A power supply system for high-altitude wind power AI servers based on magnetic components

Abstract

This invention relates to the field of integrated power supply technology, and in particular to a high-altitude wind power AI server power supply system based on integrated magnetic components. The system includes: converting wind power output fluctuations into a visualized magnetic energy distribution map to capture the spatiotemporal dynamics of energy and provide a data foundation for subsequent scheduling; locking the time interval for supply-demand matching through multi-dimensional matching of wind power energy supply and AI computing load, reducing energy loss and power supply fluctuation risks caused by supply-demand mismatch; generating differentiated magnetic control drive commands on demand to provide control basis for subsequent layered energy release, ensuring that the magnetic energy release rhythm matches the dynamic changes in load demand; improving energy transmission efficiency and reducing redundant losses in the power supply link by prioritizing and directionally allocating magnetic energy supply paths; and ensuring stable and efficient long-term operation of the system under complex conditions by real-time correction of magnetic component coupling deviations and synchronous updating of the energy map.

Description

Technical Field

[0001] This invention relates to the field of integrated power supply technology, and in particular to a high-altitude wind power AI server power supply system based on magnetic components. Background Technology

[0002] High-altitude wind power systems operate in the strong wind belt of the troposphere, where wind speed and turbulence intensity are significantly higher than in the near-surface environment, resulting in high-frequency random fluctuations in the output power of the power generation units. AI server clusters, as typical computationally intensive loads, exhibit pulse-like power consumption curves with alternating steep rising and falling edges as training tasks iterate and inference requests arrive suddenly, placing stringent demands on power supply voltage stability and power tracking response speed.

[0003] However, when the high-frequency random fluctuations in high-altitude wind power output are superimposed on the power consumption abrupt changes of the AI ​​server load over time, existing solutions face fundamental limitations: power electronic converters, limited by switching frequency and loop bandwidth, cannot complete the entire power matching process from wind energy capture to load distribution within milliseconds; lithium battery packs, due to limitations in electrochemical reaction rates, struggle to keep up with power fluctuations on the order of seconds, resulting in frequent shallow charging and discharging that accelerates cycle life decay; while supercapacitors possess rapid response capabilities, their energy density is too low to independently handle deep power support tasks. Consequently, under the dual uncertainties of drastic fluctuations in source-side output and rapid changes in load-side demand, the DC bus voltage frequently exceeds limits, triggering discharge protection or frequency reduction of computing nodes. Furthermore, existing solutions lack a unified state representation and scheduling coordination mechanism between the energy buffer unit and both the power generation and consumption sides, and the management strategy of the energy storage unit is disconnected from the dynamic changes in load demand, making it difficult for the system to achieve a balance between fluctuation suppression and energy utilization efficiency. To address this, this invention proposes a high-altitude wind power AI server power supply system based on integrated magnetic components. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the background art and to propose a high-altitude wind power AI server power supply system based on magnetic component integration.

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides a high-altitude wind power AI server power supply system based on magnetic component integration, comprising: Energy Tide Mapping Module: Maps the collected real-time power output fluctuation data to the hysteresis loop trajectory pattern of each magnetic element unit in the magnetic element integrated array, and constructs the energy tide map. Load Resonance Window Identification Module: Acquires real-time load demand data of each computing node in the AI ​​server cluster, and performs supply and demand coupling analysis in combination with the energy tide map to identify load resonance windows in the energy tide map that match the real-time load demand data in a preset dimension. Magnetic flux pulse sequence generation module: Based on the identification result of the load resonance window, it generates a magnetic flux pulse encoding sequence for driving the magnetic element integrated array to perform energy release, and performs hierarchical energy release scheduling for each magnetic element unit in the magnetic element integrated array. Layered energy release scheduling module: During the execution of layered energy release scheduling, it continuously monitors the change in magnetic coupling depth of each magnetic element unit in the magnetic element integrated array and determines whether to trigger an adaptive coupling reconstruction operation for the deviated magnetic element unit. Adaptive Coupling Reconstruction Update Module: After performing the adaptive coupling reconstruction operation, update the energy tidal map.

[0006] Furthermore, the preset dimensions include a time dimension and an energy dimension.

[0007] Furthermore, the process of mapping the collected real-time output fluctuation data to the hysteresis loop trajectory pattern of each magnetic element unit in the integrated array of magnetic elements, and constructing an energy tidal map, includes: Real-time power output fluctuation data is continuously collected through the output ports of each power generation unit in the high-altitude wind power array; the real-time power output fluctuation data includes the output power value and power change direction of each power generation unit at different sampling times; The real-time output fluctuation data of each power generation unit is mapped to the hysteresis loop trajectory pattern of the corresponding magnetic element unit in the magnetic element integrated array according to the preset output-magnetization mapping relationship; wherein, the hysteresis loop trajectory pattern represents the magnetization state and remaining magnetic energy reserve level of the corresponding magnetic element unit under the current output conditions. Based on the hysteresis loop trajectory patterns of all magnetic element units, energy tidal maps are generated by aggregating the spatial topological positions of each magnetic element unit in the integrated array of magnetic elements.

[0008] Further, the step of mapping the real-time output fluctuation data of each power generation unit to the hysteresis loop trajectory pattern of the corresponding magnetic element unit in the magnetic element integrated array according to a preset output-magnetization mapping relationship includes: The output power value sequence and power change direction sequence of each power generation unit within a continuous sampling period are jointly encoded to obtain the output characteristic code of each power generation unit. Acquire historical operating data and construct a mapping table between power output feature codes and hysteresis loop trajectory types to characterize the power output-magnetization mapping relationship; Based on the output-magnetization mapping relationship, find the target hysteresis loop trajectory type corresponding to the current output feature code; Based on the target hysteresis loop trajectory type, the magnetic element unit with energy receiving relationship with the corresponding power generation unit is determined in the magnetic element integrated array, and the target hysteresis loop trajectory type is assigned to the magnetic element unit as the current hysteresis loop trajectory mode of the magnetic element unit.

[0009] Furthermore, each node in the energy tidal map corresponds to a magnetic element unit, and the node state value of the map node is determined by the hysteresis loop trajectory mode of the corresponding magnetic element unit. The connection relationship between map nodes reflects the magnetic coupling path relationship between magnetic element units.

[0010] Furthermore, the step of acquiring real-time load demand data of each computing node in the AI ​​server cluster, and combining it with energy tidal maps to perform supply and demand coupling analysis, and identifying load resonance windows in the energy tidal maps that match the real-time load demand data in a preset dimension, includes: Obtain real-time load demand data for each computing node in the AI ​​server cluster within the prediction time window; the real-time load demand data includes the expected power consumption value and power consumption trend of each computing node. The expected power consumption value is matched with the node state value of each node in the energy tide map according to the energy dimension, and a candidate map node set is selected. In the candidate graph node set, the power consumption change trend in the real-time load demand data is matched with the node state value change trend of each graph node in the time dimension. Graph nodes whose node state value change trend is consistent with the power consumption change trend in the time dimension and whose change rhythm is synchronized are identified as target matching graph nodes. The time intervals corresponding to the target matching map nodes are combined with the magnetic element units to identify the load resonance window.

[0011] Further, the step of generating a magnetic flux pulse-coded sequence for driving the integrated array of magnetic elements to perform energy release based on the identification result of the load resonance window, and performing hierarchical energy release scheduling for each magnetic element unit in the integrated array of magnetic elements, includes: Based on the energy supply correspondence between the target matching spectrum nodes corresponding to the load resonance window and each computing node, the energy release priority of each magnetic element unit is determined. Based on the priority of energy release, a corresponding magnetic flux pulse coding sequence is generated for each magnetic element unit; wherein, the magnetic flux pulse coding sequence includes the pulse trigger time, pulse duration and pulse intensity level; According to the order of pulse triggering times in the magnetic flux pulse coding sequence, pulse drive signals are sent to the corresponding magnetic element units in sequence, driving each magnetic element unit to perform hierarchical energy release scheduling according to the pulse duration and pulse intensity level. The energy released by high-priority magnetic element units is preferentially delivered to computing nodes whose power consumption trend is in the rising edge stage of the load resonance window, while the energy released by low-priority magnetic element units is delivered to computing nodes whose power consumption trend is in the stable stage.

[0012] Furthermore, the continuous monitoring of the magnetic coupling depth change state of each magnetic element unit in the integrated magnetic element array, and the determination of whether to trigger an adaptive coupling reconstruction operation for the deviating magnetic element unit, includes: The real-time magnetic coupling depth value of each magnetic element unit in the magnetic element integrated array is continuously collected at a preset monitoring cycle. The real-time magnetic coupling depth value of each magnetic element unit is compared with the expected magnetic coupling depth value at the corresponding time in the expected coupling trajectory to obtain the coupling deviation state of each magnetic element unit. At the same time, based on the coupling deviation state, it is determined whether the magnetic element unit has deviated from the magnetic coupling depth; For a magnetic element unit that has deviated from its magnetic coupling depth, obtain the current magnetic coupling depth value and the remaining magnetic energy reserve level of its adjacent magnetic element units; Based on the current magnetic coupling depth and remaining magnetic energy reserve level of adjacent magnetic element units, a coupling compensation scheme is determined, and an adaptive coupling reconstruction operation is performed on the magnetic element units that have deviated according to the coupling compensation scheme.

[0013] Furthermore, the adaptive coupling reconfiguration operation includes adjusting the magnetic circuit connection between the deviated magnetic element unit and its adjacent magnetic element units, and redistributing the magnetic flux transmission path.

[0014] Furthermore, the updated energy tidal map includes: After the adaptive coupling reconstruction operation is completed, the updated hysteresis loop trajectory pattern of each magnetic element unit in the magnetic element integrated array is obtained. Based on the updated hysteresis loop trajectory pattern, the node state values ​​of the corresponding nodes in the energy tidal map are regenerated, and the connection relationships between the nodes that have changed due to the change in the magnetic circuit connection method are updated synchronously. The updated energy tide map is used as the input for the supply and demand coupling analysis in the next scheduling cycle, replacing the energy tide map of the previous scheduling cycle in the identification of the load resonance window. Based on the matching results between the updated energy tidal map and the real-time load demand data, a magnetic flux pulse coding sequence for the next scheduling cycle is generated.

[0015] Compared with existing technologies, the advantages of this invention in providing a high-altitude wind power AI server power supply system based on integrated magnetic components are as follows: 1) By mapping the power output fluctuations of high-altitude wind power arrays to hysteresis loop trajectories and constructing energy tidal maps, the magnetic energy distribution status at different times and spatial locations is characterized, thereby improving the utilization efficiency of intermittent wind power energy; by combining the real-time load demand of AI server clusters to complete multi-dimensional supply and demand coupling analysis, identifying load resonance windows, and enabling deep matching between the energy supply rhythm of the wind power side and the fluctuation trend of AI computing power load, the stability of computing power operation is ensured. 2) By relying on magnetic flux pulse coding sequences to complete layered energy release and allocating magnetic energy supply paths differently according to load priority, highly adaptable magnetic energy is preferentially delivered to nodes with multiple computing power requirements, realizing directional and efficient energy transmission and reducing unnecessary losses in the power supply link; by monitoring the changes in magnetic coupling depth of magnetic components throughout the entire cycle, abnormal units with coupling deviation can be identified in a timely manner, and adaptive coupling reconstruction operations can be quickly triggered, completing the dynamic adjustment of the magnetic circuit and magnetic flux path without shutdown, reducing the probability of failure under complex operating conditions; by synchronously updating the energy tidal map after reconstruction, state input is provided for the next round of scheduling, continuously adapting to the dynamic changes in high-altitude wind power output and AI computing power load, and long-term operating energy efficiency can be steadily improved. Attached Figure Description

[0016] Figure 1 This is a block diagram of a high-altitude wind power AI server power supply system based on magnetic components proposed in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides a high-altitude wind power AI server power supply system based on integrated magnetic components, comprising: Energy Tide Mapping Module: Maps the collected real-time power output fluctuation data to the hysteresis loop trajectory pattern of each magnetic element unit in the magnetic element integrated array, and constructs the energy tide map. Load Resonance Window Identification Module: Acquires real-time load demand data of each computing node in the AI ​​server cluster, and performs supply and demand coupling analysis in combination with energy tide map to identify load resonance windows in energy tide map that match the real-time load demand data in preset dimensions; wherein, preset dimensions include time dimension and energy dimension; Magnetic flux pulse sequence generation module: Based on the identification result of the load resonance window, it generates a magnetic flux pulse encoding sequence for driving the magnetic element integrated array to perform energy release, and performs hierarchical energy release scheduling for each magnetic element unit in the magnetic element integrated array. Layered energy release scheduling module: During the execution of layered energy release scheduling, it continuously monitors the change in magnetic coupling depth of each magnetic element unit in the magnetic element integrated array and determines whether to trigger an adaptive coupling reconstruction operation for the deviated magnetic element unit. Adaptive Coupling Reconstruction Update Module: After performing the adaptive coupling reconstruction operation, update the energy tidal map.

[0019] In this embodiment of the invention, the specific implementation method for mapping the collected real-time power output fluctuation data to the hysteresis loop trajectory pattern of each magnetic element unit in the magnetic element integrated array, and constructing an energy tidal map, is as follows: S11. Real-time power output fluctuation data is continuously collected through the output ports of each power generation unit in the high-altitude wind power array; wherein, the real-time power output fluctuation data includes the output power value and power change direction of each power generation unit at different sampling times; In the high-altitude wind power array, each power generation unit's energy conversion device output port is equipped with a power sensor and a direction discrimination component; the power sensor continuously collects the power output value of the power generation unit at a fixed sampling period of 10 milliseconds to ensure that it can capture the rapid fluctuations in power output caused by sudden changes in wind speed; each sampling moment records the raw data point containing the absolute timestamp and power value; Simultaneously, the direction discrimination component continuously compares the power value at the current sampling moment with the power value at the previous sampling moment; it calculates the difference between the current power value and the power value at the previous moment. If the absolute value of the difference exceeds the preset minimum fluctuation criterion threshold, the direction of power change is determined according to the sign of the difference: if the difference is positive, the direction of power change is determined to be upward, and the direction identifier is recorded as positive; if the difference is negative, the direction of power change is determined to be downward, and the direction identifier is recorded as negative; if the absolute value of the difference is within the range of the minimum fluctuation criterion threshold, the direction of power change is determined to be flat, and the direction identifier is recorded as zero. The minimum fluctuation criterion threshold is pre-calibrated based on the rated power of the power generation unit and the sensor noise level, taking 3 times the peak-to-peak value of the noise and converting it into a power percentage. In this embodiment, it is set to 0.5% of the rated power of the power generation unit to filter out false direction changes caused by sensor noise. The collected data is encapsulated into real-time power output fluctuation data packets. Each real-time power output fluctuation data packet contains the power generation unit number, the timestamp of the sampling time, the output power value at that time, and the power change direction identifier.

[0020] S12. The real-time output fluctuation data of each power generation unit is mapped to the hysteresis loop trajectory pattern of the corresponding magnetic element unit in the magnetic element integrated array according to the preset output-magnetization mapping relationship; wherein, the hysteresis loop trajectory pattern represents the magnetization state and remaining magnetic energy reserve level of the corresponding magnetic element unit under the current output conditions. S121. Jointly encode the output power value sequence and power change direction sequence of each power generation unit within a continuous sampling period to obtain the output characteristic code of each power generation unit. Specifically, after obtaining the real-time power output fluctuation data of each power generation unit, the data within the continuous sampling period is extracted and encoded in the form of a sliding time window; the length of the sliding time window is set to 2 seconds, which can cover the typical micro-scale cycle of wind power fluctuation; the window contains 200 sampling points and the sampling period is 10 milliseconds; For the sequence of 200 continuous output power value sampling points contained within the sliding time window, it is segmented and discretized: the power output range of the power generation unit from 10% to 100% of its rated power is divided into 10 equally spaced power intervals, and each power interval is assigned a power level represented by a number from 0 to 9; the arithmetic mean of the power values ​​within the window is calculated, and the power level to which the arithmetic mean falls is determined; at the same time, the standard deviation of the power change within the window is recorded as the fluctuation intensity level, and the fluctuation intensity level is divided into three levels: low, medium, and high, represented by the letters L, M, and H, respectively; Then, process the power change direction sequence: count the number of times the direction identifier is positive, negative, and zero within the window, and take the direction that appears most frequently as the dominant change direction of the window, using the letters U to represent rising, D to represent falling, and S to represent flat; if the number of occurrences is equal, the most recently occurring change direction is preferred. Next, the power level number, the fluctuation intensity level letter, and the dominant change direction letter are combined into a 5-character code string, which serves as the output feature code for this window. The window slides forward 20 sampling points each time to generate a continuous output feature code sequence. For example, if the average power value of a power generation unit falls into level 5 within a 2-second window, the fluctuation intensity is medium, and the dominant direction is upward, then its output feature code is 5_M_U.

[0021] S122. Obtain historical operating data, analyze the correlation between the power output characteristics of the power generation unit and the hysteresis loop morphology of the magnetic element unit in the historical operating data, and construct a mapping table between the power output characteristic code and the hysteresis loop trajectory type. The specific steps to characterize the power output-magnetization mapping relationship include: Retrieve historical operating data containing at least 1,000 hours of operating time from the database; the historical operating data includes the output characteristic coding sequence of each power generation unit within a continuous period of time, and the hysteresis loop trajectory data of the corresponding magnetic element unit obtained by the hysteresis loop measurement device within the same time period. When analyzing correlation patterns, all output feature codes in the historical operation data are extracted to form a feature code set. Morphological clustering is then performed on the hysteresis loop trajectories of the magnetic element units corresponding to the time periods in which each code appears. The hysteresis loop trajectory morphological clustering uses K-means clustering based on four parameters: loop area, remanent point location, coercive point location, and loop rectangle ratio. This classifies the hysteresis loop trajectories into 12 typical types, labeled H1 to H12. For each output feature code, the probability of each trajectory type occurring under that output feature code is calculated using conditional probability: in the historical operation data, the frequency of each hysteresis loop trajectory type occurring in all time periods where that output feature code appears is counted, and this frequency is divided by the total frequency of that output feature code to obtain the conditional probability of each trajectory type occurring. The hysteresis loop trajectory type with the highest probability (above 60%) is selected as the dominant corresponding type for this encoding. If the highest probability is below 60%, the two hysteresis loop trajectory types with the highest probability are selected as optional types, and the probability ratio between the two is recorded. This establishes a mapping table with the output feature code as the key and the value as an identifier of one or two hysteresis loop trajectory types, which is the output-magnetization mapping relationship, stored in the controller in the form of a database table. For example, statistical analysis shows that the output feature code 5_M_U corresponds to the H3 type hysteresis loop trajectory type in 78% of cases. Therefore, 5_M_U in the mapping table directly points to the H3 type, ensuring the robustness of the mapping relationship.

[0022] S123. Based on the output-magnetization mapping relationship, find the target hysteresis loop trajectory type corresponding to the current output feature code.

[0023] S124. Based on the target hysteresis loop trajectory type, determine the magnetic element unit in the magnetic element integrated array that has an energy receiving relationship with the corresponding power generation unit, and assign the target hysteresis loop trajectory type to the magnetic element unit as the current hysteresis loop trajectory mode of the magnetic element unit. Specifically, each power generation unit in the high-altitude wind power array is connected to a specific subarray in the magnetic element integrated array via a DC bus. This connection relationship is predetermined and recorded in an energy receiving association table, which records a one-to-one or one-to-many correspondence between the power generation unit number and the magnetic element unit group number. After obtaining the target hysteresis loop trajectory type by looking up the output characteristic code of any power generation unit in the table, the energy receiving association table is immediately queried to locate one or more magnetic element units associated with that power generation unit. If the energy receiving association relationship is one-to-many, the magnetic element unit with the lowest current remaining magnetic energy reserve and not in the process of energy release is selected as the target magnetic element unit to achieve balanced utilization. After positioning is completed, a one-byte mode setting instruction is written to the control register of the magnetic element unit. The high 4 bits of this instruction represent the hysteresis loop trajectory type code, and the low 4 bits represent the operating mode flag. After receiving the mode setting instruction, the control register of the magnetic element unit immediately adjusts the operating point to the magnetization state that matches the target hysteresis loop trajectory type, thereby completing the process of assigning the hysteresis loop trajectory mode. For example, the power generation unit UG03 corresponds to magnetic element unit groups M05 to M08. The current target trajectory type is H3. By comparison, it is found that the residual magnetic energy reserve of magnetic element unit M06 is the lowest. Therefore, the mode register of magnetic element unit M06 is updated to H3, so that it enters the hysteresis loop trajectory mode that is synchronized with the upstream power output fluctuation.

[0024] S13. Based on the hysteresis loop trajectory patterns of all magnetic element units in the integrated array of magnetic elements, aggregate them according to the spatial topological positional relationship of each magnetic element unit in the integrated array of magnetic elements to generate an energy tidal map; wherein, each map node in the energy tidal map corresponds to a magnetic element unit, the node state value of the map node is determined by the hysteresis loop trajectory pattern of the corresponding magnetic element unit, and the connection relationship between map nodes reflects the magnetic coupling path relationship between magnetic element units: Specifically, the magnetic element integrated array adopts a topology combining three-dimensional stacking and planar staggered arrangement. Each magnetic element unit has unique spatial coordinates in the magnetic element integrated array, which are represented by row index, column index, and layer index. After all magnetic element units are assigned the current hysteresis loop trajectory pattern, an aggregation process is performed. During aggregation, the spatial coordinates are used as the organization index to transform the hysteresis loop trajectory pattern of each magnetic element unit into a quantized energy state description at that spatial coordinate. Specifically, three key parameters are extracted from the hysteresis loop trajectory pattern: the current magnetization intensity vector angle of the magnetic element unit, the remaining magnetic energy reserve, and the magnetization branch flag. A three-dimensional matrix is ​​constructed using the three key parameters. Each element of the matrix corresponds to a magnetic element unit, and the element value includes the encoded combination of the magnetization intensity vector angle, the remaining magnetic energy reserve, and the magnetization branch flag. Simultaneously, based on the pre-laid magnetic coupling path topology diagram in the integrated array of magnetic components, which is a directed graph reflecting the physical magnetic circuit connection relationship and magnetic flux flow path between each magnetic component unit, determined by the magnetic core layout and magnetically conductive connectors during assembly, connecting edges are established between matrix elements. The weight of the connecting edge represents the magnetic circuit conduction capability between adjacent magnetic component units, which is determined by the magnetic core spacing and the permeability of the intermediate medium. The aggregation result forms a dynamic graph containing all unit position information, energy state information, and magnetic circuit connection relationship, which is the energy tidal map. In the energy tide map, each node corresponds to a magnetic element unit. The node state value is a multi-dimensional state vector. The first component of this vector is the magnetization intensity vector angle, which ranges from 0 degrees to 360 degrees, reflecting the macroscopic orientation of the magnetic domains inside the unit. The second component is the remaining magnetic energy reserve. The third component is the magnetization branch flag, which is 0 or 1, representing the rising or falling branch of the main magnetization curve, respectively. All three components are directly extracted in real time from the current hysteresis loop trajectory mode of the unit. The connection between the nodes in the graph is presented in the form of directed edges. The direction of the edge represents the physical direction that allows the transmission of magnetic flux. The weight of the edge is equal to the coupling coefficient of the magnetic coupling path between two adjacent magnetic element units. The coupling coefficient is obtained by calculation and actual measurement calibration based on the air gap width between adjacent magnetic element units, the overlap of the magnetic core cross-sectional area, and the instantaneous permeability of the intermediate magnetic conductive material. These parameters are determined by the magnetic circuit structure setting and are measured and recorded during assembly. The energy tidal map is refreshed on a time axis. Each refresh updates the state values ​​and connections of all map nodes. Within the refresh cycle of the energy tidal map, when the hysteresis loop trajectory pattern of any magnetic element unit is updated, the state vector of the corresponding node is updated accordingly. When the magnetic circuit switch state changes, the weight of the corresponding edge is updated, thus ensuring that the map remains synchronized with the physical state of the integrated array of magnetic elements. For example, a magnetic element unit with row index 2, column index 3, and layer index 1 has a magnetization intensity vector angle of 45 degrees, a remaining magnetic energy reserve of 80 joules, and a magnetization branch flag indicating the rising branch of the main magnetization curve. The corresponding position in the map will then display the corresponding energy state flag, and the weights of the connection edges between this magnetic element unit and its four adjacent magnetic element units will be recorded simultaneously.

[0025] In this embodiment of the invention, the detailed implementation steps of obtaining real-time load demand data of each computing node in the AI ​​server cluster, combining it with energy tide maps for supply and demand coupling analysis, and identifying load resonance windows in the energy tide maps that match the real-time load demand data in a preset dimension include: S21. Obtain real-time load demand data for each computing node in the AI ​​server cluster within the prediction time window; wherein, the real-time load demand data includes the expected power consumption value and power consumption change trend of each computing node. Understandably, the AI ​​server cluster's scheduler is configured with a load prediction component based on time series analysis. This component takes the historical power consumption sequence of each computing node as input and outputs a sequence of expected power consumption values ​​every second over a future period. The prediction time window is set to 3 minutes. The value of the prediction time window is determined jointly based on the average time of a single iteration step in the training task and the average queuing time of the cluster task queue. The smaller of 6 times the average iteration step time and 2 times the average queuing time is taken as the window length to ensure that the window can cover typical load fluctuation cycles without introducing excessively long invalid prediction intervals. The real-time load demand data collection process is as follows: The load prediction component performs a prediction every 5 seconds, generating an array of expected power consumption values ​​for 180 time points within 3 minutes. Simultaneously, by performing point-by-point difference calculations on the expected power consumption value sequence, the direction of power consumption change information at each time point is obtained. The difference between the expected power consumption values ​​of two adjacent sampling points is calculated point by point. If the difference between 10 consecutive sampling points is positive and the cumulative increase exceeds 1% of the rated power, it is determined to be an upward trend; if the difference between 10 consecutive sampling points is negative and the cumulative decrease exceeds 1%, it is determined to be a downward trend; the rest are considered a stable trend. Thus, each calculation node generates a data structure containing a timestamp, expected power consumption value, and power consumption change trend label as real-time load demand data.

[0026] S22. Match the expected power consumption value in the real-time load demand data with the node status value of each graph node in the energy tide graph in terms of energy dimension, and select a set of candidate graph nodes whose node status values ​​meet the energy supply capacity required by the real-time load demand data. When performing energy dimension matching, the expected power consumption value of the computing node is converted into the total energy demand. Specifically, the expected power consumption value sequence of the computing node within the prediction time window is taken, and the expected power consumption value sequence is integrated over time to obtain the total electrical energy required within the window period. All graph nodes in the energy tide map are traversed, the remaining magnetic energy reserve of each graph node is read, and the remaining magnetic energy reserve is multiplied by the maximum magnetic energy capacity of the corresponding magnetic element unit of the graph node to obtain the total releaseable magnetic energy of the graph node. Then, an energy supply safety factor is introduced, which is determined based on historical power supply deviation statistics, i.e., an energy reserve higher than the demand must be provided. The energy supply safety factor is set to 1.2, which requires that the total releaseable magnetic energy of the node must be greater than or equal to 1.2 times the total energy demand. Graph nodes that meet this condition are initially included in the candidate graph node set. At the same time, it checks whether the graph node is currently in a schedulable state, that is, its magnetization branch flag must be the descending branch of the main magnetization curve, because only the descending branch can correspond to the external release of magnetic energy, while the ascending branch represents that energy storage is in progress; after comparing the total energy and filtering the states, the remaining graph nodes constitute the candidate graph node set.

[0027] S23. In the candidate graph node set, the power consumption change trend in the real-time load demand data is matched with the node state value change trend of each graph node in the time dimension, and the graph node whose node state value change trend is consistent with the power consumption change trend in the time dimension and whose change rhythm is synchronized is identified as the target matching graph node. Specifically, the rhythmic characteristics of power consumption trends of each computing node in real-time load demand data are extracted. These rhythmic characteristics include the start time of the rising edge of power consumption, the peak time of power consumption, and the end time of the falling edge of power consumption. When extracting these rhythmic characteristics, inflection point detection is performed on the power consumption trend time series of each computing node. A sliding window zero-crossing detection method is used to identify the start time of the rising edge of power consumption: a sliding window with a width of 5 sampling points is applied to the power change trend, and the point where the difference value changes from negative to positive within the window is the start time of the rising edge. The moment when the first difference value changes from negative to positive during the transition from a stable phase to an upward trend is marked as the start time. The peak time of power consumption is the point at which the power value reaches a local maximum after the upward trend ends and then begins to decline continuously. The end time of the falling edge of power consumption is the moment when the power value returns to a stable low level and the fluctuation amplitude is less than 2% of the rated power. These three moments constitute a triplet of rhythmic characteristics. The energy release rhythm characteristics corresponding to the changing trends of node state values ​​of each node in the candidate graph node set are extracted. These energy release rhythm characteristics include the start time, peak time, and end time of magnetic energy release. When extracting these characteristics, for each graph node in the candidate graph node set, based on the curve portion of its current hysteresis loop trajectory pattern where magnetization decreases with magnetic energy release, the same time scale is used to extract the start time of magnetic energy release (i.e., the moment when magnetization begins to decrease significantly from the remanence point); the peak time of magnetic energy release (i.e., the moment when the magnetic energy release power reaches its maximum); and the end time of magnetic energy release (i.e., the moment when magnetization drops below the coercivity point and the releaseable magnetic energy is exhausted). The coercivity point refers to the magnetic field strength position on the hysteresis loop trajectory where the magnetization returns to zero. When the magnetization falls below this point, the magnetic element unit can no longer release useful magnetic energy. The temporal alignment of the change rhythm characteristics and the energy release rhythm characteristics is compared. If the start time of the rising edge of power consumption is no earlier than the start time of magnetic energy release, and the end time of the falling edge of power consumption is no later than the end time of magnetic energy release, then the corresponding spectral node is considered to meet the matching condition in the time dimension. In other words, when performing the temporal alignment comparison, the change rhythm characteristic triplet and the energy release rhythm characteristic triplet are placed on the same time axis. The condition is that the start time of the rising edge of power consumption is no earlier than the start time of magnetic energy release, indicating that when the load begins to increase, the magnetic element... The energy release of the device is ready and there will be no power supply delay; at the same time, the termination time of the falling edge of power consumption is no later than the termination time of magnetic energy release, requiring that the duration of magnetic energy supply must be sufficient to cover the high power consumption phase of the load until it completely decreases, and cannot be interrupted prematurely; if the above two timing inclusion relationships are satisfied, the graph node is determined to match in the time dimension; for example, the magnetic energy release start time of a graph node is at second 0 and the termination time is at second 4; while the load rising edge starts at second 0.3 and the falling edge terminates at second 3.5, which is completely included, and the match is determined to be successful.

[0028] S24. The process of combining the time interval corresponding to the target matching spectrum node with the magnetic element unit and identifying it as the load resonance window is as follows: After completing the time dimension matching, each target matching graph node and its successfully matched time interval are extracted together. The start point of this time interval is the earlier of the start time of magnetic energy release and the start time of the rising edge of power consumption, and the end point is the later of the end time of magnetic energy release and the end time of the falling edge of power consumption, forming a complete power supply coverage window. Then, this time interval is combined with the magnetic element unit number corresponding to the target matching graph node to establish a data tuple with the time interval as the validity period and the magnetic element unit as the main power supply body. This data tuple is a load resonance window. If multiple target matching graph nodes jointly power a computing node, the intersection of their time intervals is taken as the overall window, and all participating magnetic element unit numbers are formed into a unit set, which together constitute the load resonance window description of the computing node. The identification result of the load resonance window includes fields such as the absolute start and end time of the window, the list of participating units, the total amount of releaseable magnetic energy provided to the corresponding computing node, and the expected power consumption sequence of the node within the window.

[0029] In this embodiment of the invention, the specific implementation method for generating a magnetic flux pulse coding sequence for driving the integrated array of magnetic components to perform energy release based on the identification result of the load resonance window, and for performing hierarchical energy release scheduling on each magnetic component unit in the integrated array of magnetic components includes: S31. Determine the energy release priority ranking of each magnetic element unit based on the energy supply correspondence between the target matching spectrum nodes corresponding to the load resonance window and each calculation node. Specifically, an energy supply correspondence table is established based on the load resonance window information. Each row in the table records the association between a computing node and its matching target matching spectrum node set and the corresponding magnetic element unit number. When determining the priority ranking, each computing node is assigned a weight coefficient, which combines three factors: the priority label of the computing node's current running task, which is divided into three levels: critical task, important task, and ordinary task, and assigned numerical weights of 3, 2, and 1 respectively; the ratio of the computing node's average power consumption value within the prediction window to its rated power; and the remaining time before the task deadline for the computing node. The shorter the duration, the higher the urgency factor. The urgency factor is a numerical coefficient dynamically calculated based on the remaining time of the computing node's task deadline. Specifically, the remaining time is divided into three intervals: when the remaining time is greater than 10 minutes, the urgency factor is 1.0; when the remaining time is between 2 and 10 minutes, the urgency factor is 1.5; and when the remaining time is less than 2 minutes, the urgency factor is 2.0. The shorter the remaining time, the higher the urgency factor, thereby increasing the overall weight of the computing node and obtaining a higher scheduling priority in the energy release priority ranking. The comprehensive weight of a computing node is obtained by multiplying the task priority weight coefficient, the average power consumption ratio, and the urgency factor. For each target matching graph node, the comprehensive weights of all computing nodes powered by it are accumulated, and the accumulated value is the basic priority score of the magnetic element unit corresponding to that graph node. At the same time, a penalty factor is added: if the current remaining magnetic energy reserve of the magnetic element unit is less than 40%, the basic priority score is multiplied by the penalty factor of 0.8 to obtain the final priority score, so as to reduce the release order of units with lower energy storage levels and avoid their over-discharge. Finally, all magnetic element units participating in the load resonance window are sorted from high to low priority scores to form an energy release priority queue, with the unit with the highest priority releasing energy first.

[0030] S32. Based on the priority of energy release, generate a corresponding magnetic flux pulse coding sequence for each magnetic element unit; wherein, the magnetic flux pulse coding sequence includes the pulse trigger time, pulse duration and pulse intensity level; Specifically, based on the start and end times of the load resonance window and the expected power consumption sequence of each computing node, the load resonance window period is divided into multiple micro-time slots, each with a width of 1 millisecond. Starting with the highest priority magnetic element unit, energy release actions are sequentially allocated to each unit within the micro-time slots. For each magnetic element unit, the total magnetic energy to be released is first determined. This total magnetic energy is equal to the sum of the energy demands of all computing nodes powered by that magnetic element unit within the load resonance window multiplied by an energy supply safety factor of 1.2. Based on the hysteresis loop trajectory mode and the core saturation flux density limit of the magnetic element unit, an optimal combination of pulse duration and pulse current amplitude is calculated without entering magnetic saturation. The pulse current amplitude level is quantized into 8 levels, from level 1 to level 8, corresponding to... The rated drive current is 12.5% ​​to 100%; the micro-timeslot grid is a time grid formed by dividing the load resonance window period into 1-millisecond intervals. The pulse trigger time is selected as the micro-timeslot start point where the corresponding magnetic element unit in the grid begins to release energy. The principle for determining the micro-timeslot start point is: the pulse release start time must be earlier than the start time of the rising edge of the power consumption of the computing node, and the advance is set to 500 microseconds to reserve energy transmission delay; thus, a sequence of multiple pulses is generated for each unit, namely the magnetic flux pulse coding sequence; each pulse has a trigger time, duration and intensity level; for example, a high-priority unit needs to release 800 joules, generating 10 pulse segments, where the trigger time of the first pulse segment is 0.5 milliseconds before the load window opens, the duration is 200 microseconds, and the intensity level is 6.

[0031] S33. According to the order of pulse triggering times in the magnetic flux pulse coding sequence, pulse drive signals are sent sequentially to the corresponding magnetic element units to drive each magnetic element unit to perform hierarchical energy release scheduling according to the pulse duration and pulse intensity level. The specific implementation method is as follows: All pending magnetic flux pulses are globally sorted according to their trigger times to form a sequential pulse event queue. The current time is continuously monitored, and when the time advances to the trigger time of a magnetic flux pulse, a pulse drive signal is immediately sent to the control winding of the corresponding magnetic element unit via the magnetic element drive bus. The pulse drive signal is converted from digital to analog output according to the pulse intensity level, and its duration is controlled by a timer interrupt. During the effective period of the drive signal, the excitation current of the magnetic element unit changes accordingly, causing magnetic energy to be released to the DC bus through the coupling path. The hierarchical energy release scheduling is manifested as follows: the magnetic flux pulse sequence of high-priority units is advanced on the time axis, prioritizing energy injection into the bus; the magnetic flux pulse sequence of low-priority units is arranged to be released in subsequent time periods, thus naturally forming a stepped sequence of energy release. When the complete magnetic flux pulse sequence of a magnetic element unit is completed, the hysteresis loop trajectory mode of that magnetic element unit is updated to the end of the loop, marking the completion of this round of energy release.

[0032] S34. Prioritize the delivery of energy released by high-priority magnetic element units to computing nodes where the power consumption trend is in the rising edge stage of the load resonance window, and deliver the delivery of energy released by low-priority magnetic element units to computing nodes where the power consumption trend is in the stable stage. Specifically, after energy is collected via the DC bus, it is directionally delivered through a controlled energy distribution switch matrix. Each output port of the switch matrix is ​​connected to a computing node. The conduction strategy of the switch matrix is ​​dynamically set according to the power consumption change trend label in the real-time load demand data. For computing nodes marked as rising edge phases, their input ports are configured to maintain a hard connection with the bus partition where the high-priority magnetic element unit is located, and the power supply priority weight of this port is increased, so that the switch matrix first allocates the energy released by the high-priority unit to these computing nodes. At this time, the corresponding power MOSFET bridge arm inside the switch matrix is ​​turned on with low impedance, and the energy transmission loss is extremely low. For computing nodes in the steady phase, their input ports are connected to the bus partition containing the energy of the low-priority unit, and the remaining energy is only received after the voltage drop requirement of the computing node in the rising edge phase is met. The distribution process uses hysteresis comparison control: the input voltage of the rising edge node is detected in real time. If it is higher than the set rated operating voltage reference value, the excess energy is allowed to bypass to the steady phase node. Through the directional delivery mechanism, the limited pulse magnetic matching is ensured during the phase of the most drastic load change, thereby improving power supply efficiency.

[0033] In this embodiment of the invention, the detailed implementation steps of continuously monitoring the magnetic coupling depth change state of each magnetic element unit in the magnetic element integrated array during the hierarchical energy release scheduling process, and determining whether to trigger the adaptive coupling reconfiguration operation for the deviating magnetic element unit, include: S41. Continuously collect the real-time magnetic coupling depth value of each magnetic element unit in the magnetic element integrated array at a preset monitoring cycle. Understandably, on the core pillar of each magnetic element unit, a set of miniature Hall sensors and a set of auxiliary induction coils are installed around the cross-section through which the magnetic flux flows. The Hall sensor outputs a voltage proportional to the local magnetic flux density, and the induced electromotive force of the auxiliary induction coil is proportional to the rate of change of magnetic flux. The signal conditioning circuit performs high-speed synchronous sampling of the voltage signal output by the Hall sensor and the induced electromotive force signal output by the auxiliary induction coil, with the sampling frequency set to 10 kHz. The sampling value is read from each magnetic element unit at a preset monitoring period, and the real-time magnetic coupling depth value is obtained by calculating the ratio of mutual inductance to self-inductance between adjacent magnetic element units. The monitoring period is set to 200 microseconds, which is determined based on the relaxation time of the ferrite material magnetic domain rotation in the magnetic element unit. The relaxation time is approximately several microseconds to tens of microseconds. Setting the monitoring period to about 10 times the upper limit of the relaxation time ensures timely detection of coupling changes. Therefore, the real-time magnetic coupling depth value of all magnetic element units is captured every 200 microseconds.

[0034] S42. Compare the real-time magnetic coupling depth value of each magnetic element unit with the expected magnetic coupling depth value at the corresponding time in the expected coupling trajectory to obtain the coupling deviation state of each magnetic element unit. Understandably, the expected coupling trajectory is a time-series curve obtained through offline simulation and calibration experiments. For each typical magnetic flux pulse-coded sequence and standard magnetic circuit state, the coupling depth between magnetic component units is simulated over time using finite element electromagnetic simulation software. The simulation results are stored as a time-indexed expected coupling depth data table, forming the expected coupling trajectory. The expected magnetic coupling depth value at each time point in the expected coupling trajectory represents the degree of magnetic coupling that should be achieved between units under ideal conditions of no faults and no interference. When comparing, the corresponding expected magnetic coupling depth value is found from the expected coupling trajectory table based on the timestamp of the current scheduling cycle, and the difference between it and the real-time magnetic coupling depth value is used to obtain the absolute deviation. Further, the absolute deviation is divided by the expected magnetic coupling depth value to obtain the relative deviation percentage. The relative deviation percentage, together with the deviation direction, forms the coupling deviation status word. The high byte indicates the deviation magnitude level, and the low byte indicates the positive or negative direction. For example, if the real-time magnetic coupling depth value is 0.85, the expected magnetic coupling depth value is 0.90, and the relative deviation is -5.6%, then the coupling deviation status word is recorded as a slight negative deviation.

[0035] S43. At the same time, based on the coupling deviation state, it is determined whether the magnetic element unit has deviated from the magnetic coupling depth: when the coupling deviation state of any magnetic element unit indicates that the real-time magnetic coupling depth value deviates from the expected magnetic coupling depth value by more than the preset coupling deviation threshold, it is determined that the magnetic element unit has deviated from the magnetic coupling depth.

[0036] Understandably, the coupling deviation threshold is set to 15% of the expected magnetic coupling depth. The 15% value is determined by considering the percentage decrease in the maximum allowable transmission power of the magnetic element unit and the probability of false protection: a 15% decrease in coupling depth will result in a decrease in magnetic energy transmission power of approximately 20%. During comparison, if the coupling deviation status of three consecutive monitoring cycles indicates a relative deviation exceeding 15% and the directions are the same, it is determined that the magnetic element unit has deviated from the magnetic coupling depth, and a deviation alarm signal is immediately issued. If a deviation is determined, the adaptive coupling reconfiguration preparation process is triggered, the number of the deviating unit is locked, and its current pulse release task is frozen.

[0037] S44. For a magnetic element unit that has deviated from its magnetic coupling depth, obtain the current magnetic coupling depth value and remaining magnetic energy reserve level of its adjacent magnetic element units, including: Immediately access the magnetic coupling path topology file of the integrated magnetic element array. Centered on the spatial coordinates of the offset element, retrieve all adjacent magnetic element elements with which there is a direct magnetic coupling path. Identify a total of 8 or up to 26 adjacent magnetic element elements along the row, column, and diagonal directions. Read the real-time magnetic coupling depth values ​​of these adjacent magnetic element elements at the same time. Simultaneously, read the node state vectors corresponding to the adjacent magnetic element elements in the energy tidal map and extract the current value of the remaining magnetic energy reserve. Combine this information of adjacent magnetic element elements into a neighborhood state list, each item containing the adjacent magnetic element element number, real-time magnetic coupling depth value, remaining magnetic energy reserve, and whether it is in a schedulable state. This neighborhood state list will serve as direct input data for determining the compensation scheme.

[0038] S45. Based on the current magnetic coupling depth value and the remaining magnetic energy reserve level of the adjacent magnetic element units, determine the coupling compensation scheme, and perform an adaptive coupling reconstruction operation on the magnetic element units that have deviated according to the coupling compensation scheme; wherein, the adaptive coupling reconstruction operation includes adjusting the magnetic circuit connection mode between the deviated magnetic element unit and its adjacent magnetic element units and redistributing the magnetic flux transmission path. Specifically, from the neighboring magnetic element units of the magnetic element unit that has deviated from the magnetic coupling depth, usable compensation units are selected that have a current magnetic coupling depth value within the normal coupling range and a remaining magnetic energy reserve level higher than the reserve threshold. The normal coupling range is defined as a range of 10% above and below the expected magnetic coupling depth value, that is, within ±10% of the expected magnetic coupling depth value is considered a normal coupling state. The reserve threshold is set to 30% of the maximum magnetic energy capacity of the magnetic element unit. Magnetic element units below the reserve threshold are considered to have insufficient remaining energy and are not used as compensation units. After obtaining the neighborhood state list, magnetic element units with real-time magnetic coupling depth values ​​exceeding the normal coupling range and magnetic element units with remaining magnetic energy reserve levels lower than the reserve threshold are filtered out. The remaining magnetic element units are the usable compensation units. Based on the relative relationship between the number of available compensation units and their respective residual magnetic energy reserve levels, the coupling compensation share undertaken by each available compensation unit is determined: the compensation share of any available compensation unit is equal to the residual magnetic energy reserve of that unit divided by the sum of the residual magnetic energy reserves of all available compensation units. According to the coupling compensation ratio, a magnetic circuit reconstruction instruction is generated for the magnetic element unit that has deviated; wherein, the magnetic circuit reconstruction instruction indicates that the original magnetic flux transmission path of the magnetic element unit that has deviated be switched to the backup magnetic flux transmission path corresponding to the available compensation unit. The magnetic circuit reconfiguration command is sent to the magnetic circuit switching actuator of the integrated magnetic element array to complete the adaptive coupling reconfiguration operation. The magnetic circuit reconfiguration command is transmitted to the magnetic circuit switching actuator of the integrated magnetic element array via a digital signal line. The actuator consists of a set of high-speed magnetic latching relays and adjustable magnetoresistive units. The relay switching disconnects the magnetic circuit input and output terminals of the deviating unit and closes the bypass switch of the corresponding compensation unit, allowing the magnetic flux to flow through the new path. The entire operation is completed within 200 microseconds, achieving a seamless switching of power supply continuity.

[0039] In this embodiment of the invention, the specific implementation of updating the energy tidal map after performing the adaptive coupling reconstruction operation is as follows: S51. After the adaptive coupling reconstruction operation is completed, obtain the updated hysteresis loop trajectory pattern of each magnetic element unit in the magnetic element integrated array.

[0040] S52. Based on the updated hysteresis loop trajectory pattern, regenerate the node state values ​​of the corresponding nodes in the energy tidal map, and synchronously update the connection relationships between map nodes that have changed due to the change in the magnetic circuit connection method.

[0041] S53. Use the updated energy tide map as the input for the supply and demand coupling analysis in the next scheduling cycle, so as to replace the energy tide map of the previous scheduling cycle in the identification of the load resonance window.

[0042] S54. After the load resonance window identification for the next scheduling cycle is completed, the magnetic flux pulse coding sequence for the next scheduling cycle is generated based on the matching results of the updated energy tidal map and the real-time load demand data.

[0043] Specifically, after the adaptive coupling reconstruction operation is completed and a transient stabilization delay has elapsed, the system initiates the energy tidal map update process; it reads the updated hysteresis loop trajectory patterns of all affected units in the magnetic element integrated array; because the magnetic circuit reconstruction changes the magnetic flux path, the previously deviated units enter an isolated, awaiting-repair state, and their hysteresis loop trajectories are frozen; while the units undertaking compensation, due to the inflow of additional magnetic flux, will have their magnetization state move along the new local hysteresis loop, so their hysteresis loop trajectory patterns will be remeasured and refreshed; based on the updated hysteresis loop trajectory patterns, the multidimensional state vectors of these magnetic element units corresponding to the map nodes in the energy tidal map are recalculated: the magnetic... The intensity vector angle, remaining magnetic energy reserve, and magnetization branch flag are updated, and the new values ​​overwrite the old values. At the same time, the weights of the connecting edges that have changed due to the magnetic circuit switching in the topology file are updated synchronously. Strong connecting edges that originally pointed to the deviating unit in the energy tidal map are marked as invalid, and guiding connecting edges from the main bus to the compensation unit are added. The updated energy tidal map is submitted to the supply and demand coupling analysis of the next scheduling cycle, replacing the old energy tidal map of the previous cycle. At the beginning of the next scheduling cycle, the supply and demand coupling analysis will use this map, which already reflects the real magnetic energy distribution after reconstruction, to identify the load resonance window with the new real-time load demand data and generate a magnetic flux pulse coding sequence adapted to the new topology.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Various changes made within the scope of knowledge possessed by those skilled in the art without departing from the concept of the present invention still fall within the scope of protection of the present invention.

[0045] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0046] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A high-altitude wind power AI server power supply system based on magnetic element integration, characterized by, include: Energy Tide Mapping Module: Maps the collected real-time power output fluctuation data to the hysteresis loop trajectory pattern of each magnetic element unit in the magnetic element integrated array, and constructs the energy tide map. Load Resonance Window Identification Module: Acquires real-time load demand data of each computing node in the AI ​​server cluster, and performs supply and demand coupling analysis in combination with energy tide map to identify load resonance windows in the energy tide map that match the real-time load demand data in a preset dimension. Magnetic flux pulse sequence generation module: Based on the identification result of the load resonance window, it generates a magnetic flux pulse encoding sequence for driving the magnetic element integrated array to perform energy release, and performs hierarchical energy release scheduling for each magnetic element unit in the magnetic element integrated array. Layered energy release scheduling module: During the execution of layered energy release scheduling, it continuously monitors the change in magnetic coupling depth of each magnetic element unit in the magnetic element integrated array and determines whether to trigger an adaptive coupling reconfiguration operation for the deviated magnetic element unit. Adaptive Coupling Reconstruction Update Module: After performing the adaptive coupling reconstruction operation, update the energy tidal map.

2. The system of claim 1, wherein, The preset dimensions include the time dimension and the energy dimension.

3. The system of claim 1, wherein, The process of mapping the collected real-time power fluctuation data into the hysteresis loop trajectory pattern of each magnetic element unit in the integrated array of magnetic elements, and constructing an energy tidal map, includes: Real-time power output fluctuation data is continuously collected through the output ports of each power generation unit in the high-altitude wind power array; the real-time power output fluctuation data includes the output power value and power change direction of each power generation unit at different sampling times; The real-time output fluctuation data of each power generation unit is mapped to the hysteresis loop trajectory pattern of the corresponding magnetic element unit in the magnetic element integrated array according to the preset output-magnetization mapping relationship; wherein, the hysteresis loop trajectory pattern represents the magnetization state and remaining magnetic energy reserve level of the corresponding magnetic element unit under the current output conditions. Based on the hysteresis loop trajectory patterns of all magnetic element units, energy tidal maps are generated by aggregating the spatial topological positions of each magnetic element unit in the integrated array of magnetic elements.

4. The system of claim 3, wherein, The step of mapping the real-time output fluctuation data of each power generation unit to the hysteresis loop trajectory pattern of the corresponding magnetic element unit in the magnetic element integrated array according to the preset output-magnetization mapping relationship includes: The output power value sequence and power change direction sequence of each power generation unit within a continuous sampling period are jointly encoded to obtain the output characteristic code of each power generation unit. Acquire historical operational data and construct a mapping table between power output feature codes and hysteresis loop trajectory types to characterize the power output-magnetization mapping relationship; Based on the output-magnetization mapping relationship, find the target hysteresis loop trajectory type corresponding to the current output feature code; Based on the target hysteresis loop trajectory type, the magnetic element unit with energy receiving relationship with the corresponding power generation unit is determined in the magnetic element integrated array, and the target hysteresis loop trajectory type is assigned to the magnetic element unit as the current hysteresis loop trajectory mode of the magnetic element unit.

5. The system of claim 3, wherein, In the energy tidal map, each map node corresponds to a magnetic element unit. The node state value of the map node is determined by the hysteresis loop trajectory mode of the corresponding magnetic element unit. The connection relationship between map nodes reflects the magnetic coupling path relationship between magnetic element units.

6. The system of claim 1, wherein, The process of acquiring real-time load demand data for each computing node in the AI ​​server cluster and performing supply-demand coupling analysis in conjunction with energy tidal maps to identify load resonance windows in the energy tidal maps that match the real-time load demand data on a preset dimension includes: Obtain real-time load demand data for each computing node in the AI ​​server cluster within the prediction time window; the real-time load demand data includes the expected power consumption value and power consumption trend of each computing node. The expected power consumption value is matched with the node state value of each node in the energy tide map according to the energy dimension, and a candidate map node set is selected. In the candidate graph node set, the power consumption change trend in the real-time load demand data is matched with the node state value change trend of each graph node in the time dimension. Graph nodes whose node state value change trend is consistent with the power consumption change trend in the time dimension and whose change rhythm is synchronized are identified as target matching graph nodes. The time intervals corresponding to the target matching map nodes are combined with the magnetic element units to identify the load resonance window.

7. The system of claim 1, wherein, The step of generating a magnetic flux pulse-coded sequence to drive the integrated magnetic element array to perform energy release based on the identification result of the load resonance window, and performing hierarchical energy release scheduling on each magnetic element unit in the integrated magnetic element array, includes: Based on the energy supply correspondence between the target matching spectrum nodes corresponding to the load resonance window and each computing node, the energy release priority of each magnetic element unit is determined. Based on the priority of energy release, a corresponding magnetic flux pulse coding sequence is generated for each magnetic element unit; wherein, the magnetic flux pulse coding sequence includes the pulse trigger time, pulse duration and pulse intensity level; According to the order of pulse triggering times in the magnetic flux pulse coding sequence, pulse drive signals are sent to the corresponding magnetic element units in sequence, driving each magnetic element unit to perform hierarchical energy release scheduling according to the pulse duration and pulse intensity level. The energy released by high-priority magnetic element units is preferentially delivered to computing nodes whose power consumption trend is in the rising edge stage of the load resonance window, while the energy released by low-priority magnetic element units is delivered to computing nodes whose power consumption trend is in the stable stage.

8. The system of claim 1, wherein, The continuous monitoring of the magnetic coupling depth changes of each magnetic element unit in the integrated magnetic element array, and the determination of whether to trigger an adaptive coupling reconstruction operation for the deviating magnetic element unit, includes: The real-time magnetic coupling depth value of each magnetic element unit in the magnetic element integrated array is continuously collected at a preset monitoring cycle. The real-time magnetic coupling depth value of each magnetic element unit is compared with the expected magnetic coupling depth value at the corresponding moment in the expected coupling trajectory to obtain the coupling deviation state of each magnetic element unit. At the same time, based on the coupling deviation state, it is determined whether the magnetic element unit has deviated from the magnetic coupling depth; For a magnetic element unit that has deviated from its magnetic coupling depth, obtain the current magnetic coupling depth value and the remaining magnetic energy reserve level of its adjacent magnetic element units; Based on the current magnetic coupling depth and remaining magnetic energy reserve level of adjacent magnetic element units, a coupling compensation scheme is determined, and an adaptive coupling reconstruction operation is performed on the magnetic element units that have deviated according to the coupling compensation scheme.

9. The system of claim 8, wherein, The adaptive coupling reconfiguration operation includes adjusting the magnetic circuit connection between the deviated magnetic element unit and its adjacent magnetic element units, and redistributing the magnetic flux transmission path.

10. The system of claim 1, wherein, The updated energy tidal map includes: After the adaptive coupling reconstruction operation is completed, the updated hysteresis loop trajectory pattern of each magnetic element unit in the magnetic element integrated array is obtained. Based on the updated hysteresis loop trajectory pattern, the node state values ​​of the corresponding nodes in the energy tidal map are regenerated, and the connection relationships between the nodes that have changed due to the change in the magnetic circuit connection method are updated synchronously. The updated energy tide map is used as the input for the supply and demand coupling analysis in the next scheduling cycle, replacing the energy tide map of the previous scheduling cycle in the identification of the load resonance window. Based on the matching results between the updated energy tidal map and the real-time load demand data, a magnetic flux pulse coding sequence for the next scheduling cycle is generated.

Images