolant flow rate, or pump frequency over time. Subsequently, the system uses the last sampling moment before the coordinated control action is executed as the extension starting point. Based on the direction and magnitude of change of each sequence within the current control cycle, the system extends each sequence to the future prediction window, forming uncontrolled load extension sequences, uncontrolled power extension sequences, uncontrolled power supply load extension sequences, uncontrolled temperature extension sequences, and uncontrolled refrigeration output extension sequences. These sequences represent the possible continued development of various resource states without executing the current round of coordinated control action. When selecting uncontrolled risk extension sequences... When extending the sequence, the system selects the appropriate sequence based on the risk type. If the risk type is load concentration risk, the uncontrolled load extension sequence is selected as the uncontrolled risk extension sequence; if the risk type is power increase risk, the uncontrolled power extension sequence is selected as the uncontrolled risk extension sequence; if the risk type is power supply branch approaching its upper limit risk, the uncontrolled power supply load extension sequence is selected as the uncontrolled risk extension sequence; if the risk type is temperature increase risk, the uncontrolled temperature extension sequence is selected as the uncontrolled risk extension sequence; if the risk type is insufficient cooling capacity risk, both the uncontrolled cooling output extension sequence and the uncontrolled temperature extension sequence are selected as the uncontrolled risk extension sequence to reflect the cooling output... The system can detect a continuous rise in temperature, but the temperature still shows an upward trend. For example, if the predicted risk result shows that there is a risk of temperature rise in area A within the next ten minutes, the system will extract the corresponding temperature evolution result from the uncontrolled temperature extension sequence of area A as the uncontrolled risk extension sequence, and bind it with area A, the temperature rise risk, and the prediction window for the next ten minutes to generate the original risk image result. If the predicted risk result shows that there is a risk of power supply branches in area B approaching their upper limit, the uncontrolled power supply load extension sequence of area B will be selected as the uncontrolled risk extension sequence, so that the subsequent actual operation feedback data can be compared with the uncontrolled risk extension sequence to determine whether the feedback stability state is caused by the current round of coordinated control actions.

[0023] In one embodiment, S3, the steps of generating a coordinated control strategy for this round based on the predicted risk results, executing corresponding coordinated control actions according to the coordinated control strategy for this round, and recording the target, type, and intensity of the coordinated control actions to obtain the control action recording results are as follows: Based on the target risk area and risk type in the predicted risk results, determine the resource objects to be regulated corresponding to the target risk area. The resource objects to be regulated include at least one of the target server, target cabinet, target power supply branch, target cooling equipment, and target business task. When the risk type is load concentration risk, the business tasks to be migrated are selected from the target risk area, and the receiving servers that meet the requirements of computing margin, power supply margin and cooling margin are selected from the non-target risk area, and a task migration strategy containing the business tasks to be migrated and the receiving servers is generated. When the risk type is power increase risk, identify the target server or target cabinet in the target risk area whose power increase exceeds the preset range, and generate a power limiting policy that includes the power limiting object and the power limiting range. When the risk type is that the power supply branch is close to the upper limit risk, the target power supply branch corresponding to the target risk area is determined, and the allocation of new business tasks to the server covered by the target power supply branch is prohibited. A power supply protection policy containing the target power supply branch and the task allocation restriction range is generated. When the risk type is temperature rise risk, the target refrigeration equipment corresponding to the target risk area is identified, and a refrigeration enhancement strategy is generated that includes the target refrigeration equipment, the object to be enhanced in refrigeration output, and the magnitude of the enhancement in refrigeration output. When the risk type is insufficient cooling capacity risk, identify the target servers and corresponding target cooling equipment in the target risk area whose heat load needs to be reduced, and generate a joint heat reduction strategy that includes task migration objects, power limiting objects and cooling output enhancement objects. The task migration strategy, power limiting strategy, power supply protection strategy, cooling enhancement strategy or joint cooling strategy are used as the collaborative control strategy for this round, and corresponding control commands are sent to the business scheduling system, server management system, power supply control system or cooling control system according to the collaborative control strategy for this round to execute the collaborative control action. When executing coordinated control actions, the target, action type, action intensity, start time, and end time of each coordinated control action are recorded to obtain the control action record results. The target includes at least one of the following: the business task being migrated, the receiving server, the server with limited power, the power supply branch with limited task allocation, and the cooling equipment with increased output. The action type includes at least one of the following: task migration, power limitation, task allocation limitation, and cooling output enhancement. The action intensity includes the number of migrated tasks, the computing power occupancy of the migrated tasks, the power limitation range, the task allocation limitation range, the increase in fan speed, the increase in coolant flow rate, or the increase in chilled water valve opening.

[0024] It should be noted that, specifically: after obtaining the predicted risk results, the system first reads the target risk area and risk type, and determines which resource objects the risk needs to affect, such as servers, racks, power supply branches, cooling equipment, or business tasks; if the risk type is a load concentration risk, then selects migrateable low-priority tasks, batch processing tasks, or non-real-time computing tasks from the target risk area, and selects servers in other areas that simultaneously meet the requirements of idle computing power, power supply branches not approaching their limits, and sufficient cooling capacity as receiving servers, forming a task migration strategy; if the risk type is a power increase risk, then locks servers or racks with significant power increases and issues power limiting or task reduction instructions; if the risk type is a power supply branch approaching its limit risk, then locks the corresponding PDU, UPS, or distribution branch, restricting new business tasks from being assigned to servers covered by that branch; if the risk type is a temperature increase risk, then identifies the precision air conditioners, fans, or CDUs covering the area and increases fan speed, coolant flow, or chilled water valve opening; if the risk type is a insufficient cooling capacity risk, then simultaneously executes task migration, power limiting, and cooling enhancement to reduce the heat load of the target area. For example, if the system predicts that there is a risk of temperature rise in area A in the next ten minutes, it identifies the precision air conditioner and CDU corresponding to area A as the target refrigeration equipment, generates a refrigeration enhancement strategy to increase fan speed and coolant flow, and sends control commands to the refrigeration control system for execution. At the same time, it records the action applied to the precision air conditioner and CDU in area A, the action type as refrigeration output enhancement, the action intensity as the increase in fan speed and coolant flow, and the corresponding start and end times of execution. This forms the control action record result, which is used to judge whether the actual feedback has been changed by the control action.

[0025] In one embodiment, S4, after the coordinated control action is executed, the actual operation feedback data of the power data center within the future prediction window is collected, and the actual operation feedback data is compared with the original risk image result to obtain the feedback offset result. Between the start of the coordinated control action and the end of the future prediction window, the actual feedback sequence corresponding to the target risk area is collected according to the preset sampling interval. The actual feedback sequence includes at least one of the following: actual server load sequence, actual rack power sequence, actual power supply branch load sequence, actual temperature change sequence, and actual cooling output sequence. Extract the target risk area, risk type, future prediction window, and uncontrolled risk extension sequence from the original risk snapshot results; Based on the future forecast window, the actual feedback sequence and the uncontrolled risk extension sequence are aligned according to the sampling time to obtain the aligned feedback sequence and the aligned image sequence; Based on the risk type, select the actual risk feedback sequence from the aligned feedback sequence that corresponds to the uncontrolled risk extension sequence; By comparing the numerical differences between the actual risk feedback sequence and the aligned image sequence at the same sampling time, multiple single-point feedback offset values ​​are obtained. Multiple single-point feedback offset values ​​are arranged in the order of sampling time to obtain a feedback offset sequence. The feedback offset sequence, target risk area, risk type and future prediction window are then bound together to generate the feedback offset result.

[0026] It should be noted that, specifically, step S4 is used to determine "how much the actual operating state has changed compared to the original uncontrolled state" after the current round of coordinated control actions has been executed. The system first uses the start time of the coordinated control action as the starting point for actual feedback collection and the end time of the future prediction window as the collection endpoint. It then continuously collects actual operating data of the target risk area at fixed sampling intervals. For example, if the target risk area is area A and the risk type is temperature rise risk, after the system executes the cooling enhancement action in area A, it collects actual temperature data such as the cabinet inlet air temperature, cold aisle temperature, and CDU return liquid temperature in area A every minute to form an actual temperature change sequence. At the same time, the system reads the previously generated uncontrolled risk extension sequence from the original risk snapshot results. This uncontrolled risk extension sequence indicates the possible temperature change state of area A within the future prediction window if the current round of control actions were not executed. Subsequently, the system aligns the actual feedback sequence with the uncontrolled risk extension sequence at the same sampling time, for example, aligning them to the 1st minute, 2nd minute, 3rd minute, and up to the 10th minute to avoid comparison distortion caused by inconsistent sampling times. After alignment, the system selects the corresponding actual risk feedback sequence based on the risk type. If the risk type is temperature rise risk, the actual temperature change sequence is selected; if the risk type is power supply branch approaching its upper limit risk, the actual power supply branch load sequence is selected; if the risk type is power rise risk, the actual cabinet power sequence is selected. Then, the system compares the differences between the actual risk feedback sequence and the aligned image sequence at the same sampling time. For example, if the original risk image result shows that the temperature in area A is expected to rise to 28℃ in the 5th minute, but the actual feedback shows that the temperature in the 5th minute after control is 25℃, then the single-point feedback offset value corresponding to the 5th minute is 3℃; if the image temperature in the 6th minute is 29℃, and the actual temperature is 25.5℃, then the single-point feedback offset value corresponding to the 6th minute is 3.5℃. The system arranges the single-point feedback offset values ​​obtained at each sampling time in chronological order to form a feedback offset sequence. This sequence is then linked to the target risk area, risk type, and future prediction window to generate a feedback offset result. This result allows subsequent steps to determine whether the actual steady state is largely consistent with the original risk evolution or significantly deviates from it. A large offset usually indicates that the actual feedback has been significantly affected by coordinated control actions, and the predicted risk can no longer be simply proven by "temperature stabilization" or "power decrease."

[0027] In one embodiment, S5: Based on the correspondence between the feedback offset result and the control action recording result, the step of identifying the degree to which the actual operation feedback data is affected by the coordinated control action to obtain the control masking identification result is as follows: Based on the correspondence between the feedback offset results and the control action recording results, the action landing echo index (to determine whether the feedback offset closely follows the control action) and the risk trajectory reversal index (to determine whether the actual feedback is pulled back relative to the original risk shadow) are calculated. The action landing echo index and the risk trajectory reversal index are added together to obtain the influence index. Based on the influence index and the preset influence index threshold, the degree to which the actual operation feedback data is affected by the coordinated control action is identified, and the control masking identification result is obtained.

[0028] In one implementation, the steps for calculating the echo index of the action landing point are as follows: Read the feedback offset sequence from the feedback offset result. The feedback offset sequence includes multiple sampling times and the feedback offset value corresponding to each sampling time. Take the absolute value of each feedback offset value in the feedback offset sequence, and use the maximum value among all feedback offset values ​​as the normalization benchmark. Divide the absolute value of each feedback offset value by the normalization benchmark to obtain the feedback offset normalization sequence. When the normalization benchmark is zero, set the normalized feedback offset value corresponding to each sampling time in the feedback offset normalization sequence to zero. Read the start time of each coordinated control action from the control action record results, and determine the first sampling time no earlier than the start time of execution as the action landing point of the coordinated control action in the feedback offset normalization sequence; For any coordinated control action, with the action landing point of the coordinated control action as the center, the same number of sampling points are selected forward and backward respectively to form multiple front and back mirror sampling groups. Each front and back mirror sampling group includes a normalized feedback offset value located before the action landing point and a normalized feedback offset value located after the action landing point, and the sampling interval between the two from the action landing point is the same. For each before-and-after mirror sampling group, the normalized feedback offset value after the action landing point is compared with the normalized feedback offset value before the action landing point. When the normalized feedback offset value after the action landing point is greater than the normalized feedback offset value before the action landing point, the difference between the two is divided by the larger of the two values ​​to obtain the dominance value of the mirror back side corresponding to the before-and-after mirror sampling group. When the normalized feedback offset value after the action landing point is not greater than the normalized feedback offset value before the action landing point, the dominance value of the mirror back side corresponding to the before-and-after mirror sampling group is set to zero. For any coordinated control action, subtract the corresponding mirror back-side dominance value from each mirror back-side dominance value by one to obtain multiple mirror back-side non-dominant values; multiply all mirror back-side non-dominant values ​​to obtain a non-dominant product value; perform a square root operation on the non-dominant product value, with the square root number being the number of front and rear mirror sampling groups corresponding to the coordinated control action; then subtract the square root result from one to obtain the action back-side echo dominance value corresponding to the coordinated control action; when there are no front and rear mirror sampling groups for the coordinated control action, set the action back-side echo dominance value corresponding to the coordinated control action to zero. Using the echo dominance value behind the action as the echo point screening criterion for the coordinated control action, we determine whether the echo dominance value behind the mirror of each front and rear mirror sampling group is not less than the echo dominance value behind the action, and determine the front and rear mirror sampling groups that are not less than the echo dominance value behind the action as the landing echo points. Starting from the first front and rear mirror sampling group immediately after the action landing point, count the number of consecutive front and rear mirror sampling groups that are identified as landing point echo points to obtain the landing point close echo length; when the first front and rear mirror sampling group immediately after the action landing point is not identified as landing point echo point, set the landing point close echo length to zero. Divide the length of the echo close to the landing point by the number of front and rear mirror sampling groups corresponding to the coordinated control action to obtain the landing point closeness corresponding to the coordinated control action; when there are no front and rear mirror sampling groups for the coordinated control action, the landing point closeness is set to zero. Multiply the echo dominance value behind the action by the proximity of the landing point, and then perform a square root operation on the multiplication result to obtain the single action landing point echo value corresponding to the coordinated control action. When multiple coordinated control actions exist in the control action recording results, the echo value of the single action landing point corresponding to each coordinated control action is subtracted by one to obtain multiple single action non-echo values; all single action non-echo values ​​are multiplied together to obtain a non-echo multiplication value; the non-echo multiplication value is square-rooted, and the number of square roots is the number of coordinated control actions; the result after square rooting is subtracted by one to obtain the action landing point echo index; when no coordinated control actions exist in the control action recording results, the action landing point echo index is set to zero.

[0029] It should be noted that the action landing echo index is a quantitative indicator used to measure whether the feedback offset result shows obvious posterior enhancement, close proximity, and continuous diffusion characteristics after the execution landing point of the coordinated control action. In essence, it judges whether the enhanced position of the feedback offset is close to the start time of the coordinated control action, whether the enhancement direction is concentrated after the action landing point, and whether the enhanced segment is continuous. In other words, this index is used to measure whether the time trace of the coordinated control action is left in the actual operation feedback data. The reason why a larger action landing point echo index indicates a greater degree of influence of coordinated control actions on the actual operational feedback data is that, without significant influence from coordinated control actions, the deviation of the actual feedback sequence relative to the original risk image result usually does not stably concentrate after the start of the control action, nor does it show that the deviation value after the action landing point is consistently stronger than that before the action landing point. On the contrary, if the system has just executed cooling enhancement, task migration, power limiting, or power supply protection actions, and the feedback deviation value immediately increases after the action landing point, and multiple consecutive sampling points show that the deviation after the action landing point is stronger than that before the action landing point, it indicates that the actual feedback is no longer the result of the natural evolution of the original operational risk, but has been significantly pulled away from the original risk image result by the control action. For example, the original risk projection results show that the temperature in area A will gradually rise from 26 degrees Celsius to 29 degrees Celsius within the next ten minutes. After the system executes a cooling enhancement action in the second minute, the actual temperature becomes significantly lower than the original risk projection temperature from the second minute onwards. Furthermore, the feedback offset values ​​in the third, fourth, and fifth minutes all continuously increase. This offset is not a random fluctuation but corresponds to the execution point of the cooling enhancement action, resulting in a higher action landing echo index. The higher this index, the more the feedback offset resembles an action echo triggered by coordinated control actions, and the less suitable the actual operational feedback data is to be directly used as a basis for judging the authenticity of the predicted risk. Correspondingly, the control masking identification results tend to be identified as strong masking, meaning that the actual operational feedback data has been significantly rewritten by coordinated control actions. Subsequently, the system cannot simply judge the existence of the predicted risk or the necessity of the control strategy based on feedback phenomena such as "temperature stability, power decrease, and load drop." Instead, it should combine the original risk projection results to further determine whether there is over-control. Conversely, when the echo index of the action landing point is low, it indicates that the feedback offset is not significantly concentrated after the execution of the coordinated control action, or the difference in offset before and after the action landing point is not obvious. The actual operation feedback data is less directly affected by the coordinated control action in this round, and the corresponding control masking identification results tend to be weak or low masking. At this time, the actual operation feedback data still has high reference value for the original predicted risk.

[0030] In one implementation, the calculation steps for the risk trajectory reversal index are as follows: The calculation steps for the risk trajectory reversion index include: Read the original risk extension sequence and the actual risk feedback sequence from the feedback offset results, and determine the risk propagation direction according to the risk type; when the risk value increases, indicating that the risk is more serious, the risk propagation direction is determined to be positive, and when the risk value decreases, indicating that the risk is more serious, the risk propagation direction is determined to be negative. Based on the direction of risk propagation, the original risk extension sequence and the actual risk feedback sequence are homogenized to obtain the original homogenized risk trajectory and the actual homogenized feedback trajectory. The homogenization process includes: calculating the change in risk value at each sampling time in the original risk extension sequence relative to the risk value at the first sampling time, and calculating the change in feedback value at each sampling time in the actual risk feedback sequence relative to the feedback value at the first sampling time; when the risk propagation direction is positive, the direction of the change is retained; when the risk propagation direction is negative, the change is reversed; then, using the maximum absolute value of the change after all orientations as the normalization benchmark, the change after each orientation is normalized to obtain the original homogenized risk trajectory and the actual homogenized feedback trajectory; when the normalization benchmark is zero, the risk trajectory reversal index is set to zero. For two adjacent sampling moments in the original unidirectional risk trajectory and the actual unidirectional feedback trajectory, calculate the original risk advance, the actual feedback retreat, and the trajectory separation. Specifically, subtract the original unidirectional risk value from the previous sampling moment from the original unidirectional risk value at the later sampling moment, and set the result to zero when the difference is less than zero to obtain the original risk advance. Subtract the actual unidirectional feedback value from the previous sampling moment from the actual unidirectional feedback value at the later sampling moment, and set the result to zero when the difference is less than zero to obtain the actual feedback retreat. Subtract the actual unidirectional feedback value from the original unidirectional risk value at the later sampling moment from the actual unidirectional feedback value at the later sampling moment, and set the result to zero when the difference is less than zero to obtain the trajectory separation. The original risk advance amount, actual feedback retreat amount, and trajectory separation amount are compressed by dividing the original amount by one and adding the original amount, respectively, to obtain the risk advance compression value, feedback retreat compression value, and trajectory separation compression value. The risk advance compression value, feedback retreat compression value, and trajectory separation compression value are multiplied together, and the cube root operation is performed on the multiplication result to obtain the single-segment return deduction value corresponding to two adjacent sampling times. Adjacent time periods with a single-segment deduction value greater than zero are identified as valid turnaround segments, and temporally consecutive valid turnaround segments are merged into continuous turnaround segments. For any continuous turnaround segment, multiply the sum of all single-segment turnaround values ​​within the continuous turnaround segment, and take the square root of the sum of the effective turnaround segments within the continuous turnaround segment to obtain the segment turnaround intensity; divide the number of effective turnaround segments within the continuous turnaround segment by the number of all adjacent time periods to obtain the segment duration ratio; multiply the segment turnaround intensity by the segment duration ratio, and take the square root of the multiplication result to obtain the segment turnaround value of the continuous turnaround segment; The maximum value among all consecutive return segments is used as the risk trajectory return index; when there are no consecutive return segments, the risk trajectory return index is set to zero.

[0031] It should be noted that the Risk Trajectory Reversal Index is a quantitative indicator used to measure the degree of divergence between the original risk extension sequence and the actual risk feedback sequence, where "the original state continues to develop in the direction of risk, while the actual feedback retreats in the direction of safety." It does not focus on whether the actual feedback result has stabilized, nor does it simply compare the difference between the actual value and the predicted value. Instead, it judges whether the risk trajectory, which should have continued to worsen without the implementation of the current round of coordinated control actions, has been pulled back in the opposite direction by the actual feedback trajectory after the implementation of coordinated control actions. Specifically, if the original risk extension sequence shows that the temperature, power, load, or power supply branch pressure in the target risk area continues to advance in the direction of risk, while the actual risk feedback sequence shows a decrease, retreat, or movement away from the risk boundary within the same time period, and this reversal is not an accidental fluctuation at a single sampling point but persists continuously over multiple adjacent time periods, then it indicates that the actual operational feedback has significantly deviated from the original risk evolution path under uncontrolled conditions. The reason why a larger risk trajectory reversal index indicates a greater degree of influence of coordinated control actions on actual operational feedback data is that the index requires three conditions to be met simultaneously: First, the original risk trajectory still exhibits a risk propagation, indicating that the risk would continue to worsen if the current round of coordinated control actions were not implemented; second, the actual feedback trajectory exhibits a feedback pullback, indicating that the actual state did not continue to develop in the risk direction after the implementation of coordinated control actions, but rather retreated in the safe direction; third, there is a trajectory separation between the actual feedback trajectory and the original risk trajectory, indicating that the actual feedback has not merely slowed down naturally, but has fallen below or deviated from the original risk trajectory. Only when these three conditions are met simultaneously and consecutively will the risk trajectory reversal index increase. Therefore, this index can reflect the "pull-back" or "rewriting" effect of coordinated control actions on the actual feedback trajectory. For example, the original risk extension sequence shows that the temperature in area A should gradually rise from 26 degrees Celsius to 29 degrees Celsius within the next ten minutes, indicating that the temperature risk is still advancing. However, after the system performs cooling enhancement and task migration in the second minute, the actual temperature drops from 26 degrees Celsius to 25 degrees Celsius and remains below the original risk extension sequence at multiple subsequent sampling times. This indicates that the actual temperature trajectory has turned back from the original warming trajectory to a safe direction. This turnback is likely due to coordinated control actions rather than the natural disappearance of the original risk. The larger the risk trajectory turnback index, the more obvious the persistence and intensity of this "original risk continues to advance, actual feedback reverses" phenomenon. The actual operational feedback data is less likely to be used directly to prove whether the predicted risk itself actually exists. Correspondingly, the more the control masking identification results tend to be strong masking, that is, the subsequent collected feedback results such as temperature stability, power reduction, or load drop have been significantly affected by coordinated control actions. The system cannot simply assume that the prediction model is accurate or that the control strategy is necessarily necessary. Instead, it should further combine the original risk imprint results to determine whether there is over-control.Conversely, when the risk trajectory reversal index is low, it indicates that the actual feedback trajectory has not deviated significantly from the original risk trajectory, or that the original risk itself has not continued to advance. In this case, the actual operational feedback data is less affected by the trajectory rewriting of coordinated control actions, and the corresponding control masking identification results tend to be weakly masked or lowly masked.

[0032] In one embodiment, the step of identifying the degree to which actual operational feedback data is affected by coordinated control actions based on an impact index and a preset impact index threshold, and obtaining the control masking identification result, includes: The preset impact index thresholds include a first impact index threshold and a second impact index threshold, and the first impact index threshold is less than the second impact index threshold; When the impact index is less than the first impact index threshold, the degree to which the actual operation feedback data is affected by the coordinated control action is determined as low occlusion, and this is used as the low occlusion identification result. When the impact index is not less than the first impact index threshold and less than the second impact index threshold, the degree to which the actual operation feedback data is affected by the coordinated control action is determined as medium occlusion, and this is used as the occlusion identification result. When the impact index is not less than the second impact index threshold, the degree to which the actual operation feedback data is affected by the coordinated control action is determined as high occlusion, and this is used as the high occlusion identification result.

[0033] It should be noted that when the impact index is less than the first impact index threshold, it indicates that the echo of the action landing point and the risk trajectory reversal are not obvious, and the actual operation feedback data is less affected by the coordinated control action, so it is determined to be low masking; when the impact index is between the first impact index threshold and the second impact index threshold, it indicates that the feedback result has been affected by the control to a certain extent, but has not yet reached the level of significantly rewriting the feedback trajectory, so it is determined to be medium masking; when the impact index is not less than the second impact index threshold, it indicates that the actual feedback data shows obvious action echo or risk trajectory reversal after the control action, and the feedback result has been largely rewritten by the coordinated control action, so it is determined to be high masking.

[0034] In one embodiment, S6, the steps of modifying the coordinated control strategy for the next control cycle based on the control masking identification result, and executing resource coordinated control for the next control cycle according to the modified coordinated control strategy are as follows: Read the occlusion level, target risk area and risk type from the control occlusion identification results, and obtain the initial collaborative control strategy corresponding to the target risk area and risk type in the next control cycle; When the occlusion level is low, the initial coordinated control strategy is used as the revised coordinated control strategy for the next control cycle, and resource coordinated control is performed according to the revised coordinated control strategy. When the shading level is medium shading, the intensity of the coordinated control action in the initial coordinated control strategy is reduced by one level to obtain the revised coordinated control strategy for the next control cycle, and resource coordinated control is executed according to the revised coordinated control strategy. Among them, reducing the intensity of the coordinated control action by one level includes at least one of the following: reducing the number of business tasks to be migrated, reducing the power limit range, narrowing the task allocation limit range, reducing the fan speed increase range, reducing the coolant flow rate increase range, or reducing the chilled water valve opening increase range. When the occlusion level is high, the coordinated control actions in the initial coordinated control strategy are changed from direct execution to segmented execution to obtain the revised coordinated control strategy for the next control cycle, and resource coordinated control is executed according to the revised coordinated control strategy. The segmented execution method includes first executing the coordinated control action intensity of the preset proportion in the initial coordinated control strategy, collecting stage feedback data in the next control cycle, and then determining whether to continue executing the remaining coordinated control action intensity based on the stage feedback data. After the resource coordination and regulation is completed in the next control cycle, record the strategy type, target, action intensity, and execution time of the revised coordination and regulation strategy, and use the recorded results as the record results for the next round of regulation actions.

[0035] It should be noted that the purpose of step S6 is to transform the control shielding identification results obtained in the previous control cycle into the actual control correction rules for the next control cycle, preventing the system from continuing to mechanically execute premature cooling, task migration, or power supply restrictions according to the original intensity. Specifically, the system first reads the shielding level, target risk area, and risk type from the control shielding identification results. For example, if area A was identified as having a temperature rise risk in the previous control cycle, and the control shielding identification result was high shielding, it indicates that the subsequent temperature stability in area A was largely caused by enhanced cooling and task migration actions, and cannot directly prove that the original temperature rise risk actually existed. After entering the next control cycle, if the system generates an initial collaborative control strategy again for the temperature rise risk in area A, such as originally planning to directly migrate 30% of the low-priority tasks in area A to area B, and increase the speed of the precision air conditioner fan in area A by 20% and the CDU coolant flow... If the quantity is increased by 15%, the system will not execute the strategy all at once. Instead, it will adjust to a segmented execution method. For example, it will first migrate 10% of the low-priority tasks, while only increasing the fan speed by 8% and the CDU coolant flow rate by 5%. Then, in the next control cycle, it will continue to collect stage feedback data such as temperature, power, load, and CDU return temperature in area A. If the stage feedback data still shows that the temperature in area A continues to develop in a risky direction, the remaining task migration and cooling enhancement actions will continue to be executed. If the stage feedback data shows that the temperature in area A has stabilized or naturally decreased, the remaining intensity will be stopped to avoid over-regulation. If the shading level of the previous control cycle is medium shading, it means that the feedback results have been affected by some regulation, but not to the extent of significant rewriting. In this case, the system will not completely cancel the regulation, but will reduce the intensity of the original regulation action by one level. For example, it may reduce the number of tasks to be migrated, reduce the power limit ratio, narrow the power supply branch limit range, or reduce the fan speed increase before execution. If the occlusion level is low, it indicates that the actual feedback is less affected by the control action, and the feedback data still reflects the original risk state relatively well. Therefore, the initial collaborative control strategy can continue to be used in the next control cycle. In this way, the system can take corrective measures of maintaining, downgrading, and segmenting execution according to the three results of low occlusion, medium occlusion, and high occlusion, respectively. This means that the collaborative control in the next control cycle will no longer simply follow the judgment logic of "stability equals effectiveness" from the previous round, but will dynamically adjust the control intensity according to the degree to which the feedback is rewritten by the control action, reducing invalid task migration, excessive cooling, and excessive power supply reservation. After each execution, the strategy type, target, action intensity, and execution time will continue to be recorded as the basic data for occlusion identification in the next round of control.

[0036] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should still fall within the scope of the claims of the present invention.

Claims

1. A method for coordinated regulation and control of power data center resources based on operational status prediction, characterized in that, Includes the following steps: The system acquires the operational status data of the power data center within the current control cycle, and predicts the operational risks within the future prediction window based on the operational status data, thereby obtaining the predicted risk results. Based on the predicted risk results, the original risk image results are generated under the condition that the current round of coordinated control actions are not implemented. The original risk image results are used to characterize the expected evolution state of the operational risk when it is not intervened by the current round of coordinated control actions. Based on the predicted risk results, generate the current round of coordinated regulation strategy, and execute the corresponding coordinated regulation actions according to the current round of coordinated regulation strategy. At the same time, record the target, type and intensity of the coordinated regulation actions to obtain the regulation action record results. After the coordinated control action is executed, the actual operation feedback data of the power data center within the future prediction window is collected, and the actual operation feedback data is compared with the original risk image result to obtain the feedback offset result. Based on the correspondence between the feedback offset results and the control action recording results, the degree to which the actual operation feedback data is affected by the coordinated control action is identified, and the control masking identification results are obtained. Based on the results of the control masking identification, the coordinated control strategy for the next control cycle is revised, and the resource coordinated control for the next control cycle is executed according to the revised coordinated control strategy.

2. The method for coordinated regulation and control of power data center resources based on operational status prediction according to claim 1, characterized in that, Based on the predicted risk results, the steps to generate the original risk snapshot under the condition that the current round of coordinated control actions has not been implemented are as follows: Based on the predicted risk results, determine the target risk area, risk type, and future prediction window; Extract the server load sequence, rack power sequence, power supply branch load sequence, temperature change sequence, and cooling output sequence corresponding to the target risk area within the current control cycle from the operational status data. Based on the last sampling moment before the coordinated control action is executed, the server load sequence, cabinet power sequence, power supply branch load sequence, temperature change sequence, and cooling output sequence are extended to the future prediction window to obtain the uncontrolled load extension sequence, uncontrolled power extension sequence, uncontrolled power supply load extension sequence, uncontrolled temperature extension sequence, and uncontrolled cooling output extension sequence. Based on the risk type, select the uncontrolled risk extension sequence corresponding to the risk type from the uncontrolled load extension sequence, uncontrolled power extension sequence, uncontrolled power supply load extension sequence, uncontrolled temperature extension sequence, and uncontrolled cooling output extension sequence; By binding the uncontrolled risk extension sequence with the target risk area, risk type, and future prediction window, the original risk image result is generated.

3. The method for coordinated regulation and control of power data center resources based on operational status prediction according to claim 1, characterized in that, The steps for generating a coordinated regulation strategy based on the predicted risk results, executing corresponding coordinated regulation actions according to the strategy, and recording the target, type, and intensity of the coordinated regulation actions to obtain the regulation action recording results include: Based on the target risk area and risk type in the predicted risk results, determine the resource objects to be regulated corresponding to the target risk area; When the risk type is load concentration risk, generate a task migration strategy that includes the business tasks to be migrated and the hosting servers; When the risk type is power increase risk, a power limiting strategy is generated that includes the power limiting object and the power limiting range; When the risk type is that the power supply branch is close to the upper limit risk, a power supply protection strategy is generated that includes the target power supply branch and the task allocation limit range. When the risk type is temperature rise risk, generate a cooling enhancement strategy that includes the target cooling equipment and the increase in cooling output; When the risk type is insufficient cooling capacity risk, a joint heat reduction strategy is generated that includes task migration objects, power limitation objects, and cooling output enhancement objects. The task migration strategy, power limiting strategy, power supply protection strategy, cooling enhancement strategy, or combined heat reduction strategy are used as the collaborative control strategy for this round, and control commands are sent to the corresponding business scheduling system, server management system, power supply control system, or cooling control system to execute the collaborative control action. When performing coordinated control actions, the target, type, intensity, start time, and end time of each coordinated control action are recorded to obtain the control action recording results.

4. The method for coordinated regulation and control of power data center resources based on operational status prediction according to claim 1, characterized in that, The steps to compare the actual operational feedback data with the original risk snapshot results to obtain the feedback offset results are as follows: Between the start of the coordinated control action and the end of the future prediction window, the actual feedback sequence corresponding to the target risk area is collected according to the preset sampling interval. Extract the target risk area, risk type, future prediction window, and uncontrolled risk extension sequence from the original risk snapshot results; Based on the future forecast window, the actual feedback sequence and the uncontrolled risk extension sequence are aligned according to the sampling time to obtain the aligned feedback sequence and the aligned image sequence; Based on the risk type, select the actual risk feedback sequence from the aligned feedback sequence that corresponds to the uncontrolled risk extension sequence; By comparing the numerical differences between the actual risk feedback sequence and the aligned image sequence at the same sampling time, multiple single-point feedback offset values ​​are obtained. Multiple single-point feedback offset values ​​are arranged in the order of sampling time to obtain a feedback offset sequence. The feedback offset sequence, target risk area, risk type and future prediction window are then bound together to generate the feedback offset result.

5. The method for coordinated regulation and control of power data center resources based on operational status prediction according to claim 1, characterized in that, Based on the correspondence between the feedback offset results and the control action recording results, the steps to identify the degree to which the actual operation feedback data is affected by the coordinated control actions and obtain the control masking identification results are as follows: Based on the correspondence between the feedback offset results and the control action recording results, the action landing point echo index and the risk trajectory reversal index are calculated. The action landing point echo index and the risk trajectory reversal index are added together to obtain the influence index. Based on the influence index and the preset influence index threshold, the degree of influence of the actual operation feedback data on the coordinated control action is identified, and the control masking identification result is obtained.

6. A method for coordinated control of power data center resources based on operational status prediction according to claim 5, characterized in that, The steps for calculating the echo index of the impact point include: Read the feedback offset sequence from the feedback offset result, and divide the absolute value of each feedback offset value by the absolute value of the maximum feedback offset value to obtain the feedback offset normalization sequence. When the absolute value of the maximum feedback offset value is zero, all normalized feedback offset values ​​are set to zero. Read the start time of each coordinated control action from the control action record results, and determine the first sampling time that is no earlier than the corresponding start time as the action landing point; For any coordinated control action, construct a mirror sampling group with equal distances before and after the action landing point as the center. When the normalized feedback offset value behind the action landing point is greater than the normalized feedback offset value in front, divide the difference between the two by the larger of the two values ​​to obtain the dominant value behind the mirror. Otherwise, set the dominant value behind the mirror to zero. Complementary multiplication and compression are performed on the dominance values ​​of each mirror image corresponding to the coordinated control action to obtain the echo dominance value of the action. Based on the echo dominance value behind the action, the number of consecutive mirror sampling groups that satisfy the echo dominance value behind the action not less than the echo dominance value behind the action is counted to obtain the echo length close to the landing point, and then divided by the number of mirror sampling groups to obtain the landing point proximity. The echo value of a single action is obtained by multiplying the echo dominance value of the action by the proximity of the landing point and taking the square root. The echo values ​​of the single action landing points corresponding to each coordinated control action are then multiplied and compressed to obtain the action landing point echo index.

7. The method for coordinated regulation and control of power data center resources based on operational status prediction according to claim 5, characterized in that, The calculation steps for the risk trajectory reversion index include: Read the original risk extension sequence and the actual risk feedback sequence from the feedback offset results, and determine the risk advancement direction according to the risk type; According to the direction of risk advancement, the changes of each sampling time in the two sequences relative to the first sampling time are oriented and normalized to obtain the original same-direction risk trajectory and the actual same-direction feedback trajectory. For adjacent sampling times, calculate the risk advance amount of the original same-direction risk trajectory, the feedback retreat amount of the actual same-direction feedback trajectory, and the trajectory separation amount of the two respectively. Then, compress the three by dividing the original value by one and adding the original value, multiply them by the cube root, and obtain the single-segment return deduction value. Adjacent time periods with a single-segment return deduction value greater than zero and continuous time are merged into a continuous return segment. The segment return value is calculated based on the square root of the product of all single-segment return deduction values ​​within the continuous return segment and the continuous proportion of the continuous return segment. The maximum value among all segment return values ​​is used as the risk trajectory return index. When continuous return segments cannot be formed, the risk trajectory return index is set to zero.

8. A method for coordinated control of power data center resources based on operational status prediction according to claim 5, characterized in that, The steps for identifying the degree to which actual operational feedback data is affected by coordinated control actions based on the impact index and a preset impact index threshold, and obtaining the control masking identification result, include: The preset impact index thresholds include a first impact index threshold and a second impact index threshold, and the first impact index threshold is less than the second impact index threshold; When the impact index is less than the first impact index threshold, the degree to which the actual operation feedback data is affected by the coordinated control action is determined as low occlusion, and this is used as the low occlusion identification result. When the impact index is not less than the first impact index threshold and less than the second impact index threshold, the degree to which the actual operation feedback data is affected by the coordinated control action is determined as medium occlusion, and this is used as the occlusion identification result. When the impact index is not less than the second impact index threshold, the degree to which the actual operation feedback data is affected by the coordinated control action is determined as high occlusion, and this is used as the high occlusion identification result.

9. A method for coordinated regulation and control of power data center resources based on operational status prediction according to claim 1, characterized in that, The steps for revising the coordinated control strategy for the next control cycle based on the control masking identification results, and then executing the resource coordinated control for the next control cycle according to the revised coordinated control strategy are as follows: Read the occlusion level, target risk area and risk type from the control occlusion identification results, and obtain the initial collaborative control strategy corresponding to the target risk area and risk type in the next control cycle; When the occlusion level is low, the initial coordinated control strategy is used as the revised coordinated control strategy for the next control cycle, and resource coordinated control is performed according to the revised coordinated control strategy. When the shading level is medium shading, the intensity of the coordinated control action in the initial coordinated control strategy is reduced by one level to obtain the revised coordinated control strategy for the next control cycle, and resource coordinated control is executed according to the revised coordinated control strategy. Among them, reducing the intensity of the coordinated control action by one level includes at least one of the following: reducing the number of business tasks to be migrated, reducing the power limit range, narrowing the task allocation limit range, reducing the fan speed increase range, reducing the coolant flow rate increase range, or reducing the chilled water valve opening increase range. When the occlusion level is high, the coordinated control actions in the initial coordinated control strategy are changed from direct execution to segmented execution to obtain the revised coordinated control strategy for the next control cycle, and resource coordinated control is executed according to the revised coordinated control strategy. The segmented execution method includes first executing the coordinated control action intensity of the preset proportion in the initial coordinated control strategy, collecting stage feedback data in the next control cycle, and then determining whether to continue executing the remaining coordinated control action intensity based on the stage feedback data. After the resource coordination and regulation is completed in the next control cycle, record the strategy type, target, intensity of action, and execution time of the revised coordination and regulation strategy, and use the recorded results as the record results for the next round of regulation actions.

Images