An intelligent power efficiency scheduling method and system based on load prediction

Abstract

The application discloses an intelligent power supply energy efficiency scheduling method and system based on load prediction, relates to the field of power management and control, and comprises the following steps: inputting an output voltage sequence and an output current sequence of a power converter into a machine learning load interval prediction model to output a lower limit and an upper limit of future load power; determining the lower limit as certain power, controlling a main power branch to lock at a preset efficiency optimal working point with the certain power as a target; determining the difference between the upper limit and the lower limit as to-be-responded power, querying a standby configuration template library to configure the number of auxiliary power branches and standby current values; when the to-be-responded power is smaller than a preset standby exemption threshold, turning off all auxiliary power branches and entering an energy-saving mode; and the part of actual load power exceeding the certain power is borne by the auxiliary power branches maintaining the standby current values. The application solves the problem that the power supply is in a low-efficiency high-response state for a long time to cope with load surge.

Description

Technical Field

[0001] This application relates to the field of power management and control technology, and in particular to an intelligent power energy efficiency scheduling method and system based on load prediction. Background Technology

[0002] Processors in AI servers and data centers are powered by switching power converters. The power consumption of these loads changes far faster than in typical electronic devices; during a single computational task, the load current can rise from tens of amperes to hundreds of amperes within microseconds. If the supply voltage drops too sharply during a load surge, the processor may reset or malfunction, thus requiring the power converter to maintain a rapid response capability at all times. To handle high currents, server power supplies typically consist of multiple parallel power branches, each of which can be independently switched on or off. Re-energizing a switched-off power branch requires a process involving power-on of the drive circuit, inductor current building up from zero, and control loop recapture, taking tens of microseconds. A power branch that is already running and maintaining current can increase its output in just a few switching cycles. Furthermore, the conversion efficiency of the power branches is not constant; there exists a switching frequency that maximizes conversion efficiency at a given output power, and deviations from this frequency result in increased losses.

[0003] For the aforementioned loads, existing technologies have developed two main approaches. One approach adjusts the number of power branches activated based on the current load. Patent CN104660019A proposes a multiphase parallel converter and its control method, which shuts off one phase when the load current is low to reduce switching losses and improve system efficiency. Smooth phase switching is achieved by superimposing a time-varying bias voltage signal onto the inductor current sampling signal. This approach relies on the already occurring load current, and the adjustment of the number of branches lags behind load changes. To avoid insufficient time to restart power branches during sudden load increases, a significant number of power branches must remain operational in practical applications. The other approach introduces prediction into the control. Patent CN111313687A proposes a power converter that collects operating point samples such as input and output current and voltage, uses machine learning to predict events such as load changes, and updates the regulator's control parameters accordingly. This approach advances the adjustment timing before load changes occur, but the predicted output is a single expected value, and the controller can only determine the energy-saving depth based on this single value.

[0004] Both approaches are constrained by the same physical constraint: a trade-off between fast response capability and conversion efficiency. When all power branches are operational and the loop bandwidth is increased, the converter can handle sudden load surges, at the cost of significantly increased losses during the stable load phase. When branches are shut down and the response configuration is reduced, efficiency increases, but the ability to handle sudden load surges decreases. Take a multi-phase power supply powering a processor as an example. The predictive controller judges that the load will be stable for a period of time, so it shuts down most power branches to reduce losses. Since the prediction is based on historical operating data, a certain computational task is not predicted, and the load current rises by hundreds of amperes within microseconds. At this time, the shut-down power branches need tens of microseconds to rebuild current, and during this period, the excess current can only be borne by the output capacitor discharge. This causes the output voltage to drop beyond the allowable range, resulting in a processor power supply failure. To avoid this consequence, practical engineering either abandons predictive energy saving or adds a sufficiently large fixed margin on top of the predicted value, thus losing most of the energy-saving benefits.

[0005] The above solutions collectively overlook the following technical issues: the efficiency cost of rapid response is borne by the entire power of the machine, while only a portion of the load power is truly uncertain. A consistently stable base power does not require rapid response capability, yet it is still placed in a low-efficiency, high-response operating state. Furthermore, a single numerical prediction cannot express the degree of certainty of the prediction; the controller neither knows where the power floor for reliable energy saving is, nor how wide the power range requiring rapid response capability is. For example, a server power supply's output power is consistently 300 watts for a long time, but occasional computing tasks may cause it to briefly surge to 600 watts. Of this, 300 watts almost certainly persists; the truly uncertain part is the potential additional 300 watts. With the current approach, the entire 600-watt power supply capacity must maintain a high-response state to accommodate the uncertainty of 300 watts, making the efficiency loss borne by the stable portion unnecessary.

[0006] Therefore, how to distinguish between the deterministic and indeterminate parts of the load power, so that the cost of rapid response is borne only by the indeterminate part, while the deterministic part operates at its highest efficiency, has become an urgent problem to be solved in the field of data center power supply. Summary of the Invention

[0007] The technical problem to be solved by this application is to provide an intelligent power energy efficiency scheduling method and system based on load prediction, which is used to solve the problem that the power converter is in a long-term low-efficiency high-response state in order to cope with the sudden increase in load, and the output voltage drops when the load prediction is wrong.

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

[0009] A smart power efficiency scheduling method based on load prediction is applied to a power converter that includes a main power branch and at least one auxiliary power branch. The method includes: collecting the output voltage sequence and output current sequence of the power converter, inputting the output voltage sequence and the output current sequence into a pre-trained machine learning load interval prediction model, and outputting the lower bound and upper bound of the load power in the future prediction period.

[0010] The lower bound is determined as the necessary power, and the main power branch is controlled to perform energy efficiency scheduling with the necessary power as the output target and locked at the preset optimal efficiency operating point.

[0011] The difference between the upper bound and the lower bound is determined as the power to be responded to. The standby configuration template library is queried using the power to be responded to to obtain the number of auxiliary power branches that need to be put into operation and the standby current value of each auxiliary power branch that is put into operation.

[0012] When the power to be responded to is greater than or equal to a preset standby exemption threshold, the auxiliary power branch that has been put into operation is controlled to maintain the standby current value.

[0013] When the power to be responded to is less than the preset standby exemption threshold, all auxiliary power branches are shut down, so that the power converter enters an energy-saving mode powered only by the main power branch.

[0014] When the auxiliary power branch maintains the standby current value, and the actual load power is detected to be greater than the necessary power, the portion of the actual load power exceeding the necessary power is borne by the auxiliary power branch maintaining the standby current value.

[0015] Optionally, the machine learning load interval prediction model is a quantile regression model, which includes a shared time-series feature extraction layer and two parallel quantile output layers. The time-series feature extraction layer is used to extract the time-domain variation features of the load power from the output voltage sequence and the output current sequence. The first quantile output layer outputs a load power value corresponding to a first quantile level as the lower bound, and the second quantile output layer outputs a load power value corresponding to a second quantile level as the upper bound. Both the first and second quantile levels are quantile values ​​between 0 and 1, and the second quantile level is greater than the first quantile level.

[0016] Optionally, the training process of the quantile regression model includes: collecting multiple output voltage sequences and output current sequences of the power converter during its historical operation; for each output voltage sequence and output current sequence, using the measured load power values ​​obtained within the future prediction period after the end of the collection of that segment as load power labels to form training samples; setting quantile loss for the first quantile output layer according to the first quantile level, setting quantile loss for the second quantile output layer according to the second quantile level, and performing supervised training on the parameters of the quantile regression model using the training samples.

[0017] Optionally, the standby configuration template library includes multiple standby configuration templates. Each standby configuration template includes a power value range to be responded to, the number of auxiliary power branches corresponding to the power value range to be responded to, and a standby current value corresponding to the power value range to be responded to. The power value ranges to be responded to in the multiple standby configuration templates do not overlap and continuously cover a preset power value range to be responded to. Querying the standby configuration template library includes: determining the power value range to be responded to that falls into, and outputting the number of auxiliary power branches and the standby current value in the standby configuration template to which the power value range to be responded to belongs. The number of auxiliary power branches in each standby configuration template is not less than the power value range to be responded to. The minimum integer value obtained by dividing the upper limit of the power value range by the rated power of a single auxiliary power branch; the standby current value in each standby configuration template is determined by a standby current calibration process, which is as follows: activating auxiliary power branches according to the number of auxiliary power branches in the standby configuration template, applying a load step to the power converter with an amplitude equal to the upper limit of the power value range to be responded to, and successively increasing the standby current value of each activated auxiliary power branch until the output voltage of the power converter drops after the load step does not exceed the preset drop amplitude upper limit, and writing the first standby current value that makes the drop amplitude not exceed the preset drop amplitude upper limit during the successive increase process into the standby configuration template.

[0018] Optionally, controlling the main power branch to perform energy efficiency scheduling with the inevitable power as the output target and locking it at the preset optimal efficiency operating point includes: keeping the inductor current of the main power branch in a continuous conduction mode, keeping the synchronous rectifier tube of the main power branch in synchronous rectification operation, and fixing the switching frequency of the main power branch to the switching frequency value corresponding to the preset optimal efficiency operating point; the preset optimal efficiency operating point is determined by an efficiency calibration curve, which is a curve obtained by calibration under multiple output power values, showing the change of the conversion efficiency of the power converter with the switching frequency, and the preset optimal efficiency operating point is the point on the efficiency calibration curve corresponding to the output power value closest to the inevitable power where the conversion efficiency reaches its maximum value.

[0019] Optionally, the acquisition of the output voltage sequence and output current sequence of the power converter further includes: sampling the output voltage of the power converter at the midpoint phase in each switching cycle of the main power branch to obtain a midpoint sampling value; and sampling the output voltage of the power converter at a preset sampling phase before the synchronous rectifier tube of the main power branch is turned off to obtain a valley sampling value; subtracting the valley sampling value and the midpoint sampling value in the same switching cycle, arranging the differences of each switching cycle in chronological order to form a divergence sequence, and inputting the divergence sequence, the output voltage sequence, and the output current sequence into the machine learning load range prediction model.

[0020] Optionally, the preset sampling phase is determined by a sampling phase calibration process, which is as follows: when the power converter is running under no-load, the output voltage is sampled by multiple candidate phases before the synchronous rectifier is turned off, and the candidate phase with the smallest sample value is determined as the preset sampling phase; before inputting the deviation sequence into the machine learning load interval prediction model, the deviation sequence is further processed as follows: the mean of the deviation sequence and the maximum value of the difference between two adjacent terms of the deviation sequence are calculated within a sliding statistical window of a preset length, and the mean and the maximum value are used as input features of the deviation sequence and input into the machine learning load interval prediction model.

[0021] Optionally, it also includes a bypass step: subtracting the voltage drop from the valley baseline voltage to obtain the valley voltage lower limit, and setting the valley voltage lower limit as the threshold of the bypass trigger comparator; the valley baseline voltage is the average value of the valley sampled value within a preset baseline statistical window, and the voltage drop is the quotient obtained by dividing the product of the necessary power and the switching cycle of the main power branch by the product of the output capacitor value of the power converter and the output voltage setting value of the power converter; when the valley sampled value is lower than the threshold of the bypass trigger comparator, the output of the bypass trigger comparator directly bypasses the feedback control path of the machine learning load range prediction model and the power converter, and enables all auxiliary power branches with a preset bypass drive signal in the current switching cycle.

[0022] Optionally, the machine learning load range prediction model outputs the quantile prediction value of the load power at the first quantile level as the lower bound; the intelligent power energy efficiency scheduling method based on load prediction further includes a quantile self-tuning step: each trigger of the out-of-bounds trigger comparator is recorded as a missed event, and the occurrence frequency of the missed event is counted within a preset missed event statistics window; when the occurrence frequency is greater than the preset upper limit, the first quantile level is lowered by a preset quantile step size; when the occurrence frequency is less than the preset lower limit, the first quantile level is increased by the preset quantile step size; the adjusted first quantile level is written back to the machine learning load range prediction model, so that the machine learning load range prediction model outputs the lower bound at the adjusted first quantile level.

[0023] This application also discloses an intelligent power energy efficiency scheduling system based on load prediction, used to implement the intelligent power energy efficiency scheduling method based on load prediction described in this application, including a power converter and a digital controller. The power converter includes a main power branch and at least one auxiliary power branch. The digital controller includes a sampling module, an interval prediction module, an energy efficiency scheduling module, a standby configuration module, and a switching module.

[0024] The sampling module is used to collect the output voltage sequence and output current sequence of the power converter;

[0025] The interval prediction module has a built-in pre-trained machine learning load interval prediction model, which is used to input the output voltage sequence and the output current sequence into the machine learning load interval prediction model, output the lower bound and upper bound of the load power in the future prediction period, determine the lower bound as the necessary power, and determine the difference between the upper bound and the lower bound as the power to be responded to.

[0026] The energy efficiency scheduling module is used to control the main power branch to perform energy efficiency scheduling with the necessary power as the output target and lock it at the preset optimal efficiency operating point.

[0027] The standby configuration module has a built-in standby configuration template library, which is used to query the standby configuration template library with the power to be responded to, to obtain the number of auxiliary power branches to be put into operation and the standby current value of each auxiliary power branch to be put into operation. When the power to be responded to is greater than or equal to a preset standby exemption threshold, the auxiliary power branches to be put into operation are controlled to maintain the standby current value. When the power to be responded to is less than the preset standby exemption threshold, all auxiliary power branches are turned off, so that the power converter enters an energy-saving mode in which only the main power branch is powered.

[0028] The load switching module is used to, when the actual load power exceeds the necessary power while the auxiliary power branch maintains the standby current value, transfer the portion of the actual load power exceeding the necessary power to the auxiliary power branch that maintains the standby current value.

[0029] The beneficial effects of this application are as follows:

[0030] This application ensures that the efficiency cost of a power converter for fast response no longer increases with the overall power output, but is only related to the uncertainty of the load. This is achieved by dividing the load power into two parts based on its degree of certainty. Specifically, the machine learning load range prediction model outputs a lower and upper bound for the load power. The lower bound, determined as the inevitable power, is locked to the preset optimal operating point by the main power branch for energy efficiency scheduling. The difference between the upper and lower bounds is determined as the power to be responded to, and the standby configuration template library is consulted to determine the number of auxiliary power branches and the standby current value. This division is based on an objective characteristic of load power: the continuous stable portion almost inevitably repeats within adjacent time periods, while the occurrence and magnitude of sudden bursts are difficult to determine. Separate processing balances efficiency and response, relying on the asymmetry of power branch responses. Branches with current can increase their output in just a few switching cycles, while branches that have been turned off require tens of microseconds to rebuild current. Therefore, the power supply section does not need to reserve response capacity for sudden changes, and the entire conversion process operates at its highest efficiency. The standby loss only applies to the portion corresponding to uncertainty. When the load is stable, the standby power decreases, and the standby loss decreases accordingly. When the standby power is less than the preset standby exemption threshold, all auxiliary power branches are shut down, and the power converter enters energy-saving mode. The cost of rapid response approaches zero. Taking a server power supply that is consistently stable at 300 watts and occasionally surges to 600 watts as an example, the 300-watt base section always converts at maximum efficiency, and only a few auxiliary power branches covering the uncertain portion of the other 300 watts are on standby. Predictions show that even this part of the standby is exempted during periods of stable load.

[0031] This application achieves deep energy savings while maintaining the ability to handle sudden load surges. This capability stems from the fact that the surge is not handled by the already switched-off branches, but rather by the auxiliary power branches that have maintained their standby current values. When the actual load power exceeds the necessary power, the excess is handled by the auxiliary power branches maintaining their standby current values. This ability to handle surges is based on the physical laws of inductor current. Branches already carrying current in the inductor can continue to increase their output current within several switching cycles simply by increasing the duty cycle, while branches with zero current must first undergo the process of being driven and having their current built up from zero. Therefore, the output voltage does not need to wait for the already switched-off branches to slowly rebuild their current, and energy saving and rapid load handling are no longer mutually exclusive.

[0032] This application also enables load prediction to detect signs of load increase earlier without adding any hardware. The implementation involves collecting midpoint and trough sample values ​​within the same switching cycle, subtracting them to form a divergence sequence, and inputting this sequence along with the output voltage and current sequences into a machine learning load range prediction model. The maximum difference between the mean and adjacent terms within a sliding statistical window is used as the input feature of the divergence sequence. This allows for earlier prediction because of an objective fact at the circuit level: the decoupling capacitor near the load chip discharges first when the load begins to increase, causing the trough of the output voltage to change before the cycle average. In existing technologies, the two sample values ​​are used separately, and this leading information is discarded. Subtracting reveals this leading information. The model can thus capture early signs of load increase before the average output voltage changes significantly, the prediction range is closer to the actual load, and standby losses are further reduced. The sampling capabilities used are already available in digital power supplies, requiring no additional components.

[0033] To address prediction errors, this application establishes a hardware-level protection path and uses the recording of protection actions to correct predictions. The protection is implemented by subtracting the voltage drop calculated based on the necessary power from the valley baseline voltage to obtain the valley voltage lower limit, which is set as the threshold for the over-limit trigger comparator. Once the valley sample value falls below the threshold, the over-limit trigger comparator directly bypasses the machine learning load range prediction model and feedback control path, enabling all auxiliary power branches with a preset direct drive signal within the current switching cycle. The reason it can maintain the output voltage when prediction and feedback are insufficient is that the analog comparator's response speed to voltage over-limit is much faster than that of the digital control loop; this is an inherent property of the circuit. The prediction correction is implemented by recording each trigger as a missed event and automatically adjusting the first quantile level based on the frequency of occurrence within the missed event statistics window. The correction effect stems from the fact that each over-limit is a true record of the narrowing prediction range; the frequency thus constitutes a correction basis from actual operation. When the frequency is too high, the first quantile level is automatically lowered, the necessary power is reduced, the standby range is widened, and the aggressiveness of the prediction converges to a conservative level. Therefore, the predictor is responsible for energy saving under normal conditions, while the over-limit trigger comparator is responsible for protection when the predictor fails. The two share the valley sampling value as the basis for judgment, and the consequences of the predictor failure are limited to a short-term voltage fluctuation and do not develop into a power supply failure.

[0034] The above design works together to ensure that the depth of energy saving is determined by the degree of predictability, the responsiveness is guaranteed by the standby configuration, and the consequences of failure are limited by hardware protection. It is suitable for powering data centers and artificial intelligence servers with drastic load fluctuations, as well as other multi-branch switching power supplies. Attached Figure Description

[0035] Figure 1 A schematic diagram of the structure of an intelligent power energy efficiency scheduling system based on load prediction provided in an embodiment of this application;

[0036] Figure 2 A flowchart illustrating the intelligent power energy efficiency scheduling method based on load prediction provided in this application embodiment;

[0037] Figure 3 A schematic diagram showing the lower and upper bounds of the load power and the power to be responded to, provided for embodiments of this application;

[0038] Figure 4 This is a schematic diagram of the structure of the quantile regression model provided in the embodiments of this application;

[0039] Figure 5 A schematic diagram illustrating the training process of the quantile regression model provided in the embodiments of this application;

[0040] Figure 6 A schematic diagram illustrating the composition of the standby configuration template library provided in this application embodiment;

[0041] Figure 7 A flowchart illustrating the standby current calibration process provided in this application embodiment;

[0042] Figure 8 A schematic diagram of the efficiency calibration curve and the preset optimal efficiency operating point provided in the embodiments of this application;

[0043] Figure 9 A schematic diagram illustrating the divergence sequence formed by subtracting the midpoint sample value and the valley sample value from the embodiments of this application;

[0044] Figure 10 A schematic diagram of the sampling phase calibration process provided in the embodiments of this application;

[0045] Figure 11 A timing diagram illustrating the boundary bypass step provided in an embodiment of this application;

[0046] Figure 12 This is a schematic diagram of the signal path for the out-of-bounds pass-through step provided in an embodiment of this application;

[0047] Figure 13 A flowchart illustrating the quantile self-tuning steps provided in an embodiment of this application;

[0048] Figure 14 A block diagram of a smart power energy efficiency scheduling system based on load prediction provided in an embodiment of this application. Detailed Implementation

[0049] 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.

[0050] 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.

[0051] Figure 1This is a schematic diagram of the intelligent power efficiency scheduling system based on load prediction provided in an embodiment of this application. The system includes a power converter 10 and a digital controller 40. The power converter 10 is used to convert the input voltage into the supply voltage required by a load 60 such as a processor. To handle the large current in data center and artificial intelligence server scenarios, the power stage of the power converter 10 consists of multiple parallel power branches that can be independently turned on or off, one of which serves as the main power branch 20, and at least one of the others serves as an auxiliary power branch 30. Figure 1 The auxiliary power branch 30 is indicated by two boxes with ellipses, suggesting that there can be one or more auxiliary power branches 30; this application does not limit the number. After the outputs of each power branch are combined, an output capacitor 50 is connected in parallel and then connected to the load 60. The output capacitor 50 provides or absorbs charge when the load current changes instantaneously, thus smoothing the output voltage.

[0052] The digital controller 40 collects the output voltage and current of the power converter 10, and determines the on / off state and operating point of each power branch accordingly. It then sends control signals to the main power branch 20 and the auxiliary power branch 30. This division is based on separating the branches that handle basic power from those that handle burst power, allowing for different control objectives to be applied to them. This is the foundation of the entire scheme in the energy efficiency scheduling discussed later. The digital controller 40 can be implemented using a digital signal processor, a field-programmable gate array (FPGA), or a dedicated digital power control chip. The sampling and computation capabilities of these chips are inherent, eliminating the need for additional components in this scheme.

[0053] This section first explains the power characteristics of load type 60, which forms the physical basis for all subsequent scheduling decisions. The power variation rate of this type of load is much higher than that of general electronic devices. When a computing task is issued, the load current can rise from tens of amperes to hundreds of amperes within microseconds. Simultaneously, its power is not chaotic, but rather consists of a long-term, continuous base component and an occasional surge. Taking a power supply for a processor as an example, its output power may remain stable around 300 watts for a long time, while an occasional computing task might cause it to briefly spike to 600 watts. Of this, approximately 300 watts almost certainly persist; the only uncertainty lies in the potential for another approximately 300 watts. The base component almost inevitably repeats within adjacent time periods, while the occurrence and magnitude of the surge are difficult to predict in advance. It is precisely this difference in certainty that allows this application to treat the power in two distinct ways.

[0054] Reference Figure 2The intelligent power efficiency scheduling method based on load prediction provided in this application includes steps S101 to S105. The overall idea of ​​the method is to first obtain a range of future load power rather than a single value from the prediction; use the lower edge of the range to indicate the power that is certain to exist, and use the width of the range to indicate the power that is uncertain; then apply efficiency-first and response-first control to these two parts respectively; finally, use a hardware path that is faster than prediction and feedback to cover prediction errors. An overview of each step is given below, followed by a detailed explanation.

[0055] In step S101, the sampling module 41 acquires the output voltage sequence and output current sequence of the power converter 10, inputs them into the pre-trained machine learning load range prediction model 47, and outputs the lower and upper bounds of the load power within the future prediction period. In step S102, the lower bound is determined as the necessary power, and the main power branch 20 is controlled to perform energy efficiency scheduling with the necessary power as the output target and locked at the preset optimal efficiency operating point.

[0056] In step S103, the difference between the upper and lower bounds is determined as the power to be responded to. The standby configuration template library 46 is then queried using this power to obtain the number of auxiliary power branches 30 to be activated and the standby current value of each activated auxiliary power branch 30. Subsequently, it is determined whether the power to be responded to is greater than or equal to a preset standby exemption threshold. If not, all auxiliary power branches 30 are shut down, causing the power converter 10 to enter an energy-saving mode powered only by the main power branch 20. If yes, the process proceeds to step S104, controlling the activated auxiliary power branches 30 to maintain their standby current values. In step S105, while the activated auxiliary power branches 30 maintain their standby current values, if the actual load power is detected to be greater than a certain power, the portion of the actual load power exceeding the certain power is handled by the auxiliary power branches 30 maintaining their standby current values.

[0057] The following steps will be explained in detail, one by one, following the steps outlined above. For ease of understanding, we will first combine... Figure 3 This illustrates how the load power is divided into two parts. Figure 3 This diagram illustrates the lower and upper bounds of the load power and the power to be responded to, as provided in this embodiment. The horizontal axis represents time, and the vertical axis represents the load power. The two dashed lines in the diagram represent the lower and upper bounds output by the machine learning load interval prediction model 47, respectively, while the solid line in the middle represents the actual load power. It can be seen that the actual load power always falls between the lower and upper bounds throughout the entire time axis. The lower bound is the level that the load power will almost certainly reach, and the upper bound is the upper limit that the load power may reach. The distance between the two is the power to be responded to. This application uses the lower bound to lock in the definite part and uses this distance to cover the uncertain part.

[0058] Let's first look at the acquisition and prediction in step S101. The sampling module 41 reads the output voltage and output current at a fixed sampling interval on the output side of the power converter 10, arranging the values ​​from several consecutive sampling intervals into output voltage and output current sequences. The sampling interval can range from 50 nanoseconds to 2 microseconds, typically 200 nanoseconds; the length of each sequence can range from 16 to 256 points, typically 64 points. A sequence that is too short will make it difficult to show the trend, while a sequence that is too long will increase the input dimension and inference delay. A compromise is made to cover several to a dozen switching cycles. The reason for using sequences instead of single instantaneous values ​​is that the trend of load power is inherent in its time variation. Only segmented waveforms can allow the model to see whether the power is rising, stabilizing, or beginning to fall. Both sequences are from the voltage and current loops already present on the output side of the power converter 10. These are quantities that digital power supplies inherently possess for achieving closed-loop control, requiring no additional sensors. The signal source is clear and does not increase hardware costs.

[0059] The machine learning load range prediction model 47 receives the two sequences mentioned above and outputs not a predicted power value, but rather the lower and upper bounds of the load power within the future prediction period. The future prediction period refers to a time period looking forward from the current moment, with a length that can range from 10 microseconds to 10 milliseconds, typically 100 microseconds. If the prediction period is too short, there won't be enough time to complete the standby preset of the auxiliary power branch 30 before the actual load surge, increasing the frequency of hardware protection intervention. If the prediction period is too long, the range must be widened to cover further uncertainties, thus reducing energy-saving benefits. Therefore, the prediction period is selected as a compromise between the typical setup time of a load surge and the standby preset time of the auxiliary power branch 30. The fact that the output is a range is the fundamental difference between this application and single-value prediction. A single value cannot express the degree of certainty in the prediction; the controller neither knows where the power floor for reliable energy saving is, nor how wide the power range requiring reserved response capability is. However, with the lower and upper bounds given together, the floor and the width are simultaneously determined, allowing the two subsequent control branches to each have their specific targets.

[0060] The machine learning load interval prediction model 47 is invoked once every prediction period, rather than running in every switching cycle. Therefore, its inference only needs to be completed within one prediction period. Loop structures with a large number of parameters can be deployed on field-programmable gate arrays or digital signal processors to meet this latency requirement, while one-dimensional convolutions with a smaller number of parameters can be completed smoothly on dedicated digital power control chips at this pace. Real-time actions for each switching cycle are handled by the switching and out-of-bounds pass-through mechanisms described later. These mechanisms only perform lightweight operations such as comparisons and table lookups, and are not constrained by model inference latency. Therefore, the heavy prediction and fast response operate on two separate timescales in terms of computing power without competing with each other.

[0061] Next, let's look at the handling of the inevitable power in step S102, which prioritizes efficiency. Determining the lower bound as the inevitable power means that the portion of power almost certainly used within the prediction interval is assigned to the main power branch 20 for long-term operation. The energy efficiency scheduling module 43 controls the main power branch 20 to use the inevitable power as its output target and locks its operating point at the preset optimal efficiency operating point. The reason why the main power branch 20 can operate at its optimal efficiency point for a long time without reserving margin for sudden increases is precisely because it only handles the inevitable power, which has extremely low uncertainty and almost never requires a significant temporary increase in output. By eliminating the cost of this part for rapid response, it can operate in a state of minimum loss throughout the entire process. If, according to the existing practice, the branch handling all the power is configured to maintain a high response for possible sudden increases, then the basic power portion is also forced into an inefficient state, which is precisely the waste that this application aims to avoid.

[0062] Next is step S103, which processes the power to be responded to. This is the response-first approach. The difference between the upper and lower bounds is determined as the power to be responded to, which measures the size of the uncertain portion of the load power. The standby configuration template library 46 is consulted using the power to be responded to to obtain the number of auxiliary power branches 30 that need to be activated and the standby current value of each activated auxiliary power branch 30. A lookup table is used here instead of online solving to ensure real-time performance. The correspondence between the number of auxiliary power branches 30 and the standby current value is calibrated and stored in the standby configuration template library 46 offline. During runtime, only one lookup table is needed to obtain the configuration result, avoiding iterative calculations in each control cycle. Therefore, the cycle time of the digital controller 40 is not affected. The standby current value refers to the inductor current that the activated auxiliary power branch 30 maintains in advance. The activated auxiliary power branch 30 maintains its switching action, keeps the inductor current, and keeps the drive powered on, thus being in a standby state where the output can be increased at any time. The standby current value is usually small. When the auxiliary power branch 30 is in standby mode, it only outputs the corresponding small power. The remaining necessary power is still borne by the main power branch 20. The sharing of the two is coordinated by the normal voltage feedback loop, and the output voltage is maintained at the set value.

[0063] Next, the power to be responded to is compared with the preset standby exemption threshold, which determines whether the auxiliary power branch 30 needs to be on standby. When the power to be responded to is greater than or equal to the preset standby exemption threshold, it means that the uncertainty is large enough to warrant the standby loss for prevention. At this time, the auxiliary power branch 30 is controlled to maintain the standby current value according to the lookup table result. When the power to be responded to is less than the preset standby exemption threshold, it means that the uncertainty is small enough that even if there is a sudden surge, it can be easily absorbed by the main power branch 20 and the output capacitor 50. At this time, all auxiliary power branches 30 are turned off, so that the power converter 10 enters the energy-saving mode where it is powered only by the main power branch 20.

[0064] The preset standby exemption threshold can be set to 10% to 50% of the rated power of a single auxiliary power branch 30, typically 20%. If this threshold is too high, the auxiliary power branch 30 will be prematurely shut down under light burst conditions where it should be on standby, weakening its load-bearing capacity. If the threshold is too low, the auxiliary power branch 30 will still waste standby current under deep light load conditions with almost no uncertainty, unnecessarily increasing losses. Therefore, calculating this threshold based on the power corresponding to the amount of charge that the output capacitor 50 can absorb in one control cycle is the most prudent approach. After entering energy-saving mode, rapid relay is no longer available. If an unforeseen surge occurs, only the output capacitor 50 can support the current control cycle until a direct overload occurs, re-activating the shut-down auxiliary power branch 30. Therefore, the preset standby exemption threshold must be small enough that the capacitor discharge within this control cycle is sufficient to withstand the surge without causing the output voltage to exceed the limit. This step allows standby losses to decrease as the load stabilizes and to return to zero when the load is sufficiently stable, making the cost of rapid response approach zero.

[0065] This section supplements the voltage regulation under normal conditions, which is handled by the feedback control path mentioned in the subsequent over-limit bypass step. Main power branch 20 is locked at the preset optimal efficiency operating point, meaning its switching frequency is fixed at the optimal efficiency value and it maintains continuous conduction mode and synchronous rectification. Its duty cycle is still adjusted in real-time by the feedback control path according to the deviation between the output voltage and the set value. This allows the actual output power of main power branch 20 to be finely adjusted near the necessary power level to follow the base load, stabilizing the output voltage at the set value. The reason for fixing the frequency while still adjusting the duty cycle is that conversion efficiency is mainly determined by the switching frequency and operating mode. Small adjustments to the duty cycle hardly change the efficiency but are necessary to maintain voltage regulation. Therefore, when the actual load power is slightly lower than the necessary power, the feedback control path correspondingly reduces the output of main power branch 20 and standby auxiliary power branch 30, preventing the output voltage from rising due to oversupply. When the actual load power is higher than the necessary power, the standby auxiliary power branch 30 first increases its output to handle the load, and the feedback control path simultaneously eliminates the deviation. The power splitting of the feedforward determines the role and standby capacity of each branch, while the feedback voltage loop is responsible for stabilizing the output voltage at the set value. The division of labor between the two does not conflict.

[0066] Finally, step S105 involves the load switching, ensuring that the system can still handle sudden load increases after deep energy saving. While the auxiliary power branch 30 maintains its standby current value, if the actual load power is detected to exceed the required power, the portion exceeding the required power is handled by the auxiliary power branch 30 maintaining the standby current value. The actual load power is obtained by multiplying the output voltage and output current obtained in real time by the sampling module 41, and this comparison with the required power is performed cycle-by-cycle within the digital controller 40. The reason for using the auxiliary power branch 30 maintaining the standby current value instead of temporarily waking up the switched-off branch stems from the physical fact of asymmetrical power branch responses. A branch with existing current in the inductor only needs to increase its duty cycle to continue increasing the output current within several switching cycles; while a branch with zero current must first undergo the process of powering on the drive circuit and establishing the inductor current from zero, which takes tens of microseconds.

[0067] If the shut-off branch is only woken up when the load has already surged, the excess current during those tens of microseconds can only be supported by the discharge of the output capacitor 50, and the output voltage is very likely to drop below the allowable range. This application allows the relay to occur between the auxiliary power branches 30 that already have current, so the output voltage does not need to wait for any branch to slowly rebuild the current, and energy saving and rapid relay are no longer mutually exclusive.

[0068] The following quantitative explanation explains the basis for the takeover being completed within several switching cycles. The auxiliary power branch 30, which maintains the standby current value, already has an inductor with a current close to the standby current value before a sudden load increase. During takeover, the inductor current is simply increased from the standby current value to the current required to handle the excess. The increase in inductor current within each switching cycle is approximately equal to the difference between the input and output voltages multiplied by the conduction time and divided by the inductance. Therefore, the smaller the inductance, the greater the current increase per cycle. However, excessively small inductance will increase current ripple and turn-off losses. Thus, the inductance is chosen as a trade-off between ripple and efficiency, typically allowing a single auxiliary power branch 30 to increase its current by 5% to 20% of its rated current within one switching cycle. Accordingly, the portion of the actual load power that exceeds the required power is first divided by the output voltage setting value to obtain the additional output current required, and then divided by the total current that can be added per cycle in the auxiliary power branch 30, which is the number of switching cycles required to complete the takeover. Under normal circumstances where the power to be responded to is correctly preset, this number of cycles usually falls between 3 and 15 switching cycles, which is much smaller than the tens of microseconds required for the reconstruction current of the already turned-off branch.

[0069] It is necessary to supplement the division of labor under the normal condition where the actual load power exceeds the necessary power but has not exceeded the upper limit. At this time, the part exceeding the necessary power falls within the power range to be responded to. The auxiliary power branch 30, which is put into operation and maintains the standby current value, is configured to cover this range. Its quantity and standby current value have been pre-prepared by the standby configuration template library 46 according to the size of the power to be responded to, so it can immediately handle this part of the power. Only when the magnitude or speed of a sudden increase exceeds the range that the auxiliary power branch 30 can immediately handle, causing a deep drop in output voltage, is the hardware protection path to be intervened, as described later. This part is left to be carried out in the over-limit pass-through step.

[0070] The following describes the specific structure of the machine learning load interval prediction model 47, which determines how the lower and upper bounds are given simultaneously. Figure 4 This is a schematic diagram of the quantile regression model provided in an embodiment of this application. In one embodiment, the machine learning load interval prediction model 47 is a quantile regression model, including a shared time-series feature extraction layer and two parallel quantile output layers. The time-series feature extraction layer is used to extract the time-domain variation features of the load power from the output voltage sequence and the output current sequence. Here, the time-domain variation features refer to those quantities that can reflect whether the power is rising, stabilizing, or falling, such as the recent mean, slope, and fluctuation amplitude of the sequence. The time-series feature extraction layer can be implemented using a one-dimensional convolutional network, or it can be implemented using a recurrent structure such as a gated recurrent unit or a long short-term memory network. The former has low computational cost and low latency, while the latter is better at capturing dependencies over longer time spans. The appropriate layer can be selected based on the computing power of the digital controller 40.

[0071] The two quantile output layers share the same time-domain variation characteristics and each outputs a load power value. The first quantile output layer outputs the load power value corresponding to the first quantile level as the lower bound, and the second quantile output layer outputs the load power value corresponding to the second quantile level as the upper bound. The quantile level is a number between 0 and 1, representing the degree of confidence that the actual load power will not exceed the output value. A larger value indicates higher confidence and a higher corresponding power value. The first quantile level is set relatively low, making the lower bound a level where the load power is highly likely to be reached and can therefore be considered certain. The second quantile level is set relatively high, making the upper bound an upper limit that the load power is unlikely to exceed. For the interval to hold, the second quantile level must be greater than the first quantile level. The outputs of the two output layers also need to be clamped sequentially to ensure that the upper bound is not less than the lower bound, avoiding quantile overlap that could lead to an inverted interval.

[0072] The first quantile level can be set between 0.05 and 0.2, typically 0.1. It is not fixed during implementation, and its self-tuning range is given in the quantile self-tuning steps below. If the first quantile level is set too low, the lower bound will be conservative, inevitably resulting in underestimating power, with more power being allocated to standby power, increasing standby losses. If it is set too high, the lower bound will approach the actual load, increasing the probability that the main power branch 20 cannot handle the load, and increasing the triggering frequency of the hardware protection path. Therefore, the first quantile level is determined by a trade-off between acceptable false negative rate and standby loss budget. The second quantile level can be set between 0.9 and 0.98, typically 0.95. If the second quantile level is set too low, the upper bound will be too narrow, and the standby configuration will be insufficient to cover the actual surge, reducing the handling capacity. If it is set too high, the upper bound will be artificially high, resulting in an excessive number of standby auxiliary power branches 30, increasing standby losses. Therefore, it is selected based on the high quantile of the statistical distribution of the surge magnitude.

[0073] Further, we will explore the training process of the machine learning load interval prediction model 47, which answers the question of how the output layer of these two quantiles is learned. Figure 5 This is a schematic diagram illustrating the training process of the quantile regression model provided in this application embodiment. Training samples are accumulated from the historical operation of the power converter 10, collecting multiple output voltage and current sequences from its historical operation as input features. For each sequence, the measured load power values ​​within the future prediction period after the end of that sampling period are used as load power labels. The input features and load power labels together constitute the training samples. The reason for defining the labels in this way is that the model aims to predict the load trend after the sampling window ends. Using the actual power at that time as a monitoring signal, the model learns the mapping from historical waveforms to future power.

[0074] Training employs supervised learning. The first quantile output layer is given a quantile loss based on the first quantile level, and the second quantile output layer is given a quantile loss based on the second quantile level. Supervised training of the quantile regression model's parameters is performed using training samples. The quantile loss is the standard loss for quantile regression, characterized by an asymmetric penalty for underestimating or overestimating predictions. The bias of the penalty is determined by the quantile level. This asymmetry ensures that each output layer converges to its desired quantile, one learning a high-probability lower bound and the other a low-probability upper bound. Training can be performed offline using conventional optimization algorithms such as stochastic gradient descent. The convergence criterion is that the quantile loss no longer decreases on the validation set. The total number of parameters in the model ranges from tens of thousands to millions, determined by the size of the temporal feature extraction layer within the storage and computing power constraints of the digital controller 40. One-dimensional convolutional implementations typically have tens of thousands to hundreds of thousands of parameters, while recurrent implementations can have millions of parameters.

[0075] Regarding cold start and updates, in the initial installation phase when samples are insufficient, a model can be pre-trained using historical data from similar power supplies or typical load curves from within the factory. During operation, local data is continuously collected, and after accumulating to a certain scale, offline retraining is performed. Thereafter, offline incremental updates are performed weekly or monthly. If the prediction remains consistently narrow, a quantile self-tuning step (described later) allows for online fine-tuning of the first quantile level as a rapid correction in addition to offline updates. Thus, the machine learning load range prediction model 47 has clear implementation methods for each stage, from sample definition, features and labels, to loss setting and training algorithms, and then to parameter magnitude and cold start updates. Those skilled in the art can reproduce it accordingly without needing to guess how the model operates internally.

[0076] The following describes the standby configuration template library 46, which solidifies the mapping of the power to be responded to to the auxiliary power branch 30 configuration into a lookable table. Figure 6 This is a schematic diagram illustrating the composition of the standby configuration template library provided in this application embodiment. The standby configuration template library 46 includes multiple standby configuration templates. Each standby configuration template includes a power value range to be responded to, the number of auxiliary power branches 30 corresponding to that range, and a standby current value corresponding to that range. The power value ranges to be responded to in the multiple standby configuration templates do not overlap and continuously cover a preset range of power values ​​to be responded to, such as... Figure 6 As shown on the horizontal axis, the entire value range is divided into several segments that are continuous and non-overlapping. Any power to be responded to will fall into exactly one of these segments. The requirement for non-overlapping and continuous coverage ensures that the lookup result is unique and without gaps. Any power to be responded to will find a definite configuration, preventing situations where two templates match simultaneously or no template matches. During the query, the range of power values ​​to be responded to is determined, and the number of auxiliary power branches 30 and the standby current value in the standby configuration template corresponding to that range are output.

[0077] The number of auxiliary power branches 30 in the standby configuration template is determined according to a clear rule: it is the smallest integer obtained by dividing the upper limit of the range of power to be responded to by the rated power of a single auxiliary power branch 30. This selection is based on the consideration that the number of auxiliary power branches 30 deployed must be sufficient to cover the maximum power to be responded to in this range. Rounding up ensures that even if the power to be responded to reaches the upper limit of the range, there are enough branches to handle it, while avoiding excessive deployment and wasted power. The rated power of a single auxiliary power branch 30 is determined according to the power level design, and can range from 50 watts to 200 watts, typically 100 watts, selected based on device current stress and heat dissipation conditions. The rated power of the main power branch 20 is set according to the upper limit of the necessary power it needs to bear over a long period, usually several times the rated power of a single auxiliary power branch 30. The main power branch 20 itself can also be composed of multiple phases connected in parallel to bear a larger necessary power. The density of the range of power values ​​to be responded to can be selected according to the required configuration precision. If the range is too coarse, it will leave a large margin between the actual standby capacity and the needs, resulting in the loss of some energy saving. If the range is too fine, the number of templates will increase and the calibration workload will increase. In practice, it is more natural to use an integer multiple of the rated power of 30 for a single auxiliary power branch as the range boundary.

[0078] The standby current value in each standby configuration template is determined by the standby current calibration process. Figure 7 The flowchart illustrating the standby current calibration process provided in this application embodiment includes steps S401 to S404. In step S401, a corresponding number of auxiliary power branches 30 are activated according to the number of auxiliary power branches 30 in the standby configuration template. In step S402, a load step with an amplitude equal to the upper limit of the power value range to be responded to is applied to the power converter 10 to reproduce the most severe surge that the template must handle. In step S403, the standby current value of each activated auxiliary power branch 30 is increased sequentially. Then, it is determined whether the voltage drop of the power converter 10 after the load step does not exceed the preset upper limit of the drop amplitude. If the determination is no, the process returns to step S403 to continue increasing the standby current value; if the determination is yes, the process proceeds to step S404, where the first standby current value that causes the drop amplitude to not exceed the preset upper limit of the drop amplitude is written into the standby configuration template.

[0079] This calibration yields a standby current value that is just sufficient, ensuring that the output voltage drop remains within acceptable limits under the most severe surge, without unnecessarily increasing standby losses due to an excessively large standby current value. The preset upper limit of the drop amplitude can be set to 1% to 5% of the output voltage setting, typically 3%. If it is too large, the processor may malfunction due to insufficient supply voltage; if it is too small, the standby current value will be forced to be very high, increasing standby losses. Therefore, it is selected with a certain margin based on the supply voltage tolerance of the load chip. When a load step is applied, the main power branch 20 maintains its rated baseline operating point, ensuring that the calibration conditions are consistent with actual operation.

[0080] The following section details the implementation of energy efficiency scheduling in step S102, explaining how the main power branch 20 is locked at the preset optimal efficiency operating point. Controlling the main power branch 20 to achieve a certain output power and locking it at the preset optimal efficiency operating point involves energy efficiency scheduling. Specifically, this includes maintaining the inductor current of the main power branch 20 in continuous conduction mode, maintaining the synchronous rectifier tube of the main power branch 20 in synchronous rectification operation, and fixing the switching frequency of the main power branch 20 to the value corresponding to the preset optimal efficiency operating point. This switching frequency is typically in the hundreds of kilohertz range, such as 500 kilohertz. The statistical windows measured in switching cycles mentioned earlier are converted using this as the time base. Continuous conduction mode and synchronous rectification are mature operating methods with low losses in switching power supplies under medium to high loads. They are explicitly fixed here because they correspond to the operating conditions in the neighborhood of the optimal efficiency point. Fixing the switching frequency at the optimal value avoids the efficiency fluctuations caused by frequency reduction and cycle skipping common under light loads, ensuring that the main power branch 20, which bears the certain power, is stably at the point of lowest loss.

[0081] The optimal operating point for preset efficiency is determined by the efficiency calibration curve. Figure 8 This diagram illustrates the efficiency calibration curves and the preset optimal operating point provided in this embodiment. The horizontal axis represents the switching frequency, and the vertical axis represents the conversion efficiency. The diagram shows multiple efficiency calibration curves, corresponding to the relationship between the conversion efficiency of the power converter 10 and the switching frequency, calibrated under different output power values. Under a given output power, the conversion efficiency first increases and then decreases with the switching frequency. There exists a frequency at which the conversion efficiency reaches its maximum value; this is determined by the inverse relationship between switching losses and conduction losses with frequency. The preset optimal operating point is taken as the point on the efficiency calibration curve corresponding to the output power value closest to the required power value, where the conversion efficiency reaches its maximum value. Figure 8 The point marked with a solid dot and extended vertically to the horizontal axis at the switching frequency value. The reason for selecting the curve based on the output power value closest to the required power is that the main power branch 20 actually bears the required power, and only the frequency determined by the curve that matches it is its current optimal efficiency.

[0082] The "preset" here refers to the efficiency calibration curve and the optimal point of each curve being pre-determined and stored in a table during the production line calibration stage. During operation, the nearest value is used based on the current necessary power, without the need for online searching. The grid composed of multiple output power values ​​can be divided at equal intervals, ranging from 10% to 100% of the rated power. If the grid is too sparse, the curve obtained from the nearest value will deviate significantly from the actual necessary power, and the determined frequency will deviate slightly from the true optimum. If the grid is too dense, the calibration workload will increase. In practice, a dozen or so power points are sufficient to balance accuracy and workload.

[0083] The following section discusses an enhancement to the acquisition process: introducing a divergence sequence in addition to the output voltage and current sequences, enabling load prediction to detect load increases earlier without adding any hardware. Figure 9 This diagram illustrates the divergence sequence formed by subtracting the midpoint and valley sampling values ​​from the midpoint sampling values ​​provided in this embodiment. The upper part shows the output voltage waveform and two sampling points within one switching cycle, while the lower part shows the divergence sequence arranged in chronological order. The output voltage of the power converter 10 is sampled at the midpoint phase of each switching cycle of the main power branch 20 to obtain the midpoint sampling value; and the output voltage is sampled at a preset sampling phase before the synchronous rectifier diode of the main power branch 20 is turned off to obtain the valley sampling value.

[0084] The midpoint phase refers to the phase in which the output voltage is close to its periodic average value within a switching cycle. In processor power supplies with a small duty cycle, it usually falls near the midpoint of the switching cycle. If necessary, it can also be calibrated for each unit according to the phase that makes the sampled value closest to the periodic average value, similar to the valley phase. The preset sampling phase falls before the synchronous rectifier tube turns off, where the output voltage is close to its periodic trough. The difference between the valley sampled value and the midpoint sampled value within the same switching cycle is calculated, and the differences of each switching cycle are arranged in chronological order to form a divergence sequence. The divergence sequence, along with the output voltage sequence and the output current sequence, are then input into the machine learning load range prediction model 47.

[0085] The reason why the difference can make predictions earlier is due to an objective fact at the circuit level. The decoupling capacitor near the load chip discharges first when the load begins to increase. As a result, the trough of the output voltage changes before the periodic average value. The trough sample value moves first, while the midpoint sample value remains approximately unchanged. The difference between the two widens early in the process of load increase.

[0086] In existing technologies, midpoint and valley sampling values ​​are used separately, and the prior information hidden in their difference has been discarded. After subtraction, this prior information is revealed, allowing the model to detect signs of increased load before the average output voltage changes significantly. The typical magnitude of the divergence difference is a few thousandths of the output voltage setpoint, on the same order of magnitude as switching ripple, but with a discernible direction. To avoid it being drowned out by noise, sampling must be performed after the preset sampling phase calibration described below to improve signal-to-noise ratio. The analog-to-digital converter uses 12 bits or more and is supplemented with multi-cycle oversampling, so that the difference of a few thousandths is higher than the quantization noise and can be distinguished. The divergence sequence is only used as one input feature for model weighting, and is not used as a hard criterion to trigger action. Therefore, even if individual differences are disturbed, they are simply absorbed by the model and will not directly cause malfunctions. The midpoint and valley sampling used are capabilities inherent in digital power supplies for ripple monitoring, and no additional components are required.

[0087] The following sections cover two processing steps that accompany the divergence sequence: calibration of the preset sampling phase and extraction of the input features of the divergence sequence. Figure 10 This diagram illustrates the sampling phase calibration process provided in this application embodiment. The horizontal axis represents time, and the vertical axis represents output voltage. The diagram shows the waveform of the output voltage during one cycle under no-load operation and sampling points at several candidate phases. The preset sampling phase is determined by the sampling phase calibration process. When the power converter 10 is under no-load operation, the output voltage is sampled at multiple candidate phases before the synchronous rectifier is turned off. The candidate phase with the smallest sample value is determined as the preset sampling phase. Figure 10 The phase is marked with a solid dot and a vertical line drawn to the horizontal axis. The reason for calibrating under no-load and taking the minimum sample value is that the output voltage waveform is the cleanest under no-load conditions, and the valley position best reflects the valley phase of the circuit itself. Taking the candidate phase with the minimum sample value is to take the true valley, so that the valley sample value is stably taken on the valley rather than on its slope.

[0088] Candidate phases are arranged at equal intervals within a window before the synchronous rectifier tube is turned off. The preset sampling phase usually falls at 85% to 98% of the switching cycle, typically at 95%. If the preset sampling phase deviates too much from the trough, the trough sampling value will be raised, and the deviation difference will be weakened. Therefore, the minimum sampling value is used as the standard for calibration of each unit.

[0089] Before inputting the divergence sequence into the machine learning load interval prediction model 47, a statistical compression step is performed on the divergence sequence. Within a sliding statistical window of a preset length, the mean of the divergence sequence and the maximum difference between two adjacent terms are calculated. These two quantities are used as input features of the divergence sequence into the model. The mean reflects the overall level of recent divergence, and the maximum difference between two adjacent terms reflects the intensity of the divergence jump. The former corresponds to a gradual increase in load, and the latter corresponds to a sharp increase in load. Together, these two quantities represent both trends and abrupt changes. The sliding statistical window can be 8 to 64 switching cycles, typically 16 switching cycles. If the window is too short, the statistics will be significantly affected by single-cycle noise and prone to spikes; if the window is too long, recent changes will be diluted by historical values, and the leading nature will be weakened. Therefore, the number of switching cycles corresponding to the typical establishment time of a load surge is selected. Replacing the entire divergence sequence with two statistics in the model not only preserves the leading information carried by the divergence but also reduces the input dimensionality, thus alleviating the computational burden of the machine learning load interval prediction model 47.

[0090] The following describes the over-boundary bypass procedure, which is a hardware-level protection that prevents prediction and feedback from being timely, thus catching up with occasional prediction errors. Figure 11 This is a timing diagram illustrating the out-of-bounds pass-through step provided in this application embodiment. The horizontal axis represents time, and the vertical axis represents output voltage. A dashed line represents the valley baseline voltage, and a dotted line represents the valley voltage lower limit. The distance between the two lines represents the voltage drop. The sequence of several sampling points represents the valley sampling values ​​for each cycle. The valley voltage lower limit is obtained by subtracting the voltage drop from the valley baseline voltage, and this lower limit is set as the threshold of the out-of-bounds trigger comparator. The valley baseline voltage is the average value of the valley sampling values ​​within a preset baseline statistical window. It indicates the position where the valley should be under the current operating condition and moves slowly with the gradual change in load. The preset baseline statistical window can be 64 to 1024 switching cycles, typically 256 switching cycles. If the window is too short, the baseline will fluctuate with sudden changes, averaging out actual out-of-bounds events without reaching the threshold. If the window is too long, the baseline will not keep up with the normal drift of the operating condition and will not provide a reference for deviation after the load changes gradually. Therefore, a length much longer than the duration of a single sudden change and much shorter than the time scale of the operating condition change is chosen.

[0091] The voltage sag is given as an analytically calculable quantity, which is the product of the necessary power and the switching cycle of the main power branch 20, divided by the product of the capacitance of the output capacitor 50 of the power converter 10 and the output voltage setting. The physical meaning of this conversion is that if a power gap of approximately the necessary power order occurs within one switching cycle and is entirely filled by the discharge of the output capacitor 50, the depth to which the output voltage will sag is used as the drop from the baseline to the lower limit. This is equivalent to setting the threshold at a position that the capacitor's natural fluctuations within a normal cycle cannot reach, but only when a true out-of-bounds gap occurs. To ensure that the threshold is reliably outside the normal ripple band, a safety factor can be multiplied by this voltage sag. The safety factor can be 1.1 to 2, typically 1.5. Simultaneously, a minimum trigger depth is set, meaning the difference between the valley voltage lower limit and the valley baseline voltage is not less than a lower limit percentage of the output voltage setting. This percentage can be 0.5% to 2%, typically 1%. If the safety factor is too small or the minimum trigger depth is too shallow, the threshold will slip into the normal ripple band under light load, causing false triggering; if it is too large or too deep, the actual over-limit will fall deeper before being detected, and the protection intervention will be too late. Therefore, the threshold is selected with a margin based on the peak-to-peak value of the light load ripple.

[0092] When the valley sampling value is lower than the threshold of the over-limit trigger comparator, the output of the over-limit trigger comparator directly bypasses the feedback control path of the machine learning load range prediction model 47 and the power converter 10, and enables all auxiliary power branches 30 with a preset direct drive signal in the current switching cycle. Figure 12 This is a schematic diagram of the signal path for the out-of-bounds pass-through step provided in this application embodiment. Under normal operating conditions, the valley sampling value and the valley voltage lower limit are sent together to the out-of-bounds trigger comparator, and the remaining control is issued through the machine learning load range prediction model 47 and the feedback control path; once the valley sampling value exceeds the limit, the output of the out-of-bounds trigger comparator is... Figure 12 The direct path in the circuit reaches the auxiliary power branch 30 directly, bypassing the first two slower paths.

[0093] The reason why an analog comparator can bypass prediction and feedback and act directly is that the analog comparator's response to voltage over-limit is on the order of nanoseconds to microseconds, which is much faster than the digital control loop that requires sampling, calculation, and output. This is an inherent property of the circuit. When the prediction fails to anticipate the problem and the feedback is still integrating, only this hardware path has enough time to drive the auxiliary power branch 30 to full output to support the output voltage. The preset shoot-through drive signal is designed to rapidly increase the current of the auxiliary power branch 30. It can be a drive with a near full duty cycle or a drive set at a higher peak current limit. Its duration is limited to the current switching cycle. Once the sampling value at the bottom of the next cycle rises back above the threshold, the shoot-through is exited and the control is returned to normal. Thus, the circuit is only pushed during the few cycles when the over-limit is actually exceeded, without long-term disturbance to the normal scheduling.

[0094] The following section describes the quantile self-tuning process, which allows the recording of protection actions to correct the predictions, automatically reducing the aggressiveness of the predictions to a safe level. Figure 13 The flowchart illustrating the quantile self-tuning steps provided in this application embodiment includes steps S901 to S905. The machine learning load range prediction model 47 outputs the quantile prediction value of the load power at the first quantile level as the lower bound. In step S901, each triggering of the out-of-bounds trigger comparator is recorded as a missed event. The out-of-bounds trigger comparator is triggered when the valley sampling value is lower than the valley voltage lower limit, i.e., when the output voltage drops significantly. Each trigger means that this surge was not immediately caught by the standby configuration, reflecting that the prediction range does not reserve enough for this surge. Therefore, the number of triggers is the true count of the predicted narrowing. In step S902, the frequency of missed events is counted within a preset-length missed event statistics window. The window for underreporting statistics can be 1,000 to 100,000 switching cycles, typically 10,000 switching cycles. If the window is too short, the frequency statistics will have high noise and be easily misled by occasional jitter. If the window is too long, the setting will be slow and unable to keep up with the change of operating conditions. Therefore, the length should be chosen to cover enough potential underreporting times without being too smooth.

[0095] The process then proceeds to two levels of judgment. In step S903, when the occurrence frequency is greater than the preset upper limit, the level of the first quantile is lowered by the preset quantile step size; in step S904, when the occurrence frequency is less than the preset lower limit, the level of the first quantile is raised by the preset quantile step size; when the occurrence frequency falls between the preset lower limit and the preset upper limit, the level of the first quantile remains unchanged.

[0096] The preset upper limit of frequency can be set to 5 to 50 times per missed report statistical window, typically 20 times; the preset lower limit of frequency can be set to 0 to 5 times per missed report statistical window, typically 2 times; the preset quantile step size can be set to 0.005 to 0.02, typically 0.01. The fixed interval between the upper and lower limits constitutes hysteresis, preventing adjustments from being triggered when the frequency fluctuates within a small range, thus avoiding back-and-forth fluctuations in the first quantile level. A higher upper limit will tolerate too many missed reports before convergence begins, protecting the intervention frequency bias; a lower limit will tolerate excessive conservatism for a long time, resulting in insufficient energy saving. A step size that is too large will cause overshoot in tuning, while a step size that is too small will cause slow tuning. The three factors are compromised based on the acceptable missed report rate and tuning speed. In step S905, the adjusted first quantile level is written back to the machine learning load interval prediction model 47, so that it outputs the lower bound according to the adjusted first quantile level. Figure 13 The code also shows a double negative branch, where if neither of the two-level judgments is satisfied, the code returns to step S901 via the upper right loop and continues the next round of statistics without making any adjustments.

[0097] This adjustment is a convergent negative feedback, not a divergent positive feedback. When there are too many missed reports and the frequency exceeds the upper limit, the first quantile level is lowered. This lowering has two effects in the same direction: firstly, the lower bound is lowered, and the necessary power is reduced, so the fixed portion borne by the main power branch 20 becomes smaller; secondly, the difference between the upper and lower bounds of the power to be responded to is widened, and the standby configuration module 44 accordingly deploys more auxiliary power branches 30 or increases the standby current value, thereby enhancing the standby capacity. Both effects aim to make it more difficult for the actual load to exceed the necessary power, thus reducing subsequent missed reports and lowering the frequency. The high frequency is thus suppressed by the self-triggered adjustment, which is a characteristic of negative feedback.

[0098] Conversely, when false negatives are extremely rare and the frequency is below the lower limit, raising the first quantile level inevitably increases power, allowing more power to be efficiently handled by the main power branch 20. The standby band narrows to save on standby losses, at the cost of a slight increase in false negatives and a rebound in frequency, which is then pulled back to the stationary range. The system eventually stabilizes in a state where the false negative frequency falls between the upper and lower limits, avoiding both uncontrolled false negatives due to energy conservation and excessively conservative standby. Combined with the hysteresis stationary range, the first quantile level converges without oscillation. As for the overall retraining of the machine learning load range prediction model 47 itself, it is still performed offline periodically as described above. Online self-tuning only fine-tunes the first quantile level; the two are separated in time scale and do not interfere with each other. Only the first quantile level is fine-tuned online, while the second quantile level remains unchanged, because the lower bound corresponding to the first quantile level directly determines the necessary power and false negatives and needs to be quickly adjusted according to operating conditions. The upper bound corresponding to the second quantile level only affects the upper limit of standby, and its deviation can be corrected by offline periodic retraining without frequent online modifications.

[0099] The following is a system embodiment corresponding to the above method, illustrating how each step is implemented on the module of the digital controller 40. Figure 14 This is a block diagram of a load-prediction-based intelligent power efficiency scheduling system provided in an embodiment of this application. The system includes a power converter 10 and a digital controller 40. The power converter 10 includes a main power branch 20 and at least one auxiliary power branch 30. The digital controller 40 includes a sampling module 41, an interval prediction module 42, and an energy efficiency scheduling module 43, as well as a standby configuration module 44 and a switching module 45. Each module corresponds one-to-one with the aforementioned method steps. The responsibilities of each module are explained below; their internal processing details are the same as the method section and will not be repeated.

[0100] Sampling module 41 is used to acquire the output voltage sequence and output current sequence of power converter 10, corresponding to the acquisition stage in step S101. It also provides two sampling points: the midpoint sampling value and the valley sampling value of the divergent sequence. Figure 14 The power converter 10 is connected to the sampling module 41 via a dashed line, as shown.

[0101] The interval prediction module 42 incorporates a machine learning load interval prediction model 47. This model receives the output voltage and current sequences as inputs and outputs a lower and upper bound. The lower bound is determined as the necessary power, and the difference between the upper and lower bounds is determined as the power to be responded to, corresponding to step S101 and the subsequent power bisection. The energy efficiency scheduling module 43 controls the main power branch 20 to perform energy efficiency scheduling with the necessary power as the output target and locked at a preset optimal efficiency operating point, corresponding to step S102. Figure 14 The power obtained by the mid-range prediction module 42 is sent to the energy efficiency scheduling module 43 via a vertical arrow, and the energy efficiency scheduling module 43 then issues control to the main power branch 20 based on this.

[0102] The standby configuration module 44 has a built-in standby configuration template library 46, which is used to query the standby configuration template library 46 based on the power to be responded to, to obtain the number of auxiliary power branches 30 to be put into operation and the standby current value. When the power to be responded to is greater than or equal to a preset standby exemption threshold, the auxiliary power branches 30 put into operation are controlled to maintain the standby current value. When the power to be responded to is less than the preset standby exemption threshold, all auxiliary power branches 30 are turned off, so that the power converter 10 enters the energy-saving mode, corresponding to steps S103 and S104. Figure 14 The power of the waiting response is bypassed by the interval prediction module 42 and sent to the standby configuration module 44.

[0103] The switching module 45 is used to, when the auxiliary power branch 30 is maintaining a standby current value, detect that the actual load power exceeds a certain power, and then have the portion of the actual load power exceeding the certain power handled by the auxiliary power branch 30 maintaining the standby current value, corresponding to step S105. Figure 14 The actual load power obtained by the sampling module 41 is sent to the switching module 45 via the outer side for comparison with the required power.

[0104] The above-mentioned over-boundary direct-through step and quantile self-tuning step can be implemented as additional functions of the interval prediction module 42 and the switching module 45. The former directly enables the auxiliary power branch 30 based on the over-boundary trigger comparator, and the latter writes back the first quantile level in the interval prediction module 42 according to the missed reporting frequency. The two share the valley sampling value obtained by the sampling module 41.

[0105] Thus, a closed-loop correspondence is formed from the method to the system: the sampling module 41 acquires data, the interval prediction module 42 divides the power, the energy efficiency scheduling module 43 manages the efficiency of the deterministic part, the standby configuration module 44 prepares the response for the uncertain part, the switching module 45 completes the immediate relay for sudden increases, and the over-limit pass-through and quantile self-tuning further cover and correct the prediction from both hardware and tuning levels. This system is suitable for powering data centers and artificial intelligence servers with drastic load fluctuations, and is also suitable for other multi-branch switching power supplies that include a main power branch 20 and at least one auxiliary power branch 30.

[0106] 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 smart power efficiency scheduling method based on load prediction, applied to a power converter including a main power branch and at least one auxiliary power branch, characterized in that, include: The output voltage sequence and output current sequence of the power converter are collected, and the output voltage sequence and output current sequence are input into a pre-trained machine learning load range prediction model to output the lower and upper bounds of the load power in the future prediction period. The lower bound is determined as the necessary power, and the main power branch is controlled to perform energy efficiency scheduling with the necessary power as the output target and locked at the preset optimal efficiency operating point. The difference between the upper bound and the lower bound is determined as the power to be responded to. The standby configuration template library is queried using the power to be responded to to obtain the number of auxiliary power branches that need to be put into operation and the standby current value of each auxiliary power branch that is put into operation. When the power to be responded to is greater than or equal to a preset standby exemption threshold, the auxiliary power branch that has been put into operation is controlled to maintain the standby current value. When the power to be responded to is less than the preset standby exemption threshold, all auxiliary power branches are shut down, so that the power converter enters an energy-saving mode powered only by the main power branch. When the auxiliary power branch maintains the standby current value, and the actual load power is detected to be greater than the necessary power, the portion of the actual load power exceeding the necessary power is borne by the auxiliary power branch maintaining the standby current value.

2. The intelligent power energy efficiency scheduling method based on load prediction according to claim 1, characterized in that, The machine learning load interval prediction model is a quantile regression model, which includes a shared temporal feature extraction layer and two parallel quantile output layers. The time-series feature extraction layer is used to extract the time-domain variation features of the load power from the output voltage sequence and the output current sequence. The first quantile output layer outputs a load power value corresponding to the first quantile level as the lower bound, and the second quantile output layer outputs a load power value corresponding to the second quantile level as the upper bound. Both the first quantile level and the second quantile level are quantile values ​​between 0 and 1, and the second quantile level is greater than the first quantile level.

3. The intelligent power energy efficiency scheduling method based on load prediction according to claim 2, characterized in that, The training process of the quantile regression model includes: collecting multiple output voltage and output current sequences of the power converter during its historical operation; for each output voltage and output current sequence, using the measured load power values ​​obtained within the future prediction period after the end of the collection of the segment as load power labels to form training samples. The first quantile output layer is given a quantile loss at the first quantile level, the second quantile output layer is given a quantile loss at the second quantile level, and supervised training of the parameters of the quantile regression model is performed using the training samples.

4. The intelligent power energy efficiency scheduling method based on load prediction according to claim 1, characterized in that, The standby configuration template library includes multiple standby configuration templates. Each standby configuration template includes a power value range to be responded to, the number of auxiliary power branches corresponding to the power value range to be responded to, and a standby current value corresponding to the power value range to be responded to. The power value ranges to be responded to in the multiple standby configuration templates do not overlap and continuously cover the preset power value range to be responded to. The query standby configuration template library includes: determining the range of values ​​of the power to be responded to that falls into, and outputting the number of auxiliary power branches and the standby current value in the standby configuration template to which the power to be responded to falls into the range of values ​​of the power to be responded to. The number of auxiliary power branches in each of the standby configuration templates is the smallest integer not less than the upper limit of the range of power values ​​to be responded to, divided by the rated power of a single auxiliary power branch. The standby current value in each standby configuration template is determined by a standby current calibration process, which is as follows: auxiliary power branches are activated according to the number of auxiliary power branches in the standby configuration template; a load step with an amplitude equal to the upper limit of the range of the power value to be responded to is applied to the power converter; the standby current value of each activated auxiliary power branch is increased successively until the drop amplitude of the output voltage of the power converter after the load step does not exceed the preset drop amplitude upper limit; and the first standby current value that makes the drop amplitude not exceed the preset drop amplitude upper limit during the successive increase process is written into the standby configuration template.

5. The intelligent power energy efficiency scheduling method based on load prediction according to claim 1, characterized in that, The control of the main power branch to perform energy efficiency scheduling with the required power as the output target and locked at the preset optimal efficiency operating point includes: keeping the inductor current of the main power branch in a continuous conduction mode, keeping the synchronous rectifier tube of the main power branch in synchronous rectification operation, and fixing the switching frequency of the main power branch to the switching frequency value corresponding to the preset optimal efficiency operating point. The preset optimal operating point is determined by an efficiency calibration curve, which is a curve obtained by calibrating the conversion efficiency of the power converter with the switching frequency under multiple output power values. The preset optimal operating point is the point on the efficiency calibration curve where the conversion efficiency reaches its maximum value corresponding to the output power value that is closest to the required power.

6. The intelligent power energy efficiency scheduling method based on load prediction according to claim 1, characterized in that, The process of acquiring the output voltage sequence and output current sequence of the power converter further includes: sampling the output voltage of the power converter at the midpoint phase during each switching cycle of the main power branch to obtain the midpoint sampling value; The output voltage of the power converter is sampled at a preset sampling phase before the synchronous rectifier tube of the main power branch is turned off, and the valley sampling value is obtained. The difference between the valley sampling value and the midpoint sampling value within the same switching cycle is calculated, and the differences of each switching cycle are arranged in chronological order to form a divergence sequence. The divergence sequence, along with the output voltage sequence and the output current sequence, are then input into the machine learning load range prediction model.

7. The intelligent power energy efficiency scheduling method based on load prediction according to claim 6, characterized in that, The preset sampling phase is determined by a sampling phase calibration process, which is as follows: when the power converter is running under no-load, the output voltage is sampled by multiple candidate phases before the synchronous rectifier is turned off, and the candidate phase with the smallest sample value is determined as the preset sampling phase. Before inputting the divergent sequence into the machine learning load interval prediction model, the divergent sequence is further processed as follows: the mean of the divergent sequence and the maximum value of the difference between two adjacent terms of the divergent sequence are calculated within a sliding statistical window of a preset length, and the mean and the maximum value are used as input features of the divergent sequence and input into the machine learning load interval prediction model.

8. The intelligent power energy efficiency scheduling method based on load prediction according to claim 6, characterized in that, It also includes a cross-boundary pass-through step: subtracting the voltage drop from the valley baseline voltage to obtain the valley voltage lower limit, and setting the valley voltage lower limit as the threshold of the cross-boundary trigger comparator; The valley baseline voltage is the average value of the valley sampled value within a preset baseline statistical window, and the voltage drop is the quotient obtained by dividing the product of the necessary power and the switching cycle of the main power branch by the product of the output capacitor value of the power converter and the output voltage setting value of the power converter. When the valley sampling value is lower than the threshold of the over-boundary trigger comparator, the output of the over-boundary trigger comparator directly bypasses the feedback control path of the machine learning load range prediction model and the power converter, and enables all auxiliary power branches with a preset direct drive signal in the current switching cycle.

9. The intelligent power energy efficiency scheduling method based on load prediction according to claim 8, characterized in that, The machine learning load interval prediction model outputs the quantile prediction value of the load power at the first quantile level as the lower bound; The intelligent power energy efficiency scheduling method based on load prediction also includes a quantile self-tuning step: each time the out-of-bounds trigger comparator is triggered, it is recorded as a missed event, and the frequency of the missed events is counted within a missed event statistics window of a preset length. When the occurrence frequency is greater than the preset frequency upper limit, the first quantile level is lowered by the preset quantile step size; When the occurrence frequency is less than the preset frequency lower limit, the preset quantile step size is increased by the first quantile level. The adjusted first quantile level is written back to the machine learning load interval prediction model, so that the machine learning load interval prediction model outputs the lower bound according to the adjusted first quantile level.

10. A smart power energy efficiency scheduling system based on load forecasting, used to implement the smart power energy efficiency scheduling method based on load forecasting as described in claim 1, characterized in that, It includes a power converter and a digital controller. The power converter includes a main power branch and at least one auxiliary power branch. The digital controller includes a sampling module, an interval prediction module, an energy efficiency scheduling module, a standby configuration module, and a switching module. The sampling module is used to collect the output voltage sequence and output current sequence of the power converter; The interval prediction module has a built-in pre-trained machine learning load interval prediction model, which is used to input the output voltage sequence and the output current sequence into the machine learning load interval prediction model, output the lower bound and upper bound of the load power in the future prediction period, determine the lower bound as the necessary power, and determine the difference between the upper bound and the lower bound as the power to be responded to. The energy efficiency scheduling module is used to control the main power branch to perform energy efficiency scheduling with the necessary power as the output target and lock it at the preset optimal efficiency operating point. The standby configuration module has a built-in standby configuration template library, which is used to query the standby configuration template library with the power to be responded to, to obtain the number of auxiliary power branches to be put into operation and the standby current value of each auxiliary power branch to be put into operation. When the power to be responded to is greater than or equal to a preset standby exemption threshold, the auxiliary power branches to be put into operation are controlled to maintain the standby current value. When the power to be responded to is less than the preset standby exemption threshold, all auxiliary power branches are turned off, so that the power converter enters an energy-saving mode in which only the main power branch is powered. The load switching module is used to, when the actual load power exceeds the necessary power while the auxiliary power branch is maintaining the standby current value, transfer the portion of the actual load power exceeding the necessary power to the auxiliary power branch that is maintaining the standby current value.

Images