A data center green energy driven computing power optimization method

Abstract

The application provides a data center green energy driven computing power optimization method, comprising: obtaining real-time tidal energy supply data and computing load demand data, analyzing the matching degree of tidal energy periodic supply and computing load fluctuation demand, and determining the peak misalignment period of tidal energy supply and computing load demand; according to the peak misalignment period of tidal energy supply and computing load demand, evaluating the execution window of high energy consumption computing tasks of the data center, and obtaining the available resource list of the energy sufficient period; if the balanced degree of the migrated computing power resource allocation does not reach the preset balanced threshold, obtaining the current load of the server cluster of the data center, determining the energy low period activation condition; analyzing the relief degree of resource mismatch from the available capacity index of the buffer energy, reallocating the idle computing power resources according to the relief degree, and determining the supplementary demand of the tidal energy insufficient period.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, in particular to a data center green energy driven computing power optimization method. BACKGROUND

[0002] Ocean tidal energy, as a clean and renewable energy, has attracted much attention in the energy field due to its stability and predictability, especially in promoting the construction of green data centers. Tidal energy is converted into electrical energy through tidal movement, providing sustainable power for data centers, directly affecting the stable supply of computing power resources. However, the mismatch between the periodic supply of tidal energy and the dynamic demand for computing power in data centers may lead to insufficient or excessive allocation of computing power, affecting operational efficiency. This mismatch not only reduces energy utilization efficiency, but also increases the operating cost and stability risk of data centers due to waste or shortage of computing power resources. Existing methods mainly rely on static energy scheduling strategies or simple backup power switching to address the mismatch between tidal energy and data center demand. These methods usually cannot dynamically adapt to the periodic fluctuations of tidal energy supply, nor can they flexibly adjust according to the real-time demand of computing tasks. For example, some solutions allocate energy in fixed time periods, ignoring the changes in computing power demand in data centers at different times, resulting in low energy utilization efficiency. In addition, the use of backup power often relies on traditional energy, which weakens the green and environmentally friendly advantages of tidal energy, and does not fundamentally solve the dynamic matching problem between energy and computing power. The power generation capacity of tidal energy is affected by the tidal cycle, showing a fixed peak and trough every day, while the computing task demand of data centers fluctuates due to user access, task type, and other factors. For example, in the video streaming processing scenario, due to high user activity, traffic surge, frequent content updates, and other reasons, the evening user access peak may overlap with the tidal energy generation trough period. This is because in some geographical locations, tidal power generation relies on the potential energy difference formed by the rise and fall of seawater, and the tidal cycle is not completely synchronized with the day-night cycle, so the tidal level at night may be at low tide or flat tide stage, resulting in a significant decrease in tidal energy generation during the evening period, which cannot meet the peak demand, leading to energy shortage to support high-load tasks; while in the tidal energy peak period, the computing task may be less, causing energy waste. This time misalignment leads to a resource matching problem, which is due to the lack of a mechanism to dynamically adjust the time distribution of computing tasks and the running state of servers. Therefore, how to dynamically adjust the allocation of computing tasks and the running scale of servers in data centers according to the periodic supply characteristics of tidal energy to realize real-time matching between energy supply and computing power demand has become a key problem in the utilization of tidal energy in data centers. SUMMARY

[0003] The present application provides a data center green energy driven computing power optimization method, mainly comprising:

[0004] The matching degree of the tidal energy supply data and the computing load demand data is analyzed to determine a peak misalignment period of the tidal energy supply and the computing load demand; an execution window of a high-energy-consumption computing task is evaluated according to the peak misalignment period, and an available resource list of an energy sufficient period is generated; a priority sequence of the high-energy-consumption task is extracted from the available resource list, and the task matching the tidal energy supply period is dynamically migrated to the corresponding period for execution, and the balance degree of the computing power resource allocation after migration is judged; the balance degree of the computing power resource allocation and the current load of the server cluster are used to adjust the server cluster size to a target active node number, and a trough period activation condition is determined; a buffer energy available capacity index is generated according to the trough period activation condition; the idle computing power resources are redistributed according to the buffer energy available capacity index, and a supplement demand in a tidal energy insufficient period is determined; the task execution efficiency and energy consumption data in the task execution log after dynamic migration are analyzed, and a computing power optimization scheme is generated in combination with the supplement demand in the tidal energy insufficient period.

[0005] Further, the matching degree of the tidal energy supply data and the computing load demand data is analyzed to determine a peak misalignment period of the tidal energy supply and the computing load demand, comprising:

[0006] The tidal height, flow rate and power generation data of the tidal power station are collected, and the CPU usage rate, memory occupancy rate and network bandwidth consumption data of the data center server are obtained; the tidal power station data is subjected to time series decomposition to extract a periodic characteristic curve; the computing load demand data is processed by using a sliding window method according to the periodic characteristic curve, the load mean and variance are calculated, and the load peak time is identified; a supply-demand time sequence deviation matrix is generated by comparing the tidal energy supply peak time with the load peak time; a period with a time offset exceeding a threshold value is extracted from the supply-demand time sequence deviation matrix and marked as a misalignment period to generate a misalignment period record table containing the misalignment start time, duration and energy difference.

[0007] Further, the execution window of the high-energy-consumption computing task is evaluated according to the peak misalignment period, and an available resource list of an energy sufficient period is generated, comprising:

[0008] The power demand and runtime of video transcoding, batch data processing and machine learning training tasks are identified according to the peak misalignment period; the tidal energy supply data and the power demand are compared to determine a time interval in which the energy supply exceeds the task demand as an execution window; the net power value after deducting the infrastructure power consumption from the tidal power generation power in the execution window is calculated to determine the maximum number of nodes supported and the total storage capacity; the network device configuration table is queried according to the maximum number of nodes, the node network card rate is accumulated as the bandwidth demand value, the bandwidth quota is adjusted, and an available resource list containing the number of computing nodes, storage capacity and bandwidth allocation is generated.

[0009] Further, the priority sequence of high-energy-consumption tasks is extracted from the available resource list, and tasks matching the tidal energy supply period are dynamically migrated to the corresponding period for execution, and the balance degree of computing resource allocation after migration is judged, including:

[0010] The number of computing nodes, storage capacity and bandwidth allocation data are read from the available resource list, and the priority sequence of high-energy-consumption tasks is generated by combining the difference between the task deadline and the current time; the execution time of the tasks in the priority sequence is compared with the tidal energy supply period, and the matching tasks are marked to construct a task migration candidate set; the tasks in the task migration candidate set are migrated to the computing nodes in the corresponding period, and the CPU utilization, memory usage and bandwidth occupancy after migration are recorded; the balance degree of computing resource allocation is evaluated by the variance value of the CPU utilization.

[0011] Further, after generating the available resource list of the energy sufficient period, including:

[0012] The execution requirement parameters of the tasks in the task migration candidate set are extracted, including minimum runtime, maximum allowed delay, resource occupancy and interruptability; the overlap ratio of the time window of the tasks in the task migration candidate set and the tidal energy supply period is calculated to determine the fitting degree; according to the fitting degree, the task scheduling plan is adjusted, and the task execution time is migrated to the start time of the tidal energy supply period; according to the minimum runtime, the task end time is set to generate the task execution time table after migration adjustment.

[0013] Further, according to the balance degree of computing resource allocation and the current load of the server cluster, the size of the server cluster is adjusted to the target active node number, and the low valley period activation condition is determined, including:

[0014] According to the balance degree of computing resource allocation, the CPU usage, memory occupancy and task queue length of the server cluster are obtained, and the current load value is calculated; if the current load value exceeds the threshold value, the node load value is sorted, and the sleep operation is executed to reach the target active node number; according to the target active node number, the minimum load threshold, node running efficiency and energy supply stability are set to generate the low valley period activation condition.

[0015] Further, the available capacity index of the buffer energy generated according to the low valley period activation condition, including:

[0016] According to the low tide activation condition, the difference between the tidal power generation power and the data center power consumption is detected, and the excess energy is allocated to the electrochemical or mechanical energy storage unit;Perform the charging or mechanical energy storage process, record the energy storage growth rate and conversion efficiency;When the tidal energy supply data is lower than the power consumption, control the energy storage unit to release energy, count the remaining capacity, conversion efficiency and release rate, and generate a buffer energy available capacity index.

[0017] Further, the generation of the buffer energy available capacity index comprises:

[0018] Compare the tidal energy supply data with the data center power consumption, mark the energy gap period and the power difference;According to the power difference, control the energy storage unit to output electric energy, adjust the release rate to match the gap;Collect the remaining electric quantity of the energy storage unit, the ratio of the output electric quantity to the consumed energy and the instantaneous maximum output power;Statistical average conversion efficiency and release rate peak value in the release period, generate buffer energy available capacity index.

[0019] Further, the redistribution of idle computing resources according to the buffer energy available capacity index to determine the tidal energy shortage period supplement demand comprises:

[0020] Extract the energy storage capacity and conversion efficiency from the buffer energy available capacity index, calculate the actual available energy;Compare the actual available energy with the demand energy, determine the resource mismatch relief degree value;According to the resource mismatch relief degree value, allocate idle computing resources to low priority tasks;Statistical additional start node number, task urgency and average waiting time, generate a supplement demand including computing power supplement, energy supply priority and task scheduling delay.

[0021] Further, the analysis of the task execution efficiency and energy consumption data in the task execution log after dynamic migration, combined with the tidal energy shortage period supplement demand, generates a computing power optimization scheme, comprising:

[0022] Extract the task identification, execution time and resource utilization rate from the task execution log to generate a task execution efficiency index set;According to the task execution efficiency index set, calculate the unit task energy consumption;According to the tidal energy shortage period supplement demand, arrange high energy consumption tasks to energy sufficient period, compile power quota and energy storage release plan;According to the task load, determine the number of activated or dormant nodes, and generate a computing power optimization scheme.

[0023] The technical scheme provided by the embodiment of the application can include the following beneficial effects:

[0024] The application discloses a data center green energy driven computing power optimization method, and aims at a resource mismatch problem caused by a poor matching degree between a tidal energy periodic supply and a computing load fluctuation demand, through analyzing a peak mismatch period of the tidal energy supply and the load demand, dynamically adjusting an execution window of a high energy consumption task, migrating the task to an energy sufficient period, and combining an available resource list to optimize a computing power distribution balance degree. BRIEF DESCRIPTION OF DRAWINGS

[0025] Fig. 1 A flowchart of the data center green energy driven computing power optimization method of the application.

[0026] Fig. 2 A schematic diagram of step S101 in the data center green energy driven computing power optimization method of the application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical scheme and advantages of the application clearer, the application is described in detail below with reference to the drawings and specific embodiments.

[0028] As Figs. 1-2 , the data center green energy driven computing power optimization method of the embodiment specifically can include:

[0029] Step S101, real-time tidal energy supply data and computing load demand data are acquired, a matching degree between a tidal energy periodic supply and a computing load fluctuation demand is analyzed, and a peak mismatch period of the tidal energy supply and the load demand is determined.

[0030] The tidal height, flow rate and power generation data are collected by the sensor array deployed in the tidal power station, and the server CPU usage, memory occupancy and network bandwidth consumption records are obtained from the data center management platform. The tidal energy data collected is decomposed by time series, the main period components of 24 hours and 12.42 hours are extracted, and the periodic characteristic curve of tidal energy supply is obtained. According to the periodic characteristic curve, the sliding window method is used to segment the load calculation data, the mean and variance of the load in each period are calculated, the peak time and duration of the load fluctuation are identified, the time difference between the peak time of the tidal energy supply and the peak time of the load calculation is calculated, the time offset of the two is judged, and the supply-demand time sequence deviation matrix is obtained. The time offset greater than the preset deviation threshold is extracted from the supply-demand time sequence deviation matrix, wherein the preset deviation threshold is one fourth of the tidal period length, the period exceeding the threshold is marked as a misalignment period, the energy gap value and the energy surplus value in each misalignment period are calculated, and the misalignment severity is determined according to the ratio of the gap value or surplus value to the average energy supply in the period. The misalignment period record table containing the misalignment start time, duration and energy difference is obtained. If there is a misalignment period with a duration greater than the preset duration threshold in the misalignment period record table, the historical load data in the period is processed by K-means clustering, the periodically recurring high-load task type and its occurrence rule are identified according to the clustering result, and the high peak misalignment period of tidal energy supply and calculation load demand is determined combined with the energy difference data in the misalignment period record table.

[0031] Specifically, in an embodiment, the sensor array of the tidal power station adopts a hierarchical deployment architecture, and pressure sensors, flow rate sensors and power monitoring sensors are arranged at the seawater inlet, turbine set and power output end respectively. The pressure sensor obtains tidal height data by measuring the change of seawater pressure, samples once a minute, and records the real-time change of tidal level. The flow rate sensor is installed at the front end of the turbine blade and monitors the speed of seawater flowing through the turbine. The power monitoring sensor is directly connected to the output end of the generator and collects real-time power generation data. The time series decomposition adopts the Fourier transform method to analyze the frequency domain of the tidal energy data for 30 consecutive days, and extracts the dominant frequency component. The 24-hour period corresponds to the diurnal tidal change caused by the earth's rotation, and the 12.42-hour period corresponds to the semidiurnal tidal change caused by the moon's gravity. These two period components constitute the main characteristics of tidal energy supply.

[0032] Specifically, the data center management platform collects resource usage of each server node in real time through the deployed monitoring agent. CPU usage is obtained by reading the processor performance counter, memory occupancy is obtained by querying the memory management interface of the operating system, and network bandwidth consumption is recorded through the network card traffic statistics module. These data are sampled at 5-minute intervals to form a continuous computing load time series.

[0033] For example, in the implementation of the sliding window method, the window width is set to 1 hour and the sliding step is 15 minutes. For the load data in each window, the arithmetic mean is calculated as the average load of the period, and the standard deviation is calculated as a measure of load fluctuation. When the load value of a period exceeds 1.5 times the average of the previous and subsequent periods, it is identified as a load peak time. By recording the start and end time points of the peak value, the peak duration is determined. The calculation of the time offset is the difference between the timestamp of the tidal energy supply peak and the timestamp of the calculated load peak, forming a two-dimensional matrix, where the rows represent different dates and the columns represent different time periods within a day.

[0034] In one possible implementation, the determination of the preset deviation threshold is based on the characteristics of the tidal cycle. Since the semidiurnal tide cycle is 12.42 hours, one quarter of which is about 3.1 hours, which means that when the time offset between energy supply and load demand exceeds 3.1 hours, they are basically in anti-phase state and the matching degree is the worst. The quantification of misalignment severity uses the ratio method, and the formula is misalignment degree equals to energy gap value or surplus value divided by the average energy supply of the period. When the ratio exceeds 0.3, it indicates that the misalignment is serious; the ratio between 0.1 and 0.3 is moderate misalignment; less than 0.1 is mild misalignment.

[0035] It should be noted that the specific application of K-means clustering in load pattern recognition is to construct a feature vector according to the time dimension and load intensity dimension. Each feature vector contains attributes such as load peak, duration, peak occurrence time, etc. The number of clusters K is determined by the elbow rule, usually set to 3 to 5 categories, corresponding to video transcoding tasks, data backup tasks, user interaction tasks, and other types of computing loads. The cluster center represents the typical load pattern of each type of task, and by calculating the Euclidean distance between the new load and each cluster center, the type of task it belongs to is determined.

[0036] Preferably, in identifying high-load task types, special patterns in video streaming processing scenarios are focused on. During the peak user viewing period from 20:00 to 23:00, video transcoding and content distribution tasks surge, server CPU utilization can reach more than 85%, and memory occupancy exceeds 70%. At this time, the tidal energy is at a low tide period, and the power generation is only about 30% of the peak. This typical misalignment scenario needs to be solved through task scheduling and energy storage system coordination.

[0037] In one embodiment, the data structure of the misalignment period record table contains multiple fields. The misalignment start time is accurate to the minute level, recording the specific time when the misalignment starts. The duration is in hours, representing the time span of the misalignment state maintenance. The energy difference is divided into two categories: gap value and surplus value. The gap value represents the amount of power that the load demand exceeds the energy supply, and the surplus value represents the amount of power that the energy supply exceeds the load demand, both in units of kilowatt-hours. Through this structured record, subsequent task scheduling optimization and energy storage system deployment are facilitated.

[0038] For example, when detecting that the continuous misalignment period exceeds the preset duration threshold of 4 hours, the deep analysis mechanism is triggered. At this time, not only the current load pattern is analyzed, but also the load data of the same period in the past 7 days is traced back to find periodicity. If it is found that a certain type of high-load task repeatedly occurs at a fixed period every day, and there is a stable phase difference with the tidal energy supply cycle, the period is marked as a key misalignment period, which needs to be focused on for energy scheduling optimization.

[0039] It can be understood that the final determination of the peak misalignment period considers multiple dimensions of energy supply and demand. Not only includes the matching degree of energy quantity, but also includes the stability of energy quality, the urgency of task execution, the economy of system operation and other factors. By establishing a multi-dimensional evaluation system, it is ensured that the identified misalignment period really affects the running efficiency of the data center, and provides accurate decision basis for subsequent optimization scheduling.

[0040] Step S102, according to the peak misalignment period of tidal energy supply and computing load demand, the execution window of high-energy consumption computing task of the data center is evaluated, and the available resource list of the energy sufficient period is obtained.

[0041] According to the start time and duration of the peak dislocation period, the execution characteristics of video transcoding, batch data processing, and machine learning training tasks in the data center are identified, the power demand value, running time, and resource occupation rate of each type of high-energy-consumption task are extracted, and the power generation value of the tidal energy supply curve in the same period is compared to determine the time interval in which the energy supply exceeds the power demand of the task as the task execution window. By subtracting the power consumption of the data center infrastructure from the hourly tidal power generation in the task execution window, the net power value available for computing tasks is obtained. According to the power consumption record of a single computing node under full load, the maximum number of nodes supported by the net power value is calculated, and the memory size and hard disk capacity of each node are simultaneously counted to obtain the total storage capacity of the period. According to the maximum number of nodes and the total storage capacity, the network card rate of each computing node is queried from the data center network device configuration table, and the total sum of the network card rates of all nodes is accumulated as the bandwidth demand value. If the bandwidth demand value exceeds the upper limit of the core switch port bandwidth, the bandwidth quota of each node is redistributed according to the task priority, and an energy sufficient period available resource list containing the number of computing nodes, storage capacity, and bandwidth allocation is generated.

[0042] Specifically, in an embodiment, the identification process of high-energy-consumption computing tasks is based on historical power consumption records. When performing H.265 encoding, the power demand of a single video transcoding task reaches 200 to 300 watts, and the running time is usually 2 to 4 hours. Batch data processing tasks involve the parsing and statistics of massive log files, with a power demand of 150 to 200 watts and a running time of 1 to 3 hours fluctuating according to the data volume. When performing deep neural network training, the power demand of a GPU cluster for machine learning training tasks can reach 500 to 800 watts, and the training period ranges from several hours to several days.

[0043] Specifically, the calculation of the net power value requires accurate evaluation of infrastructure power consumption. The infrastructure power consumption of a data center includes the power consumption of the refrigeration system, the loss of the uninterruptible power supply, the lighting power consumption, and the basic power consumption of network devices. The refrigeration system adjusts dynamically according to the server heat dissipation, usually accounting for 30% to 40% of the total power consumption. The uninterruptible power supply generates about 10% energy loss during charging and discharging conversion. After deducting these basic power consumptions from the total tidal power generation, the remaining part is the net power value available for computing tasks.

[0044] It should be noted that the power consumption record of a single computing node is obtained by real-time monitoring of an intelligent power distribution unit. The power consumption of each node differs significantly under idle, half-load, and full-load conditions, with about 100 watts under idle, 180 watts under half-load, and 250 watts under full-load. The calculation of the maximum number of nodes uses the full-load power consumption value as a reference to ensure sufficient energy supply under extreme load conditions.

[0045] Preferably, the bandwidth allocation adopts a hierarchical management mechanism. The core switch has multiple Gigabit ports, with a total bandwidth capacity of 160 Gbps. When the cumulative bandwidth demand exceeds this upper limit, reallocation is made according to task priority. Real-time video stream processing tasks get the highest priority, with 40% of the bandwidth resources allocated; batch data transmission tasks come second, with 35% of the bandwidth allocated; the remaining tasks share the remaining 25% of the bandwidth. Through this dynamic allocation mechanism, the network transmission of critical tasks is ensured to be unaffected.

[0046] In an embodiment, the energy sufficient period available resource list is presented in a structured table form, with each period corresponding to a row record containing four fields of period start and end time, available computing node number, total storage capacity and allocated bandwidth.

[0047] Step S103, extract the priority sequence of high energy consumption tasks from the available resource list, if there is a task in the priority sequence that matches the tidal energy supply period, dynamically migrate the task that matches the tidal energy supply period to the corresponding period for execution, and judge the balance degree of the computing resource allocation after migration according to the available resource list.

[0048] Read the computing node number, storage capacity and bandwidth allocation data of each period from the available resource list, and respectively obtain the difference between the deadline and the current time of the video transcoding, data backup and machine learning training tasks as the urgency index, and arrange them in descending order according to the product of the urgency index and the task power consumption value, to obtain the priority sequence of high energy consumption tasks. According to the expected execution time of each task in the priority sequence and the duration of the tidal energy supply period, if the task execution time is less than the preset proportion threshold of the supply period duration, mark the task as a matchable task, extract the original execution time and resource demand of all matchable tasks, and construct a task migration candidate set. For each task in the task migration candidate set, read its current running state and memory image data, migrate the task execution environment to the computing node corresponding to the tidal energy sufficient period through the container orchestration tool, record the CPU utilization, memory usage and network bandwidth occupation rate values of each period after the migration operation is completed, and obtain the resource utilization record of the period after migration. Use the variance value of the CPU utilization of all periods in the period resource utilization record to evaluate the difference degree of the allocation of computing resources in the time dimension, if the variance value exceeds the preset balance threshold, calculate the difference percentage between the period with the highest resource utilization rate and the period with the lowest resource utilization rate, and judge the balance degree of the computing resource allocation after migration according to the difference percentage.

[0049] Specifically, in an embodiment, the construction of the high-energy-consumption task priority sequence adopts a two-dimensional evaluation mechanism. The urgency index is obtained by calculating the difference between the task deadline and the current system time, and the smaller the difference, the more urgent the task. Optionally, the assignment of the urgency index can be specifically set in a segmented scoring manner according to the task type and the remaining time.

[0050] For example, for a live task in video transcoding, if the remaining time is within 2 hours, it is considered extremely urgent, and the urgency index is assigned a value of 100; if the remaining time is between 2 and 12 hours, it is considered relatively urgent, and the value is assigned a value of 60; if the remaining time exceeds 12 hours, it is considered not urgent, and the value is assigned a value of 20. For on-demand video transcoding tasks, if the remaining time is within 6 hours, it is considered urgent, and the urgency index is assigned a value of 40; if the remaining time is between 6 and 24 hours, it is considered general, and the value is assigned a value of 10; if the remaining time exceeds 24 hours, it is considered relatively relaxed, and the value is assigned a value of 5. For data backup tasks, if the remaining time is within 1 hour, it is considered urgent, and the value is assigned a value of 80; if the remaining time is between 1 and 8 hours, it is considered moderately urgent, and the value is assigned a value of 30; if the remaining time exceeds 8 hours, it is considered low priority, and the value is assigned a value of 10. For machine learning training tasks, because they usually have high power consumption and long computation periods, if the remaining time is within 1 hour, they must be completed as soon as possible, and the urgency index is assigned a value of 90; if the remaining time is between 1 and 4 hours, attention should be paid to the execution timing, and the value is assigned a value of 50; if the remaining time exceeds 4 hours, the processing can be appropriately delayed, and the value is assigned a value of 20. For video transcoding tasks, if a transcoding task for a live platform has only 2 hours left before broadcast, the urgency index is assigned a value of 100; if it is an offline transcoding task for an on-demand video, the deadline is 24 hours later, and the urgency index is assigned a value of 10. The task power consumption value is extracted from historical execution records, the average power consumption of video transcoding tasks is 250 watts, the average power consumption of data backup tasks is 180 watts, and the average power consumption of machine learning training tasks is 650 watts. After multiplying the urgency index and the power consumption value, they are arranged in descending order to ensure that tasks that are both urgent and high in energy consumption have higher priority.

[0051] Specifically, the estimation of task execution duration is based on statistical analysis of historical execution data. The system records the execution duration of each type of task under different data sizes to construct a task duration prediction table. A 4K video transcoding task takes 30 minutes to process 1 hour of material, batch data backup takes 15 minutes per 100 GB, and deep learning model training takes 45 minutes per epoch. The duration of the tidal energy supply period is extracted from the tidal cycle curve, and the duration of the high tide period is usually 3 to 4 hours. The preset proportion threshold is set to 0.8, meaning that a task is only marked as a matchable task if its execution duration does not exceed 80% of the supply period, leaving 20% buffer time to handle execution delays.

[0052] Exemplarily, the task migration process of the container orchestration tool involves multiple technical links. With the Kubernetes orchestration platform, the dynamic migration of tasks is realized through the Pod scheduler. Before migration, the system creates a checkpoint of the task, saving the current execution state, memory data and intermediate results. The checkpoint data is synchronized to the target node through a distributed storage system, ensuring data consistency during the migration process. The migration operation is completed by modifying the node affinity label of the Pod, and the scheduler automatically migrates the Pod from the source node to the target node. The downtime of the entire migration process is controlled within seconds, and for interruptible tasks such as video transcoding, it has little effect on the overall execution efficiency.

[0053] It should be noted that the collection of period resource utilization rate records covers three-dimensional monitoring indicators. CPU utilization is obtained by reading / proc / stat file to obtain the usage time slice of each core, and the ratio of busy time to total time is calculated. The memory usage is read from the / proc / meminfo file to obtain the used memory and total memory, and the percentage of occupation is calculated. The network bandwidth occupancy rate is obtained through the network card driver interface to obtain the real-time transmission rate, and the maximum rate of the network card is compared to obtain the value. These indicators are collected every 5 minutes to form continuous time series data. Each period contains 12 sampling points, and the average value is taken as the resource utilization rate of the period.

[0054] In one possible implementation, the calculation of the variance value is used to quantify the degree of imbalance of resource allocation. The CPU utilization data of all periods is collected, and the average value is calculated as the baseline. The difference between the utilization rate of each period and the average value is squared and summed, and then divided by the total number of periods to obtain the variance value. The larger the variance value, the more uneven the distribution of resources in the time dimension. The preset balance threshold is set according to the operation and maintenance experience of the data center, and is usually 0.15. When the variance value exceeds this threshold, it indicates that there is a significant resource utilization peak and valley, which needs to be further optimized.

[0055] Preferably, the calculation of the difference percentage provides a more intuitive evaluation of the balance degree. The system identifies the period with the highest resource utilization rate, which may reach 90% of the CPU utilization rate; at the same time, it finds the lowest period, which may have a utilization rate of only 20%. The difference between the two is 70%, and this percentage directly reflects the degree of resource allocation. According to the difference percentage, the balance degree is divided into three levels: the difference is less than 30% for balance, 30% to 60% for mild imbalance, and more than 60% for serious imbalance. For example, during the user access peak period from 20:00 to 23:00, the tidal energy is at a low point, and batch data processing and offline analysis tasks scheduled for the evening can be migrated to the tidal energy peak period to balance resource allocation.

[0056] In an embodiment, the construction of the task migration candidate set takes into account various constraints. In addition to the execution duration matching, the interruptibility of the task, data dependency and service level agreement requirements are also evaluated. Tasks with high real-time requirements are not included in the migration candidates to ensure that the quality of service is not affected. For task chains with data dependencies, the feasibility of migration needs to be evaluated as a whole to avoid breaking the dependencies between tasks.

[0057] It can be understood that the judgment of the balance degree of computing power resource allocation not only focuses on the resource utilization at a single moment, but more importantly, evaluates the rationality of resource allocation within the entire running period. Through dynamic task migration, the supply curve of tidal energy and the demand curve of computing load are better matched, reducing energy waste and resource idling, and improving the overall energy efficiency and operation stability of the data center.

[0058] The execution demand and time window of the task matching the tidal energy supply period are identified, the original start time, deadline and interruptibility of the task are matched with the tidal energy supply period to obtain the matching degree of the task and the energy supply period, the scheduling plan of the task is adjusted according to the matching degree of the task and the energy supply period, the task execution time is migrated from the original time period to the peak period of tidal energy supply, the task execution time table after migration adjustment is obtained, and the start and end time of each matched task in the new period is determined.

[0059] The task matching the tidal energy supply period is identified, the execution requirement parameters of each task are extracted, including the minimum running time, the maximum allowed delay, the resource occupation, and the original start time and the deadline of the task are obtained, the interruptibility attribute of the task is determined according to the task type, if it is a video transcoding task, it is marked as interruptible, if it is a real-time data processing task, it is marked as non-interruptible, and the task execution requirement feature set is obtained. The start and end points of the task time window are determined according to the original start time and the minimum running time in the task execution requirement feature set, the overlapping length of the window and the tidal energy peak period is calculated, the time overlap ratio is obtained by dividing the overlapping length by the minimum running time of the task, and whether the deadline of the task is later than the end time of the energy peak is judged, and the fit degree of the task and the energy supply period is obtained by comprehensively considering the two indexes. According to the fit degree, if the value exceeds the preset threshold, the scheduling plan of the task is adjusted, the original execution record is deleted from the task scheduling table, the task execution start time is modified to the start time of the tidal energy peak period, and the expected end time of the task in the new period is set according to the minimum running time in the task execution requirement feature set, and the adjusted task time arrangement is obtained. The specific start time of each matching task in the tidal energy peak period is determined by using the adjusted task time arrangement, if the expected end time exceeds the peak period range, the number of parallel processing nodes is increased to shorten the task execution time, the end time point is recalculated, the new start time and the end time of each task are recorded, and the task execution time table after migration adjustment is obtained.

[0060] Specifically, in an embodiment, the extraction of task execution requirement parameters is based on historical operation data and task configuration files. The minimum running time is calculated from the number of basic processing units of the task and the processing speed of the unit. Video transcoding task needs fixed coding time for processing a GOP group, and the processing time of all GOP groups is the minimum running time. The maximum allowed delay is determined according to the service level agreement. The maximum allowed delay of live transcoding task is usually 30 minutes, the delay of on-demand video can be 24 hours, and the delay of batch data backup can be 48 hours. The resource occupation includes three dimensions of CPU core number, memory capacity and storage space. 4K video transcoding requires 8 CPU cores, 16 GB memory and 100 GB temporary storage space.

[0061] Specifically, the task interruptibility determination follows strict classification criteria. Video transcoding tasks can be interrupted safely at key frame positions, and the processing is resumed from the nearest key frame after interruption without affecting the final output quality. Batch data processing tasks support sharding, and each shard is completed independently to save intermediate results, and the processing is continued from the unfinished shard after interruption. Machine learning training tasks save model parameters at the end of each training cycle, and can be interrupted and resumed at any cycle boundary. Real-time data stream processing tasks are not allowed to be interrupted, as it will cause data loss and processing delay accumulation. Transaction processing tasks have transactional requirements and must be executed completely or rolled back completely, and are also marked as non-interruptible.

[0062] By way of example, the calculation process of the time overlap ratio involves intersection operation of time intervals. The task time window is determined by the original start time and the minimum runtime, forming a time interval. The tidal energy peak period is extracted from the tidal power generation curve, usually in the middle period of the rising tide and the ebbing tide, lasting for 3 to 4 hours. The intersection of the two time intervals is determined by comparing the start and end time points, and the start point of the intersection is the larger value of the start points of the two intervals, and the end point is the smaller value of the end points of the two intervals. If the end point is greater than the start point, there is an effective intersection, and the intersection length is equal to the difference between the end point and the start point. The time overlap ratio is equal to the intersection length divided by the minimum runtime of the task, and the higher the ratio indicates that the energy supply is more sufficient during the task execution.

[0063] It should be noted that the comprehensive evaluation of the fitting degree not only considers the time overlap ratio, but also includes the satisfaction degree of the deadline constraint. If the task deadline is earlier than the energy peak end time, it means that the task can be completed during the energy sufficient period, and the fitting degree weighting coefficient is set to 1.2. If the deadline is later than the peak end but earlier than the next peak start, the weighting coefficient is 1.0. If the deadline is very urgent, even if there is overlap with the energy peak, it is not suitable for migration, and the weighting coefficient is reduced to 0.5. The fitting degree is equal to the time overlap ratio multiplied by the weighting coefficient, which comprehensively reflects the feasibility and benefit of task migration.

[0064] Preferably, the adjustment of the task scheduling plan adopts a transactional update mechanism to ensure the consistency of the scheduling table. The system first locks the scheduling record of the task to be adjusted to prevent concurrent modification. When the original execution record is deleted from the scheduling table, the computing resources reserved in the original time period are released simultaneously, including CPU quota, memory allocation and network bandwidth. The task execution start time is modified to the start time of the tidal energy peak period, and the corresponding resources are reserved in the new time period. After the scheduling table is updated, a scheduling change notification is sent to the task executor to ensure that the task is executed according to the new time plan.

[0065] In one possible implementation, when the predicted end time exceeds the peak period range, the system compresses the execution time length by increasing the number of parallel processing nodes. The implementation of parallelization depends on the decomposability of the task. Video transcoding tasks can be divided by scene, and each scene is assigned to a different node for parallel processing. Data backup tasks are divided by file directory, and multiple nodes simultaneously backup different directories. Machine learning training adopts a data parallel strategy, which divides the training data set and distributes it to multiple GPU nodes. Each node independently calculates the gradient and then updates the model parameters. The upper limit of parallelism is limited by the parallelization efficiency of the task itself, and is usually set to 2 to 4 times the original number of nodes.

[0066] In one embodiment, the adjusted task execution schedule after migration is recorded in the form of a Gantt chart, which shows the time arrangement of each task. The horizontal axis represents time, and the vertical axis represents different computing nodes. Each task is represented by a rectangular bar, with the starting point corresponding to the start time and the end point corresponding to the end time. The height of the rectangle represents the number of nodes occupied. Through visual display, operation and maintenance personnel can intuitively see the task distribution and timely detect resource conflicts and scheduling abnormalities.

[0067] In step S104, if the balance degree of the computing resource allocation after migration does not reach the preset balance threshold, the current load of the server cluster of the data center is obtained, and the energy valley period activation condition is determined.

[0068] If the balance degree of the computing resource allocation after migration does not reach the preset balance threshold, the CPU usage, memory occupancy, and task queue length of all nodes in the cluster are obtained from the server management platform. The CPU usage is multiplied by a preset weight coefficient, the memory occupancy is multiplied by a corresponding weight coefficient, and the task queue length is normalized. The current load value of the server cluster is calculated. According to the current load value of the server cluster, if it exceeds the preset load threshold, the number of nodes with a load value below the preset low load standard is counted, a node list is generated in order from low to high load value, and the first node in the list is selected to perform a sleep operation until the number of remaining active nodes reaches the target value. The number list of the nodes that are put to sleep is recorded. Through the number list of the nodes that are put to sleep and the target number of active nodes, the node activation condition in the energy valley period is set, wherein the minimum load threshold is set to a preset lower limit of the average load value of the cluster, the node running efficiency requirement reaches a preset efficiency threshold, the tidal energy supply power fluctuation amplitude is less than a preset stability threshold, and the duration exceeds a preset monitoring duration. The energy valley period activation condition includes the minimum load threshold, node running efficiency, and energy supply stability.

[0069] Specifically, in an embodiment, the calculation of the current load value of the server cluster adopts a three-dimensional weighted evaluation method. The weight coefficient of CPU usage rate is set to 0.5 because the processor resource is the core element determining the computing power. The weight coefficient of memory occupancy rate is set to 0.3, reflecting the demand for data cache and processing space. The task queue length needs to be normalized by dividing the queue length by the upper limit of the queue capacity to obtain a normalized value, and the weight coefficient is set to 0.2. The weighted sum of the three dimensions forms a comprehensive load value, the numerical range of which is between 0 and 1.

[0070] Specifically, the execution of the node sleep operation follows a gradual closing process. A preparation signal is first sent to the node to be slept, and the node stops receiving new tasks after receiving the signal and completes the task being executed. After the task is completed, the node synchronizes the key data in the memory to the shared storage to ensure that the data is not lost. Then the node enters the sleep state, closes the CPU frequency, reduces the memory refresh frequency, stops the network card data transmission, and only retains the wake-up listening function. The entire sleep process usually takes less than 30 seconds.

[0071] It should be noted that the setting of the low load standard is based on statistical analysis of historical running data. The load values of each node in different time periods in the past 30 days are collected, and the 25th percentile of the load distribution is calculated as the low load standard. When the load value of the node is lower than this standard, it indicates that the node is in a light load state and is the preferred object of the sleep operation.

[0072] Exemplarily, the three dimensions of the energy valley period activation condition are related to each other. The minimum load threshold ensures that the sleep node is activated only when the overall load of the cluster is reduced to a low level, avoiding frequent sleep-wake switching. The node running efficiency requires evaluating the resource utilization efficiency of the activated node, and when the efficiency value is above the preset threshold, it indicates that the existing node can efficiently process the current task and does not need to wake up more nodes. The stability of energy supply is judged by monitoring the fluctuation of tidal power. The power fluctuation amplitude remains within the preset range within the continuous monitoring time, indicating that the energy supply enters the stable period, which is suitable for adjusting the node running scale.

[0073] Preferably, in practical applications, the data center reduces energy consumption through the node sleep mechanism during the tidal energy low tide period. When the tide is in the low tide period at 2 am, the power generation power is reduced to 40% of the peak value, and at this time the user access volume is also relatively small, and the cluster load value is reduced to below 0.3. According to the load sorting, 30% of the nodes are selected to execute sleep, and the number of active nodes is reduced from 100 to 70.

[0074] Step S105, according to the determined trough period activation condition, the excess tidal energy is converted into electrochemical energy storage or mechanical energy storage form storage, preferentially allocated to high efficiency energy storage unit, and dynamically released in the low valley period to supplement the energy gap, to store excess tidal energy, get the available capacity index of buffer energy.

[0075] According to the determined trough period activation condition, the difference between the tidal power generation power and the real-time power consumption of the data center is detected, if the difference is positive, it is identified as excess energy, and the excess tidal energy is allocated according to the rated input power ratio of the electrochemical energy storage unit and the mechanical energy storage unit, to obtain the input power allocation value of each energy storage channel. The input power allocation value is used to perform the charging process of the electrochemical energy storage unit, monitor the state of charge of the battery pack, and drive the flywheel of the mechanical energy storage unit to accelerate or charge the compressed air storage tank. Record the energy storage rate and conversion efficiency of each energy storage unit, and preferentially allocate more energy to the energy storage unit with conversion efficiency exceeding the preset threshold. By continuously storing excess energy, when it is detected that the tidal energy enters the trough period and the data center power consumption exceeds the tidal power generation power, the energy gap value is calculated, the release order is determined according to the remaining energy storage of each energy storage unit and the discharge efficiency, and the output power of the energy storage unit is controlled to dynamically release the energy to supplement the gap. The current remaining capacity of each energy storage unit, the overall conversion efficiency from storage to release, and the instantaneous release rate peak value are calculated to form the available capacity index of the buffer energy, which includes the buffer energy capacity, conversion efficiency, and release rate.

[0076] Specifically, in an embodiment, the detection of excess tidal energy is based on real-time power difference calculation. The output power of the tidal power station is collected by a power transmitter every second, and the total power consumption of the data center is monitored in real time by the total meter of the intelligent power distribution cabinet. When the tidal power generation power is 15 megawatts and the data center only consumes 10 megawatts, the difference of 5 megawatts is excess energy. These excess energies need to be stored in time to avoid energy loss caused by direct feedback to the power grid.

[0077] Specifically, the power allocation of the energy storage channel follows the capacity matching principle. The electrochemical energy storage unit uses lithium iron phosphate battery packs, and the rated input power of a single battery cabinet is 500 kilowatts. The entire energy storage station is configured with 10 battery cabinets, with a total rated input power of 5 megawatts. The mechanical energy storage adopts flywheel energy storage and compressed air energy storage, with a rated input power of 2 megawatts for the flywheel energy storage system and 3 megawatts for the compressed air energy storage system. According to the rated input power ratio of each energy storage unit, the excess energy is allocated to the three energy storage channels in the ratio of 5:2:3. When the excess energy is 5 megawatts, the electrochemical energy storage is allocated 2.5 megawatts, the flywheel energy storage is allocated 1 megawatt, and the compressed air energy storage is allocated 1.5 megawatts.

[0078] Exemplarily, the charging process of the electrochemical energy storage adopts a three-stage charging strategy. The first stage is constant current charging, and the charging current is set to 0.5 times the capacity of the battery. In this stage, the battery quickly absorbs energy, and the state of charge rises from 20% to 80%. The second stage is constant voltage charging. When the battery voltage reaches the rated value, the voltage remains unchanged, and the charging current gradually decreases, and the state of charge rises from 80% to 95%. The third stage is trickle charging, which maintains the full charge state of the battery with a very small current. During the entire charging process, the battery management system monitors the voltage, temperature and internal resistance of each battery cell in real time. When the temperature exceeds 45 degrees Celsius, the charging power is reduced to ensure charging safety.

[0079] It should be noted that the working principles of mechanical energy storage systems have their own characteristics. Flywheel energy storage drives the flywheel to rotate at high speed in a vacuum chamber, converting electrical energy into rotational kinetic energy. Flywheels are made of carbon fiber composite materials, with a maximum speed of 20,000 rpm. A single flywheel can store 50 kilowatt-hours of energy. Compressed air energy storage uses a compressor to compress air into a high-pressure tank, with a pressure of up to 30 megapascals. The heat generated during compression is recovered through a heat exchanger and stored in thermal storage materials, improving overall energy conversion efficiency. When energy needs to be released, high-pressure air expands to do work through a turbine generator, while stored heat is used to heat the air, improving power generation efficiency.

[0080] Preferably, real-time monitoring and optimization of conversion efficiency is the key to improving the performance of the energy storage system. The charging and discharging efficiency of electrochemical energy storage is calculated by comparing the input and output energy. The round-trip efficiency of lithium iron phosphate batteries is usually between 90% and 95%. The efficiency of flywheel energy storage is mainly affected by bearing friction and wind resistance, and can reach 85% to 90% in a vacuum environment. The efficiency of compressed air energy storage is relatively low, but can be improved to 70% to 75% through heat recovery technology. According to the real-time efficiency of each energy storage unit, the power distribution is dynamically adjusted. When the electrochemical energy storage efficiency is higher than the preset threshold of 92%, increase its allocation proportion to 60%; when the flywheel energy storage efficiency is lower than 85%, reduce its allocation proportion to 15%.

[0081] In one possible implementation, the dynamic supplement of energy gap adopts a hierarchical release strategy. When the tidal energy enters the trough period, the power generation power drops to 8 megawatts, while the data center load is 12 megawatts, forming a 4 megawatt energy gap. First, evaluate the remaining capacity and discharge capacity of each energy storage unit. Electrochemical energy storage is fast in response and stable in output, and is the first priority release source. Flywheel energy storage responds in milliseconds and is suitable for responding to transient power fluctuations, making it the second priority. Compressed air energy storage has a longer start-up time but a longer continuous power supply time, making it the third priority. According to the size and duration of the gap, the system automatically combines different energy storage units to achieve smooth energy supplement.

[0082] For example, during the low tide period in the early morning, the tidal power generation power drops to the lowest point, and the data center still has batch computing tasks to be executed. After detecting a 3-megawatt energy gap, the electrochemical energy storage is immediately started to release 2 megawatts of power, and the compressed air energy storage system is preheated. After 5 minutes, the compressed air energy storage starts to release 1 megawatt of power, and the electrochemical energy storage correspondingly reduces the output, maintaining a total output of 3 megawatts. This coordinated release strategy not only meets the immediate energy demand, but also extends the power supply time of the energy storage system.

[0083] In one embodiment, the statistics of the buffer energy available capacity index provide decision basis for system operation and maintenance. The energy storage capacity is obtained by accumulating the remaining power of each energy storage unit, the electrochemical energy storage has a remaining capacity of 2000 kilowatt-hours, the flywheel energy storage has a remaining capacity of 100 kilowatt-hours, and the compressed air energy storage has a remaining capacity of 1500 kilowatt-hours, with a total capacity of 3600 kilowatt-hours. The conversion efficiency is the weighted average of each energy storage unit, and the weight is determined according to the actual release power ratio. The release rate records the maximum release power of the system in the past 24 hours, reflecting the instantaneous power supply capacity of the system.

[0084] It can be understood that by establishing a perfect energy storage system and dynamic control mechanism, the data center realizes the time shift utilization of tidal energy. In the energy sufficient period, the excess energy is stored, and in the energy deficient period, the energy storage is released to supplement the gap, forming a stable and reliable green energy supply system. The three-dimensional capacity index reflects the running state of the energy storage system in real time, providing quantitative support for energy dispatching decision-making, and ensuring that the data center can still operate stably in complete dependence on tidal energy.

[0085] During the low tide period, first receive real-time energy demand data and tidal energy supply data, identify the time period and size of the energy gap, then dynamically release the excess tidal energy stored in the energy storage unit, adjust the release rate to match the gap demand, and monitor the changes of the remaining capacity, energy conversion efficiency and release rate of the energy storage unit in real time during the adjustment process, and convert it into the available capacity index of the buffer energy, evaluate the conversion efficiency, and combine the release rate record to obtain the comprehensive index including energy storage capacity, conversion efficiency and release rate.

[0086] In the low valley period, real-time energy demand data from the data center and tidal energy supply data from the tidal power station are received, the difference between demand power and supply power at each time point is calculated by comparing the time series curves of the two sets of data, and if the demand is greater than the supply, the period is marked as a gap period, and the gap start time, duration and power difference value are recorded to obtain the period distribution and scale of the energy gap. According to the period distribution and scale of the energy gap, the current storage amount of the electrochemical energy storage and mechanical energy storage in the energy storage unit is read, the gap power value is taken as the release target, the energy storage inverter and the generator set are controlled to output electric energy according to the target power, the pulse width modulation duty cycle of the inverter and the speed of the generator set are adjusted to make the actual release rate match the gap demand, and a dynamic release power curve is obtained. The energy release is performed through the dynamic release power curve, the remaining capacity of the energy storage unit is collected in real time as the remaining capacity, the ratio of output power to consumed energy storage is calculated as the conversion efficiency, and the instantaneous maximum output power is recorded as the release rate. The three parameters are updated and recorded in real time to obtain the energy storage operation state data set. Using the energy storage operation state data set, the average conversion efficiency in the entire release period is calculated, the peak and average of the release rate are extracted, and the data in three dimensions of energy storage capacity, conversion efficiency and release rate are combined to obtain a comprehensive index containing energy storage capacity, conversion efficiency and release rate.

[0087] Specifically, in an embodiment, the identification of energy gap is based on accurate comparison and analysis of time series curves. The energy demand data of the data center is collected by a distributed sensor network, and each server rack is equipped with a power monitoring module to report instantaneous power values at a frequency of seconds. The supply data of the tidal power station comes from the output end of the generator set, and the three-phase electric parameters are measured by current transformers and voltage transformers to calculate the real-time power generation. The two data streams converge to the energy management platform to form time-aligned power curves. When the demand power is 12 megawatts and the supply power is only 8 megawatts at a certain time, the 4-megawatt difference is the energy gap at that time.

[0088] Specifically, the adjustment of the pulse width modulation duty cycle realizes the accurate control of the energy storage release power. The energy storage inverter adopts a full-bridge topology structure, and the effective value of the output voltage is adjusted by controlling the conduction time ratio of the switching tube. The duty cycle is continuously adjustable from 0 to 100%, and the output power changes linearly from zero to rated power. When 2 megawatts of power needs to be released, the controller calculates the required duty cycle value according to the current DC bus voltage and load impedance, which is usually between 60% and 80%. The modulation frequency is set to 20 kHz, which is higher than the human hearing range to avoid noise interference. The speed control of the generator set uses variable frequency speed regulation technology to adjust the speed by changing the power supply frequency of the stator winding, thereby controlling the output power.

[0089] Exemplarily, the forming process of the dynamic release power curve involves the coordination of multiple control links. First, the total release power target is determined according to the gap size, and then the release task is allocated according to the current capacity and efficiency of each energy storage unit. Electrochemical energy storage undertakes the basic load part due to its fast response speed, and maintains a relatively stable output. Mechanical energy storage is responsible for coping with fluctuations, and dynamically adjusts the output according to real-time gap changes. The controller updates the power distribution instructions every 100 milliseconds to ensure that the total output power always matches the gap demand. During the release process, if it is detected that a certain energy storage unit is close to depletion, the system automatically transfers its load to other units to achieve seamless switching. This dynamic coordination mechanism ensures the continuity and stability of energy supply.

[0090] It should be noted that the construction of the energy storage operating state data set relies on multi-dimensional real-time monitoring. The monitoring of the remaining capacity is for electrochemical energy storage, which accumulates the charge and discharge energy by coulomb counting method, and combines the open circuit voltage method to correct the accurate state of charge. For mechanical energy storage, the flywheel system calculates the kinetic energy storage by measuring the speed, and the compressed air system calculates the gas storage by pressure and temperature sensors. The calculation of conversion efficiency requires simultaneous measurement of input and output energy. The input energy is obtained by integrating the charging power, and the output energy is obtained by integrating the discharging power. The ratio of the two is the conversion efficiency. The release rate is directly measured by the power transducer to measure the instantaneous output power, and the peak value is recorded.

[0091] Preferably, the data sampling and processing in the real-time monitoring process adopts a hierarchical architecture. The bottom layer acquisition device samples the original signal at a millisecond level, and removes high-frequency noise through digital filtering. The intermediate layer aggregation node performs preliminary processing on the data, calculates the moving average and standard deviation. The top layer management platform receives the processed data, performs trend analysis and anomaly detection. When a sudden decrease in conversion efficiency of more than 5% is detected, the system triggers the diagnostic program to check the health status of the energy storage unit.

[0092] In one possible implementation, the statistical analysis of the comprehensive index provides a quantitative basis for system optimization. The entire release cycle usually lasts for 3 to 4 hours, corresponding to the length of the tidal low period. The system records a set of operating parameters every minute, including the output power, remaining capacity, internal temperature, etc. of each energy storage unit. After the release cycle ends, all data are statistically processed. The average conversion efficiency is obtained by arithmetic averaging of the efficiency values at each time, reflecting the overall performance of the energy storage system. The peak and average values of the release rate reflect the maximum power supply capacity and average power supply level of the system, respectively.

[0093] For example, during a typical trough period, the system detects a 3 MW energy gap. The electrochemical storage releases 2 MW immediately, the flywheel storage releases 0.5 MW, and the compressed air storage releases 0.5 MW after 2 minutes of preheating. In the following 3 hours, the gap size fluctuates between 2 and 4 MW, and the system maintains the balance by dynamically adjusting the output of each storage unit. Throughout the process, the conversion efficiency of the electrochemical storage is maintained at 92%, the overall efficiency of the mechanical storage is 78%, and the total system efficiency reaches 87%.

[0094] In an embodiment, the visualization of the three-dimensional comprehensive index helps the operation and maintenance personnel to intuitively understand the system status. The storage capacity is displayed in a column chart, and different colors represent different types of storage units. The conversion efficiency is shown in a line chart to show the historical trend and facilitate the discovery of performance degradation. The release rate is presented in a heat map, and the color depth represents the power size.

[0095] Step S106, analyze the degree of relief of resource mismatch from the available capacity index of the buffer energy, and reallocate the idle computing resource according to the degree of relief, to determine the supplement demand of the tidal energy shortage period.

[0096] The storage capacity, conversion efficiency and release rate values are extracted from the available capacity index of the buffer energy, the storage capacity is multiplied by the conversion efficiency to obtain the actual available energy, and compared with the theoretical demand energy of the data center in the corresponding period. If the ratio exceeds the preset threshold, it is determined that the resource mismatch is relieved, and a resource mismatch relief degree value is obtained. According to the resource mismatch relief degree value, the computing nodes in the idle state in the data center are identified, and the idle computing power corresponding to the period with high relief degree is reallocated to the low priority tasks in the task queue to be executed, to obtain a computing resource reallocation table. Through the computing resource reallocation table, the number of computing nodes that need to be started additionally in the tidal energy shortage period is counted as the computing power supplement amount, the energy supply priority is determined according to the task urgency ranking, the task average waiting time is calculated as the task scheduling delay, and the tidal energy shortage period supplement demand including the computing power supplement amount, the energy supply priority and the task scheduling delay is determined.

[0097] Specifically, in an embodiment, the evaluation of the resource mismatch relief degree is based on the actual supply capacity of the storage system. The actual available energy is calculated by multiplying the storage capacity by the conversion efficiency. For example, if the storage capacity is 1000 kWh and the conversion efficiency is 85%, the actual available energy is 850 kWh. The theoretical demand energy of the data center is calculated based on historical operation data, which is usually 800 to 1200 kWh during the trough period. When the ratio of the actual available energy to the theoretical demand energy exceeds 0.8, it is determined that the resource mismatch is effectively relieved.

[0098] Specifically, the identification and redistribution of idle computing power follows the load balancing principle. The resource usage of all computing nodes is scanned, and the nodes with CPU utilization lower than 30% and memory occupancy lower than 40% are marked as idle. According to the degree of relief value, the system classifies the management of these idle computing power. When the relief degree is higher than 0.9, the idle computing power can be allocated to low-priority tasks such as batch data processing and log analysis; when the relief degree is between 0.7 and 0.9, only interruptible background tasks are allocated. This differentiated allocation ensures the resource guarantee of critical tasks.

[0099] It should be noted that the calculation of the amount of computing power supplement takes into account the real-time state of the task queue. The resource requirements of each task in the task queue to be executed are monitored, and the total demand is accumulated. Compared with the total computing power of the current active nodes, the difference is the number of nodes that need to be started additionally. If there are 50 video transcoding tasks in the queue, each requiring 2 CPU cores, and only 80 cores are available, the amount of computing power supplement is 20 cores, corresponding to the need to start 5 four-core nodes.

[0100] Exemplarily, the determination of energy supply priority is based on the business attributes of the tasks. Real-time live transcoding tasks are set to the highest priority due to their high timeliness requirements; user interactive queries follow; batch processing and data backup tasks have the lowest priority. The task scheduling delay is calculated by recording the time interval from task submission to execution, and the system maintains a sliding time window to calculate the average waiting time of all tasks in the past hour. In actual application, when the tidal energy enters the insufficient period, the system allocates resources according to the determined supplement demand. If the amount of computing power supplement is 10 nodes, the new type of server with high energy efficiency ratio is preferentially awakened; the energy supply is allocated according to the priority order to ensure that high-priority tasks obtain sufficient power; for tasks that can accept a certain delay, their scheduling delay is appropriately extended, and they are executed after the energy is sufficient.

[0101] Step S107, obtaining the task execution log matched with the period of sufficient tidal energy supply after dynamic migration, analyzing the task execution efficiency and energy consumption data in the task execution log, and combining the supplement demand in the period of insufficient tidal energy to generate a computing power optimization scheme for the data center driven by green energy.

[0102] The task execution log matching the period of sufficient tidal energy supply after dynamic migration is obtained, the task identifier, start time, end time, CPU occupancy, memory usage data are extracted from the log file, the execution duration is obtained by subtracting the start time from the end time, the CPU occupancy and memory usage are averaged as the resource utilization, and the task execution efficiency index set is obtained. According to the task execution efficiency index set, the corresponding period electric energy consumption data recorded by the data center power distribution cabinet is read, the total electric energy consumption is divided by the number of completed computing tasks to obtain the unit task energy consumption, and the computing power supplement amount, energy supply priority and task scheduling delay parameter in the tidal energy insufficient period supplement demand are extracted to obtain the energy consumption evaluation result. Through the energy consumption evaluation result, high energy consumption tasks are arranged to be executed in the period of sufficient tidal energy, and low energy consumption tasks are dispersed in each period, and the hourly power quota and energy storage release plan are prepared according to the tidal energy supply curve to obtain the periodical scheduling and energy distribution rule. According to the scheduling and energy distribution rule, the number of server nodes that need to be activated or hibernated is determined according to the task load in each period, and the task execution period arrangement, energy quota in each period and node increase and decrease plan are integrated to generate a data center computing power optimization scheme under green energy driving containing task scheduling, energy distribution plan and node dynamic adjustment.

[0103] Specifically, in an embodiment, the acquisition and analysis of task execution logs constitute the data basis of the entire optimization scheme. The task management system of the data center records detailed logs for each executed task, including task unique identifier, submission time, start execution time, completion time, allocated CPU core number, actual CPU usage, memory allocation, actual memory occupancy, disk IO throughput and other multi-dimensional information. These log files are stored in a distributed file system in chronological order, with one log file generated every hour, facilitating periodical retrieval and analysis.

[0104] Specifically, the calculation of task execution efficiency involves the comprehensive evaluation of multiple performance indicators. The execution duration is obtained by subtracting the start execution time from the completion time, reflecting the time efficiency of the task. The resource utilization rate needs to be calculated for CPU and memory respectively. CPU utilization rate is equal to the actual used CPU time divided by the allocated CPU time, and memory utilization rate is equal to the peak memory occupancy divided by the allocated memory capacity. For video transcoding tasks, the typical execution duration is 30 to 120 minutes, and the CPU utilization rate is usually between 70% and 90%; the execution duration of data backup tasks is longer, up to 3 to 5 hours, but the CPU utilization rate is relatively low, about 40% to 60%. Through statistical analysis of these indicators, the execution characteristics and resource consumption patterns of different types of tasks can be identified.

[0105] Exemplarily, the collection and analysis of energy consumption data need to be accurate to the task level. The intelligent power distribution system of the data center installs intelligent power meters on each server rack to record power consumption with second-level precision. When a task starts to execute, the system records the initial power meter reading of the server where the task is located; when the task is completed, the power meter value is read again, and the difference between the two readings is the total energy consumption of the task. The calculation of unit task energy consumption needs to consider the computing amount of the task, usually taking the number of floating-point operations or the amount of data processed as the measurement unit of computing amount.

[0106] For example, a transcoding task that processes 1 TB of video data consumes 500 kilowatt-hours of electricity, so the unit energy consumption is 0.5 kilowatt-hours per GB. This fine-grained energy consumption analysis provides a quantitative basis for subsequent task scheduling optimization.

[0107] It should be noted that the three dimensions of the demand for energy supplement during the tidal energy shortage period are related to each other. The amount of computing power supplement reflects the number of additional computing resources that need to be started, which is obtained by calculating the difference between the total resource demand in the task queue and the currently available resources. The energy supply priority is determined according to the business importance and timeliness requirements of the task, and real-time streaming processing has the highest priority, while batch processing has the lowest priority. The task scheduling delay represents the maximum waiting time that the task can tolerate. For real-time tasks, the delay requirement is in seconds; for batch processing tasks, a delay of several hours can be accepted.

[0108] Preferably, the formulation of the periodical scheduling rule is based on the periodic characteristics of the energy supply curve. A day is divided into 24 periods, each corresponding to one hour. According to historical data analysis, tidal energy peaks from 14:00 to 17:00, which is marked as an energy sufficient period. High-energy-consuming tasks such as large-scale data analysis and machine learning model training are preferentially arranged to be executed during these periods. Low-energy-consuming tasks such as log archiving and configuration backup are scattered in various periods to fully utilize the basic energy supply.

[0109] In one possible implementation, the preparation of the energy allocation plan needs to balance multiple constraints. The power quota of each period is determined according to the tidal power prediction value, usually with a 10% safety margin. The energy storage release plan is formulated according to the current capacity of the energy storage system and the expected energy gap. When the task load of a period exceeds the tidal energy supply, the electrochemical energy storage is preferentially used for supplement due to its fast response speed; if the gap lasts for a long time, the mechanical energy storage system is started. The quota allocation follows the priority principle to ensure that critical tasks have sufficient energy supply.

[0110] For example, in a typical operation cycle, the data center has scheduled 50 video transcoding tasks and 20 data backup tasks during the low-traffic period in the early morning. According to the efficiency analysis of task execution, these tasks are expected to consume 8 megawatt-hours of electricity. At the same period, the tidal power prediction value is 10 megawatt-hours, leaving a 2 megawatt-hour margin. Based on this, the energy allocation plan is formulated: 6 megawatt-hours are allocated to video transcoding tasks, 2 megawatt-hours are allocated to backup tasks, and the remaining 2 megawatt-hours are used for energy storage system charging to prepare for subsequent energy shortage periods.

[0111] In one embodiment, the decision of node dynamic adjustment is based on real-time load monitoring and prediction. A node status table is maintained, recording the running state, current load and energy consumption level of each server. When the task load of a certain period increases, the system activates the dormant nodes in order of energy efficiency ratio from high to low; when the load decreases, the old servers with lower energy efficiency ratio are selected for hibernation first. This dynamic adjustment mechanism ensures that the energy consumption is minimized while meeting the computing demand.

[0112] It can be understood that the computing power optimization scheme is a comprehensive solution, including three core components: task scheduling strategy, energy allocation plan and node management. The task scheduling strategy specifies the execution arrangement of various tasks at different periods; the energy allocation plan specifies the power quota and energy storage usage scheme for each period; the node dynamic adjustment scheme determines the start and stop time and number of servers. These three parts work together to achieve efficient use of computing resources in the tidal energy-constrained data center, achieving the dual goals of green energy saving and stable operation.

[0113] It is apparent to those skilled in the art that the present application is not limited to the details of the foregoing exemplary embodiments, and can be implemented in other concrete forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all respects as illustrative and not restrictive, and the scope of the present application is defined by the appended claims rather than the foregoing description, and all changes falling within the meaning and range of equivalents of the elements of the claims are intended to be embraced therein. Any reference signs in the claims should not be considered as limiting the claims involved.

Claims

1. A method for optimizing computing power driven by green energy in a data center, characterized in that, The method comprises the following steps: analyzing the matching degree of the tidal energy supply data and the computing load demand data to determine the peak misalignment period of the tidal energy supply and the computing load demand; According to the peak misalignment period, the execution window of the high-energy-consumption computing task is evaluated, and an available resource list of the energy sufficient period is generated; the priority sequence of the high-energy-consumption task is extracted from the available resource list, and the task matching the tidal energy supply period is dynamically migrated to the corresponding period for execution, and the balance degree of the computing power resource allocation after migration is judged; According to the balance degree of the computing power resource allocation and the current load of the server cluster, the size of the server cluster is adjusted to the target active node number, and the low valley period activation condition is determined; according to the low valley period activation condition, the available capacity index of the buffer energy is generated; according to the available capacity index of the buffer energy, the idle computing power resources are redistributed, and the supplement demand of the tidal energy insufficient period is determined; the task execution efficiency and energy consumption data in the task execution log after dynamic migration are analyzed, and the tidal energy insufficient period supplement demand is combined to generate an algorithm optimization scheme; The method comprises the following steps: Collecting tidal height, flow rate and power generation data of a tidal power station, and simultaneously collecting CPU usage rate, memory occupancy rate and network bandwidth consumption data of a data center server; performing time series decomposition on the tidal power station data to extract periodic characteristic curves; according to the periodic characteristic curves, the computing load demand data is processed by using a sliding window method, the load mean and variance are calculated, and the load peak time is identified; by comparing the tidal energy supply peak time with the load peak time, a supply-demand time sequence deviation matrix is generated; from the supply-demand time sequence deviation matrix, a period with a time offset greater than a threshold value is extracted and marked as a misalignment period, and a misalignment period record table containing the misalignment start time, duration and energy difference is generated; The method comprises the following steps: According to the peak misalignment period, the power demand and running time of video transcoding, batch data processing and machine learning training tasks are identified; comparing the tidal energy supply data with the power demand, the time interval in which the energy supply exceeds the task demand is determined as the execution window; calculating the net power value after deducting the infrastructure power consumption from the tidal power generation power in the execution window, determining the maximum number of nodes supported and the total storage capacity; according to the maximum number of nodes, querying the network device configuration table, accumulating the node network card rate as the bandwidth demand value, adjusting the bandwidth quota, and generating an available resource list containing the number of computing nodes, storage capacity and bandwidth allocation.

2. The computing power optimization method of claim 1, wherein, The method comprises the following steps: Read the number of computing nodes, storage capacity and bandwidth allocation data from the available resource list, combine the difference between the task deadline and the current time, generate the priority sequence of high energy consumption tasks; compare the execution time of the tasks in the priority sequence with the tidal energy supply period, mark the matching tasks, and construct the task migration candidate set; migrate the tasks in the task migration candidate set to the corresponding computing nodes, and record the CPU utilization, memory usage and bandwidth occupancy after migration; evaluate the balance degree of computing resource allocation through the variance value of CPU utilization.

3. The computing power optimization method of claim 2, wherein, After generating the available resource list of the energy sufficient period, it includes: Extract the execution requirement parameters of the tasks in the task migration candidate set, including minimum running time, maximum allowed delay, resource occupancy and interruptability; calculate the overlap ratio of the time window of the tasks in the task migration candidate set and the tidal energy supply period, and determine the fitting degree; according to the fitting degree, adjust the task scheduling plan, and migrate the task execution time to the starting time of the tidal energy supply period; according to the minimum running time, set the task end time, and generate the task execution time table after migration adjustment.

4. The computing power optimization method of claim 1, wherein, According to the balance degree of computing resource allocation and the current load of the server cluster, adjust the server cluster size to the target active node number, and determine the low valley period activation condition, including: According to the balance degree of computing resource allocation, obtain the CPU usage, memory occupancy and task queue length of the server cluster, and calculate the current load value; if the current load value exceeds the threshold value, sort the node load value, and perform the sleep operation to reach the target active node number; according to the target active node number, set the minimum load threshold, node running efficiency and energy supply stability, and generate the low valley period activation condition.

5. The computing power optimization method of claim 1, wherein, According to the low valley period activation condition, the generated buffer energy available capacity index includes: According to the low valley period activation condition, detect the difference between the tidal power generation and the data center power consumption, and allocate the excess energy to the electrochemical or mechanical energy storage unit; perform the charging or mechanical energy storage process, record the energy storage growth rate and conversion efficiency; when the tidal energy supply data is lower than the power consumption, control the energy storage unit to release energy, and count the remaining capacity, conversion efficiency and release rate to generate the buffer energy available capacity index.

6. The computing power optimization method of claim 5, wherein, The generation of the buffer energy available capacity index includes: Compare the tidal energy supply data and the data center power consumption, mark the energy gap period and the power difference; according to the power difference, control the energy storage unit to output electric energy, adjust the release rate to match the gap; collect the ratio of the remaining electric quantity, output electric quantity and consumed energy quantity of the energy storage unit and the instantaneous maximum output power; count the average conversion efficiency and release rate peak value in the release period, and generate the buffer energy available capacity index.

7. The computing power optimization method of claim 1, wherein, According to the buffer energy available capacity index, redistribute the idle computing resource, determine the supplement demand in the tidal energy insufficient period, including: Extracting energy storage capacity and conversion efficiency from the buffer energy available capacity index, calculating the actual available energy; comparing the actual available energy with the demand energy to determine the resource mismatch relief degree value; according to the resource mismatch relief degree value, allocating idle computing power to low priority tasks; counting the number of additional start nodes, task urgency and average waiting time to generate a supplementary demand containing computing power supplement, energy supply priority and task scheduling delay.

8. The computing power optimization method of claim 1, wherein, The analysis of the task execution efficiency and energy consumption data in the task execution log after dynamic migration, combined with the tidal energy shortage period supplementary demand, generates a computing power optimization scheme, including: Extracting task identification, execution duration and resource utilization from the task execution log to generate a task execution efficiency index set; calculating the unit task energy consumption according to the task execution efficiency index set; according to the tidal energy shortage period supplementary demand, arranging high energy consumption tasks to energy sufficient period, compiling power quota and energy storage release plan; according to the task load, determine the number of node activation or dormancy, generate a computing power optimization scheme.

Images