A method and system for collaborative intelligent control of algorithm-electricity storage coordination oriented to an intelligent computing center

Abstract

The application provides an algorithm and electricity storage collaborative intelligent control method and system for a calculation center, relates to the technical field of power storage, and comprises the following steps: reading operation data of the calculation center, and constructing a collaborative state representation; performing time sequence structure analysis on calculation task execution state data and algorithm resource occupation data; performing function partition matching on the electricity impact representation result and the storage system operation state data; generating an electricity storage collaborative control strategy based on the algorithm structure mutation event, the electricity impact representation result and the storage function configuration structure, and delivering the electricity storage collaborative control strategy to a storage converter and a distribution unit to execute power regulation. The application solves the technical problem that the calculation center cluster in the prior art is prone to causing power grid low-frequency oscillation and power supply voltage flicker due to algorithm structure mutation, and affecting the power supply continuity and power quality stability of the calculation center under high dynamic algorithm load, thereby ensuring the power supply continuity and power quality stability of the calculation center under high dynamic algorithm load.

Description

Technical Field

[0001] This application relates to the field of power energy storage technology, specifically to a computing-power-storage collaborative intelligent control method and system for intelligent computing centers. Background Technology

[0002] As the core carrier of artificial intelligence infrastructure, intelligent computing centers are experiencing a continuous increase in computing power and power density. Furthermore, the training tasks for large AI models exhibit pulsed parallel computing characteristics. The synchronous start-up and shutdown of computing tasks in the time dimension and the resource contention in the spatial dimension cause drastic and non-periodic topological changes in computing load within millisecond timescales. Currently, intelligent computing centers generally use the power management architecture of traditional data centers. Their perception layer can only monitor the total power at the rack level and cannot detect the millisecond-level computing topological changes at the chip level caused by the reconstruction of parallel task relationships, creating a perception blind spot. Their decision-making layer relies on fixed-threshold energy storage charging and discharging strategies, with response times far exceeding the rhythm of computing load changes, resulting in a severe mismatch between capacity allocation and surge demand.

[0003] Because the spatiotemporal distribution of power surges caused by sudden changes in computing power structure cannot be predicted in advance, when GPU clusters synchronously switch tasks, triggering instantaneous power jumps, the uncompensated power deficit is propagated upwards along the power supply topology to the upper-level power distribution network. This may exacerbate the risk of weakly damped low-frequency oscillations in the power supply system and cause voltage spikes and drops and excessive flicker at the point of common coupling, thereby interfering with the stable operation of other computing nodes within the same power supply zone, and even triggering over / under voltage protection on the high-voltage side of the server, leading to task interruption or hardware damage. Simultaneously, the existing energy storage configurations are limited and cannot provide differentiated power support based on the type of surge, further exacerbating the dynamic mismatch between energy storage peak-shaving capabilities and computing power valley-filling needs. This severely restricts the power supply continuity guarantee capability and comprehensive power quality control level of intelligent computing centers in the context of high green electricity access.

[0004] In summary, existing technologies suffer from technical problems such as sudden changes in the computing power structure of intelligent computing center clusters, blind spots and lags in the response of traditional power storage systems to intelligent computing loads, which can easily cause low-frequency oscillations in the power grid and voltage flicker, further affecting the power supply continuity and power quality stability of intelligent computing centers under high dynamic computing loads. Summary of the Invention

[0005] The purpose of this application is to provide a computing, power, and storage collaborative intelligent control method and system for intelligent computing centers, in order to solve the technical problems in the prior art where, due to sudden changes in the computing power structure of intelligent computing center clusters, traditional power and storage systems have blind spots and response delays to intelligent computing loads, which easily cause low-frequency oscillations in the power grid and power supply voltage flicker, further affecting the power supply continuity and power quality stability of intelligent computing centers under high dynamic computing loads.

[0006] To achieve the above objectives, this application provides a computing, power, and storage collaborative intelligent control method and system for intelligent computing centers.

[0007] Firstly, this application provides a computational-power-storage collaborative intelligent control method for intelligent computing centers. This method is implemented through a computational-power-storage collaborative intelligent control system for intelligent computing centers. The method includes: reading operational data from the intelligent computing center, including computational task execution status data, computing power resource occupancy data, power supply and demand status data, and energy storage system operational status data; constructing a collaborative status representation after time alignment and spatial mapping of the operational data; and performing temporal structure analysis on the computational task execution status data and computing power resource occupancy data based on the collaborative status representation to identify computational load spikes in task parallelism and resource occupancy topology at the current moment. The system identifies structural disturbances and their intensity. Based on the structural disturbance intensity, it maps the computing power structural mutation events to power supply and demand status data, identifies the propagation range and local power fluctuation characteristics of the corresponding power supply and demand pressure, and establishes power impact characterization results. It then uses the power impact characterization results and energy storage system operation status data for functional partition matching to generate an energy storage functional configuration structure adapted to the power impact characterization results. This energy storage functional configuration structure includes a transient power vibration suppression unit, a time-shifted energy smoothing unit, and a local power supply compensation unit. Based on the computing power structural mutation events, power impact characterization results, and energy storage functional configuration structure, it generates a power-storage collaborative control strategy. This control strategy is then parsed into pulse width modulation signals and switching commands, and sent to the energy storage converter and distribution unit for power regulation.

[0008] Optionally, a unified clock source is used to timestamp and align the computing task execution status data, computing power resource occupancy data, power supply and demand status data, and energy storage system operation status data; based on the physical deployment topology of the intelligent computing center, computing power resource nodes, power supply units, and energy storage units are mapped to the same spatial coordinate system to construct a computing-power-storage collaborative mapping table.

[0009] Optionally, in response to the scheduling cycle division of computation task execution status data, a task parallel relationship graph is constructed in each time slice. Nodes in the task parallel relationship graph represent computation task instances, and edges represent parallel execution relationships or resource competition relationships between tasks. Each node is assigned a corresponding resource occupancy vector based on computational resource occupancy data. The graph structure difference is calculated based on the task parallel relationship graphs of adjacent time slices. The graph structure difference includes at least the node activity change rate, the edge connection relationship reconstruction ratio, and the resource occupancy vector distribution offset. When the graph structure difference exceeds a preset mutation judgment threshold, the corresponding time slice is marked as the time of the computational power structure mutation event. A structural disturbance intensity index is constructed based on the computational power structure mutation event. The structural disturbance intensity index is constructed by normalizing and weighting the node activity change rate, the edge connection relationship reconstruction ratio, and the resource occupancy vector distribution offset.

[0010] Optionally, the resource occupancy vector distribution offset is obtained by performing a weighted centroid migration calculation on the resource occupancy vector of each node.

[0011] Optionally, based on the structural disturbance intensity corresponding to the computing power structure mutation event, a power mapping association is performed on the set of associated computing task instances to establish a mapping correspondence between task nodes and power supply units. The power supply unit includes at least a rack-level power distribution unit, a node-level power supply module, and a bus distribution unit. The structural disturbance intensity is decomposed into node-level power disturbance factors according to the mapping correspondence. Then, according to the topological position of the task in the task parallel relationship graph, propagation calculations are performed level by level along the power supply unit hierarchy to generate a set of power supply and demand pressure propagation paths. A pressure diffusion coefficient is calculated based on the load accumulation effect of the path nodes on the power supply and demand pressure propagation path set. The pressure diffusion coefficient is used to configure the propagation range of power supply and demand pressure. Local power fluctuation features are extracted from the real-time power change sequence of each power supply unit within the propagation range. These local power fluctuation features include at least the power transition amplitude, fluctuation duration, and fluctuation concentration.

[0012] Optionally, based on the rated capacity, current load rate and line impedance characteristics of each power supply unit, the pressure attenuation factor or amplification factor at that power supply unit is calculated along the power supply and demand pressure propagation path; the attenuation factor or amplification factor of each level is accumulated to establish the pressure diffusion coefficient.

[0013] Optionally, based on the local power fluctuation characteristics in the power impact characterization results, the power impact is divided into transient impact components, continuous fluctuation components, and local off-center load components, and the energy regulation demand type corresponding to each component is determined respectively; the energy storage system operation status data is decomposed to construct an energy storage adjustable capability space, which includes at least instantaneous power response capability, energy time-shift regulation capability, and local power supply output capability, and an energy storage capability vector set is established; the matching fit degree between the power impact component and the energy storage unit is calculated based on the matching relationship between the energy regulation demand type and the energy storage capability vector set; the energy storage unit is functionally partitioned according to the fit degree to generate an energy storage functional partition set corresponding to different power impact components; and corresponding energy storage regulation execution mechanisms are configured in the energy storage functional partition set to generate an energy storage functional configuration structure.

[0014] Optionally, a transient power vibration suppression unit is configured for the energy storage zone corresponding to the transient impact component to perform short-time high-power fast response regulation; a time-shifted energy smoothing unit is configured for the energy storage zone corresponding to the continuous fluctuation component to perform energy timing migration and fluctuation smoothing control; and a local power supply compensation unit is configured for the energy storage zone corresponding to the local off-center load component to perform local power supply redistribution and load balancing compensation.

[0015] Optionally, when the energy storage system operation status data indicates that some energy storage units are unavailable or their health status is below a safe threshold, a functional partition reallocation instruction is generated; based on the functional partition reallocation instruction, under the constraint of ensuring the continuity of power supply for critical computing tasks, the transient power suppression, time-shifted energy smoothing, and local power supply compensation functions are downgraded.

[0016] Secondly, this application also provides a computing-power-storage collaborative intelligent control system for intelligent computing centers, used to execute a computing-power-storage collaborative intelligent control method for intelligent computing centers as described in the first aspect. The system includes: a running data processing module, used to read running data from the intelligent computing center, including computing task execution status data, computing power resource occupancy data, power supply and demand status data, and energy storage system running status data; after time alignment and spatial mapping processing of the running data, a collaborative state representation is constructed; and a time-series structure parsing module, used to perform time-series structure parsing on the computing task execution status data and computing power resource occupancy data based on the collaborative state representation, identify computing power structure mutation events in the task parallelism and resource occupancy topology of the computing load at the current moment, and identify the structure. The system comprises the following modules: a disturbance intensity module; a data mapping module, used to map the computing power structure mutation event to the power supply and demand status data based on the structural disturbance intensity, identify the propagation range and local power fluctuation characteristics of the corresponding power supply and demand pressure, and establish the power impact characterization result; a functional partition matching module, used to perform functional partition matching with the energy storage system operation status data using the power impact characterization result, and generate an energy storage functional configuration structure adapted to the power impact characterization result, wherein the energy storage functional configuration structure includes a transient power vibration suppression unit, a time-shifted energy smoothing unit, and a local power supply compensation unit; and a strategy generation module, used to generate an energy storage collaborative control strategy based on the computing power structure mutation event, the power impact characterization result, and the energy storage functional configuration structure, and to parse the control strategy into pulse width modulation signals and switching instructions, which are then sent to the energy storage converter and distribution unit to perform power regulation.

[0017] One or more technical solutions provided in this application have at least the following technical effects or advantages: spatiotemporal unified mapping of multi-source heterogeneous data to construct a collaborative state representation, identification of computing power structure mutation events and structural disturbance intensity, mapping along the power supply physical level to the propagation range and local power fluctuation characteristics of power supply and demand pressure, decomposing power impacts, and dynamically matching them with the energy storage system in functional partitions, directly parsing the comprehensive control strategy into pulse width modulation signals and switching instructions and issuing them to the underlying physical devices for execution, realizing a one-to-one targeted correspondence between the energy storage functional characteristics and different power impact components, and ensuring the power supply continuity and power quality stability of the intelligent computing center under high dynamic computing load.

[0018] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a collaborative intelligent control method for computing, power, and storage centers, as proposed in this application.

[0021] Figure 2 This is a time-series line graph showing the identification of sudden changes in computing power structure and the intensity of disturbances in the computing, power, and storage collaborative intelligent control method for intelligent computing centers, as described in this application.

[0022] Figure 3 This is a schematic diagram of the structure of a computing, power, and storage collaborative intelligent control system for intelligent computing centers, as described in this application.

[0023] Figure labeling: 11 running data processing module, 12 time series structure parsing module, 13 data mapping module, 14 function partition matching module, 15 strategy generation module. Detailed Implementation

[0024] This application provides a collaborative intelligent control method and system for computing, power, and energy storage in intelligent computing centers. It addresses the technical problems in existing technologies where sudden changes in the computing power structure of intelligent computing center clusters lead to blind spots and response delays in traditional energy storage systems, easily causing low-frequency oscillations in the power grid and voltage flicker, further affecting the power supply continuity and power quality stability of intelligent computing centers under high dynamic computing loads. The method performs a unified spatiotemporal mapping of multi-source heterogeneous data to construct a collaborative state representation, identifies sudden computing power structure events and the intensity of structural disturbances, and maps them step-by-step along the power supply physical level to the propagation range and local power fluctuation characteristics of power supply and demand pressures. It decomposes power impacts and dynamically matches them with the energy storage system's functional partitions. The comprehensive control strategy is directly parsed into pulse width modulation signals and switching commands, which are then sent to the underlying physical devices for execution. This achieves a targeted one-to-one correspondence between energy storage functional characteristics and different power impact components, ensuring the power supply continuity and power quality stability of intelligent computing centers under high dynamic computing loads.

[0025] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0026] Example 1, please refer to the appendix. Figure 1 This application provides a computing, power, and storage collaborative intelligent control method for intelligent computing centers. The method is applied to a computing, power, and storage collaborative intelligent control system for intelligent computing centers, and specifically includes the following steps: Read the operational data of the intelligent computing center, which includes computing task execution status data, computing power resource occupancy data, power supply and demand status data, and energy storage system operation status data. After time alignment and spatial mapping processing of the operational data, construct a collaborative status representation.

[0027] Furthermore, this application also includes the following steps: using a unified clock source to timestamp-align the computing task execution status data, computing power resource occupancy data, power supply and demand status data, and energy storage system operation status data; and mapping computing power resource nodes, power supply units, and energy storage units to the same spatial coordinate system based on the physical deployment topology of the intelligent computing center, thereby constructing a computing-power-storage collaborative mapping table.

[0028] Specifically, the data acquisition subsystem deployed within the intelligent computing center reads computing task execution status data, computing resource occupancy data, power supply and demand status data, and energy storage system operation status data from the computing task scheduler, computing resource monitoring agent, power quality monitoring terminal, and energy storage system battery management system, respectively. The computing task execution status data is a collection of information describing the current execution progress, task ID, task type (synchronous / asynchronous), priority, user, or queue of each running computing task within the intelligent computing center. The computing resource occupancy data is a collection describing the proportion and allocation relationship of various computing resources currently occupied by each task relative to the total amount. The power supply and demand status data is a collection describing the real-time voltage, current, active power, reactive power, power factor, and supply and demand balance status of each power supply node in the intelligent computing center, as well as the upstream power grid. The energy storage system operation status data is a collection describing the state of charge, health status, maximum discharge / charge power, current charge / discharge status, and temperature of each energy storage unit in the energy storage system.

[0029] The raw sampling frequencies of computing task execution status data, computing resource usage data, power supply and demand status data, and energy storage system operation status data are different. For example, computing resource monitoring may be collected once per second, power quality monitoring may be collected once every 20ms, and energy storage BMS data may be updated once every 100ms. In addition, the clocks within each subsystem drift to varying degrees.

[0030] A unified clock source is used to align the timestamps of data on computing task execution status, computing resource usage, power supply and demand, and energy storage system operation. Each original data record is re-marked with a standardized timestamp. The time interval with the highest sampling frequency among all data sources is used as the unified time grid. For data sources with lower sampling frequencies, nearest neighbor interpolation is used to map their values ​​to each unified time grid point, ensuring that all four types of data exist and are time-consistent at any given time point. The unified clock source is a time synchronization device that provides a globally consistent time reference for the entire intelligent computing center control system. It typically uses BeiDou satellite timing signals to ensure that data collected by different subsystems are comparable within the same time coordinate system.

[0031] A physical deployment topology map, constructed based on the physical design drawings of the intelligent computing center, is pre-stored. This map details the complete electrical hierarchy, starting from the transformer outgoing line, through the low-voltage main cabinet, segmented busbars, floor distribution boxes, rack header cabinets, rack power distribution units, and finally to the power input of each server. It also records the specific locations of each energy storage battery cluster and energy storage converter connected to the busbar. Based on this physical deployment topology map, each computing resource node is assigned unique spatial coordinates, typically represented by (room number, rack row number, rack column number, rack U-position number). Each power supply unit is assigned corresponding electrical hierarchy coordinates, and each energy storage unit is assigned the coordinates of its connected busbar node.

[0032] Based on the dual principles of proximity by electrical distance and matching power supply hierarchy, a computing-power-storage collaborative mapping table is constructed. Each row records a computing resource node, whose associated power supply unit is the upstream PDU port powered by that node and the bus segment it is connected to. The associated energy storage units are several energy storage clusters connected to the same bus segment and closest in electrical distance. The computing-power-storage collaborative mapping table is a structured data table that records the relationship between the power supply unit corresponding to each computing resource node and the nearby deployed energy storage units.

[0033] After the computing-power-storage collaborative mapping table is constructed, the four types of data that have undergone time alignment are merged according to the relationships in the mapping table to form a collaborative state representation. This is a multi-dimensional data matrix with the time axis as the first dimension, spatial coordinates and device ID as the second dimension, and the values ​​of each data attribute as the third dimension. The collaborative state representation integrates the four types of heterogeneous data that have undergone time alignment and spatial mapping into a unified multi-dimensional data structure.

[0034] By aligning timestamps, four types of data with different sampling periods are unified to the same time base, eliminating misjudgments caused by clock deviations. Through the construction of spatial mapping and collaborative mapping tables, abstract computing tasks are hard-linked to specific power supply units and energy storage units, so that when a computing node experiences a sudden change, the affected PDU ports and buses can be directly indexed, and the nearest available energy storage unit can be quickly located.

[0035] Based on the collaborative state representation, the temporal structure of the computing task execution state data and computing resource occupancy data is analyzed to identify the computing power structure mutation events in the current moment in terms of the computing load in terms of task parallel relationship and resource occupancy topology, and to identify the intensity of structural disturbance.

[0036] Furthermore, this application also includes the following steps: in response to the scheduling cycle division of the computation task execution status data, a task parallel relationship graph is constructed in each time slice, wherein the nodes of the task parallel relationship graph represent computation task instances, and the edges represent parallel execution relationships or resource competition relationships between tasks; a corresponding resource occupancy vector is assigned to each node based on the computational resource occupancy data; the graph structure difference degree is calculated based on the task parallel relationship graphs of adjacent time slices, wherein the graph structure difference degree includes at least the node activity change rate, the edge connection relationship reconstruction ratio, and the resource occupancy vector distribution offset; when the graph structure difference degree exceeds a preset mutation judgment threshold, the corresponding time slice is marked as the time when the computational power structure mutation event occurs; a structural disturbance intensity index is constructed based on the computational power structure mutation event, wherein the structural disturbance intensity index is constructed by normalizing and weighting the node activity change rate, the edge connection relationship reconstruction ratio, and the resource occupancy vector distribution offset.

[0037] Furthermore, this application also includes the following steps: the resource occupancy vector distribution offset is obtained by performing a weighted centroid migration calculation on the resource occupancy vector of each node.

[0038] Specifically, according to a preset scheduling period, such as 100ms per time slice, the execution status and computing resource usage data of all computing tasks within the current time slice are extracted from the cooperative state representation. Assuming the current time slice is the k-th time slice, all active computing task instances within this time slice are treated as nodes. Edges are established between nodes based on task dependencies, synchronization barrier information, and shared resource access records in the task scheduling log. If two tasks are scheduled to run independently on different GPUs at the same time without data exchange, a parallel execution relationship edge is established between these two nodes; if two tasks share the same GPU's memory bandwidth or the same PCIe channel, a resource contention relationship edge is established. The task parallel relationship graph is a directed or undirected graph data structure used to depict the interdependencies between all running computing tasks within a given time slice. Nodes represent independent computing tasks, and edges are divided into parallel execution relationship edges and resource contention relationship edges. Parallel execution relationship edges indicate that two tasks can run independently at the same time without data dependency; resource contention relationship edges indicate that two tasks are competing for the same type of limited computing resources, such as sharing memory bandwidth. Each node is assigned a resource occupancy vector, which is usually determined based on the computing resource occupancy data corresponding to the task in the collaborative state representation. The resource occupancy vector is a multi-dimensional numerical vector assigned to each task node. Each dimension represents the quantity of various computing resources occupied by the task at a certain moment, such as 4 GPU cores, 36.2GB of video memory, 8 CPU cores, 320GB / s of memory bandwidth, and 12.5Gbps of network bandwidth.

[0039] The task parallel relationship graphs of adjacent time slices are compared to calculate the graph structure difference. Specifically, the task parallel relationship graphs of the previous and current time slices are taken, and the node activity change rate is calculated. This is calculated by summing the number of newly added and disappeared nodes in the current time slice relative to the previous time slice, dividing by the total number of nodes in the previous time slice, and obtaining a percentage value. For example, 12.5% ​​means that 12.5% ​​of the tasks have changed in start and stop between adjacent time slices. The edge connection relationship reconstruction ratio is calculated by summing the number of newly added, disappeared, and type-changed edges between the two graphs, dividing by the total number of edges in the previous time slice, and obtaining another percentage value. The weighted centroids of the resource occupancy vectors of all nodes in both time slices are calculated. For each node, its resource occupancy vector is taken, and a weight is assigned to it based on the computing power proportion of that task. , where vi is the i-th resource occupancy vector and wi is the weight corresponding to the i-th resource occupancy vector. The Euclidean distance between the weighted centroids of the two time slices is calculated to obtain the resource occupancy vector distribution offset.

[0040] The graph structure difference score is a composite score that integrates the node activity change rate, edge connection reconstruction ratio, and resource consumption vector distribution offset based on preset weights. For example, the weights of the three sub-indicators are 0.3, 0.4, and 0.3, respectively. The calculated graph structure difference score is then compared with a preset mutation judgment threshold. The preset mutation judgment threshold is set to four times the historical average graph structure difference score under normal system operation. For example, if the normal operation average is 0.08, the preset mutation judgment threshold is set to 0.32. If the current score exceeds the preset mutation judgment threshold, the current time slice is marked as the time of occurrence of the computing power structure mutation event, and the timestamp of the mutation, the graph structure snapshot before and after the mutation, and the detailed values ​​of each sub-indicator are recorded.

[0041] Based on the labeled mutation events, a structural perturbation intensity index is constructed. The three sub-indices—node activity change rate, edge connection reconstruction ratio, and resource occupancy vector distribution offset—are each min-max normalized, mapping all values ​​to the [0,1] interval. These values ​​are then weighted and summed according to a preset weight vector to obtain the final structural perturbation intensity value. The structural perturbation intensity is a dimensionless quantification index normalized to the [0,1] interval, used to characterize the severity of the potential power impact on the downstream power supply system caused by the identified computing power structural mutation events. Values ​​closer to 1 indicate a more severe impact, while values ​​closer to 0 indicate a more gradual change.

[0042] For example, as shown in the appendix Figure 2The figure shows the changes in the computing power structure disturbance intensity of the intelligent computing center over 24 time slices (120 minutes in total). By constructing a task parallel relationship graph, the change rate of node activity, the reconstruction ratio of edge connection relationships, and the offset of resource occupation vector distribution in adjacent time slices were calculated, and the structural disturbance intensity index was obtained by normalization and weighting. When the disturbance intensity exceeds the threshold of 0.35, it is determined to be a computing power structure mutation event, triggering subsequent power impact mapping and energy storage collaborative control. The figure marks three mutation times: T8 (0.478), T12 (0.543), and T22 (0.369), where T12 is the maximum disturbance intensity. Specifically, there are 8 server racks (numbered 0-7), 4 nodes per rack, a total of 32 computing nodes, a rated capacity of 50kW per rack, a rated capacity of 12.5kW per node, a total power supply capacity of 400kW, an energy storage capacity of 200kWh (190kWh available, SOH=95%), a maximum energy storage power of 150kW, a time slice granularity of 5 minutes, and a total duration of 24 time slices = 120 minutes. A task parallelism graph is constructed, where at time T7, there are 32 nodes, with nodes 0-27 active and nodes 28-31 idle. Edges exist between active nodes within the same rack, indicating competition. The number of edges is 42 (6 edges each for racks 0-6) and 1 edge for rack 7 (only 2 active nodes), for a total of 43 edges. The resource occupancy vector is 3.0-6.0kW for each node, with a weighted centroid of 3.72kW. At time T8, the task parallel graph has 32 nodes, with nodes 0-23 active and nodes 24-31 idle. The number of edges is as follows: racks 0-3 each have 24 edges; racks 4-5 each have 12 edges; racks 6-7 have no active nodes, resulting in a total of 36 edges. The resource occupancy vector for nodes 0-15 is 12.0kW, for nodes 16-23 it is 6.0-9.5kW, and for nodes 24-31 it is 0.5kW. The weighted centroid is 8.53kW. During T0-T7 (stable period), each node is randomly in an idle / lightly loaded / medium loaded state. The task parallel graph structure is stable, with node activity approximately 28 / 32 and approximately 45 edge connections. The resource occupancy vector centroid is approximately 3.8kW, and the disturbance intensity is 0.10-0.23, which is lower than the preset mutation threshold of 0.35. In T8, the large-scale distributed training started with all racks 0-3 fully loaded (20 nodes, state = 4), racks 4-5 partially overloaded, and racks 6-7 idle. The node activity change rate was 0.312, the edge connection reconstruction ratio was 0.286, the resource occupancy vector distribution offset was 0.54, and the computational structure disturbance intensity was 0.478, exceeding the preset mutation threshold of 0.35, thus triggering a computational power structure mutation event. T9-T10 showed stable load patterns with no significant changes in the graph structure, and a disturbance intensity of 0.004, far below the preset mutation threshold. In T12, the load shifted from racks 0-3 to racks 4-7, the graph structure was reconstructed again, and the edge connection relationships changed drastically, with a disturbance intensity of 0.543.The timeframes from T13 to T21 stabilized again. During T22, local task scheduling changed, with a disturbance intensity of 0.369, exceeding the preset mutation threshold of 0.35.

[0043] By reconstructing the ratio of edge connections and the offset of resource consumption vector distribution, structural changes that cause I / O storms even though computation is paused can be identified, accurately distinguishing between benign fluctuations and malignant mutations, and avoiding false alarms based on simple thresholds.

[0044] Based on the structural disturbance intensity, the computing power structural mutation event is mapped to the power supply and demand status data, the propagation range and local power fluctuation characteristics of the corresponding power supply and demand pressure are identified, and the power impact characterization results are established.

[0045] Furthermore, this application also includes the following steps: power mapping association of the set of associated computing task instances based on the structural disturbance intensity corresponding to the computing power structural mutation event, establishing a mapping correspondence between task nodes and power supply units, wherein the power supply unit includes at least a rack-level power distribution unit, a node-level power supply module, and a bus distribution unit; decomposing the structural disturbance intensity into node-level power disturbance factors according to the mapping correspondence, and performing step-by-step propagation calculations along the power supply unit hierarchy according to the topological position of the task in the task parallel relationship graph, generating a set of power supply and demand pressure propagation paths; calculating the pressure diffusion coefficient based on the load accumulation effect of the path nodes on the set of power supply and demand pressure propagation paths, and configuring the propagation range of power supply and demand pressure using the pressure diffusion coefficient; extracting local power fluctuation features from the real-time power change sequence of each power supply unit within the propagation range, wherein the local power fluctuation features include at least the power transition amplitude, fluctuation duration, and fluctuation concentration.

[0046] Furthermore, this application also includes the following steps: calculating the attenuation factor or amplification factor of the pressure on each power supply unit based on the rated capacity, current load rate and line impedance characteristics of each power supply unit along the power supply and demand pressure propagation path; accumulating the attenuation factors or amplification factors of each level to establish a pressure diffusion coefficient.

[0047] Specifically, based on the set of task instances associated with abrupt changes in the computing power structure, including newly added tasks, disappeared tasks, and existing tasks whose resource occupancy vectors have changed significantly, the unique identifiers of these tasks and their corresponding spatial mapping information are read from the collaborative state representation. By searching the computing-power-storage collaborative mapping table, a mapping relationship is established between each abruptly associated task and its power supply unit at each level of the power supply link. The power supply unit is the power equipment at each level in the power supply link of the intelligent computing center, including at least rack-level power distribution units, node-level power supply modules, and bus distribution units. The rack-level power distribution unit is the terminal power distribution equipment that distributes power to multiple servers in a single rack; the node-level power supply module is the power module inside the server that converts AC power to DC power to supply the motherboard and GPU; the bus distribution unit is the bus section and switching equipment in the low-voltage distribution cabinet that distributes the transformer output to multiple feeder circuits.

[0048] For each task node, determine the node-level power supply module number of the server it belongs to, the port number of the rack-level power distribution unit to which the server is connected, the bus distribution unit and bus segment number to which the rack-level power distribution unit belongs, and the transformer number to which the bus is finally connected, forming a complete power supply hierarchy link from the task node to the transformer.

[0049] Based on the mapping relationship, the structural disturbance intensity is decomposed into node-level power disturbance factors. The power change of each task in the mutation-related task set is calculated, that is, based on the difference in resource occupancy vector before and after the mutation, the power consumption coefficient of each computing resource type is converted into power increment. The proportion of the power increment to the total power increment of all mutation-related tasks is multiplied by the structural disturbance intensity to obtain the node-level power disturbance factor of the task node, which quantifies the contribution of each task to the overall power impact.

[0050] Based on the task's topological location and power supply hierarchy in the task parallel relationship graph, a step-by-step propagation calculation is performed. For each mutation-related task node, starting from its corresponding server power input, the path proceeds sequentially upwards through the rack-level power distribution unit, node-level power supply module, and bus distribution unit to the transformer. The node-level power disturbance factor corresponding to each task is propagated step-by-step along this path. At each power supply unit, the attenuation factor for the disturbance component is calculated based on the unit's rated capacity, current load rate, and impedance parameters of the connecting lines. The attenuation factor equals the available capacity margin of the power supply unit at that level (rated capacity minus current load) divided by the rated capacity, multiplied by a loss coefficient related to line impedance. The higher the impedance, the higher the loss, and the smaller the attenuation factor. The amplification factor is used when disturbances may be amplified when passing through a certain power supply unit under certain special operating conditions. The amplification factor is greater than 1. If an abnormal operating condition is detected, such as a sudden increase in harmonic content or a sharp fluctuation in power factor, and the resonant amplification trigger condition is met, the backup resonant model is invoked, and the impedance component is replaced with an amplification factor greater than 1 to reflect the disturbance amplification effect. This scenario is an occasional special case.

[0051] The pressure diffusion coefficient along the propagation path is obtained by multiplying and accumulating the attenuation factors / amplification factors at all levels. The pressure diffusion coefficient is a comprehensive coefficient obtained by multiplying and accumulating the attenuation factors / amplification factors at each level along the propagation path, representing the total attenuation of the initial disturbance by the entire power supply link from the mutation task node to the transformer output. For example, if the attenuation factors are f1, f2, f3, and f4, then the pressure diffusion coefficient K = f1 × f2 × f3 × f4. The initial structural disturbance intensity is the proportion of the remaining intensity to the initial intensity after propagating from the task node through the entire power supply link to the highest reference point along the path. The pressure diffusion coefficient is calculated for the propagation path of all mutation-related tasks, and finally, the coefficients of each path are weighted and averaged. The weight is the proportion of the task power increment corresponding to that path to the total power increment, resulting in a comprehensive pressure diffusion coefficient that characterizes the upstream propagation degree of the overall electrical energy impact of this mutation event.

[0052] By utilizing the pressure diffusion coefficient, the propagation range of power supply and demand pressure is configured. The propagation is iterated upwards from the task node. When the remaining disturbance intensity at a certain power supply unit (i.e., the initial disturbance intensity multiplied by the cumulative attenuation coefficient up to that level) drops below 5% of the initial disturbance intensity, that level and higher levels are no longer included in the propagation range. The upper boundary of the propagation range ends at the level below that level. The propagation range determines the upstream boundary of the impact.

[0053] Within a defined propagation range, real-time power change sequences of each power supply unit are extracted within a continuous event window before and after the mutation. The sampling time interval is consistent with the cooperative state characterization, such as a sampling interval of 20ms, covering 200ms before the mutation to 500ms after the mutation, for a total duration of 700ms and 36 sampling points. Statistical analysis is performed on the real-time power change sequences to extract power transition amplitude, fluctuation duration, and fluctuation concentration. Power transition amplitude is the absolute value of the difference between the steady-state average power after the mutation and the steady-state average power before the mutation in the power change sequence of a power supply unit before and after the mutation. Fluctuation duration is the total time from the moment the mutation occurs until the power change sequence re-enters a new steady state. Fluctuation concentration describes the degree of concentration of power fluctuations among power supply units within the propagation range, which can be quantified by the variance of the fluctuation energy distribution in each level of power supply unit. A higher value indicates that the fluctuation is concentrated at a lower level (closer to the task end), and a lower value indicates that the fluctuation has spread to a higher level (closer to the transformer end).

[0054] The propagation range and local power fluctuation characteristics of the corresponding power supply and demand pressure are integrated and encapsulated into a power impact characterization result, which is a structured data object containing complete information such as propagation range, power transition amplitude at each level, fluctuation duration, fluctuation concentration and pressure diffusion coefficient.

[0055] Functional partitioning is performed by using the results of electrical energy impact characterization and the operating status data of the energy storage system to generate an energy storage functional configuration structure that is adapted to the results of electrical energy impact characterization. The energy storage functional configuration structure includes a transient power vibration suppression unit, a time-shifted energy smoothing unit, and a local power supply compensation unit.

[0056] Furthermore, this application also includes the following steps: based on the local power fluctuation characteristics in the power impact characterization results, the power impact is divided into transient impact components, continuous fluctuation components, and local off-center load components, and the energy regulation demand type corresponding to each component is determined respectively; the energy storage system operation status data is decomposed to construct an energy storage adjustable capability space, wherein the energy storage adjustable capability space includes at least instantaneous power response capability, energy time-shift regulation capability, and local power supply output capability, and an energy storage capability vector set is established; based on the matching relationship between the energy regulation demand type and the energy storage capability vector set, the matching fit degree between the power impact component and the energy storage unit is calculated; the energy storage unit is functionally partitioned according to the fit degree to generate an energy storage functional partition set corresponding to different power impact components; and corresponding energy storage regulation execution mechanisms are configured in the energy storage functional partition set to generate an energy storage functional configuration structure.

[0057] Furthermore, this application also includes the following steps: configuring a transient power vibration suppression unit for the energy storage zone corresponding to the transient impact component, for performing short-time high-power fast response regulation; configuring a time-shifted energy smoothing unit for the energy storage zone corresponding to the continuous fluctuation component, for performing energy timing migration and fluctuation smoothing control; and configuring a local power supply compensation unit for the energy storage zone corresponding to the local off-center load component, for performing local power supply redistribution and load balancing compensation.

[0058] Specifically, local power fluctuation characteristics are extracted from the power impact characterization results, and the power impact is divided into transient impact components, continuous fluctuation components, and local off-center load components. The transient impact component is determined by the part with the fastest rate of change in the power transition amplitude, typically manifested as a pulse-like spike. Its amplitude is defined by the magnitude of the high-frequency components in the power transition amplitude, and its duration is defined by the components in the fluctuation duration that are less than 100ms. The continuous fluctuation component consists of the remaining stable deviation after deducting the transient component from the power transition amplitude, and its duration corresponds to the components in the fluctuation duration that are greater than 100ms. The local off-center load component is determined by the spatial distribution information of the fluctuation concentration. When the entropy value is low, i.e., the fluctuation concentration is less than 0.6, it indicates that the fluctuation energy is concentrated in a certain local power supply level, and the power transition amplitude of this concentrated part is classified as the local off-center load component. For each component, a corresponding energy regulation demand type is determined, including the required response time, power amplitude, and duration, such as the instantaneous power support demand for transient impacts, the energy time-shifting demand for continuous fluctuations, and the local power supply redistribution demand for local off-center loads.

[0059] Real-time operating status data of each energy storage unit is read from the coordinated state representation, including State of Charge (SOC), State of Health (SOH), current maximum discharge / chargeable power, remaining available energy, and connection bus location. A three-dimensional capability vector is constructed for each energy storage unit, including instantaneous power response capability, energy time-shift regulation capability, and local power output capability. Instantaneous power response capability represents the maximum power that the energy storage unit can release or absorb in a short period, determined by the energy storage type and current SOC. Energy time-shift regulation capability represents the total energy that the energy storage unit can process over a longer time scale, determined by the energy storage capacity and SOH. Local power output capability represents the spatial accessibility of the energy storage unit for local power replenishment, determined by its electrical distance from the connection bus and its power supply radius. The capability vectors of all energy storage units are aggregated to form an energy storage capability vector set.

[0060] For each type of energy regulation demand, it is transformed into a three-dimensional demand vector, namely demand response power, demand throughput energy, and demand supply radius. The normalized cosine similarity is calculated between this vector and the capacity vector of each energy storage unit. The resulting similarity value represents the matching degree between the energy storage unit and the impact component, with a value range of [0,1]. For example, the demand vectors corresponding to transient impact components, continuous fluctuation components, and local off-center load components, after normalization, have the following values: transient impact demand vector: 0.95, 0.1, 0.5; continuous fluctuation demand vector: 0.6, 0.9, 0.5; local off-center load demand vector: 0.7, 0.6, 0.95. The instantaneous power response capability, energy time-shift regulation capability, and local power output capability corresponding to the capability vector of a certain energy storage cluster are normalized to 0.7, 0.85, and 0.6, respectively. The cosine similarity between the cluster and the three types of demand is calculated. The fit with transient impact is 0.777, the fit with continuous fluctuation is 0.993, and the fit with local off-center load is 0.946. The maximum fit with continuous fluctuation is 0.993, and the cluster is classified into the continuous fluctuation partition.

[0061] The energy storage units are functionally partitioned according to the principle of maximum fit priority. For each energy storage unit, its fit with three impact components—transient, sustained, and local—is compared, and it is assigned to the functional partition corresponding to the impact component with the highest fit. When the fit of an energy storage unit with multiple components is close (e.g., the difference < 0.1), the standby strategy is invoked to assign it to the partition with the most scarce capacity under the current configuration, thus achieving load balancing. Specifically, all energy storage units are first pre-allocated according to the conventional maximum fit priority principle, and the total amount of energy storage resources currently obtained by each functional partition is recorded, including total available power and total available energy. The three fit values ​​of each energy storage unit are checked one by one. If the difference between the fit values ​​of any two components is < 0.1, the energy storage unit is marked as a unit with a fit conflict, and the standby strategy is triggered. For the three functional partitions, their scarcity index is calculated, which is the ratio of the total demand capacity of the partition to the currently allocated capacity. The larger the value, the more scarce the current configuration of the partition. The marked energy storage units are assigned to the functional partition with the highest scarcity. If the scarcity values ​​are the same, priority will be given to allocating to the transient power suppression zone, because transient shocks have the most stringent response time requirements, and the consequences of insufficient supply are the most severe, potentially leading to equipment power outages. The allocated energy storage units will then be removed from the candidate lists of other zones, and the total available power and total available energy values ​​for each zone will be updated. The scarcity level for each zone will be recalculated for the next round of allocation decisions.

[0062] After determining the set of energy storage functional zones, a corresponding energy storage regulation execution mechanism is configured for each functional zone. For energy storage zones corresponding to transient impact components, a transient power suppression unit is configured, employing high-power-density energy storage devices such as supercapacitors to perform short-term high-power rapid response regulation. For energy storage zones corresponding to continuous fluctuation components, a time-shifted energy smoothing unit is configured, employing high-energy-density energy storage devices such as lithium batteries to perform energy time-series migration and fluctuation smoothing control. For energy storage zones corresponding to local off-center load components, a local power supply compensation unit is configured, employing locally deployed distributed energy storage devices to perform local power supply redistribution and load balancing compensation.

[0063] The controller parameters of energy storage devices assigned to the transient power suppression zone are uniformly configured to transient suppression mode. The core function is to absorb or release short-duration high-power pulses within a millisecond timescale, directly suppressing the impact of power spikes on the bus voltage. Key parameters of the transient power suppression unit include a voltage outer loop response bandwidth of 500Hz or higher, a current inner loop response time of <2ms, a power change rate limit set to the maximum allowable value for the device (typically 200%~300% of rated power), and an activation threshold set to trigger immediately when the bus voltage change rate exceeds 5V / ms or the power change rate exceeds 0.5kW / ms. When a transient impact is detected, power is injected into or absorbed from the bus within <10ms at maximum power output capacity (e.g., 680kW), for a duration equal to the impact duration (e.g., 60ms), thus offsetting the impact energy locally before it is transmitted to the upstream transformer.

[0064] The controller parameters of energy storage devices assigned to the time-shifted energy smoothing zone are uniformly configured to smooth tracking mode. The core function is to smooth out power peaks and valleys during abrupt changes through charging and discharging operations over a longer timescale. The core parameters of the time-shifted energy smoothing unit include a response time set to <100ms, a power change rate limit set to <5kW / ms, a charging / discharging power upper limit set to 80% of the lithium battery's rated power, and a control target set to reduce the power change rate to <0.05kW / ms. When a sustained fluctuation is detected, it is activated after the transient vibration suppression unit completes its operation, i.e., 60ms after the transient impact is absorbed. Through a peak-shaving and valley-filling strategy, when the bus power is higher than the target value, charging absorbs the surplus; when the bus power is lower than the target value, discharging replenishes the deficit, thus smoothing out the bus power change until the fluctuation ends, such as at 240ms.

[0065] The controller parameters of energy storage devices assigned to local power supply compensation zones are uniformly configured to local balancing mode. The core function is to perform power compensation locally at specific power supply nodes, resolving local overload or underload issues caused by uneven task scheduling leading to significant power deviations from the average at specific load points. The core parameters of the local power supply compensation unit include setting the load rate of the power supply node to which the energy storage device is connected as the local control target. The target load rate is set to 70%~75% of the node's rated capacity, and the control dead zone is set to ±3% of the target value. When the load rate of the node deviates from the target value beyond the dead zone, the local power supply compensation unit activates locally. When the load is too heavy, it discharges to compensate; when the load is too light, it charges to absorb the load, bringing the node's load rate back to the target range and achieving local load balancing. The response time of the local power supply compensation unit is set to <50ms to ensure compensation is completed before uneven load causes line overheating or voltage drops.

[0066] The three functional zones are configured as corresponding energy storage execution units: a transient power suppression unit, a time-shifted energy smoothing unit, and a local power supply compensation unit. For each unit, a list of included energy storage device IDs, total available power, total available energy, scheduling priority, and response time specifications are clearly listed and integrated into a standardized energy storage function configuration structure data object. The electrical characteristics of the energy storage units are adapted to the characteristics of power surges. Through dynamic partitioning, the same group of energy storage hardware can play different roles at different times. The precise intervention of the local power supply compensation unit effectively suppresses three-phase imbalance and voltage flicker caused by high-concentration loads, prevents cascading accidents caused by single-point failures, and significantly enhances the resilience of the intelligent computing center's power supply.

[0067] Based on the computing power structure mutation event, the power impact characterization results and the energy storage function configuration structure, a power storage collaborative control strategy is generated. The control strategy is parsed into pulse width modulation signals and switching instructions, which are then sent to the energy storage converter and power distribution unit to perform power regulation.

[0068] Furthermore, this application also includes the following steps: when the energy storage system operation status data indicates that some energy storage units are unavailable or their health status is below a safety threshold, a functional partition reallocation instruction is generated; based on the functional partition reallocation instruction, under the constraint of ensuring the continuity of power supply for critical computing tasks, the transient power vibration suppression, time-shifted energy smoothing, and local power supply compensation functions are downgraded.

[0069] Specifically, using the sudden change event in computing power structure as the time base, the power impact characterization result as the control objective, and the energy storage function configuration structure as the resource constraint, a time-power two-dimensional collaborative control strategy table is comprehensively generated. This table is divided into multiple control stages along the time axis, with each stage clearly defining the start time, end time, functional partition name for the action, target output power value, power change rate limit, and collaborative constraint markers. The time resolution of the time-power two-dimensional collaborative control strategy table is consistent with the collaborative state characterization, covering the entire process. Using the duration of the fluctuation as the total control period, the power transition amplitude is allocated to the output curves of each functional partition along the time axis, ensuring that the sum of the outputs of each partition equals the total target compensation power, and that the output of each partition does not exceed its available power limit.

[0070] For each time-power instruction in the time-power two-dimensional coordinated control strategy table, signal parsing and conversion are performed. For example, for a PWM signal, the target power value at that moment is read, and combined with the DC bus voltage, AC side line voltage RMS value, and rated capacity of the corresponding energy storage converter, the required modulation index M is calculated, with a value ranging from 0 to 1. The modulation index is then converted into the PWM duty cycle D. The duty cycle D and the modulation index M have a linear relationship under sinusoidal pulse width modulation (SPWM), usually D=(1+M) / 2. For example, if the target power is 54.70kW and the PCS rated capacity is 350kW, then the modulation index M=54.70 / 350=0.156, and the duty cycle D=0.578, i.e., 58.8%. Then, a square wave signal with a duty cycle of 58.8% is generated with a fixed carrier frequency, typically 10kHz, and a period of 100µs. For switching instructions, an N-bit (N equals the total number of energy storage clusters) contactor on / off bitmap is generated based on the partitioned device list in the energy storage function configuration structure. 1 indicates closed, and 0 indicates open. For example, if clusters 7 and 8 need to be activated in this strategy, then bits 7 and 8 of the bitmap are set to 1, and the rest are 0. The PWM duty cycle value and the switching bitmap together constitute a complete set of underlying drive instructions.

[0071] Pulse Width Modulation (PWM) signals are square wave electrical signals with a fixed frequency and adjustable duty cycle, used to drive the switching on and off of IGBTs inside the energy storage converter. The duty cycle directly determines the average voltage and power output of the energy storage converter; a larger duty cycle results in higher output power. Switching commands are digital instructions that control the on / off switching of contactors or solid-state relays within the power distribution unit, used to physically connect or disconnect the electrical connection between a specific energy storage cluster and the busbar.

[0072] The pulse width modulation (PWM) signal and switching instructions are encapsulated into standard data frames according to the industrial Ethernet communication protocol and sent to the local controllers of each energy storage converter and the switch input modules of the power distribution unit. A timestamp synchronization mechanism is used, with each data frame containing an absolute execution time, ensuring that each energy storage converter and power distribution unit operates simultaneously or in a predetermined sequence under the same global time base. After receiving the PWM signal duty cycle, each energy storage converter updates the IGBT drive pulse width at the start of the next carrier cycle, achieving continuous power regulation. After receiving the switch bitmap, each power distribution unit completes the contactor engagement or disengagement operation within 5ms of receiving the instruction, completing the activation or deactivation of the energy storage cluster. The total delay from strategy generation to the start of physical power change is controlled within 15ms, ensuring compatibility with the millisecond-level rapid changes in computing power.

[0073] The system scans the operational status data of all energy storage units in the collaborative status characterization at fixed intervals. For each unit, communication status verification and health status threshold verification are performed. If no heartbeat data or status refresh frame is received from a certain energy storage unit for three consecutive scan cycles, it is marked as unusable. If the health status SOH value of a unit is lower than a preset safety threshold, such as 80%, it is marked as having limited health status SOH, and its maximum discharge / chargeable power must be linearly reduced according to the percentage of health status SOH. For example, for a unit with SOH=75%, its usable power limit is reduced to 75% of the rated value. All abnormal units are marked and their failure type and remaining capacity are recorded to form an abnormal unit list.

[0074] The abnormal unit is removed from the current energy storage function configuration structure. The remaining available power and energy of the three functional zones are recalculated and compared with the demand target of this power surge. The capacity gap of each zone is calculated, which is the demand power / energy minus the remaining available value. A positive gap indicates that the zone's capacity is insufficient and needs to be allocated from other zones or accepted for degradation.

[0075] Based on the gap calculation results, a zone redistribution decision is executed. Priority is given to ensuring the capacity of the transient power suppression zone, as uncontrolled transient shocks will directly cause voltage drops below the server disconnection threshold, posing the most direct threat to power supply continuity. After the transient zone gap is filled, if there are still remaining healthy energy storage units, they are preferentially allocated to the time-shifted energy smoothing zone. Local zones, as the lowest priority, can have resources borrowed from other zones when resources are insufficient, or receive remaining allocations when resources are abundant. The specific redistribution method involves allocating some energy storage units from non-core zones with good health and relatively high adaptability to the zone with the largest capacity gap, and updating the zone affiliation relationships in the collaborative mapping table.

[0076] If, after redistribution, the capacity gap in some zones still cannot be fully filled—meaning the total capacity of the healthy energy storage units is insufficient to meet all impact demands—then a downgrade configuration will be implemented, prioritizing a reduction in the effective range and balancing accuracy of the local power supply compensation units. If this is still insufficient, the duration of the time-shifted energy smoothing units will be shortened, while the response time and suppression amplitude of the transient power vibration suppression units will be increased. All downgrade parameters will be recorded in a downgrade configuration list.

[0077] The reassigned partition affiliation and degraded configuration list are integrated to generate a new degraded energy storage function configuration structure to replace the current operating configuration. This changes take effect immediately, driving subsequent PWM signal generation and command issuance processes. Simultaneously, an alarm notification is sent to the upper-level operation and maintenance platform, prompting manual intervention for maintenance or replacement of faulty energy storage units.

[0078] In summary, the intelligent control method for computing, power, and energy storage collaboration provided in this application has the following technical effects: it performs spatiotemporal unified mapping of multi-source heterogeneous data to construct a collaborative state representation, identifies computing power structural mutation events and structural disturbance intensity, maps the propagation range and local power fluctuation characteristics of power supply and demand pressure level by level along the power supply physical layer, decomposes the power impact, and performs dynamic functional partitioning matching with the energy storage system, directly parses the comprehensive control strategy into pulse width modulation signals and switching instructions and sends them to the underlying physical devices for execution, realizes the targeted one-to-one correspondence between the energy storage functional characteristics and different power impact components, and ensures the power supply continuity and power quality stability of the intelligent computing center under high dynamic computing load.

[0079] Example 2: Based on the same inventive concept as the computing-power-storage collaborative intelligent control method for intelligent computing centers in Example 1, this application also provides a computing-power-storage collaborative intelligent control system for intelligent computing centers. Please refer to the appendix. Figure 3The aforementioned intelligent control system for computing, power, and energy storage in a smart computing center includes: a running data processing module 11, used to read the running data of the smart computing center, including computing task execution status data, computing power resource occupancy data, power supply and demand status data, and energy storage system running status data; after time alignment and spatial mapping of the running data, a collaborative state representation is constructed; a time-series structure analysis module 12, used to perform time-series structure analysis on the computing task execution status data and computing power resource occupancy data based on the collaborative state representation, identify computing power structure mutation events in the task parallel relationship and resource occupancy topology of the computing load at the current moment, and mark the intensity of the structural disturbance; and a data mapping module 13, used to map computing power structure mutation events based on the intensity of the structural disturbance. Mapping to power supply and demand status data, identifying the propagation range and local power fluctuation characteristics of corresponding power supply and demand pressures, and establishing power impact characterization results; Functional partition matching module 14 is used to perform functional partition matching with energy storage system operating status data using power impact characterization results, and generate an energy storage functional configuration structure adapted to the power impact characterization results. The energy storage functional configuration structure includes a transient power vibration suppression unit, a time-shifted energy smoothing unit, and a local power supply compensation unit; Strategy generation module 15 is used to generate an energy storage collaborative control strategy based on computing power structure mutation events, power impact characterization results, and energy storage functional configuration structure, and parse the control strategy into pulse width modulation signals and switching instructions, and send them to the energy storage converter and distribution unit to perform power regulation.

[0080] Furthermore, the operation data processing module 11 in the intelligent control system for computing, power, and energy storage in a computing center is also used to: align the timestamps of computing task execution status data, computing power resource occupancy data, power supply and demand status data, and energy storage system operation status data using a unified clock source; and map computing power resource nodes, power supply units, and energy storage units to the same spatial coordinate system based on the physical deployment topology of the intelligent computing center, thereby constructing a computing-power-energy storage collaborative mapping table.

[0081] Furthermore, the time-series structure parsing module 12 in the intelligent control system for computing, power, and storage collaboration oriented towards intelligent computing centers is also used for: responding to the scheduling cycle division of computing task execution status data, constructing a task parallel relationship graph in each time slice, wherein the nodes of the task parallel relationship graph represent computing task instances, the edges represent parallel execution relationships or resource competition relationships between tasks, and assigning corresponding resource occupancy vectors to each node based on computing power resource occupancy data; calculating the graph structure difference degree based on the task parallel relationship graphs of adjacent time slices, wherein the graph structure difference degree includes at least the node activity change rate, the edge connection relationship reconstruction ratio, and the resource occupancy vector distribution offset; when the graph structure difference degree exceeds a preset mutation judgment threshold, marking the corresponding time slice as the time when the computing power structure mutation event occurs; constructing a structural disturbance intensity index based on the computing power structure mutation event, wherein the structural disturbance intensity index is constructed by normalizing and weighting the node activity change rate, the edge connection relationship reconstruction ratio, and the resource occupancy vector distribution offset.

[0082] Furthermore, the time-series structure parsing module 12 in the intelligent control system for computing, power, and storage in a computing center is also used to: obtain the resource occupancy vector distribution offset by performing a weighted centroid migration calculation on the resource occupancy vector of each node.

[0083] Furthermore, the data mapping module 13 in the intelligent control system for computing, power, and storage collaboration oriented towards intelligent computing centers is also used for: power mapping association of the set of associated computing task instances based on the structural disturbance intensity corresponding to the computing power structural mutation event, establishing a mapping correspondence between task nodes and power supply units, wherein the power supply unit includes at least a rack-level power distribution unit, a node-level power supply module, and a bus distribution unit; decomposing the structural disturbance intensity into node-level power disturbance factors according to the mapping correspondence, and performing step-by-step propagation calculations along the power supply unit hierarchy according to the topological position of the task in the task parallel relationship graph, generating a set of power supply and demand pressure propagation paths; calculating the pressure diffusion coefficient based on the load accumulation effect of the path nodes on the set of power supply and demand pressure propagation paths, and configuring the propagation range of power supply and demand pressure using the pressure diffusion coefficient; extracting local power fluctuation features from the real-time power change sequence of each power supply unit within the propagation range, wherein the local power fluctuation features include at least the power transition amplitude, fluctuation duration, and fluctuation concentration.

[0084] Furthermore, the data mapping module 13 in the intelligent control system for computing, power storage and computing centers is also used to: calculate the attenuation factor or amplification factor of the pressure on the power supply and demand pressure propagation path based on the rated capacity, current load rate and line impedance characteristics of each power supply unit; and accumulate the attenuation factor or amplification factor of each level to establish a pressure diffusion coefficient.

[0085] Furthermore, the functional partition matching module 14 in the intelligent control system for computing, power, and energy storage in a smart computing center is also used for: dividing the power impact into transient impact components, continuous fluctuation components, and local off-center load components based on the local power fluctuation characteristics in the power impact characterization results, and determining the energy regulation demand type corresponding to each component; decomposing the energy storage system operation status data to construct an energy storage adjustable capability space, wherein the energy storage adjustable capability space includes at least instantaneous power response capability, energy time-shift regulation capability, and local power supply output capability, and establishing an energy storage capability vector set; calculating the matching fit degree between the power impact component and the energy storage unit based on the matching relationship between the energy regulation demand type and the energy storage capability vector set; dividing the energy storage unit into functional partitions according to the fit degree, generating an energy storage functional partition set corresponding to different power impact components; configuring corresponding energy storage regulation execution mechanisms in the energy storage functional partition set, and generating an energy storage function configuration structure.

[0086] Furthermore, the functional partition matching module 14 in the intelligent control system for computing, power, and storage collaboration of the intelligent computing center is also used to: configure a transient power vibration suppression unit for the energy storage partition corresponding to the transient impact component, for performing short-time high-power fast response adjustment; configure a time-shifted energy smoothing unit for the energy storage partition corresponding to the continuous fluctuation component, for performing energy time-series migration and fluctuation smoothing control; and configure a local power supply compensation unit for the energy storage partition corresponding to the local off-center load component, for performing local power supply redistribution and load balancing compensation.

[0087] Furthermore, the strategy generation module 15 in the intelligent control system for computing, power, and energy storage collaboration for intelligent computing centers is also used to: generate a functional partition reallocation instruction when the energy storage system operation status data indicates that some energy storage units are unavailable or their health status is below a safety threshold; and, based on the functional partition reallocation instruction, degrade the transient power suppression, time-shifted energy smoothing, and local power supply compensation functions under the constraint of ensuring the continuity of power supply for critical computing tasks.

[0088] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The computing, power, and storage collaborative intelligent control method and specific examples for intelligent computing centers described in Embodiment 1 above are also applicable to the computing, power, and storage collaborative intelligent control system for intelligent computing centers in this embodiment. Through the foregoing detailed description of the computing, power, and storage collaborative intelligent control method for intelligent computing centers, those skilled in the art can clearly understand the computing, power, and storage collaborative intelligent control system for intelligent computing centers in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0089] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0090] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A collaborative intelligent control method for computing, power, and storage in intelligent computing centers, characterized in that, The method includes: Read the operational data of the intelligent computing center, which includes computing task execution status data, computing power resource occupancy data, power supply and demand status data, and energy storage system operation status data. After time alignment and spatial mapping processing of the operational data, construct a collaborative status representation. Based on the collaborative state representation, the temporal structure of the computing task execution state data and computing resource occupancy data is analyzed to identify the computing power structure mutation events in the task parallel relationship and resource occupancy topology of the computing load at the current moment, and to identify the intensity of the structural disturbance. Based on the structural disturbance intensity, the computing power structural mutation event is mapped to the power supply and demand status data, the propagation range and local power fluctuation characteristics of the corresponding power supply and demand pressure are identified, and the power impact characterization results are established. Functional partitioning is matched by the results of electrical energy impact characterization and the operating status data of the energy storage system to generate an energy storage functional configuration structure that is adapted to the results of electrical energy impact characterization. The energy storage functional configuration structure includes a transient power vibration suppression unit, a time-shifted energy smoothing unit, and a local power supply compensation unit. Based on the computing power structure mutation event, the power impact characterization results and the energy storage function configuration structure, a power storage collaborative control strategy is generated. The control strategy is parsed into pulse width modulation signals and switching instructions, which are then sent to the energy storage converter and power distribution unit to perform power regulation. Based on the structural disturbance intensity, abrupt structural events of computing power are mapped to power supply and demand state data. The propagation range and local power fluctuation characteristics of the corresponding power supply and demand pressures are identified, and power impact characterization results are established, including: Based on the power mapping association of the set of computing task instances corresponding to the structural disturbance intensity of the computing power structure mutation event, a mapping correspondence between task nodes and power supply units is established. The power supply unit includes at least a rack-level power distribution unit, a node-level power supply module, and a bus distribution unit. Based on the mapping correspondence, the structural disturbance intensity is decomposed into node-level power disturbance factors, and according to the topological position of the task in the task parallel relationship diagram, the propagation calculation is carried out step by step along the power supply unit hierarchical structure to generate a set of power supply and demand pressure propagation paths. The pressure diffusion coefficient is calculated based on the load accumulation effect of the path nodes on the set of power supply and demand pressure propagation paths, and the propagation range of power supply and demand pressure is configured using the pressure diffusion coefficient. Local power fluctuation features are extracted from the real-time power change sequence of each power supply unit within the propagation range. The local power fluctuation features include at least the power transition amplitude, fluctuation duration, and fluctuation concentration.

2. The intelligent control method for computing, power, and storage collaboration for intelligent computing centers as described in claim 1, characterized in that, Identify abrupt changes in the computing power structure in terms of task parallelism and resource occupancy topology at the current moment, and characterize the intensity of these structural disturbances, including: In response to the scheduling cycle division based on the computing task execution status data, a task parallel relationship graph is constructed on each time slice. In the task parallel relationship graph, the nodes represent computing task instances, and the edges represent the parallel execution relationship or resource competition relationship between tasks. Based on the computing power resource usage data, each node is assigned a corresponding resource usage vector. The graph structure difference is calculated based on the task parallel relationship graph of adjacent time slices. The graph structure difference includes at least the node activity change rate, the edge connection relationship reconstruction ratio and the resource occupation vector distribution offset. When the graph structure difference exceeds the preset mutation judgment threshold, the corresponding time slice is marked as the time when the computing power structure mutation event occurs. Based on the aforementioned computing power structural mutation event, a structural disturbance intensity index is constructed. The structural disturbance intensity index is constructed by normalizing and weighting the node activity change rate, the edge connection relationship reconstruction ratio, and the resource occupation vector distribution offset.

3. The intelligent control method for computing, power, and storage collaboration for intelligent computing centers as described in claim 2, characterized in that, The resource occupancy vector distribution offset is obtained by performing a weighted centroid migration calculation on the resource occupancy vector of each node.

4. The intelligent control method for computing, power, and storage collaboration in intelligent computing centers as described in claim 1, characterized in that, The pressure diffusion coefficient is calculated based on the load accumulation effect of path nodes on the set of power supply and demand pressure propagation paths, including: Based on the rated capacity, current load rate and line impedance characteristics of each power supply unit, the attenuation factor or amplification factor of the pressure on that power supply unit is calculated along the power supply and demand pressure propagation path. The pressure diffusion coefficient is established by accumulating the attenuation factors or amplification factors at each level.

5. The intelligent control method for computing, power, and storage collaboration for intelligent computing centers as described in claim 1, characterized in that, Functional partitioning is performed using the results of electrical energy impact characterization and the operating status data of the energy storage system to generate an energy storage functional configuration structure adapted to the results of electrical energy impact characterization, including: Based on the local power fluctuation characteristics in the power impact characterization results, the power impact is divided into transient impact component, continuous fluctuation component and local off-center load component, and the energy regulation demand type corresponding to each component is determined respectively. The energy storage system's operating status data is decomposed to construct an energy storage adjustable capability space, which includes at least instantaneous power response capability, energy time-shift adjustment capability, and local power supply output capability, and an energy storage capability vector set is established. The matching degree between the electrical energy surge component and the energy storage unit is calculated based on the matching relationship between the energy regulation demand type and the energy storage capacity vector set. The energy storage units are functionally partitioned according to their adaptability, generating a set of energy storage functional partitions corresponding to different electrical energy impact components. Within the set of energy storage functional zones, configure the corresponding energy storage regulation and execution mechanisms to generate the energy storage functional configuration structure.

6. The intelligent control method for computing, power, and storage collaboration for intelligent computing centers as described in claim 5, characterized in that, Within each energy storage functional zone set, a corresponding energy storage regulation execution mechanism is configured, including: Transient power vibration suppression units are configured for the energy storage zones corresponding to transient impact components to perform short-time high-power rapid response regulation; A time-shifted energy smoothing unit is configured for the energy storage partition corresponding to the continuous fluctuation component, which is used to perform energy time-series migration and fluctuation smoothing control; Local power supply compensation units are configured for energy storage zones corresponding to local off-center load components to perform local power supply redistribution and load balancing compensation.

7. The intelligent control method for computing, power, and storage collaboration for intelligent computing centers as described in claim 1, characterized in that, The processing of runtime data time alignment and spatial mapping includes: A unified clock source is used to align the timestamps of computing task execution status data, computing resource usage data, power supply and demand status data, and energy storage system operation status data. The physical deployment topology based on the intelligent computing center maps computing resource nodes, power supply units and energy storage units to the same spatial coordinate system, and constructs a computing-power-storage collaborative mapping table.

8. The intelligent control method for computing, power, and storage collaboration for intelligent computing centers as described in claim 1, characterized in that, Based on the results of computing power structural mutation events, power impact characterization, and energy storage functional configuration structure, a collaborative control strategy for energy storage and electricity generation is generated, including: When the energy storage system's operating status data indicates that some energy storage units are unavailable or their health status is below a safe threshold, a functional partition reallocation instruction is generated. Based on the functional partitioning reallocation instructions, and under the constraint of ensuring the continuity of power supply for critical computing tasks, the transient power suppression, time-shifted energy smoothing, and local power supply compensation functions are downgraded.

9. A computing, power, and storage collaborative intelligent control system for intelligent computing centers, characterized in that: The steps for implementing the intelligent computing center-oriented computing, power, and storage collaborative intelligent control method according to any one of claims 1 to 8, wherein the intelligent computing center-oriented computing, power, and storage collaborative intelligent control system comprises: The data processing module is used to read the operational data of the intelligent computing center. The operational data includes computing task execution status data, computing power resource occupancy data, power supply and demand status data, and energy storage system operation status data. After time alignment and spatial mapping processing of the operational data, a collaborative status representation is constructed. The temporal structure parsing module is used to perform temporal structure parsing on the execution status data and computing resource occupancy data of computing tasks based on the cooperative state representation, identify computing load in the task parallel relationship and resource occupancy topology at the current moment, and identify the intensity of structural disturbance. The data mapping module is used to map computing power structural mutation events to power supply and demand status data based on structural disturbance intensity, identify the propagation range and local power fluctuation characteristics of the corresponding power supply and demand pressure, and establish power impact characterization results. The functional partition matching module is used to perform functional partition matching with the energy storage system operation status data using the results of the power impact characterization, and generate an energy storage functional configuration structure that is adapted to the results of the power impact characterization. The energy storage functional configuration structure includes a transient power vibration suppression unit, a time-shifted energy smoothing unit, and a local power supply compensation unit. The strategy generation module is used to generate an energy storage collaborative control strategy based on the computing power structure mutation event, the power impact characterization result and the energy storage function configuration structure. The control strategy is parsed into pulse width modulation signals and switching instructions, and sent to the energy storage converter and power distribution unit to perform power regulation.

Images