Intelligent adjustment method and system for peak load of supercapacitor energy storage system of computing power center

Abstract

The application provides a peak load intelligent adjustment method and system for a supercapacitor energy storage system of a computing center, relates to the technical field of energy management, and comprises the following steps: a coupling decision mechanism is constructed between a task scheduling layer and an energy storage scheduling layer, a plurality of candidate task execution time sequences are generated based on to-be-executed task information, a charging and discharging time window is correspondingly decomposed from a load peak value period, charging and discharging power and a time sequence are allocated, residual peak power is calculated and an optimal scheme is selected, and real-time monitoring and dynamic adjustment are performed, so that the peak load of the computing center is effectively reduced, the energy utilization efficiency is improved, and power grid fluctuation is reduced.

Description

Technical Field

[0001] This invention relates to energy management technology, and more particularly to a method and system for intelligent peak load adjustment of a supercapacitor energy storage system in a computing center. Background Technology

[0002] With the rapid development of big data and artificial intelligence technologies, the scale of computing centers is constantly expanding, and their energy consumption and power demand are showing a rapid growth trend. Computing centers typically have a large number of computing devices, which generate significant power fluctuations and peak loads when running high-load computing tasks. This not only increases the pressure on the power supply system but also leads to an increase in electricity costs. To address this challenge, supercapacitor energy storage systems, due to their high power density, rapid charging and discharging capabilities, and long cycle life, have become an important means for computing centers to smooth out peak loads and reduce load fluctuations.

[0003] The existing task scheduling and energy storage system management are independent of each other, lacking an effective coordination mechanism. This results in energy storage resources being unable to be optimally configured according to computing load characteristics, reducing the overall system efficiency. Secondly, existing energy storage control strategies often use fixed thresholds or simple rules for charging and discharging decisions, which cannot adapt to the complex and ever-changing load characteristics of computing centers. In particular, when facing sudden high-load tasks, it is difficult to achieve precise peak load adjustment. Summary of the Invention

[0004] This invention provides a method and system for intelligent peak load adjustment of a supercapacitor energy storage system in a computing center, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for intelligent peak load adjustment of a supercapacitor energy storage system in a computing center, comprising:

[0006] Obtain information on tasks to be executed from the computing center, status information of the supercapacitor energy storage system, and historical load characteristic data;

[0007] A coupled decision-making mechanism is constructed between the task scheduling layer and the energy storage scheduling layer. The task scheduling layer generates multiple candidate task execution sequences based on the information of the tasks to be executed. The energy storage scheduling layer decomposes the peak period in the corresponding load power curve into multiple charge and discharge time windows for each candidate task execution sequence. Based on the available capacity in the status information and the historical load characteristic data, it allocates charge and discharge power and timing to each charge and discharge time window and calculates the remaining peak power after charge and discharge. The task scheduling layer selects the target task execution sequence and the corresponding target charge and discharge scheme based on the remaining peak power.

[0008] During execution, the actual load power is monitored. When the deviation between the actual load power and the expected value exceeds a set threshold, the coupled decision-making mechanism is re-executed to generate a correction scheme, and the actual load power is added to the historical load characteristic data. Tasks are allocated according to the correction scheme, and the supercapacitor energy storage system is controlled to perform charging and discharging.

[0009] The steps of generating multiple candidate task execution sequences by the task scheduling layer include:

[0010] Based on the resource requirements and time constraints of each task in the task information to be executed, the schedulable time range of each task is determined; within the schedulable time range, multiple candidate task execution sequences are generated by adjusting the task execution order and start time.

[0011] For each candidate task execution sequence, a corresponding load power curve is calculated based on the resource demand characteristics. The load power curve represents the power demand corresponding to the total resource occupancy of all tasks at each time point. The load power curve is then transmitted to the energy storage scheduling layer for the energy storage scheduling layer to perform charging and discharging power allocation and peak period decomposition.

[0012] The energy storage scheduling layer decomposes the peak period in the corresponding load power curve into multiple charge / discharge time windows for each candidate task execution sequence. Based on the available capacity in the status information and the historical load characteristic data, it allocates charge / discharge power and timing to each charge / discharge time window. The steps for calculating the remaining peak power after charge / discharge execution include:

[0013] The period in the load power curve where the power exceeds the reference power threshold is identified as the peak period; the peak period is divided into multiple charge and discharge time windows according to the time granularity.

[0014] Based on the peak-valley period distribution pattern in the historical load characteristic data, the pre-charging period of the supercapacitor before the peak period is determined, and the energy that can be stored is calculated based on the upper limit of charging power in the status information and the duration of the pre-charging period.

[0015] Based on the storable energy and the available capacity, calculate the excess power of the load power exceeding the reference power threshold at the corresponding moment of each charge and discharge time window, and use the excess power as the target discharge power;

[0016] The charging sequence is determined based on the end time of the pre-charging period, and the discharging sequence is determined based on the start and end times of each charging and discharging time window within the peak period, thus forming a charging and discharging sequence.

[0017] Based on the target discharge power of each charge / discharge time window and the charge / discharge sequence, calculate the load power reduction at the corresponding moment after discharge in each charge / discharge time window; subtract the corresponding load power reduction from the power value at each moment in the load power curve to obtain the remaining peak power after charge / discharge.

[0018] The steps of the task scheduling layer selecting the target task execution sequence and the corresponding target charging and discharging scheme based on the remaining peak power include:

[0019] The remaining peak power is compared with the upper limit of the grid capacity, and candidate task execution sequences with remaining peak power not exceeding the upper limit of the grid capacity are selected as a set of feasible sequences.

[0020] For each candidate task execution sequence in the feasible timing set, the grid power supply margin is calculated based on the difference between the remaining peak power and the grid capacity limit; the capacity decay is calculated based on the charge / discharge depth, charge / discharge power in the target charge / discharge scheme, and the current available capacity in the status information; based on the capacity decay, the task arrival frequency distribution and peak load periodicity characteristics in the historical load characteristic data, the available capacity decay trajectory of the supercapacitor in the future multiple decision cycles is deduced; the number of decision cycles required for the capacity to drop to the capacity failure threshold is determined based on the available capacity decay trajectory; and the time length from the current moment to the moment corresponding to the number of decision cycles is calculated as the remaining lifetime of the energy storage system.

[0021] Based on the power grid supply margin, the task completion time of the candidate task execution sequence, and the remaining lifespan of the energy storage system, a comprehensive evaluation index is calculated. The candidate task execution sequence with the best comprehensive evaluation index is selected as the target task execution sequence, and the corresponding charging and discharging power and timing are extracted as the target charging and discharging scheme.

[0022] The steps for calculating the comprehensive evaluation index based on the power grid supply margin, the task completion time of the candidate task execution sequence, and the remaining lifetime of the energy storage system include:

[0023] Based on the available capacity decay trajectory, obtain the predicted available capacity value for each future decision cycle; extract the peak load statistics for each future decision cycle from the historical load characteristic data, and calculate the minimum capacity requirement for peak shaving in each period; calculate the number of periods where the predicted available capacity value is lower than the minimum capacity requirement as the number of periods with capacity shortage risk.

[0024] The third weighting coefficient is determined based on the number of periods of insufficient capacity risk; the first weighting coefficient is determined based on the ratio of the power grid supply margin to the upper limit of the power grid capacity; and the second weighting coefficient is determined based on the ratio of the task completion time to the time constraint.

[0025] The comprehensive evaluation index is calculated by weighting and summing the power grid supply margin, the reciprocal of the task completion time, and the remaining lifespan of the energy storage system with the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively.

[0026] During execution, the actual load power is monitored. When the deviation between the actual load power and the expected value exceeds a set threshold, the step of re-executing the coupled decision mechanism to generate a correction scheme includes:

[0027] Calculate the deviation between the actual load power and the expected load power predicted based on the execution timing of the target task;

[0028] When the absolute value of the deviation exceeds the set threshold, the current time is recorded as the correction trigger time, and the time period from the correction trigger time to the end of the task execution is extracted as the time window to be corrected.

[0029] Obtain the remaining available capacity of the supercapacitor and the task allocation status of the executed task corresponding to the correction trigger time. Based on the deviation, correct the expected load power within the time window to be corrected to obtain the corrected load power curve. Use the corrected load power curve, the remaining tasks to be executed within the time window to be corrected, and the remaining available capacity of the supercapacitor as inputs to re-execute the coupling decision mechanism and generate the corrected task execution sequence and the corrected charging and discharging scheme.

[0030] The task allocation and charging / discharging actions executed before the correction trigger time remain unchanged. Starting from the correction trigger time, the remaining tasks are executed according to the correction task execution sequence and the correction charging / discharging scheme. The actual load power and its corresponding timestamp are added to the historical load characteristic data.

[0031] The steps of allocating tasks and controlling the supercapacitor energy storage system to perform charging and discharging according to the modified scheme include:

[0032] Generate supercapacitor charge and discharge control commands based on the charge and discharge timing sequence in the target charge and discharge scheme or the modified charge and discharge scheme.

[0033] The charging and discharging control command is sent to the power control unit of the supercapacitor energy storage system. During the charging and discharging process, the voltage and current parameters of the supercapacitor are collected in real time. The actual available capacity change is calculated based on the voltage and current parameters, and the actual available capacity change is updated to the current available capacity in the status information.

[0034] A second aspect of the present invention provides a peak load intelligent adjustment system for a supercapacitor energy storage system in a computing center, comprising:

[0035] The data acquisition module is used to acquire information on tasks to be executed from the computing center, status information of the supercapacitor energy storage system, and historical load characteristic data.

[0036] The decision planning module is used to construct a coupled decision-making mechanism between the task scheduling layer and the energy storage scheduling layer. The task scheduling layer generates multiple candidate task execution sequences based on the information of the tasks to be executed. The energy storage scheduling layer decomposes the peak period in the corresponding load power curve into multiple charge / discharge time windows for each candidate task execution sequence. Based on the available capacity in the status information and the historical load characteristic data, it allocates charge / discharge power and timing to each charge / discharge time window and calculates the remaining peak power after charge / discharge. The task scheduling layer selects the target task execution sequence and the corresponding target charge / discharge scheme based on the remaining peak power.

[0037] The deviation monitoring module is used to monitor the actual load power during execution. When the deviation between the actual load power and the expected value exceeds a set threshold, the coupled decision-making mechanism is re-executed to generate a correction scheme, and the actual load power is added to the historical load characteristic data.

[0038] The execution control module is used to allocate tasks according to the modified scheme and control the supercapacitor energy storage system to perform charging and discharging.

[0039] A third aspect of the present invention provides an electronic device, comprising:

[0040] processor;

[0041] Memory used to store processor-executable instructions;

[0042] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0043] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0044] This invention achieves collaborative optimization of computing tasks and energy storage resources by constructing a coupled decision-making mechanism between the task scheduling layer and the energy storage scheduling layer. This overcomes the limitations of traditional methods where task scheduling and energy management are separated, enabling more efficient smoothing of peak loads. By decomposing peak periods and dividing charging and discharging time windows based on the load power curves of each candidate task execution sequence, the charging and discharging strategy of the supercapacitor energy storage system can more accurately match load characteristics, fully leveraging the rapid response advantage of supercapacitors. Attached Figure Description

[0045] Figure 1This is a flowchart illustrating the intelligent peak load adjustment method for a supercapacitor energy storage system in a computing center, according to an embodiment of the present invention.

[0046] Figure 2 A flowchart for selecting the execution timing and charging / discharging scheme for the target task. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes will not be repeated in some embodiments.

[0049] Figure 1 This is a flowchart illustrating the intelligent peak load adjustment method for a supercapacitor energy storage system in a computing center, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0050] Obtain information on tasks to be executed from the computing center, status information of the supercapacitor energy storage system, and historical load characteristic data;

[0051] A coupled decision-making mechanism is constructed between the task scheduling layer and the energy storage scheduling layer. The task scheduling layer generates multiple candidate task execution sequences based on the information of the tasks to be executed. The energy storage scheduling layer decomposes the peak period in the corresponding load power curve into multiple charge and discharge time windows for each candidate task execution sequence. Based on the available capacity in the status information and the historical load characteristic data, it allocates charge and discharge power and timing to each charge and discharge time window and calculates the remaining peak power after charge and discharge. The task scheduling layer selects the target task execution sequence and the corresponding target charge and discharge scheme based on the remaining peak power.

[0052] During execution, the actual load power is monitored. When the deviation between the actual load power and the expected value exceeds a set threshold, the coupled decision-making mechanism is re-executed to generate a correction scheme, and the actual load power is added to the historical load characteristic data. Tasks are allocated according to the correction scheme, and the supercapacitor energy storage system is controlled to perform charging and discharging.

[0053] In one optional implementation, the step of the task scheduling layer generating multiple candidate task execution sequences includes: determining the schedulable time range of each task based on the resource requirement characteristics and time constraints of each task in the task information to be executed; generating multiple candidate task execution sequences by adjusting the task execution order and start time within the schedulable time range; calculating the corresponding load power curve for each candidate task execution sequence based on the resource requirement characteristics, wherein the load power curve represents the power demand corresponding to the total resource occupancy of all tasks at each time point; and transmitting the load power curve to the energy storage scheduling layer for the energy storage scheduling layer to perform charging and discharging power allocation and peak period decomposition.

[0054] For example, during task scheduling in a computing center, the schedulable time range for each task is determined based on its resource requirements and time constraints as described in the task information. Resource requirements include GPU core utilization, video memory requirements, CPU utilization, memory requirements, and storage bandwidth usage, which directly translate into corresponding power requirements. Deep learning training tasks typically require high GPU utilization, with power requirements reaching 300 to 500 watts per card; data preprocessing tasks primarily consume CPU and memory resources, with power requirements of approximately 150 to 250 watts per server. Time constraints include the earliest start time, latest finish time, and task duration; these constraints are usually specified by the user when submitting the task or automatically set based on business priorities.

[0055] Obtain the earliest start time, latest finish time, and task execution duration for each task. For task i, its schedulable time range is calculated as the interval from the earliest start time to the latest finish time minus the task duration, representing the time range within which the task can start after the earliest start time and is guaranteed to be completed before the latest finish time. For example, if the earliest start time for an AI model training task is 10:00 AM, the latest finish time is 6:00 PM, and the task duration is 5 hours, then the schedulable time range is from 10:00 AM to 1:00 PM. If the latest finish time minus the task duration is earlier than the earliest start time, then the schedulable time range for that task is empty, indicating that the task cannot be completed under the current constraints, and the constraints need to be adjusted or the task abandoned.

[0056] After determining the schedulable time range, multiple candidate task execution sequences are generated by adjusting the task execution order and start time. When sorting based on priority rules, tasks are prioritized according to their urgency, business importance, or deadline urgency. An earliest deadline-first strategy can be used to schedule the tasks with the most urgent deadlines first; a shortest processing time-first strategy can be used to prioritize tasks with shorter durations to quickly release resources; or a weighted priority strategy can be used to comprehensively consider task importance scores and time urgency. For example, inference service tasks that need to be completed within a specific time window are assigned higher priority to ensure their priority execution.

[0057] When using the time window method, the scheduling cycle is divided into multiple time windows, and tasks with compatible resource requirements are scheduled within each window. Different task allocation schemes can be generated by sliding the start and end times of the time windows or adjusting the window size. For each time window, a greedy strategy is used to select the most suitable task for execution, prioritizing tasks whose power requirements match the available power supply capacity. For example, during low-load periods at night, multiple large-scale training tasks can be scheduled to run in parallel; during high-load periods during the day, lightweight inference tasks are prioritized.

[0058] When using heuristic search methods, a greedy strategy combined with random perturbation is employed to generate multiple candidate schemes. First, tasks are sorted from highest to lowest priority, with high-priority tasks scheduled earlier within their schedulable timeframes to form a baseline scheduling scheme. Then, the baseline scheme is randomly adjusted by selecting several pairs of tasks and swapping their execution times while satisfying time constraints, generating variant schemes. This random adjustment is repeated several times to obtain multiple different candidate task execution sequences. For example, for 10 tasks to be scheduled, the baseline scheme is arranged in priority order; three pairs of tasks are randomly selected and swapped to generate variant scheme 1; another three pairs are randomly selected and swapped to generate variant scheme 2; and so on, generating 5 to 10 candidate schemes. For each candidate scheme, corresponding peak load power, average task completion time, and other metrics are calculated for subsequent decision-making.

[0059] When considering parallel execution under resource constraints, the system determines which tasks can be executed in parallel based on GPU slot dependencies, network bandwidth limitations, and the number of available servers. By splitting or merging tasks, the degree of parallelism is adjusted to generate various resource utilization methods. For groups of tasks that can be parallelized, different GPU card allocation ratios are explored to further enrich the candidate solutions. For example, a large-scale training task can be split into four sub-tasks, each occupying four GPU cards for parallel execution, or it can be split into eight sub-tasks, each occupying eight GPU cards. Different splitting methods correspond to different power demand curves and completion times.

[0060] The load power curve represents the power demand corresponding to the total resource consumption of all tasks at each time point. The calculation process is as follows: For each task i, calculate its runtime power demand based on its resource demand characteristics. This can be obtained through historical running data statistics or the mapping relationship between resource demand and power. For example, when the GPU utilization reaches over 90%, the power demand of a single card is approximately 400 watts; when the CPU utilization reaches 80%, the power demand of a single processor is approximately 150 watts; memory access frequency, storage read / write operation frequency, etc., can all be converted into corresponding power demands. Based on the candidate execution sequence, determine which tasks are running at each time point t. For each time point t, accumulate the power demands of all currently executing tasks to obtain the total load power at that time. In specific calculations, simply add the power values ​​of all currently executing tasks corresponding to time t.

[0061] To improve computational accuracy, time can be discretized by sampling at sufficiently small time intervals to generate a discrete sequence of load power curves. The choice of time interval needs to balance computational complexity and accuracy requirements; for computing center scenarios, it is typically set to the level of 1 minute or 5 minutes. Too small a time interval will result in excessive computation, while too large a time interval will miss short-term power peaks.

[0062] When calculating load power, power fluctuations at task startup and termination must also be considered. For example, deep learning training tasks experience brief storage bandwidth spikes during the startup phase when loading model weights and datasets, with corresponding power spikes reaching 1.3 to 1.5 times that of steady-state operation; similar power spikes occur when saving checkpoints at the end of training. These dynamic characteristics are characterized using a task power curve template, making the calculated load power curve more accurate. The task power curve template is generated based on power monitoring data from similar historical tasks and includes typical power change patterns during startup, stable operation, and termination phases.

[0063] After the load power curve calculation is completed, the load power curve data is transmitted to the energy storage scheduling layer for charging and discharging power allocation and peak period decomposition. Based on the characteristics of the load power curve, the energy storage scheduling layer identifies power peak and valley periods and formulates charging and discharging strategies for the supercapacitor energy storage device to smooth the load curve and reduce power peaks.

[0064] This invention achieves coordinated optimization from task scheduling to energy storage scheduling, effectively utilizing energy storage resources to cope with power fluctuations caused by task scheduling, and provides a foundation for energy optimization management of computing centers.

[0065] In one optional implementation, the energy storage scheduling layer decomposes the peak period in the corresponding load power curve into multiple charge / discharge time windows for each candidate task execution sequence. Based on the available capacity in the status information and the historical load characteristic data, it allocates charge / discharge power and timing to each charge / discharge time window. The step of calculating the remaining peak power after charge / discharge includes: identifying the period in the load power curve where the power exceeds a reference power threshold as the peak period; dividing the peak period into multiple charge / discharge time windows according to time granularity; determining the pre-charging period of the supercapacitor before the peak period arrives based on the peak-valley period distribution pattern in the historical load characteristic data; and determining the pre-charging period of the supercapacitor based on the charging power upper limit in the status information and... The duration of the pre-charging period is used to calculate the energy that can be stored; based on the energy that can be stored and the available capacity, the excess power of the load power at each charging and discharging time window that exceeds the reference power threshold is calculated, and the excess power is used as the target discharge power; the charging sequence is determined based on the end time of the pre-charging period, and the discharging sequence is determined based on the start and end times of each charging and discharging time window within the peak period, forming a charging and discharging sequence; based on the target discharge power of each charging and discharging time window and the charging and discharging sequence, the load power reduction amount at the corresponding time after discharging in each charging and discharging time window is calculated; the power value at each time in the load power curve is subtracted from the corresponding load power reduction amount to obtain the remaining peak power after charging and discharging.

[0066] For example, during the operation of the energy storage scheduling layer, the peak period in the corresponding load power curve is decomposed into multiple charging and discharging time windows for each candidate task execution sequence. Based on the current status information and historical load characteristic data, charging and discharging power and timing are allocated to each charging and discharging time window, and then the remaining peak power after charging and discharging is calculated.

[0067] A baseline power threshold is set at 80% of the grid transformer capacity. When the power value in a certain period of the load power curve exceeds this threshold, that period is marked as a peak period. For example, if the computing center is configured with a 2500 kVA transformer, the corresponding baseline power threshold is set to 2000 kW. The identified peak periods are divided into multiple charging and discharging time windows according to time granularity. The time granularity is determined based on the minimum time unit of task scheduling, typically set to 5 minutes, 10 minutes, or 15 minutes, depending on the specific application scenario. For example, for a peak period lasting 60 minutes, if the time granularity is 15 minutes, it is divided into 4 charging and discharging time windows, corresponding to 4 consecutive time periods within the peak period.

[0068] Historical load characteristic data contains trends in load power changes over a past period. Analyzing this data can predict future peak and valley load power distribution. By statistically analyzing the average load power for each time period of the past 30 days, typical peak and valley distribution patterns can be identified. For example, the computing center experiences high load periods from 9:00 AM to 11:00 AM and 2:00 PM to 5:00 PM daily, and low load periods from 11:00 PM to 6:00 AM the following day. After determining the expected start time of the peak period, the low load period before the peak period is selected as the pre-charging period. Assuming the peak period is expected to start at 2:00 PM, then 12:00 PM to 2:00 PM can be selected as the pre-charging period.

[0069] The upper limit of supercapacitor charging power is limited by the rated power of the power module and the charging characteristics of the capacitor, and is typically set to 0.5 to 1 times the power corresponding to the rated capacity. Assuming the upper limit of the supercapacitor energy storage system's charging power is 200 kW and the pre-charging period is 2 hours, the energy that can be stored is calculated by multiplying the upper limit of the charging power by the duration, i.e., 200 kW multiplied by 2 hours equals 400 kWh. This value represents the maximum electrical energy that the supercapacitor can store during the pre-charging period.

[0070] Compare the available 400 kWh of stored energy with the current available capacity of the supercapacitor, and take the smaller value as the upper limit of the actual callable energy. For example, if the current available capacity is 350 kWh, then the actual callable energy is 350 kWh. For each charge / discharge time window, calculate the excess power of the load power exceeding the baseline power threshold at that moment. If the load power at a certain moment is 2150 kW and the baseline power threshold is 2000 kW, then the excess power is 150 kW. Sum the excess power of all time windows to obtain the total excess power demand. Allocate the actual callable energy to each time window according to the proportion of the excess power of each time window to the total excess power. For example, if the excess power of the first time window is 150 kW for 15 minutes, corresponding to 37.5 kWh, and the excess power of the second window is 120 kW for 15 minutes, corresponding to 30 kWh, then the discharge energy allocated to the first window is 350 kWh multiplied by 37.5 divided by the total demand energy. The target discharge power is calculated based on the discharge energy allocated to each time window and the window duration.

[0071] The charging sequence begins at the start of the pre-charging period and continues until the end of the pre-charging period; the discharging sequence is arranged according to the time order of each charging and discharging time window, starting from the start of the first window and ending at the end of the last window, forming a complete charging and discharging sequence. For example, the charging sequence is from 12:00 to 14:00, and the discharging sequence is from 14:00 to 15:00, corresponding to four 15-minute time windows.

[0072] For each discharge time window, the load power reduction is equal to the actual discharge power performed in that window. For example, if the target discharge power allocated for a certain time window is 150 kW, then the load power reduction after the discharge in that window is also 150 kW. Subtracting the corresponding load power reduction from the power value at each moment in the load power curve yields the remaining peak power after charging and discharging. For each moment in the load power curve, if the original power value is 2150 kW and the load power reduction is 150 kW, then the remaining peak power after charging and discharging is calculated as 2150 kW minus 150 kW, which equals 2000 kW.

[0073] In practical applications, there are situations where the available capacity of a supercapacitor is insufficient to cover all peak periods. In such cases, a priority strategy is adopted, allocating discharge power first to the time windows with larger excess power. For example, if the excess power of two time windows is 180 kW and 100 kW respectively, but the available capacity can only support 60 kWh of discharge, then the discharge power is allocated first to reduce the higher peak value.

[0074] This invention effectively utilizes the energy storage capacity of supercapacitors to smooth out peak values ​​in the power curve of a computing center load, reducing peak power consumption, improving energy efficiency, and minimizing additional capacity costs incurred due to peak power exceeding transformer capacity limits. Simultaneously, the predictive charging and discharging strategy based on historical load characteristic data can prepare sufficient reserve energy in advance for peak periods, ensuring effective peak reduction.

[0075] In an optional implementation, the step of the task scheduling layer selecting the target task execution sequence and the corresponding target charging and discharging scheme based on the remaining peak power includes: comparing the remaining peak power with the upper limit of the grid capacity, and filtering out candidate task execution sequences whose remaining peak power does not exceed the upper limit of the grid capacity as a feasible sequence set; for each candidate task execution sequence in the feasible sequence set, calculating the grid power supply margin based on the difference between the remaining peak power and the upper limit of the grid capacity; calculating the capacity decay based on the charging and discharging depth, charging and discharging power in the target charging and discharging scheme and the current available capacity in the status information; and based on the capacity decay, The task arrival frequency distribution and peak load periodicity characteristics in the historical load characteristic data are used to deduce the available capacity decay trajectory of the supercapacitor in multiple decision cycles in the future. Based on the available capacity decay trajectory, the number of decision cycles required for the capacity to drop to the capacity failure threshold is determined. The time length from the current moment to the moment corresponding to the number of decision cycles is calculated as the remaining lifespan of the energy storage system. Based on the grid power supply margin, the task completion time of the candidate task execution sequence, and the remaining lifespan of the energy storage system, a comprehensive evaluation index is calculated. The candidate task execution sequence with the best comprehensive evaluation index is selected as the target task execution sequence, and the corresponding charging and discharging power and timing are extracted as the target charging and discharging scheme.

[0076] Combination Figure 2 The flowchart illustrating the target task execution timing and charging / discharging scheme selection is provided below. For example, the remaining peak power is compared with the grid capacity limit, and candidate task execution timings whose remaining peak power does not exceed the grid capacity limit are selected as the feasible timing set. For instance, if the current remaining peak power is 1800 kW, and the grid transformer capacity limit is 2500 kVA, corresponding to approximately 2000 kW of active power, then this task execution timing satisfies the grid capacity constraint and is included in the feasible timing set.

[0077] For each candidate task in the feasible time series, the grid supply margin is calculated based on the difference between the remaining peak power and the grid capacity limit. The grid supply margin is expressed as the grid capacity limit minus the remaining peak power. For example, a grid supply margin of 200 kW is equivalent to 2000 kW minus 1800 kW. A larger supply margin indicates a more sufficient grid carrying capacity, higher system stability, and a lower risk of overload tripping.

[0078] Capacity decay is positively correlated with the depth of charge / discharge and the charge / discharge power. The depth of charge / discharge is defined as the percentage of energy change during a single charge / discharge cycle relative to the rated capacity of the supercapacitor. For example, if a supercapacitor has a rated capacity of 500 kWh and releases 150 kWh of energy during a discharge, the depth of charge / discharge is 150 divided by 500, which equals 30%. The depth of charge / discharge coefficient is determined using a piecewise function: 0.001 for depths not exceeding 30%, 0.002 for depths between 30% and 60%, and 0.003 for depths exceeding 60%. The power stress coefficient is determined based on the ratio of charge / discharge power to rated power, with a linear mapping from 0.001 to 0.005. For example, if the charge / discharge power is 200 kW and the rated power is 400 kW, the ratio is 0.5, corresponding to a power stress coefficient of approximately 0.003. Capacity decay is calculated by multiplying the current available capacity by the depth of charge / discharge factor and then by the power stress factor. For example, when the current available capacity of a supercapacitor is 500 kW, a depth of charge / discharge of 30% corresponds to a depth factor of 0.001, and a charge / discharge power of 200 kW corresponds to a power stress factor of 0.003, the calculated capacity decay is 500 multiplied by 0.001 multiplied by 0.003, which equals 0.0015 kWh.

[0079] Based on the capacity attenuation calculated in the preceding steps, and combined with the task arrival frequency distribution and peak load periodicity characteristics in historical load characteristic data, the available capacity attenuation trajectory of the supercapacitor in multiple future decision-making cycles is deduced. Specifically, the previously calculated capacity attenuation is used as the baseline attenuation, with 0.0015 kWh as the baseline attenuation. An attenuation pattern lookup table is established based on historical load characteristic data. The timing and duration of peak load occurrences over the past 30 days are statistically analyzed to identify peak load periodicity characteristics. For example, 14:00 to 17:00 daily is identified as a high-frequency peak period and an accelerated attenuation interval. The task arrival frequency distribution is extracted from the historical load characteristic data, and the number of tasks arriving per hour over the past 30 days is statistically analyzed, calculating an average task arrival frequency of 5 tasks per hour. For each future decision-making cycle, the attenuation amount for that cycle is determined by combining the above characteristics: First, it is determined whether the time of the decision-making cycle falls within the accelerated attenuation interval. If the decision period falls between 14:00 and 17:00, the attenuation rate for that period is 1.5 times the baseline attenuation, i.e., 0.00225 kWh; if it does not fall within this range, the attenuation rate is the baseline attenuation of 0.0015 kWh. Secondly, the attenuation is adjusted based on the task arrival frequency during that period. If the task arrival frequency during that period is higher than the average, the attenuation is increased proportionally to the frequency. For example, if the task arrival frequency during a certain period is 1.2 times the average (i.e., 6 tasks per hour), the attenuation for that period is multiplied by 1.2. Starting from the current available capacity of 500 kWh, the attenuation is cumulatively reduced according to each decision period to obtain a sequence of predicted capacity values ​​for each period. For example, if the first cycle is not in the acceleration range and the task frequency is the average, the capacity becomes 500 - 0.0015 = 499.9985 kWh; if the second cycle is in the acceleration range and the task frequency is 1.2 times the average, the capacity becomes 499.9985 - 0.00225 × 1.2 = 499.9958 kWh. This pattern continues, forming a usable capacity decay trajectory for multiple future decision cycles. When the capacity value of a certain cycle in the sequence first falls below the capacity failure threshold, the corresponding time is the end of the remaining lifetime. The capacity failure threshold is set to 80% of the initial capacity, i.e., 400 kWh. Using the aforementioned decay trajectory, the number of decision cycles required for the capacity to decrease to 400 kWh is determined and converted into actual time. The time length from the current time to the end of the remaining lifetime is calculated as the remaining lifetime of the energy storage system. For example, if the decision cycle is 1 hour, the average decay is 0.002 kWh per cycle, and it is estimated that 50,000 decision cycles are needed to reach the failure threshold, then the time from the current moment to the failure moment, i.e. the remaining lifespan of the energy storage system, is about 2083 days.

[0080] A lookup table for attenuation patterns is established based on historical load characteristic data. The times and durations of peak loads over the past 30 days are statistically analyzed, identifying the period from 14:00 to 17:00 daily as the high-frequency peak period and the accelerated attenuation interval. For each future decision cycle, it is determined whether it falls within the accelerated attenuation interval. If the decision cycle falls within the 14:00 to 17:00 range, the attenuation rate for that cycle is 1.5 times the baseline attenuation of 0.0015 kWh, i.e., 0.00225 kWh; if it does not fall within this interval, the attenuation rate is the baseline attenuation of 0.0015 kWh. Starting from the current available capacity of 500 kWh, the attenuation is cumulatively reduced for each decision cycle to obtain a sequence of predicted capacity values ​​for each cycle. For example, if the first cycle is not within the accelerated attenuation interval, the capacity becomes 499.9985 kWh; if the second cycle is within the accelerated attenuation interval, the capacity becomes 499.996 kWh. The frequency of task arrivals affects the charging and discharging frequency. Historical data shows an average of 5 tasks arriving per hour. The number of charging and discharging cycles during high-frequency task arrival periods decreases linearly with this frequency. When the capacity value in a certain period of the sequence first falls below the capacity failure threshold, the corresponding time is the end of the remaining lifetime. The capacity failure threshold is set at 80% of the initial capacity, i.e., 400 kWh. Using the aforementioned decay trajectory, the number of decision cycles required for the capacity to drop to 400 kWh is determined and converted into actual time. For example, if the decision cycle is 1 hour, with an average decay of 0.002 kWh per cycle, it is estimated that 50,000 decision cycles are needed to reach the failure threshold, resulting in a remaining lifetime of approximately 2083 days.

[0081] The comprehensive evaluation index is calculated as follows: First weighting coefficient × Grid power supply margin + Second weighting coefficient × Reciprocal of task completion time + Third weighting coefficient × Remaining lifespan of the energy storage system. Each weighting coefficient reflects the importance of different factors, and the sum of the three weighting coefficients equals 1. For example, if the first weighting coefficient is set to 0.3, the second weighting coefficient to 0.4, and the third weighting coefficient to 0.3, and a candidate scheme has a grid power supply margin of 200 kW, a task completion time of 3 hours, and a remaining lifespan of 2083 days, the calculated evaluation index is 0.3 × 200 + 0.4 × (1 / 3) + 0.3 × 2083 = 685.03.

[0082] By comparing the evaluation metrics of all candidate solutions, the solution with the highest metric value is selected. In practical applications, when multiple deep learning training tasks arrive simultaneously in a computing center's GPU cluster, multiple servers need to be started to process the tasks concurrently, resulting in a significant increase in power demand. The above method can achieve efficient energy management for the computing center while meeting task time constraints and considering grid stability and energy storage system lifespan.

[0083] In an optional implementation, the step of calculating a comprehensive evaluation index based on the grid power supply margin, the task completion time of the candidate task execution sequence, and the remaining lifetime of the energy storage system includes: obtaining the predicted available capacity value at the corresponding time of each future decision cycle according to the available capacity decay trajectory; extracting the peak load statistics of the corresponding time period of each future decision cycle from the historical load characteristic data, and calculating the minimum capacity requirement required for peak shaving in each time period; calculating the number of time periods where the predicted available capacity value is lower than the minimum capacity requirement as the number of time periods with capacity insufficiency risk; determining a third weighting coefficient based on the number of time periods with capacity insufficiency risk, determining a first weighting coefficient based on the ratio of the grid power supply margin to the grid capacity limit, and determining a second weighting coefficient based on the ratio of the task completion time to the time constraint condition; and weighting and summing the grid power supply margin, the reciprocal of the task completion time, and the remaining lifetime of the energy storage system with the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively, to calculate the comprehensive evaluation index.

[0084] For example, the available capacity decay trajectory has been calculated in the previous steps by cumulatively reducing the decay amount cycle by cycle. This trajectory contains a sequence of capacity predictions for several future decision cycles. For instance, the next 720 decision cycles, starting from the current moment, correspond to the next 30 days. Each cycle corresponds to a duration of 1 hour. The predicted capacity for the first cycle is 499.998 kWh, the predicted capacity for the 100th cycle is 499.8 kWh, and so on, forming a complete sequence. Peak load statistics for each time period corresponding to the future decision cycles are extracted from the historical load characteristic data. The historical load characteristic data records the actual load power at each moment in the past period. By statistically analyzing the load power distribution of the same time period over multiple historical days, the mean and standard deviation of the peak load for that time period are calculated. For example, by statistically analyzing the peak load from 14:00 to 15:00 each day over the past 30 days, the mean peak load for that time period is calculated to be 2100 kW, and the standard deviation is 150 kW. The future decision cycles are mapped to the corresponding time periods, and the peak load statistics for that time period are extracted as the expected load.

[0085] Calculate the minimum capacity requirement for peak shaving in each time period. The capacity required for peak shaving equals the excess power of the peak load exceeding the baseline power threshold for that time period multiplied by the duration of the time period. For example, if the peak load for a time period is 2100 kW, the baseline power threshold is 2000 kW, the excess power is 100 kW, and the time period lasts for 1 hour, then the minimum capacity required for peak shaving is 100 kW × 1 hour = 100 kWh. Calculate the capacity required for peak shaving for each future decision cycle. Calculate the number of time periods where the predicted available capacity is lower than the minimum capacity requirement as the number of capacity insufficiency risk periods. Iterate through each future decision cycle, comparing the predicted available capacity for that cycle with the minimum capacity required for peak shaving. If the predicted available capacity is lower than the minimum capacity requirement, then that time period is marked as a capacity insufficiency risk period. Count the number of all marked time periods to obtain the number of capacity insufficiency risk periods. For example, if in the next 720 decision cycles, there are 15 cycles where the predicted available capacity is lower than the peak shaving capacity required for the corresponding time period, then the number of capacity insufficiency risk periods is 15.

[0086] The third weighting coefficient is determined based on the number of periods of capacity insufficiency risk. This coefficient reflects the importance of the remaining lifespan of the energy storage system to the overall evaluation. A higher number of periods of capacity insufficiency risk indicates a greater impact of system degradation on future peak-shaving capabilities, thus necessitating a higher third weighting coefficient to emphasize lifespan factors. A segmented mapping rule is adopted: when the number of periods of capacity insufficiency risk does not exceed 2% of the total number of cycles, the third weighting coefficient is 0.2; when the number of risk periods is between 2% and 5%, the coefficient is 0.3; and when the number of risk periods exceeds 5%, the coefficient is 0.4. For example, if the total number of cycles is 720, and the number of risk periods is 15 (2.08%), falling within the 2% to 5% range, then the third weighting coefficient is 0.3.

[0087] The first weighting coefficient is determined based on the ratio of the grid supply margin to the grid capacity limit. The grid supply margin reflects the grid's carrying capacity; a smaller margin indicates the grid is approaching its capacity limit, increasing the risk of overload. Therefore, the first weighting coefficient should be increased to emphasize the supply margin factor. The margin ratio is calculated by dividing the grid supply margin by the grid capacity limit. For example, if the grid supply margin is 200 kW and the grid capacity limit is 2000 kW, the margin ratio is 0.1. A linear mapping rule is used: when the margin ratio is greater than 0.2, the first weighting coefficient is 0.2; when the margin ratio is between 0.1 and 0.2, the first weighting coefficient is linearly interpolated to a range of 0.2 to 0.4; when the margin ratio is less than 0.1, the first weighting coefficient is 0.4. In the case of a margin ratio of 0.1, the first weighting coefficient is 0.4.

[0088] Task completion time reflects task execution efficiency. The closer the completion time is to the upper limit of the time constraint, the smaller the time margin, and the higher the second weighting coefficient should be to emphasize task completion timeliness. The time ratio is calculated by dividing the task completion time by the latest completion time in the time constraint. For example, a candidate solution has a task completion time of 3 hours, and the time constraint requires a latest completion time of 4 hours, resulting in a time ratio of 0.75. A linear mapping rule is used: when the time ratio is less than 0.7, the second weighting coefficient is 0.2; when the time ratio is between 0.7 and 0.9, the second weighting coefficient is linearly interpolated to a range of 0.2 to 0.4; when the time ratio is greater than 0.9, the second weighting coefficient is 0.4. With a time ratio of 0.75, the second weighting coefficient is approximately 0.3.

[0089] Ensure the sum of the three weighting coefficients equals 1. If the sum of the coefficients determined by the aforementioned rules does not equal 1, normalize proportionally. For example, if the first weighting coefficient is 0.4, the second weighting coefficient is 0.3, and the third weighting coefficient is 0.3, the sum of the coefficients is 1, and no adjustment is needed. Calculate the comprehensive evaluation index by weighting and summing the grid power supply margin, task completion time, and remaining lifespan of the energy storage system with the first, second, and third weighting coefficients, respectively. The comprehensive evaluation index is calculated as follows: First weighting coefficient × Grid power supply margin + Second weighting coefficient × (Normalization factor / (Reciprocal of task completion time)) + Third weighting coefficient × Remaining lifespan of the energy storage system. The normalization factor is used to balance the index values ​​of different dimensions, making the contributions of each index to the comprehensive evaluation comparable. The reciprocal of the task completion time is multiplied by 100 to obtain the normalized value. The remaining lifespan of the energy storage system, in days, is directly used in the calculation. For example, if the grid power supply margin is 200 kW, the task completion time is 3 hours, the remaining lifespan of the energy storage system is 2083 days, the first weighting coefficient is 0.4, the second weighting coefficient is 0.3, and the third weighting coefficient is 0.3, the calculated comprehensive evaluation index is 0.4 × 200 + 0.3 × (100 / 3) + 0.3 × 2083 = 714.9. This comprehensive evaluation index is used to rank different candidate schemes, and the scheme with the optimal index value is selected as the final execution scheme.

[0090] This invention quantifies the sustainability of future peak shaving capacity by measuring the number of periods of insufficient capacity risk. By dynamically adjusting the weighting coefficients through margin ratio and time ratio, it achieves an adaptive balance between grid security, task timeliness, and energy storage lifespan, thereby improving the scientific nature and robustness of energy dispatch decisions in computing centers.

[0091] In one optional implementation, the step of monitoring the actual load power during execution and re-executing the coupled decision mechanism to generate a correction scheme when the deviation between the actual load power and the expected value exceeds a set threshold includes: calculating the deviation between the actual load power and the expected load power predicted based on the execution sequence of the target task; when the absolute value of the deviation exceeds the set threshold, recording the current time as the correction trigger time and extracting the time period from the correction trigger time to the end of task execution as the time window to be corrected; obtaining the remaining available capacity of the supercapacitor and the assigned status of the executed tasks corresponding to the correction trigger time, and correcting the expected load power within the time window to be corrected based on the deviation to obtain the corrected load power curve; using the corrected load power curve, the remaining tasks to be executed within the time window to be corrected, and the remaining available capacity of the supercapacitor as input, re-executing the coupled decision mechanism to generate the corrected task execution sequence and the corrected charging and discharging scheme; keeping the task assignments and charging and discharging actions executed before the correction trigger time unchanged, and executing the remaining tasks from the correction trigger time according to the corrected task execution sequence and the corrected charging and discharging scheme; and supplementing the actual load power and its corresponding timestamp into the historical load feature data.

[0092] For example, during actual task execution, the actual load power may deviate from the expected value due to factors such as sudden task arrivals at the computing center, server failures, or fluctuations in GPU utilization. When this deviation is too large, real-time correction is required to ensure the system can operate efficiently and complete all tasks. During system operation, the actual load power of the system is monitored in real time by power sensors deployed in the power distribution cabinet, with a sampling frequency set to once per second. Simultaneously, the expected load power at the current moment is obtained from the target task execution timeline. For each monitoring moment, the actual power and expected power are obtained, and then the deviation between them is calculated. The deviation is equal to the actual power minus the expected power.

[0093] A power deviation threshold is pre-set. When the absolute value of the deviation exceeds this threshold, it indicates that the actual load differs too much from the expected load, requiring a correction mechanism to be triggered. In practical applications, this threshold is set based on the transformer capacity of the computing center and the importance of the task; for example, it might be set to 10% of the system's rated power, or 200 kilowatts. When the absolute value of the detected deviation exceeds 200 kilowatts, the current time is recorded as the correction trigger time, and the time period from that time to the task's scheduled execution end time is extracted as the correction window. For example, if the current time is 15:30 and the task's scheduled execution end time is 17:00, then the correction window is 15:30 to 17:00, a total of 1.5 hours. The task execution plan within this time window will be re-evaluated and adjusted.

[0094] After determining the correction trigger time, it is necessary to obtain the current status information, including the remaining available capacity of the supercapacitor and the status of the executed tasks. The remaining capacity of the supercapacitor is calculated in real time by the capacitor voltage sensor and the capacity management module; for example, the current remaining capacity is 320 kWh. The status of the executed tasks includes information such as which tasks have been completed, which tasks are currently being executed, their running time, and remaining time. For example, task A has been completed, task B is being executed and has been running for 20 minutes with 10 minutes remaining, and tasks C and D have not yet started.

[0095] Revised version:

[0096] The expected load power within the correction time window is corrected based on the deviation, resulting in the corrected load power curve. The correction method uses an exponential decay model: deviation impact = current deviation value × e^(-decay coefficient × time difference). The decay coefficient is determined based on historical deviation duration statistics. The average duration is calculated by analyzing the duration distribution of deviation events over the past 30 days. If the average duration is 30 minutes, the decay coefficient is set to ln2 / 30 minutes ≈ 0.023 / minute, corresponding to a half-life of 30 minutes. For each moment within the correction time window, the time difference between that moment and the correction trigger moment is calculated and substituted into the decay formula to calculate the deviation impact. For example, if the deviation at the correction trigger moment is +250 kW, the deviation impact 10 minutes later = 250 × e^(-0.023 × 10) ≈ 200 kW, and the deviation impact 20 minutes later = 250 × e^(-0.023 × 20) ≈ 160 kW. The expected power at each moment is then added to the corresponding deviation impact to obtain the corrected load power curve.

[0097] The revised load power curve, the remaining tasks to be executed within the time window to be revised, and the remaining available capacity of the supercapacitor are used as inputs to re-execute the coupled decision-making mechanism. This decision-making mechanism comprehensively considers factors such as the time constraints, power requirements, and available capacity of the remaining tasks to generate revised task execution timing and revised charging and discharging schemes. For example, if task C was originally scheduled to start at 15:45 and task D at 16:15, the revised mechanism, considering the actual high load, will postpone the start of task C to 16:00 and task D to 16:30. At the same time, the discharge power of the supercapacitor during the period from 15:30 to 16:00 will be increased from 150 kW to 200 kW to offset the additional load peak.

[0098] When implementing the revised plan, the task assignments and charging / discharging actions already executed before the revision trigger time remain unchanged. From the revision trigger time, the remaining tasks are executed according to the revised task execution sequence and the revised charging / discharging plan. This ensures the continuity of system operation and avoids unnecessary interruptions and backtracking. Completed task A remains unchanged, currently executing task B continues until completion, and from 15:30, tasks C and D are executed according to the revised plan. Simultaneously, the supercapacitor adjusts its discharge power according to the revised charging / discharging plan.

[0099] The actual load power and its corresponding timestamp are added to the historical load feature data. This data will be used for future power prediction and decision optimization, improving the prediction accuracy for similar task scenarios. These historical data are analyzed periodically to identify power variation patterns and optimize the power prediction algorithm. For example, if a deep learning training task is found to have significant power fluctuations during the data loading phase, a power margin for that phase can be increased in the power prediction.

[0100] Through the above correction mechanism, dynamic adjustments can be made during task execution to adapt to actual load changes, ensuring that all computing tasks are completed efficiently under energy constraints.

[0101] In one optional implementation, the steps of allocating tasks and controlling the supercapacitor energy storage system to perform charging and discharging according to the modified scheme include: generating a supercapacitor charging and discharging control command according to the charging and discharging sequence in the target charging and discharging scheme or the modified charging and discharging scheme; sending the charging and discharging control command to the power control unit of the supercapacitor energy storage system; collecting the voltage and current parameters of the supercapacitor in real time during the charging and discharging process; calculating the actual available capacity change based on the voltage and current parameters; and updating the actual available capacity change to the current available capacity in the status information.

[0102] For example, charging and discharging control commands for the supercapacitor are generated based on the target charging and discharging scheme or the charging and discharging sequence in the modified charging and discharging scheme. The charging and discharging sequence includes charging periods, discharging periods, and the corresponding power setpoints for each period. The control commands are encapsulated using a standard communication protocol format, including a command type field, a timestamp field, a power setpoint field, and a duration field. For example, the command type field for a charging command is set to 0x01, the discharging command to 0x02, and the standby command to 0x00. The timestamp field records the absolute moment the command takes effect, using a Unix timestamp format accurate to the second. The power setpoint field records the target value of the charging and discharging power, in watts, represented by a 32-bit integer, ranging from -200 kW to +200 kW, with negative values ​​indicating charging and positive values ​​indicating discharging. The duration field records the duration for which the power setpoint is maintained, in seconds, represented by a 16-bit integer. For a charging period from 12:00 to 14:00 with a charging power of 150 kW, the generated control command is command type 0x01, timestamp corresponding to the Unix timestamp of 12:00, power setting value of -150000 watts, and duration of 7200 seconds. For a discharging period from 14:00 to 15:00, divided into four 15-minute windows with different discharging powers, four consecutive discharging commands are generated.

[0103] Charging and discharging control commands are sent to the power control unit of the supercapacitor energy storage system via industrial Ethernet or CAN bus. The power control unit is responsible for receiving commands and driving the bidirectional DC-DC converter to perform charging and discharging operations. The communication interface uses the TCP protocol to establish a reliable connection, and a timeout retransmission mechanism is set to ensure command delivery. Before sending a command, a CRC checksum is calculated and appended to the end of the command. The receiving end verifies the checksum to confirm data integrity. After receiving the command, the power control unit returns an acknowledgment message. If no acknowledgment is received within 500 milliseconds, the command is retransmitted, with a maximum of 3 retries. Command transmission uses a time synchronization mechanism. The control system and the power control unit maintain clock synchronization error of less than 10 milliseconds via the NTP protocol to ensure that charging and discharging actions are triggered at precise times.

[0104] The voltage and current parameters of the supercapacitor are acquired in real time during charging and discharging. The voltage of a single supercapacitor cell is measured using a high-precision voltage sensor with an accuracy of 0.1% full-scale error, a sampling resolution of 1 mV, and a sampling frequency of 10 times per second. The current parameter is measured using a Hall effect current sensor with an accuracy of 0.5% full-scale error, a sampling resolution of 0.1 amperes, and a sampling frequency of 10 times per second. The acquired voltage and current data are converted into digital signals by an analog-to-digital converter, and after high-frequency noise is removed by a digital filter, they are stored in a data buffer. The data buffer uses a circular queue structure to store the sampling data from the most recent hour for capacity calculation and anomaly detection.

[0105] The actual usable capacity change is calculated based on voltage and current parameters. The stored energy of a supercapacitor is proportional to the square of the voltage; the capacity change is calculated using the energy integration method. For each sampling moment, the instantaneous power at that moment is calculated as voltage × current. The average instantaneous power between two adjacent sampling moments is taken and multiplied by the sampling time interval of 0.1 seconds to obtain the energy change over that time interval. The energy changes over all sampling time intervals are summed to obtain the total energy change from the start of charging / discharging to the current moment, which is the actual usable capacity change. For example, in a discharge process, the initial voltage is 150 volts and the initial current is 1000 amperes. After 10 seconds, the voltage drops to 145 volts and the current is 950 amperes. During this time period, the average voltage is (150 + 145) / 2 = 147.5 volts, the average current is (1000 + 950) / 2 = 975 amperes, and the average power is 147.5 × 975 = 143.8 kilowatts. The energy change is 143.8 kilowatts × 10 seconds / 3600 = 0.4 kilowatt-hours. Considering the energy loss due to the internal resistance of the supercapacitor, the actual energy efficiency is approximately 95%. Therefore, the actual usable capacity change is 0.4 kilowatt-hours × 0.95 = 0.38 kilowatt-hours.

[0106] The actual change in available capacity is updated to the current available capacity in the status information. Status information is stored in the system database and includes fields such as current available capacity, charge / discharge cycle count, and cumulative charge / discharge energy. The update operation uses atomic transactions to ensure data consistency. First, the old value of the current available capacity is read, then the capacity change due to discharge is subtracted, or the capacity change due to charging is added, to obtain the new current available capacity value. This new value is then written back to the database and the transaction is committed. For example, if the current available capacity is 320 kWh before charging / discharging begins, and the actual capacity change during discharge is 0.38 kWh, the updated current available capacity is 319.62 kWh. Simultaneously, the charge / discharge cycle count is incremented by 1, and the cumulative charge / discharge energy increases by 0.38 kWh. After the status information is updated, a capacity degradation assessment process is triggered. If the current available capacity is found to have decreased by more than 20% compared to the rated capacity, a capacity warning is issued to maintenance personnel to check the health status of the supercapacitor.

[0107] By collecting voltage and current parameters in real time and accurately calculating capacity changes, this invention enables accurate control of the supercapacitor energy storage system's status, ensuring closed-loop control of energy dispatching in the computing center.

[0108] A second aspect of the present invention provides a peak load intelligent adjustment system for a supercapacitor energy storage system in a computing center, comprising:

[0109] The data acquisition module is used to acquire information on tasks to be executed from the computing center, status information of the supercapacitor energy storage system, and historical load characteristic data.

[0110] The decision planning module is used to construct a coupled decision-making mechanism between the task scheduling layer and the energy storage scheduling layer. The task scheduling layer generates multiple candidate task execution sequences based on the information of the tasks to be executed. The energy storage scheduling layer decomposes the peak period in the corresponding load power curve into multiple charge / discharge time windows for each candidate task execution sequence. Based on the available capacity in the status information and the historical load characteristic data, it allocates charge / discharge power and timing to each charge / discharge time window and calculates the remaining peak power after charge / discharge. The task scheduling layer selects the target task execution sequence and the corresponding target charge / discharge scheme based on the remaining peak power.

[0111] The deviation monitoring module is used to monitor the actual load power during execution. When the deviation between the actual load power and the expected value exceeds a set threshold, the coupled decision-making mechanism is re-executed to generate a correction scheme, and the actual load power is added to the historical load characteristic data.

[0112] The execution control module is used to allocate tasks according to the modified scheme and control the supercapacitor energy storage system to perform charging and discharging.

[0113] A third aspect of the present invention provides an electronic device, comprising:

[0114] processor;

[0115] Memory used to store processor-executable instructions;

[0116] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0117] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0118] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent peak load adjustment of a supercapacitor energy storage system in a computing center, characterized in that, include: Obtain information on tasks to be executed from the computing center, status information of the supercapacitor energy storage system, and historical load characteristic data; A coupled decision-making mechanism is constructed between the task scheduling layer and the energy storage scheduling layer, wherein the task scheduling layer generates multiple candidate task execution sequences based on the information of the tasks to be executed; The energy storage scheduling layer decomposes the peak period in the corresponding load power curve into multiple charging and discharging time windows for each candidate task execution sequence. Based on the available capacity in the status information and the historical load characteristic data, it allocates charging and discharging power and timing to each charging and discharging time window, and calculates the remaining peak power after charging and discharging. The task scheduling layer selects the target task execution sequence and the corresponding target charging and discharging scheme based on the remaining peak power. Specifically, this includes: comparing the remaining peak power with the grid capacity limit, and selecting candidate task execution sequences whose remaining peak power does not exceed the grid capacity limit as a feasible sequence set; for each candidate task execution sequence in the feasible sequence set, calculating the grid power supply margin based on the difference between the remaining peak power and the grid capacity limit. Based on the charge / discharge depth, charge / discharge power in the target charge / discharge scheme, and the current available capacity in the status information, the capacity decay is calculated. Based on the capacity decay, the task arrival frequency distribution in the historical load characteristic data, and the peak load periodicity, the available capacity decay trajectory of the supercapacitor over multiple future decision cycles is deduced. The number of decision cycles required for the capacity to drop to the capacity failure threshold is determined based on the available capacity decay trajectory. The time length between the current moment and the moment corresponding to the number of decision cycles is calculated as the remaining lifetime of the energy storage system. Based on the grid power supply margin, the task completion time of the candidate task execution sequence, and the remaining lifetime of the energy storage system, a comprehensive evaluation index is calculated. The candidate task execution sequence with the optimal comprehensive evaluation index is selected as the target task execution sequence, and the corresponding charge / discharge power and sequence are extracted as the target charge / discharge scheme. During execution, the actual load power is monitored. When the deviation between the actual load power and the expected value exceeds a set threshold, the coupled decision-making mechanism is re-executed to generate a correction scheme, and the actual load power is added to the historical load characteristic data. Tasks are allocated according to the correction scheme, and the supercapacitor energy storage system is controlled to perform charging and discharging.

2. The method according to claim 1, characterized in that, The steps of generating multiple candidate task execution sequences by the task scheduling layer include: Based on the resource requirements and time constraints of each task in the task information to be executed, the schedulable time range of each task is determined; within the schedulable time range, multiple candidate task execution sequences are generated by adjusting the task execution order and start time. For each candidate task execution sequence, a corresponding load power curve is calculated based on the resource requirement characteristics. The load power curve represents the power requirement corresponding to the total resource occupancy of all tasks at each time point.

3. The method according to claim 1, characterized in that, The energy storage scheduling layer decomposes the peak period in the corresponding load power curve into multiple charge / discharge time windows for each candidate task execution sequence. Based on the available capacity in the status information and the historical load characteristic data, it allocates charge / discharge power and timing to each charge / discharge time window. The steps for calculating the remaining peak power after charge / discharge execution include: The period in the load power curve where the power exceeds the reference power threshold is identified as the peak period; the peak period is divided into multiple charge and discharge time windows according to the time granularity. Based on the peak-valley period distribution pattern in the historical load characteristic data, the pre-charging period of the supercapacitor before the peak period is determined, and the energy that can be stored is calculated based on the upper limit of charging power in the status information and the duration of the pre-charging period. Based on the storable energy and the available capacity, calculate the excess power of the load power exceeding the reference power threshold at the corresponding moment of each charge and discharge time window, and use the excess power as the target discharge power; The charging sequence is determined based on the end time of the pre-charging period, and the discharging sequence is determined based on the start and end times of each charging and discharging time window within the peak period, thus forming a charging and discharging sequence. Based on the target discharge power of each charge / discharge time window and the charge / discharge sequence, calculate the load power reduction at the corresponding moment after discharge in each charge / discharge time window; subtract the corresponding load power reduction from the power value at each moment in the load power curve to obtain the remaining peak power after charge / discharge.

4. The method according to claim 1, characterized in that, The steps for calculating the comprehensive evaluation index based on the power grid supply margin, the task completion time of the candidate task execution sequence, and the remaining lifetime of the energy storage system include: Based on the available capacity decay trajectory, obtain the predicted available capacity value for each future decision cycle; extract the peak load statistics for each future decision cycle from the historical load characteristic data, and calculate the minimum capacity requirement for peak shaving in each period; calculate the number of periods where the predicted available capacity value is lower than the minimum capacity requirement as the number of periods with capacity shortage risk. The third weighting coefficient is determined based on the number of periods of insufficient capacity risk; the first weighting coefficient is determined based on the ratio of the power grid supply margin to the upper limit of the power grid capacity; and the second weighting coefficient is determined based on the ratio of the task completion time to the time constraint. The comprehensive evaluation index is calculated by weighting and summing the power grid supply margin, the reciprocal of the task completion time, and the remaining lifespan of the energy storage system with the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively.

5. The method according to claim 1, characterized in that, When the deviation between the actual load power and the expected value exceeds a set threshold, the step of re-executing the coupled decision mechanism to generate a correction scheme includes: Calculate the deviation between the actual load power and the expected load power predicted based on the execution timing of the target task; When the absolute value of the deviation exceeds the set threshold, the current time is recorded as the correction trigger time, and the time period from the correction trigger time to the end of the task execution is extracted as the time window to be corrected. Obtain the remaining available capacity of the supercapacitor and the task allocation status of the executed task corresponding to the correction trigger time. Based on the deviation, correct the expected load power within the time window to be corrected to obtain the corrected load power curve. Use the corrected load power curve, the remaining tasks to be executed within the time window to be corrected, and the remaining available capacity of the supercapacitor as inputs to re-execute the coupling decision mechanism and generate the corrected task execution sequence and the corrected charging and discharging scheme. The task allocation and charging / discharging actions already executed before the correction trigger time remain unchanged. Starting from the correction trigger time, the remaining tasks are executed according to the correction task execution sequence and the correction charging / discharging scheme.

6. The method according to claim 5, characterized in that, The steps of allocating tasks and controlling the supercapacitor energy storage system to perform charging and discharging according to the modified scheme include: Based on the charging and discharging timing sequence in the target charging and discharging scheme or the modified charging and discharging scheme, generate charging and discharging control commands; The charging and discharging control command is sent to the power control unit of the supercapacitor energy storage system. During the charging and discharging process, the voltage and current parameters of the supercapacitor are collected in real time. The actual available capacity change is calculated based on the voltage and current parameters, and the actual available capacity change is updated to the current available capacity in the status information.

7. A peak load intelligent adjustment system for a supercapacitor energy storage system in a computing center, used to implement the method described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to acquire information on tasks to be executed from the computing center, status information of the supercapacitor energy storage system, and historical load characteristic data; The decision planning module is used to construct a coupled decision-making mechanism between the task scheduling layer and the energy storage scheduling layer. The task scheduling layer generates multiple candidate task execution sequences based on the information of the tasks to be executed. The energy storage scheduling layer decomposes the peak period in the corresponding load power curve into multiple charge / discharge time windows for each candidate task execution sequence. Based on the available capacity in the status information and the historical load characteristic data, it allocates charge / discharge power and timing to each charge / discharge time window and calculates the remaining peak power after charge / discharge execution. The task scheduling layer selects the target task execution sequence and the corresponding target charge / discharge scheme based on the remaining peak power. The deviation monitoring module is used to monitor the actual load power during execution. When the deviation between the actual load power and the expected value exceeds a set threshold, the coupled decision-making mechanism is re-executed to generate a correction scheme, and the actual load power is added to the historical load characteristic data. The execution control module is used to allocate tasks according to the modified scheme and control the supercapacitor energy storage system to perform charging and discharging.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images