A power supply parameter adaptive optimization method and system based on time series data

Abstract

The application discloses a power supply parameter adaptive optimization method and system based on time sequence data, relates to the field of power management and control, and comprises the following steps: collecting voltage and current time sequence data of input and output of a switching power supply, dividing current recurrence intervals according to output current, marking intervals as uplink crossing or downlink crossing according to positive and negative output current change rates; obtaining energy efficiency samples from the ratio of output power to input power during the validity period of candidate modulation parameters, classifying the energy efficiency samples into uplink and downlink sample sets according to directions and counting energy efficiency averages, counting uplink and downlink cumulative residence time lengths a and b under target operating conditions, taking the sum of the two averages multiplied by a and b as an interval optimization target, summing up each interval to obtain a total optimization target, and determining target modulation parameters by a machine learning optimization model under the constraint that the two averages are not lower than an energy efficiency threshold and the total optimization target is maximum, and writing the target modulation parameters into a modulator. The application makes the power supply parameter adaptive optimization based on time sequence data no longer dependent on the composition of the load during sampling.

Description

Technical Field

[0001] This application relates to the field of power management and control technology, and in particular to a method and system for adaptive optimization of power parameters based on time-series data. Background Technology

[0002] In a data center rack, a server power supply experiences vastly different load conditions throughout the day: during the day, when handling inference loads, the output current rapidly fluctuates between medium and heavy loads; at night, when performing offline tasks, the output current remains at a light load for extended periods, occasionally jumping to full load. Under these two load conditions, the junction temperature, core temperature, and magnetization history of the power switching transistors differ, and the overall energy efficiency of a switching power supply is precisely sensitive to these parameters.

[0003] The modulation parameters of a switching power supply, including switching frequency, duty cycle, dead time, and number of phases, determine the distribution ratio between switching losses, conduction losses, and magnetic component losses, thus directly determining the overall energy efficiency. Early practices involved setting these parameters to fixed values ​​and embedding them in the controller before shipment. Later, a lookup table approach based on operating point emerged. The scheme in publication number CN101006406A maps the optimal dead time of the synchronous rectification circuit according to the circuit's operating range and stores it in memory. During operation, it looks up the value in the table based on monitored circuit parameters, and periodically updates the dead time in the table using changes in duty cycle as a proxy for power consumption changes. In recent years, a further approach has emerged to optimize energy efficiency by adjusting parameters online based on the measured load. The scheme in publication number CN107078647A dynamically adjusts the sleep duration of burst mode in the resonant converter according to the measured output load current, maintaining the converter efficiency at the highest level. A more general approach is to continuously collect the time-series data of input voltage, input current, output voltage and output current, divide the running time into several load levels according to the output current, statistically analyze the measured energy efficiency in each load level, and then use an optimization algorithm to search for the set of parameters that maximizes the measured energy efficiency in the parameter space and write it into the modulator.

[0004] The common premise of the above online optimization methods is that as long as the output current falls within the same load range, the collected efficiency samples are comparable. However, this premise does not hold true under real-world conditions of dynamic load changes: the on-resistance of the power switch increases with junction temperature, core losses first decrease and then increase with core temperature, and both junction and core temperatures are determined by the load history over a previous period, with thermal time constants ranging from milliseconds to tens of seconds. The same output current value may rise from a light load or fall from a heavy load, and the thermal state of the device differs greatly in these two cases. Therefore, the efficiency samples within the same load range are mixed with the two thermal states, and the statistically obtained average efficiency no longer reflects only the modulation parameters themselves, but also the composition of the load source during the sampling period.

[0005] None of these solutions considered the contamination of energy efficiency samples by the load path, and therefore failed to address the dependence of the optimization conclusion on the load path as a problem to be solved. Taking the aforementioned server power supply as an example, during the day, most samples are from higher load levels, resulting in hotter components, and the optimization biases towards parameters with lower energy loss at high temperatures. At night, most samples are from lower load levels, resulting in cooler components, and the optimization biases towards a different set of parameters. Thus, the same machine yields different optimal parameters on different shifts, and the optimization conclusion fluctuates with the sampling period. The selected parameters are often most efficient only under a certain thermal condition. In real-world operating conditions, components constantly fluctuate between hot and cold, and these parameters only realize their optimal efficiency during certain periods, resulting in cumulative energy consumption higher than predicted by the optimization conclusion.

[0006] Furthermore, to obtain energy efficiency samples for online optimization, candidate parameters must be actually applied to powered equipment for a period of time. However, whether candidate parameters will cause the power switching transistor current or temperature rise to exceed limits cannot be known before they take effect. Therefore, online optimization functions are often disabled in the field, or the exploration range is narrowed, resulting in the loss of adaptive capabilities. After several months of equipment operation, component aging will cause an overall decrease in energy efficiency. This decrease is similar to the decrease caused by temperature in the energy efficiency samples. Optimization will treat aging as a deterioration of modulation parameters and continuously adjust in the wrong direction. Although adding multiple temperature sensors to directly measure junction temperature and core temperature can partially alleviate the problem, the space and cost budget of server power supplies do not allow for the installation of sensors on each power device, and mass-produced models do not have the conditions for calibrating thermal models on a unit-by-unit basis.

[0007] Therefore, how to make the adaptive optimization of power parameters based on time-series data independent of the load path during sampling without adding sensors, and how to maintain the safety boundary of devices during live exploration and distinguish the effects of aging and temperature during long-term operation are unsolved problems in the field of power management and control. Summary of the Invention

[0008] The technical problem to be solved by this application is to provide a method and system for adaptive optimization of power parameters based on time-series data, so that the adaptive optimization of power parameters based on time-series data no longer depends on the load path during the sampling period.

[0009] To achieve the above objectives, this application adopts the following technical solution:

[0010] An adaptive optimization method for power parameters based on time-series data is applied to a switching power supply containing a power conversion circuit. The method includes: collecting time-series data of the input voltage, input current, output voltage and output current of the switching power supply; dividing the output current into multiple current reproduction intervals according to the range of output current values; and marking each entry into the current reproduction interval as an upward crossover or a downward crossover based on the sign of the rate of change of the output current with respect to time.

[0011] During the period when a set of candidate modulation parameters are in effect, a segment of the time-series data is extracted for each crossover, and an energy efficiency sample is obtained by the ratio of output power to input power. The energy efficiency samples obtained from the uplink and downlink crossovers are respectively assigned to the uplink sample set and the downlink sample set according to the current reproduction interval.

[0012] Within the same current reproduction interval, the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set are statistically analyzed, and the cumulative dwell time of the current uplink crossing and downlink crossing under the target operating condition is statistically analyzed, which are a and b respectively.

[0013] With the constraint that both of the energy efficiency average values ​​are not lower than the energy efficiency threshold, the sum of the energy efficiency average value of the uplink sample set multiplied by a and the energy efficiency average value of the downlink sample set multiplied by b is used as the interval optimization objective. The sum of the interval optimization objectives of all current reproduction intervals is used as the total optimization objective. The optimization objective is maximized as the optimization objective. The target modulation parameter is determined from multiple candidate modulation parameters by the machine learning optimization model.

[0014] The target modulation parameters are written into the modulator of the power conversion circuit, and the power switching transistor in the power conversion circuit is driven according to the target modulation parameters.

[0015] Optionally, marking each entry into the current reproduction interval as an upward or downward crossing includes marking an upward crossing when the output current changes from below the lower boundary of the current reproduction interval to falling into the current reproduction interval, and marking a downward crossing when the output current changes from above the upper boundary of the current reproduction interval to falling into the current reproduction interval. For any crossing, if the number of consecutive sampling points of the output current falling within the current reproduction interval is less than the lower limit of the number of sampling points, or the difference between the maximum and minimum values ​​of the input voltage within the consecutive sampling points is greater than the upper limit of the input voltage fluctuation, the energy efficiency sample corresponding to the crossing is discarded.

[0016] Optionally, the constraints also include output voltage ripple constraints and load step constraints. The output voltage ripple constraint is that the peak-to-peak value of the output voltage in a segment of time-series data corresponding to the energy efficiency sample is not greater than the ripple upper limit. During the effective period of the candidate modulation parameters, when the change in output current within a set time period is greater than the step threshold, a load step event is determined to have occurred. A segment of output voltage time-series data after the load step event is extracted, and the maximum amplitude of the output voltage deviating from the output voltage set value is taken as the overshoot amplitude. The time taken for the output voltage to return to the allowable deviation band from the load step event to the output voltage set value is taken as the recovery time. The load step constraint is that the overshoot amplitude corresponding to the cumulative set number of load step events is not greater than the overshoot upper limit and the recovery time is not greater than the recovery time upper limit.

[0017] Optionally, the training process of the machine learning optimization model includes: using the values ​​of candidate modulation parameters that have already taken effect and the corresponding current reproduction interval identifiers as inputs; using the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set as outputs; and using the average energy efficiency actually calculated from the uplink sample set and the downlink sample set as the average energy efficiency labels for the uplink sample set and the downlink sample set, respectively, to perform supervised training on the machine learning optimization model; determining the target modulation parameter from multiple sets of candidate modulation parameters includes inputting candidate modulation parameters that have not yet taken effect into the trained machine learning optimization model, to obtain... For each current reproduction interval, the predicted values ​​of the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set are used. Candidate modulation parameters whose predicted average energy efficiency is lower than the energy efficiency threshold are eliminated. For the remaining candidate modulation parameters, a set number of candidate modulation parameters are selected in descending order of the overall optimization target calculated from the two predicted average energy efficiency values ​​and are applied sequentially. The two average energy efficiency values ​​and the overall optimization target are recalculated from the uplink sample set and the downlink sample set actually obtained during the application period. The candidate modulation parameter with the largest overall optimization target and whose two average energy efficiency values ​​are not lower than the energy efficiency threshold is determined as the target modulation parameter.

[0018] Optionally, the candidate modulation parameters include at least two of the following: switching frequency, duty cycle, dead time, and number of phases; the multiple sets of candidate modulation parameters are generated from a neighborhood centered on the currently effective target modulation parameter, and the neighborhood expands to both sides on the respective value axes of each parameter included in the candidate modulation parameters by a set step size; after one round of optimization, when the increase in the total optimization target of the newly determined target modulation parameter relative to the total optimization target of the target modulation parameter determined in the previous round is less than the convergence threshold, the set step size is reduced by a set ratio and the next round of optimization is entered; the candidate modulation parameters also include a mode switching threshold, which is the output current value on which the power conversion circuit switches between continuous conduction mode and intermittent conduction mode.

[0019] Optionally, before activating a set of candidate modulation parameters, the maximum value of the input current actually collected within the most recent statistical period is taken as the current envelope value, and the maximum value of the rate of change of the chassis temperature with respect to time actually collected within the most recent statistical period is taken as the temperature rise rate envelope value. The current envelope value and the temperature rise rate envelope value constitute a trial stress envelope. After the rate of change of the output current with respect to time changes from negative to zero, the set of candidate modulation parameters is activated.

[0020] Optionally, during the effective period of the set of candidate modulation parameters, the real-time acquired input current is continuously compared with the current envelope value, and the real-time acquired rate of change of chassis temperature over time is compared with the temperature rise rate envelope value; when the input current is greater than the current envelope value, or the rate of change of chassis temperature over time is greater than the temperature rise rate envelope value, the target modulation parameters that are currently effective are rewritten into the modulator, the energy efficiency samples generated by the set of candidate modulation parameters during this effective period are discarded, and the set of candidate modulation parameters is marked as prohibited from being tested within the corresponding current reproduction interval.

[0021] Optionally, for any uplink pass, if the output current is continuously lower than the lower boundary of the lowest current reproduction interval before the uplink pass, and the duration is not less than the lower limit of the resting time, and the difference between the chassis temperature and the ambient temperature is not greater than the upper limit of the temperature difference, the energy efficiency sample corresponding to the uplink pass is taken as the cold anchor point sample; the cold anchor point samples within the same current reproduction interval and under the same set of modulation parameters are arranged in chronological order to form a cold anchor point sequence.

[0022] Optionally, the difference between the median of the cold anchor sequence in the most recent aging statistical window and the median of the cold anchor sequence in the first aging statistical window is taken as the aging component; before calculating the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set, the difference between each energy efficiency sample in each current reproduction interval and the aging component is taken as the corrected energy efficiency sample, and the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set are calculated from the corrected energy efficiency sample; when the absolute value of the aging component is greater than the aging threshold, the multiple sets of candidate modulation parameters are regenerated and optimization is re-executed.

[0023] This application also discloses a power supply parameter adaptive optimization system based on time-series data, applied to a switching power supply containing a power conversion circuit, used to implement the power supply parameter adaptive optimization method based on time-series data described in this application, including: a time-series acquisition module, used to acquire time-series data of the input voltage, input current, output voltage and output current of the switching power supply;

[0024] The crossing marker module is used to divide multiple current reproduction intervals according to the range of output current values, and to mark each entry into the current reproduction interval as an upward crossing or a downward crossing based on the sign of the rate of change of the output current with respect to time.

[0025] The sample collection module is used to extract a segment of the time-series data for each crossover during the effective period of a set of candidate modulation parameters, obtain energy efficiency samples from the ratio of output power to input power, establish uplink sample sets and downlink sample sets according to the current reproduction interval, and classify the energy efficiency samples obtained from the uplink crossover and the downlink crossover into the uplink sample set and the downlink sample set respectively.

[0026] The target synthesis module is used to calculate the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set within the same current reproduction interval, and to calculate the cumulative dwell time of the uplink and downlink current crossings, a and b, respectively, under the target operating conditions. The sum of the average energy efficiency of the uplink sample set multiplied by a and the average energy efficiency of the downlink sample set multiplied by b is used as the interval optimization target, and the sum of the interval optimization targets of all current reproduction intervals is used as the overall optimization target.

[0027] An optimization decision module is set in the controller of the switching power supply. The controller is a microcontroller, a digital signal processor, or a field-programmable gate array. It is used to determine the target modulation parameter from multiple sets of candidate modulation parameters by a machine learning optimization model, with the constraint that the two average energy efficiency values ​​are not lower than the energy efficiency threshold and the optimization objective being maximized.

[0028] The parameter delivery module includes a modulator and a power switch driver circuit of the power conversion circuit, used to write the target modulation parameters into the modulator and drive the power switch in the power conversion circuit according to the target modulation parameters.

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

[0030] This application enables adaptive optimization of power parameters based on timing data to no longer depend on the load path composition during the sampling period. The on-resistance of power switches increases with junction temperature, and core losses initially decrease and then increase with core temperature. Since junction and core temperatures are determined by previous load history, with thermal time constants ranging from milliseconds to tens of seconds, devices tend to be cooler when the same output current is applied and hotter when it is applied. Based on this, this application divides the energy efficiency samples into an uplink sample set and a downlink sample set according to the crossing direction. The average energy efficiency of each set corresponds to only one thermal state and is essentially free of load path components, providing a clean measure of the losses of this set of parameters in both cold and hot states. Furthermore, this application does not mix and average the samples during the sampling period. Instead, it uses the cumulative dwell time 'a' and 'b' of the current crossing the uplink and downlink intervals as weights. The sum of the average energy efficiency of the uplink sample set multiplied by 'a' and the average energy efficiency of the downlink sample set multiplied by 'b' is used as the interval optimization objective. The sum of all intervals is used as the overall optimization objective. Under the constraint that neither of the two average energy efficiency values ​​is lower than the energy efficiency threshold, the optimization objective is to maximize the overall optimization objective. Since a and b represent the actual dwell time in two directions under the target operating conditions, the interval optimization objective is the average energy efficiency calculated based on the proportion of the two thermal states according to the actual time. The optimization conclusion does not fluctuate with the sampling period. The selected parameters meet the standards in both cold and hot states and are the most energy-efficient under the actual cold-hot round-trip ratio. Taking a server power supply as an example, the two sets of parameters have similar energy efficiency on a constant temperature test bench, but one set is more sensitive to temperature. After being installed in the rack, it only achieves the required energy efficiency when the device temperature is close to that of the test bench. This application directly measures this difference from the operating data and calculates the true average energy efficiency.

[0031] Using this application, online optimization can be truly initiated on powered equipment. This application uses the maximum value of the input current within the most recent statistical period as the current envelope value and the maximum value of the chassis temperature change rate within the same period as the temperature rise rate envelope value, forming a test stress envelope. After the output current change rate changes from negative to zero, a set waiting period is set to activate the candidate modulation parameters, and the real-time value is continuously compared with the envelope value. Once exceeded, the previously activated target modulation parameter is written back, and the current energy efficiency sample is discarded. The stress upper limit is reliable because the recent peak input current and chassis temperature rise rate values ​​are quantities that the hardware has already withstood without triggering protection. The test is scheduled to start after the natural downward edge of the output current, because the device's temperature rise margin is at its maximum at this time. Since the stress allowed by the exploration does not exceed the levels the equipment has recently withstood, the exploration will not bring the equipment into a state that has never occurred before, and online optimization no longer needs to rely on shutting down functions for safety.

[0032] This application distinguishes between energy efficiency degradation caused by component aging and energy efficiency degradation caused by temperature, preventing optimization from continuously adjusting in the wrong direction. This application takes the first upward crossover sample after the output current has remained below the minimum current reproduction range for an extended period and the difference between the chassis temperature and ambient temperature falls within the upper limit of the temperature difference limit as the cold-state anchor point sample. The change in the median of the cold-state anchor point sequence over time is taken as the aging component. Before calculating the average energy efficiency, the aging component is subtracted from each energy efficiency sample within each current reproduction range to obtain the corrected energy efficiency sample. This degradation can isolate aging because the effects of temperature and aging on losses are different: temperature is reversible, and the energy efficiency returns to its original value as the device cools down; aging is irreversible, and the energy efficiency does not return to its original value. The cold-state anchor point samples have a consistent thermal state, and their differences can only come from irreversible changes. Therefore, the aging component essentially contains no temperature component. After subtracting it, optimization is only responsible for the modulation parameters. When the aging component exceeds the aging threshold, candidate modulation parameters are regenerated and optimization is performed again.

[0033] The combined design features of the above components enable this application to complete all discrimination without adding sensors. The crossing direction is determined by the sign of the output current change rate; the energy efficiency under each thermal state is derived from the average energy efficiency of the uplink and downlink sample sets; the operating condition weight is derived from the actual cumulative dwell time of the current during uplink and downlink cycles; and the stress upper limit and aging component are derived from the measured peak value and cold anchor point sequence, respectively. These criteria require no additional hardware because the input voltage, input current, output voltage, output current, and chassis temperature are already collected by the switching power supply for closed-loop control and protection. The sampling channels and computing power overhead are not increased by this application. Since no junction temperature or core temperature sensors are introduced, and there is no need to build thermal models for each unit, this application can run directly on a microcontroller, digital signal processor, or field-programmable gate array, making it suitable for server power supplies, base station power supplies, and charging modules with drastic load dynamics. Attached Figure Description

[0034] Figure 1 A schematic diagram of the structure of the power parameter adaptive optimization system based on time-series data provided in the embodiments of this application;

[0035] Figure 2 A flowchart illustrating the adaptive optimization method for power parameters based on time-series data provided in this application embodiment;

[0036] Figure 3 A schematic diagram illustrating the output current timing and current reproduction interval crossing provided in the embodiments of this application;

[0037] Figure 4 A schematic diagram of the data flow for the energy efficiency sample collection and the synthesis of the interval optimization objective and the overall optimization objective provided in the embodiments of this application;

[0038] Figure 5A schematic diagram illustrating the difference in average energy efficiency between the uplink and downlink sample sets within the same current reproduction interval provided in this application embodiment;

[0039] Figure 6 A block diagram of a power parameter adaptive optimization system based on time-series data provided in an embodiment of this application;

[0040] Figure 7 A schematic diagram of the data flow for training the machine learning optimization model and determining the target modulation parameters provided in the embodiments of this application;

[0041] Figure 8 A schematic diagram illustrating the effective timing of the trial stress envelope and candidate modulation parameters provided in the embodiments of this application;

[0042] Figure 9 A schematic diagram illustrating the extraction of cold anchor point sequences and aging components provided in an embodiment of this application;

[0043] Figure 10 This is a schematic diagram showing the output voltage overshoot amplitude and recovery time under a load step event provided in an embodiment of this application;

[0044] Figure 11 This is a schematic diagram illustrating the generation of candidate modulation parameter neighborhoods and the setting of step size shrinkage provided in an embodiment of this application. Detailed Implementation

[0045] 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.

[0046] 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.

[0047] The solution provided in this application is applied to switching power supplies containing power conversion circuits, and is particularly suitable for server power supplies, base station power supplies, and charging modules with drastic load dynamics. Taking a server power supply in a data center rack as an example, it experiences very different load paths throughout the day. Here, the load path refers to whether the same output current value is increased from a lighter load or decreased from a heavier load, that is, the time direction before the current reaches the current value. During the day when handling inference batches, the output current rapidly shuttles between medium and heavy loads; at night when performing offline tasks, the output current remains at a light load for a long time, occasionally jumping to full load. The junction temperature of the power switching transistor and the core temperature are not the same under these two paths, and the overall energy efficiency of the switching power supply is precisely sensitive to the junction temperature and core temperature. This is the starting point of this application to find the optimal energy efficiency under real operating conditions in real load cycles.

[0048] Figure 1 A schematic diagram of the structure of a power parameter adaptive optimization system based on time-series data provided in an embodiment of this application. Figure 1 As shown, the main circuit of the switching power supply is a power conversion circuit 10. The power switching transistor 11 in the power conversion circuit 10 is switched on and off under the control of the drive signal output by the modulator 12, converting the electrical energy from the input side and sending it to the load 50. The controller 20 is connected to the power conversion circuit 10; it can be a microcontroller, a digital signal processor, or a field-programmable gate array (FPGA). The power parameter adaptive optimization system 40 operates within the controller 20. Its timing acquisition module 41 is connected to the input and output terminals of the power conversion circuit 10, acquiring the input voltage, input current, output voltage, and output current, while also acquiring the chassis temperature.

[0049] The following explains the sources of each signal. Input voltage, input current, output voltage, and output current are quantities that the switching power supply needs to collect in real time to achieve closed-loop voltage regulation and overcurrent protection. Chassis temperature is also a quantity that the overall thermal protection monitors. Therefore, the timing acquisition module 41 only takes data from the existing sampling channels, and no additional sensors are needed for this application. The input and output voltage and current used for statistical energy efficiency samples and determining the direction of travel can be sampled at a period of tens to hundreds of microseconds, corresponding to a sampling rate of several kilohertz to tens of kilohertz. This is because the energy efficiency samples are taken from the average over a period of time, and the direction is only observed by the slow rise and fall of the current. This rate is already available in conventional power supply control loops.

[0050] The peak-to-peak value of the output voltage ripple, the overshoot amplitude after a load step, and the recovery time—these fast dynamic quantities—are obtained by the high-speed capture channels already equipped in the switching power supply for overvoltage protection and undervoltage detection, such as peak hold circuits or high-speed comparators. Their responses are sufficient to cover switching frequencies and microsecond-level transients, and are also existing components of the equipment. The short-window moving average of the four voltage and current channels described later is only used in the energy efficiency sample and direction determination channel and does not affect the capture of ripple and transients by these high-speed channels. In this way, both slow energy efficiency quantities and fast dynamic quantities are obtained from existing hardware, and the sampling channels and computing power overhead are not increased due to this application. The controller 20 calculates the target modulation parameters based on the collected data, which are then applied to the power switch 11 via the modulator 12, forming a closed loop from acquisition to control together with the power conversion circuit 10 and the load 50.

[0051] Figure 2 This is a flowchart illustrating the power parameter adaptive optimization method based on time-series data provided in an embodiment of this application. (Refer to...) Figure 2 The method includes steps S101 to S106. Step S101 involves acquiring time-series data of input voltage, input current, output voltage, and output current. Multiple current reproduction intervals are divided according to the range of output current values. Based on the sign of the rate of change of output current with respect to time, each entry into an interval is marked as an uplink or downlink crossover. Step S102 involves obtaining an energy efficiency sample for each crossover based on the ratio of output power to input power during the effective period of a set of candidate modulation parameters. The energy efficiency samples obtained from the uplink and downlink crossovers are then assigned to the uplink sample set and downlink sample set, respectively, according to the current reproduction interval.

[0052] Step S103: Within the same current reproduction interval, calculate the average energy efficiency of the uplink and downlink sample sets, and under the target operating condition, calculate the cumulative dwell time of the uplink and downlink current crossings, denoted as a and b, respectively. Step S104: Multiply the average energy efficiency of the uplink sample set by a and the sum of the average energy efficiency of the downlink sample set by b as the interval optimization objective. Then, sum the interval optimization objectives of all current reproduction intervals as the overall optimization objective. Step S105: With the constraint that both average energy efficiency values ​​are not lower than the energy efficiency threshold, and with the goal of maximizing the overall optimization objective, a machine learning optimization model determines the target modulation parameter from multiple candidate modulation parameters. Step S106: Write the target modulation parameter into the modulator 12 of the power conversion circuit 10, and drive the power switch 11 in the power conversion circuit 10 according to the target modulation parameter. The following is combined with... Figures 3 to 11 The implementation methods and design considerations for each of the above steps will be discussed one by one.

[0053] First, let's explain the most fundamental step in the core process: the acquisition and processing of time-series data. The time-series acquisition module 41 records the input voltage, input current, output voltage, and output current into a time-stamped sequence at a fixed sampling period, retaining the acquisition time for each sampling point. The reason for retaining the time stamp is that time information is needed later for determining the crossing direction and calculating the cumulative dwell time; without time information, it's impossible to determine the rate of change and dwell time based solely on current values. The acquired four voltage and current signals undergo a light noise reduction process, such as a moving average from 3 to 7 points, typically using 5 points. Too few points won't filter out sampling glitches, while too many points will smooth out the rate of load change and slow down direction determination. Therefore, the goal is to filter out glitches while preserving the trend, laying a solid foundation for subsequent direction determination.

[0054] Next, the entire operating range of the output current 71 is divided into several current reproduction intervals. For example... Figure 3 As shown, the range of the output current 71 is divided into multiple intervals from low to high, including the first current recurrence interval 51, the second current recurrence interval 52, and the third current recurrence interval 53. The same load current value will repeatedly fall into its respective interval during equipment operation, which is why it is called a current recurrence interval. The number of intervals is determined according to the load distribution of the equipment and the curvature of the efficiency curve, and is generally divided into 4 to 16 intervals, typically 8 intervals. If the division is too coarse, the current range spanned in a single interval is too wide, the operating points of devices within the intervals vary greatly, and the energy efficiency samples themselves are not uniform. If the division is too fine, the number of times each interval is crossed is too small, and the statistical average energy efficiency is unstable. To prevent the output current from fluctuating back and forth near the interval boundary and being repeatedly recorded as crossing, a hysteresis band is set at each boundary. Its width can be 5% to 20% of the interval width, typically 10%. If the hysteresis band is too narrow, it will not block the boundary jitter, and if it is too wide, normal crossings close to the boundary will be missed. Only when the output current crosses the hysteresis band is it determined to be a real entry or exit, so as to avoid the boundary jitter from falsely increasing the number of crossings and misrecording the cumulative dwell time.

[0055] After defining the intervals, the determination of whether each entry into the interval is upward or downward is based on the sign of the rate of change of the output current with respect to time. The difference between the output currents of adjacent sampling points is divided by the sampling period to obtain the rate of change of the output current. A positive rate of change indicates that the current is increasing, and this entry into the interval is recorded as an upward crossing; a negative rate of change indicates that the current is decreasing, and this entry into the interval is recorded as a downward crossing. Because the differential is sensitive to noise, in actual calculations, the output current is first smoothed using the aforementioned short window before subtracting, or the difference between several consecutive sampling points is averaged, and only the sign is used, not the absolute magnitude. Therefore, individual glitches will not reverse the direction of the determination. Figure 3The rising segment of the output current 71 corresponds to the upward crossing, and the falling segment corresponds to the downward crossing. The same current reproduction interval usually has both upward and downward crossings within a certain operating time.

[0056] The reason for separating the efficiency samples by uplink and downlink, instead of mixing and averaging them as long as they fall within the same load range, lies in the thermal behavior of the devices. The on-resistance of power switches increases with junction temperature, and core losses initially decrease and then increase with core temperature. Neither junction nor core temperature is solely determined by the current at the current moment; rather, they are accumulated over a period of load history, with thermal time constants ranging from milliseconds to tens of seconds. This leads to an easily overlooked consequence: the same output current value could originate from a lighter load or a heavier load. When the current rises from a lower range into a certain range, the device has previously operated at a lower current and its temperature is still low; conversely, when the current falls from a higher range into the same range, the device has previously operated at a higher current and its temperature is still high due to residual heat. Therefore, within the same range, samples taken during uplink crossings correspond to cooler devices, while samples taken during downlink crossings correspond to warmer devices.

[0057] If these two types of samples are averaged without distinction, the resulting average energy efficiency value will simultaneously reflect the composition ratio of uplink and downlink samples during the sampling period. This ratio changes with the load source during the day and night, so the optimization conclusion also fluctuates with the sampling period. This is the root cause of the instability of conventional online optimization based on load level statistics under dynamic load.

[0058] Figure 5 This is a schematic diagram illustrating the difference in average energy efficiency between the uplink and downlink sample sets within the same current reproduction interval, as provided in an embodiment of this application. Figure 5 As shown, within the same current reproduction interval, the average energy efficiency of the uplink sample set 72 is higher than that of the downlink sample set 73, and both are higher than the energy efficiency threshold 74.

[0059] The average energy efficiency of 72 in the uplink sample set mainly corresponds to the relatively cool thermal state of the device, while the average energy efficiency of 73 in the downlink sample set mainly corresponds to the relatively hot thermal state of the device. Compared with the mixed average that does not distinguish between directions, both have significantly removed the influence of the load path, and thus are closer to the clean measure of the loss of the modulation parameters in the cold and hot states, respectively. Strictly speaking, for both uplink crossings, those rising from deep idle are cooler than those rising from shallow dips, and there is still a layer of second-order difference in the direction due to the different depths before the rise. However, this is much smaller than the first-order oscillation caused by the change in the ratio of uplink and downlink in the mixed average over time. If necessary, a single, fully cooled thermal benchmark can be provided by the cold state anchor point mentioned later for further normalization. Measuring the energy efficiency of the two thermal states separately is the basis for this application to calculate the average energy efficiency according to the actual operating conditions and significantly reduce the oscillation of the optimization conclusion with the sampling period.

[0060] The following examples, representing a range of magnitudes, further illustrate the above mechanism. The on-resistance of power switching transistors typically has a positive temperature coefficient. For every 100 degrees Celsius increase in junction temperature, the on-resistance may increase by approximately 1.3 to 2 times, depending on the device type, leading to a corresponding increase in conduction losses. Core losses initially decrease and then increase with core temperature. Most ferrite materials experience their lowest losses around 80 to 100 degrees Celsius; losses increase beyond this temperature range. The thermal time constant of the junction temperature is often in the range of milliseconds to seconds, while the thermal time constant between the core and the chassis / heatsink can reach tens of seconds. Therefore, after a load change, it takes this considerable time for the device temperature to gradually catch up with the load. Reflecting on efficiency, with the same modulation parameters and the same current reproduction range, the measured efficiency might be 97.8% during uplink crossing when the device is cooler, and 97.2% during downlink crossing when the device is hotter, a difference of 0.6 percentage points. If there are more upward samples in a certain period, the mixed average will be biased towards 97.8%, and if there are more downward samples in another period, the mixed average will be biased towards 97.2%. The optimal conclusion will fluctuate back and forth within this 0.6 percentage point. This application measures the two separately, and 97.8% and 97.2% are stable respectively, so the optimal conclusion no longer fluctuates.

[0061] Furthermore, the issue of measurement accuracy needs to be addressed. Efficiency is the ratio of two large numbers, output power and input power. To stably resolve a small difference of 0.6 percentage points around 97% seems to require very high measurement accuracy. The reason this application is not sensitive to absolute measurement accuracy is that optimization truly relies on the difference between the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set within the same current reproduction interval. These two sets of samples are taken from the same interval, the same batch of voltage and current sensors, and have similar sampling times. The sensor gain error and systematic errors such as zero-point bias are basically canceled out when the two are subtracted, leaving only the real difference caused by different thermal states. Therefore, even if the absolute accuracy of the sampling channel itself is only at the level of conventional protection level, this small difference between uplink and downlink can still be stably resolved, and averaging through multiple passes further reduces random noise.

[0062] The following explains the applicable boundaries of this scheme. The reason why uplink and downlink can differentiate in terms of thermal state is that the rate of load change is comparable to or slower than the thermal time constant of the device. For example, during a nighttime period of significant, slow changes between light and full load, the slowly heating magnetic core and chassis have sufficient time to form a temperature difference between the rise and fall phases, maximizing the benefit in each direction. When the load rapidly cycles between medium and heavy loads, with a cycle much shorter than the thermal time constant of the slowly heating material, the magnetic core and chassis do not have enough time to cool down or heat up during the rise and fall. In this case, the distinction between uplink and downlink mainly lies in the junction temperature on the order of milliseconds, and the benefit of separation decreases accordingly but does not disappear. For scenarios where the load is relatively stable year-round with almost no bidirectional crossover, samples in both directions are scarce, and the separation mechanism of this application degrades. In this case, the cold anchor point and aging correction discussed later are more crucial to ensure that the optimization does not deviate from the correct path.

[0063] The following explains the sampling step. During the effective period of a set of candidate modulation parameters, for each crossover, an energy efficiency sample is calculated by taking the output power and input power from the time series data corresponding to that crossover. The output power is obtained by multiplying the output voltage 80 and output current point by point during this time series, and the input power is obtained by multiplying the input voltage and input current point by point. The two are then integrated or averaged over this time period, and the ratio obtained by dividing the output power by the input power is the energy efficiency sample for this crossover. It is an efficiency reading of the set of modulation parameters in this interval and in this direction.

[0064] It's important to note that the data used for calculating energy efficiency samples is taken from a period where the current remains relatively stable within the range after entering it, rather than the period when the current is rapidly rising or falling. This is because when the current changes rapidly, the energy in the energy storage inductor also increases or decreases rapidly with the square of the current. This energy entering and leaving the inductor is mixed into the input and output power, causing the calculated efficiency to deviate from the true value. In the relatively stable current residence period, the inductor's energy storage remains almost unchanged, and this deviation is negligible. The rule of discarding samples with excessively short residence periods, as described later, ensures that the remaining samples all have sufficiently long stable residence periods. Using integration over a period of time instead of single-point division is to average out instantaneous fluctuations within the switching cycle, making the energy efficiency samples more representative of the stable efficiency within that range. In embodiments with stricter requirements for energy storage deviation, the change in inductor energy storage can be estimated from the change in current and subtracted from the power as further correction. The extracted time-series data is based on a continuous sampling period where the current remains stable within this range.

[0065] After calculating the energy efficiency samples, the sample collection module 43 places them into the corresponding sets according to the current reproduction interval and crossing direction. The same current reproduction interval corresponds to two sets: the uplink sample set contains only the energy efficiency samples obtained from the uplink crossing within that interval, and the downlink sample set contains only the energy efficiency samples obtained from the downlink crossing within that interval. Only after a set of candidate modulation parameters has been in effect for a sufficiently long time, and the accumulated sample count in each of the two sample sets for each interval has reached the lower limit of the sample count, are the energy efficiency samples in the uplink and downlink sample sets averaged within the same current reproduction interval to obtain the average energy efficiency value 72 of the uplink sample set and the average energy efficiency value 73 of the downlink sample set. This lower limit of the sample count can be between 10 and 30, typically 20. If the number is too small, the average will still be affected by random fluctuations and will not be stable enough; if the number is too large, it will prolong the evaluation time for each set of candidate parameters. The reason for taking the mean instead of a single sample is to reduce the impact of random noise in a single pass. Since the energy efficiency samples entering the sample set have already been eliminated by the previous two rounds of dwell time and bus fluctuation, there are very few obvious outliers. Therefore, the mean can be taken directly. If there are still scattered points in individual working condition samples, the truncated mean or median can be used to further enhance robustness.

[0066] The two energy efficiency averages characterize the efficiency in cold and hot states, respectively. However, their respective weights depend on how much time the equipment spends in upward and downward traversal during actual operation. This necessitates the introduction of cumulative dwell times 'a' and 'b'. Under a pre-defined target operating condition, the duration of time segments where the current falls within a certain current reproducibility interval and is in the upward traversal phase is accumulated segment by segment to obtain the cumulative upward dwell time 'a' for that interval. Similarly, the duration of time segments where the current falls within the same interval and is in the downward traversal phase is accumulated segment by segment to obtain the cumulative downward dwell time 'b'. The cumulative dwell time is obtained by subtracting the sampling time scales and then summing them up, without the need for an additional timer. The target operating condition can be the equipment's rated operating condition, a typical daily load profile measured on-site, or an operating profile specified by the user according to the application scenario, as long as it represents the actual daily load composition of the equipment. To ensure that 'a' and 'b' keep up with the slow drift of the operating condition, they can be statistically analyzed on a daily or shift basis, and the results of adjacent periods can be smoothed to avoid individual abnormal days skewing the weights.

[0067] In other embodiments, when it is necessary to accurately account for the difference in output power between different intervals, a and b can be replaced with the cumulative output energy during the period when the current flows through the interval in the corresponding direction, that is, the output power is integrated over time. When the output voltage is controlled and constant, the cumulative output energy is proportional to the cumulative charge, so the cumulative charge obtained by integrating the output current over time can also be used as an equivalent substitute. It should be noted that all current reproduction intervals in the same optimization must use the same weighting method, either using the cumulative dwell time or the cumulative output energy. Otherwise, the dimensions of the interval optimization objectives of each interval will be inconsistent, and adding them together will be meaningless. Weighting with cumulative dwell time and weighting with cumulative output energy are both about assigning weights to two thermal states according to the actual working conditions. The difference between the two becomes more negligible as the interval is narrower, and the embodiments of this application do not impose any restrictions on this.

[0068] There is another scenario that needs clarification: a certain current recurrence interval is almost traversed only in one direction under the target operating condition, meaning that either a or b is close to zero, or even the weaker direction cannot accumulate enough samples to calculate an average energy efficiency value for a long time. In this case, on the one hand, the interval optimization objective naturally tends to be determined by the dominant direction, and the weaker direction contributes very little to the overall optimization objective due to its extremely small weight; on the other hand, for weak directions where the energy efficiency average cannot be calculated due to insufficient samples, the predicted value of the energy efficiency average in that direction can be used by a machine learning optimization model to replace it in the threshold judgment, or when the number of samples in the weak direction is consistently lower than the lower limit of the sample number, the threshold constraint on that direction can be temporarily not imposed until enough samples are accumulated before including it, thus avoiding the optimization being blocked by a lack of samples while retaining the fallback requirement for that direction.

[0069] Figure 4This is a schematic diagram illustrating the data flow for the energy efficiency sample collection and the synthesis of the interval optimization objective and the overall optimization objective, as provided in the embodiments of this application. Figure 4 As shown, energy efficiency samples are obtained from the ratio of output power to input power. These energy efficiency samples are merged into the uplink and downlink sample sets according to their direction. The average energy efficiency of the uplink and downlink sample sets is calculated separately, and then combined with the cumulative dwell time 'a' and 'b' under the target operating condition. The target synthesis module 44 multiplies the average energy efficiency of the uplink sample set by 'a' and the average energy efficiency of the downlink sample set by 'b', and adds the two together as the interval optimization target for this current reproduction interval. Then, the interval optimization targets of all current reproduction intervals are summed to obtain the overall optimization target. The interval optimization target is constructed in this way because the energy efficiency of the two thermal states is weighted and summed according to their time proportion in the actual operating condition, which is the time-averaged energy efficiency of this interval under the actual hot and cold round trip. After summing the intervals, the overall optimization target represents the comprehensive energy efficiency level of the whole machine under the target operating condition. Maximizing this target means making the whole machine the most energy-efficient under the actual load round trip ratio.

[0070] The following numerical example illustrates this weighting. Suppose that in a certain current reproduction interval, under the target operating condition, the cumulative dwell time *a* of the upward current crossing is 40 minutes, and the cumulative dwell time *b* of the downward current crossing is 20 minutes. The average energy efficiency of the upward sample set is 97.8%, and the average energy efficiency of the downward sample set is 97.2%. The optimization objective for this interval is 97.8% multiplied by 40 minutes plus 97.2% multiplied by 20 minutes. Converting this to time-averaged energy efficiency is equivalent to weighting the two energy efficiency averages by a 40:20 time ratio, resulting in approximately 97.6%, which is the average energy efficiency per unit time under real hot-cold cycle conditions in this interval. If a different device is used, with the downward cycle predominating in the same interval, for example, *a* is 15 minutes and *b* is 45 minutes, the time-averaged energy efficiency calculated from these two energy efficiency averages is approximately 97.35%. The optimization will then bias towards the parameters that perform better on the hotter side. It is evident that the optimization objective of the interval is automatically adjusted according to the direction of the actual working conditions, and does not change with whether a particular sampling happens to sample the upward or downward direction more than the expected. This is precisely the numerical manifestation that the optimization conclusion no longer fluctuates with the sampling period.

[0071] After synthesizing the overall optimization objective, a constraint needs to be added to the optimization process: the average energy efficiency of both the uplink and downlink sample sets within the same interval must not be lower than the energy efficiency threshold of 74. This constraint is to prevent the optimization process from focusing solely on increasing the energy efficiency of the direction with the higher time proportion while suppressing the energy efficiency of the other direction with the lower proportion, thus causing a significant drop in efficiency when the load path occasionally reverses. The energy efficiency threshold of 74 can be set according to the rated operating efficiency specification of the equipment, for example, 95% to 99% of the rated efficiency, typically 97%. If the threshold is set too high, most candidate modulation parameters will be excluded from the constraint, resulting in a narrow optimization space; if the threshold is set too low, it will lose its function of filtering out inferior parameters and will not provide a safety net. When changing equipment models, simply reset the energy efficiency threshold of 74 according to the rated efficiency specification of the new model.

[0072] To maximize the overall optimization objective under constraints, it is necessary to select the best set of modulation parameters from many candidate sets. This step is accomplished by a machine learning optimization model. The machine learning optimization model learns the mapping from a set of candidate modulation parameters to the average energy efficiency of the uplink and downlink sample sets achievable across various current reproduction intervals. This allows it to predict the merits of each candidate modulation parameter without actually testing them on the device, significantly reducing the number of parameter sets requiring actual effectiveness testing. Its specific construction, training, and inference processes will be discussed later. Figure 7 Let's elaborate. First, let's explain its computational cost. One inference operation only involves performing a forward calculation on the candidate parameters, then calculating the overall optimization target and comparing it with the energy efficiency threshold of 74. The computational load is within a few thousand floating-point operations, which can be easily completed within the millisecond-level control cycle of a microcontroller, digital signal processor, or field-programmable gate array without the need for additional acceleration hardware.

[0073] After the target modulation parameters are optimally determined, the parameter distribution module 46 writes the target modulation parameters into the modulator 12 of the power conversion circuit 10. The modulator 12 then generates a new drive signal to drive the power switch 11 in the power conversion circuit 10 to switch on and off according to the target modulation parameters. Thus, from acquiring timing data and separating energy efficiency samples by direction to synthesizing the optimization target and distributing the optimization, a complete adaptive optimization closed loop is formed. The device continuously adjusts the modulation parameters towards a more energy-efficient direction under real-world operating conditions while maintaining normal power supply. Furthermore, the target modulation parameters are written to the modulator 12 using a gradual and smooth transition method to avoid disturbances in the output voltage or output current caused by sudden parameter changes.

[0074] The following section, based on the core solution, details the further design aspects. First, there's the precise determination of the crossing marker and the removal of invalid samples. Regarding the entry interval mentioned earlier, clear criteria must be provided at the boundaries. When the output current changes from below the lower boundary of a certain current reproduction interval to falling into that interval, this entry is marked as an upward crossing; when the output current changes from above the upper boundary of that interval to falling into it, this entry is marked as a downward crossing. Combined with... Figure 3 It can be seen that when the output current 71 crosses the lower boundary of the current reproduction interval from bottom to top, it is considered an upward crossing, and when it crosses the upper boundary from top to bottom, it is considered a downward crossing. Combined with the boundary hysteresis band mentioned earlier, each entry can be cleanly attributed to a single direction.

[0075] Not every energy efficiency sample obtained during a current pass is reliable, therefore two additional discard rules are necessary. First, if the number of consecutive sampling points where the output current falls within the current reproduction interval is less than the lower limit of the sampling point count, it indicates that the current merely glides through the interval, with a very short dwell time and insufficient sampling points reflecting stable efficiency; such energy efficiency samples should be discarded. The lower limit of the sampling point count can be 3 to 8 points, typically 5. Too few points won't effectively eliminate glimpsed samples, while too many points will mistakenly discard many effective fast-pass samples. Second, if the difference between the maximum and minimum values ​​of the input voltage within this continuous sampling period exceeds the upper limit of input voltage fluctuation, it indicates that the input bus itself is fluctuating significantly during the sampling period. The calculated energy efficiency sample in this case is influenced by bus disturbances and should also be discarded. The upper limit for input voltage fluctuation can be set to 2% to 8% of the nominal input voltage value, typically 5% of the nominal value. Setting it too low will cause normal bus ripple to be treated as abnormal and too many samples will be discarded, while setting it too high will miss bus disturbances that could actually contaminate the samples. After these two rounds of rejection, the samples entering the sample set are clean samples that have stayed long enough and whose buses are stable enough.

[0076] Secondly, the quality of the output voltage is also incorporated into the optimization constraints, divided into output voltage ripple constraints and load step constraints. These, along with the aforementioned energy efficiency threshold 74, constitute the optimization constraints, preventing the sacrifice of power quality for energy saving. The output voltage ripple constraint requires that the peak-to-peak value of the output voltage in the time series data corresponding to the energy efficiency sample does not exceed the ripple upper limit. The ripple upper limit can be set from 0.5% to 2% of the nominal output voltage value, typically 1% of the nominal value. If it is too strict, some feasible parameters with ripple within the specification will be misjudged as exceeding the limit; if it is too lenient, the ripple control will be lost. Any candidate modulation parameter whose ripple exceeds the ripple upper limit for the corresponding time period will not be selected, even if the energy efficiency is high.

[0077] The load step constraint targets the dynamic response when the load changes abruptly. Figure 10This diagram illustrates the output voltage overshoot amplitude and recovery time under a load step event, as provided in an embodiment of this application. During the effective period of the candidate modulation parameters, a load step event is determined to have occurred when the change in output current within a set time exceeds the step threshold. The set time can be from 0.1 milliseconds to 2 milliseconds, typically 0.5 milliseconds. If the set time is too short, even a normal rapid load adjustment may be counted as a step; if it is too long, a true step will be smoothed out. The step threshold can be from 20% to 50% of the rated output current, typically 30%. If the step threshold is too small, ordinary load fluctuations may be misjudged as steps; if it is too large, true step events will be missed.

[0078] like Figure 10 As shown, timing data of the output voltage 80 is extracted after a load step event. The maximum deviation of the output voltage 80 from the output voltage setpoint is taken as the overshoot amplitude, and the time taken for the output voltage 80 to return to the allowable deviation band from the step event is taken as the recovery time. The half-width of the allowable deviation band can be 0.5% to 2% of the output voltage setpoint, typically 1% of the setpoint. If the deviation band is too narrow, the recovery time will be calculated as artificially long; if it is too wide, the recovery time will be artificially short.

[0079] The load step constraint requires that the overshoot amplitude corresponding to a set number of load step events should not exceed the overshoot limit, and the recovery time should not exceed the recovery time limit. The set number can be 5 to 30 times, typically 10 times. Too few times, and a single anomaly is enough to negate a set of parameters; too many times, and it will prolong the time required for dynamic constraint judgment. The overshoot limit can be 1% to 5% of the output voltage setting, typically 3% of the setting; the recovery time limit can be 50 microseconds to 500 microseconds, typically 200 microseconds, depending on the load's tolerance for voltage drop and recovery speed. Too strict a limit will mistakenly invalidate parameters that are slightly slower but still meet the load requirements; too lenient a limit will lose control over the dynamic response. The reason for accumulating a number of step events before judgment, rather than drawing conclusions based on a single step, is that the overshoot and recovery of a single step are random due to the influence of the operating conditions at the time. Accumulating multiple events can filter out occasional extreme values, making the dynamic constraint judgment more stable. Candidate modulation parameters that do not meet the load step constraint are not selected, thus ensuring that the optimal parameters maintain the bottom line of dynamic response while saving power.

[0080] The following details the process of building and training a machine learning optimization model, as well as the process of determining its parameters. Figure 7This is a schematic diagram of the data flow for training the machine learning optimization model and determining the target modulation parameters provided in the embodiments of this application. Let's look at the training first. The training samples come from candidate modulation parameters that have actually taken effect during device operation. For each set of parameters that have taken effect, an uplink sample set and a downlink sample set are left in each current reproduction interval, which also leaves two actually calculated average energy efficiency values.

[0081] During training, the values ​​of candidate modulation parameters that have already taken effect and the corresponding current recurrence intervals are used as input features. The average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set are used as the output values ​​of the model. The two average energy efficiency values ​​actually calculated from the uplink and downlink sample sets are used as labels to perform supervised training on the machine learning optimization model. Figure 7 As shown, the two calculated average energy efficiency values ​​are used as labels and fed into the model. The model gradually learns to predict these two average energy efficiency values ​​from the parameters and interval labels. During training, the mean square error between the model's predicted average energy efficiency value and the measured average energy efficiency value used as labels is used as the loss. Gradient descent or the training algorithm built into the selected model is used to reduce the loss. Training stops when the loss no longer decreases significantly on the reserved validation samples or reaches the set iteration limit.

[0082] In the input features, the candidate modulation parameters include switching frequency, duty cycle, dead time, and number of phases, which are given directly as normalized values. The current recurrence interval can be identified using interval number encoding or one-hot encoding, allowing the model to distinguish different intervals. The machine learning optimization model can use existing gradient boosting trees, Gaussian process regression, or shallow neural networks. These models are good at fitting the mapping from low-dimensional input to continuous output when the sample size is small. Taking gradient boosting trees as an example, the model size ranges from tens to hundreds of trees, and the model size of shallow neural networks ranges from hundreds to thousands of weights, all within the storage and computing power of the controller 20. These small-to-medium-scale existing models are chosen instead of larger networks because the input dimension is low, and the samples come from the accumulated data of a single device, so the quantity is not large. A model that is too large would be prone to overfitting.

[0083] The following explains the timing of training and how to handle insufficient samples. When the device is powered on for the first time and has not yet accumulated enough effective data, the machine learning optimization model cannot be trained. At this time, it enters the cold start mode. Instead of using the model to predict, a small number of candidate modulation parameters are selected in the parameter space according to the preset sparse grid and are actually applied in sequence. Their average energy efficiency is directly collected and calculated. This maintains the availability of optimization and accumulates initial training samples for the model.

[0084] Upon initial power-up, no prior target modulation parameters are available for rollback. Therefore, a set of conservative parameters set at the factory is used as the initial safe rollback parameters. The sparse grid probing during the cold start phase is also protected by the stress envelope of subsequent probing. If the limit is exceeded, the system reverts to this set of factory parameters, ensuring that the device does not enter an over-limit state during the cold start. Once the accumulated samples reach a set scale, the first offline training is completed and the system is put into use. This set scale is generally several hundred effective records, such as 200 to 1000 sets. Too few sets will result in underfitting and inaccurate predictions, while too many sets will prolong the cold start. Subsequently, for every additional set number of effective samples, such as every 50 to 200 new samples, incremental training is triggered to update the model, ensuring that the model keeps up with the slow changes in device status and operating conditions, preventing inaccurate predictions due to aging or environmental changes. Training and updates are performed during the idle periods of controller 20, without consuming the computing power of real-time control.

[0085] The trained model is used to determine the parameters, i.e. Figure 7 The right half describes the reasoning and parameter selection process. Candidate modulation parameters that have not yet taken effect are input one by one into the trained machine learning optimization model to obtain the predicted values ​​of the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set for each current reproduction interval. A coarse screening is performed first; any candidate modulation parameter whose predicted average energy efficiency is lower than the energy efficiency threshold of 74 is directly eliminated and no longer considered.

[0086] For the remaining candidate modulation parameters, they are sorted from largest to smallest according to the overall optimization objective calculated from the two predicted energy efficiency averages. A set number of the top candidates are selected and implemented sequentially. This set number is generally between 3 and 10. Too few might miss the true optimal value when the model prediction is biased, while too many would increase the number of actual parameter trials on the device. The reason for implementing the parameters after prediction is that the predicted values ​​are only model estimates, and the final result should be based on the actual implementation. After these selected candidate parameters are implemented sequentially, the two energy efficiency averages and the overall optimization objective are recalculated from the uplink and downlink sample sets obtained during the implementation period. Finally, the group of candidate modulation parameters with the largest overall optimization objective and whose two energy efficiency averages are not lower than the energy efficiency threshold of 74 is determined as the target modulation parameters. Model prediction first filters out a large number of obviously undesirable parameters, and only a few promising parameters are verified through actual implementation. This preserves the accuracy of the optimization search while keeping the number of parameter trials on the device very low.

[0087] The following explains how multiple sets of candidate modulation parameters are generated, and the optimization process with round-by-round convergence. The candidate modulation parameters include at least two of the following: switching frequency, duty cycle, dead time, and number of phases. In other words, it is not necessary to adjust all four parameters; only the two or three with the greatest impact can be adjusted. Figure 11This is a schematic diagram illustrating the generation of candidate modulation parameter neighborhoods and the setting of step size shrinkage, as provided in an embodiment of this application. Figure 11 As shown, the multiple sets of candidate modulation parameters are not randomly scattered throughout the parameter space, but rather, with the currently active target modulation parameter as the center, several candidates are expanded to both sides on the respective value axes of switching frequency, duty cycle, dead time, and number of phases according to a set step size. Figure 11 Solid dots represent the currently active target modulation parameter, while hollow dots represent candidate modulation parameters obtained by expanding outwards from it. Local expansion centered on the current optimal parameter is used because the optimal parameters in adjacent rounds are usually not far apart, making searching near the current optimal parameter much more efficient than a blind search across the entire space.

[0088] The step size determines the density of the search in each round. For continuously variable parameters such as switching frequency, duty cycle, dead time, and mode switching threshold, the step size can be set to 2% to 10% of the current parameter value, typically 5%. If the step size is too large, it is easy to skip the optimal point in one step and oscillate back and forth around the optimal point; if the step size is too small, the movement in each round is too small, and the convergence is too slow. The number of phases is a discrete parameter that can only take integer values, so percentage step sizes are not applicable. It is treated separately, with one or more phases expanded outwards from the current number of phases as candidates, that is, one phase is used as its minimum step size. The optimization is carried out round by round, and in each round, the candidates are expanded again around the target modulation parameter selected in the previous round, and a new target modulation parameter is selected.

[0089] When the increase in the total optimization target of the newly determined target modulation parameters relative to the total optimization target of the previous round is less than the convergence threshold, it indicates that the search is close to optimal. Further large-step searches are no longer meaningful, so the set step size is reduced proportionally before entering the next round for a more refined search. The convergence threshold can be set to 0.5% to 2% of the total optimization target of the previous round, typically 1%. A threshold that is too large will prematurely stop refinement, while a threshold that is too small will result in wasted exploration even after near convergence. The set ratio can be set to 0.3 to 0.7, typically 0.5. A ratio that is too small will cause the step size to drop too drastically, potentially locking the search near a local optimum before sufficient exploration is achieved; a ratio that is too large will result in too slow step size contraction, increasing the number of refinement rounds. Figure 11 As shown in the second row, the candidate points are denser after the convergence threshold is triggered than in the first row, which is the result of scaling down the step size proportionally.

[0090] In addition to the aforementioned parameters, candidate modulation parameters may also include a mode switching threshold. The mode switching threshold is the output current value upon which the power conversion circuit 10 switches between continuous conduction mode and intermittent conduction mode. If the output current is higher than this threshold, it operates in continuous conduction mode; if it is lower, it switches to intermittent conduction mode. Switching to intermittent conduction mode under light load can reduce switching losses, but the location of the switching point directly affects the energy efficiency at the boundary between light and medium loads. Therefore, it is also included as a parameter to be optimized. The candidate value for the mode switching threshold can be selected within the range of 10% to 30% of the rated output current, typically 20%. If the threshold is too high, the section that should enter intermittent conduction mode for energy saving will remain in continuous conduction mode, resulting in insufficient energy efficiency under light loads. If it is too low, it will switch prematurely under medium loads, causing efficiency fluctuations near the mode switching point.

[0091] The following explains the safety issues of testing candidate parameters on energized equipment. To obtain energy efficiency samples through online optimization, candidate modulation parameters must actually take effect on energized equipment for a period of time. However, whether the candidate parameters will cause current or temperature rise to exceed limits is unknown before they take effect, which is why online optimization is often simply shut down in the field. This application uses a test stress envelope to define a safe testing boundary. Figure 8 This is a schematic diagram illustrating the timing of the activation of the trial stress envelope and candidate modulation parameters provided in this application embodiment. Before activating a set of candidate modulation parameters, the maximum value of the input current 77 actually collected within the most recent statistical period is taken as the current envelope value 75, and the maximum value of the rate of change of chassis temperature with respect to time 78 actually collected within the most recent statistical period is taken as the temperature rise rate envelope value 76. The current envelope value 75 and the temperature rise rate envelope value 76 together constitute the trial stress envelope.

[0092] Chassis temperature is a very slow-changing quantity. Its rate of change over time (78%) is not obtained by directly subtracting two adjacent sampling points. Instead, the net temperature change is taken over a time window of seconds, divided by the window length, and then smoothed to prevent the point-by-point difference from being overwhelmed by the quantization noise of temperature sampling. The statistical duration can range from 1 minute to 30 minutes, with 10 minutes being typical. If the statistical duration is too short, the envelope is determined only by the most recent few minutes, which is too conservative. If it is too long, old stresses from long ago will be included, failing to keep up with changes in operating conditions.

[0093] Using these two quantities as the upper limit for testing is reliable because the current envelope value of 75 and the temperature rise rate envelope value of 76 are peak values ​​that the equipment has recently experienced without triggering any protection. In other words, the hardware has proven itself capable of withstanding these levels. Ensuring that candidate parameters do not exceed these two envelopes strictly limits the testing to the stress range that the equipment has recently endured, preventing it from entering unprecedented states. Therefore, online optimization no longer needs to rely on disabling functions for safety. Besides limiting the stress amplitude, the timing of the testing is also crucial. For example... Figure 8 As shown, the candidate modulation parameters are only activated within a set waiting period after the rate of change of the output current (71) with respect to time changes from negative to zero. This moment is chosen because the change of the output current rate of change from negative to zero signifies that a downlink cycle has just ended and the load has fallen to a relatively low position. At this time, the device has the largest temperature rise margin, ensuring maximum thermal safety even if the tested parameters are aggressive. The set waiting period can be set from 5 milliseconds to 50 milliseconds, typically 20 milliseconds. If the waiting period is too short, the test will be initiated immediately after the downlink edge has passed and before the thermal inertia has stabilized; if it is too long, the favorable window of low load will be missed.

[0094] During the period when the candidate modulation parameters are in effect, the process is not idle. Instead, the real-time input current 77 is continuously compared with the current envelope value 75, and the real-time chassis temperature change rate over time 78 is compared with the temperature rise rate envelope value 76. If the input current 77 exceeds the current envelope value 75, or the chassis temperature change rate over time 78 exceeds the temperature rise rate envelope value 76, the attempt is immediately terminated. The target modulation parameters that are currently in effect are rewritten into the modulator 12, allowing the device to revert to known safe parameters. Simultaneously, the energy efficiency samples generated during this period of candidate modulation parameters are discarded, as these samples were collected during an abnormal process that was interrupted and are unusable. Figure 8 As shown, when the input current 77 exceeds the current envelope value 75, it is judged as exceeding the limit, triggering the rewriting of the target modulation parameters into the modulator 12 and discarding the energy efficiency sample. However, the rate of change of the chassis temperature with respect to time 78 in the figure never exceeds the temperature rise rate envelope value 76. In addition, this set of candidate modulation parameters is marked as prohibited from being tested in the corresponding current reproduction range, and this set of parameters will not be tested in this range again to avoid repeatedly touching the same dangerous point.

[0095] Finally, the issue of the intertwining of aging and temperature during long-term operation is explained. After several months of equipment operation, component aging will cause an overall decrease in energy efficiency, while increased temperature will also cause a decrease in energy efficiency. The two show similar performance in energy efficiency samples. If optimization is misjudged as a deterioration in parameters, it will lead to continuous adjustments in the wrong direction. This application uses cold anchor points to measure aging separately. Figure 9 This is a schematic diagram illustrating the extraction of cold anchor sequence and aging component provided in an embodiment of this application.

[0096] For any given uplink pass, if prior to this uplink pass, the output current has been continuously below the lower boundary of the lowest current recurrence interval for a duration not less than the lower limit of the resting time, and the difference between the chassis temperature and the ambient temperature is not greater than the upper limit of the temperature difference, then the energy efficiency sample corresponding to this uplink pass is taken as the cold anchor point sample. The ambient temperature can be obtained from the air inlet temperature monitoring system already installed in the system. Server power supply racks typically have air inlet temperature sensors, and reading them does not require adding any new components to this application. In situations where there is no independent air inlet temperature sensor, the ambient temperature can be approximated by the chassis temperature when the equipment has been under minimum load for an extended period and the chassis temperature has stopped decreasing.

[0097] These two conditions together lock in a baseline thermal state where the device is fully cooled and returns to near ambient temperature. The lower limit of the resting time can be 30 to 300 seconds, typically 120 seconds. It needs to be longer than the upper limit of the device's thermal time constant to ensure that the device is truly cooled. The upper limit of the temperature difference can be 3 to 10 degrees Celsius, typically 5 degrees Celsius. If it is too small, there will be too few cold anchor points that meet the conditions, and if it is too large, samples that have not been cooled completely will be mixed in.

[0098] Arranging the cold anchor point samples taken within the same current reproduction interval and under the same set of modulation parameters in chronological order constitutes the cold anchor point sequence 79. For example... Figure 9 As shown, the samples in the cold anchor sequence 79 decrease slowly over time. The median of the cold anchor sequence 79 within the most recent aging statistical window is subtracted from the median of the cold anchor sequence 79 within the first aging statistical window; this difference is taken as the aging component. Since energy efficiency decreases slowly with aging, and the most recent median is lower than the first median, the aging component is a negative value. Its absolute value represents the energy efficiency lost due to aging during this period.

[0099] The aging statistical window can be 3 to 30 days, typically 7 days. It is much longer than the thermal time constant to filter out temperature fluctuations, and shorter than the aging time scale to keep up with the aging process.

[0100] The median is used instead of the average to avoid being biased by a few outliers. The reason this difference can separate aging is fundamentally because the effects of temperature and aging on losses are fundamentally different. The effect of temperature is reversible; once the device cools down, the efficiency returns to its original value. However, the effect of aging is irreversible; the efficiency will not return to its original value. The cold-state anchor samples are all taken from the same fully cooled thermal state. Since the thermal state is consistent, the differences between them over time cannot come from temperature but only from irreversible aging. Therefore, the aging component essentially contains no temperature component.

[0101] After measuring the aging component, it is used to correct the energy efficiency samples. Before calculating the average energy efficiency of the uplink and downlink sample sets, the aging component is subtracted from each energy efficiency sample within each current reproduction interval to obtain the corrected energy efficiency sample. Then, the two energy efficiency averages are calculated from the corrected energy efficiency samples. Since the aging component is negative, subtracting it from each sample is equivalent to making up for the energy efficiency lost due to aging. In this way, the overall downward shift caused by aging is deducted, and the remaining energy efficiency difference only reflects the quality of the modulation parameters themselves. The optimization method is only responsible for the modulation parameters and will not misinterpret aging as a deterioration of a certain set of parameters and adjust in the wrong direction.

[0102] The cold anchor point only appears in the lowest current recurrence interval. Therefore, the aging component is a device-level reference quantity measured from the light-load operating point. It is applied to each current recurrence interval in the same way, based on the approximation that device aging mainly manifests as overall characteristic degradation, and the energy efficiency reduction caused at each load point is roughly equivalent. When more precision is needed, the aging component can be scaled proportionally to the square of the representative current of each interval before being subtracted to account for the portion of aging loss that increases with increasing current. Subtracting the aging component before calculating the mean is to ensure that the subtraction treats the uplink and downlink sample sets equally, without changing the relative levels between the two directions.

[0103] Here, it's important to explain the coordination between the cold anchor point sequence 79 and the optimization process. The cold anchor point sequence 79 accumulates according to the same set of modulation parameters. Only when the optimization converges and the target modulation parameters remain stable for a long period will the cold anchor points under the same set of parameters accumulate one after another, forming a sufficiently long sequence. Therefore, aging monitoring actually occurs in the long-term operation phase after optimization convergence. Each time the optimization converges and locks onto a set of target modulation parameters, a segment of cold anchor points at that time is re-anchored as the first aging statistical window, serving as the energy efficiency baseline for this set of parameters.

[0104] When the absolute value of the aging component exceeds the aging threshold, it indicates that device aging has accumulated to a non-negligible level. At this point, multiple sets of candidate modulation parameters are regenerated and the optimization process is repeated to adapt the parameters to the aged device. After re-optimizing and determining the new parameters, the cold anchor sequence 79 and the first aging statistical window are reset together and recalculated from the baseline of the new parameters. Therefore, there will be no situation where the old baseline is immediately used to judge the device as exceeding the limit immediately after aging is deducted, repeatedly triggering re-optimization. During the period from changing parameters to when the new sequence is sufficient, no aging correction is performed. This only affects the correction accuracy during this period and does not affect the optimization itself. The aging threshold can be set to 0.5% to 2% of the rated efficiency, typically 1% of the rated efficiency. If the threshold is set too small, global re-optimization will be frequently triggered due to even a small amount of aging. If it is set too large, aging will be allowed to drag the parameters away from the optimal state for too long. Since the two conditions of the lower limit of the resting time and the upper limit of the temperature difference will be repeatedly met during long-term operation, the cold anchor point sequence 79 will be continuously replenished. When there are fewer anchor points at a certain time, the update of the aging component will only be slower, and it will not cause false triggering.

[0105] The above explains each step and its design considerations from a methodological perspective. The following explains each functional module from a system perspective, and they correspond one-to-one with the previous methodological steps. Figure 6 A block diagram of a power parameter adaptive optimization system based on time-series data provided in an embodiment of this application. (Refer to...) Figure 6 The power parameter adaptive optimization system 40 includes a timing acquisition module 41, a crossing marker module 42, and a sample collection module 43. The power parameter adaptive optimization system 40 also includes a target synthesis module 44, an optimization decision module 45, and a parameter distribution module 46. The timing acquisition module 41 acquires the voltage and current timing data of the input and output of the power conversion circuit 10, corresponding to the previous step of acquiring timing data. The crossing marker module 42 divides the output current into multiple current reproduction intervals according to the range of output current values, and marks each entry into the interval as an upward or downward crossing based on the sign of the rate of change of the output current with respect to time; this process involves dividing the interval and determining the direction.

[0106] The sample collection module 43 obtains energy efficiency samples for each crossover based on the ratio of output power to input power during the effective period of the candidate modulation parameters. It establishes uplink and downlink sample sets according to the current reproduction interval, and assigns the energy efficiency samples obtained from the uplink and downlink crossovers to the uplink and downlink sample sets respectively. This process involves sample acquisition and collection. The target synthesis module 44 calculates the average energy efficiency of the uplink and downlink sample sets within the same current reproduction interval. Under the target operating condition, it calculates the cumulative dwell times 'a' and 'b' for uplink and downlink. The sum of the average energy efficiency of the uplink sample set multiplied by 'a' and the average energy efficiency of the downlink sample set multiplied by 'b' is used as the interval optimization target. The sum of the interval optimization targets for all current reproduction intervals is used as the overall optimization target, corresponding to the target synthesis step.

[0107] The optimization decision module 45 is located within the controller 20 of the switching power supply. The controller 20 can be a microcontroller, digital signal processor, or field-programmable gate array. The optimization decision module 45 uses two energy efficiency averages not lower than the energy efficiency threshold 74 as constraints, and maximizes the overall optimization objective as the optimization goal. A machine learning optimization model determines the target modulation parameter from multiple candidate modulation parameters, thus undertaking the optimization decision. The parameter distribution module 46 includes the modulator 12 and the drive circuit for the power switching transistor 11 in the power conversion circuit 10. It writes the target modulation parameter into the modulator 12 and drives the power switching transistor 11 in the power conversion circuit 10 according to the target modulation parameter, thus implementing parameter distribution. The specific processing performed by each module is the same as in the previous section and will not be repeated here. This system does not require junction temperature and core temperature sensors, nor does it need to build thermal models for each unit; it can run directly on the controller 20.

[0108] It should be noted that the design steps described above are interconnected and work collaboratively within a complete optimization process, and their order cannot be reversed. In terms of sequence, after each batch of energy efficiency samples is collected, the aging component is first subtracted using the method described in the previous section on aging to obtain the corrected energy efficiency samples. Then, the energy efficiency average of the uplink sample set and the downlink sample set are statistically analyzed using the corrected samples. In other words, aging is deducted before the average is calculated, ensuring that neither load-related nor aging components are included in the two energy efficiency averages. When testing candidate parameters under load, the stress envelope of the test is monitored throughout the entire effective period. If the threshold is crossed midway, the sample is immediately discarded, and the parameter set is marked as prohibited from testing in that interval. The discarded sample will not enter the sample set and will not contaminate the energy efficiency average. When the aging component crosses the aging threshold, triggering a global re-optimization, the newly generated multiple sets of candidate modulation parameters still undergo the effective process with stress envelope monitoring and are subject to the combined constraints of the energy efficiency threshold 74, output voltage ripple constraints, and load step constraints. Several design elements each manage their own section and interlock with each other to ensure that the optimal parameters are both energy-saving and safe, and are not skewed by the load path or dragged down by aging.

[0109] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Those skilled in the art should understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features, and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A power supply parameter adaptive optimization method based on time-series data, applied to a switching power supply including a power conversion circuit, characterized in that, include: The timing data of the input voltage, input current, output voltage and output current of the switching power supply are collected. Multiple current reproduction intervals are divided according to the value range of the output current. Based on the sign of the rate of change of the output current with respect to time, each process of entering the current reproduction interval is marked as an upward crossing or a downward crossing. During the period when a set of candidate modulation parameters are in effect, a segment of the time-series data is extracted for each crossover, and an energy efficiency sample is obtained by the ratio of output power to input power. The energy efficiency samples obtained from the uplink and downlink crossovers are respectively assigned to the uplink sample set and the downlink sample set according to the current reproduction interval. Within the same current reproduction interval, the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set are statistically analyzed, and the cumulative dwell time of the current uplink crossing and downlink crossing under the target operating condition is statistically analyzed, which are a and b respectively. With the constraint that both of the energy efficiency average values ​​are not lower than the energy efficiency threshold, the sum of the energy efficiency average value of the uplink sample set multiplied by a and the energy efficiency average value of the downlink sample set multiplied by b is used as the interval optimization objective. The sum of the interval optimization objectives of all current reproduction intervals is used as the total optimization objective. The optimization objective is maximized as the optimization objective. The target modulation parameter is determined from multiple candidate modulation parameters by the machine learning optimization model. The target modulation parameters are written into the modulator of the power conversion circuit, and the power switching transistor in the power conversion circuit is driven according to the target modulation parameters.

2. The power parameter adaptive optimization method based on time-series data according to claim 1, characterized in that, Marking each entry into the current reproduction interval as an upward crossing or a downward crossing includes marking an upward crossing when the output current changes from below the lower boundary of the current reproduction interval to falling into the current reproduction interval, and marking a downward crossing when the output current changes from above the upper boundary of the current reproduction interval to falling into the current reproduction interval. For any given crossover, if the number of consecutive sampling points where the output current falls within the current reproduction interval is less than the lower limit of the number of sampling points, or if the difference between the maximum and minimum values ​​of the input voltage within the consecutive sampling points is greater than the upper limit of the input voltage fluctuation, the energy efficiency sample corresponding to the crossover is discarded.

3. The power supply parameter adaptive optimization method based on time-series data according to claim 1, characterized in that, The constraints also include output voltage ripple constraints and load step constraints. The output voltage ripple constraint is that the peak-to-peak value of the output voltage in a time series data corresponding to the energy efficiency sample is not greater than the upper limit of ripple. During the effective period of the candidate modulation parameters, when the change in output current within a set time period is greater than the step threshold, a load step event is determined to have occurred. The timing data of the output voltage after the load step event is extracted, and the maximum amplitude of the output voltage deviation from the output voltage set value is taken as the overshoot amplitude. The time taken for the output voltage to return to the allowable deviation band from the load step event to the output voltage set value is taken as the recovery time. The load step constraint is that the overshoot amplitude corresponding to the cumulative set number of load step events is not greater than the overshoot upper limit and the recovery time is not greater than the recovery time upper limit.

4. The power supply parameter adaptive optimization method based on time-series data according to claim 1, characterized in that, The training process of the machine learning optimization model includes: taking the values ​​of the candidate modulation parameters that have already taken effect and the corresponding current reproduction interval identifiers as inputs, taking the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set as outputs, and taking the average energy efficiency calculated from the uplink sample set and the downlink sample set as the average energy efficiency labels of the uplink sample set and the downlink sample set, respectively, to perform supervised training on the machine learning optimization model; Determining the target modulation parameter from multiple sets of candidate modulation parameters includes: inputting candidate modulation parameters that have not yet been activated into the trained machine learning optimization model to obtain the predicted values ​​of the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set for each current reproduction interval; eliminating candidate modulation parameters whose predicted value of any average energy efficiency is lower than the energy efficiency threshold; selecting a set number of candidate modulation parameters to be activated sequentially according to the total optimization target calculated from the two predicted average energy efficiency values, and recalculating the two average energy efficiency values ​​and the total optimization target from the uplink sample set and the downlink sample set actually obtained during the activation period; and determining the candidate modulation parameter with the largest total optimization target and whose two average energy efficiency values ​​are not lower than the energy efficiency threshold as the target modulation parameter.

5. The power supply parameter adaptive optimization method based on time-series data according to claim 1, characterized in that, The candidate modulation parameters include at least two of the following: switching frequency, duty cycle, dead time, and number of phases; The multiple sets of candidate modulation parameters are generated from a neighborhood centered on the currently effective target modulation parameter. The neighborhood expands to both sides on the respective value axes of each parameter included in the candidate modulation parameters with a set step size. After one round of optimization, when the increase in the total optimization target of the newly determined target modulation parameters relative to the total optimization target of the target modulation parameters determined in the previous round is less than the convergence threshold, the set step size is reduced by a set ratio and the next round of optimization is entered. The candidate modulation parameters also include a mode switching threshold, which is the output current value on which the power conversion circuit switches between continuous conduction mode and intermittent conduction mode.

6. The power supply parameter adaptive optimization method based on time-series data according to claim 1, characterized in that, Before activating a set of candidate modulation parameters, the maximum value of the input current actually collected within the most recent statistical period is taken as the current envelope value, and the maximum value of the rate of change of the chassis temperature with respect to time actually collected within the most recent statistical period is taken as the temperature rise rate envelope value. The current envelope value and the temperature rise rate envelope value constitute the trial stress envelope. And within a set waiting period after the rate of change of the output current with respect to time changes from negative to zero, the set of candidate modulation parameters are made effective.

7. The power supply parameter adaptive optimization method based on time-series data according to claim 6, characterized in that, During the period when the set of candidate modulation parameters are in effect, the real-time acquired input current is continuously compared with the current envelope value, and the real-time acquired rate of change of chassis temperature with respect to time is compared with the temperature rise rate envelope value. When the input current is greater than the current envelope value, or the rate of change of chassis temperature with respect to time is greater than the temperature rise rate envelope value, the target modulation parameters that are currently effective are rewritten into the modulator, the energy efficiency samples generated by the set of candidate modulation parameters during this effective period are discarded, and the set of candidate modulation parameters is marked as prohibited from being tested in the corresponding current reproduction interval.

8. The power supply parameter adaptive optimization method based on time-series data according to claim 1, characterized in that, For any uplink pass, if the output current is continuously lower than the lower boundary of the lowest current reproduction interval before the uplink pass and the duration is not less than the lower limit of the resting time, and the difference between the chassis temperature and the ambient temperature is not greater than the upper limit of the temperature difference, the energy efficiency sample corresponding to the uplink pass is taken as the cold anchor point sample. The cold anchor point samples within the same current reproduction interval and under the same set of modulation parameters are arranged chronologically to form a cold anchor point sequence.

9. The power supply parameter adaptive optimization method based on time-series data according to claim 8, characterized in that, The difference between the median of the cold anchor sequence in the most recent aging statistical window and the median of the cold anchor sequence in the first aging statistical window is taken as the aging component; Before calculating the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set, the difference between each energy efficiency sample in each current reproduction interval and the aging component is used as the corrected energy efficiency sample, and the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set are calculated from the corrected energy efficiency sample. When the absolute value of the aging component is greater than the aging threshold, the multiple sets of candidate modulation parameters are regenerated and optimization is re-executed.

10. A power supply parameter adaptive optimization system based on time-series data, applied to a switching power supply including a power conversion circuit, for implementing the power supply parameter adaptive optimization method based on time-series data as described in claim 1, characterized in that, include: The timing acquisition module is used to acquire timing data of the input voltage, input current, output voltage, and output current of the switching power supply. The crossing marker module is used to divide multiple current reproduction intervals according to the range of output current values, and to mark each entry into the current reproduction interval as an upward crossing or a downward crossing based on the sign of the rate of change of the output current with respect to time. The sample collection module is used to extract a segment of the time-series data for each crossover during the effective period of a set of candidate modulation parameters, obtain energy efficiency samples from the ratio of output power to input power, establish uplink sample sets and downlink sample sets according to the current reproduction interval, and classify the energy efficiency samples obtained from the uplink crossover and the downlink crossover into the uplink sample set and the downlink sample set respectively. The target synthesis module is used to calculate the average energy efficiency of the uplink sample set and the average energy efficiency of the downlink sample set within the same current reproduction interval, and to calculate the cumulative dwell time of the uplink and downlink current crossings, a and b, respectively, under the target operating conditions. The sum of the average energy efficiency of the uplink sample set multiplied by a and the average energy efficiency of the downlink sample set multiplied by b is used as the interval optimization target, and the sum of the interval optimization targets of all current reproduction intervals is used as the overall optimization target. An optimization decision module is set in the controller of the switching power supply. The controller is a microcontroller, a digital signal processor, or a field-programmable gate array. It is used to determine the target modulation parameter from multiple sets of candidate modulation parameters by a machine learning optimization model, with the constraint that the two average energy efficiency values ​​are not lower than the energy efficiency threshold and the optimization objective being maximized. The parameter delivery module includes a modulator and a power switch driver circuit of the power conversion circuit, used to write the target modulation parameters into the modulator and drive the power switch in the power conversion circuit according to the target modulation parameters.

Images