A Switching Power Supply Control Method and System Based on Dynamic Load Tracking

By employing a dynamic load tracking control method, the problems of parameter uncertainty and sensor anomalies in multiphase interleaved parallel buck switching power supplies under dynamic load conditions are solved, achieving accuracy and stability in current distribution and voltage regulation, and forming a closed-loop parameter correction mechanism.

CN123122243APending Publication Date: 2026-11-17XIAN SAIERCOM CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611621209.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The circuit parameters of multiphase interleaved parallel buck switching power supplies are difficult to balance dynamic load event detection, parameter uncertainty constraints, common output voltage regulation, and current distribution of each phase branch under changing operating conditions and rapid load current changes. They are also easily affected by erroneous feedback from drive saturation or sensor malfunctions.

Method used

A control method based on dynamic load tracking is adopted. By synchronously sampling the input voltage, common output voltage, inductor current of each phase branch, and load current of a multi-phase interleaved parallel buck switching power supply, fast event detection data and slow parameter identification data are generated. Online parameter status records are established, predictive constraint margin and differential mode duty cycle adjustment are generated, parameter confidence domain conversion and energy regulation are performed, and the event feedback correction module is combined to handle model mismatch and sensor anomalies.

Benefits of technology

It effectively handles parameter uncertainties and sensor anomalies under dynamic load conditions, ensures the accuracy of current distribution and voltage regulation, reduces response deviations caused by model mismatch and drive saturation, and forms a closed loop of input data, control results and parameter correction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122243A_ABST
    Figure CN123122243A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of digital control technology for switching power supplies, specifically disclosing a switching power supply control method and system based on dynamic load tracking. The method includes: synchronously sampling the input voltage, common output voltage, inductor current of each phase branch, and load current of a multi-phase interleaved parallel buck switching power supply, aligning the case temperature of the power switch transistors in each phase branch, dividing the sampled data into fast event detection data and slow parameter identification data, identifying the parameters of each phase branch and the parameters of the common output node based on the slow parameter identification data, and generating an online parameter status record. The technical problem this invention aims to solve is how to balance dynamic load event detection, parameter uncertainty constraints, common output voltage regulation, and current distribution in each phase branch under the condition that the circuit parameters of a multi-phase interleaved parallel buck switching power supply change with operating conditions and the load current changes rapidly, while limiting erroneous feedback from drive saturation or sensor malfunctions to online parameter identification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of digital control technology for switching power supplies, and specifically to a switching power supply control method and system based on dynamic load tracking. Background Technology

[0002] Multiphase interleaved parallel buck switching power supplies typically utilize multiple phase branches to supply power to a common output terminal. Phase interleaving reduces current ripple at the common output terminal, and adjusting the duty cycle of each phase branch maintains the voltage distribution at the common output terminal and the current distribution across phase branches. During operation, the inductance parameters of each phase branch, the equivalent conduction loss resistance of each phase branch, the parameters of the common output filter capacitor, and the equivalent series resistance of the common output filter capacitor are affected by device variability, temperature variations, and operating time. Control models established with fixed parameters may gradually deviate from the actual circuit state. Low-pass filtering of synchronously sampled data can suppress switching ripple, but it increases the detection delay of dynamic load events; directly using unfiltered data to identify circuit parameters introduces switching noise and sampling errors into parameter updates. Some control methods only use parameter estimates, failing to propagate the uncertainty represented by the parameter covariance to duty cycle, current, and temperature constraints. The common output voltage regulation and the current distribution in each phase branch have different control objectives. When the inductance parameters of each phase branch are inconsistent, if the current sharing regulation with a zero algebraic sum is superimposed on the common mode duty cycle, the total rate of change of the inductor current in each phase branch may still be affected by the differential mode control. The deviation between the actual response and the model prediction results may also be caused by sensor anomalies, duty cycle boundaries, current constraints, or temperature constraints. If the online parameters are adjusted directly without distinguishing the source of the deviation, the parameter estimates may deviate from the actual circuit state. Therefore, it is necessary to establish a continuous control relationship between sampling data splitting, online parameter identification, parameter uncertainty propagation, common mode control, differential mode constraint allocation, and event feedback correction. Summary of the Invention

[0003] This invention provides a switching power supply control method and system based on dynamic load tracking. The technical problem to be solved by this invention is how to balance dynamic load event detection, parameter uncertainty constraints, common output voltage regulation and current distribution of each phase branch under the condition that the circuit parameters of a multi-phase interleaved parallel buck switching power supply change with the operating conditions and the load current changes rapidly, while limiting the erroneous feedback caused by drive saturation or sensor abnormalities in the identification of online parameters.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The switching power supply control method based on dynamic load tracking includes: synchronously sampling the input voltage, common output voltage, inductor current of each phase branch, and load current of a multi-phase interleaved parallel buck switching power supply, aligning the case temperature of the power switch transistors in each phase branch, dividing the sampled data into fast event detection data and slow parameter identification data, identifying the parameters of each phase branch and the parameters of the common output node based on the slow parameter identification data, and generating online parameter status records; establishing a physical model and parameter confidence domain based on the online parameter status records, converting the parameter confidence domain into predictive constraint margins, generating a load prediction sequence based on the fast event detection data, and solving the common mode energy regulation duty cycle under the charge balance constraint of the common output filter capacitor. The process involves comparing component and model prediction results; generating the original differential mode duty cycle adjustment amount based on the inductor current of each phase branch, the estimated junction temperature of the power switch, the duty cycle component of the common mode energy regulation, and the model prediction results. This amount is then projected onto a feasible region that satisfies the inductor parameter weighted zero-sum constraint and control constraints, generating the duty cycle component of the differential mode current redistribution for each phase. After quantization and verification, the final duty cycle and drive signal are generated. Based on the actual response data corresponding to the final duty cycle, the model prediction results, the drive saturation state, and the effective state of the sensors, an event-level model mismatch index and event cause category are generated. Based on the event cause category, a bounded identification state adjustment amount is generated, and an online parameter state record used in subsequent control cycles is generated.

[0005] In one aspect of the invention, synchronous sampling and generation of fast event detection data and slow parameter identification data includes: The sample-and-hold circuits that simultaneously trigger the input voltage, common output voltage, inductor current of each phase branch, and load current using a unified sampling clock are used to add a unified timestamp to the synchronous sampling data and align the case temperature of the power switch transistors of each phase branch with the synchronous sampling data according to the timestamp. The synchronous sampling data is sent into the fast event detection path. Local linear fitting is performed on the continuous load current sampling values. The load current change rate, load current increment, event direction, event occurrence time and event valid marker are determined based on the fitting slope and the difference between the fitting endpoints, and fast event detection data is generated. The synchronous sampling data is sent into the slow parameter identification path, and the switching frequency component is suppressed by the linear phase finite length unit impulse response low-pass filter. Based on the effective timestamp of the group delay correction filtering result of the low-pass filter, the slow parameter identification data is generated.

[0006] In one aspect of the invention, identifying the parameters of each phase branch and the parameters of the common output node includes: Align the slow parameter identification data with the historical final duty cycle that has actually been executed at the corresponding time. Construct regression data of each phase branch parameter based on the input voltage, historical final duty cycle, common output voltage, inductor current of each phase branch and its rate of change, and identify the inductor parameters of each phase branch and the equivalent conduction loss resistance of each phase branch. The current of the common output filter capacitor is determined based on the sum of the inductor currents of all phase branches, the load current, and the voltage at the common output terminal. The equivalent series resistance of the common output filter capacitor and the parameters of the common output filter capacitor are identified by using a short data window that includes the edge of load change and a long data window that includes the change of charge accumulation, respectively. A normalized information matrix is ​​constructed based on the slow parameter identification data. When the normalized information matrix satisfies the continuous excitation condition and the sensor is in an effective state, a recursive least squares parameter update is performed. When the update condition is not met, the previous effective parameter estimate is maintained and the corresponding parameter covariance is increased to generate an online parameter status record.

[0007] In one aspect of the invention, establishing a parameter confidence domain and converting it into a prediction constraint margin includes: The parameter vector and parameter covariance are normalized in terms of dimensions based on the estimated values ​​of each parameter. The normalized parameter covariance is decomposed, and boundary parameter points are generated along the positive and negative directions of each principal axis of covariance. The boundary parameter points are then converted into parameter confidence regions with the original parameter units. The nominal parameter points and each boundary parameter point are used to predict the estimated values ​​of the common output voltage, the inductor current of each phase branch and the junction temperature of the power switch of each phase branch in the future control time domain. The prediction deviations corresponding to each covariance principal axis are then used to form a joint prediction deviation envelope. The duty cycle prediction constraint margin is generated based on the joint prediction deviation of the common output voltage and the duty cycle sensitivity; the current prediction constraint margin is generated based on the joint prediction deviation of the inductor current of each phase branch; and the temperature prediction constraint margin is generated based on the joint prediction deviation of the junction temperature and the thermal model calibration residual.

[0008] In one aspect of the invention, solving for the common-mode energy regulation duty cycle component and model prediction results includes: Based on the final duty cycle, input voltage, common output voltage, parameters of each phase branch, and parameters of the common output node actually executed in the previous control cycle, an average physical model of the multiphase interleaved parallel buck switching power supply is established, and a junction temperature prediction relationship is established based on the case temperature of the power switch tubes in each phase branch, switching loss data, and thermal impedance data. A load prediction sequence is generated based on the load current change rate, load current increment, event direction, and event occurrence time in the fast event detection data. The common output filter capacitor charge demand is accumulated cycle by cycle based on the difference between the load prediction sequence and the predicted inductor current values ​​of all phase branches. Under the constraints of charge balance of the common output filter capacitor, duty cycle constraint after deducting the prediction constraint margin, current constraint, and temperature constraint, the common mode energy regulation duty cycle component is solved, and the predicted sequence of common output terminal voltage, predicted sequence of inductor current of each phase branch, predicted sequence of common output filter capacitor current, predicted sequence of junction temperature of power switch of each phase branch, predicted recovery time, and constraint activation state are output.

[0009] In one aspect of the invention, generating the original differential mode duty cycle adjustment amount includes: The total target current is determined based on the load prediction sequence. The available current margin is determined based on the difference between the upper limit of the inductor current of each phase branch and the current inductor current of each phase branch. The available temperature margin is determined based on the difference between the upper limit of the junction temperature of the power switch and the current estimated junction temperature of the power switch. The target distribution current of each phase branch is generated based on the available current margin and the available temperature margin. A current deviation control component is generated based on the deviation between the inductor current of each phase branch and the corresponding target distribution current. A temperature deviation control component is generated based on the deviation between the estimated junction temperature of each phase branch power switch and the estimated average junction temperature. When the drive saturation state is effective, the integral state of the current deviation control component is corrected based on the actual execution duty cycle. The current deviation control component and temperature deviation control component of the same phase branch are combined to generate the original differential mode duty cycle adjustment of each phase branch.

[0010] In one aspect of the invention, projecting the original differential mode duty cycle adjustment amount onto the feasible region includes: Based on the inductance parameters of each phase branch in the online parameter status record, a weighted zero-sum constraint for inductance parameters is established, and the feasible region is constructed based on the common mode energy regulation duty cycle component, the prediction constraint margin, the final duty cycle boundary, the upper limit of the inductance current of each phase branch, the upper limit of the junction temperature of the power switch tube, and the boundary of the differential mode duty cycle component. With the minimum weighted deviation between the original differential mode duty cycle adjustment and the current redistribution duty cycle components of each phase differential mode as the projection objective, the current redistribution duty cycle components of each phase differential mode are solved within the feasible region. When the feasible region is empty, the common mode energy regulation duty cycle component is first moved to the feasible boundary closest to the original solution value and reprojected. When there is still no feasible solution after two consecutive control cycles, the duty cycle component of the differential mode current of each phase is decayed cycle by cycle and the common mode energy regulation duty cycle component is reduced when there is still a risk of overcurrent or overtemperature after decaying to zero.

[0011] In one aspect of the invention, generating the final duty cycle and drive signal includes: The common mode energy regulation duty cycle component is combined with the differential mode current redistribution duty cycle component of each phase to generate the target duty cycle of each phase branch. The duty cycle is quantized according to the pulse width modulation time resolution. The quantized duty cycle is then used to re-examine the inductor parameter weighted zero-sum constraint and control constraint to generate the final duty cycle of each phase branch. The interleaved phase between the drive signals of each phase branch is determined according to the number of phase branches. The pulse width modulation signal of each phase branch is generated based on the final duty cycle, switching frequency, minimum on time, minimum off time and dead time of each phase branch, and the final duty cycle of the actual execution is recorded. The system detects whether the final duty cycle has reached the duty cycle boundary and whether the current constraint, temperature constraint, differential mode duty cycle component boundary, or solution backoff condition has been activated. Based on the detection results, it generates the drive saturation state.

[0012] In one aspect of the invention, generating an online parameter status record for subsequent control cycles includes: Based on the event occurrence time in the fast event detection data, the actual response data after the final duty cycle operation is time-aligned with the model prediction results. An event-level model mismatch index is generated based on the deviation of the minimum value of the common output voltage, the recovery time deviation, the peak value deviation of the inductor current of each phase branch, the steady-state current distribution deviation, and the net charge deviation of the common output filter capacitor. Following the priority order of sensor anomaly, drive saturation, model mismatch, and normal response, event cause categories are generated based on sensor effective state, drive saturation state, and event-level model mismatch index. The parameter group to be adjusted is determined based on the contribution of each deviation that constitutes the event-level model mismatch index. Based on the event cause category and the parameter group to be adjusted, the forgetting factor, parameter covariance, parameter estimation correction boundary, and parameter freeze state used in recursive identification are adjusted to generate the bounded identification state adjustment amount. The adjusted data is then written into the online parameter state record used in the next control cycle after the bounded identification state adjustment amount is completed.

[0013] In another aspect, the present invention also relates to a switching power supply control system based on dynamic load tracking, comprising: The synchronous sampling and online parameter identification module is used to synchronously sample the input voltage, common output voltage, inductor current and load current of each phase branch of the multi-phase interleaved parallel buck switching power supply, align the case temperature of the power switch tubes of each phase branch, divide the sampled data into fast event detection data and slow parameter identification data, identify the parameters of each phase branch and the parameters of the common output node, and generate online parameter status records. The common mode prediction control module is used to establish a physical model and parameter confidence domain based on online parameter status records, convert the parameter confidence domain into prediction constraint margin, generate a load prediction sequence based on fast event detection data, and generate common mode energy regulation duty cycle components and model prediction results under the common output filter capacitor charge balance constraint. The differential mode constraint allocation and drive module is used to generate the original differential mode duty cycle adjustment amount based on the inductor current of each phase branch, the estimated junction temperature of the power switch, the duty cycle component of the common mode energy regulation and the model prediction results. It projects the original differential mode duty cycle adjustment amount to the feasible region that satisfies the weighted zero-sum constraint of the inductor parameters and the control constraint. After duty cycle quantization and constraint verification, it generates the duty cycle component of the differential mode current redistribution of each phase, the final duty cycle and the drive signal. The event feedback correction module is used to generate an event-level model mismatch index and event cause category based on the actual response data corresponding to the final duty cycle, model prediction results, drive saturation state and sensor effective state. It also generates a bounded identification state adjustment amount based on the event cause category and generates online parameter state records for subsequent control cycles.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention processes synchronous electrical quantity sampling data into fast event detection data and slow parameter identification data, enabling dynamic load event detection to bypass slow low-pass filtering. Simultaneously, the filtered data is used to identify parameters of each phase branch and the common output node. The online parameter status record simultaneously stores parameter estimates and parameter covariance, allowing the common mode predictive control module to generate duty cycle, current, and temperature prediction constraint margins based on the current parameter uncertainty. The common mode energy regulation duty cycle component is generated under the charge balance constraint of the common output filter capacitor and is used to adjust the total energy change of all phase branches. The differential mode duty cycle component is weighted and projected according to the online inductance parameters, used to adjust the current distribution among phase branches without changing the trend of the first-order total inductance current. The event feedback correction module aligns the actual response with the model prediction results and, combined with the effective sensor state and drive saturation state, determines the source of deviation, handling response deviations caused by model mismatch and those caused by sensor anomalies or limited control capabilities separately. The online parameter status record generated in S4 is reread in subsequent control cycles, forming a closed loop between input data, control results, actual response, and parameter correction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the switching power supply control method based on dynamic load tracking of the present invention.

[0017] Figure 2 This is a flowchart of a sub-step in step 1 of the switching power supply control method based on dynamic load tracking of the present invention.

[0018] Figure 3 This is a flowchart of the sub-step of step 2 in the switching power supply control method based on dynamic load tracking of the present invention.

[0019] Figure 4 This is a flowchart of the sub-step of step 3 in the switching power supply control method based on dynamic load tracking of the present invention.

[0020] Figure 5 This is a flowchart of step 4 in the switching power supply control method based on dynamic load tracking of the present invention.

[0021] Figure 6 This is a framework diagram of the switching power supply control system based on dynamic load tracking of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to embodiments. These embodiments are only a part of the embodiments of the present invention, and not all of them. Adjustments to device specifications, parameter ranges, and execution cycles made by those skilled in the art without departing from the technical principles of the present invention can be applied to the present invention.

[0023] This invention discloses a switching power supply control method and system based on dynamic load tracking. The switching power supply is a multi-phase interleaved parallel buck switching power supply, including an input power supply, at least two parallel phase branches, a common output node, a common output filter capacitor, and a load. Each phase branch is a synchronous buck phase branch, including an upper-arm power switch, a lower-arm power switch, inductors for each phase branch, current sampling circuits for each phase branch, and a case temperature sampling circuit for the power switches. Each phase branch is connected between the input power supply and the common output node, and the common output filter capacitor is connected between the common output node and reference ground.

[0024] This specification refers to the inductor in each phase branch as the phase branch inductor, and the current flowing through the phase branch inductor as the phase branch inductor current. The equivalent loss parameter formed by the on-resistance of the power switch, the winding resistance of the phase branch inductor, the printed circuit board trace resistance, and the connector contact resistance at the current operating point is called the equivalent on-resistance resistance of the phase branch. The equivalent on-resistance resistance of each phase branch remains constant within a predicted time domain and is redefined after subsequent parameter updates.

[0025] The capacitance of the common output filter capacitor is called the common output filter capacitor parameter, and the series loss parameter corresponding to the common output filter capacitor is called the equivalent series resistance of the common output filter capacitor. The parameters of the common output node include both the common output filter capacitor parameter and the equivalent series resistance of the common output filter capacitor.

[0026] The case temperature of the power switches in each phase branch is obtained directly through a temperature sampling circuit. The estimated junction temperature of the power switches in each phase branch is calculated based on the case temperature, conduction loss, switching loss, and thermal impedance data of the power switches. When a phase branch has both upper and lower bridge arm power switches, the larger of the two estimated junction temperatures is taken as the estimated junction temperature of the power switches in that phase branch.

[0027] Fast event detection data is an event feature record generated from load current samples without slow low-pass filtering. It includes the load current change rate, load current increment, event direction, event occurrence time, and event validity marker. The load current change rate is in A / μs, the load current increment is in A, the event direction is either load increase or load decrease, and the event validity marker is a Boolean type. Fast event detection data is generated in S1 and used in S2 to generate the load prediction sequence.

[0028] Slow parameter identification data is a data sequence obtained by suppressing switching ripple and correcting group delay timestamps on synchronously sampled data. The slow parameter identification data retains the original units of each sampled physical quantity and is generated in S1 for identifying the parameters of each phase branch and the parameters of the common output node.

[0029] The online parameter status record is a structured data record with a unified time base, including estimated inductance parameters for each phase branch, estimated equivalent conduction loss resistance for each phase branch, estimated parameters for the common output filter capacitor, estimated equivalent series resistance of the common output filter capacitor, corresponding parameter covariance, estimated junction temperature of the power switches for each phase branch, valid identification markers, drive saturation state, valid status of each channel sensor, historical actual duty cycle, and valid timestamps. The online parameter status record is generated in S1 or S4 and used by S2 and subsequent parameter identification processes.

[0030] The confidence region of a parameter is the range of possible values ​​for the parameter, generated from the parameter estimate, parameter covariance, and confidence coefficient. It is represented by a set of nominal parameter points and boundary parameter points in units of the original parameter, and is used in S2 to predict state deviations caused by parameter uncertainty.

[0031] The prediction constraint margins include duty cycle prediction constraint margin, current prediction constraint margin, and temperature prediction constraint margin. The duty cycle prediction constraint margin is a dimensionless value, the current prediction constraint margin is in amperes (A), and the temperature prediction constraint margin is in degrees Celsius (°C). The prediction constraint margins are generated based on the joint prediction bias obtained from parameter confidence domain propagation and are used in S2 and S3 to tighten the corresponding control constraints.

[0032] The common-mode energy regulation duty cycle component is a dimensionless absolute duty cycle base value applied to all phase branches within the current control cycle. It is determined by the common-mode duty cycle actually executed in the previous control cycle, the model predictive control correction, and the bounded load feedforward. The common-mode energy regulation duty cycle component is used to regulate the overall trend of inductor current variation in all phase branches.

[0033] The original differential mode duty cycle adjustment is a dimensionless duty cycle vector generated based on the current deviation of each phase branch and the junction temperature deviation of the power switch transistors, and has not yet been projected by control constraints. The current redistribution duty cycle component of each phase is a dimensionless duty cycle vector obtained by projecting the original differential mode duty cycle adjustment after applying inductor parameter weighted zero-sum constraints and control constraints. It is used to adjust the current distribution between each phase branch.

[0034] The model prediction results are structured prediction records generated by S2, including the predicted sequence of the common output voltage, the predicted sequence of the inductor current of each phase branch, the predicted sequence of the common output filter capacitor current, the predicted sequence of the junction temperature of the power switches of each phase branch, the predicted recovery time, and the constraint activation state in the future control time domain. The model prediction results are used by S3 to construct the feasible region and by S4 to compare with the actual response data.

[0035] The event-level model mismatch index is the dimensionless comprehensive deviation between the actual response and the model prediction in a single dynamic load event. Event cause categories include sensor anomaly, drive saturation, model mismatch, and normal response. The bounded identification state adjustment is an identification control record generated based on the event cause category and the contributions of various deviations, including the forgetting factor target value, parameter covariance amplification factor, parameter estimation correction boundary, and parameter freeze state.

[0036] like Figure 1 As shown, in some embodiments, the control method is executed by a synchronous sampling and online parameter identification module, a common mode prediction control module, a differential mode constraint allocation and driving module, and an event feedback correction module, including the following four steps.

[0037] S1. Synchronously sample the input voltage, common output voltage, inductor current and load current of each phase branch, align the case temperature of the power switch tube of each phase branch according to the timestamp, process the sampled data into fast event detection data and slow parameter identification data respectively, use the slow parameter identification data to identify the parameters of each phase branch and the parameters of the common output node respectively, and generate online parameter status records.

[0038] S2. Read the online parameter status record generated by S1, establish the average physical model of the multiphase interleaved parallel buck switching power supply based on the parameter estimate, establish the parameter confidence domain and generate the prediction constraint margin based on the parameter covariance, generate the load prediction sequence based on the fast event detection data, and generate the common mode energy regulation duty cycle component and model prediction results under the common output filter capacitor charge balance constraint and the tightened control constraint.

[0039] S3. Read the common mode energy regulation duty cycle component and model prediction results generated in S2. Generate the original differential mode duty cycle regulation amount based on the inductor current of each phase branch and the estimated junction temperature of the power switch tube of each phase branch. Project the original differential mode duty cycle regulation amount to the feasible region that satisfies the inductor parameter weighted zero-sum constraint, duty cycle constraint, current constraint and temperature constraint. Generate the differential mode current redistribution duty cycle component of each phase. After duty cycle quantization and constraint verification, generate the final duty cycle of each phase branch and the corresponding drive signal.

[0040] S4. Obtain the actual response data after running according to the final duty cycle generated in S3. Time-align the actual response data with the model prediction results generated in S2. Generate the event-level model mismatch index and event cause category. Generate the bounded identification state adjustment amount based on the event cause category and the contribution of each deviation. Write the adjustment result into the online parameter status record used in the next control cycle after the bounded identification state adjustment amount is completed.

[0041] In this embodiment, it should be noted that in S1, electrical quantity sampling and temperature acquisition use a unified time base, but the temperature sampling channel and the electrical quantity sampling channel are not required to have the same sampling frequency. Input voltage, common output voltage, inductor current of each phase branch, and load current are simultaneously latched by the same hardware trigger signal. The case temperature of the power switch transistor in each phase branch is written to a timestamp according to its own sampling time and aligned with the most recent synchronous electrical quantity sampling record. The fast event detection path directly uses the load current sample value without slow low-pass filtering to shorten the load change recognition time. The slow parameter identification path suppresses the switching frequency component through low-pass filtering and maps the filtered output to the center time of the filtering window. The online parameter status record output by S1 simultaneously includes parameter estimates and parameter covariance, enabling S2 to handle nominal parameter status and parameter uncertainty separately.

[0042] In S2, the average physical model is used to describe the changes in inductor current, common output filter capacitor current, and common output terminal voltage near the current operating point. The parameter confidence region is generated from the parameter estimates and parameter covariance in the online parameter state record. Different parameters have different units, so dimensional normalization is performed before parameter covariance decomposition. Nominal parameter points and boundary parameter points are respectively entered into the future state prediction process, and the prediction biases corresponding to the principal axes of each covariance form a joint prediction bias envelope. The joint prediction bias of the common output terminal voltage is converted into a duty cycle prediction constraint margin through local duty cycle sensitivity, and the joint prediction biases of current and junction temperature form current prediction constraint margins and temperature prediction constraint margins, respectively. The model prediction results generated by S2 contain all the prediction sequences required for S3 to construct the feasible region and S4 to execute event feedback.

[0043] In S3, the common-mode energy regulation duty cycle component is used to determine the energy change shared by all phase branches, while the original differential-mode duty cycle regulation is used to determine the current migration direction between each phase branch. When the inductance parameters of each phase branch differ, the differential-mode duty cycle with a normal algebraic sum of zero cannot make its first-order action on the rate of change of the total inductor current zero. Therefore, S3 establishes weighted zero-sum constraints based on the inductance parameters of each phase branch obtained through online identification, and incorporates the duty cycle prediction constraint margin, current prediction constraint margin, and temperature prediction constraint margin into the feasible region. The continuous solution results are quantized using pulse width modulation and then the constraint check is re-executed. Parameter identification and event feedback use the final duty cycle actually executed after quantization.

[0044] In S4, the event feedback correction module does not immediately modify parameters after an event occurs. The module first collects the actual response data within the complete event evaluation window, then aligns it with the model prediction results based on a unified timestamp. Event cause categories are determined in the order of sensor malfunction, drive saturation, model mismatch, and normal response. Output deviations caused by sensor malfunction and drive saturation are not used to directly correct circuit parameter estimates. When the event cause category is model mismatch, the parameter group to be adjusted is determined based on the contribution of each deviation to the event-level model mismatch index. The updated online parameter status record takes effect from the first control cycle after S43 is completed.

[0045] like Figure 2 As shown, in some embodiments, S1 includes: S11 is used to complete the synchronous sampling of electrical quantities, time alignment of shell temperature, and generation of fast event detection data and slow parameter identification data.

[0046] S12 is used to identify the parameters of each phase branch and the parameters of the common output node, and to update the control parameters according to the continuous excitation conditions and the effective state of the sensor.

[0047] S13 is used to perform parameter physical boundary checks, parameter covariance updates, and data encapsulation to generate online parameter status records.

[0048] like Figure 3 As shown, in some embodiments, S2 includes: S21 is used to establish an average physical model, parameter confidence domain, and prediction constraint margin based on online parameter status records.

[0049] S22 is used to generate a load prediction sequence based on fast event detection data and determine the charge requirement of the common output filter capacitor in the future control time domain.

[0050] S23 is used to solve the common mode energy regulation duty cycle component under the common output filter capacitor charge balance constraint and the tightened control constraint, and to generate the model prediction results simultaneously.

[0051] like Figure 4 As shown, in some embodiments, S3 includes: S31 is used to generate the original differential mode duty cycle adjustment amount based on the current deviation of each phase branch and the junction temperature deviation of the power switch tube.

[0052] S32 is used to construct a feasible region that includes inductance parameter weighted zero-sum constraints and control constraints, and to generate the duty cycle components of the differential mode current redistribution for each phase.

[0053] S33 is used to combine the common mode and differential mode duty cycle components, perform duty cycle quantization, constraint verification and interleaved pulse width modulation, and generate the final duty cycle, drive signal and drive saturation state.

[0054] like Figure 5 As shown, in some embodiments, S4 includes: S41 is used to align the actual response data with the model prediction results in time and generate an event-level model mismatch index.

[0055] S42 is used to generate event cause categories based on sensor effective state, drive saturation state and event-level model mismatch index, and to determine the parameter group to be adjusted.

[0056] S43 is used to generate bounded identification state adjustment quantities based on the event cause category, and to generate online parameter state records for use in subsequent control cycles.

[0057] In some embodiments, S1 is described in detail.

[0058] S11. Using a unified sampling clock, the sample-and-hold circuits of input voltage, common output voltage, inductor current of each phase branch, and load current are simultaneously triggered to add a unified timestamp to the synchronous sampling data. The case temperature of the power switch tube of each phase branch is aligned with the synchronous sampling data according to the timestamp, and the synchronous sampling data is sent to the fast event detection path and the slow parameter identification path respectively.

[0059] In this embodiment, the input voltage, common output voltage, inductor current of each phase branch, and load current sampling channels are each equipped with independent sample-and-hold circuits, latched by the same hardware trigger signal, instead of using a sequential polling analog-to-digital conversion method. The channel start-up time difference is determined by inputting the same square wave edge to all channels and comparing the latching times of each channel. In calibration data containing 1000 triggers, the maximum channel start-up time difference is 18 ns, therefore the allowable upper limit is set to 20 ns. This upper limit accounts for 1% of the 2 μs switching cycle corresponding to the 500 kHz switching frequency, and is used to limit the channel time deviation in the calculation of the phase branch current change rate.

[0060] This embodiment employs a 2.5MHz electrical quantity sampling frequency and a 12-bit analog-to-digital converter (ADC). The 2.5MHz sampling frequency allows for five sampling moments per 500kHz switching cycle, providing continuous sample values ​​for rapid event detection. The current sampling circuit uses a 10mΩ sampling resistor and a differential current sampling amplifier with a gain of 20. The rated power of the sampling resistor is no less than 3W. This rated power is determined based on the upper limit of the inductor current in each phase branch of 12A and the short-time inductor saturation current of 15A. At 12A current, the sampling resistor dissipates 1.44W, the voltage across the sampling resistor is 120mV, and the amplified voltage is 2.4V, which is within the input range of the 3.3V ADC. At a short-time current of 15A, the sampling resistor dissipates 2.25W, which does not exceed the 3W rated power.

[0061] The fast event detection path stores the five most recent load current samples. Using the sampling timestamp as the independent variable and the load current as the dependent variable, it performs local linear fitting, taking the fitting slope as the load current change rate and the difference between the first and last timestamps of the fitted line as the load current increment. When the absolute value of the load current change rate reaches the dynamic load event judgment threshold for two consecutive sliding windows, the end time of the first window reaching the threshold is recorded as the event occurrence time. The condition of two consecutive windows is derived from 300 sets of steady-state, switching ripple, and load change calibration data. Single-window judgment is susceptible to occasional false triggering by switching noise, and three consecutive windows increase the recognition time by more than 0.4 μs; therefore, two consecutive windows are used.

[0062] The dynamic load event determination threshold is set at 0.8 A / μs. This threshold is determined based on 300 sets of finite historical calibration data. In the calibration data, the 99.5th percentile of the absolute value of the load current change rate under non-load event conditions is 0.53 A / μs. After adding sampling noise and temperature drift margin to this value, the threshold is set to 0.8 A / μs. For example, when five load current sampling values ​​at 0.4 μs intervals are 5.0 A, 5.3 A, 5.7 A, 6.1 A, and 6.5 A, the change rate obtained by local linear fitting is approximately 0.94 A / μs, which exceeds the threshold, thus generating a load increase event.

[0063] The slow parameter identification path employs a linear-phase finite-length unit impulse response low-pass filter with 64 taps. The filter coefficients are generated using a window function design method, with the DC gain normalized to 1. The amplitude-frequency response at the 50kHz passband boundary and the 500kHz switching frequency is calculated before deployment. In a set of filter design data, with 64 taps, the attenuation at 500kHz is 64.7dB, meeting the design requirement of at least 60dB. Reducing the number of taps to 48 fails to consistently achieve 60dB attenuation at the switching frequency; therefore, 64 taps are used. This filter is of order 63, with a group delay of 31.5 sampling periods, corresponding to 12.6μs at a 2.5MHz sampling frequency. The module corrects the effective timestamp of the filtering result to the center time of the filtering window, but this timestamp correction is not used for fast event detection.

[0064] S12. Based on the slow parameter identification data, construct the regression data of each phase branch parameter and the regression data of the common output node parameter respectively. When the normalized information matrix satisfies the continuous excitation condition and the sensor is in an effective state, perform recursive least squares parameter update.

[0065] In this implementation, the parameter identification for each phase branch first maps the slow parameter identification data to the historical final duty cycle that has already been executed, according to the corrected timestamp. Then, local linear fitting is performed on the five most recent filtered inductor current samples for each phase branch, with the fitting slope used as the rate of change of the inductor current for each phase branch. The module uses the input voltage, historical actual execution duty cycle, common output voltage, inductor current of each phase branch, and its rate of change to form the current regression vector and observations, and sequentially performs predicted value calculation, regression residual calculation, recursive gain update, parameter estimate update, parameter covariance update, and physical boundary check.

[0066] The inductance parameters and equivalent conduction loss resistance of each phase branch are stored separately, without assuming that all phase branches have the same parameters. Initial parameter values ​​are taken from the device's nominal value or measurements taken under shutdown conditions, and the initial parameter covariance is determined based on the device tolerance. For example, if the nominal inductance parameter is 20 μH and the device tolerance is 20%, the initial estimate is set to 20 μH, the initial standard deviation is set to 2 μH, and the allowable physical range is set to 14 μH to 26 μH based on the device tolerance and the calibration measurement range. Updates exceeding this physical range are rejected, and the previous valid parameter estimate is retained.

[0067] The common output node parameter identification first uses the sum of the inductor currents of all phase branches minus the synchronously sampled load current to obtain the common output filter capacitor current. The equivalent series resistance identification of the common output filter capacitor uses a short data window that includes the edge of the load current change, comparing the rapid change in the common output terminal voltage with the rapid change in the common output filter capacitor current. The common output filter capacitor parameter identification uses a long data window that includes the cumulative change in charge, comparing the charge change formed by the cumulative change in the common output filter capacitor current over time with the low-frequency change in the common output terminal voltage.

[0068] In this embodiment, the equivalent series resistance identification window contains 16 slow parameter identification data points, and the common output filter capacitor parameter identification window contains 128 slow parameter identification data points. The window length is determined based on 200 sets of load step and steady-state change calibration data. When there are fewer than 16 data points, the equivalent series resistance estimation is more sensitive to single-point noise; when there are more than 16 data points, the voltage change inside the common output filter capacitor begins to enter the rapid change window. 128 data points are used for the common output filter capacitor parameter identification because the charge accumulation corresponding to 64 data points is insufficient under some small load change conditions, and using 256 data points would result in an update delay exceeding the set parameter identification cycle.

[0069] Before constructing the normalized information matrix, the module calculates the root mean square (RMS) value of each regression variable in the most recent 64 data points, and uses the larger of the RMS value and the minimum normalized baseline value determined based on sampling noise as the normalization scale. The window length is determined based on the comparison results of 32, 64, and 128 data windows. In the calibration data, the minimum eigenvalue of the 32 data window fluctuates significantly, the parameter update delay of the 128 data window is relatively long, and the 64 data window balances the stability of the minimum eigenvalue and the 20μs update time limit; therefore, 64 data points are used.

[0070] The continuous excitation threshold is set to 0.015. This threshold is derived from 200 sets of load current variation, input voltage variation, and steady-state operation calibration data. At a threshold of 0.015, the parameter convergence time does not exceed 12ms, and the 95th percentile of the steady-state parameter relative error does not exceed 5%. When the minimum eigenvalue of a normalized information matrix is ​​0.018, parameter updates are allowed; when the minimum eigenvalue is 0.011, the previous effective parameter estimate is retained, and the parameter covariance is increased according to the process noise.

[0071] The effective state of the sensor is determined based on range checks, rate of change checks, and redundancy checks. The residual threshold for inductor current sampling in each phase branch is set to 0.35A. This threshold is determined based on the combined standard deviation of sampling resistor error, amplifier gain error, analog-to-digital converter quantization error, time alignment error, and switching noise. In a set of calibration data, the combined standard deviation corresponding to the above errors is approximately 0.056A; multiplying this by 6 and rounding up yields 0.35A. The residual threshold for common output voltage sampling is set to 60mV. This threshold is determined based on the combined standard deviation of voltage divider error, quantization error, time alignment error, and filter residual; this combined standard deviation is approximately 10mV; multiplying this by 6 yields 60mV.

[0072] S13. Perform physical boundary checks on the parameter estimates, update the parameter covariance and identify valid markers based on the parameter regression residuals, and write the parameter estimates, parameter covariance, power switch junction temperature estimates, identified valid markers, drive saturation state, sensor valid state, historical actual execution duty cycle and valid timestamp into the online parameter status record.

[0073] In this implementation, if the parameter regression residual is greater than 3.5 times the standard deviation of the calibration residual, the current parameter update is rejected, and the corresponding parameter covariance increases according to the process noise. The 3.5-fold threshold is derived from the regression residual distribution of 200 sets of parameter identification calibration data; below this threshold, the non-abnormal data retention rate is higher than 99.5%. When the parameter regression residual is within the allowable range, the module accepts the current parameter update result and corrects the parameter covariance based on the current regression residual.

[0074] When the common-mode predictive control module reads online parameter status records, it calculates the time difference between the current control time and the record's valid timestamp. If this time difference is less than 100 μs, the record is considered fresh; if it exceeds 100 μs, it is considered expired. 100 μs is determined by five times the 20 μs parameter identification cycle. When a record expires, the module retains the previous valid parameter estimate, increases the parameter covariance according to the number of elapsed parameter identification cycles, and prohibits the addition of bounded load feedforward. For example, a record timestamp 88 μs from the current time is allowed, while a difference of 112 μs is treated as an expired record.

[0075] In one numerical embodiment, the multiphase interleaved parallel buck switching power supply has four phase branches, an input voltage of 24V, a common output terminal voltage reference of 12V, and a switching frequency of 500kHz for each phase branch. After processing by S1, the estimated inductance parameters of each phase branch from phase 1 to phase 4 are 18.2μH, 19.1μH, 20.3μH, and 21.0μH, respectively; the estimated equivalent conduction loss resistance of each phase branch is 36.0mΩ, 39.2mΩ, 43.1mΩ, and 45.8mΩ, respectively; the estimated parameter of the common output filter capacitor is 466μF; and the estimated equivalent series resistance of the common output filter capacitor is 12.4mΩ.

[0076] The standard deviations of the inductance parameters for each phase branch from phase 1 to phase 4 are 0.32μH, 0.35μH, 0.37μH, and 0.40μH, respectively. The standard deviation of the common output filter capacitor parameter is 9μF, and the standard deviation of the equivalent series resistance of the common output filter capacitor is 0.6mΩ. The above parameter estimates, parameter covariances, and data validity status are written into the same online parameter status record for use by S2.

[0077] In some embodiments, S2 will be described in detail.

[0078] S21. Read the online parameter status record, establish the average physical model of the multiphase interleaved parallel buck switching power supply based on the parameter estimate, establish the parameter confidence domain based on the parameter covariance, and propagate the parameter confidence domain to the future control time domain to generate predictive constraint margin.

[0079] In this embodiment, the multiphase interleaved parallel buck switching power supply adopts the following continuous-time average physical model:

[0080] In the formula, Indicates time, in seconds; This indicates taking the derivative with respect to time; Indicates the branch sequence number; This represents the set of phase indices consisting of all phase branch numbers; Indicates the first Branch roads in time The inductance current of each phase branch, in A; Indicates the first Branch roads in time The final duty cycle actually executed after minimum pulse width limitation, dead zone processing and pulse width modulation quantization is a dimensionless value. Indicates time The input voltage, in volts (V); Indicates time The common output voltage, in volts (V); Indicates the first Branch roads in time The equivalent conduction loss resistance of each phase branch, in Ω; Indicates the first Branch roads in time The inductance parameters of each phase branch are given in ohms (H). Indicates time The common output filter capacitor current, in amperes (A); Indicates time The load current, in amperes (A); Indicates time The internal voltage of the common output filter capacitor, in volts (V). Indicates time The parameters of the common output filter capacitor are in F; Indicates time The equivalent series resistance of the common output filter capacitor, in Ω; This represents the summation of data in the phase index set. Adjacent physical quantities in the formula represent a multiplication relationship.

[0081] Equation 1 applies to the multiphase interleaved parallel step-down switching power supply disclosed in this embodiment. The common mode predictive control module uses the parameter estimates in the online parameter status record as the nominal parameters. The equivalent conduction loss resistance of each phase branch represents the equivalent value at the current operating point, remains unchanged within the 80μs prediction time domain, and is re-determined after subsequent parameter updates.

[0082] Before forming the parameter confidence region, the module uses the current valid estimates of each parameter as the normalization scale to convert parameter estimates and parameter covariances of different dimensions into dimensionless parameter vectors and dimensionless parameter covariances. When identifying the parameters of each phase branch and the parameters of the common output node independently, the parameter covariances are arranged into a block diagonal matrix according to the identification object. The module decomposes the normalized parameter covariance, generates boundary parameter points along the positive and negative directions of each principal axis of covariance, and then converts the boundary parameter points back to the original units of each parameter.

[0083] The confidence coefficient is set at 3.0. This confidence coefficient is determined based on standardized calibration data of the parameter estimation error relative to the parameter standard deviation. When the calibration residuals approximate a normal distribution, 3.0 corresponds to approximately 99.7% coverage; if this distribution condition is not met, the 99.7% measured quantile of the standardized parameter estimation error is used instead of 3.0. During the calibration process, reference parameter values ​​are obtained using precision inductance, resistance, and capacitance measurement equipment, and then the deviations between the online estimates and the reference parameter values ​​are compared to avoid directly using untransformed regression residuals to determine the confidence coefficient.

[0084] For the same control candidate sequence, the module calculates the prediction deviation of each covariance principal axis positive and negative boundary parameter point relative to the nominal parameter point, and accumulates the larger absolute deviation corresponding to each principal axis axis by axis to form a joint prediction deviation envelope used when multiple parameters change simultaneously. The joint prediction deviation of the common output voltage is converted into a duty cycle prediction constraint margin through local voltage and duty cycle sensitivity. The sensitivity is obtained by increasing and decreasing the duty cycle of the common mode candidate by one pulse width modulation minimum resolution at the current operating point and comparing the corresponding predicted values ​​of the common output voltage.

[0085] In one parameter propagation example, the joint prediction deviation of the common output voltage is 27mV. At the current operating point, the change in common output voltage per unit duty cycle is approximately 3.4V. Dividing 27mV by the aforementioned sensitivity and rounding up, the duty cycle prediction constraint margin is 0.008. The maximum joint prediction deviation of the inductor current in each phase branch is 0.45A, therefore the current prediction constraint margin is 0.45A. The joint prediction deviation of the junction temperature caused by changes in electrical parameters is 0.7℃, and the upper limit of the thermal model calibration residual is 2.5℃. Adding these two together yields a temperature prediction constraint margin of 3.2℃.

[0086] The junction temperature prediction for each phase branch power switch does not directly use the equivalent on-state loss resistance of each phase branch, which includes inductor windings, printed circuit board traces, and connector losses. The module reads the case temperature of the upper and lower bridge arm power switches, the current RMS value of the inductor current in each phase branch, the final duty cycle, the switching frequency, the on-resistance temperature characteristics of the power switches, and the turn-on and turn-off energy calibration table to obtain the on-state loss and switching loss of the power switches. The losses are then input into the heat dissipation impedance network to generate a junction temperature prediction sequence.

[0087] The thermal impedance data was obtained based on a limited number of historical calibration data. Total load current tests of 5A, 10A, and 15A were performed at ambient temperatures of 25℃, 55℃, and 85℃, respectively. The case temperature, calibration junction temperature, switching frequency, and RMS current of the power switch were recorded, and the thermal impedance coefficient was fitted according to different time scales. For example, for a power switch with a case temperature of 55℃ and a calculated power loss of 0.245W in the current control cycle, the estimated junction temperature at a time scale of 12℃ / W is 57.94℃. The 99.7th percentile of the junction temperature estimation residual in the calibration data is 2.5℃; therefore, 2.5℃ was used as the upper limit of the thermal model calibration residual.

[0088] S22. Generate a load prediction sequence based on fast event detection data, historical load current data, and changes in the common output filter capacitor current, and determine the charge requirement of the common output filter capacitor based on the difference between the load prediction sequence and the predicted inductor current values ​​of all phase branches.

[0089] In this embodiment, each element of the load prediction sequence represents the predicted load current value within the corresponding prediction control period, in amperes (A). When a dynamic load event is valid, the module starts from the last valid load current value before the event occurs and recursively calculates the load current change rate cycle by cycle according to the upper limit limit. When the absolute value of the load current change rate of five consecutive sampled values ​​is less than half of the dynamic load event determination threshold, the median of the five sampled values ​​is determined as the new steady-state load current, and this value is maintained in subsequent prediction periods. When a dynamic load event is invalid, the most recent valid load current value is used as the load prediction value for the entire prediction time domain.

[0090] The upper limit for the load current change rate is set at 1.5 A / μs. This upper limit is determined based on the maximum measured load current change rate of 1.32 A / μs during the target load device's operating mode switching, and is adjusted to 1.5 A / μs after adding measurement error and operating condition change margin. The same upper limit for the absolute value of the change rate is used for load increase events and load decrease events, and the recursive result for load decrease events must not be lower than 0 A.

[0091] When the load current sensor is invalid, the event validity flag of the fast event detection data is set to invalid, the load prediction sequence retains the most recent valid load current value, and the current prediction constraint margin is increased. This process does not use the operation mode without a configured independent load current sensor to maintain consistency with the technology of synchronously sampling load current.

[0092] The prediction time domain comprises 10 control cycles, each lasting 8 μs, for a total prediction time domain of 80 μs. The control time domain includes the first four control cycles, with the control values ​​from the fifth to the tenth prediction cycles remaining the same as those from the fourth control cycle. This prediction time domain is derived from the power supply requirement that the voltage recovery time at the common output terminal after a load step change does not exceed 80 μs. The 8 μs control cycle is determined based on the time required for the control processor to complete state updates, parameter point predictions, and constraint solving. During calibration, the calculation time for a single common mode control operation does not exceed 6 μs, reserving 2 μs for pulse width modulation updates.

[0093] The charge demand of the common output filter capacitor is accumulated cycle by cycle according to the difference between the load prediction sequence and the predicted values ​​of the inductor currents of all phase branches. When the load current is greater than the sum of the inductor currents of all phase branches, the common output filter capacitor releases charge to the load; when the load current is less than the sum of the inductor currents of all phase branches, the common output filter capacitor receives replenished charge.

[0094] The net charge deviation limit for the common output filter capacitor is determined based on the lower boundary of the common output filter capacitor parameter within the parameter confidence region. With an estimated common output filter capacitor parameter of 466 μF, a standard deviation of 9 μF, and a confidence coefficient of 3.0, the lower boundary of the capacitor parameter is 439 μF. To control the internal voltage change of the common output filter capacitor caused by the net charge deviation within 10 mV, the allowable net charge deviation is 4.39 μC, rounded down to the nearest integer, setting the limit to 4 μC. At 439 μF, a 4 μC net charge deviation corresponds to a voltage change of approximately 9.11 mV.

[0095] S23. Solve for the common mode energy regulation duty cycle component under the common output filter capacitor charge balance constraint and the tightened control constraint, and output the model prediction results corresponding to the common mode energy regulation duty cycle component.

[0096] In this embodiment, the common mode predictive control module uses an 8μs discrete control period. It discretizes the continuous-time average model at the current operating point using the current parameter estimates, forming a state recursive relationship for the next 10 control periods. The optimization variable is the common mode energy regulation duty cycle component for the first four control periods, while the duty cycle of the fourth control period is maintained for the fifth to tenth prediction periods.

[0097] The permissible deviation of the common output voltage is 1% of the reference value. This deviation range is derived from the power supply voltage specification of the target load device. Common mode control performs state prediction on the nominal parameter point and all boundary parameter points, requiring that the predicted common output voltage values ​​for each parameter point be within the permissible range. Duty cycle constraints, current constraints, and temperature constraints are each reduced by their corresponding prediction constraint margins.

[0098] The control objectives include reducing the common output voltage deviation, reducing the net charge deviation of the common output filter capacitor, and limiting the duty cycle period variation in the common mode. The common output voltage deviation is normalized to a 120mV allowable deviation, the net charge deviation is normalized to a 4μC limit, and the duty cycle variation is normalized to the allowable variation. In one embodiment, the weights of the three normalization costs are 1.00, 0.35, and 0.12, respectively. These weights are determined based on a finite number of calibration data sets under 300 different load step and input voltage conditions. Among the candidate weights for keeping the common output voltage within limits, the set with the shortest recovery time and the discontinuous oscillation of the duty cycle is selected.

[0099] The discrete-state recursive relation, control objective, common output filter capacitor charge constraint, and tightened control constraints together form a quadratic programming problem. The solver performs a maximum of 20 iterations per control cycle. The upper limit of 20 iterations is determined based on the maximum number of stable iterations that the control processor can complete within 6 μs. The solution terminates when the duty cycle change between two adjacent iterations is less than 0.0001 and all constraint residuals are within the set tolerance. The 0.0001 termination threshold is determined based on the pulse width modulation time resolution and the impact of continued iteration on the predicted value of the common output voltage; if the output voltage prediction does not change by more than 1 mV after continuing iteration below this threshold, further iteration will not cause the output voltage prediction to change by more than 1 mV.

[0100] The common-mode energy regulation duty cycle component consists of the actual common-mode duty cycle executed in the previous control cycle, the model predictive control correction, and the bounded load feedforward. The absolute upper limit of the bounded load feedforward is 0.025. This upper limit is determined based on load step scan data with an input voltage of 24V and inductance parameters of each phase branch in the range of 18μH to 21μH. When the upper limit exceeds 0.025, some parameter boundary points experience an increase in common output voltage overshoot without a further reduction in recovery time; therefore, the upper limit is set to 0.025.

[0101] When the solver times out or no feasible solution is found, the energy adjustment duty cycle component of the feasible common mode from the previous control cycle is used to halve the bounded load feedforward cycle by cycle and set the drive saturation state to active. This backoff method retains the common mode duty cycle that has passed the boundary parameter point check in the previous cycle and gradually reduces the load feedforward that may cause constraint conflicts.

[0102] In one numerical embodiment, the load current increases from 5A to 8A within 3μs, with a load current change rate of 1.0A / μs, exceeding the dynamic load event judgment threshold of 0.8A / μs. Within the first 8μs after the load current reaches 8A, if the sum of the inductor currents of all phase branches remains at 5A, the common output filter capacitor needs to provide approximately 24μC of charge. Based on the nominal value of a 470μF common output filter capacitor, this charge corresponds to an internal voltage drop of approximately 51mV; based on the nominal value of a 12mΩ equivalent series resistance, the 3A discharge current corresponds to an instantaneous voltage drop of approximately 36mV, totaling approximately 87mV.

[0103] The duty cycle prediction constraint margin obtained from parameter confidence domain propagation is 0.008, the current prediction constraint margin is 0.45A, and the temperature prediction constraint margin is 3.2℃. The actual common mode duty cycle executed in the previous control cycle was 0.503, the model predictive control correction was 0.006, and the bounded load feedforward was 0.012. These three factors combined yield the current common mode energy regulation duty cycle component of 0.521. The lowest predicted value of the common output voltage corresponding to the nominal parameter point is 11.913V, and the lowest boundary predicted value corresponding to the joint prediction deviation envelope is not lower than 11.886V, both of which are not lower than the 1% allowable deviation lower limit of 11.88V for the 12V reference value. The time for the predicted common output voltage to recover to 1% of the reference value and remain continuously for 5 control cycles is 68μs.

[0104] In some embodiments, S3 is described in detail.

[0105] S31. Based on the deviation of the inductor current of each phase branch relative to the target distribution current and the deviation of the estimated junction temperature of each phase branch power switch relative to the average estimated junction temperature, generate the original differential mode duty cycle adjustment.

[0106] In this embodiment, the differential mode constraint allocation and drive module determines the total target current in the future control time domain based on the load prediction sequence. The module subtracts the current inductor current of each phase branch from the upper limit of the inductor current to obtain the available current margin; it subtracts the estimated current junction temperature of the power switch from the upper limit of the junction temperature to obtain the available temperature margin. These two types of margins are normalized according to their corresponding upper limits and combined to form non-negative allocation weights. Then, each allocation weight is normalized so that the sum of the target allocated currents of each phase branch equals the total target current.

[0107] For example, with a total target current of 8A, when the normalized allocation weights of the four phase branches are 0.95, 1.05, 1.10, and 0.90 respectively, the corresponding target allocation currents are 1.90A, 2.10A, 2.20A, and 1.80A respectively. Phase branches with higher estimated junction temperatures of the power switches or smaller available current margins receive smaller target allocation currents, while phase branches with lower estimated junction temperatures and larger available current margins receive larger target allocation currents.

[0108] The current deviation control component employs proportional-integral (PI) control. The proportional gain is set to 0.018 duty cycle adjustment per 1A current deviation, and the integral gain is set to 120 duty cycle adjustment per 1A·s cumulative current deviation. The temperature deviation control component is set to 0.0008 duty cycle adjustment per 1℃ temperature deviation. These control parameters were determined based on a finite number of current sharing calibrations performed under 300 different total load currents, phase branch inductance parameter deviations, and ambient temperatures. The selection criteria are that the steady-state current deviation of the phase branch does not exceed 5% of the average phase branch current, and that no duty cycle oscillation with an increasing amplitude in the same direction occurs for five consecutive control cycles during the control process.

[0109] The integral state of the proportional-integral controller is limited to the value corresponding to the duty cycle adjustment. The range is from 0.03 to 0.03. This range is determined based on half of the upper limit of the absolute value of the differential mode duty cycle component, which is 0.06, reserving duty cycle space for proportional control and temperature deviation control. When the projected result is inconsistent with the original differential mode duty cycle adjustment or when the drive saturation state is valid, the module corrects the integral state according to the actually executed differential mode duty cycle component to prevent the integral state from continuing to accumulate during constraint activation.

[0110] S32. Construct a feasible region based on the inductance parameters of each phase branch, the duty cycle component of the common mode energy regulation, the prediction constraint margin, and the model prediction results in the online parameter status record. Project the original differential mode duty cycle regulation amount onto the feasible region to obtain the duty cycle component of the differential mode current redistribution for each phase.

[0111] In this embodiment, the following projection relationship is adopted:

[0112] In the formula, This represents the vector of the phase differential mode current redistribution duty cycle components obtained after projection. This represents the duty cycle vector of the difference pattern to be solved; Indicates the first The duty cycle components of the difference mode to be solved; Indicates the first The original differential mode duty cycle adjustment amount; Indicates the first The phase is a dimensionless projection weight, and it is a positive number; Represents the set of phase indices; Indicates the first The estimated inductance parameters obtained from online phase identification, in units of H; Indicates the duty cycle component of the common mode energy regulation; and These represent the lower and upper limits of the final duty cycle, respectively. This indicates the duty cycle prediction constraint margin; This represents the upper limit of the absolute value of the duty cycle quantity of current redistribution in each phase differential mode; This indicates the prediction obtained after using the entire differential mode duty cycle vector. The inductor current in the next control cycle of a phase, in amperes (A). Indicates the first Upper limit of phase inductor current, in amperes (A); Indicates the first Phase current prediction constraint margin, in A; This indicates the prediction obtained after using the entire differential mode duty cycle vector. Estimated junction temperature of the power switch transistor in the next control cycle of the phase, in °C; Indicates the first Upper limit of junction temperature of phase power switching transistor, in °C; Indicates the first Phase temperature prediction constraint margin, in °C; superscript Indicates the raw, unprojected quantity; superscript Indicates the result of the projection solution; superscript Indicates the next control cycle; This represents the vector corresponding to the minimum value of the objective function; This indicates that the vector to be solved should satisfy the constraints listed below; This indicates summing the data in the set of phase indices; This indicates that the corresponding constraint applies to each phase branch in the phase index set.

[0113] The weighted zero-sum constraint on inductor parameters in Equation 2 is set based on the correlation between the effect of the duty cycle change of each phase branch on the rate of change of inductor current and the reciprocal of the inductor parameters of each phase branch. When the inductor parameters of each phase branch are different, the ordinary algebraic sum of the duty cycle components of the differential mode current redistribution in each phase is zero, which cannot make its first-order action on the rate of change of total inductor current zero. By dividing each phase differential mode duty cycle component by the corresponding estimated inductor parameter and constraining their sum to zero, the effect of the differential mode duty cycle on the first-order rate of change of total inductor current can be limited.

[0114] The projection weights are determined based on the inductance parameter covariance, available current margin, and available temperature margin of each phase branch. The module first normalizes the three types of inputs. Phase branches with larger parameter covariance, smaller available current margin, or smaller available temperature margin are assigned larger projection weights to reduce adjustments to those phase branches. The projection weights are updated every 1ms. This 1ms update period is determined based on calibration data of the power switch junction temperature change rate and parameter covariance change rate. With update periods shorter than 1ms, the projection weights are significantly affected by current ripple; with update periods longer than 1ms, the weights may not reflect changes in temperature margin in a timely manner during some load changes.

[0115] Final duty cycle lower limit The lower limit is set to 0.05. This lower limit is determined based on a 2μs switching period and a minimum on-time of 100ns at a switching frequency of 500kHz. The final upper limit of the duty cycle is... The duty cycle is set to 0.92. The total budget for the minimum turn-off time of the drive circuit is 160ns, which includes a dead time of 40ns for the two commutation processes. Therefore, the upper limit of the duty cycle is 0.92.

[0116] Upper limit of inductor current in each phase branch A value of 12A is selected. This upper limit is determined based on the saturation current of a 15A inductor and the continuous safe current of an 18A power switch, with 80% of the smaller value (15A) used as the operating upper limit. Upper limit of power switch junction temperature. The upper limit is set to 105℃. This upper limit is determined based on the maximum junction temperature of the device at 150℃, the upper limit of the estimated residual junction temperature at 2.5℃, and the temperature derating corresponding to the target lifespan, so that a temperature margin of more than 42.5℃ is maintained between the upper limit of control and the maximum junction temperature of the device.

[0117] The upper limit of the absolute value of the duty cycle of current redistribution in each phase differential mode The upper limit is set to 0.06. This upper limit is determined based on duty cycle step scan data under different total load currents and deviations in the inductance parameters of each phase branch. When the absolute value exceeds 0.06, the peak-to-peak value of the inductor current in some phase branches increases by more than 15% relative to the calibration reference, therefore the upper limit is set to 0.06.

[0118] S33. Combine the common mode energy regulation duty cycle component and the differential mode current redistribution duty cycle component of each phase into the target duty cycle of each phase branch, perform duty cycle quantization, constraint verification and interleaved pulse width modulation, and perform backoff control when the feasible domain is empty or the driving conditions are not met.

[0119] In this embodiment, Equation 2 is first solved in the continuous duty cycle domain. The interleaved pulse width modulation unit converts the continuous duty cycle into a pulse width modulation count value according to the actual hardware time resolution, and then uses the quantized duty cycle to recheck the weighted zero-sum constraints of the duty cycle, current, temperature, and inductance parameters. If the quantized weighted zero-sum residual exceeds the quantization tolerance, one pulse width modulation count is added or subtracted on the phase branch with a larger available current margin and available temperature margin, and all constraints are rechecked. The quantization tolerance is determined based on the maximum impact of one pulse width modulation count on the weighted zero-sum residual of each phase branch.

[0120] Each phase branch uses the same 500kHz switching frequency. The staggered phases of the four phase branches are 0°, 90°, 180°, and 270°, respectively. The phase interval is determined by dividing 360° by the number of phase branches. The gate drive voltage is maintained at 5V. During the control process, the duty cycle of the pulse width modulation signal is adjusted, but the gate drive voltage amplitude is not adjusted.

[0121] When the feasible region is empty, the module first calculates the intersection of the allowable duty cycle intervals of each phase branch, moves the common mode energy regulation duty cycle component to the intersection boundary closest to the original solution value, and then reprojects it. If no feasible solution is found after two consecutive control cycles, the differential mode current redistribution duty cycle component of each phase decays by 50% of its current value cycle by cycle. The condition of two consecutive control cycles is used to eliminate quantization errors and solver transients within a single cycle, ensuring that backoff control starts within 16μs. The 50% decay ratio is determined based on backoff experiments with different decay rates, ensuring that the differential mode regulation amount approaches zero within several control cycles while avoiding sudden changes in duty cycle.

[0122] If overcurrent or overtemperature risks still exist after the duty cycle components of the differential mode current redistribution have decayed to zero, the module reduces the duty cycle component of the common mode energy regulation and triggers hardware overcurrent or overtemperature protection. The drive saturation state records the duty cycle boundary, current constraint, temperature constraint, differential mode duty cycle component boundary, solution timeout, and active items in the rollback control.

[0123] In a numerical embodiment, the common-mode energy regulation duty cycle component is 0.521. The estimated inductance parameters of each phase branch from phase 1 to phase 4 are 18.2 μH, 19.1 μH, 20.3 μH, and 21.0 μH, respectively. After projection by S32, the differential-mode current redistribution duty cycle components of each phase are as follows: 0.006000, 0.004000, 0.007000 and 0.004716208. The sum of the duty cycle components of the above differential mode, after being divided by the corresponding estimated inductance parameters, is zero within the solver equality constraint tolerance.

[0124] The target duty cycles for the continuous domain from phase 1 to phase 4 are 0.515000, 0.525000, 0.528000, and 0.516283792, respectively. The interleaved pulse width modulation unit quantizes the target duty cycle according to the hardware time resolution, re-checks all constraints, and writes the final duty cycle actually executed after quantization into the online parameter status record. Parameter identification and event feedback both use the final duty cycle actually executed.

[0125] In some embodiments, S4 is described in detail.

[0126] S41. Collect the actual response data after running according to the final duty cycle, align the actual response data with the model prediction results output by S2 in time, extract the event response deviation and generate the event-level model mismatch index.

[0127] In this embodiment, the event feedback correction module determines the evaluation starting point based on the event occurrence time in the fast event detection data. The evaluation window includes baseline data for 32 μs before the event and response data for 256 μs after the event. The data for the 32 μs before the event is used to determine the baseline for the common output voltage and the inductor current of each phase branch before the event. The data for the 256 μs after the event covers the 80 μs prediction time domain, the 80 μs common output voltage recovery requirement, and the steady-state confirmation stage after recovery. The window length is determined based on 300 sets of load step response data, and the proportion of events whose cause category is not changed by continuing to increase the window after 256 μs exceeds 99%.

[0128] The actual response data includes the actual sequence of the common output voltage, the actual sequence of the inductor current in each phase branch, the actual sequence of the load current, and the sampling sequence of the power switch case temperature. The event feedback correction module compensates for the time difference caused by sampling, filtering, calculation, and drive execution based on the timestamp, so that the actual response data corresponds to the same control moment as the model prediction results.

[0129] The event feedback correction module extracts the minimum voltage deviation at the common output terminal, recovery time deviation, peak inductor current deviation in each phase branch, steady-state current distribution deviation, and net charge deviation of the common output filter capacitor. The recovery time is defined as the difference between the moment the common output terminal voltage first enters the 1% range of the reference value and remains there for 5 consecutive control cycles and the event occurrence time. The condition of 5 consecutive control cycles is determined based on load step calibration data, which can exclude the possibility of the common output terminal voltage briefly crossing the allowable range and then leaving it again.

[0130] Each deviation was divided by the corresponding standard deviation of the calibration residual to obtain five dimensionless deviations. These five dimensionless deviations were then multiplied by fixed weights of 0.30, 0.20, 0.25, 0.15, and 0.10, respectively. The square root of the sum of the weighted squared values ​​yielded the event-level model mismatch index. These weights were determined based on 300 sets of cross-validation data, including normal response, parameter drift, drive saturation, and sensor anomalies, ensuring that the minimum common output voltage and the peak inductor current of each phase branch had high weights in the mismatch index.

[0131] For example, when the five dimensionless biases are 0.25, 0.333, 0.30, 0.20, and 0.25 respectively, the event-level model mismatch index is approximately 0.275, or 0.28 when recorded to two decimal places. When the five dimensionless biases are 1.10, 0.90, 1.50, 1.20, and 0.90 respectively, the event-level model mismatch index is approximately 1.18.

[0132] S42. Generate event cause categories based on sensor effective state, drive saturation state, and event-level model mismatch index, and determine the parameter group to be adjusted based on the contribution of each deviation.

[0133] In this implementation, the event cause category is determined in the order of sensor malfunction, drive saturation, model mismatch, and normal response. When the sensor validity state of any critical voltage or current sampling channel is invalid, the event cause category is set to sensor malfunction, and the affected parameters are not corrected using the current event. When the sensor validity state is valid, but the drive saturation state indicates that duty cycle boundaries, current constraints, temperature constraints, differential mode duty cycle component boundaries, or backoff control are activated, the event cause category is set to drive saturation.

[0134] When the sensor is in an effective state, the drive is in an ineffective state, and the event-level model mismatch index reaches 1.00, the event cause category enters the model mismatch state. The entry threshold of 1.00 is determined based on 300 sets of cross-validation data; at this threshold, the proportion of normal events judged as model mismatches is less than 1%. When the event-level model mismatch index is below 0.45 and this condition is met for 5 consecutive effective evaluation windows, the event exits the model mismatch state and enters the normal response state. The exit threshold of 0.45 and the 5 consecutive window condition are determined based on the same set of cross-validation data to reduce repeated switching of the mismatch index near the boundary.

[0135] When the event-level model mismatch index is between 0.45 and 1.00, it only maintains the previous model mismatch or normal response state. Sensor anomalies and drive saturation states are reassessed each time and do not participate in the model mismatch hysteresis.

[0136] The event feedback correction module retains the weighted contributions of each component constituting the mismatch index of the event-level model. When the deviation of the minimum value of the common output voltage and the deviation of the net charge of the common output filter capacitor contribute significantly, the parameters of the common output filter capacitor and the equivalent series resistance of the common output filter capacitor are determined as the parameter set to be adjusted. When the peak deviation of the inductor current or the steady-state current distribution deviation of a certain phase branch contributes significantly, the inductor parameters of that phase branch and the equivalent conduction loss resistance of each phase branch are determined as the parameter set to be adjusted.

[0137] S43. Generate bounded identification state adjustment amount based on event cause category and parameter group to be adjusted, adjust forgetting factor, parameter covariance, parameter estimation correction boundary and parameter freeze state used in recursive identification, and generate online parameter state record for subsequent control cycles.

[0138] In this implementation, when the event cause category is model mismatch, the recursive identification forgetting factor is adjusted from the normal operating value to the fast tracking lower limit, and the parameter covariance of the parameter group to be adjusted is amplified. The normal operating value of the forgetting factor is 0.995, the fast tracking lower limit is 0.985, and the steady-state noise immunity upper limit is 0.999. The above values ​​are determined based on 200 sets of parameter identification calibration data, and the selection conditions are that the convergence time does not exceed 12ms after a 15% change in the parameter reference value, and the 95th percentile of the steady-state parameter relative error does not exceed 5%.

[0139] When the event-level model mismatch index is between 1.00 and 1.50, the parameter covariance of the parameter group to be adjusted is amplified by a factor of 2; when it is between 1.50 and 2.00, it is amplified by a factor of 3; and when it is not lower than 2.00, it is amplified by a factor of 4. The amplification factor is determined based on the comparison between convergence time and estimated fluctuation in the parameter drift calibration data. When the amplification factor exceeds 4, the periodic fluctuation of the parameter estimate increases while the convergence time does not continue to shorten; therefore, 4 is set as the upper limit.

[0140] The correction amount for a single parameter estimate shall not exceed 20% of the previous valid parameter estimate. This boundary is determined based on device tolerance, temperature variation range, and online anomaly identification data, and is used to limit the correction magnitude of a single estimator. It does not represent a 20% actual change in device parameters within a control cycle. When the current update result exceeds the boundary, it is limited to the boundary value, and a large parameter covariance is maintained for further correction with subsequent valid data.

[0141] When the event cause category is a normal response, the forgetting factor is gradually adjusted towards 0.999, and the parameter covariance shrinks according to the parameter regression residual, but not below the lower limit of covariance determined by sensor noise and quantization error. When the event cause category is drive saturation, no additional parameter correction or covariance amplification caused by the current event is performed, but normal time updates continue to be performed. When the event cause category is sensor malfunction, the measurement updates of the affected parameters are frozen, and their parameter covariance is increased according to the elapsed time.

[0142] Event feedback assessment requires response data for 256 μs after the event occurs; therefore, the updated online parameter status record does not take effect in the first control cycle after the event. From the event occurrence until S43 completion, the common mode predictive control module continues to use the most recent online parameter status record. After S43 completion, the updated online parameter status record takes effect from the first subsequent control cycle.

[0143] In one numerical embodiment, S2 predicts a minimum common output voltage of 11.913V and a recovery time of 68μs. After operating according to the final duty cycle generated by S3, the actual minimum common output voltage is 11.910V, and the actual recovery time is 72μs. The peak inductor currents of each phase branch from phase 1 to phase 4 are 2.36A, 2.30A, 2.27A, and 2.25A, respectively, none of which reach the 12A inductor current limit.

[0144] The event-level model mismatch index obtained by the event feedback correction module is 0.28, indicating that the sensor is in an effective state and the drive is in an ineffective state. Therefore, the event cause category is normal response. The recursive identification forgetting factor is adjusted to 0.999, the parameter covariance shrinks according to the regression residual, and the update result is written into the effective online parameter status record in the first control cycle after S43 is completed.

[0145] In another numerical embodiment, the inductance parameters of the first-phase branch shift due to changes in temperature and the operating point of the magnetic material, resulting in a persistent deviation between the peak inductance current of the first-phase branch and the model prediction. The event-level model mismatch index is 1.18. The sensor is in an effective state, while the drive is in an ineffective state; therefore, the event cause category is model mismatch. Since the deviation of the peak inductance current of the first-phase branch contributes significantly to the event-level model mismatch index, the inductance parameters of the first-phase branch and the equivalent conduction loss resistance are identified as the parameter set to be adjusted. The event feedback correction module adjusts the forgetting factor to 0.985, amplifies the parameter covariance of the parameter set to be adjusted by a factor of 2, and updates the online parameter status record according to the boundary that the correction amount for a single parameter estimate does not exceed 20%.

[0146] like Figure 6 As shown, in some embodiments, this embodiment describes the control system.

[0147] The control system includes a synchronous sampling and online parameter identification module, a common mode predictive control module, a differential mode constraint allocation and drive module, and an event feedback correction module.

[0148] The synchronous sampling and online parameter identification module is connected to the input voltage sampling circuit, the common output voltage sampling circuit, the current sampling circuits for each phase branch, the load current sampling circuit, and the power switch case temperature sampling circuit. This module includes a unified sampling clock, an independent sample-and-hold circuit, a fast event detection path, a slow parameter identification path, a recursive parameter identification processing unit, and an online parameter status record storage area. The fast event detection path generates the load current change rate, load current increment, event direction, event occurrence time, and event validity marker. The slow parameter identification path performs switching ripple suppression and valid timestamp correction. The recursive parameter identification processing unit identifies the parameters of each phase branch and the parameters of the common output node.

[0149] The common-mode predictive control module includes a physical model calculation unit, a parameter confidence domain generation unit, a predictive constraint margin generation unit, a load prediction unit, and a common-mode control solution unit. The physical model calculation unit establishes the state prediction relationship according to Equation 1. The parameter confidence domain generation unit reads the parameter estimates and parameter covariances, and performs dimensional normalization, covariance decomposition, and boundary parameter point generation. The predictive constraint margin generation unit generates duty cycle, current, and temperature prediction constraint margins based on the joint prediction bias envelope. The common-mode control solution unit outputs the common-mode energy regulation duty cycle component and the model prediction results under the common output filter capacitor charge balance constraint.

[0150] The differential mode constraint allocation and drive module includes a raw differential adjustment generation unit, a feasible region projection unit, a duty cycle quantization verification unit, and an interleaved pulse width modulation unit. The raw differential adjustment generation unit generates the raw differential mode duty cycle adjustment based on the current deviation and junction temperature deviation. The feasible region projection unit generates the differential mode current redistribution duty cycle components for each phase according to Equation 2. The duty cycle quantization verification unit converts the continuous domain target duty cycle into a pulse width modulation count and re-checks all constraints. The interleaved pulse width modulation unit generates power switch drive signals with predetermined interleaved phases and outputs the drive saturation state.

[0151] The event feedback correction module includes a data alignment unit, an event-level model mismatch index generation unit, an event cause classification unit, and a bounded identification state adjustment unit. The data alignment unit aligns the actual response data with the model prediction results according to the event occurrence time. The event-level model mismatch index generation unit extracts five types of event response bias and generates a dimensionless event-level model mismatch index. The event cause classification unit generates event cause categories and parameter groups to be adjusted. The bounded identification state adjustment unit adjusts the forgetting factor, parameter covariance, parameter estimation correction boundary, and parameter freeze state, and writes the results to the online parameter status record storage area.

[0152] In one hardware configuration, synchronous sampling, timestamp generation, duty cycle quantization, and interleaved pulse width modulation are performed by a field-programmable gate array (FPGA), while parameter identification, parameter confidence domain propagation, load prediction, common-mode control solving, feasible region projection, and event feedback correction are performed by the main control processor. The FPGA and the main control processor exchange online parameter status records, model prediction results, and final duty cycles via a dual-port memory. Synchronous sampling data is stored in a circular buffer, enabling the event feedback correction module to read data before and after an event occurs.

[0153] In another hardware configuration, the above functions are performed by a digital signal processor with a high-resolution pulse width modulation unit. Synchronous sampling interrupts have the highest execution priority, interleaved pulse width modulation updates are executed at the switching cycle boundaries, common mode predictive control is executed every 8 μs, slow parameter identification is executed every 20 μs, and event feedback correction is triggered based on valid dynamic load events. Data from different execution cycles are timestamped.

[0154] To enable those skilled in the art to understand and implement the technical solutions described in this specification, the execution process of the method will be further explained below in the context of a power supply scenario for a communication base station power amplifier module.

[0155] This embodiment uses a four-phase interleaved parallel step-down switching power supply to power the communication base station power amplifier module. The input voltage is 24V, the common output voltage reference value is 12V, the switching frequency of each phase branch is 500kHz, and the drive phase interval between adjacent phase branches is 90°. The nominal inductance parameters of each phase branch from phase 1 to phase 4 are 18μH, 19μH, 20μH, and 21μH, respectively. The nominal value of the common output filter capacitor is 470μF, and the nominal value of the equivalent series resistance of the common output filter capacitor is 12mΩ. The above device parameters are determined based on the maximum load current, output voltage ripple requirements, and load step recovery time requirements of the communication base station power amplifier module.

[0156] After the switching power supply enters a stable operating state, the synchronous sampling and online parameter identification module synchronously acquires the input voltage, common output voltage, inductor current of each phase branch, and load current at a sampling frequency of 2.5MHz. It then correlates the case temperature of the power switch transistors in each phase branch with the sampled electrical quantities based on a unified timestamp. The synchronously sampled data simultaneously enters both the fast event detection path and the slow parameter identification path. The fast event detection path calculates the load current change rate using the five most recent load current samples. The slow parameter identification path employs a 64-tap linear phase finite-length unit impulse response low-pass filter to suppress switching frequency components and corrects the effective timestamp of the filtering result to the center of the filtering window.

[0157] When the power amplifier module of the communication base station switches from standby mode to signal transmission mode, the load current increases from 5A to 8A within 3μs. The load current change rate calculated by the fast event detection path is 1.0A / μs, which exceeds the dynamic load event judgment threshold of 0.8A / μs, generating valid fast event detection data. The data includes a load increase event marker, the event occurrence time, the 1.0A / μs load current change rate, and the 3A load current increment.

[0158] The slow parameter identification path constructs regression data for parameters of each phase branch and the common output node. When the minimum eigenvalue of the normalized information matrix reaches 0.015 and the sensor is valid, recursive least squares processing is used to update the corresponding parameters. After parameter identification, the estimated inductance parameters of each phase branch from phase 1 to phase 4 are 18.2μH, 19.1μH, 20.3μH, and 21.0μH, respectively; the estimated equivalent conduction loss resistance of each phase branch is 36.0mΩ, 39.2mΩ, 43.1mΩ, and 45.8mΩ, respectively; the estimated parameter of the common output filter capacitor is 466μF; and the estimated equivalent series resistance of the common output filter capacitor is 12.4mΩ. The above parameter estimates, parameter covariance, junction temperature estimates, and data validity status form an online parameter status record.

[0159] The common-mode predictive control module reads online parameter status records, normalizes the dimensions of parameter estimates and parameter covariance, and generates nominal parameter points and boundary parameter points. Each parameter point is then entered into an 80μs prediction time domain. A joint prediction bias envelope transformation yields a 0.008 duty cycle prediction constraint margin, a 0.45A current prediction constraint margin, and a 3.2℃ temperature prediction constraint margin.

[0160] The common-mode predictive control module generates a load prediction sequence that varies from 5A to 8A based on fast event detection data. Within the first 8μs after the load current reaches 8A, if the sum of the inductor currents in all phase branches is still 5A, the common output filter capacitor needs to provide approximately 24μC of charge to the load. Based on an initial estimate using a nominal capacitance of 470μF and an equivalent series resistance of 12mΩ, the total voltage change within the common output filter capacitor and the voltage drop across the equivalent series resistance is approximately 87mV.

[0161] The common-mode predictive control module generates a 0.521 common-mode energy regulation duty cycle component, ensuring that the predicted common output voltage at all boundary parameter points is within 1% of the allowable deviation of the 12V reference value and satisfies charge, current, and temperature constraints. The lowest predicted common output voltage at the nominal parameter point is 11.913V, the lowest boundary predicted value corresponding to the joint prediction deviation envelope is not lower than 11.886V, and the prediction recovery time is 68μs. The common-mode energy regulation duty cycle component and the model prediction results are fed into the differential mode constraint allocation and drive module.

[0162] The differential mode constraint allocation and drive module generates the original differential mode duty cycle adjustment based on the inductor current deviation of each phase branch and the junction temperature deviation of the power switch. After projection using Equation 2, the differential mode current redistribution duty cycle components for phases 1 to 4 are as follows: 0.006000, 0.004000, 0.007000 and 0.004716208. The duty cycle components of the current redistribution in each phase differential mode are weighted according to the estimated online inductance parameters of 18.2μH, 19.1μH, 20.3μH and 21.0μH, and their sum is within the solver equality constraint tolerance.

[0163] The target duty cycles in the continuous domain are 0.515000, 0.525000, 0.528000, and 0.516283792, respectively. The interleaved pulse width modulation unit quantizes and verifies these target duty cycles, generating the final executed duty cycles and phases of the power switch drive signals at 0°, 90°, 180°, and 270°, respectively. The gate drive voltage is maintained at 5V.

[0164] After operating at the final duty cycle, the event feedback correction module acquires the actual sequence of the common output voltage, the actual sequence of the inductor current in each phase branch, and the actual sequence of the load current, and aligns them with the model prediction results based on the timestamps. The actual minimum value of the common output voltage is 11.910V, and the actual recovery time is 72μs. The peak values ​​of the inductor currents in each phase branch from phase 1 to phase 4 are 2.36A, 2.30A, 2.27A, and 2.25A, respectively, all of which do not reach the 12A inductor current upper limit.

[0165] The event feedback correction module generates a 0.28 event-level model mismatch index based on the deviations of five standardized event categories. The sensor's effective state is considered effective, the drive saturation state is considered ineffective, and the event cause category is normal response. The event feedback correction module adjusts the forgetting factor to 0.999 and shrinks the parameter covariance based on the parameter regression residuals. After S43 is completed, the updated parameter estimates, parameter covariance, identified effective markers, drive saturation state, sensor effective state, historical actual execution duty cycle, and effective timestamp are written into the online parameter status record used in the first subsequent control cycle.

[0166] Finally, it should be noted that in this embodiment, the online parameter status record generated by S1 is used by S2, the common mode energy regulation duty cycle component and model prediction results generated by S2 are used by S3, the actual response data corresponding to the final duty cycle generated by S3 is used by S4, and the bounded identification state adjustment amount generated by S4 is written into the online parameter status record of subsequent control cycles. The above data relationships form a continuous control process.

[0167] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A switching power supply control method based on dynamic load tracking, characterized in that, include: The input voltage, common output voltage, inductor current and load current of each phase branch of the multi-phase interleaved parallel step-down switching power supply are sampled synchronously, and the case temperature of the power switch tubes of each phase branch is aligned. The sampled data is divided into fast event detection data and slow parameter identification data. Based on the slow parameter identification data, the parameters of each phase branch and the parameters of the common output node are identified, and an online parameter status record is generated. A physical model and parameter confidence domain are established based on online parameter status records. The parameter confidence domain is converted into prediction constraint margin. A load prediction sequence is generated based on fast event detection data. The common mode energy regulation duty cycle component and model prediction results are solved under the common output filter capacitor charge balance constraint. The original differential mode duty cycle adjustment is generated based on the inductor current of each phase branch, the estimated junction temperature of the power switch, the duty cycle component of the common mode energy regulation, and the model prediction results. It is then projected onto the feasible region that satisfies the weighted zero-sum constraint of the inductor parameters and the control constraint to generate the duty cycle component of the differential mode current redistribution for each phase. After quantization and verification, the final duty cycle and drive signal are generated. Based on the actual response data corresponding to the final duty cycle, model prediction results, drive saturation state, and sensor effective state, an event-level model mismatch index and event cause category are generated. Based on the event cause category, a bounded identification state adjustment amount is generated, and an online parameter state record for subsequent control cycles is generated.

2. The switching power supply control method based on dynamic load tracking according to claim 1, characterized in that, Simultaneous sampling and generation of fast event detection data and slow parameter identification data, including: The sample-and-hold circuits that simultaneously trigger the input voltage, common output voltage, inductor current of each phase branch, and load current using a unified sampling clock are used to add a unified timestamp to the synchronous sampling data and align the case temperature of the power switch transistors of each phase branch with the synchronous sampling data according to the timestamp. The synchronous sampling data is sent into the fast event detection path. Local linear fitting is performed on the continuous load current sampling values. The load current change rate, load current increment, event direction, event occurrence time and event valid marker are determined based on the fitting slope and the difference between the fitting endpoints, and fast event detection data is generated. The synchronous sampling data is sent into the slow parameter identification path, and the switching frequency component is suppressed by the linear phase finite length unit impulse response low-pass filter. Based on the effective timestamp of the group delay correction filtering result of the low-pass filter, the slow parameter identification data is generated.

3. The switching power supply control method based on dynamic load tracking according to claim 2, characterized in that, Identify the parameters of each phase branch and the parameters of the common output node, including: Align the slow parameter identification data with the historical final duty cycle that has actually been executed at the corresponding time. Construct regression data of each phase branch parameter based on the input voltage, historical final duty cycle, common output voltage, inductor current of each phase branch and its rate of change, and identify the inductor parameters of each phase branch and the equivalent conduction loss resistance of each phase branch. The current of the common output filter capacitor is determined based on the sum of the inductor currents of all phase branches, the load current, and the voltage at the common output terminal. The equivalent series resistance of the common output filter capacitor and the parameters of the common output filter capacitor are identified by using a short data window that includes the edge of load change and a long data window that includes the change of charge accumulation, respectively. A normalized information matrix is ​​constructed based on the slow parameter identification data. When the normalized information matrix satisfies the continuous excitation condition and the sensor is in an effective state, a recursive least squares parameter update is performed. When the update condition is not met, the previous effective parameter estimate is maintained and the corresponding parameter covariance is increased to generate an online parameter status record.

4. The switching power supply control method based on dynamic load tracking according to claim 3, characterized in that, Establish the parameter confidence region and convert it into prediction constraint margins, including: The parameter vector and parameter covariance are normalized in terms of dimensions based on the estimated values ​​of each parameter. The normalized parameter covariance is decomposed, and boundary parameter points are generated along the positive and negative directions of each principal axis of covariance. The boundary parameter points are then converted into parameter confidence regions with the original parameter units. The nominal parameter points and each boundary parameter point are used to predict the estimated values ​​of the common output voltage, the inductor current of each phase branch and the junction temperature of the power switch of each phase branch in the future control time domain. The prediction deviations corresponding to each covariance principal axis are then used to form a joint prediction deviation envelope. The duty cycle prediction constraint margin is generated based on the joint prediction deviation of the common output voltage and the duty cycle sensitivity; the current prediction constraint margin is generated based on the joint prediction deviation of the inductor current of each phase branch; and the temperature prediction constraint margin is generated based on the joint prediction deviation of the junction temperature and the thermal model calibration residual.

5. The switching power supply control method based on dynamic load tracking according to claim 4, characterized in that, Solve for the common model energy regulation duty cycle component and model prediction results, including: Based on the final duty cycle, input voltage, common output voltage, parameters of each phase branch, and parameters of the common output node actually executed in the previous control cycle, an average physical model of the multiphase interleaved parallel buck switching power supply is established, and a junction temperature prediction relationship is established based on the case temperature of the power switch tubes in each phase branch, switching loss data, and thermal impedance data. A load prediction sequence is generated based on the load current change rate, load current increment, event direction, and event occurrence time in the fast event detection data. The common output filter capacitor charge demand is accumulated cycle by cycle based on the difference between the load prediction sequence and the predicted inductor current values ​​of all phase branches. Under the constraints of charge balance of the common output filter capacitor, duty cycle constraint after deducting the prediction constraint margin, current constraint, and temperature constraint, the common mode energy regulation duty cycle component is solved, and the predicted sequence of common output terminal voltage, predicted sequence of inductor current of each phase branch, predicted sequence of common output filter capacitor current, predicted sequence of junction temperature of power switch of each phase branch, predicted recovery time, and constraint activation state are output.

6. The switching power supply control method based on dynamic load tracking according to claim 5, characterized in that, Generate the original differential mode duty cycle adjustment, including: The total target current is determined based on the load prediction sequence. The available current margin is determined based on the difference between the upper limit of the inductor current of each phase branch and the current inductor current of each phase branch. The available temperature margin is determined based on the difference between the upper limit of the junction temperature of the power switch and the current estimated junction temperature of the power switch. The target distribution current of each phase branch is generated based on the available current margin and the available temperature margin. A current deviation control component is generated based on the deviation between the inductor current of each phase branch and the corresponding target distribution current. A temperature deviation control component is generated based on the deviation between the estimated junction temperature of each phase branch power switch and the estimated average junction temperature. When the drive saturation state is effective, the integral state of the current deviation control component is corrected based on the actual execution duty cycle. The current deviation control component and temperature deviation control component of the same phase branch are combined to generate the original differential mode duty cycle adjustment of each phase branch.

7. The switching power supply control method based on dynamic load tracking according to claim 6, characterized in that, Projecting the original differential mode duty cycle adjustment amount onto the feasible region includes: Based on the inductance parameters of each phase branch in the online parameter status record, a weighted zero-sum constraint for inductance parameters is established, and the feasible region is constructed based on the common mode energy regulation duty cycle component, the prediction constraint margin, the final duty cycle boundary, the upper limit of the inductance current of each phase branch, the upper limit of the junction temperature of the power switch tube, and the boundary of the differential mode duty cycle component. With the minimum weighted deviation between the original differential mode duty cycle adjustment and the current redistribution duty cycle components of each phase differential mode as the projection objective, the current redistribution duty cycle components of each phase differential mode are solved within the feasible region. When the feasible region is empty, the common mode energy regulation duty cycle component is first moved to the feasible boundary closest to the original solution value and reprojected. When there is still no feasible solution after two consecutive control cycles, the duty cycle component of the differential mode current of each phase is decayed cycle by cycle and the common mode energy regulation duty cycle component is reduced when there is still a risk of overcurrent or overtemperature after decaying to zero.

8. The switching power supply control method based on dynamic load tracking according to claim 7, characterized in that, Generate the final duty cycle and drive signal, including: The common mode energy regulation duty cycle component is combined with the differential mode current redistribution duty cycle component of each phase to generate the target duty cycle of each phase branch. The duty cycle is quantized according to the pulse width modulation time resolution. The quantized duty cycle is then used to re-examine the inductor parameter weighted zero-sum constraint and control constraint to generate the final duty cycle of each phase branch. The interleaved phase between the drive signals of each phase branch is determined according to the number of phase branches. The pulse width modulation signal of each phase branch is generated based on the final duty cycle, switching frequency, minimum on time, minimum off time and dead time of each phase branch, and the final duty cycle of the actual execution is recorded. The system detects whether the final duty cycle has reached the duty cycle boundary and whether the current constraint, temperature constraint, differential mode duty cycle component boundary, or solution backoff condition has been activated. Based on the detection results, it generates the drive saturation state.

9. The switching power supply control method based on dynamic load tracking according to claim 8, characterized in that, Generate online parameter status records for use in subsequent control cycles, including: Based on the event occurrence time in the fast event detection data, the actual response data after the final duty cycle operation is time-aligned with the model prediction results. An event-level model mismatch index is generated based on the deviation of the minimum value of the common output voltage, the recovery time deviation, the peak value deviation of the inductor current of each phase branch, the steady-state current distribution deviation, and the net charge deviation of the common output filter capacitor. Following the priority order of sensor anomaly, drive saturation, model mismatch, and normal response, event cause categories are generated based on sensor effective state, drive saturation state, and event-level model mismatch index. The parameter group to be adjusted is determined based on the contribution of each deviation that constitutes the event-level model mismatch index. Based on the event cause category and the parameter group to be adjusted, the forgetting factor, parameter covariance, parameter estimation correction boundary, and parameter freeze state used in recursive identification are adjusted to generate the bounded identification state adjustment amount. The adjusted data is then written into the online parameter state record used in the next control cycle after the bounded identification state adjustment amount is completed.

10. A switching power supply control system based on dynamic load tracking, characterized in that, include: The synchronous sampling and online parameter identification module is used to synchronously sample the input voltage, common output voltage, inductor current and load current of each phase branch of the multi-phase interleaved parallel buck switching power supply, align the case temperature of the power switch tubes of each phase branch, divide the sampled data into fast event detection data and slow parameter identification data, identify the parameters of each phase branch and the parameters of the common output node, and generate online parameter status records. The common mode prediction control module is used to establish a physical model and parameter confidence domain based on online parameter status records, convert the parameter confidence domain into prediction constraint margin, generate a load prediction sequence based on fast event detection data, and generate common mode energy regulation duty cycle components and model prediction results under the common output filter capacitor charge balance constraint. The differential mode constraint allocation and drive module is used to generate the original differential mode duty cycle adjustment amount based on the inductor current of each phase branch, the estimated junction temperature of the power switch, the duty cycle component of the common mode energy regulation and the model prediction results. It projects the original differential mode duty cycle adjustment amount to the feasible region that satisfies the weighted zero-sum constraint of the inductor parameters and the control constraint. After duty cycle quantization and constraint verification, it generates the duty cycle component of the differential mode current redistribution of each phase, the final duty cycle and the drive signal. The event feedback correction module is used to generate an event-level model mismatch index and event cause category based on the actual response data corresponding to the final duty cycle, model prediction results, drive saturation state and sensor effective state. It also generates a bounded identification state adjustment amount based on the event cause category and generates online parameter state records for subsequent control cycles.