Method for allocating virtual capacity of shared energy storage of multi-computing cluster

Abstract

The present application relates to the technical field of data center power supply and electrochemical energy storage collaborative control, and discloses a multi-computing cluster shared energy storage virtual capacity allocation method applied to an industrial cloud computing platform. The operating state parameters and power consumption data of the computing cluster are collected to form a load time series data set; the load fluctuation, peak value and cross-cluster load complementary characteristics are extracted to calculate the energy storage virtual capacity demand; the energy storage virtual capacity demand and the energy storage health state are written into the alliance value, the capacity allocation proportion is generated by using the kernel solution distribution algorithm and converted into the given power of the energy storage converter; the tracking residual is formed according to the actual output power, the tracking residual is used to update the alliance value, and the tracking residual, the proportion change and the health change meet the termination condition until the output charge and discharge power instruction is output. The present application can reduce the power allocation deviation and the charge and discharge state switching frequency, and reduce the power supply interruption risk of the computing cluster.

Description

Technical Field

[0001] This invention relates to the field of collaborative control technology for power supply and electrochemical energy storage in data centers, specifically to a method for allocating shared virtual energy storage capacity among multi-computing power clusters. Background Technology

[0002] With the rapid development of industrial cloud computing platforms, data centers, as the core carriers of computing infrastructure, are experiencing a continuous increase in power consumption. Industrial cloud platforms integrate dispersed computing resources into a unified computing resource pool through virtualization technology, providing a platform foundation for the collaborative scheduling of multiple computing clusters. Existing industrial cloud platforms primarily focus on data acquisition, monitoring, and statistical analysis in the field of energy management. Their energy storage scheduling functions typically perform simple power allocation based on state of charge or historical maximum load, lacking refined modeling of the time-domain complementary characteristics of loads across multiple computing clusters and the physical attenuation of energy storage units. Under the centralized scheduling architecture of cloud platforms, the power execution deviation of energy storage converters is difficult to feed back to the upper-level capacity allocation decision, resulting in a significant discrepancy between the mathematical allocation results and the actual execution effect.

[0003] Data center campuses typically house multiple computing clusters designed for compute-intensive tasks, graphics processing tasks, storage access tasks, or a combination of both. The CPU utilization, GPU utilization, task queue length, and storage access frequency of different computing clusters fluctuate in the time domain as tasks arrive, resulting in varying durations, peak times, and rates of change in active power load across different power supply areas. While shared energy storage pools can provide charging and discharging power to multiple power supply areas via a common AC or DC bus, energy storage capacity is energy-constrained, while the instantaneous load of computing clusters is power-constrained. Allocating capacity solely based on rated power ratios or historical maximum loads can easily lead to redundant reserves or power supply gaps.

[0004] Existing shared energy storage capacity optimization schemes can determine capacity configuration and output plans based on load forecasting, operating costs, or multi-entity benefits; cooperative game theory schemes can allocate alliance benefits using kernel-alliance solutions, Shapley values, or negotiation solutions; data center power management schemes can adjust task scheduling based on CPU utilization or rack power; and battery management schemes can perform energy storage unit power balancing based on state of charge, peak power state, internal resistance, or energy storage health status. These technologies typically handle capacity planning, benefit allocation, computing power management, or energy storage converter control, respectively. The mathematical allocation ratio obtained from the kernel-alliance solution may not be executed proportionally when energy storage unit degradation, temperature changes, charging / discharging power limits, and energy storage converter tracking deviations exist; the actual power deviation generated by the energy storage converter control layer does not reversely change the alliance value or the capacity allocation ratio in the next window iteration. Directly using correlation coefficients for computing loads under different sampling cycles can also identify misaligned peaks as unrelated loads, causing distortion in the calculation of cross-cluster complementarity.

[0005] Therefore, the technical problem that needs to be solved is: how to form an executable data closed loop in the scenario of shared energy storage power supply by multiple computing power clusters, especially in the large-scale computing power scheduling environment of industrial cloud computing platforms, to combine the load time-domain complementary characteristics, physical attenuation of energy storage units, kernel capacity allocation and energy storage converter power tracking feedback, so that the virtual capacity allocation ratio of energy storage can simultaneously meet the peak and valley demand of computing power load, the actual available capacity of shared energy storage and the power boundary of energy storage converter. Summary of the Invention

[0006] The purpose of this invention is to provide a method for sharing virtual energy storage capacity among multiple computing power clusters. The technical problem to be solved by this invention is to reduce power allocation deviation and the frequency of charging and discharging state switching, thereby reducing the risk of power outages in computing power clusters.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A method for allocating shared virtual energy storage capacity across multiple computing power clusters, characterized by comprising:

[0009] Collect the operating status parameters and power consumption data of each computing cluster to form a multi-dimensional load time series dataset;

[0010] Load fluctuation features and peak load features are extracted from the multidimensional load time series dataset. Local peaks and valleys of different computing power clusters are aligned using time warping paths. Cross-cluster load complementarity features are formed based on time warping distance and aligned load correlation coefficient, generating the energy storage virtual capacity requirements of each computing power cluster.

[0011] Collect the physical attenuation parameters of each energy storage unit in the shared energy storage resource pool to generate the energy storage health status and actual available capacity;

[0012] The virtual capacity demand for energy storage, the complementary characteristics of cross-cluster loads, the health status of energy storage, and the actual available capacity are written into the alliance value, and the capacity allocation ratio is generated using the kernel-based allocation algorithm.

[0013] The capacity allocation ratio is converted into the given power of the energy storage converter and the shared energy storage resource pool is controlled to perform charging and discharging.

[0014] Based on the actual output power of the energy storage converter and the given power of the energy storage converter, a tracking residual is formed. The health penalty coefficient of the alliance value is updated using the tracking residual and the capacity allocation ratio is regenerated until the tracking residual, the change in the capacity allocation ratio and the change in the energy storage health status meet the termination conditions, and then the charging and discharging power command is output.

[0015] Furthermore, forming a multidimensional load time series dataset includes: collecting task queue length and active power respectively; performing resampling and normalization according to the sampling timestamp; writing the most recent valid sample and attaching a missing confidence flag for missing sampling points; and dividing the multidimensional load time series dataset according to the window length covering at least one load peak-valley cycle and the overlap rate of adjacent windows.

[0016] Furthermore, the extraction of load fluctuation characteristics and peak load characteristics includes: calculating the mean, standard deviation, and maximum value of active power within the calculation window; forming a load fluctuation coefficient by the ratio of the standard deviation to the mean active power; forming a peak load factor by the ratio of the maximum value to the mean active power; and outputting the load fluctuation coefficient and peak load factor to the energy storage virtual capacity demand calculation process.

[0017] Furthermore, the cross-cluster load complementarity characteristics and the virtual capacity demand for energy storage are calculated according to the following formula:

[0018] ;

[0019] ;

[0020] in, For the first The computing cluster in the first... The maximum active load of each window For the first Window duration, For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window is calculated using the following formula:

[0021] ;

[0022] in, For the first The first window The computing cluster and the first Normalized time warp distance of a computing cluster For time-ordered paths, For the set of allowed time-warped paths, This represents the number of aligned sample position pairs contained in the time-warped path. and These are the sampling positions in the time warping path. For the first The computing cluster in the first... The first window Normalized active load at each sampling location For the first The computing cluster in the first... The first window Normalized active load at each sampling location To achieve cross-cluster load complementarity, The load correlation coefficient after alignment along the time-warped path. This is the upper limit of the allowed normalized time-normalized distance. For logical AND, For logical OR, For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window For the first The discharge efficiency of the power supply branch corresponding to each computing cluster. To preset the minimum capacity factor, As load fluctuation weight, This is the load fluctuation coefficient. Peak load weighting, Peak load factor For load complementarity weights, For the first The average cross-cluster load complementarity between a computing cluster and other computing clusters.

[0023] Furthermore, the physical degradation parameters include the current internal resistance of the energy storage unit, the initial internal resistance during commissioning, the capacity retention rate, the self-discharge rate, and the surface temperature. The changes in the current internal resistance relative to the initial internal resistance during commissioning, the capacity retention rate, the self-discharge rate, and the surface temperature deviation are normalized and weighted to form the energy storage health status. The rated capacity of the energy storage unit, the capacity retention rate, the energy storage health status, and the temperature correction coefficient are multiplied together to form the actual usable capacity.

[0024] Furthermore, the candidate energy supply capacity, coalition value, and transcendence degree of the kernel-based allocation algorithm are determined according to the following formula:

[0025] ;

[0026] ;

[0027] ;

[0028] in, For the first One energy storage unit to the alliance Provided candidate capacity, For energy storage unit index, Sum the indexes for energy storage units that are in an available state. For the first A collection of energy storage units whose windows are in an available state. For the first The energy storage unit in the first The energy storage health status of each window. For the first The energy storage unit in the first The actual usable capacity of each window A non-empty subset of the set of members of the computing power cluster. For the first The computing cluster in the ... The virtual energy storage capacity requirement for each window For the first The total actual available capacity of the shared energy storage resource pool is shared by all windows. For the first Window Alliance The value of the alliance For amplitude limiting calculation, For complementary gain weights, and For computing cluster indexing, For the first The computing cluster and the first The computing cluster in the ... Cross-cluster load complementarity of each window, For the first Health penalty coefficient for each window, For the alliance Relative to the allocation vector The degree of transcendence, For the first A capacity allocation vector for each window. For the first The virtual energy storage capacity obtained by each computing cluster; the kernel allocation algorithm ensures that the total capacity does not exceed... Furthermore, under the constraint that each capacity component is non-negative, the lexicographical minimization is performed on the vector formed by all transcendence degrees in descending order to obtain the capacity allocation ratio.

[0029] Furthermore, converting the capacity allocation ratio into the given power of the energy storage converter includes: determining the charging or discharging direction based on the difference between the active load of the computing cluster and the window reference load; forming the total power target of the shared energy storage resource pool based on the capacity allocation ratio, the virtual capacity demand of energy storage, and the window duration; forming the power weight of each energy storage unit based on its actual available capacity, energy storage health status, and peak power status; decomposing the total power target of the shared energy storage resource pool according to the power weight of the energy storage units, and applying voltage, state of charge, temperature, and rated power limit of the energy storage converter to the decomposition results.

[0030] Furthermore, the given power, tracking residual, health penalty coefficient update, and termination condition of the energy storage converter are determined according to the following formula:

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] in, For the first Iteration cycles within a window, For the first The next iteration round within each window For the first The energy storage unit corresponds to the energy storage converter in the first... The first window Given power in the next iteration, For amplitude limiting calculation, For the first The energy storage unit in the first The first window Energy storage health status in the next iteration. For the first The energy storage unit in the first The actual usable capacity of each window Sum the indexes for energy storage units that are in an available state. For the first A collection of energy storage units whose windows are in an available state. For the first The computing cluster in the first... The first window Capacity allocation ratio for each iteration For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window For the first Window duration, For symbolic operations, For the first The difference between the active load of a computing cluster and the window baseline load. For the first The energy storage unit in the first The maximum charging power of each window. For the first The energy storage unit in the first The upper limit of discharge power for each window. For the first The first window The tracking residual of the next iteration, For the first The energy storage unit corresponds to the energy storage converter in the first... The first window The actual output power of the next iteration. The upper limit of power corresponding to the current charging / discharging direction. To prevent positive numbers with a denominator of zero, For the first The first window The health penalty coefficient for the next iteration. For residual feedback gain, This represents the lower limit of the health penalty coefficient. This represents the upper limit of the health penalty coefficient. The power residual threshold, For the first The capacity allocation ratio vector for the next iteration round of each window. For the first The capacity allocation ratio vector for the current iteration round of each window. The threshold for proportional change, For the first The energy storage unit in the first The energy storage health status in the next iteration of the window. For health change threshold, For logical AND.

[0036] Furthermore, before issuing charging and discharging power commands, power balance constraints, energy storage unit terminal voltage constraints, state of charge constraints, temperature constraints, and computing cluster power supply gap constraints are established; the capacity allocation ratio and the constraint violation amount corresponding to the given power of the energy storage converter are calculated using Lagrange multipliers; when the constraint verification fails, the issuance of updated charging and discharging power commands is prohibited, the charging and discharging power commands that satisfy the constraints in the previous window are maintained, and the alliance value calculation is re-executed.

[0037] Furthermore, the multi-computing clusters are set up in different power supply areas of the data center park, and the shared energy storage resource pool is connected to different power supply areas through a common AC bus or a common DC bus; each power supply area is equipped with computing nodes, power metering devices and uninterruptible power supplies; charging and discharging power commands are sent to the energy storage converter through the local controller of the energy storage system.

[0038] Furthermore, when the power supply gap of the computing power cluster remains positive within a continuous window and the actual available capacity of the shared energy storage resource pool is insufficient, the capacity allocation ratio of the computing power cluster where the high-priority task is located is maintained, and a scheduling instruction is generated to migrate the low-priority task container to the computing power cluster with remaining power supply margin; after the task migration is completed, the power consumption data is collected again and the virtual capacity requirement of energy storage is regenerated.

[0039] Furthermore, the virtual energy storage capacity that has been allocated but has not generated actual charging and discharging power within a preset continuous window is reclaimed; the reclaimed capacity is restored to the unallocated capacity of the shared energy storage resource pool and the total actual available capacity of the shared energy storage resource pool is updated; the virtual energy storage capacity demand is regenerated based on the updated load fluctuation characteristics and peak load characteristics, and the alliance value is recalculated based on the regenerated virtual energy storage capacity demand and the updated energy storage health status, and the charging and discharging power command for the next window is generated.

[0040] The present invention also discloses a multi-computing power cluster shared energy storage virtual capacity allocation system for executing the multi-computing power cluster shared energy storage virtual capacity allocation method as described above, comprising: a computing power cluster status acquisition device, an energy storage status acquisition device, a virtual capacity allocation decision center, an energy storage system local controller, and multiple energy storage converters;

[0041] The computing power cluster status acquisition device is connected to the system management interface and power metering device of multiple computing power clusters. It is used to receive operating status parameters and power consumption data and output multi-dimensional load time series dataset to the virtual capacity allocation decision center.

[0042] The energy storage status acquisition device is connected to the battery management system of the shared energy storage resource pool to receive internal resistance, capacity retention rate, self-discharge rate, temperature, state of charge and peak power status, and output the energy storage health status and actual available capacity to the virtual capacity allocation decision center.

[0043] The virtual capacity allocation decision center includes a processor and a memory. The processor executes instructions in the memory to form the virtual capacity demand of energy storage and the cross-cluster load complementarity characteristics based on the multi-dimensional load time series dataset. The virtual capacity demand of energy storage, the cross-cluster load complementarity characteristics, the energy storage health status and the actual available capacity are written into the alliance value. The kernel de-allocation algorithm is used to generate the capacity allocation ratio and convert the capacity allocation ratio into the given power of the energy storage converter.

[0044] The local controller of the energy storage system receives the given power from the energy storage converter and sends charging and discharging power commands to multiple energy storage converters. It also collects the actual output power of multiple energy storage converters and transmits the actual output power of the energy storage converters back to the virtual capacity allocation decision center.

[0045] The virtual capacity allocation decision center forms a tracking residual based on the actual output power of the energy storage converter and the given power of the energy storage converter. It uses the tracking residual to update the health penalty coefficient of the alliance value and regenerate the capacity allocation ratio until the tracking residual, the change in the capacity allocation ratio, and the change in the energy storage health status meet the termination conditions.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] This invention connects the operating status parameters of multi-computing power clusters, power consumption data, the physical state of shared energy storage resource pools, and the power execution results of energy storage converters into a continuous data closed loop, extending the allocation of virtual energy storage capacity from load characteristic analysis to the execution of charging and discharging power commands. After resampling, normalization, and processing of missing sampling points, the multi-dimensional load time-series dataset maintains the temporal correspondence between computing resource status and active power. Load fluctuation characteristics, peak load characteristics, and cross-cluster load complementarity characteristics jointly participate in the calculation of virtual energy storage capacity demand. The time warping path aligns local peaks and valleys of different computing power clusters, reducing the impact of load peak misalignment on cross-cluster load complementarity calculation and minimizing redundant capacity reservations. The current internal resistance, capacity retention rate, self-discharge rate, and surface temperature of energy storage units constitute the energy storage health state. This health state is simultaneously incorporated into the actual available capacity, alliance value, and energy storage unit power weight, subjecting the capacity allocation ratio generated by the kernel-based allocation algorithm to constraints imposed by the physical attenuation of energy storage units and power boundaries. After the capacity allocation ratio is converted into the given power of the energy storage converter, the actual output power of the energy storage converter forms a tracking residual. The tracking residual updates the health penalty coefficient and is returned to the alliance value calculation process, which can correct the deviation between the capacity allocation result and the execution capability of the energy storage converter. Power balance constraints, energy storage unit terminal voltage constraints, state of charge constraints, surface temperature constraints, and computing cluster power supply gap constraints can prevent the issuance of charging and discharging power commands that do not meet the physical boundaries. Task migration and virtual capacity reclamation of energy storage further improve the utilization continuity of the shared energy storage resource pool and reduce the risk of continuous accumulation of power supply gaps in the computing cluster. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0049] Figure 1 This is a flowchart of the method of the present invention.

[0050] Figure 2 This is a flowchart of the cross-cluster load complementarity feature calculation process of the present invention.

[0051] Figure 3 This is a flowchart of the multi-level fact credibility assessment process of the present invention.

[0052] Figure 4 This is a flowchart of the context dependency graph and semantic consistency correction process of the present invention.

[0053] Figure 5 This is a flowchart illustrating the process of generating the load status of the manual review queue in this invention.

[0054] Figure 6 This is a flowchart of the credibility weighted correlation matrix construction process of the present invention.

[0055] Figure 7 This is a flowchart of the nucleolus solution and consortium negotiation process for this invention.

[0056] Figure 8 This is a system architecture diagram of the present invention. Detailed Implementation

[0057] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0058] The following is in conjunction with the appendix Figures 1-8 The embodiments of the present invention will be described in detail below.

[0059] This embodiment discloses a method for allocating shared virtual energy storage capacity among multiple computing power clusters, including:

[0060] Step S1: Generate a multidimensional load time series dataset

[0061] The computing cluster status acquisition device reads the task queue length through the system management interface of each computing node and the active power through the power metering device of the power supply branch. The task queue length is normalized using the maximum value within the corresponding window; the active power is normalized using the rated power of the computing cluster power supply branch.

[0062] When the original sampling periods for different computing power clusters are 15s, 30s, and 60s respectively, a unified sampling period of 60s can be used. The task queue length is taken as the effective value at the end of the unified sampling interval, and the active power is resampled within the unified sampling interval by dividing the energy integral by the interval duration. This process can maintain the time correspondence between energy calculation and task status.

[0063] The window length is determined based on the load peak-valley cycle. If the historical load shows at least one local peak and one local trough within 1 hour, the window length can be 1 hour; if the load peak-valley cycle is 4 hours, the window length can be 4 hours. The overlap rate between adjacent windows can be 50%, allowing adjacent windows to share half of the sampling points. When the proportion of missing samples within a window is no higher than 10%, the most recent valid samples are used to fill the gaps, and the window confidence level is reduced; when the proportion of missing samples within a window is higher than 10%, the charging and discharging power commands that satisfy the physical constraints of the previous window are maintained, and no new coalition allocation is executed.

[0064] With a uniform sampling period of 1 minute, a window length of 60 minutes, and an overlap rate of 50% between adjacent windows, each window contains 60 sampling points, and adjacent windows move 30 sampling points. When a window is missing 4 sampling points, the missing rate is 6.67%, which is less than 10%, and the window continues to participate in the calculation; when a window is missing 8 sampling points, the missing rate is 13.33%, which is greater than 10%, and the window does not generate a new capacity allocation ratio.

[0065] Step S2: Extract load characteristics and calculate virtual energy storage capacity demand.

[0066] No. The computing cluster in the first... The mean active power, standard deviation, load fluctuation coefficient, and peak load factor for each window are calculated using the following formulas:

[0067]

[0068] in, For the first The computing cluster in the first... The average active power of each window. For the computing power cluster index, For window indexing, For the first The number of valid sampling points per window For the index of sampling points within the window, For the first The computing cluster in the first... The first window Active power at each sampling point For the first The computing cluster in the first... The standard deviation of active power in each window. For the first The computing cluster in the first... The load fluctuation coefficient for each window, To prevent positive numbers with a denominator of zero, For the first The computing cluster in the first... Peak load factor for each window, For the first The computing cluster in the first... The maximum active power of each window.

[0069] Local peak loads from different computing power clusters may experience time shifts. The normalized time warp distance and cross-cluster load complementarity are calculated using the following formula:

[0070] ;

[0071] ;

[0072] in, For the first The first window The computing cluster and the first Normalized time warp distance of a computing cluster For the computing power cluster index, For time-ordered paths, For the set of allowed time-warped paths, This represents the number of aligned sample position pairs contained in the time-warped path. and These are the sampling positions in the time warping path. For the first The computing cluster in the first... The first window Normalized active load at each sampling location For the first The computing cluster in the first... The first window Normalized active load at each sampling location For the first The computing cluster and the first The computing cluster in the first... Cross-cluster load complementarity of each window, The load correlation coefficient after alignment along the time-warped path. This is the upper limit of the allowed normalized time-normalized distance. For logical AND, For logical OR.

[0073] Time-warped path set The restricted path extends from the start to the end of both windows, allowing only movement to adjacent sampling positions at a time, and limiting the alignment offset to no more than one-quarter of the window length. The two load sequences are rearranged along the path corresponding to the minimum normalized time warp distance, and the alignment load correlation coefficient is calculated. . Not higher than and When the value is negative, the two load sequences have an exploitable peak-shifting relationship. Forming positive collaborative contributions; Higher than or When not lower than zero, Take zero.

[0074] The virtual capacity requirement for energy storage is calculated using the following formula:

[0075] ;

[0076] in, For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window For the first The computing cluster in the first... The maximum active load of each window For the first Window duration, For the first The discharge efficiency of the power supply branch corresponding to each computing cluster. To preset the minimum capacity factor, As load fluctuation weight, For the first The computing cluster in the ... The load fluctuation coefficient for each window, Peak load weighting, For the first The computing cluster in the first... Peak load factor for each window, For load complementarity weights, For the first The average cross-cluster load complementarity between the existing computing power cluster and the remaining computing power cluster.

[0077] Load fluctuation weight Peak load weight Complementary weights for load All are non-negative numbers.

[0078] The above weights are calibrated as follows: Load data of the computing cluster and corresponding measured values ​​of power supply gap energy are collected from at least 30 historical windows. The sum of the squares of the differences between the virtual energy storage capacity demand of each window and the actual energy storage discharge energy required to eliminate the power supply gap is used as the optimization objective. A coarse-to-fine grid search strategy is employed within the [0,1] interval to determine the optimal weight combination. The validity of the calibration results is verified using an independent validation set representing at least 30% of the total number of windows. Parameter calibration only changes the weight values ​​and does not alter the input-output relationship.

[0079] when For 20MW, For 1 hour 0.95 For 0.20, 0.40, 0.15 1.60, It is 0.10 and When the value is 0.50, substituting the aforementioned parameters into the formula for virtual energy storage capacity demand, the result within the parentheses is 1.12, and the virtual energy storage capacity demand is 23.5789 MWh.

[0080] Step S3: Establishing a healthy energy storage status and actual usable capacity

[0081] The energy storage state acquisition device obtains the first data from the battery management system. Current internal resistance of each energy storage unit Initial internal resistance during commissioning Capacity retention rate Self-discharge rate Surface temperature State of charge (SOC) and peak power state. Current internal resistance is obtained through online pulse current response measurement of the battery management system. Capacity retention rate is the ratio of current dischargeable capacity to rated capacity, and self-discharge rate is the ratio of charge loss during rest to charge before rest.

[0082] The health status of energy storage is determined by the following formula:

[0083]

[0084] in, For the first The energy storage unit in the first The energy storage health status of each window. For energy storage unit index, For window indexing, As the internal resistance weight, For the first The initial internal resistance of each energy storage unit during operation For the first The energy storage unit in the first The current internal resistance of each window. As a weight for capacity retention rate, For the first The energy storage unit in the first Capacity retention rate of each window Weighted by self-discharge rate, For the first The energy storage unit in the first The self-discharge rate of each window, Temperature weighting, For the first The energy storage unit in the first Temperature correction amount for each window To allow participation in the allocation of health status as a minimum, For the first The energy storage unit in the first The surface temperature of each window For reference temperature, To allow for temperature deviation scale, This is for amplitude limiting calculation.

[0085] Internal resistance weight Capacity retention rate weight Self-discharge rate weight and temperature weight The sum is 1. An increase in current internal resistance will decrease... A decrease in capacity retention will reduce An increase in self-discharge rate will decrease The surface temperature will decrease if it deviates from the reference temperature. .

[0086] The sum of the actual usable capacity of the energy storage unit and the actual usable capacity of the shared energy storage resource pool is calculated according to the following formula:

[0087]

[0088] in, For the first The energy storage unit in the first The actual usable capacity of each window For the first The rated capacity of each energy storage unit For the first The energy storage unit in the first Capacity retention rate of each window For the first The energy storage unit in the first The energy storage health status of each window. For the first The energy storage unit in the first Temperature correction amount for each window For the first The total actual available capacity of the shared energy storage resource pool is shared by all windows. For the first A set of energy storage units whose windows are in normal communication, protection not locked, and energy storage health status not lower than the lower limit of the health status. This is the index for the energy storage unit.

[0089] With a rated capacity of 40 MWh, a capacity retention rate of 0.95, an energy storage health status of 0.92, and a temperature correction of 0.98, the actual usable capacity is 34.2608 MWh. Since the actual usable capacity is lower than the rated capacity, participation in the alliance's total capacity constraints at the rated capacity level is not permitted.

[0090] Step S4: Form the health constraint alliance value and generate capacity allocation ratio.

[0091] The set of members of the computing power cluster is denoted as Any non-empty subset is denoted as a union. . No. One energy storage unit to the alliance The provided candidate capacity is calculated according to the following formula:

[0092] ;

[0093] in, For the first One energy storage unit to the alliance Provided candidate capacity, For energy storage unit index, A non-empty subset of the set of members of the computing power cluster. For window indexing, For the first The energy storage unit in the first The energy storage health status of each window. For the first The energy storage unit in the first The actual usable capacity of each window Sum the indexes for energy storage units that are in an available state. For the first A collection of energy storage units whose windows are in an available state. For the computing power cluster index, For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window For the first The total actual available capacity of the shared energy storage resource pool across all windows.

[0094] Alliance value and transcendence are determined by the following formula:

[0095] ;

[0096] ;

[0097] in, For the first Window Alliance The value of the alliance and For the computing power cluster index, For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window For the first The total actual available capacity of the shared energy storage resource pool is shared by all windows. For complementary gain weights, For the first The computing cluster and the first The computing cluster in the first... Cross-cluster load complementarity of each window, For the first Health penalty coefficient for each window, For energy storage unit index, For the first A collection of energy storage units whose windows are in an available state. For the first The energy storage unit in the first The energy storage health status of each window. For the first One energy storage unit to the alliance Provided candidate capacity, For amplitude limiting calculation, For the alliance Relative to the allocation vector The degree of transcendence, For the first A capacity allocation vector for each window. For the first The virtual energy storage capacity obtained by each computing cluster.

[0098] The basic term is the smaller of the sum of the virtual capacity requirements of alliance members' energy storage and the sum of the actual available capacity of the shared energy storage resource pool, ensuring that the alliance's committed capacity does not exceed the available physical capacity. The collaborative gain term accumulates the product of the cross-cluster load complementarity of paired computing clusters within the alliance and the smaller virtual capacity requirement of energy storage, enabling alliances with high peak-shifting capabilities to obtain capacity reuse gains. The health penalty term accumulates the product of the health loss of energy storage units and the candidate energy supply capacity, ensuring that energy storage branches with lower energy storage health status are not treated as equivalent healthy capacity in the alliance value.

[0099] The kernel-based allocation algorithm calculates the transcendence degree for each coalition. The transcendence degree represents the degree to which the value of a coalition is not satisfied relative to the sum of the allocated capacities of its members. During the solution process, a total capacity constraint, a non-negativity constraint, and a minimum guaranteed capacity constraint for each member are established to minimize the maximum transcendence degree across all coalitions. After fixing the maximum transcendence degree in the first round, the second-largest transcendence degree in the remaining coalitions is minimized. Iteration continues until the capacity allocation vector is unique or the change in allocation vectors between adjacent rounds is below the solution tolerance.

[0100] The capacity allocation ratio is calculated according to the following formula:

[0101] ;

[0102] in, For the first The computing cluster in the first... The capacity allocation ratio of each window, For the first The computing cluster in the first... The virtual energy storage capacity obtained from each window For summation index of computing power cluster, For the first The computing cluster in the first... The virtual energy storage capacity obtained from each window.

[0103] The kernel solution calculation is converted into a linear programming sequence. Auxiliary variables are introduced in the first round. And solve:

[0104] ;

[0105] ;

[0106] ;

[0107] in, This represents the upper limit of the maximum transcendence in the first round. For the first Window Alliance The value of the alliance For the computing power cluster index, For the first The computing cluster in the first... The virtual energy storage capacity obtained from each window For the set of computing power cluster members non-empty subset For the set of members of the computing power cluster, It is an empty set. For the first The alliance value of the window-based alliance.

[0108] The alliance constraints that achieve the equality are fixed in the next round, and the upper bound of the transcendence of the remaining alliances is minimized in the next round. The number of computing power clusters is usually less than the number of computing nodes participating in energy storage scheduling. Alliance computing uses computing power clusters as members rather than individual computing nodes as members, which can control the number of alliance subsets.

[0109] The virtual energy storage capacity requirements of the three computing clusters are 48MWh, 30MWh, and 36MWh, respectively. When the total actual available capacity of the shared energy storage resource pool is 96MWh, the candidate energy supply capacity, cross-cluster load complementarity, and energy storage health status corresponding to all non-empty alliances are written into the alliance value. A capacity allocation vector is generated using a linear programming sequence, and the capacity allocation vector is normalized to form the capacity allocation ratio. All components of the capacity allocation vector are non-negative, and the sum of the capacity allocation vector does not exceed the total actual available capacity of the shared energy storage resource pool.

[0110] Step S5: Convert the capacity allocation ratio into the given power of the energy storage converter.

[0111] No. The window baseline load for each computing cluster is the average active power within that window. When the current active load is higher than the window reference load, When the value is positive, the shared energy storage resource pool discharges to the corresponding computing power cluster; when the current active load is lower than the window baseline load, A negative value indicates that the shared energy storage resource pool absorbs electrical energy. (Symbolic operation) Maintain the charging / discharging direction.

[0112] No. The first window The given power of the energy storage converter in the next iteration is formed according to the following formula:

[0113] ;

[0114] in, For the first The energy storage unit corresponds to the energy storage converter in the first... The first window Given power in the next iteration, For energy storage unit index, For window indexing, For the first Iteration cycles within a window, For amplitude limiting calculation, For the first The energy storage unit in the first The first window Energy storage health status in the next iteration. For the first The energy storage unit in the first The actual usable capacity of each window Sum the indexes for energy storage units that are in an available state. For the first A collection of energy storage units whose windows are in an available state. For the first The computing cluster in the first... The first window Capacity allocation ratio for each iteration For the computing power cluster index, For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window For the first Window duration, For symbolic operations, For the first The difference between the active load of a computing cluster and the window baseline load. For the first The energy storage unit in the first The maximum charging power of each window. For the first The energy storage unit in the first The upper limit of discharge power for each window.

[0115] The power weight of an energy storage unit is formed by the ratio of the product of its energy storage health status and actual available capacity to the sum of the products of all available energy storage units. Energy storage units with higher energy storage health status and actual available capacity receive a larger power share, and their power share decreases accordingly when either energy storage health status or actual available capacity decreases. The upper limits of charging power and discharging power are jointly determined by the peak power state output by the battery management system, the rated power of the energy storage converter, and the current state of charge.

[0116] After the energy storage converters are given a power limit for charging or discharging, the virtual capacity allocation decision center calculates the power difference between the total power target of the shared energy storage resource pool and the sum of the given power of each energy storage converter. If a power difference exists, it is redistributed to the energy storage units that have not reached their corresponding power limits, according to the proportion of their remaining power margin. If no energy storage unit can accommodate the power difference, it is written into the power supply gap constraint of the computing cluster.

[0117] The products of the energy storage health status and actual available capacity of the three energy storage units are 32MWh, 28MWh, and 20MWh, respectively. When the total discharge power target of the shared energy storage resource pool is 20MW, the unlimited given power is 8MW, 7MW, and 5MW, respectively. When the discharge power limit of the third energy storage unit is 4MW, the given power limit of the third energy storage unit is 4MW, and the remaining 1MW is redistributed according to the remaining power margin of the first and second energy storage units.

[0118] The local controller of the energy storage system converts the given power of the energy storage converter into a reference value for the active power of the energy storage converter. The current inner loop, pulse width modulation, and overcurrent protection of the energy storage converter are existing control functions of the energy storage converter; this invention does not change the current inner loop structure. This invention provides the energy storage converter with an active power reference value constrained by the capacity allocation ratio and the energy storage health status, and reads the actual output power of the energy storage converter.

[0119] Step S6: Generate tracking residuals and update coalition value

[0120] Actual output power of energy storage converter Obtained from AC-side active power measurement. The difference between the given power and the actual output power of the energy storage converter is divided by the current power limit and a positive number. The sum of these values ​​forms the normalized deviation of a single energy storage converter; the maximum value of the normalized deviations of all energy storage converters forms the tracking residual.

[0121] The tracking residual is calculated using the following formula:

[0122] ;

[0123] in, For the first The first window The tracking residual of the next iteration For window indexing, For the first Iteration cycles within a window, For energy storage unit index, For the first The energy storage unit corresponds to the energy storage converter in the first... The first window The actual output power of the next iteration. For the first The energy storage unit corresponds to the energy storage converter in the first... The first window Given power in the next iteration, The upper limit of power corresponding to the current charging / discharging direction. To prevent positive numbers with a denominator of zero.

[0124] Using the maximum normalized deviation of all energy storage converters avoids the average result of multiple normal energy storage branches from masking a single unworkable energy storage branch. Tracking residuals are used to update the health penalty coefficient.

[0125] ;

[0126] in, For the first The health penalty coefficient for the next iteration of the window. For amplitude limiting calculation, For the first The health penalty coefficient for the current iteration round of the window. For residual feedback gain, For the first The first window The tracking residual of the next iteration This represents the lower limit of the health penalty coefficient. This represents the upper limit of the health penalty coefficient. For window indexing, For the first Iteration rounds within a window.

[0127] When the tracking residual increases, the health penalty coefficient increases, and the penalty corresponding to the health loss in the alliance value of the next iteration round increases; when the tracking residual decreases, the health penalty coefficient remains within the allowable range. The updated health penalty coefficient proceeds to step S4, where the total alliance value is reconstructed and the capacity allocation ratio is calculated. At the start of a new window, the health penalty coefficient is initialized with the health penalty coefficient at which the termination condition of the previous window was met, and remains between the lower and upper limits of the health penalty coefficient.

[0128] Tracking residuals is also used to identify energy storage unit anomalies. If the normalized deviation of a single energy storage converter exceeds the power residual threshold for three consecutive control cycles, the local controller of the energy storage system checks the communication status, protection status, and temperature status. In case of communication timeout, the previous safe power level is maintained; in case of protection lockout, the corresponding energy storage unit is removed from the set. When the temperature exceeds the limit, reduce the upper limit of charging power and the upper limit of discharging power. After the abnormal handling changes the actual available energy storage unit set or power limit, re-execute steps S3 to S5.

[0129] The given power ratings of the three energy storage converters are 8MW, 7MW, and 4MW, respectively. Their actual output power ratings are 7.8MW, 6.9MW, and 3.2MW, respectively, corresponding to upper power caps of 10MW, 8MW, and 4MW. When the value is 0.01MW, the normalized biases are 0.0200, 0.0125, and 0.1995, respectively, and the tracking residual is 0.1995. When the health penalty coefficient is 0.20 and the residual feedback gain is 0.50, the health penalty coefficient for the next iteration without limiting is 0.21995.

[0130] Step S7: Loop closure termination, constraint verification, and instruction output

[0131] The closed-loop termination condition simultaneously constrains power execution, capacity allocation stability, and energy storage health state stability. The closed-loop termination condition is determined according to the following formula:

[0132] ;

[0133] in, For the first The first window The tracking residual of the next iteration The power residual threshold, For the first The capacity allocation ratio vector for the next iteration round of each window. For the first The capacity allocation ratio vector for the current iteration round of each window. The threshold for proportional change. For energy storage unit index, For the first The energy storage unit in the first The energy storage health status in the next iteration of the window. For the first The energy storage unit in the first The energy storage health status of the current iteration in each window. For health change threshold, For logical AND.

[0134] Power residual threshold Determined based on the allowable steady-state power error of the energy storage converter; proportional change threshold. Determined by dividing the capacity scheduling resolution by the total actual available capacity of the shared energy storage resource pool; health change threshold. The resolution is determined based on the battery management system's health status estimation. Iteration does not terminate if only one condition is met.

[0135] Before issuing each round of charging and discharging power commands, verify the total power balance of the shared energy storage resource pool, the terminal voltage of the energy storage units, the state of charge, the surface temperature, the charging and discharging power, and the power supply gap of the computing cluster. The constraints are expressed as follows:

[0136] ;

[0137] in, For the first The constraint function corresponding to each physical constraint For constrained indexes, For the first The first window Capacity allocation ratio vector for each iteration. For the first The first window The vector consisting of the given power of all energy storage converters in the next iteration.

[0138] The Lagrangian function used for constraint verification is formed according to the following formula:

[0139] ;

[0140] in, For Lagrange functions, The verification targets are the changes in capacity allocation ratio and the energy composition of power supply gaps. For the first The first window Capacity allocation ratio vector for each iteration. For the first The first window The vector consisting of the given power of all energy storage converters in the next iteration. For constrained indexes, For the first The non-negative Lagrange multipliers corresponding to each physical constraint For the first The constraint function corresponding to each physical constraint.

[0141] The Lagrange multiplier is used to calculate the capacity allocation ratio and the constraint violation corresponding to the given power of the energy storage converter, and to filter out charging and discharging power commands that violate physical boundaries without changing the inner current loop structure of the energy storage converter.

[0142] When the constraint verification passes and the closed-loop termination condition is met, the virtual capacity allocation decision center outputs the charging and discharging power command for this window; when the constraint verification passes but the closed-loop termination condition is not met, the health penalty coefficient is updated and the process returns to step S4; when the constraint verification fails, the charging and discharging power command that satisfies the constraints in the previous window is retained, and the process returns to step S4 to re-execute the alliance value calculation.

[0143] Setting a maximum number of iterations prevents infinite iterations caused by communication anomalies. When the maximum number of iterations is reached, the charging / discharging power command that satisfies all physical constraints and has the smallest tracking residual is selected from the completed iteration records in this window; if there is no charging / discharging power command that satisfies all physical constraints in this window, the charging / discharging power command that satisfies the physical constraints in the previous window is retained.

[0144] The missing sample proportion threshold and the upper limit of the allowable normalized time warping distance are determined based on the impact of missing historical window data on the load peak-valley identification results; the load fluctuation weight, peak load weight, load complementarity weight, and complementarity gain weight are determined based on the error between the virtual capacity demand of energy storage and the power supply gap energy in the historical window; the health status weight, health status limit, reference temperature, and allowable temperature deviation scale are determined based on the operating boundaries given by the energy storage unit manufacturer and the output range of the battery management system; the health penalty coefficient boundary, residual feedback gain, nucleolus solution tolerance, abnormal duration period, and maximum number of iterations are determined based on the allowable steady-state power error of the energy storage converter, the control cycle, and the closed-loop convergence results of the historical window.

[0145] Example 1: This example sets up 3 computing clusters and 3 energy storage units. Computing cluster 1 is a compute-intensive computing cluster, computing cluster 2 is a storage-intensive computing cluster, and computing cluster 3 is a hybrid computing cluster. The uniform sampling period is 1 minute, the window length is 60 minutes, and the overlap rate of adjacent windows is 50%. The maximum active power load of the three computing cluster windows are 45MW, 25MW, and 35MW, respectively; the average active power is 30MW, 20MW, and 25MW, respectively; and the standard deviation of active power is 12MW, 5MW, and 9MW, respectively.

[0146] Based on step S2, the load fluctuation coefficients of the three computing clusters are 0.40, 0.25, and 0.36, respectively; the peak load factors are 1.50, 1.25, and 1.40, respectively. Time warping and alignment correlation calculations yield... 0.42, For 0.30, The average cross-cluster load complementarity of the three computing clusters is 0.38, 0.36, 0.40, and 0.34, respectively.

[0147] With a window duration of 1 hour, the discharge efficiency of all power supply branches is 0.95. For 0.20, 0.15 The value is 0.10. According to the formula for virtual energy storage capacity demand, the virtual energy storage capacity demands of computing cluster one, computing cluster two, and computing cluster three are 53.0053MWh, 27.5658MWh, and 40.4526MWh, respectively.

[0148] The three energy storage units each have a rated capacity of 40MWh, with capacity retention rates of 0.98, 0.95, and 0.90, respectively; energy storage health status of 0.97, 0.92, and 0.86, respectively; and temperature correction values ​​of 1, 0.99, and 0.98, respectively. The actual usable capacities of the three energy storage units are 38.024MWh, 34.6104MWh, and 30.3408MWh, respectively, and the total actual usable capacity of the shared energy storage resource pool is 102.9752MWh.

[0149] Take complementary gain weights The health penalty coefficient for the current iteration round is 0.08. The value is 0.20. Candidate capacity is generated for each of the three energy storage units according to the candidate energy supply capacity formula, and the virtual energy storage capacity requirement, cross-cluster load complementarity, energy storage health status, actual available capacity, and candidate energy supply capacity are written into the alliance value.

[0150] The alliance values ​​of the three single-member alliances are 52.1884 MWh, 27.1409 MWh, and 39.8291 MWh, respectively; the alliance value of the alliance composed of computing power cluster one and computing power cluster two is 80.2555 MWh; the alliance value of the alliance composed of computing power cluster one and computing power cluster three is 92.9884 MWh; the alliance value of the alliance composed of computing power cluster two and computing power cluster three is 67.8081 MWh; and the alliance value of the large alliance is 102.9752 MWh.

[0151] The alliance values ​​of the seven non-empty alliances are input into the kernel solution of a linear programming sequence. Under the constraints that the capacity components are non-negative and the sum of the capacities equals the value of the largest alliance, the capacity allocation vectors are obtained as follows: 46.8676 MWh, 21.6874 MWh, and 34.4202 MWh, corresponding to capacity allocation ratios of 0.4551, 0.2106, and 0.3343, respectively. The sum of the three capacity components is 102.9752 MWh, which does not exceed the total actual available capacity of the shared energy storage resource pool.

[0152] The virtual capacity allocation decision center determines the charging and discharging direction of each computing cluster based on the difference between the current active load and the window baseline load, and writes the capacity allocation ratio, virtual energy storage capacity demand, and window duration into the total power target of the shared energy storage resource pool. The total power target of the shared energy storage resource pool is decomposed according to the power weights formed by the energy storage health status and actual available capacity of the three energy storage units, and limits are applied to charging power ceiling, discharging power ceiling, terminal voltage, state of charge, and temperature. If a power difference exists after limiting, the power difference is redistributed according to the remaining power margin of the energy storage units that have not reached the corresponding power ceiling.

[0153] The local controller of the energy storage system sends the given power of the energy storage converter to the three energy storage converters and sends back the actual output power of the energy storage converters. The virtual capacity allocation decision center generates a tracking residual based on the actual output power of the energy storage converters and the given power of the energy storage converters, updates the health penalty coefficient, and recalculates the alliance value and capacity allocation ratio. After the tracking residual, the change in the capacity allocation ratio in adjacent iterations, and the change in the energy storage health status in adjacent iterations all meet the termination condition, the system outputs the charging and discharging power command for this window.

[0154] This embodiment illustrates the complete chain from multidimensional load time-series datasets, virtual energy storage capacity demand, energy storage health status, health constraint coalition value, capacity allocation ratios, to the feedback of the energy storage converter's given power and actual output power. Each intermediate result has downstream applications.

[0155] Example 2: This example is the same as Example 1. However, in this example, the original sampling periods of the three computing clusters are 15s, 30s, and 60s, respectively. The computing cluster status acquisition device resamples the data to 60s. The maximum allowable offset of the time warping path is 15 sampling points. The ordinary correlation coefficients of the three computing clusters before alignment are 0.05, 0.12, and -0.08, respectively. After time warping, the correlation coefficients are -0.35, -0.20, and -0.30, respectively, corresponding to normalized time warping distances of 0.15, 0.20, and 0.18, respectively.

[0156] The cross-cluster load complementarity is obtained by combining the time warp distance and the aligned load correlation coefficient. is 0.5870, is 0.5000, is 0.5508. Compared with the unaligned correlation coefficient, the cross-cluster load complementarity identifies the reusable capacity between dislocation peaks. When the maximum active power load, average active power, standard deviation of active power, window duration, discharge efficiency and load weight of Embodiment 1 are used, the total energy storage virtual capacity demand of 3 computing clusters is adjusted from 121.0237 MWh to 119.0258 MWh.

[0157] The current internal resistance of the third energy storage unit rises from the initial value of 2.5 mΩ to 2.8 mΩ after being put into operation, the capacity retention rate is 0.90, the self-discharge rate is 0.01, and the temperature correction amount is 0.98. Take is 0.30, is 0.40, is 0.10, is 0.20, and the state of health of the third energy storage unit is 0.9229. The state of health is included in the candidate energy supply capacity, the coalition value health penalty term, the actual available capacity and the power weight of the energy storage unit, and the average distribution only based on the rated capacity is not adopted.

[0158] After the virtual capacity allocation decision center generates the first round of capacity allocation ratios and the given power of the energy storage converters, the local controller of the energy storage system collects the actual output power of the 3 energy storage converters. When any energy storage converter cannot track the given power of the energy storage converter due to the limitation of temperature, state of charge or peak power state, the normalized deviation formed by the actual output power of the energy storage converter and the given power of the energy storage converter is included in the tracking residual. The tracking residual updates the health penalty coefficient, and the updated health penalty coefficient is returned to the coalition value calculation process.

[0159] In the next iteration round, the coalition values of 7 non-empty coalitions are re-formed according to the updated health penalty coefficient, and the nucleolus solution linear programming sequence is used to regenerate the capacity allocation ratio. The capacity allocation ratio is rewritten into the calculation process of the given power of the energy storage converter, so that the candidate energy supply capacity corresponding to the energy storage branch with low energy storage state of health or insufficient power tracking capability bears a higher health penalty. After the tracking residual, the change of capacity allocation ratio and the change of energy storage state of health meet the joint termination condition, the charge and discharge power command is output.

[0160] This embodiment is used to illustrate that the output of the normalized time warping distance is input into the energy storage virtual capacity demand and coalition coordination gain, the output of the energy storage state of health is input into the actual available capacity, candidate energy supply capacity, coalition health penalty and energy storage unit power weight, and the output of the tracking residual is input into the update of the health penalty coefficient, forming the interaction among the load feature algorithm, the nucleolus solution algorithm and the power execution feedback.

[0161] Example 3: This example is based on the same principles as Example 1, but in this example, the three computing clusters are respectively located in three power supply areas of the data center park. The shared energy storage resource pool is connected to the uninterruptible power supply input side of the three power supply areas via a common AC bus. Each power supply area is equipped with an energy metering device, and the local controller of the energy storage system is connected to the three energy storage converters via an industrial communication network.

[0162] Within a given window, computing cluster one contains high-priority computing tasks, computing cluster two contains deferred backup tasks, and computing cluster three contains general mixed tasks. After the third energy storage unit exceeds its temperature limit and exits the set of available energy storage units, the total actual available capacity of the shared energy storage resource pool decreases from 102.9752 MWh to 72.6344 MWh. The virtual capacity allocation decision center maintains the minimum guaranteed capacity for computing cluster one and reduces the capacity allocation ratio for deferred backup tasks and general mixed tasks.

[0163] After capacity allocation, the power supply gap for computing cluster 2 was 1.8MW, 1.6MW, and 1.5MW in three consecutive windows, respectively, and the shared energy storage resource pool had no remaining discharge power margin. The power supply gap constraint triggered task migration, generating scheduling instructions to migrate low-priority backup task containers from computing cluster 2 to idle computing nodes in computing cluster 3. After the task migration was completed, the active load of computing cluster 2 decreased by 2.1MW, while the active load of computing cluster 3 increased by 1.7MW. The 0.4MW power loss corresponding to the migration overhead is reflected in the new energy consumption data.

[0164] The energy consumption data after the task migration re-enters step S1, forming an updated multidimensional load time-series dataset. This updated dataset then sequentially undergoes load feature extraction, energy storage virtual capacity demand calculation, consortium value calculation, kernel allocation, and energy storage converter given power mapping. The energy storage system's local controller outputs charging and discharging power commands based on the updated energy storage converter given power and returns the actual output power of the energy storage converter to the virtual capacity allocation decision center.

[0165] Virtual energy storage capacity that has been allocated but has not generated actual charging / discharging power within a preset continuous window is restored to the unallocated capacity of the shared energy storage resource pool. The virtual capacity allocation decision center updates the total actual available capacity of the shared energy storage resource pool, regenerates the virtual energy storage capacity demand based on the updated load fluctuation characteristics and peak load characteristics, and recalculates the alliance value based on the regenerated virtual energy storage capacity demand and the updated energy storage health status.

[0166] This embodiment illustrates the implementation of scenarios involving a common bus, multiple power supply areas, task migration, and virtual capacity reclamation of energy storage. Task migration is not an independent business rule; it is triggered by both insufficient actual available capacity of the shared energy storage resource pool and a continuous power supply gap. The result of the task migration is returned to the capacity allocation closed loop through updated energy consumption data.

[0167] Comparative Example 1: Comparative Example 1 uses the same load data of the three computing clusters, the same physical states of the three energy storage units, the same initial parameters for the alliance value, the same power weights of the energy storage units, the same power limit of the energy storage converter, and the same power execution process as Example 1. The difference is that Comparative Example 1 does not use the tracking residual formed by the actual output power of the energy storage converter and the given power of the energy storage converter to update the health penalty coefficient of the alliance value.

[0168] Existing technology uses the peak power state ratio of the energy storage unit to form the given active power of the energy storage converter, and uses the deviation between the given power and the actual output power to perform current loop control of the energy storage converter. Comparative Example 1 retains the tracking process of the energy storage converter current loop on the given power of the energy storage converter, but the execution result of the energy storage converter current loop is not returned to the alliance value calculation process.

[0169] When the energy storage converter cannot track its given power due to limitations imposed by temperature, state of charge, or peak power state, the health penalty coefficient of Comparative 1 remains unchanged within the same window, and the kernel solution capacity allocation ratio cannot be corrected based on the actual output power of the energy storage converter. Subsequent iterations continue to generate capacity allocation ratios based on the coalition value that does not reflect power execution deviations. The deviation between the given power and the actual output power of the energy storage converter can only be handled by the energy storage converter's current inner loop.

[0170] Example 1 utilizes the tracking residual to update the health penalty coefficient and re-forms the health penalty corresponding to the candidate power supply capacity, enabling the energy storage converter power execution result to be returned to the coalition value and kernel solution calculation process. The difference between Comparative Example 1 and Example 1 is whether or not the tracking residual is used to update the health penalty coefficient of the coalition value.

[0171] Comparative Example 2: Comparative Example 2 uses the same load data from the three computing clusters, the same physical states of the three energy storage units, the same normalized time warping distance, the same cross-cluster load complementarity, the same energy storage unit power weights, and the same energy storage converter control process as Example 1. However, the alliance value of Comparative Example 2 does not include the health penalty term formed by the energy storage health status, candidate energy supply capacity, and health penalty coefficient.

[0172] The alliance value in Comparative Example 2 is formed by the basic term corresponding to the virtual capacity demand of alliance members' energy storage and the collaborative gain term corresponding to the cross-cluster load complementarity. The kernel-based allocation algorithm cannot identify the health loss of different energy storage units at the alliance value stage; the energy storage health status only participates in the calculation of energy storage unit power weights in the subsequent energy storage converter given power mapping stage.

[0173] In Example 1, before the nucleolus solution calculation, the energy storage health status, actual available capacity, and candidate energy supply capacity are jointly written into the consortium value, allowing the physical degradation of energy storage to play a role in the capacity allocation ratio formation stage. After the capacity allocation ratio is formed, the energy storage health status and actual available capacity are again involved in the calculation of energy storage unit power weights. The difference between Comparative Example 2 and Example 1 is whether the consortium value includes a health penalty term.

[0174] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further illustrated below with reference to a specific application scenario.

[0175] The maximum active power loads of the three computing clusters within a 60-minute window were 45MW, 25MW, and 35MW, respectively; the average active power was 30MW, 20MW, and 25MW, respectively; and the standard deviations of active power were 12MW, 5MW, and 9MW, respectively. According to the load fluctuation coefficient calculation process, the load fluctuation coefficients of the three computing clusters were 0.40, 0.25, and 0.36, respectively; and according to the peak load factor calculation process, the peak load factors of the three computing clusters were 1.50, 1.25, and 1.40, respectively.

[0176] The pairwise normalized time warp distances of the three computing clusters are 0.20, 0.25, and 0.18, respectively. After alignment, the load correlation coefficients are -0.40, -0.20, and -0.30. All normalized time warp distances are below the allowed upper limit, and all aligned load correlation coefficients are negative. The cross-cluster load complementarity is 0.5833, 0.4800, and 0.5508, respectively. The average cross-cluster load complementarity of the three computing clusters is 0.5317, 0.5671, and 0.5154, respectively.

[0177] The window duration is set to 1 hour, the discharge efficiency to 0.95, the load fluctuation weight to 0.20, the peak load weight to 0.15, and the load complementarity weight to 0.10. The virtual energy storage capacity requirements for computing clusters 1, 2, and 3 are 52.1919 MWh, 27.1261 MWh, and 39.8064 MWh, respectively.

[0178] Three energy storage units, each with a rated capacity of 40 MWh, have capacity retention rates of 0.98, 0.95, and 0.90, respectively; energy storage health status of 0.97, 0.92, and 0.86, respectively; and temperature correction values ​​of 1, 0.99, and 0.98, respectively. The actual usable capacities of the three energy storage units are 38.024 MWh, 34.6104 MWh, and 30.3408 MWh, respectively, and the total actual usable capacity of the shared energy storage resource pool is 102.9752 MWh.

[0179] The complementary gain weight is set at 0.08, and the health penalty coefficient is set at 0.20. Based on the weights formed by the energy storage health status and the actual available capacity, the smaller value between the virtual energy storage capacity demand of each alliance member and the sum of the actual available capacity of the shared energy storage resource pool is decomposed into candidate energy supply capacity.

[0180] The alliance values ​​of the three single-member alliances are 51.3875 MWh, 26.7080 MWh, and 39.1929 MWh, respectively; the alliance value of the alliance composed of computing power cluster one and computing power cluster two is 79.3613 MWh; the alliance value of the alliance composed of computing power cluster one and computing power cluster three is 92.1090 MWh; the alliance value of the alliance composed of computing power cluster two and computing power cluster three is 67.0962 MWh; and the alliance value of the large alliance is 102.9752 MWh.

[0181] The alliance values ​​of the seven non-empty alliances were input into the kernel solution of a linear programming sequence, yielding capacity allocation vectors of 46.7510 MWh, 21.7383 MWh, and 34.4859 MWh, with capacity allocation ratios of 0.4540, 0.2111, and 0.3349, respectively. All components of the capacity allocation vectors are non-negative, and the sum of the capacity allocation vectors is 102.9752 MWh, which does not exceed the total actual available capacity of the shared energy storage resource pool.

[0182] The virtual capacity allocation decision center determines the charging or discharging direction based on the difference between the current active load of each computing cluster and the window baseline load. It then forms the total power target for the shared energy storage resource pool based on the capacity allocation ratio, virtual energy storage capacity demand, and window duration. The total power target for the shared energy storage resource pool is decomposed according to the power weights formed by the energy storage health status and actual available capacity of the three energy storage units.

[0183] The energy storage converter is given a power limit that applies to the upper limit of charging power, the upper limit of discharging power, terminal voltage, state of charge, and temperature. If there is an energy storage unit that can absorb the power difference after the limit is applied, the power difference is redistributed according to the remaining power margin; if there is no energy storage unit that can absorb the power difference after the limit is applied, the power difference is included in the power supply gap constraint of the computing cluster.

[0184] The local controller of the energy storage system executes the limited setpoint power of the energy storage converter and sends back the actual output power of the energy storage converter. The actual output power of the energy storage converter and the setpoint power of the energy storage converter form a tracking residual, which updates the health penalty coefficient. The updated health penalty coefficient is returned to the coalition value calculation process to regenerate the capacity allocation ratio and the setpoint power of the energy storage converter.

[0185] When the power tracking residual is no higher than the power residual threshold, the change in capacity allocation ratio over adjacent iterations is no higher than the ratio change threshold, and the change in energy storage health status over adjacent iterations is no higher than the health change threshold, the closed loop terminates and a charge / discharge power command is output. If any condition is not met, the process returns to the alliance value calculation process and continues iterating. When the maximum number of iterations is reached, the charge / discharge power command that satisfies all physical constraints and has the smallest tracking residual within this window is output. If there is no charge / discharge power command that satisfies all physical constraints in this window, the charge / discharge power command that satisfies the physical constraints in the previous window is retained.

[0186] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0187] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for allocating shared virtual energy storage capacity across multiple computing power clusters, characterized in that: include: Collect the operating status parameters and power consumption data of each computing cluster to form a multi-dimensional load time series dataset; Load fluctuation features and peak load features are extracted from the multidimensional load time series dataset. Local peaks and valleys of different computing power clusters are aligned using time warping paths. Cross-cluster load complementarity features are formed based on time warping distance and aligned load correlation coefficient, generating the energy storage virtual capacity requirements of each computing power cluster. Collect the physical attenuation parameters of each energy storage unit in the shared energy storage resource pool to generate the energy storage health status and actual available capacity; The virtual capacity demand for energy storage, the complementary characteristics of cross-cluster loads, the health status of energy storage, and the actual available capacity are written into the alliance value, and the capacity allocation ratio is generated using the kernel-based allocation algorithm. The capacity allocation ratio is converted into the given power of the energy storage converter and the shared energy storage resource pool is controlled to perform charging and discharging. Based on the actual output power of the energy storage converter and the given power of the energy storage converter, a tracking residual is formed. The health penalty coefficient of the alliance value is updated using the tracking residual and the capacity allocation ratio is regenerated until the tracking residual, the change in the capacity allocation ratio and the change in the energy storage health status meet the termination conditions, and then the charging and discharging power command is output.

2. The method for allocating shared virtual energy storage capacity among multi-computing power clusters according to claim 1, characterized in that, The process of creating a multidimensional load time series dataset includes: collecting task queue length and active power separately; performing resampling and normalization according to the sampling timestamp; writing the most recent valid sample and attaching a missing confidence flag for missing sampling points; and dividing the multidimensional load time series dataset according to the window length covering at least one load peak-valley cycle and the overlap rate of adjacent windows.

3. The method for allocating shared virtual energy storage capacity among multi-computing power clusters according to claim 1, characterized in that, Extracting load fluctuation characteristics and peak load characteristics includes: calculating the mean, standard deviation, and maximum value of active power within the calculation window; forming a load fluctuation coefficient by the ratio of the standard deviation to the mean active power; forming a peak load factor by the ratio of the maximum value to the mean active power; and outputting the load fluctuation coefficient and peak load factor to the energy storage virtual capacity demand calculation process.

4. The method for allocating shared virtual energy storage capacity among multi-computing power clusters according to claim 1, characterized in that, The cross-cluster load complementarity characteristics and the virtual capacity requirement for energy storage are calculated according to the following formula: ； ； in, For the first The computing cluster in the first... The maximum active load of each window For the first Window duration, For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window is calculated using the following formula: ； in, For the first The first window The computing cluster and the first Normalized time warp distance of a computing cluster For time-ordered paths, For the set of allowed time-warped paths, This represents the number of aligned sample position pairs contained in the time-warped path. and These are the sampling positions in the time warping path. For the first The computing cluster in the first... The first window Normalized active load at each sampling location For the first The computing cluster in the first... The first window Normalized active load at each sampling location To achieve cross-cluster load complementarity, The load correlation coefficient after alignment along the time-warped path. This is the upper limit of the allowed normalized time-normalized distance. For logical AND, For logical OR, For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window For the first The discharge efficiency of the power supply branch corresponding to each computing cluster. To preset the minimum capacity factor, As load fluctuation weight, This is the load fluctuation coefficient. Peak load weighting, Peak load factor For load complementarity weights, For the first The average cross-cluster load complementarity between a computing cluster and other computing clusters.

5. The method for allocating shared virtual energy storage capacity among multi-computing power clusters according to claim 1, characterized in that, The physical degradation parameters include the current internal resistance of the energy storage unit, the initial internal resistance during commissioning, the capacity retention rate, the self-discharge rate, and the surface temperature. The changes in the current internal resistance relative to the initial internal resistance during commissioning, the capacity retention rate, the self-discharge rate, and the surface temperature deviation are normalized and weighted to form the energy storage health status. The rated capacity of the energy storage unit, the capacity retention rate, the energy storage health status, and the temperature correction coefficient are multiplied to form the actual usable capacity.

6. The method for allocating shared virtual energy storage capacity among multi-computing power clusters according to claim 1, characterized in that, The candidate power supply capacity, coalition value, and transcendence of the kernel-based allocation algorithm are determined by the following formula: ； ； ； in, For the first One energy storage unit to the alliance Provided candidate capacity, For energy storage unit index, Sum the indexes for energy storage units that are in an available state. For the first A collection of energy storage units whose windows are in an available state. For the first The energy storage unit in the first The energy storage health status of each window. For the first The energy storage unit in the first The actual usable capacity of each window A non-empty subset of the set of members of the computing power cluster. For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window For the first The total actual available capacity of the shared energy storage resource pool is shared by all windows. For the first Window Alliance The value of the alliance For amplitude limiting calculation, For complementary gain weights, and For the computing power cluster index, For the first The computing cluster and the first The computing cluster in the first... Cross-cluster load complementarity of each window, For the first Health penalty coefficient for each window, For the alliance Relative to the allocation vector The degree of transcendence, For the first A capacity allocation vector for each window. For the first The virtual energy storage capacity obtained by each computing cluster; the kernel allocation algorithm ensures that the total capacity does not exceed... Furthermore, under the constraint that each capacity component is non-negative, the lexicographical minimization is performed on the vector formed by all transcendence degrees in descending order to obtain the capacity allocation ratio.

7. The method for allocating shared virtual energy storage capacity among multi-computing power clusters according to claim 1, characterized in that, Converting the capacity allocation ratio into the given power of the energy storage converter includes: determining the charging or discharging direction based on the difference between the active load of the computing cluster and the window reference load; forming the total power target of the shared energy storage resource pool based on the capacity allocation ratio, the virtual capacity demand of energy storage, and the window duration; forming the power weight of each energy storage unit based on its actual available capacity, energy storage health status, and peak power status; decomposing the total power target of the shared energy storage resource pool according to the power weight of the energy storage units, and applying voltage, state of charge, temperature, and rated power limit of the energy storage converter to the decomposition results.

8. The method for allocating shared virtual energy storage capacity among multi-computing power clusters according to claim 1, characterized in that, The given power, tracking residual, health penalty coefficient update, and termination condition of the energy storage converter are determined according to the following formula: ； ； ； ； in, For the first Iteration cycles within a window, For the first The next iteration round within each window For the first The energy storage unit corresponds to the energy storage converter in the first... The first window Given power in the next iteration, For amplitude limiting calculation, For the first The energy storage unit in the first The first window Energy storage health status in the next iteration. For the first The energy storage unit in the first The actual usable capacity of each window Sum the indexes for energy storage units that are in an available state. For the first A collection of energy storage units whose windows are in an available state. For the first The computing cluster in the first... The first window Capacity allocation ratio for each iteration For the first The computing cluster in the first... The virtual energy storage capacity requirement for each window For the first Window duration, For symbolic operations, For the first The difference between the active load of a computing cluster and the window baseline load. For the first The energy storage unit in the first The maximum charging power of each window. For the first The energy storage unit in the first The upper limit of discharge power for each window. For the first The first window The tracking residual of the next iteration, For the first The energy storage unit corresponds to the energy storage converter in the first... The first window The actual output power of the next iteration. The upper limit of power corresponding to the current charging / discharging direction. To prevent positive numbers with a denominator of zero, For the first The first window The health penalty coefficient for the next iteration. For residual feedback gain, This represents the lower limit of the health penalty coefficient. This represents the upper limit of the health penalty coefficient. The power residual threshold, For the first The capacity allocation ratio vector for the next iteration round of each window. For the first The capacity allocation ratio vector for the current iteration round of each window. The threshold for proportional change, For the first The energy storage unit in the first The energy storage health status in the next iteration of the window. For health change threshold, For logical AND.

9. The method for allocating shared virtual energy storage capacity in a multi-computing power cluster according to claim 1, characterized in that, Before issuing charging and discharging power commands, establish power balance constraints, energy storage unit terminal voltage constraints, state of charge constraints, temperature constraints, and power supply gap constraints for the computing power cluster; use Lagrange multipliers to calculate the capacity allocation ratio and the constraint violation amount corresponding to the given power of the energy storage converter; when the constraint verification fails, prohibit the issuance of updated charging and discharging power commands, maintain the charging and discharging power commands that satisfy the constraints in the previous window, and re-execute the alliance value calculation.

Images