A control method and electronic device for a multiphase power supply

Abstract

This application discloses a control method and electronic device for a multiphase power supply, relating to the field of server power supply technology. The method includes acquiring dynamic response data of the multiphase power supply under a target load condition, using a first pre-trained neural network to obtain an output based on the dynamic response data of the target load condition, ensuring the power supply's dynamic response meets preset standards. Actual load dynamic response indicators such as load current ratio, current change rate, and load stabilization time in real-world business scenarios are collected. A corresponding predicted switching frequency is obtained using a second pre-trained neural network, trained based on load dynamic response indicators and adapted switching frequencies from different business simulation scenarios. Through coordinated control of the output and the predicted switching frequency, the switching frequency of the multiphase power supply is dynamically adjusted, solving the problems of high high-frequency loss under light loads, poor low-frequency response under heavy loads, and voltage instability, achieving a balance between efficiency and stability under different load conditions.

Description

Technical Field

[0001] This application relates to the field of server power supply technology, and in particular to a control method and electronic device for a multiphase power supply. Background Technology

[0002] The development of Artificial Intelligence (AI) technology has led to a surge in demand for high-power chips such as Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs). Multiphase power supplies are widely used because they can provide high current and low ripple voltage. However, AI loads fluctuate greatly, and computing unit loads switch frequently, requiring multiphase power supplies to operate efficiently under different operating conditions and quickly stabilize voltage during sudden load changes to ensure server stability.

[0003] Currently, multiphase power supplies for AI servers collect data through frequency sweep testing using a voltage regulation test tool (VRTT Tool) to pinpoint the "worst-case operating point," and then adjust the load frequency and duty cycle until they meet technical standards. However, this approach has limitations due to the fixed switching frequency (FSF). Under light loads, high-frequency operation increases losses and reduces efficiency; under heavy loads, low-frequency operation struggles to handle sudden changes, affecting voltage stability. This results in power supply systems that cannot meet the high energy efficiency requirements of AI servers, especially in large-scale data centers, significantly increasing energy consumption and costs. Summary of the Invention

[0004] This application provides a control method and electronic device for a multiphase power supply, which at least solves the problems of efficiency and dynamic response contradiction, fixed proportional-integral-derivative parameters, and low efficiency due to manual debugging in related technologies.

[0005] This application provides a control method for a multiphase power supply, comprising: acquiring dynamic response data of the multiphase power supply output at a target load operating point; obtaining an output quantity based on the dynamic response data corresponding to the target load operating point and a first pre-trained neural network; wherein the output quantity is used to adjust the dynamic response data output by the multiphase power supply to meet a preset standard; collecting actual load dynamic response indicators of the multiphase power supply under actual business scenarios; the actual load dynamic response indicators include load current ratio, current change rate, and load stabilization time; obtaining a predicted switching frequency corresponding to the actual load dynamic response indicators based on the actual load dynamic response indicators and a second pre-trained neural network; wherein the second pre-trained neural network is trained based on the simulated load dynamic response indicators and corresponding switching frequencies of the multiphase power supply under different business simulation scenarios; and controlling the multiphase power supply based on the output quantity and the predicted switching frequency.

[0006] This application also provides a control device for a multiphase power supply, including: The acquisition module is used to acquire the dynamic response data output by the multiphase power supply under the target load condition point; The proportional-integral-derivative parameter prediction module is used to obtain the output quantity based on the dynamic response data corresponding to the target load operating point and the first pre-trained neural network; wherein, the output quantity is used to adjust the dynamic response data of the multiphase power supply output to meet the preset standard. The data acquisition module is used to collect actual load dynamic performance indicators of multiphase power supplies in real business scenarios; the actual load dynamic performance indicators include load current ratio, current change rate and load stabilization time. The switching frequency prediction module is used to obtain the predicted switching frequency corresponding to the actual load dynamic response index based on the actual load dynamic response index and the second pre-trained neural network. The second pre-trained neural network is trained based on the simulated load dynamic response index and the corresponding switching frequency of the multi-phase power supply under different business simulation scenarios. The control module is used to control the multiphase power supply based on the output and the predicted switching frequency.

[0007] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described multiphase power supply control method.

[0008] This application also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the above-described multiphase power supply control method.

[0009] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described multiphase power supply control method.

[0010] This application first acquires dynamic response data of a multiphase power supply under target load conditions, identifying the target load conditions that reflect light / heavy load characteristics, thus establishing a load matching foundation for subsequent parameter adjustments. Next, using a first pre-trained neural network, the output is obtained based on the dynamic response data of the target load conditions. This neural network, after training, ensures that the output ensures the power supply's dynamic response meets preset standards, providing fundamental algorithmic support for stable control under different loads. By collecting actual load dynamic indicators such as load current ratio, current change rate, and load stabilization time in real-world business scenarios, and then mapping these to a second pre-trained neural network, the corresponding predicted switching frequency is obtained. This neural network, trained based on load dynamic indicators and adapted switching frequencies from different business simulation scenarios, enables dynamic matching of load characteristics and switching frequency. Simultaneously, the optimized output further ensures voltage stability. Finally, through the coordinated control of the output and the predicted switching frequency, the switching frequency of the multiphase power supply is no longer fixed but dynamically adjusted according to the actual load dynamic indicators. This overcomes the pain point of high high-frequency losses under light loads and solves the problems of poor low-frequency response and voltage instability under heavy loads, achieving a balance between efficiency and stability under different load conditions. Attached Figure Description

[0011] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1A A schematic diagram of the power detection circuit on which the execution of a control method for a multiphase power supply provided in this application depends; Figure 1B This is a schematic diagram of the internal logic of a current sensor. Figure 1C This is a schematic diagram of the internal logic of a trigger. Figure 2 A flowchart illustrating a control method for a multiphase power supply provided in an embodiment of this application; Figure 3 A flowchart illustrating another control method for a multiphase power supply provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of a control device for a multiphase power supply provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0014] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0015] To more clearly illustrate the embodiments of this application, the technical terms used in the embodiments will be briefly introduced below: A multiphase voltage regulator (VR) is a power supply device that uses multiple single-phase power modules connected in parallel to work together to achieve voltage regulation and power supply. It is a key power supply component for the core computing unit of an AI server. Its core principle is to split the total output current into multiple single-phase modules, with each module working alternately according to a specific phase difference (such as 2-phase, 4-phase, 8-phase, etc.). The current ripple is canceled by phase superposition, while reducing the current load of individual modules.

[0016] Fixed Switching Frequency (FSF) refers to a multiphase power supply where the switching frequency of the switching transistors (MOSFETs, etc.) in a single-phase module remains constant and does not adjust with load changes. It is a typical design in proportional-integral-derivative (PID) parameter optimization schemes for multiphase power supplies. Its core features are: the switching frequency is preset (e.g., 300kHz, 500kHz), and during debugging, only PID parameters need to be matched for the fixed frequency, eliminating the need to consider parameter adaptation at different frequencies, thus simplifying the initial development process.

[0017] PID parameters refer to the three key parameters used in a PID controller to adjust control accuracy, response speed, and stability. They correspond to the proportional (P), integral (I), and derivative (D) control loops, respectively, and are the core adjustment objects for dynamic performance optimization of multiphase power supplies in AI servers. Its core working principle is as follows: by comparing the actual output voltage of the multiphase power supply with the target voltage in real time, the P, I, and D loops are used to collaboratively calculate the adjustment amount, controlling the on / off state of the power switching transistors, ultimately stabilizing the output voltage within the target range to meet the power supply requirements of chips such as GPUs and TPUs.

[0018] The Voltage Regulator Test Tool (VRTT Tool) simulates the dynamic load changes of core chips such as CPUs and GPUs, collects power output data under different load conditions, and helps technicians locate the "worst operating point" and optimize parameters.

[0019] In recent years, the rapid development of AI technology has driven an explosive growth in the computing demands of AI servers. As core computing units in AI servers, the power consumption of high-performance GPUs, TPUs, and other chips has continued to rise. To meet the power supply requirements of these high-power chips, multiphase power supplies, with their characteristics of providing high current and low-ripple stable voltage, are widely used in AI server power supply systems. However, AI workloads exhibit significant dynamic fluctuations; that is, during server operation, the load of core computing units frequently switches between light and heavy loads. This poses a severe challenge to the efficiency and dynamic response capabilities of multiphase power supplies, requiring them to maintain high efficiency under different load conditions and respond quickly to sudden load changes to maintain stable output voltage, preventing AI server malfunctions due to power instability.

[0020] Currently, the core of the PID parameter optimization solution for multiphase power supplies in AI servers relies on the VRTT Tool for initial data acquisition. During actual debugging, technicians first use this tool to perform a full-load frequency sweep test on the multiphase power supply system, focusing on capturing the dynamic response data of the power supply output under different load conditions. Based on this data, they pinpoint the "worst-case scenario" where the system performance is weakest. To ensure coverage of various load fluctuation scenarios that may occur during AI server operation, technicians need to repeatedly adjust the load range and test frequency of the frequency sweep test to obtain accurate data to support subsequent parameter optimization. After determining the "worst-case scenario," technicians focus on adjusting the load frequency and duty cycle at that point: based on the dynamic response curve obtained from the frequency sweep test, they gradually adjust the load frequency and duty cycle of the multiphase power supply, while simultaneously monitoring key indicators such as output voltage stability and ripple coefficient in real time using testing equipment, until the dynamic performance of the "worst-case scenario" meets the preset technical specifications (Spec), thereby compensating for the performance shortcomings at that point and ensuring the basic operational stability of the server.

[0021] However, the relevant technical solutions have significant limitations: the switching frequency of multiphase power supplies remains fixed. While a fixed switching frequency simplifies the debugging process of PID parameters and reduces the complexity of parameter matching in the early stages of development, it cannot adapt to the dynamically changing load requirements of AI servers. Specifically, under light load conditions, fixed high-frequency operation leads to unnecessary switching and drive losses in the multiphase power supply, resulting in a significant decrease in power efficiency. Under heavy load conditions, fixed low-frequency operation may fail to meet the dynamic response speed requirements during load changes, leading to excessive overshoot or excessively long recovery time in the output voltage, affecting the stable operation of the core computing units. These problems ultimately make it difficult to maximize the overall efficiency of the multiphase power supply system, failing to meet the high energy efficiency requirements of current AI servers. Especially in large-scale AI data center scenarios, insufficient power efficiency will significantly increase overall energy consumption and operating costs.

[0022] To address some or all of the aforementioned technical problems, this application provides a control method for a multiphase power supply. First, it acquires the dynamic response data of the multiphase power supply under a target load condition point, identifying the target load condition point that reflects the characteristics of light / heavy loads, thus establishing a load matching basis for subsequent parameter adjustments. Next, using a first pre-trained neural network, it obtains the output quantity based on the dynamic response data corresponding to the target load condition point. This neural network, after training, ensures that the output quantity allows the power supply's dynamic response to meet preset standards, providing basic algorithmic support for stable control under different loads. By collecting actual load dynamic response indicators such as load current ratio, current change rate, and load stabilization time in real-world business scenarios, and then using a second pre-trained neural network to map and obtain the corresponding predicted switching frequency, this method achieves dynamic matching between load characteristics and switching frequency. Simultaneously, in conjunction with the optimized output quantity, it further ensures voltage stability. Ultimately, by coordinating the output and the predicted switching frequency, the switching frequency of the multiphase power supply is no longer fixed, but dynamically adjusted according to the actual load dynamic indicators. This not only overcomes the pain point of high high-frequency loss under light load, but also solves the problems of poor low-frequency response and voltage instability under heavy load, achieving a balance between efficiency and stability under different load conditions.

[0023] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 1A The diagram shows a framework of the power detection circuit upon which the control method for a multiphase power supply depends.

[0025] The current detection circuit includes: a current sensor and a microcontroller unit (MCU), a multiphase power controller, a power conversion module (Powerstage), a trigger, a VRTT tool, a precision resistor R2, an insulated gate field-effect transistor Q1, an output load R1, and a personal computer (PC) / cloud server.

[0026] The communication protocol, data format, data flow, and synchronization mechanism between the system components in this application are specified as follows: The current sensor communicates with the MCU via an I2C interface, using standard ADC sampled values ​​as the data format. The MCU communicates with the multiphase power controller (Controller) via PMBus or I2C protocol. The MCU directly sends the real-time calculated load current percentage, current change rate di / dt, and load settling time to the Controller. Neural network inference can be executed locally within the MCU / Controller, and the PID parameters and predicted switching frequency obtained from the inference results are directly sent from the MCU to the Controller; alternatively, inference can be executed in the cloud, and the results can be sent to the MCU and then forwarded to the Controller. The current sensor serves two purposes: first, it promptly acquires the current magnitude on the output load path and quickly feeds it back to the MCU; second, it acquires the voltage information at the load end and promptly feeds it back to the MCU. The internal logic diagram of the current sensor is shown below. Figure 1B As shown. Voltage acquisition is achieved through voltage sampling using an analog-to-digital converter (ADC). The voltage and current information acquired by the current sensor are sent to the MCU via the Inter-Integrated Circuit Bus (I2C) interface. The V1 and V2 interfaces of the current sensor are connected to the two ends of the precision resistor R2, respectively. The current through the load R1 can be calculated using the following formula: .

[0027] The MCU is used for data acquisition and preprocessing. It is a bridge connecting the current sensor and the PC / server. The MCU is responsible for acquiring real-time data from the current sensor and performing preliminary preprocessing. It stores the initial version of PID parameters as a backup. It can also match the load type, call cloud parameters or trigger backup parameters in actual business scenarios, and participate in the local recording and feedback of fault information.

[0028] The multiphase power controller is responsible for coordinating and managing the power conversion modules (Powerstage) of each phase of the multiphase power supply. It receives PID parameters and switching frequency commands from the MCU or cloud, outputs control signals to regulate the operation of the power conversion modules, and feeds back fault information and operating status to the cloud server through the Power Management Bus (PMBUS) protocol to achieve two-way interaction with the cloud.

[0029] Powerstage is the core hardware module in a multiphase power system that directly undertakes power conversion and transmission. It is located between the multiphase power controller and the load (such as GPU / TPU). According to the controller's instructions, it adjusts the switching frequency and duty cycle to convert the input voltage into a stable output voltage required by the load. Its enable terminal (En) is controlled by a trigger and can be quickly cut off in case of abnormality to protect the hardware.

[0030] The trigger's function is to acquire voltage information from the current sensor feedback, then compare it with a threshold voltage (derived from the current information). If the voltage / current exceeds the normal voltage / current by 50%, a low-potential signal is sent to the insulated-gate field-effect transistor MOSQ1 to turn it off, thereby disconnecting the enable terminal (En) of the Powerstage and stopping the output. If the voltage is below the threshold voltage, a high-potential signal is sent to MOSQ1 to turn it on, allowing normal operation. Figure 1C As shown, the trigger internally obtains the sensor's voltage and current information via the I2C interface, converts it into analog voltage information through a digital-to-analog converter (DAC), and compares it with a set threshold. Under normal circumstances, S1 outputs a high level.

[0031] The VRTT Tool is used during the initial PID tuning of a power supply board to perform a full-load frequency sweep test, focusing on capturing the dynamic response data of the power supply output under different load conditions. This helps pinpoint the worst-case operating point where the system performance is weakest. After determining the worst-case operating point, a personal computer (PC) / cloud server automatically adjusts the PID parameters based on a neural network system, gradually adjusting the power supply's load frequency and duty cycle. Simultaneously, it monitors key indicators such as output voltage stability and ripple coefficient in real time until the dynamic performance at that point meets the preset specification requirements. This provides accurate data support for subsequent parameter optimization, ensuring that the initial version of the PID parameters can meet the full load domain requirements.

[0032] The precision resistor R2 works in conjunction with the current sensor to calculate the load current through the voltage difference. It is a key component for current detection and ensures the accuracy of current data acquisition.

[0033] The insulated gate field-effect transistor Q1 serves as the switch for the enable terminal of the power conversion module. It is controlled by a trigger. When the level is low, the enable signal is turned off, cutting off the output of the power conversion module and realizing hardware protection in case of abnormality.

[0034] The output load R1 is used to simulate the actual load of the AI ​​server's CPU, GPU, etc. Its current and voltage changes reflect the load requirements of real business scenarios and are the object of system parameter optimization and performance verification.

[0035] The role of the PC / cloud server is to train the neural network model. It mainly undertakes offline training of the neural network model, system parameter configuration and functional verification. It does not directly participate in the real-time control of the multiphase power supply of the AI ​​server, but forms a collaborative closed loop with the multiphase power supply system of the AI ​​server for offline training and online application.

[0036] The embodiments of this application provide a control method for a multiphase power supply. The method is described in detail below in conjunction with the execution flow of the control method for a multiphase power supply.

[0037] like Figure 2 As shown, the method includes the following steps: S201. Obtain the dynamic response data of the multiphase power supply at the target load operating point.

[0038] Among them, the target load operating point is characterized by the multiphase power supply facing the chip power supply specification that the voltage overshoot / undershoot exceeds ±3% when the load changes under the load frequency and duty cycle, or the voltage recovery time exceeds the design requirement of 5μs, or even problems such as excessive ripple and abnormally high power loss.

[0039] In some embodiments, when performing step S201, the dynamic response data of the multiphase power supply output under different load conditions is compared based on the pre-stored frequency sweep test results of the multiphase power supply. The sweep test results include the dynamic response data of the power supply output under different load conditions. Then, target load conditions where the overshoot or undershoot of the output voltage exceeds a preset range are selected, along with the corresponding dynamic response data; or, target load conditions where the voltage recovery time exceeds a response time threshold are selected, along with the corresponding dynamic response data.

[0040] During the initial PID debugging of the circuit board, the VRTT Tool can be used to perform a full-load frequency sweep test on the multi-phase power supply to obtain the sweep test results. By applying load change signals to the multi-phase power supply through the VRTT Tool, such as step-like current jumps from 10%-20% under light load to 100% under full load, or high-frequency pulse fluctuations, the dynamic response data of the power supply output terminal can be collected in real time, including key indicators such as output voltage overshoot, undershoot, voltage recovery time, voltage ripple, and power loss at different switching frequencies.

[0041] The core of frequency sweeping is to traverse the possible load frequency range (e.g., 300kHz-1MHz) and duty cycle combinations of a multiphase power supply. By comparing the dynamic response performance under different load frequency and duty cycle combinations, the target load condition point with the weakest system performance is identified. The duty cycle is related to the input / output voltage ratio, i.e., D=V. out / V in .

[0042] For example, when the VRTT Tool finds in a frequency sweep test that when the load frequency is 400kHz and the duty cycle is 0.3, the power supply output voltage downsurge reaches 5% (exceeding ±3% of the specification) and the response time is as long as 8μs (exceeding the 5μs requirement) during the process of the load jumping from 30% to 80%, while the performance of other load frequency and duty cycle combinations meets the specifications, then the 400kHz load frequency and 0.3 duty cycle are determined to be the target load operating point.

[0043] The above embodiments compare dynamic response data at different load conditions to ensure comprehensive capture of the performance of multiphase power supplies across the entire load range; they focus on core indicators such as voltage overshoot / undershoot exceeding preset range and voltage recovery time exceeding threshold for screening, accurately locating the target load condition, so that subsequent adjustments to PID parameters and switching frequency can be targeted to the target load condition, avoiding aimless generalization optimization and improving optimization efficiency and accuracy.

[0044] S202. Determine the output quantity based on the dynamic response data corresponding to the target load operating point and the first pre-trained neural network.

[0045] The first pre-trained neural network is built on a lightweight fully connected network (MLP), with the input layer corresponding to dynamic response data and the output layer outputting PID parameters.

[0046] The lightweight architecture reduces computational complexity, preventing excessively long model inference times from impacting power supply dynamic regulation efficiency. Simultaneously, it focuses on optimizing the proportional-integral-derivative (PID) parameters (including proportional coefficient Kp, integral coefficient Ki, and derivative coefficient Kd), ensuring more precise optimization. The first pre-trained neural network has already been trained on massive amounts of load scenario data, possessing the ability to learn the mapping relationship between load characteristics and PID parameters, enabling rapid inference from input data.

[0047] The output of the first pre-trained neural network is a PID parameter used to adjust the dynamic response data of the multi-phase power supply output to meet preset standards. Meeting these preset standards includes ensuring that the overshoot or undershoot of the output voltage meets the ±3% voltage tolerance specification for the AI ​​server chip power supply; that is, the deviation between the actual output voltage and the target voltage must not exceed this range to avoid chip performance abnormalities or protection triggering due to voltage fluctuations. Regarding voltage recovery time, it is necessary to minimize the time required for the voltage to recover to a stable range during sudden load changes. In this application, a response time threshold of 5μs is set to ensure that the power supply can quickly adapt to the high-frequency fluctuation characteristics of the AI ​​load.

[0048] In some embodiments, when performing step S202, the dynamic response data corresponding to the target load condition point is input into the first pre-trained neural network, and then the first pre-trained neural network is guided to output PID parameters based on the first loss function of the first pre-trained neural network.

[0049] When inputting the dynamic response data corresponding to the target load operating point into the first pre-trained neural network, it can first be standardized according to the input requirements of the first pre-trained neural network to extract the actual output voltage, target supply voltage, voltage recovery time, etc. under the target load operating point; after processing, these data are input into the first pre-trained neural network according to the preset feature dimensions.

[0050] In the above embodiments, when the first loss function of the first pre-trained neural network guides the output of PID parameters, the first output voltage and first voltage recovery time of the multi-phase power supply at the target load operating point are first obtained. Then, the absolute value of the first voltage deviation rate is calculated based on the first output voltage and the target supply voltage. Next, the overshoot penalty term of the first loss function is determined based on the difference between the absolute value of the first voltage deviation rate and the voltage tolerance ratio. Simultaneously, the response time penalty term is determined based on the difference between the first voltage recovery time and the response time threshold; and the parameter optimization term is determined based on the candidate quantities (i.e., candidate PID parameters) output by the first pre-trained neural network. The overshoot penalty term, response time penalty term, and parameter optimization term are further weighted and combined to calculate the first loss value of the first loss function. The candidate quantities (candidate PID parameters) are adjusted based on this first loss value until the first loss value converges, thus obtaining the output quantity.

[0051] The target supply voltage is a pre-set target value for the supply voltage. The voltage tolerance ratio is a preset tolerance range for supply voltage fluctuations. The first pre-trained neural network adjusts the internal connection weights based on the first loss value using a backpropagation algorithm. If the first loss value is too high, such as when the overshoot exceeds the limit, the weights related to voltage stability are adjusted accordingly, making the subsequent output PID parameters more biased towards suppressing overshoot. Alternatively, if parameters exceed the limits, causing an increase in the parameter optimization term, the weights of the parameter range constraint are strengthened, forcing the parameters to return to the effective range. This continues until the first loss value converges to its minimum. When the first loss value reaches the minimum threshold, i.e., output voltage overshoot / undershoot ≤ ±3%, voltage recovery time ≤ 5μs, and PID parameters within the range [100, 255], the first pre-trained neural network stops iterating and outputs the current PID parameters (K). p K i K d The PID parameter is the optimal control parameter adapted to the target load condition.

[0052] First loss function As in formula (1): (1) In formula (1), This is an overshoot penalty term used to quantify the deviation between the voltage overshoot / undershoot and the target voltage specification. in, V 实际 This is the actual output voltage of the multiphase power supply, acquired through a current sensor; V 目标 It is the target output voltage of the multiphase power supply, which is a preset value.

[0053] In formula (1), This is a penalty term for response time. If the response time exceeds 5μs, the excess portion is calculated; otherwise, this term is 0. , t 响应 This is the actual voltage recovery time, which is monitored in real time using the VRTT Tool.

[0054] In formula (1), For parameter optimization terms, where , , , This is the penalty coefficient, which can be 10, to ensure that the parameter can be executed by the Controller.

[0055] Specifically, the first step is to acquire the first output voltage of the multiphase power supply using the ADC sampling function of the current sensor, and mark it as the actual output voltage V. 实际 Call the system's preset target power supply voltage V for the chip 目标 (e.g., GPU rated at 1.2V), and simultaneously monitor the voltage recovery time t during load surges under this operating condition using the VRTT Tool. 响应 Simultaneously, obtain the candidate PID parameters (K) output by the first pre-trained neural network. p K i K d ), which clearly defines its physical effective range as [100, 255].

[0056] The second step is to calculate the overshoot penalty term L. V First, through the formula V=(V 实际 -V 目标 ) / V 目标 Calculate the absolute value of the deviation rate between the actual output voltage and the target supply voltage; then calculate the overshoot penalty term L based on the absolute value of the voltage deviation rate. V If | V|≤3% conforms to the chip power supply specification, then L V =0, if | If V|>3%, then L V =| V|-3%, which quantifies the degree of voltage overshoot / undershoot exceeding the limit.

[0057] The third step is to calculate the response time penalty term L. T Determine t 响应 Does it exceed the preset response time threshold (e.g., 5μs): If t 响应 ≤5μs indicates that the dynamic response requirement is met, then L T =0; if t 响应 >5μs, then L T =t 响应 -5μs, penalizing only the excess, balancing response speed and optimization flexibility.

[0058] The fourth step is to calculate the parameter optimization term and the total loss. First, determine if the candidate PID parameters are within the range [100, 255]. If all are within this range, the parameter optimization term P(K) is calculated. p ,K i ,K d If any parameter is out of range, P(K) = 0; p ,K i ,K d Take the difference between the value exceeding the boundary (e.g., K). p When the value is 90, P = 100 - 90 = 10); then substitute the Loss... PID The general formula is a weighted summation based on weights, i.e., the loss. PID =0.7×L V +0.3×L T +10×P(K p ,K i ,K d (10 is the penalty coefficient σ, to ensure parameter compatibility priority), and finally obtain the first loss value that reflects the comprehensive performance of the PID parameters. The smaller the first loss value, the better the parameters.

[0059] When calculating the first loss value of the first loss function, the overshoot penalty term L V A penalty is imposed on voltage fluctuations exceeding the specification by calculating the difference between the absolute value of the voltage deviation and 3% (0 when the difference is negative), and its weight (0.7) is higher than that of the response time penalty term L. T (0.3), reflecting the design logic that prioritizes voltage stability; secondly, L T Only the portion exceeding 5μs in response time is calculated; no penalty is applied when the limit is not exceeded. This ensures both dynamic response speed and avoids excessive constraints that limit the parameter optimization space. Finally, the parameter optimization term P(K) p ,K i ,K d The penalty coefficient σ (with a value of 10) constitutes a hardware compatibility constraint, and K is explicitly defined. pK i K d The physical value range is [100, 255]. This range matches the parameter execution capability of the multiphase power supply controller. If the network output parameter exceeds this range, a linear penalty will be triggered, forcing the parameter to return to the executable range to avoid control failure due to invalid parameters.

[0060] The above embodiments, through rapid inference and iterative optimization of the first pre-trained neural network, can quickly output PID parameters. Furthermore, based on constraints imposed by the first loss function, standardization of PID parameter tuning under different load conditions is achieved. The first loss function prioritizes voltage stability, secondarily considers telephone recovery time, and takes PID parameter compatibility as a baseline, guiding the PID parameters output by the first pre-trained neural network to specifically address issues such as overshoot / undershoot exceeding limits and response timeouts under the target load condition. The output PID parameters not only meet the dynamic response standards of the target load condition but also adapt to the hardware execution capabilities. Subsequent iterations are sufficient to extend to other load scenarios, reducing the cost of repeated debugging.

[0061] It should be noted that although the first pre-trained neural network optimization is based on theoretical models and data output parameters, the actual multiphase power supply system is affected by hardware characteristics (such as the switching losses of Powerstage and the on-resistance of MOSFETs), circuit noise, and other factors. The initial version of PID parameters may meet the theoretical requirements but not the actual requirements. Therefore, it is necessary to use the VRTT Tool, a professional testing tool, to verify the authenticity and feasibility of the predicted PID parameters output by the first pre-trained neural network in a real hardware environment to avoid problems such as excessive voltage fluctuations and response delays in subsequent system operation.

[0062] In some embodiments, after executing step S202, the multiphase power supply can be controlled to operate according to the output quantity, and the first dynamic response data output by the multiphase power supply at the target load condition point can be collected. Then, if the first dynamic response data does not meet the preset standard, the weights of the first loss function of the first pre-trained neural network can be adjusted until the first dynamic response data meets the preset standard.

[0063] Optionally, the initial PID parameters (proportional coefficient K) obtained by the first pre-trained neural network optimization are first... p Integral coefficient K i Differential coefficient K dThe parameters are written to the core control unit VR Controller of the multiphase power supply through the control interface, enabling the multiphase power supply to operate according to these parameters. At this time, the multiphase power supply simulates the server's operating state under extreme conditions based on the target load condition determined by the previous sweep test results. The VRTT Tool is used to focus on testing the previously located target load condition, and the dynamic response data of the power supply output under this condition is collected in real time to determine whether it meets the preset standard Spec (output voltage overshoot / undershoot ≤ ±3%, voltage recovery time ≤ 5μs). If the verification result meets the Spec, it means that the initial version of the PID parameters is adapted to the hardware characteristics and can effectively solve the performance shortcomings of the worst condition, meeting all load conditions. At this time, the parameters need to be written to the local MCU (as an offline backup to deal with abnormal cloud parameter call scenarios) and VR Controller (to ensure direct call during daily operation). If the verification result does not meet the Spec (e.g., voltage overshoot reaches 4%, response time reaches 7μs), it means that the initial version of the parameters is not adapted to the actual hardware characteristics. The network input characteristics or loss function weights need to be readjusted, and new PID parameters need to be generated again until the Spec is met.

[0064] The above hardware-level testing using VRTT Tool verifies the parameters in a real power system, ensuring that the final output PID parameters are compatible with hardware performance and improving the actual usability of the PID parameters. When the dynamic response data is substandard, the weights of the first loss function are adjusted to guide the first pre-trained neural network to specifically compensate for performance shortcomings and improve the matching degree between the PID parameters and the load conditions. The PID parameters that meet the standards are simultaneously written to the MCU and VR Controller, ensuring that the VRController can directly call the parameters during daily operation, and that in abnormal scenarios such as cloud parameter loss or controller failure, the MCU can quickly read the backup parameters to restore operation, avoiding system downtime due to parameter issues.

[0065] S203. Collect actual load dynamic response indicators of multiphase power supplies in actual business scenarios.

[0066] Among them, the actual load dynamic response indicators include the load current ratio, current change rate, and load stabilization time.

[0067] Load current percentage is the ratio of the current to the rated current. It reflects the severity of the load; a low load current percentage indicates a light load, while a high load current percentage indicates a heavy load. Load current percentage is the basis for determining whether frequency throttling for energy saving or frequency boosting to maintain response is necessary.

[0068] The rate of change of current is the ratio of the difference between two consecutive current readings to the sampling interval, focusing on the speed of load fluctuations. The rate of change of current can help a second pre-trained neural network identify the risk of sudden load changes and adjust the frequency in advance to avoid voltage overshoot / undershoot. For example, in AI model training, the rate of change is high when the current jumps from 30% to 90%, while in inference scenarios, the rate of change is low when the current fluctuates smoothly.

[0069] Load stabilization time is the duration for which the current remains within a preset fluctuation range, reflecting the continuity of the load state. A long load stabilization time indicates that the load is in a steady state, and efficiency can be optimized first; a short load stabilization time indicates frequent load switching (such as multi-task alternating inference), and dynamic response should be prioritized. When executing step S203, the actual load dynamic response indicators of multi-phase power supplies under actual business scenarios can be obtained by loading various processor load stress models.

[0070] The load stress model is used to reproduce the computing power requirements and power consumption fluctuations of CPU / GPU under different business scenarios. Examples include: continuous high GPU load during model training (power consumption maintained at 80%-100%), intermittent moderate GPU load during inference (power consumption fluctuating between 40%-60%), light CPU load and low power consumption during standby (power consumption below 20%), and sudden high CPU load during data preprocessing (power consumption jumping from 30% to 90% in a short time). Each model includes key parameters such as the magnitude of load current changes, fluctuation frequency, and duration of stability, perfectly matching the hardware operating state in actual business scenarios and avoiding discrepancies between optimization results and real-world applications due to idealized load assumptions.

[0071] After loading multiple real-world load models, dynamic data across the entire load range can be obtained. For example, the low-current steady state under light load, the smooth fluctuation state under medium load, and the high-current sudden state under heavy load. This data will serve as input features (such as load current percentage, current change rate, and load stabilization time) for the second pre-trained neural network. This will support the network in learning the correspondence between different load scenarios and the optimal switching frequency and PID parameters, ensuring that subsequent optimization can adjust parameters in real time according to actual load changes, rather than being limited to a single load condition.

[0072] By replicating the diverse load states of the server in real-world business scenarios, we can provide training data that closely resembles real-world scenarios for the subsequent adaptive optimization of switching frequency and PID parameters, ensuring that the optimization results can cover various operating conditions of the server in daily operation.

[0073] S204. Based on the actual load dynamic response index and the second pre-trained neural network, obtain the predicted switching frequency corresponding to the actual load dynamic response index.

[0074] The second pre-trained neural network is trained based on the simulated load dynamic response indicators and corresponding switching frequencies of multiphase power supplies under different business simulation scenarios. The second pre-trained neural network is a lightweight fully connected MLP.

[0075] In some embodiments, when performing step S204, the actual load dynamic index is first input into the second pre-trained neural network, and then the predicted switching frequency output by the second pre-trained neural network is guided based on the second loss function of the second pre-trained neural network.

[0076] When inputting the actual load dynamic response index into the second pre-trained neural network, the actual load dynamic response index is standardized to ensure that it meets the neural network input format requirements, and then it is transmitted to the second pre-trained neural network through the data interface.

[0077] In the above embodiments, when guiding the predicted switching frequency output by the second pre-trained neural network based on the second loss function of the second pre-trained neural network, the first load current, first duty cycle, and first voltage ripple of the multi-phase power supply under actual business scenarios are first obtained. Then, a first conduction loss term is calculated based on the first load current and the first duty cycle, and a first turn-off loss term is calculated based on the first load current, the preset voltage, and the first candidate switching frequency output by the second pre-trained neural network. The first conduction loss term and the first turn-off loss term are then summed to obtain a first switching loss term. A voltage ripple constraint term is determined based on the difference between the first voltage ripple and the preset ripple allowable threshold, and a frequency constraint penalty term is calculated by comparing the first candidate switching frequency with the preset switching frequency range. Further, the first switching loss term, the voltage ripple constraint term, and the frequency constraint penalty term are weighted and combined to obtain a second loss value of the second loss function. The candidate switching frequency is then adjusted based on the second loss value until the second loss value converges to obtain the predicted switching frequency.

[0078] The first switching loss term is used to quantify the power loss level at the first candidate switching frequency. If the second loss value is high, the internal connection weights are adjusted through the backpropagation algorithm. For example, if the switching loss term is too large under light load, the weight of reducing the switching loss term is strengthened to guide the output to a lower frequency; if the voltage ripple exceeds the standard under heavy load, the weight of suppressing voltage ripple is strengthened to drive the output to a higher frequency, until the loss value converges to the minimum value. At this point, the predicted switching frequency of the output is the optimal value adapted to the current actual load.

[0079] The second loss function is shown in formula (2): (2) In formula (2), The switching loss term for Powerstage can be calculated using the following formula: , where Lon It is the conduction loss, L off It is the shutdown loss; I R2 It is the current load current, which is collected in real time by a current sensor; The on-resistance of the MOSFET can be found in the MOSFET hardware datasheet; D is the duty cycle. ), Input voltage, For output voltage, The MOS turn-off time is obtained from the MOS transistor hardware datasheet, and f is the switching frequency to be optimized output by the second pre-trained network.

[0080] In formula (2), Voltage ripple can be collected using a current sensor.

[0081] In formula (2), The frequency constraint penalty term is calculated by comparing the switching frequency output by the second pre-trained network with a preset range. , The value interval is 100kHz, if Hz, penalty item is ,like The penalty item is Otherwise, it is 0.

[0082] In formula (2), This is a penalty coefficient, set to 500, which prioritizes ripple optimization while ensuring the frequency does not exceed the limit.

[0083] Specifically, the first step is to acquire the first load current I using a current sensor. R2 I R2 =(V1-V2) / R2, where V1 and V2 are the voltages across the precision resistor R2, and the output voltage V is... out and the first voltage ripple V ripple Call the system's preset input voltage V in Consulting the datasheet, we obtain the on-resistance R of the MOSFET. ds(on) MOSFET turn-off time t off At the same time, the first candidate switching frequency f of the output of the second pre-trained neural network is recorded.

[0084] The second step is to calculate the switching loss term L. switch First, based on V in With V out Calculate the first duty cycle D=V out / V in Then calculate the first conduction loss term and the first turn-off loss term separately. The first conduction loss term L on =I R22 ×R ds(on) ×D, First shutdown loss term L off =0.5×V in ×I R2 ×t off ×f; Finally, summing up yields the total switching loss term L. switch =L on +L off This quantifies the power loss level at the current switching frequency.

[0085] The third step is to calculate the voltage ripple constraint term. First, determine the ripple allowable threshold as 3% × V. out Then determine the first voltage ripple V. ripple Does it exceed the standard? If V ripple ≤3%×V out The value of this item is 0; if V ripple >3%×V out The value of this term = V ripple -3%×V out This ensures that frequency optimization does not sacrifice output stability.

[0086] The fourth step is to calculate the frequency constraint penalty term L(f). The effective switching frequency range is defined as 300kHz-1000kHz: if f < 300kHz, L(f) = 300kHz - f; if f > 1000kHz, L(f) = f - 1000kHz; if f is within the range, L(f) = 0, then multiplied by the penalty coefficient. =500, resulting in a frequency constraint penalty term of 500×L(f), ensuring that the frequency does not exceed the hardware's capacity.

[0087] Fifth step, calculate the second loss value L. freq Sum the three terms according to their weights, i.e., L freq =0.7×L switch +0.3×max(0,V ripple -3%×V out The weight of 0.7 prioritizes efficiency, 0.3 takes into account stability, and the high weight of 500 adheres to the bottom line of hardware constraints. Finally, the second loss value is obtained to guide the optimization of the switching frequency. The smaller the second loss value, the more the switching frequency is adapted to the current load scenario.

[0088] The above embodiments output the optimal switching frequency through a second pre-trained neural network based on actual load dynamic indicators, enabling the multi-phase power supply to balance efficiency and performance across the entire load range. The second pre-trained neural network adopts a lightweight MLP structure, which can quickly react to sudden changes in the AI ​​load. At the same time, the second loss function avoids frequency optimization biased towards a single indicator through multi-dimensional constraints, ensuring that the output switching frequency meets both hardware capabilities and power supply stability requirements. Through the dynamic constraints of the second loss function, it can adapt to the hardware characteristics of different models of multi-phase power supplies, eliminating the need to redevelop control logic for a single hardware, thus improving the versatility and scalability of the solution.

[0089] In some embodiments, before executing step S204, for any business simulation scenario of different business simulation scenarios, the second load current, preset input voltage, second output voltage, and second voltage ripple, as well as the second voltage recovery time and second duty cycle under the target load operating point, can be collected. Then, the second conduction loss term is calculated based on the second duty cycle and the second load current, and the second turn-off loss term is calculated based on the second load current, the preset input voltage, and the second candidate switching frequency output by the second pre-trained model for any business simulation scenario. The absolute value of the second voltage deviation rate for any business simulation scenario is calculated based on the second output voltage and the target supply voltage. Then, the second switching loss term, the absolute value of the second voltage deviation rate, and the second voltage recovery time are weighted and combined to obtain the weight optimization fitness of any business simulation scenario. The sum of the weight optimization fitness of different business simulation scenarios is calculated, and the network weights of the second pre-trained neural network are further adjusted according to the sum of the weight optimization fitness until the sum of the weight optimization fitness converges.

[0090] The fitness function for weight optimization in any business simulation scenario is calculated using the following equation (3): (3) In formula (3), , , These represent the switching loss, voltage deviation rate, and voltage recovery time in the i-th business simulation scenario, respectively, with weights assigned as efficiency 0.5, dynamics 0.3, and response time 0.2.

[0091] Specifically, the first step is to determine the different business simulation scenarios to be included in the calculation, such as light-load standby (load current accounts for 10%-20%), medium-load inference (accounting for 40%-60%), full-load training (accounting for 90%-100%), and load changeover (such as current jumping from 30% to 80%), etc., and delineate the scope for subsequent full-scenario data collection to avoid result deviations caused by a single scenario.

[0092] The second step is to collect basic operational data for a single scenario (the i-th scenario). For the i-th business simulation scenario, multi-dimensional data collection is completed through the hardware layer: real-time second load current I is obtained using a current sensor. R2(i) Through formula I R2(i) =(V1 i -V2 i ) / R2 calculation, V1 i V2 i The voltage across the precision resistor R2 collected by the sensor, and the preset input voltage V in(i) Second output voltage V out(i) and the second voltage ripple; the second voltage recovery time t under load abrupt changes in this scenario was monitored using the VRTT Tool. 响应(i) Simultaneously, preset parameters are invoked, such as the MOSFET on-resistance R. ds(on) Shutdown time t off and target power supply voltage V 目标 This provides raw data for subsequent loss and deviation calculations.

[0093] The third step is to calculate the switching loss term L for the i-th business simulation scenario. switch(i) First, calculate the second duty cycle D based on the input and output voltages. i Then calculate the second conduction loss term and the second turn-off loss term of the MOSFET respectively: Second conduction loss term L on(i) =I R2(i) 2 ×R ds(on) ×D i The second shutdown loss term L off(i) =0.5×V in(i) ×I R2(i) ×t off ×f i f i The switching frequency output by the weights of the current second pre-trained neural network for this business simulation scenario is given; finally, the two are summed to obtain the total switching loss L of Powerstage in the i-th scenario. switch(i) =L on(i) +L off(i) This quantifies the power efficiency level in this scenario.

[0094] Fourth step, calculate the absolute value of the voltage deviation rate for the i-th business simulation scenario. V i |Weighted value for a single scenario. First, use the formula. Vi = (V 实际(i) -V 目标 ) / V 目标 Calculate the voltage deviation rate (V) 实际(i)The actual output voltage for this scenario is collected by the current sensor; then the absolute value is taken to obtain | V i | reflects voltage stability; then, substituting into the single-scenario weighted formula, and allocating weights according to efficiency (0.5), dynamics (0.3), and response (0.2), the performance score for that scenario is calculated: Single-scenario weighted value i = 0.5 × L switch(i) +0.3×| V i |+0.2×t 响应(i) The lower the score, the better the weights of the second pre-trained neural network perform in the business simulation scenario.

[0095] Step 5: Summarize the single-scenario weighted values ​​of the n scenarios and calculate the total weighted sum. Iterate through all n load scenarios, repeating steps 2 to 4 to obtain the single-scenario weighted value (i) for each scenario. Then sum all the values ​​to obtain the total weighted value = sum i =1^n[0.5×L switch(i) +0.3×| V i |+0.2×t 响应(i) The sum reflects the overall performance of the weights of the current second pre-trained neural network under full load scenarios. The smaller the sum, the stronger the adaptability of the weights to multiple scenarios.

[0096] Step 6: Calculate the final value F of the fitness function. 权重 The total weighted value is divided by the number of scenes, n, to obtain the average value. This average value eliminates the random influence of individual scenes and becomes the core criterion for the Grey Wolf Optimizer (GWO) to judge the quality of weights. GWO will use "minimizing F" as the criterion. 权重 "With the goal of continuously adjusting the network weights of the second pre-trained neural network until F..." 权重 When the weights converge to the minimum value, they represent the optimal weight combination that balances overall efficiency, dynamic response, and response time.

[0097] The above embodiments optimize the fitness of weights based on multi-scenario data calculations, comprehensively considering switching losses, voltage stability, and response speed under all business simulation scenarios. This ensures that the adjusted weights of the second pre-trained neural network can adapt to the needs of different scenarios in a balanced way, achieving performance balance across all business scenarios. Leveraging the characteristics of the GWO algorithm, when the sum of the fitness of weights in different business simulation scenarios does not converge, the search direction is dynamically adjusted based on the changing trend of the fitness sum. This avoids stagnation in local optimum regions and allows for rapid convergence towards the optimal weight combination with the minimum fitness sum, improving the efficiency and success rate of weight optimization. Simultaneously, it ensures that the final adjusted weights can achieve multi-objective optimization in all business scenarios, minimizing switching losses, voltage deviation, and recovery time, laying a solid foundation for the accurate output of adaptive switching frequencies by the second pre-trained neural network.

[0098] S205. Control the multiphase power supply based on the output and predicted switching frequency.

[0099] The multiphase power supply is controlled based on the PID parameters output by the first pre-trained neural network and the predicted switching frequency output by the second pre-trained neural network.

[0100] In some embodiments, for different business simulation scenarios optimized through pre-trained neural networks, GWO weight adjustment, and VRTT Tool verification, such as light-load standby, medium-load inference, full-load training, and load abrupt switching, the optimal PID parameters corresponding to each business simulation scenario are extracted and associated with the feature labels of that business simulation scenario, including scenario type, load characteristic threshold, compatible hardware model, and dynamic response data. The load characteristic threshold includes the load current percentage range, current change rate range, and load stabilization time range.

[0101] By binding feature tags with PID parameters, a data set relating scene features, parameters, and performance is formed, ensuring that subsequent calls can quickly match the actual business scenario.

[0102] A secure data transmission channel between the local hardware of the AI ​​server and the cloud server can be established via the PMBUS interface of the multiphase power system or industrial Ethernet. Encryption protocols are used during transmission to prevent parameter leakage or tampering. A structured database is built in the cloud, with a dedicated data storage table structure designed. Fields in the table include scene identifiers, feature tag sets (stored in JSON format as thresholds such as load current percentage and rate of change), PID parameter groups, parameter generation time, hardware adaptation information, and performance verification results. The organized scene features and parameter association data groups are written to the cloud database row by row according to the table structure, completing the batch storage of various scene parameters and forming a parameter set covering all business scenarios. Cloud storage and local MCU backup form dual storage redundancy, preventing downtime due to missing parameters.

[0103] This application also provides a control method for a multiphase power supply. The method first obtains a first load pressure model for the current business scenario, then determines a second load pressure model that matches the first load pressure model from load pressure models of different business simulation scenarios, and then controls the multiphase power supply to operate according to the output and switching frequency corresponding to the second load pressure model under the current business scenario.

[0104] Among them, in the load pressure models of different business simulation scenarios, each pressure model is marked with the corresponding load characteristic threshold range. For example, the GPU full load training model corresponds to a load current ratio of 80%-100%, current change rate ≤5A / μs, and load stabilization time ≥50μs.

[0105] The second load pressure model is determined from the load pressure models of different business simulation scenarios to match the first load pressure model. This includes: if all three characteristics of the actual load fall within the threshold range of a certain load pressure model, the load type is determined to match. At this time, the MCU will retrieve the optimized PID parameters and switching frequency parameters corresponding to the model from the cloud and transmit them to the VR Controller to control the multiphase power supply to operate according to the set of parameters; if the actual load characteristics do not match the threshold range of all n stored models, such as extreme load fluctuations exceeding the preset range, or the addition of new untrained business scenarios, the matching is determined to fail. When the load type matching fails, the MCU will immediately read the previously stored initial version of the PID parameters. These parameters have been verified by the VRTT Tool and meet the basic operating specifications of the entire load domain, namely, output voltage overshoot / undershoot ≤ ±3% and voltage recovery time ≤ 5μs. The parameters are written to the VR Controller through the control interface to force the multiphase power supply to switch to the mode of operating according to the initial version of the PID parameters.

[0106] For example, when the AI ​​server is actually running, it collects the load characteristics (including load current ratio, current change rate and load stabilization time) of the first load pressure model of the current business scenario in real time, and uploads the feature data to the cloud in encryption; the cloud database quickly retrieves the scenario identifier that is closest to the current load characteristics through feature matching algorithm, retrieves the corresponding PID parameters and switching frequency and feeds them back to the local VR Controller, and the controller immediately adjusts the power supply operation status according to the parameters.

[0107] The above embodiments store optimized parameters for various business scenarios through a black-box memory mechanism. During actual operation, parameters can be quickly scheduled by directly matching load characteristics, reducing the risk of power supply fluctuations during scenario switching and ensuring business continuity. The black-box memory hides the optimization logic of the PID parameters from the user; it only requires triggering the call through load characteristics, lowering the technical threshold. If matching fails, the initial version of the PID parameters is invoked, ensuring that parameters conforming to basic operating standards are always available for multi-phase power supplies. This prevents the system from crashing due to missing or mismatched parameters, thus ensuring business continuity. The logic of prioritizing optimized parameters and using the initial version as a fallback maximizes the value of early optimizations while providing stable protection for unknown load scenarios, avoiding sacrificing system reliability in the pursuit of extreme optimization.

[0108] In some embodiments, during the operation of the multiphase power supply, current and voltage at the load end are collected by a current sensor and synchronously transmitted to the trigger and MCU via an I2C interface. The trigger internally stores an abnormal load judgment threshold, which is converted into an analog voltage signal by a DAC converter and compared in real-time with the real-time voltage signal fed back by the current sensor. When the trigger detects that the real-time data exceeds the abnormal threshold, it outputs a low-level signal S1 to the control terminal of the insulated-gate field-effect transistor (MOSQ1). Upon receiving the low-level S1, the control terminal of MOSQ1 quickly shuts down, thereby cutting off the enable signal of the Powerstage, causing the Powerstage to stop power conversion and output, preventing excessive current or abnormal voltage from damaging the multiphase power supply hardware and backend core chips such as the GPU / TPU. Simultaneously with the trigger cutting off the output, the VR Controller uploads fault information to the cloud server in a preset format via the PMBUS communication protocol, ensuring that staff can quickly locate the fault scenario through the cloud monitoring platform. After the staff diagnoses and resolves the fault on-site, they send a "restore operation" command through the cloud platform. The AI ​​server's local system will then re-call the PID parameters and switching frequency parameters adapted to the current normal load scenario from the cloud black-box database, write them into the VR Controller, and restart the Powerstage enable terminal, thus restoring the multi-phase power supply to normal operation.

[0109] The abnormal load judgment threshold is set based on the maximum regulation capability of the multi-phase power supply controller, such as an instantaneous current exceeding 150% of the rated current, or a voltage fluctuation exceeding the normal range by 50%. Fault information includes fault type, fault occurrence time, real-time load data (current and voltage values) when protection is triggered, and the current operating status of the controller.

[0110] The above embodiments respond directly to abnormal loads through triggers based on hardware circuits, cutting off Powerstage output within microseconds and shortening protection delays. The cloud platform allows for intuitive identification of fault causes, improving the transparency and efficiency of fault handling and reducing the difficulty of manual troubleshooting. After the fault is resolved, the parameters in the cloud-based black-box database are re-invoked, ensuring that the multi-phase power supply immediately adapts to the efficiency and dynamic response requirements of the current normal load scenario after recovery, avoiding secondary faults due to parameter incompatibility. Furthermore, the recovery process does not require restarting the entire server, only the multi-phase power supply module, shortening business interruption time, especially suitable for AI services with high continuity requirements, reducing the impact of faults on business operations.

[0111] like Figure 3 As shown, Figure 3 This is a flowchart illustrating a control method for a multiphase power supply provided in an embodiment of this application. The process begins by determining the worst-case operating point through a frequency sweep test. First, a full-load range frequency sweep test is performed on the multiphase power supply system using a VRTT Tool to locate the target load operating point where the output voltage overshoot / undershoot or voltage recovery time fails to meet the standard. Then, the dynamic response data of this operating point is input into a first pre-trained neural network. Based on a first loss function, the network outputs initial PID parameters, which are then imported into the multiphase power supply controller and verified again using the VRTT Tool. If the preset criteria are met, the parameters are written to the MCU and controller backups; otherwise, the optimization process is repeated until initial PID parameters covering the entire load are obtained.

[0112] Next, load stress models for various real-world business scenarios are loaded. Real-time collected 3D load characteristics (load current percentage, current change rate, and load settling time) are input into a second pre-trained neural network. The switching frequency is adaptively optimized based on a second loss function. Simultaneously, the weights of the second pre-trained neural network are optimized using the GWO algorithm, with a fitness function comprehensively considering switching losses, voltage deviation, and response time to ensure globally optimal weights. The optimized PID parameters and switching frequency parameters are stored in the cloud as a black-box memory. In actual business scenarios, the load type is first matched with the stored model. If a match is found, the cloud parameters are called; otherwise, the initial PID parameters from the MCU are used to ensure operation.

[0113] If an abnormal load such as a sudden large current occurs, the trigger detects it through the current sensor and outputs a low-level signal to disconnect the power conversion module and stop the output. The controller feeds back the fault information to the cloud. After the staff investigates the fault, they call the cloud black box database again to restore the system operation and complete the entire closed loop process.

[0114] Figure 3The multiphase power supply control method shown uses the VRTT Tool to locate the target load operating point, providing key data for subsequent parameter optimization and enhancing power supply stability. The first pre-trained neural network, combined with the first loss function, outputs PID parameters, which are then repeatedly verified and optimized to lay the foundation for stable operation of the multiphase power supply. The second pre-trained neural network and the GWO algorithm optimize the switching frequency and PID parameters in multiple scenarios, improving power supply efficiency and adaptability. Under abnormal loads, the trigger, sensor, and controller work together to disconnect the power supply in a timely manner, report faults, and restore the system, ensuring the safe and stable operation of the power system.

[0115] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0116] like Figure 4 As shown, embodiments of this application also provide a control device for a multiphase power supply, the device comprising: The acquisition module 401 is used to acquire the dynamic response data output by the multiphase power supply under the target load condition point; The proportional-integral-derivative parameter prediction module 402 is used to obtain the output quantity based on the dynamic response data corresponding to the target load operating point and the first pre-trained neural network; wherein, the output quantity is used to adjust the dynamic response data of the multiphase power supply output to meet the preset standard. The data acquisition module 403 is used to collect the actual load dynamic response indicators of multiphase power supply in actual business scenarios; the actual load dynamic response indicators include load current ratio, current change rate and load stabilization time. The switching frequency prediction module 404 is used to obtain the predicted switching frequency corresponding to the actual load dynamic index based on the actual load dynamic index and the second pre-trained neural network; wherein, the second pre-trained neural network is trained based on the simulated load dynamic index and the corresponding switching frequency of the multi-phase power supply under different business simulation scenarios. The control module 405 is used to control the multiphase power supply based on the output and the predicted switching frequency.

[0117] As an optional implementation provided in this application embodiment, the acquisition module 401 is specifically used for: comparing dynamic response data corresponding to different load operating points based on the pre-stored frequency sweep test results of the multiphase power supply; the frequency sweep test results include dynamic response data of the multiphase power supply output under different load operating points; filtering out target load operating points where the overshoot or undershoot of the output voltage exceeds a preset range and the dynamic response data corresponding to the target load operating points; or, filtering out target load operating points where the voltage recovery time exceeds a response time threshold and the dynamic response data corresponding to the target load operating points.

[0118] As an optional implementation provided in this application, the proportional-integral-differential parameter prediction module 402 is specifically used for: inputting the dynamic response data corresponding to the target load condition point into the first pre-trained neural network; and guiding the first pre-trained neural network to output the output quantity based on the first loss function of the first pre-trained neural network.

[0119] As an optional implementation provided in this application, the proportional-integral-differential parameter prediction module 402, when guiding the output quantity of the first pre-trained neural network based on the first loss function of the first pre-trained neural network, is specifically used for: obtaining the first output voltage and the first voltage recovery time of the multiphase power supply at the target load condition point; calculating the absolute value of the first voltage deviation rate based on the first output voltage and the target supply voltage; determining the overshoot penalty term of the first loss function based on the difference between the absolute value of the first voltage deviation rate and the voltage tolerance ratio; the voltage tolerance ratio is a preset tolerance range for supply voltage fluctuations; determining the response time penalty term based on the difference between the first voltage recovery time and the response time threshold; determining the parameter optimization term based on the candidate quantity output by the first pre-trained neural network; weighting and combining the overshoot penalty term, the response time penalty term, and the parameter optimization term to calculate the first loss value of the first loss function; adjusting the candidate quantity based on the first loss value until the first loss value converges to obtain the output quantity.

[0120] As an optional implementation provided in this application, the proportional-integral-derivative parameter prediction module 402 is further configured to: control the multiphase power supply to operate according to the output quantity, and collect the first dynamic response data output by the multiphase power supply at the target load condition point; and adjust the weight of the first loss function of the first pre-trained neural network until the first dynamic response data meets the preset standard if the first dynamic response data does not meet the preset standard.

[0121] As an optional implementation provided in this application, the switching frequency prediction module 404 is specifically used to: input the actual load dynamic index into the second pre-trained neural network; and guide the second pre-trained neural network to output the predicted switching frequency based on the second loss function of the second pre-trained neural network.

[0122] As an optional implementation provided in this application, when the switching frequency prediction module 404 guides the second pre-trained neural network to output the predicted switching frequency based on the second loss function of the second pre-trained neural network, it is specifically used for: obtaining the first load current, first duty cycle, and first voltage ripple of the multi-phase power supply in the actual business scenario; calculating the first conduction loss term based on the first load current and the first duty cycle; calculating the first turn-off loss term based on the first load current, the preset input voltage, and the first candidate switching frequency output by the second pre-trained neural network; summing the first conduction loss and the first turn-off loss to obtain the first switching loss term, which is used to quantify the power loss level at the first candidate switching frequency; determining the voltage ripple constraint term based on the difference between the first voltage ripple and the preset ripple allowable threshold; comparing and calculating the frequency constraint penalty term based on the candidate switching frequency and the preset switching frequency range; weighting and combining the first switching loss term, the voltage ripple constraint term, and the frequency constraint penalty term to obtain the second loss value of the second loss function; and adjusting the first candidate switching frequency according to the second loss value until the second loss value converges to obtain the predicted switching frequency.

[0123] As an optional implementation provided in this application, the switching frequency prediction module 404 is further configured to: for any business simulation scenario of different business simulation scenarios, collect the second load current, preset input voltage, second output voltage, and second voltage ripple, as well as the second voltage recovery time and second duty cycle under the target load operating point; calculate the second conduction loss term based on the second duty cycle and the second load current; calculate the second turn-off loss term based on the second load current, the preset input voltage, and the second candidate switching frequency output by the second pre-trained neural network for any business simulation scenario; sum the second conduction loss and the second turn-off loss to obtain the second switching loss term corresponding to any business simulation scenario; calculate the absolute value of the second voltage deviation rate for any business simulation scenario based on the second output voltage and the target supply voltage; weight the second switching loss term, the absolute value of the second voltage deviation rate, and the second voltage recovery time to obtain the weight optimization fitness of any business simulation scenario; calculate the sum of the weight optimization fitness of different business simulation scenarios; and adjust the network weights of the second pre-trained neural network based on the sum of the weight optimization fitness until the sum of the weight optimization fitness converges.

[0124] As an optional implementation provided in this application embodiment, the control module 405 is further configured to: obtain a first load pressure model of the current business scenario; the first load pressure model includes the load current ratio, current change rate and load stabilization time; determine a second load pressure model that matches the first load pressure model from the load pressure models of different business simulation scenarios; and control the multiphase power supply to operate according to the output and switching frequency corresponding to the second load pressure model under the current business scenario.

[0125] For a description of the features in the embodiment corresponding to the control device of the multiphase power supply, please refer to the relevant description of the embodiment corresponding to the control method of the multiphase power supply, which will not be repeated here.

[0126] like Figure 5 As shown, embodiments of this application also provide an electronic device, including a memory 501 and a processor 502. The memory 501 stores a computer program, and the processor 502 is configured to run the computer program to execute the steps in any of the above embodiments of the multiphase power supply control method.

[0127] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above embodiments of the control method for multiphase power supply.

[0128] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0129] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the control method for multiphase power supply.

[0130] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in any of the above embodiments of the multiphase power supply control method.

[0131] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0132] The foregoing has provided a detailed description of a multiphase power supply control method and electronic device. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only intended to help understand the method and core ideas of this application. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A control method for a multiphase power supply, characterized in that, include: Acquire dynamic response data of multiphase power supply output at target load operating point; Based on the dynamic response data corresponding to the target load operating point and the first pre-trained neural network, the output quantity is obtained; wherein, the output quantity includes a proportional coefficient, an integral coefficient and a differential coefficient, which are used to calculate the adjustment amount of the dynamic response data of the multi-phase power supply output, so as to control the on / off state of the power switch tube, so that the dynamic response data is adjusted to meet the preset standard. The actual load dynamic response indicators of the multiphase power supply are collected under actual business scenarios; the actual load dynamic response indicators include load current ratio, current change rate and load stabilization time. Based on the actual load dynamic response index and the second pre-trained neural network, the predicted switching frequency corresponding to the actual load dynamic response index is obtained; wherein, the second pre-trained neural network is trained based on the simulated load dynamic response index and corresponding switching frequency of multiphase power supply under different business simulation scenarios. The multiphase power supply is controlled based on the output quantity and the predicted switching frequency.

2. The method according to claim 1, characterized in that, The acquisition of dynamic response data output by the multiphase power supply at the target load operating point includes: Based on the pre-stored frequency sweep test results of the multiphase power supply, the dynamic response data corresponding to different load operating points are compared; the frequency sweep test results include the dynamic response data output by the multiphase power supply under different load operating points. Filter out the target load operating point where the overshoot or undershoot of the output voltage exceeds the preset range, and the corresponding dynamic response data of the target load operating point; Alternatively, the target load condition point whose voltage recovery time exceeds the response time threshold and the corresponding dynamic response data of the target load condition point can be selected.

3. The method according to claim 1, characterized in that, The step of obtaining the output quantity based on the dynamic response data of the target load condition point and the first pre-trained neural network includes: Input the dynamic response data corresponding to the target load condition point into the first pre-trained neural network; Based on the first loss function of the first pre-trained neural network, the first pre-trained neural network is guided to output the output quantity.

4. The method according to claim 3, characterized in that, The step of guiding the first pre-trained neural network to output the output based on the first loss function of the first pre-trained neural network includes: Obtain the first output voltage and the first voltage recovery time of the multiphase power supply under the target load condition. Calculate the absolute value of the first voltage deviation rate based on the first output voltage and the target supply voltage; The overshoot penalty term of the first loss function is determined based on the difference between the absolute value of the first voltage deviation rate and the voltage tolerance ratio; the voltage tolerance ratio is a preset tolerance range for power supply voltage fluctuations. The response time penalty term is determined based on the difference between the first voltage recovery time and the response time threshold. Based on the candidate values ​​output by the first pre-trained neural network, determine the parameter optimization terms; The overshoot penalty term, the response time penalty term, and the parameter optimization term are weighted and combined to calculate the first loss value of the first loss function; The candidate quantities are adjusted according to the first loss value until the first loss value converges, thereby obtaining the output quantity.

5. The method according to claim 1, characterized in that, After obtaining the output based on the dynamic response data corresponding to the target load operating point and the first pre-trained neural network, and before collecting the actual load dynamic response indicators of the multiphase power supply in the actual business scenario, the method further includes: Control the multiphase power supply to operate according to the output quantity, and collect the first dynamic response data output by the multiphase power supply under the target load condition point; If the first dynamic response data does not meet the preset standard, the weights of the first loss function of the first pre-trained neural network are adjusted until the first dynamic response data meets the preset standard.

6. The method according to claim 1, characterized in that, The step of obtaining the predicted switching frequency corresponding to the actual load dynamic response index based on the actual load dynamic response index and the second pre-trained neural network includes: The actual load dynamic response index is input into the second pre-trained neural network; Based on the second loss function of the second pre-trained neural network, the second pre-trained neural network is guided to output the predicted switching frequency.

7. The method according to claim 6, characterized in that, The second loss function based on the second pre-trained neural network guides the second pre-trained neural network to output the predicted switching frequency, including: The first load current, first duty cycle, and first voltage ripple of the multiphase power supply under the actual business scenario are obtained. Calculate the first conduction loss term based on the first load current and the first duty cycle; Calculate the first turn-off loss term based on the first load current, the preset input voltage, and the first candidate switching frequency output by the second pre-trained neural network. The first switching loss term is obtained by summing the first conduction loss and the first turn-off loss. The first switching loss term is used to quantify the power loss level at the first candidate switching frequency. The voltage ripple constraint term is determined based on the difference between the first voltage ripple and the preset ripple allowable threshold. Based on the candidate switching frequency and the preset switching frequency range, a frequency constraint penalty term is calculated by comparison. The first switching loss term, the voltage ripple constraint term, and the frequency constraint penalty term are weighted and combined to obtain the second loss value of the second loss function; The first candidate switching frequency is adjusted according to the second loss value until the second loss value converges, thereby obtaining the predicted switching frequency.

8. The method according to claim 1, characterized in that, After collecting the actual load dynamic response index of the multiphase power supply in the actual business scenario, and before obtaining the predicted switching frequency corresponding to the actual load dynamic response index based on the actual load dynamic response index and the second pre-trained neural network, the method further includes: For any of the different business simulation scenarios, the second load current, preset input voltage, second output voltage, and second voltage ripple are collected, as well as the second voltage recovery time and second duty cycle under the target load operating point. Calculate the second conduction loss term based on the second duty cycle and the second load current; The second turn-off loss term is calculated based on the second load current, the preset input voltage, and the second candidate switching frequency output by the second pre-trained neural network for any of the business simulation scenarios. The second conduction loss and the second turn-off loss are summed to obtain the second switching loss term corresponding to any of the service simulation scenarios. Calculate the absolute value of the second voltage deviation rate for any business simulation scenario based on the second output voltage and the target power supply voltage; The second switching loss term, the absolute value of the second voltage deviation rate, and the second voltage recovery time are weighted and combined to obtain the weighted optimization fitness of any business simulation scenario. Calculate the sum of the weighted optimization fitness of the different business simulation scenarios; The network weights of the second pre-trained neural network are adjusted according to the sum of the weight optimization fitness, until the sum of the weight optimization fitness converges.

9. The method according to claim 1, characterized in that, The method also includes: Obtain the first load stress model for the current business scenario; the first load stress model includes the load current ratio, current change rate, and load stabilization time. From the load stress models of the different business simulation scenarios, determine a second load stress model that matches the first load stress model; In the current business scenario, the multiphase power supply is controlled to operate according to the output and switching frequency corresponding to the second load pressure model.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the control method for a multiphase power supply as described in any one of claims 1 to 9 when executing the computer program.

Images