A method and system for improving power supply efficiency based on a predictive model

By predicting the load current sequence to identify low-disturbance windows and adjusting the number and phase of power supply modules, the efficiency and quality issues of multi-module power supply devices under load changes are solved, achieving efficient and stable power supply configuration.

CN122844065BActive Publication Date: 2026-11-17SHENZHEN ZHIJIANENG AUTOMATION CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202611329935.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17
Estimated Expiration
2046-08-31

AI Technical Summary

Technical Problem

Existing multi-module power supply devices cannot effectively identify low-disturbance periods when the load changes, resulting in frequent switching and phase reconfiguration operations, causing power supply efficiency loss and bus voltage fluctuations, and making it impossible to optimize the power supply configuration without adding hardware.

Method used

A power supply efficiency improvement system based on a prediction model is adopted. The load current sequence is collected by the current sampling unit, the low disturbance window is predicted by the load prediction model, and the number and phase of the power supply modules are adjusted by the cooperative controller within the prediction window. The power supply quality is verified by the echo judgment module, and the role rotation of redundant modules is realized.

Benefits of technology

Without adding hardware, the power supply module configuration is optimized to improve power supply efficiency, reduce the impact of switching disturbances on the load, reduce no-load losses, and balance the aging of the power supply module, thereby improving power supply quality and efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122844065B_ABST
    Figure CN122844065B_ABST
Patent Text Reader

Abstract

This application discloses a method and system for improving power supply efficiency based on a predictive model, relating to the field of power management and control technology. It includes multiple power supply modules connected in parallel to a common bus, a current sampling unit, a load prediction model, and a cooperative controller. The current sampling unit collects the load current sequence output from the common bus to the load. The load prediction model predicts a low-disturbance window and the predicted load current within the window based on this load current sequence. When the prediction duration is not less than the preset setup time of the reconfiguration operation, the cooperative controller schedules the reconfiguration operation—which changes the number of connected power supply modules to the optimal efficiency under the predicted load current and resets the equally spaced staggered phases—to be executed at the beginning of the window, so that the switching disturbance is masked by the smooth state of the load itself. This application solves the problem of existing multi-module power supply devices exhibiting hysteresis and conservative switching, resulting in idle power consumption of redundant power supply modules under light loads and difficulty in improving power supply efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of power management and control technology, specifically to a method and system for improving power supply efficiency based on a predictive model. Background Technology

[0002] In energy-efficient applications such as server power supplies, communication power supplies, data center power supplies, and electric vehicle charging modules, power supply devices typically consist of multiple power supply modules connected in parallel to the same common bus, which then supplies power to the load. Each power supply module usually switches with staggered phases, allowing switching ripples to cancel each other out on the common bus. When the load decreases to a lighter range, the number of power supply modules connected to the common bus is reduced to decrease switching and drive losses, thus maintaining high power supply efficiency under light loads. Publication number CN115242059A provides a device for multiple power supplies connected in parallel to output to a load; Publication number CN107659150A describes an automatic switching method for DC-DC power conversion modules, where a control unit detects the current output to the load, calculates the number of modules to be connected, and evenly staggers the clock phases of each connected module according to the number of modules connected. Reducing modules under light loads and increasing modules under heavy loads, and resetting the staggered phases according to the number of modules connected, is a fundamental method in this field for improving power supply efficiency.

[0003] In practical implementation, existing technologies mostly use load current magnitude as the basis for switching: several load current thresholds are preset, and when the load current drops below a threshold, a power supply module is disconnected; when it rises back above the threshold, it is reconnected. A hysteresis interval is left between the connection threshold and the disconnection threshold to prevent the power supply module from being repeatedly switched on and off when the load fluctuates slightly around the threshold. To reduce the impact on the shared bus at the moment of disconnection, some methods gradually reduce the duty cycle of the disconnected power supply module to zero; other methods require the powered object to actively send a load warning signal to the power supply device before its load changes, so that the power supply device can add or remove power supply modules in advance.

[0004] However, this type of scheduling, centered on thresholds and hysteresis, has an inherent contradiction. Each switching of power supply modules or resetting of interleaved phases is a disturbance to the shared bus, resulting in a deviation and recovery process in the bus voltage. The cost of the disturbance and the timing of the switching are mutually constrained—since switching occurs when the load may change abruptly, a significant drop or overshoot in the bus voltage could occur if the load changes coinciding with such a change. To avoid this risk, the hysteresis interval must be wide enough, and the switching must be conservative enough, only executing the switch after the load has stabilized for a sufficiently long time. This leads to situations where, during periods when the load has already entered a light load phase and the number of power supply modules connected could be reduced, the power supply unit hesitates to act rashly, allowing excess power supply modules to continue to operate under idle conditions and suffer losses. On the other hand, if the hysteresis interval is narrowed to allow frequent operations, the system will experience the jitter of repeated switching of power supply modules when the load fluctuates near the threshold, increasing bus ripple and corresponding losses. For applications like data centers and communication power supplies that operate under medium to light loads for extended periods, the cumulative efficiency loss from the aforementioned conservative scheduling is considerable. Simply increasing the output capacitor to suppress switching disturbances would increase size and cost, making it impractical in engineering.

[0005] Further analysis reveals that the aforementioned approaches, whether based on load current threshold switching or requiring the load to issue a pre-announcement signal, all focus on the impending load size and the number of power supply modules to be connected. The load announcement signal merely foreshadows the impending load change itself. These solutions fail to consider the following technical issue: loads are not always equally sensitive to power supply disturbances during operation; they have low-disturbance periods where they are insensitive to disturbances. During these periods, load current changes are gradual, and the bus disturbances caused by switching or phase reconfiguration are masked by the load's own smooth state, thus not affecting the power supply. For example, in the power supply of a central processing unit, there are brief gaps with gradual load changes between two computing task switching periods; similarly, there is a gradual load change transition plateau when a charging module switches from a constant current phase to a constant voltage phase. Existing scheduling based on load size and fixed hysteresis only answers the question of load size and the number of power supply modules to be connected, but it does not answer when the action should be taken to ensure that the disturbance cost of switching is not perceived by the load. Therefore, it is impossible to place the reconfiguration operation within these low-disturbance periods.

[0006] Therefore, how to proactively identify the low-disturbance periods of the load itself based solely on the information available to the power supply device without changing the powered load or adding additional disturbance suppression hardware, and how to incorporate power supply module switching and phase reconstruction within these periods, so that the power supply device can achieve the optimal power supply module configuration without sacrificing power quality, has become an urgent problem to be solved in improving the power supply efficiency of multi-module power supply devices. Summary of the Invention

[0007] Existing multi-module power supply devices switch power supply modules based on the load current and fixed hysteresis. This only determines how many power supply modules should be connected, but cannot determine the timing of switching to ensure that the disturbance cost is not perceived by the load. To avoid bus voltage drop or overshoot caused by switching at the moment of load change, the hysteresis range can only be wide and the switching can only be conservative. As a result, when the load has already entered a light load state, the excess power supply modules continue to consume power under no-load conditions, making it difficult to improve power supply efficiency.

[0008] To achieve the above objectives, this application adopts the following technical solution:

[0009] A power supply efficiency improvement system based on a predictive model includes multiple power supply modules, a common bus, a current sampling unit, a load prediction model, and a cooperative controller.

[0010] The multiple power supply modules are connected in parallel to the common bus and supply power to the load through the common bus. Each power supply module switches on and off according to different staggered phases.

[0011] The current sampling unit is used to collect the load current sequence output from the common bus to the load;

[0012] The load prediction model is used to take the load current sequence collected by the current sampling unit as input, and output a predicted low disturbance window and the predicted load current within the predicted low disturbance window. The predicted low disturbance window is a prediction of the low disturbance window and includes the prediction start time and prediction duration. The low disturbance window is the period in which the rate of change of the load current remains within a preset rate of change range and the duration is not less than the preset shortest window duration.

[0013] The cooperative controller is used to schedule the reconstruction operation to be executed at the start time of the prediction when the prediction duration is not less than the preset establishment time of the reconstruction operation. The reconstruction operation is to change the number of power supply modules connected to the common bus to the target number of power supply modules, and reset the staggered phase of each power supply module connected to the common bus to an equal-interval phase according to the target number of power supply modules. The target number of power supply modules is the number that makes the power supply efficiency of the system not lower than the power supply efficiency under the predicted load current.

[0014] Optionally, the load prediction model is trained based on historical load current sequences. The input during training is a sampling segment in the historical load current sequence, and the output during training is a predicted low-disturbance window and a predicted load current. The training labels are a low-disturbance window label and a load current label. The low-disturbance window label is determined by the start time and duration of an actual period in the actual load current sequence after the sampling segment, where the rate of change remains within the preset rate of change range and the duration is not less than the preset shortest window duration. The load current label is the load current within the actual period.

[0015] Optionally, the cooperative controller resets the interleaved phases of each power supply module connected to the common bus to equally spaced phases by dividing a switching cycle into the same number of phase segments according to the number of power supply modules connected to the common bus, and assigning the different phase segments to each power supply module connected to the common bus as its interleaved phase; the preset establishment time is the time required for the bus voltage of the common bus to deviate and recover to the preset voltage regulation band caused by the reconfiguration operation.

[0016] Optionally, the collaborative controller determines the number of target power supply modules by: pre-setting a module quantity lookup table, which stores the correspondence between each load current range and the number of target power supply modules; the collaborative controller retrieves the corresponding number of target power supply modules from the module quantity lookup table based on the load current range to which the predicted load current belongs.

[0017] Optionally, the collaborative controller is further configured to maintain the number of power supply modules currently connected to the shared bus unchanged when the predicted duration is less than the preset establishment time, until the predicted duration of the low-disturbance window output by the load prediction model is not less than the preset establishment time.

[0018] Optionally, it also includes a bus voltage sampling unit and an echo determination module; the bus voltage sampling unit is used to collect the bus voltage echo sequence of the common bus within a preset observation window after the reconstruction operation is executed; the echo determination module has a built-in echo template library, which includes window establishment templates and window failure templates; the echo determination module is used to extract echo features from the bus voltage echo sequence, the echo features being feature quantities characterizing the deviation amplitude and deviation recovery process of the bus voltage echo sequence, and to perform distance measurement calculations between the echo features and the window establishment template and the window failure template respectively, and to take the determination result corresponding to the template with the smallest distance measurement as the window verification result; the cooperative controller is also used to restore the number of power supply modules connected to the common bus to the number before the reconstruction operation was executed when the window verification result is that the window is failure.

[0019] Optionally, the echo characteristics include the deviation peak value, deviation recovery time, and oscillation number of the bus voltage echo sequence; the window establishment template and the window failure template are generated based on historical reconstruction operations. Each template includes a template feature vector and a template determination label. The template determination label is determined by whether the load current after the corresponding historical reconstruction operation is maintained within the preset rate of change range. If it is maintained within the preset rate of change range, the window is considered established; if it is not maintained within the preset rate of change range, the window is considered failed.

[0020] Optionally, the number of the plurality of power supply modules is greater than the number of target power supply modules. The number of power supply modules currently connected to the common bus is the number of target power supply modules, and the remaining power supply modules not connected to the common bus are used as redundant modules. The cooperative controller is further configured to determine the target power supply module under the premise that the number of target power supply modules remains unchanged, and schedule the role rotation operation to be executed within the predicted low disturbance window. The role rotation operation is to switch the target power supply module from being connected to the common bus to being a redundant module, and to switch a currently redundant module to being connected to the common bus. The target power supply module is the power supply module with the largest cumulative carrying time among the power supply modules currently connected to the common bus. The cumulative carrying time is the cumulative value of the time that the corresponding power supply module is in the state of being connected to the common bus.

[0021] Optionally, the role rotation operation includes: within the predicted low-disturbance window, first connecting a current redundant module to the common bus and participating in the switching, then disconnecting the target power supply module from the common bus and switching it to a redundant module, and resetting the staggered phases of each power supply module connected to the common bus to equally spaced phases according to the number of power supply modules currently connected to the common bus, so that the number of power supply modules connected to the common bus remains unchanged before and after the role rotation operation.

[0022] This application also discloses a power supply efficiency improvement method based on a predictive model, which is applied to the power supply efficiency improvement system based on the predictive model described in this application, including: collecting the load current sequence output by the common bus to the load;

[0023] Using the load current sequence collected by the current sampling unit as input, the load prediction model outputs a predicted low-disturbance window and the predicted load current within the predicted low-disturbance window. The predicted low-disturbance window is a prediction of the low-disturbance window and includes the prediction start time and prediction duration. The low-disturbance window is a period in which the rate of change of the load current remains within a preset rate of change range and the duration is not less than the preset shortest window duration.

[0024] When the predicted duration is not less than the preset establishment time of the reconfiguration operation, the reconfiguration operation is scheduled to be executed at the start time of the prediction. The reconfiguration operation is to change the number of power supply modules connected to the common bus to the target number of power supply modules, and reset the staggered phase of each power supply module connected to the common bus to an equal-interval phase according to the target number of power supply modules. The target number of power supply modules is the number that makes the power supply efficiency of the system not lower than the power supply efficiency under the predicted load current.

[0025] The beneficial effects of this application are as follows:

[0026] This application enables multi-module power supply devices to operate close to the optimal power supply module configuration without sacrificing power quality due to switching disturbances. During operation, loads experience low-disturbance periods that are insensitive to power supply disturbances. These periods are determined by the load's own operating rhythm and exist independently of the power supply device. This application uses the load current sequence output from the shared bus to the load as input. A load prediction model anticipates the prediction start time, prediction duration, and predicted load current within the low-disturbance window. When the prediction duration is not less than the preset setup time for the reconfiguration operation, the cooperative controller schedules the reconfiguration operation—which changes the number of connected power supply modules and resets the equally spaced interleaved phases according to the changed number—to be executed at the prediction start time. The changed number is taken as the target number of power supply modules that ensures power supply efficiency is not lower than other connected numbers under the predicted load current. Because the reconfiguration operation is scheduled during periods of moderate load change, the resulting bus voltage deviation is masked by the low-disturbance state of the load itself. Switching no longer relies on widening the hysteresis interval or delaying the action to avoid sudden load changes. The hysteresis interval is thus narrowed, and switching is timely. Furthermore, since each operation utilizes the optimal number of power supply modules, the switching losses reserved for light-load periods under conservative scheduling are reduced. Taking the CPU power supply as an example, this application predicts the moderate load gap between two tasks and reduces modules during this period. The disturbance is masked by the gap, allowing switching to a more energy-efficient configuration without waiting for a prolonged period of load stabilization.

[0027] This application requires no modification to the load side or the addition of communication, and can be deployed on existing multi-module power supply devices. Its adaptability is not limited by whether the powered object has a forecast capability. This is because the information used by this application to predict the low-disturbance window comes entirely from the power supply device's collection of the load current of the common bus: the low-disturbance window corresponds to the state of the load current change being gradual, which can be directly observed from the bus current sequence, rather than a specific change that the load is about to undergo. Therefore, the power supply device does not need to wait for the load to issue a load forecast signal before its change, and can identify this state from the bus current sequence.

[0028] This application, by incorporating the reconfiguration operation into the prediction window, enables real-time verification and correction of prediction accuracy. The shape of the voltage echo left on the shared bus after a single switching operation depends on whether the load is indeed stable at that moment: when the window is active, the echo amplitude is small and recovers quickly; when the window is inactive (i.e., the load is still fluctuating), the echo presents a different shape, and this echo itself is an unavoidable byproduct of the switching operation. This application collects the bus voltage echo sequence within the observation window after the reconfiguration operation, extracts the deviation peak value, deviation recovery time, and oscillation number to form echo characteristics, and compares them with preset window active and window inactive templates using distance measurement. When it is determined that the window has been inactive, the number of connected power supply modules is immediately restored to the level before the reconfiguration. Because the echo generated by the switching itself is used as the verification basis, occasional prediction failures can be identified and corrected on the spot. Therefore, the collaborative controller can adopt a narrower hysteresis interval and a more timely module reduction strategy, achieving efficiency gains while suppressing power quality risks.

[0029] This application also distributes the losses generated by the long-term idle of redundant modules over time, and makes the aging of each power supply module tend to be more even. If a certain power supply module is fixed to be idle for a long time as a redundant module, the fixed module will continuously generate idle losses, while the other power supply modules will accumulate load and age separately. This application does not fix the power supply module to serve as a redundant module, but selects the one with the longest cumulative load time as the target power supply module, and switches it to a redundant module through a role rotation operation. The role rotation operation is also scheduled to be completed within the predicted low disturbance window in the same way as the first connection and then the second disconnection, so that the number of power supply modules connected to the common bus remains unchanged before and after the rotation. The role rotation operation uses the same low disturbance window and adopts the order of first connection and then disconnection. The process does not introduce additional perceptible disturbances. Therefore, the redundancy is distributed by each power supply module in turn, the idle losses are evened out over time, the load time of each power supply module tends to be consistent, and the aging of the devices is balanced accordingly.

[0030] Overall, the above design adapts to each other around the same predicted low-disturbance window: fixed-point deployment determines when to adjust the number of power supply modules to the optimal efficiency, echo verification protects the power supply quality when the prediction fails, and role rotation operation incorporates redundant no-load losses and uneven aging into the same window to resolve them, so that the multi-module power supply device can recover the power supply efficiency margin in the long-term operation of medium and light load. Attached Figure Description

[0031] Figure 1 A block diagram of the overall architecture of the power supply efficiency improvement system based on a prediction model provided in the embodiments of this application;

[0032] Figure 2 A schematic diagram illustrating the power supply modules provided in this application embodiment, with ripple cancellation on the interleaved phase switches and common bus;

[0033] Figure 3 A schematic diagram showing the comparison between the number of power supply modules connected to the common bus before and after the reconfiguration operation and the equal-interval phase reconfiguration, provided in the embodiments of this application;

[0034] Figure 4 A timing diagram illustrating the load current sequence, predicted low-disturbance window, and fixed-point scheduling of reconfiguration operation provided in the embodiments of this application;

[0035] Figure 5 A flowchart illustrating the power supply efficiency improvement method based on a prediction model provided in this application embodiment;

[0036] Figure 6 A schematic diagram showing the correspondence between the load current range and the target power supply module number in the module quantity comparison table provided for embodiments of this application;

[0037] Figure 7 A schematic diagram illustrating how the power supply efficiency varies with the number of power supply modules connected to the common bus to determine the target number of power supply modules, as provided in this application embodiment.

[0038] Figure 8 A schematic diagram of the bus voltage response verification process provided in the embodiments of this application;

[0039] Figure 9 This is a schematic diagram showing the comparison between the bus voltage response sequence, response characteristics, and window establishment template and window failure template provided in the embodiments of this application.

[0040] Figure 10 This is a schematic diagram of the structure of the echo determination module provided in the embodiments of this application;

[0041] Figure 11 A timing diagram illustrating the completion of redundant module role rotation within a low-disturbance prediction window using a first-access-then-switch method, as provided in the embodiments of this application.

[0042] Figure 12 This is a schematic diagram of the data flow for training the load prediction model provided in an embodiment of this application.

[0043] Figure labeling: 10-Power supply module; 20-Common bus; 30-Current sampling unit; 40-Load prediction model; 50-Cooperative controller; 60-Bus voltage sampling unit; 70-Echo judgment module; 71-Echo template library; 72-Window creation template; 73-Window failure template; 80-Load; 90-Load current sequence; 91-Power supply efficiency curve. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below 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. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0045] The use of terms like "first," "second," etc., in the embodiments of this application is only for distinguishing similar objects and does not represent a chronological order or degree of importance. The specific parameter values ​​given in the embodiments are examples used to help understand the scheme; they can be adjusted according to equipment conditions and process requirements during actual implementation.

[0046] Figure 1 This is a block diagram of the overall architecture of a power supply efficiency improvement system based on a predictive model provided in this application embodiment. The system includes multiple power supply modules 10, a common bus 20, a current sampling unit 30, a load prediction model 40, and a cooperative controller 50. The multiple power supply modules 10 are connected in parallel to the common bus 20 and supply power to the load 80 together through the common bus 20. Figure 1 The number of modules is indicated by two power supply modules (10) followed by an ellipsis, suggesting multiple modules. The actual number is determined by the rated capacity of the power supply device during deployment. This type of multi-module parallel power supply architecture is common in energy-sensitive applications such as server power supplies, communication power supplies, data center power supplies, and electric vehicle charging modules, where the load operates in a medium-to-light load range for a considerable period of time.

[0047] In this application, power supply efficiency refers to the ratio of the power output by the power supply device to the load 80 to the power obtained from its input terminal. The losses of the power supply device mainly consist of the switching losses and drive losses of each power supply module 10. The more power supply modules 10 connected to the common bus 20, the more dispersed the load current is, and the lower the conduction loss of each individual module; however, each connected power supply module 10 will generate a small amount of switching losses and drive losses dependent on its current carrying capacity. The more modules connected, the greater these losses accumulate with the number of modules. Therefore, for a given load, there is no such thing as the more or the fewer the better number of connected modules; rather, there exists a number of connected modules that maximizes power supply efficiency, which will be referred to as the target number of power supply modules in this application.

[0048] Each power supply module 10 switches on and off according to different interleaved phases. (Refer to...) Figure 2Taking four power supply modules 10-1 to 10-4 as an example, their switching frequencies are the same, but their turn-on times are evenly staggered within one switching cycle, with adjacent modules differing by a quarter of a switching cycle, i.e., equal-interval phases. The consideration for setting the phase in this way is that the current ripples induced by each switch on the common bus 20 are out of phase with each other, and after superposition, the fundamental frequency components cancel each other out. Only a much smaller ripple remains on the bus voltage, and the main frequency is raised to near the product of the number of connected power supply modules and the switching frequency. Figure 2 The bottom shows that the superimposed bus voltage has only a small ripple. Therefore, once the number of connections and the staggered phases are determined, the bus voltage quality is also determined. This is why, in the following text, each change in the number of connections requires simultaneous resetting of the equally spaced phases.

[0049] The current sampling unit 30 is used to collect the load current sequence 90 output from the common bus 20 to the load 80. In some embodiments, the current sampling unit 30 reuses the circuit originally used for output current sampling in the power supply device, such as a sampling resistor connected in series in the bus loop in conjunction with an analog-to-digital converter circuit, or a Hall current sensor, without the need to add an additional sensor for this application. The load current sequence 90 is a series of load current values ​​output by the sampling circuit at a fixed sampling rate, which is the data on which low-disturbance window prediction and identification are based. For example, the sampling rate can be from 10 kHz to 1 MHz, typically 100 kHz. When the sampling rate is too low, short-term smooth periods of the load current may be missed due to sparse sampling points; when the sampling rate is too high, the amount of data and subsequent computation increases, and there is no substantial improvement in the identification accuracy of the low-disturbance window.

[0050] The load prediction model 40 takes the load current sequence 90 collected by the current sampling unit 30 as input and outputs a prediction low-disturbance window and the predicted load current within that window. The coordinating controller 50 is connected to the switching drive of each power supply module 10 and determines whether and when to change the number and phase of the power supply modules 10 connected to the common bus 20 based on the output of the load prediction model 40. In some embodiments, the system further includes a bus voltage sampling unit 60 and an echo determination module 70. The bus voltage sampling unit 60 collects the bus voltage of the common bus 20, and the echo determination module 70 has a built-in echo template library 71, which includes a window-established template 72 and a window-empty template 73. The bus voltage sampling unit 60, the echo determination module 70, and the coordinating controller 50 cooperate to perform post-event verification of whether each change in the number of connected modules falls precisely during a period of low load. The connection and operation of these three components will be discussed later. Figures 8 to 10 Detailed explanation.

[0051] Based on the above system, this application provides a method for improving power supply efficiency based on a predictive model. (Refer to...) Figure 5The method generally includes the following steps: In step S101, the current sampling unit 30 collects the load current sequence 90 output from the common bus 20 to the load 80. In step S102, using the load current sequence 90 as input, the load prediction model 40 outputs a prediction low-disturbance window and the prediction load current within the prediction low-disturbance window. The prediction low-disturbance window includes the prediction start time and the prediction duration.

[0052] The system then determines whether the predicted duration is not less than the preset setup time for the reconfiguration operation. If the predicted duration is not less than the preset setup time, the process proceeds to step S103, where the cooperative controller 50 schedules the reconfiguration operation to be executed at the start of the prediction. The reconfiguration operation changes the number of power supply modules 10 connected to the common bus 20 to the target number, and resets the staggered phases of each power supply module 10 connected to the common bus 20 to equally spaced phases according to the target number of power supply modules. If the predicted duration is less than the preset setup time, the process proceeds to step S104, where the cooperative controller 50 maintains the current number of power supply modules 10 connected to the common bus 20 and returns to step S102 to continue predicting the next low-disturbance window with the new load current sequence 90. The implementation details of each of the above steps are explained below in conjunction with the accompanying drawings.

[0053] The acquisition of the load current sequence 90 and the identification of the low-disturbance window are the basis for this application to determine when it can operate. The low-disturbance window refers to the period during which the rate of change of the load current remains within a preset rate of change range, and the duration of this state is not less than the preset shortest window duration. The rate of change of the load current is calculated as follows: In one embodiment, the absolute value of the difference between the load currents of two adjacent sampling points is taken, divided by the sampling interval between these two sampling points, and then divided by the rated load current of the power supply device to obtain a rate of change value in terms of percentage of the rated current per millisecond.

[0054] The reason for defining the rate of change in this way is that dividing by the sampling interval makes the rate of change reflect the rate of current change per unit time and is independent of the sampling rate; dividing by the rated current makes it a dimensionless relative quantity and independent of the capacity of the power supply device. This allows the preset rate of change range to use the same set of values ​​across devices with different capacities and sampling rates. To suppress the interference of single-point noise on the rate of change, in some embodiments, a sliding window average is first performed on the load current sequence 90 before calculating the rate of change; the length of the sliding window can be 3 to 20 sampling points, typically 8 sampling points. For individual missing or significantly deviating abnormal sampling points, they can be removed or the values ​​of adjacent sampling points can be used to avoid contaminating the rate of change and subsequent judgment. A window that is too short will result in insufficient noise suppression, while a window that is too long will smooth out the actual load changes, causing the rate of change to respond too slowly to load jumps.

[0055] A preset rate of change range is used to define what level of current variation can be considered smooth. In some embodiments, the preset rate of change range is set to an upper limit between 0.5% per millisecond and 2% per millisecond, typically 1% per millisecond, meaning that a rate of change below this upper limit is considered smooth. If this upper limit is too small, the load must be extremely stable to be considered within a low-disturbance window, resulting in a scarcity of usable windows and reduced efficiency gains; if it is too large, periods of noticeable change are also included in the low-disturbance window, and the disturbances of the reconfiguration operation may no longer be masked by the load's own state. The preset minimum window duration is used to filter out false smoothness that is only momentary and insufficient to support a complete reconfiguration. Only when the rate of change remains within the preset rate of change range for a duration not shorter than the preset minimum window duration is that period registered as a low-disturbance window. Optionally, the preset minimum window duration can be set to 0.2 milliseconds to 2 milliseconds, typically 0.5 milliseconds. If the duration is too short, a large number of fleeting, smooth spikes will be registered as windows, most of which are not enough to accommodate a reconstruction. If the duration is too long, some short windows that could have been used will be excluded, reducing the opportunities for action.

[0056] It should be noted that the preset minimum window duration only determines whether a time period can be registered as a low-disturbance window. Whether a registered low-disturbance window can actually be used to perform reconstruction operations depends on the comparison between its predicted duration and the preset establishment time mentioned later. These are two independent thresholds. Therefore, there are low-disturbance windows whose predicted duration is not less than the preset minimum window duration but is still less than the preset establishment time. These windows are the objects processed in step S104 to keep the number of accesses unchanged.

[0057] The load forecasting model 40 takes the load current sequence 90 as input and outputs a predicted low-disturbance window and the predicted load current within that window. The predicted low-disturbance window includes the prediction start time and the prediction duration; the former is the expected arrival time of the next low-disturbance window, and the latter is the expected duration of the window remaining relatively flat. (Refer to...) Figure 4 The load current sequence 90 fluctuates dramatically on the time axis, sometimes entering a low-disturbance window with gradual changes. Before the window arrives, the load prediction model 40 predicts its start time and duration, and provides the predicted load current within that window. The predicted load current is used to subsequently determine the number of target power supply modules with optimal efficiency within that window. It is represented by a representative value of the load current within the window, exemplarily taking the average value of the load current within the window to reflect the approximate load level during the window period. The framework, training, and update methods of the load prediction model 40 will be discussed later. Figure 12 Detailed explanation.

[0058] After the load forecasting model 40 predicts the low-disturbance window, the cooperative controller 50 uses this prediction to determine whether and when to take action. The cooperative controller 50 compares the predicted duration with the preset setup time of the reconfiguration operation. The preset setup time refers to the time required for the bus voltage deviation caused by a reconfiguration operation to recover to within a preset voltage regulation range; that is, the transition time from the moment the number and phase of connections are changed until the bus voltage stabilizes again within the allowable range. The preset setup time is used as a threshold because the reconfiguration operation itself requires a certain amount of time for the bus voltage to stabilize. If the low-disturbance window is shorter than this transition time, the window will end before the reconfiguration stabilizes, and the load will immediately re-enter a period of fluctuation. The bus voltage deviation will then be exposed to load changes, losing the protection of the low-disturbance window.

[0059] The preset voltage regulation band is the allowable fluctuation range within which the bus voltage is considered to have recovered and stabilized. In some embodiments, the preset voltage regulation band is ±0.5% to ±2% of the rated bus voltage, typically ±1%. If the voltage regulation band is too narrow, the bus voltage must be extremely close to the rated value to be considered recovered, thus lengthening the preset settling time and reducing the available low-disturbance window. If it is too wide, the bus voltage is considered recovered even with a perceptible deviation, and the transitional effects of reconfiguration may extend beyond the window. The preset settling time can be measured offline during installation or calibration. This is done by manually triggering several reconfiguration operations under typical operating conditions, recording the time it takes for the bus voltage to deviate and fall back into the preset voltage regulation band, and taking the larger value or adding a certain margin as the preset settling time. For example, the preset settling time is in the range of 0.2 milliseconds to 5 milliseconds, typically 1 millisecond. Its specific value varies depending on the output capacitance of the power supply unit and the bandwidth of the control loop. A value that is too small may cause an unstable reconfiguration to be mistakenly judged as feasible, while a value that is too large may exclude some windows that would otherwise be sufficient.

[0060] When the predicted duration is not less than the preset setup time, the cooperative controller 50 determines that the low-disturbance window is sufficient to accommodate a complete reconfiguration, and therefore schedules the reconfiguration operation to be executed at the start of the prediction. The significance of aligning the action time with the start of the prediction is that the load current changes gradually during the window period, and the bus voltage deviation caused by the reconfiguration is superimposed on the nearly constant load. Its amplitude and recovery process are not entangled with the changes in the load itself. Therefore, this deviation is masked by the smooth state of the load and will not be perceived as a power supply abnormality by the load side.

[0061] As a correction to the fixed-point timing, some embodiments consider that the prediction start time itself has prediction errors, and execution close to the left boundary of the window is most likely to fall outside the window due to this error. Therefore, the cooperative controller 50 shifts the actual execution time inward from the prediction start time by a protection shift amount; thus, even if the prediction start time is slightly earlier or later, the reconstruction operation still falls within the smooth period. The protection shift amount is only enabled when the prediction duration has a margin relative to the preset setup time, and its value does not exceed the difference between the prediction duration and the preset setup time. For example, it can be a portion of this difference, or the smaller of 5% to 20% of the prediction duration and this difference, typically 10% of the prediction duration. This constraint ensures that the remaining duration within the window after shifting is still not less than the preset setup time required to complete the reconstruction; when the prediction duration is comparable to the preset setup time but the margin is insufficient, the protection shift amount is zero, and the reconstruction operation is still executed aligned with the prediction start time. In embodiments where echo verification is enabled (described later), this margin is reduced by the preset observation window accordingly.

[0062] The aforementioned prediction and determination do not consume real-time resources at the switching stage. Whether the load current is smooth or not is a process that evolves on a millisecond or even longer timescale. The load prediction model 40 operates as a slow monitoring task, with its inference cycle time much lower than the switching frequency of the power supply module 10, and is decoupled from the switching timing loop that maintains interleaved phases. The switching timing loop still drives each power supply module 10 at a fixed cycle; the load prediction model 40 only updates its prediction for the next low-disturbance window on its own low-speed cycle, and the two do not interrupt each other. For a load prediction model 40 with thousands to hundreds of thousands of parameters, the computational load of a single inference is on the order of thousands to hundreds of thousands of multiply-accumulate operations. The load prediction model 40 performs inference once every prediction look-ahead, with a cycle time on the order of milliseconds, much longer than the switching cycle, which can be easily completed by the controller of a conventional power supply device. Therefore, implementing this application does not require additional high-speed computing hardware for prediction; the controller in a conventional power supply device can handle it.

[0063] This application distinguishes itself from conventional switching strategies by replacing the traditional fixed hysteresis with a predicted low-disturbance window. The conventional approach involves reserving a fixed hysteresis period after the load current drops below a threshold, ensuring sufficient load stability before reducing the number of modules. Because this waiting time must be sufficient for worst-case scenarios to prevent switching during sudden load changes, a significant amount of waiting occurs under light loads, leading to repeated delays in module reduction and continuous idle consumption of excess modules during this period. This application no longer avoids load changes by widening the hysteresis or delaying the action. Instead, the load prediction model 40 proactively identifies a low-disturbance window where the load changes smoothly, placing the switching point within this window. The disturbance is masked by the smooth load condition, thus narrowing the hysteresis range, allowing for timely module reduction, and ensuring the number of connected modules is closer to the optimal target number of power supply modules.

[0064] The reconfiguration operation involves two simultaneous actions: first, changing the number of power supply modules 10 connected to the common bus 20 to the target number of power supply modules; and second, resetting the interleaved phases of each power supply module 10 connected to the common bus 20 to equally spaced phases according to the changed target number of power supply modules. (Refer to...) Figure 3 Before the reconfiguration, four power supply modules 10 were connected to the common bus 20, and their staggered phases corresponded to four equal division points on the phase circle, namely 0 degrees, 90 degrees, 180 degrees and 270 degrees. The reconfiguration operation reduced the number of connected modules to two. The two unconnected power supply modules 10 are shown as dashed lines. At the same time, the phases of the two still connected power supply modules 10 were reset to two equal division points on the phase circle, namely 0 degrees and 180 degrees. Figure 3 The phase circle in the diagram represents one switching cycle, and the equidistant points on the circumference represent the equally spaced interleaved phases of each connected power supply module 10. The reason why the phase must be reset when the number of connected modules changes is that the ripple cancellation effect of the interleaved phase depends on the uniform distribution of each connected module within the switching cycle; once the number of connected modules changes without corresponding phase adjustment, the ripple cancellation is destroyed, and the bus voltage ripple increases significantly.

[0065] The target number of power supply modules is the number that ensures the system's power supply efficiency is no lower than that of other connected modules under the predicted load current. (Refer to...) Figure 7 Under a given predicted load current, the power supply efficiency curve 91, which shows a single-peak shape where the power supply efficiency varies with the number of power supply modules 10 connected to the common bus 20, first rises and then falls: when the number of connected modules is too small, the current carried by each module is too large, resulting in high conduction losses; when the number of connected modules is too large, the switching losses and drive losses introduced by the no-load component are too high. The number of connected modules corresponding to the peak value of the power supply efficiency curve 91 is the target number of power supply modules under the predicted load current. Due to different loads, the peak position of the power supply efficiency curve 91 shifts accordingly; the heavier the load, the greater the number of connected modules required for optimal efficiency. Therefore, adjusting the number of connected modules to the target number of power supply modules means placing the power supply device near its optimal efficiency under the current load.

[0066] The load prediction model 40 is trained based on historical load current sequences. (Refer to...) Figure 12The training data stream of the load prediction model 40 includes three stages: sample construction, label generation, and model fitting. Training samples are taken from historical load current sequences, which are divided into multiple sampling segments, each serving as a training input. The actual load current sequences following the sampling segments are used to generate training labels. Training labels include low-perturbation window labels and load current labels. The low-perturbation window label is determined by the start time and duration of the actual load current sequence following the sampling segment, where the rate of change remains within a preset range and the duration is not less than a preset minimum window length. The load current label is the load current within that actual period. Thus, the model's input is a newly observed load current segment, and the output is the subsequent low-perturbation window and its representative load current value. Training uses the regression error of the predicted start time, predicted duration, and predicted load current against the corresponding labels as the loss function, for example, the weighted mean square error of each error. Iteration continues until the window hit rate and load current error on the validation set no longer decrease significantly. To ensure consistency between the training data and the subsequent table lookup data, the load current label uses the same representative value of the load current during the actual time period.

[0067] Whether the load current is smooth or not is essentially a time-series pattern recognition problem. The load prediction model 40 can be implemented using a lightweight time-series regression or sequence model. In some embodiments, the load prediction model 40 employs a multilayer perceptron with a sliding window input or a one-dimensional convolutional network, or a lightweight recurrent network such as a gated recurrent unit. Its input features, in addition to the original load current of the sampling segment, also incorporate the aforementioned rate of change sequence and statistical quantities such as the moving mean and moving variance of the load current to help the model identify the start and end of the smooth segment. The model size does not need to be large. In one possible implementation, a network parameter count between several thousand and hundreds of thousands is sufficient to characterize the short-term smooth pattern of the load current. If the parameter count is too small, the model will struggle to fit diverse load patterns; if it is too large, it will increase inference overhead with limited benefits.

[0068] The duration of the sampling segment determines the length of observations the model uses for each prediction. In some embodiments, this can range from 5 to 100 milliseconds, with 20 milliseconds being typical. If the sampling segment is too short, the model lacks sufficient context; if it is too long, the response to load changes becomes slower. The prediction look-ahead is the lead time between the prediction start time and the end of the sampling segment. For example, it can range from 1 to 20 milliseconds, with 5 milliseconds being typical. If the look-ahead is too small, the collaborative controller 50 will not have enough time to complete scheduling preparations before the window arrives; if the look-ahead is too large, the prediction error increases. The load prediction model 40 can be initially trained using historical load current sequences collected offline before installation, and then updated incrementally with newly accumulated load current sequences periodically after deployment, allowing the model to gradually adapt to the specific load patterns at the site. During the cold start phase when there are insufficient samples, i.e. when the historical load current sequence is not enough to train a reliable model in the early stage of installation, the coordinating controller 50 does not temporarily follow the output action of the load prediction model 40. Instead, it reverts to switching according to the module number comparison table below and the traditional hysteresis strategy. After the samples have accumulated to a certain scale and the model accuracy has reached the standard, it switches to fixed-point scheduling driven by the load prediction model 40.

[0069] The aforementioned reconfiguration operation resets the staggered phases of the power supply modules 10 connected to the common bus 20 to equally spaced phases. Specifically, a switching cycle is divided into an equal number of phase segments according to the number of power supply modules 10 currently connected to the common bus 20, and the distinct phase segments are assigned to each power supply module 10 connected to the common bus 20 as its staggered phase. In one embodiment, when the number of connected modules is N, the conduction time of two adjacent power supply modules 10 differs by one-Nth of a switching cycle, so that the N switches are connected end-to-end and evenly distributed within one cycle. Figure 3 When two power supply modules 10 are connected on the right, the adjacent phases differ by half a switching cycle, which is the case when N is 2. The aforementioned preset setup time is the time required for the bus voltage deviation caused by the reconfiguration operation to recover to the preset voltage regulation range. Its value and calibration method have been explained above.

[0070] In some embodiments, the cooperative controller 50 determines the target number of power supply modules using a preset module quantity lookup table. The module quantity lookup table stores the correspondence between each load current range and the target number of power supply modules. The cooperative controller 50 retrieves the corresponding target number of power supply modules from the module quantity lookup table based on the load current range to which the predicted load current belongs. (Refer to...) Figure 6The horizontal axis represents the load current, and the vertical axis represents the number of target power supply modules. The relationship between the two increases in a stepped manner; the higher the load current falls in a certain range, the greater the number of target power supply modules required. Each load current range is defined by two adjacent boundaries. The reason for using a lookup table instead of real-time optimization is that the relationship between the optimal number of connected modules and the load current can be determined offline once during the installation or calibration phase. During runtime, only one lookup table is needed to obtain the number of target power supply modules, eliminating the need to repeatedly calculate the efficiency curve within each low-disturbance window, further reducing the real-time burden on the cooperative controller 50.

[0071] The calibration of the module quantity lookup table can be completed through actual measurement or simulation. In one embodiment, the power supply device is operated stably with different numbers of connected modules at several typical load current points, and the power supply efficiency under each number of connected modules is measured. The number of connected modules corresponding to the highest efficiency is taken as the target number of power supply modules for that load current point. Then, adjacent load current points with the same number of target power supply modules are merged into one load current interval, thus obtaining the module quantity lookup table. When the equipment model or operating condition changes, the above calibration only needs to be performed again, without modifying the decision logic of the cooperative controller 50.

[0072] Reference Figure 5 When it is determined that the predicted duration is less than the preset establishment time, the cooperative controller 50 enters step S104, maintaining the number of power supply modules 10 currently connected to the shared bus 20 unchanged until the predicted duration of the low-disturbance window output by the load prediction model 40 is not less than the preset establishment time. The purpose of this branch is to ensure that when a sufficiently long low-disturbance window does not exist in the short term, the number of connected modules is not adjusted temporarily, rather than switching within a short window that cannot accommodate the reconfiguration process, so as to prevent disturbances from falling outside the smooth period. Therefore, when there is a lack of available windows, the system automatically degrades to maintaining the status quo, and the power supply efficiency cannot be further improved temporarily; however, the power supply quality is not affected, and adjustments are made after a suitable low-disturbance window appears.

[0073] The predictive capability of the load prediction model 40 relies on the existence of identifiable operating patterns in the load. In situations where the load is predictable in the short term, such as the gradual load change interval between two computing tasks in a central processing unit, or the gradual load shifting plateau when a charging module transitions from a constant current phase to a constant voltage phase, the load prediction model 40 can accurately predict low-disturbance windows. However, in situations where the load is completely acyclic and trendless, the predicted duration generally does not reach the preset setup time. In this case, the coordinating controller 50 continues to maintain the current number of connected devices according to step S104, and the system degenerates into operating according to the module number lookup table and the traditional hysteresis strategy. The power supply function is unaffected; only the opportunity for efficiency improvement is reduced. Therefore, this application sets up two layers of protection for situations where the prediction accuracy is insufficient: first, no action is taken when the predicted duration is insufficient to avoid switching during periods that are not smooth; second, even if the predicted duration reaches the target and action is taken, the subsequent echo verification will determine whether the action truly falls within a smooth period, and if it does not, it will immediately revert. Both work together to maintain the lower limit of power quality, ensuring that the system at most loses efficiency gains without sacrificing power quality when there are occasional deviations in the prediction.

[0074] Occasional prediction failures are a situation that look-ahead prediction cannot completely avoid. In some embodiments, to correct prediction failures in a timely manner, the system also includes a bus voltage sampling unit 60 and an echo determination module 70 to perform post-event verification on whether each reconstruction operation falls precisely during a smooth period. The bus voltage sampling unit 60 is used to collect the bus voltage echo sequence of the shared bus 20 within a preset observation window after the reconstruction operation is executed, and its sampling rate can be consistent with that of the current sampling unit 30. In some embodiments, the bus voltage sampling unit 60 reuses the original bus voltage sampling circuit of the power supply device voltage control loop, without the need to add additional sensors for echo verification. The bus voltage echo sequence is a small segment of waveform showing the change of the bus voltage over time after the reconstruction operation is applied, which is an unavoidable byproduct of switching and phase resetting on the bus.

[0075] The preset observation window is a short observation period immediately following the reconstruction operation. In some embodiments, it can be 0.5 milliseconds to 5 milliseconds, typically 2 milliseconds, approximately twice the preset setup time, to cover the complete process of bus voltage deviation to recovery. If the observation window is too short, the recovery segment of the echo will not be completed before data acquisition stops; if the observation window is too long, it will delay the verification conclusion and may exceed the low-disturbance window. Accordingly, in embodiments where echo verification is enabled, the cooperative controller 50 tightens the window threshold for reconstructive operations from not less than the preset setup time to not less than the sum of the preset setup time and the preset observation window, so that switching, observation, and even possible rollback all fall within the same low-disturbance window as much as possible.

[0076] The echo determination module 70 extracts echo features from the bus voltage echo sequence. (Refer to...) Figure 10The bus voltage sampling unit 60 sends the acquired bus voltage echo sequence to the echo determination module 70. Within the echo determination module 70, the echo feature extraction stage first calculates the echo features, and then the distance measurement stage compares the echo features with the window-established template 72 and the window-empty template 73 in the echo template library 71. The echo features are characteristic quantities representing the deviation amplitude and deviation recovery process of the bus voltage echo sequence. The distance measurement is used to measure the closeness between the echo features and each template feature. In some embodiments, the distance measurement uses Euclidean distance, which is the square root of the sum of the squares of the differences between the echo feature vector and each component of the template feature vector. In other embodiments, considering the different dimensions of the deviation peak, deviation recovery time, and oscillation number, the distance measurement uses Mahalanobis distance, weighted by the statistical covariance of each feature, to avoid the large-dimensional component dominating the distance. In still other embodiments, if the sampling length of each echo is different, the distance measurement can also use dynamic time warping.

[0077] The echo determination module 70 uses the determination result corresponding to the template with the smallest distance metric as the window verification result. Optionally, a template determination threshold is also set. When the distance between the echo feature and the nearest template is still greater than the template determination threshold, the echo is considered unreliable and the echo determination module 70 handles it conservatively as a window failure. The template determination threshold can be 1 to 3 times the distance from the same sample to its template center when the template is constructed, typically 2 times. If the threshold is too small, normal echo fluctuations will also be judged as unclassifiable and will frequently trigger conservative backoff. If the threshold is too large, the ability to distinguish abnormal echoes will be lost. When the window verification result is a window failure, the collaborative controller 50 restores the number of power supply modules 10 connected to the common bus 20 to the number before the reconstruction operation, that is, cancels the reconstruction.

[0078] Reference Figure 8 The bus voltage echo verification process is as follows: In step S201, the bus voltage echo sequence is collected within the preset observation window; in step S202, the echo features consisting of the deviation peak, the deviation recovery time, and the number of oscillations are extracted; in step S203, the echo features are compared with the window establishment template 72 and the window failure template 73 to measure the distance; in step S204, the judgment result of the template with the smallest distance measurement is taken as the window verification result; then it is determined whether the window verification result is a window failure. If so, step S205 is entered to restore the number of connections to the state before reconstruction. Otherwise, step S206 is entered to keep the current number of connections unchanged.

[0079] The computational load of echo verification is very small. Extracting echo features only requires calculating the deviation peak, deviation recovery time, and number of oscillations of the bus voltage echo sequence within a preset observation window; the distance metric is simply a comparison between the echo feature vector and a finite number of template vectors. Accordingly, the computational load of a single verification is on the order of hundreds of floating-point operations, which can be completed within a preset observation window in milliseconds; echo determination also runs on a low-speed cycle decoupled from the switching timing loop, without occupying real-time resources at the switching level, and without requiring additional acceleration hardware.

[0080] Using the bus voltage response to indirectly determine the authenticity of the window has the advantage of timing compared to directly measuring the load current after reconstruction and then judging whether it is smooth. The response is generated at the moment the reconstruction operation is applied and can be collected within the preset observation window. Based on this, the window verification result can be obtained in a very short time after switching, and rollback can be performed if necessary. On the other hand, directly observing whether the load current returns to smoothness after reconstruction requires waiting for the subsequent load current sequence of the window to gradually unfold before judgment can be made. By then, even if the error of the operation is found, the disturbance has already been applied and the opportunity has been missed. Therefore, using the response accompanying the switching itself as the verification basis allows even occasional prediction failures to be detected and corrected immediately at the moment of the operation.

[0081] The reason why the echo can reflect the authenticity of the switching window lies in the fact that the bus voltage response to the same switching is modulated by the load state at that time. When the window is valid, the load is almost constant, and the bus voltage deviation caused by the switching is the only disturbance source. The voltage control loop of the power supply unit only needs to suppress this step, the bus voltage deviation is small, the recovery is fast, and the oscillation decays quickly. When the window is not valid, the load itself is still changing, and the switching deviation is superimposed on the continuously changing load disturbance. The control loop needs to deal with both disturbances at the same time, the bus voltage deviation is larger, the recovery is slower, and there are more oscillations. It is this difference that makes the echoes of the window being valid and the window being not valid exhibit separable patterns in terms of deviation peak value, deviation recovery time, and number of oscillations.

[0082] The echo characteristics include the deviation from the peak value, the deviation recovery time, and the number of oscillations in the bus voltage echo sequence. (Refer to...) Figure 9The diagram illustrates the measured bus voltage echo sequence, represented by solid lines and comparison templates, with the upper subplot representing template 72 (window established) and the lower subplot representing template 73 (window not established). The dashed lines are slightly smaller in amplitude than the solid lines and located inside them to distinguish them from the measured sequence. The upper subplot shows the window established scenario, where the bus voltage echo sequence has a small peak deviation, short recovery time, and few oscillations, with the measured sequence closely approximating template 72. The lower subplot shows the window not established scenario, where the peak deviation is large, recovery time is long, and there are many oscillations, with the measured sequence closely approximating template 73. These three characteristics characterize the amplitude, speed, and decay pattern of the echo, respectively, distinguishing between echoes during smooth and non-smooth periods. The separability of the echo characteristics is predicated on the underdamped decay of the bus voltage echo, a common scenario for multi-module parallel power supply devices under switching disturbances. In the case of overdamping, the bus voltage response does not oscillate significantly, and the number of oscillations degenerates and loses its discriminative power. At this time, the window establishment and window failure are mainly distinguished by the deviation from the peak value and the deviation from the recovery time. For boundary responses with very small amplitude that fall between the window establishment and the window failure, they are identified by the aforementioned template judgment threshold and handled in a conservative manner.

[0083] The window establishment template 72 and window failure template 73 in the echo template library 71 are both generated based on historical reconstruction operations. Each template includes a template feature vector and a template determination label. The template feature vector is a representative value of the echo features corresponding to a certain type of historical reconstruction operation, exemplarily taking the mean or median of the echo features of similar samples. The template determination label is determined by whether the load current after the corresponding historical reconstruction operation is executed remains within a preset rate of change range. If it remains within the preset rate of change range, it is determined that the window is established; if it does not remain within the preset rate of change range, it is determined that the window is failed. The generation of these templates can be regarded as a lightweight sample accumulation and statistical process. The samples are each actual reconstruction operation and its echo that the power supply device has executed in the past. They are continuously accumulated as the operation progresses and can be updated periodically with new samples, so that the templates remain relevant as the equipment ages and operating conditions change. In the cold start phase when there are no historical reconstruction samples in the early stage of installation, the echo verification is not enabled for the time being, or several reconstructions are manually triggered in the calibration phase to collect the bus voltage echo to establish an initial template. After the samples have accumulated to a certain scale, the update is switched to the running samples, which corresponds to the cold start processing of the aforementioned load prediction model 40.

[0084] The echo characteristics depend on whether the action falls within a flat period and the number of power supply modules 10 changed in the reconfiguration. The more modules changed, the greater the deviation of the bus voltage, and the stronger the echo. Therefore, in some embodiments, the echo template library 71 is grouped according to the reconfiguration magnitude, and separate window-establishment templates 72 and window-empty templates 73 are created for different reconfiguration magnitudes to avoid misjudgment caused by comparing the echo of reducing one module with the echo of reducing three modules. Furthermore, the reconfiguration magnitude is included in a signed manner, the direction of bus voltage deviation caused by adding and removing modules is opposite, and adding two modules and removing two modules do not share the same set of templates.

[0085] In other embodiments, the number of multiple power supply modules 10 is greater than the target number of power supply modules. The number of power supply modules 10 currently connected to the common bus 20 is the target number of power supply modules, and the remaining power supply modules 10 not connected to the common bus 20 are considered redundant modules. Redundant modules do not participate in power supply temporarily and are added later when the number needs to be increased or rotated. It is important to distinguish that the target number of power supply modules is a numerical value, while the target power supply module referred to below is a specific power supply module 10; the two names are similar but have different meanings.

[0086] If a few power supply modules 10 are consistently used as redundant modules, the problem lies in the uneven aging process. The modules designated as redundant operate without load for extended periods, while the remaining power supply modules 10 continuously bear the load; consequently, the cumulative load time and aging degree of each power supply module 10 differ significantly. The cumulative load time refers to the accumulated time a corresponding power supply module 10 spends connected to the shared bus 20, obtained by the coordination controller 50 by summing the connection times of each power supply module 10, reflecting the actual usage and aging process of that module. To balance the cumulative load time and aging of each power supply module 10, the coordination controller 50 does not permanently assign redundant power supply modules 10, but instead periodically performs role rotation operations.

[0087] The collaborative controller 50 determines the target power supply module while keeping the number of target power supply modules constant, and schedules the role rotation operation to be executed within the predicted low-disturbance window. The target power supply module is the power supply module 10 with the longest cumulative carrying time among the power supply modules 10 currently connected to the shared bus 20, i.e., the one with the longest current connection time. The reason for selecting the one with the longest cumulative carrying time is that it has accumulated the most carrying and aging, and cutting it off and replacing it with a redundant module can make the cumulative carrying time of each power supply module 10 more consistent. Correspondingly, when the reconfiguration operation requires reducing the number of connections rather than rotating an equal number, the collaborative controller 50 also prioritizes cutting off the power supply module 10 with the longest cumulative carrying time among the currently connected power supply modules 10; thus, the module reduction and role rotation adopt a consistent approach in selecting the modules to be cut off, avoiding the creation of new uneven aging caused by the fixed cutting off of the same few modules.

[0088] The role rotation operation is performed in the order of first connecting and then disconnecting to ensure that the number of connected roles remains constant before and after the rotation. (Refer to...) Figure 11 Taking power supply modules 10-1 to 10-4 as examples, power supply module 10-1 is the target power supply module with the longest cumulative carrying time, and power supply module 10-4 is a redundant module to be connected. Within the predicted low-disturbance window, power supply module 10-4, as a redundant module, is first connected to the common bus 20 and participates in the switching, at which point the number of connected modules briefly increases; then, the target power supply module 10-1 is disconnected from the common bus 20 and converted to a redundant module, restoring the number of connected modules. After completing the above connection and disconnection, the staggered phase of each connected module is reset to an equally spaced phase according to the current number of power supply modules 10 connected to the common bus 20, so that the number of connected modules remains unchanged before and after the role rotation operation. The reason for using the method of connecting first and then disconnecting instead of disconnecting first and then connecting is that disconnecting first will cause the number of connected modules to momentarily fall below the number of target power supply modules, causing an additional underconfiguration disturbance; connecting first ensures that the number of connected modules is always not lower than the number of target power supply modules, and the process does not introduce any additional perceptible disturbance.

[0089] Both role rotation and reconfiguration operations utilize the predicted low-disturbance window for execution, and their order must be coordinated within the same window. In some embodiments, they are mutually exclusive within the same predicted low-disturbance window, with reconfiguration taking priority. When a window requires both a change in the number of connections and reaches its rotation cycle, the reconfiguration operation is completed first within the window, and role rotation is postponed to the next low-disturbance window that meets the conditions. Role rotation is only performed within a window if it does not require a change in the number of connections. Since the number of connections remains unchanged before and after the role rotation operation, and the net disturbance caused on the bus is small, echo backoff is not enabled separately for it in some embodiments. In other embodiments, echo verification can be enabled for it, in which case the increase or decrease corresponding to the order of connection and disconnection is classified into the corresponding template group for comparison. The consideration for this arrangement is that the reconfiguration operation is directly related to the power supply efficiency under the current load and has a stronger timeliness, while the role rotation operation only concerns long-term aging balancing and can be postponed; the staggered arrangement also avoids the superposition of two consecutive disturbances within the same window.

[0090] The aforementioned point-to-point switching, echo verification, and role rotation work together around a single predicted low-disturbance window. Point-to-point switching determines the optimal number of connected power supply modules within this window; echo verification safeguards power quality when the prediction fails within the same window; and role rotation uses the same window to mitigate uneven aging of redundant modules. All three share the same prediction for periods of low load, ensuring that multi-module power supply devices, operating under medium to light loads for extended periods, closely approximate the optimal connection configuration without sacrificing power quality due to switching disturbances.

[0091] From a device perspective, the power supply efficiency improvement system based on a predictive model provided in this application is the implementing entity of the aforementioned method. (Refer to...) Figure 1 and Figure 10 The current sampling unit 30 is responsible for acquiring the load current in step S101; the load prediction model 40 is responsible for the low-disturbance window and predicted load current output in step S102; the cooperative controller 50 is responsible for comparing the predicted duration with the preset establishment time, as well as the fixed-point reconstruction in step S103 and the hold-on in step S104, and determining the target power supply module quantity according to the module quantity comparison table and the role rotation according to the cumulative bearing time. The bus voltage sampling unit 60 and the echo judgment module 70 are responsible for echo verification.

[0092] Reference Figure 10 The echo determination module 70 includes two stages: echo feature extraction and distance measurement, as well as an echo template library 71. The echo template library 71 stores window establishment templates 72 and window failure templates 73. The bus voltage echo sequence output by the bus voltage sampling unit 60 first obtains echo features through the echo feature extraction stage, and then compares them with the window establishment template 72 and window failure template 73 through the distance measurement stage. The window verification result is sent back to the collaborative controller 50, which determines whether to backtrack. Each unit and module can be implemented by discrete circuits or integrated into the same digital controller, undertaking the corresponding processing as a software module.

[0093] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Those skilled in the art should understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features, and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A power supply efficiency improvement system based on a prediction model, characterized by, It includes multiple power supply modules, as well as a shared bus, current sampling unit, load prediction model and cooperative controller; The multiple power supply modules are connected in parallel to the common bus and supply power to the load through the common bus. Each power supply module switches on and off according to different staggered phases. The current sampling unit is used to collect the load current sequence output from the common bus to the load; The load prediction model is used to take the load current sequence collected by the current sampling unit as input, and output a predicted low disturbance window and the predicted load current within the predicted low disturbance window. The predicted low disturbance window is a prediction of the low disturbance window and includes the prediction start time and prediction duration. The low disturbance window is the period in which the rate of change of the load current remains within a preset rate of change range and the duration is not less than the preset shortest window duration. The cooperative controller is used to schedule the reconstruction operation to be executed at the start time of the prediction when the prediction duration is not less than the preset establishment time of the reconstruction operation. The reconstruction operation is to change the number of power supply modules connected to the common bus to the target number of power supply modules, and reset the staggered phase of each power supply module connected to the common bus to an equal-interval phase according to the target number of power supply modules. The target number of power supply modules is the number of modules that make the power supply efficiency of the system not lower than the power supply efficiency under the predicted load current. It also includes a bus voltage sampling unit and an echo determination module; The bus voltage sampling unit is used to collect the bus voltage echo sequence of the common bus within a preset observation window after the reconstruction operation is performed; The echo determination module has a built-in echo template library, which includes window creation templates and window failure templates. The echo determination module is used to extract echo features from the bus voltage echo sequence. The echo features are feature quantities that characterize the deviation amplitude and deviation recovery process of the bus voltage echo sequence. The echo features are used to calculate the distance metric between the window establishment template and the window failure template, respectively. The determination result corresponding to the template with the smallest distance metric is used as the window verification result. The collaborative controller is also used to restore the number of power supply modules connected to the common bus to the number before the reconstruction operation was performed when the window verification result is that the window is empty.

2. The power supply efficiency improvement system based on a prediction model according to claim 1, characterized by, The load prediction model is trained based on historical load current sequences. The input during training is a sample segment from the historical load current sequence, and the output during training is a predicted low-disturbance window and a predicted load current. The training labels are a low-disturbance window label and a load current label. The low-disturbance window label is determined by the start time and duration of an actual period in the actual load current sequence after the sample segment, where the rate of change remains within the preset rate of change range and the duration is not less than the preset shortest window duration. The load current label is the load current within the actual period.

3. The power supply efficiency improvement system based on a prediction model according to claim 1, characterized by, The cooperative controller resets the interleaved phases of each power supply module connected to the common bus to equally spaced phases by dividing a switching cycle into the same number of phase segments according to the number of power supply modules connected to the common bus, and assigning the different phase segments to each power supply module connected to the common bus as its interleaved phase. The preset establishment time is the time required for the bus voltage of the common bus to deviate and recover to the preset voltage regulation range caused by the reconstruction operation.

4. The power supply efficiency improvement system based on a prediction model according to claim 1, characterized by, The collaborative controller determines the number of target power supply modules by using a preset module quantity lookup table, which stores the correspondence between each load current range and the number of target power supply modules. The collaborative controller determines the corresponding target power supply module quantity from the module quantity lookup table based on the load current range to which the predicted load current belongs.

5. The power supply efficiency improvement system based on a predictive model according to claim 1, characterized in that, The collaborative controller is also used to maintain the number of power supply modules currently connected to the shared bus unchanged when the predicted duration is less than the preset establishment time, until the predicted duration of the low-disturbance window output by the load prediction model is not less than the preset establishment time.

6. The power supply efficiency improvement system based on a predictive model according to claim 1, characterized in that, The echo characteristics include the deviation peak value, deviation recovery time, and number of oscillations of the bus voltage echo sequence; The window establishment template and the window failure template are generated based on historical reconstruction operations. Each template includes a template feature vector and a template determination label. The template determination label is determined by whether the load current after the corresponding historical reconstruction operation is maintained within the preset rate of change range. If it is maintained within the preset rate of change range, the window is determined to be established. If it is not maintained within the preset rate of change range, the window is determined to be failed.

7. The power supply efficiency improvement system based on a predictive model according to claim 1, characterized in that, The number of multiple power supply modules is greater than the target number of power supply modules. The number of power supply modules currently connected to the common bus is the target number of power supply modules. The remaining power supply modules not connected to the common bus are used as redundant modules. The cooperative controller is also used to determine the target power supply module under the premise that the number of target power supply modules remains unchanged, and to schedule the role rotation operation to be executed within the predicted low disturbance window. The role rotation operation is to switch the target power supply module from being connected to the common bus to a redundant module and to switch a current redundant module to be connected to the common bus. The target power supply module is the power supply module with the longest cumulative carrying time among the power supply modules currently connected to the common bus. The cumulative carrying time is the cumulative value of the time that the corresponding power supply module has been in the state of being connected to the common bus.

8. The power supply efficiency improvement system based on a predictive model according to claim 7, characterized in that, The role rotation operation includes: within the predicted low-disturbance window, first connecting a current redundant module to the common bus and participating in the switching, then cutting off the target power supply module from the common bus and switching it to a redundant module, and resetting the staggered phases of each power supply module connected to the common bus to equally spaced phases according to the number of power supply modules currently connected to the common bus, so that the number of power supply modules connected to the common bus remains unchanged before and after the role rotation operation.

9. A power supply efficiency improvement method based on a predictive model, applied to the power supply efficiency improvement system based on a predictive model as described in claim 1, characterized in that, include: Collect the load current sequence output from the common bus to the load; Using the load current sequence collected by the current sampling unit as input, the load prediction model outputs a predicted low-disturbance window and the predicted load current within the predicted low-disturbance window. The predicted low-disturbance window is a prediction of the low-disturbance window and includes the prediction start time and prediction duration. The low-disturbance window is a period in which the rate of change of the load current remains within a preset rate of change range and the duration is not less than the preset shortest window duration. When the predicted duration is not less than the preset establishment time of the reconfiguration operation, the reconfiguration operation is scheduled to be executed at the start time of the prediction. The reconfiguration operation is to change the number of power supply modules connected to the common bus to the target number of power supply modules, and reset the staggered phase of each power supply module connected to the common bus to an equal-interval phase according to the target number of power supply modules. The target number of power supply modules is the number that makes the power supply efficiency of the system not lower than the power supply efficiency under the predicted load current.

Citation Information

Patent Citations

  • Direct current electric energy conversion method and system with DCDC module automatic conversion function

    CN107659150A

  • Multi-path power supply parallel current output device and method and electronic equipment

    CN115242059A

  • Modularized intelligent power supply rapid switching method and system for power grid fault

    CN121566466A

  • Power management control system and control method

    CN122068595A