Permanent magnet synchronous motor parameter drift compensation device and method based on sensitivity weighting

Abstract

The application belongs to the field of permanent magnet synchronous motor predictive current control, and discloses a parameter drift compensation device and method for permanent magnet synchronous motor based on sensitivity weighting. The device aligns the predictive current state with the voltage action time through single-step delay compensation, constructs an online error index according to the error between the model predictive current and the actual sampling current and the offline parameter sensitivity weight, generates a correction factor through error dead zone, saturation boundary constraint and smoothing processing, corrects the predictive current of the next control period, and generates a reference voltage according to the corrected predictive current. The application does not need to identify motor parameters online and inject additional signals, can inhibit the accumulation of predictive errors caused by parameter drift, and reduce the steady-state current error, current fluctuation and voltage command mutation risk.

Description

Technical Field

[0001] This invention belongs to the field of predictive current control for permanent magnet synchronous motors, and specifically relates to a parameter drift compensation device and method for permanent magnet synchronous motors based on sensitivity weighting. Background Technology

[0002] Predictive current control of permanent magnet synchronous motors relies on motor model parameters to predict current and solve for the reference voltage. In actual operation, stator resistance changes with winding temperature, d-axis and q-axis inductance are affected by magnetic saturation and operating conditions, and permanent magnet flux linkage may also change with temperature and demagnetization. These parameter drifts cause deviations between the model-predicted current and the actual current, leading to increased steady-state current error, current ripple, and torque fluctuations. In actual digital control systems, there are timing links between current sampling, control law calculation, pulse width modulation (PWM) updates, and inverter driving. The voltage command calculated by the controller in the current sampling cycle usually only applies to the inverter output in the next sampling cycle, resulting in a single-sampling-cycle input delay. This delay disrupts the correspondence between the predicted current and the reference voltage in dead-beat predictive current control, further amplifying the current prediction error.

[0003] A common approach is to identify motor parameters online and update the predictive control model in real time. However, online identification typically requires additional observers, recursive estimators, or injected signals, resulting in high computational burden and parameter tuning complexity. To reduce implementation complexity, another approach directly utilizes prediction errors and uses fixed gain to correct the predicted current. However, fixed gain does not differentiate the different effects of stator resistance, d-axis inductance, q-axis inductance, and permanent magnet flux drift on the dq-axis prediction error, and it is difficult to adapt to different parameter sensitivities and measurement noise levels: when the gain setting is small, it is prone to insufficient compensation under parameter-sensitive conditions; when the gain setting is large, it is prone to overcompensation under weakly sensitive or noisy conditions, leading to voltage command abrupt changes, PWM saturation, and current oscillations. On the other hand, single-step delay compensation can only solve the problem of inconsistency between the predicted state and the actual voltage application time, but cannot eliminate model mismatch caused by parameter drift; if single parameter error correction does not consider the digital control link delay, the corrected predicted current may still not match the reference voltage application time. Therefore, timing alignment and parameter drift error correction need to be processed together in the same reference voltage solution link.

[0004] Therefore, a predictive current control scheme is needed that can simultaneously achieve predicted current state timing alignment, differentiated evaluation of parameter drift sensitivity, and limited error correction without introducing online multi-parameter identification and additional signal injection. Summary of the Invention

[0005] To address the issues of single-sampling-cycle input delay in the digital implementation of dead-beat predictive current control for permanent magnet synchronous motors, which leads to inconsistencies between the predicted current state and the actual voltage action time of the inverter, and the accumulation of prediction errors, increased steady-state current deviation, and increased current ripple when stator resistance, d-axis inductance, q-axis inductance, and permanent magnet flux linkage drift occur, this invention provides a sensitivity-weighted parameter drift compensation device and method for permanent magnet synchronous motors.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution.

[0007] The sensitivity-weighted permanent magnet synchronous motor parameter drift compensation device includes: a data acquisition unit, a delay prediction buffer unit, a single-step delay compensation unit, a prediction error acquisition unit, a sensitivity weight calibration unit, an online error index calculation unit, a correction factor generation unit, a correction current prediction unit, a reference voltage generation unit, and a limiting and stabilizing unit.

[0008] The data acquisition unit is used to acquire the phase current, rotor position and electric angular velocity, inverter output status and DC bus voltage of the permanent magnet synchronous motor.

[0009] The delay prediction cache unit is used to store the model prediction current calculated in the previous control cycle and corresponding to the current sampling time, as well as the voltage command output in the previous control cycle.

[0010] The single-step delay compensation unit is used to predict the dq axis predicted current corresponding to the actual voltage action time of the next control cycle based on the dq axis current, electric angular velocity, motor discrete prediction model, and voltage command calculated and cached in the previous control cycle and actually applied in the current control cycle, and to form a delay compensation predicted current for the reference voltage solution of the next control cycle, so that the predicted current state corresponds to the actual voltage action time of the inverter.

[0011] The prediction error acquisition unit is used to subtract the model prediction current obtained in the previous control cycle and corresponding to the current sampling time from the actual current obtained in the current sampling to obtain the dq axis prediction current error.

[0012] The sensitivity weight calibration unit is used to apply perturbations to the stator resistance, d-axis inductance, q-axis inductance and permanent magnet flux during the offline calibration stage, and to statistically analyze the predicted current error under different parameter perturbations and normalize it to obtain the parameter sensitivity weights.

[0013] The online error index calculation unit is used to construct a weighted quadratic online error index based on the dq axis predicted current error, the dq axis error weight matrix obtained by the parameter sensitivity weight, and the normalization factor, so that the parameter drift sensitivity information obtained offline can participate in the error evaluation of the current control cycle.

[0014] The correction factor generation unit is used to perform error dead zone judgment, saturation boundary constraint and smoothing processing according to the online error index, generate the correction factor for the current control cycle, and limit the correction factor to between zero and a preset maximum correction factor, wherein the preset maximum correction factor is less than 1.

[0015] The correction current prediction unit is used to perform limited reconstruction of the next control cycle prediction current obtained by single-step delay compensation based on the correction factor and the dq axis prediction current error, so as to obtain the corrected next control cycle prediction current, and the prediction error correction amount is introduced before the reference voltage is solved.

[0016] The reference voltage generation unit is used to substitute the corrected next control cycle predicted current into the predicted current control voltage equation after single-step delay compensation to generate the reference voltage for the next control cycle.

[0017] The limiting and stabilizing unit is used to limit the correction factor and the reference voltage of the next control cycle, so that the correction factor and the reference voltage are kept within a preset range and the prediction error correction amount is kept bounded.

[0018] The single-step delay compensation unit is used to compensate for the single-sampling-cycle input delay caused by sampling, calculation, and PWM update, so that the predicted current of the next control cycle is aligned with the actual voltage action time of the inverter. The sensitivity weight calibration unit is used to form a priori parameter drift sensitivity. The online error index calculation unit is used to convert the priori parameter drift sensitivity into an error evaluation quantity for the current control cycle. The correction factor generation unit is used to convert the error evaluation quantity into a constrained correction strength. The correction current prediction unit is used to reconstruct the predicted current after delay compensation based on the constrained correction strength and input the reconstructed predicted current into the reference voltage generation unit. Thus, the parameter drift sensitivity information is continuously transmitted along the link of "sensitivity calibration - error weighting - correction strength generation - predicted current reconstruction - reference voltage solution".

[0019] This invention also provides a sensitivity-weighted method for compensating for parameter drift in permanent magnet synchronous motors, comprising the following steps:

[0020] Step 1: Collect the operating status parameters of the permanent magnet synchronous motor, including phase current, rotor position and electric angular velocity, inverter output status and DC bus voltage, and establish a discrete predictive current model of the dq axis of the permanent magnet synchronous motor.

[0021] In the synchronously rotating dq coordinate system, the differential equation for the stator current of the permanent magnet synchronous motor can be expressed as:

[0022] (1);

[0023] in, For differential operators, , These are the d-axis and q-axis currents, respectively. , These are the voltages along the d-axis and q-axis, respectively. For stator resistance, , These are the d-axis and q-axis inductances, respectively. Electric angular velocity, It is a permanent magnet flux linkage.

[0024] Discretization is performed using the first-order forward Euler method, with a sampling period of... , obtained the Single-step prediction model for each sampling period

[0025] (2);

[0026] in, This is the sequence number of the current sampling time. This is the sequence number of the next sampling time. , For the current moment The actual d-axis and q-axis currents obtained from sampling , The next time step predicted by the model The d-axis and q-axis currents. , For the first The d-axis and q-axis voltages applied within each cycle. For the first The electric angular velocity is obtained by sampling in one cycle. The sampling period.

[0027] Traditional deadbeat predictive current control , ,in , The first Given the d-axis and q-axis reference currents at each sampling time, the reference voltage can be derived by inverse solving:

[0028] (3);

[0029] in, , They are respectively with the first Each control cycle corresponds to a d-axis and q-axis reference voltage command. However, in actual digital control systems, there is a time delay, and parameter drift can cause a deviation between the predicted current and the actual current.

[0030] Step 2: Based on the dq axis current, electric angular velocity, motor discrete prediction model, and voltage command calculated and cached in the previous control cycle and actually applied in the current control cycle, predict the dq axis predicted current corresponding to the actual voltage action time in the next control cycle, and form a delay compensation predicted current for solving the reference voltage in the next control cycle, thus realizing single-step delay compensation.

[0031] To compensate for the one-beat delay, this invention employs a single-step advance prediction strategy. In the k-th sampling cycle, based on the current sampling information and voltage command, a discrete prediction model is used to predict the dq-axis predicted current corresponding to the actual voltage application moment of the next control cycle.

[0032] In the presence of a single-sampling-cycle input delay, the voltage command actually applied in the current control cycle is the voltage command calculated and cached in the previous control cycle. This invention uses this voltage command and, combined with the currently sampled dq-axis current and electric angular velocity, obtains the predicted dq-axis current corresponding to the actual voltage application time in the next control cycle through a discrete prediction model.

[0033] The specific implementation of delay compensation is as follows: based on the current sampling information, the discrete model is used to predict the first... Periodic current:

[0034] (4);

[0035] Among them, in equation (4) , The reference voltage command is shown in equation (3); considering the input delay of a single sampling period, the reference voltage command is calculated and cached by the previous control period and is actually applied in the k-th control period; , According to the first The sampled value of the first control cycle and the voltage command prediction obtained above. The uncorrected d-axis and q-axis currents for each control cycle are then used to construct the reference voltage solution for the next control cycle based on the delayed compensation predicted current.

[0036] (5);

[0037] Equation (5) is used to solve for the reference voltage of the next control cycle based on the predicted current of the delay compensation. , Equations (4) and (5) above achieve control delay compensation. However, parameter drift will still cause a deviation between the predicted current and the actual current, requiring further error correction.

[0038] Step 3: Read the model-predicted current obtained from the previous control cycle and corresponding to the current sampling time, and subtract it from the actual current obtained from the current sampling to obtain the dq-axis predicted current error. Define the predicted current error vector as:

[0039] (6);

[0040] in, , The model predicts the current based on the voltage command of the previous cycle and corresponding to the current sampling time. , It refers to the predicted current error along the d and q axes.

[0041] Step 4: In the offline calibration stage, perturbations are applied to the stator resistance, d-axis inductance, q-axis inductance and permanent magnet flux linkage respectively. The predicted current error under different parameter perturbations is statistically analyzed and normalized to obtain the parameter sensitivity weights.

[0042] Since different parameters have varying degrees of influence on prediction error, this invention performs parameter sensitivity analysis offline. For stator resistance... d-axis inductance q-axis inductance and permanent magnet magnetic flux Apply relative perturbations respectively Only one parameter is changed each time, while the others remain at their nominal values. Within the steady-state window after transient decay, the root mean square value of the predicted current error vector is measured and denoted as... The sensitivity index is defined as follows:

[0043] (7);

[0044] in, This is the set of sampling sequence numbers corresponding to the steady-state data window after the transient process has decayed. For the first The relative perturbation amplitudes of each parameter are then determined. The sensitivity weights of each parameter are then obtained through normalization.

[0045] (8);

[0046] in, This represents the summation of sensitivity indices over all parameters. , These are all parameter indices. The root mean square value of the steady-state error differs under different parameter perturbations, and the weighting coefficients are calculated accordingly. Used for subsequent online error index calculation. The parameter sensitivity weights This is used to characterize the impact of different motor parameter drifts on prediction errors, and the d-axis error weight can be calculated based on the sensitivity of the d-axis and q-axis prediction error components. and q-axis error weights This forms the dq-axis error weight matrix used for online error index calculation. .

[0047] Furthermore, to convert the sensitivity weights of each parameter into dq-axis error weights for online error index calculation, the first and second parameters can be statistically analyzed separately. The root mean square values ​​of the d-axis prediction error and the root mean square value of the q-axis prediction error under each parameter perturbation are denoted as follows: and .when When this parameter perturbation contributes to the prediction error components of the d-axis and q-axis, the contribution ratio is defined as follows:

[0048] (9);

[0049] for The parameter perturbations do not participate in the dq axis error weighting calculation.

[0050] Then based on the parameter sensitivity weight Calculate the d-axis error weights and q-axis error weights:

[0051] (10);

[0052] in, , They represent the first The proportion of each parameter perturbation to the prediction error components of the d-axis and q-axis; , The error weights for the d-axis and q-axis are calculated based on the sensitivity weights of the offline parameters, thus forming the dq-axis error weight matrix for online error index calculation. .

[0053] Step 5: During the online operation phase, a weighted quadratic online error index is constructed based on the predicted current error of the dq axis, the error weight matrix of the dq axis obtained by converting the parameter sensitivity weights, and the normalization factor.

[0054] An online error index is constructed based on the prediction error vector of the current period and the offline sensitivity weights. A weighted quadratic form is used:

[0055] (11);

[0056] in, For the first Online error index for each sampling period. For the predicted current error vector The transpose of . This is the normalization factor. , These are the d-axis and q-axis error weights calculated based on the sensitivity weights of the offline parameters.

[0057] Step 6: Determine the error dead zone based on the online error index, and generate a correction factor constrained by the saturation boundary, so that the correction factor is limited to between zero and a preset maximum correction factor, and the preset maximum correction factor is less than 1, and smooth the correction factor.

[0058] Generate an adaptive correction factor based on the online error index. Furthermore, the operating range and rate of change of the correction factor are constrained by error dead zone, saturation boundary, and first-order smoothing.

[0059] (12);

[0060] in, The maximum correction factor is preset and pre-tuned based on the motor's rated parameters and the controller's stability margin. When the prediction error is large, the correction strength is increased; when the prediction error is small, it is automatically reduced to avoid noise amplification.

[0061] Step 7: Correct the predicted current of the next control cycle obtained after single-step delay compensation based on the correction factor after smoothing and the dq axis predicted current error;

[0062] The predicted current for the next control cycle is corrected using a correction factor. The present invention employs the following correction method:

[0063] (13);

[0064] in, , This is the uncorrected predicted current. , This is the corrected predicted current. This correction method utilizes the currently measurable prediction error to suppress the accumulation of current deviation in the next cycle. For ease of explanation, the following will be used for predicted current correction and boundedness analysis. All of these represent the actual correction factor after first-order smoothing.

[0065] Step 8: Substitute the corrected predicted current for the next control cycle into the predicted current control voltage equation to generate the reference voltage for the next control cycle, and output the reference voltage after limiting it to drive the inverter.

[0066] Substituting the corrected predicted current into the dead-beat predicted current control voltage equation, the reference voltage for the next cycle is obtained. , :

[0067] (14);

[0068] When the reference current changes slowly or zero-order hold is used within a sampling period, the reference current in the next control period can be made consistent with the reference current in the current sampling period. When the reference current update is predictable, the reference current in the next control period can also be used in the reference voltage solution. To limit the correction strength and keep the reference voltage within the inverter's output range, this invention limits the correction factor and the generated reference voltage. The uncorrected one-step prediction error is defined as:

[0069] (15);

[0070] in, In the first The prediction obtained in the sampling period is the first Uncorrected periodic predicted current vector. For the first The current vector obtained from the actual sampling of the period. It should be noted that in equation (15) To reach the The actual sampled current obtained after the sampling time is only used to define the uncorrected one-step prediction error and to perform boundedness analysis, and is not used as the first step prediction. The input quantity for online solution of the sampling period reference voltage.

[0071] And assume the existence of a constant. Make .Right now This is the bounded upper bound of the uncorrected one-step prediction error. The error recursion can be derived from the corrected prediction current definition:

[0072] (16);

[0073] in, For the first The periodic corrected prediction error vector, For the first The predicted current vector after period correction. This represents the uncorrected one-step prediction error.

[0074] Take the norm and use ,have:

[0075] (17);

[0076] The recursive solution yields:

[0077] (18);

[0078] therefore:

[0079] (19);

[0080] From equation (19), it can be seen that when the uncorrected one-step prediction error is a bounded quantity, and the actual correction factor after first-order smoothing satisfies... At this point, the error of the prediction error correction subsystem eventually becomes uniformly bounded. Therefore, under the above assumptions, the introduced error correction term will not cause the prediction error to diverge and will help reduce the risk of current oscillations caused by parameter drift. and They represent the first Period and the The norm of the periodic prediction current error vector. for of Power of 1.

[0081] Step 9: Enter the next sampling cycle and repeat the acquisition of running status parameters in Step 1, as well as Steps 2, 3, 5 to 8; the establishment of the discrete predictive current model in Step 1 and the calibration of parameter sensitivity weights in Step 4 are completed in the initialization or offline stage, and the parameter sensitivity weights are called as offline calibration results in the online running stage;

[0082] Specifically, the single-step delay compensation is used to align the predicted current state with the actual voltage application time; the parameter sensitivity weight is used to construct an online error index, which is used to generate a correction factor constrained by error dead zone, saturation boundary, and smoothing processing; the correction factor is used to correct the predicted current of the next control cycle obtained by single-step delay compensation, so that parameter drift sensitivity information is passed step by step through online error evaluation, correction strength determination, and predicted current correction; and the corrected predicted current of the next control cycle is formed before the reference voltage is solved, and is used to replace the uncorrected predicted current in the calculation of the predicted current control voltage equation, so that the prediction error correction is introduced in a limited manner before the reference voltage is solved, and the prediction error caused by parameter drift is suppressed before participating in the voltage solution.

[0083] The core of this invention does not lie in using delay compensation or error feedback alone, but in embedding the prior sensitivity of offline obtained parameters into the online prediction error evaluation, and using a limited correction factor to act on the prediction current reconstruction process before the reference voltage solution, so that the parameter drift sensitivity information is continuously transmitted along the link of "sensitivity calibration - error weighting - correction intensity generation - prediction current reconstruction - reference voltage solution".

[0084] Therefore, this invention utilizes single-step delay compensation to address the inconsistency between the predicted current state and the actual voltage application time. Simultaneously, it employs offline parameter sensitivity weights to characterize the influence of stator resistance, d-axis inductance, q-axis inductance, and permanent magnet flux drift on the predicted current error. Furthermore, the offline parameter sensitivity weights are converted into the correction intensity for the current control cycle through online error indices. Further, correction factors constrained by error dead zones, saturation boundaries, and smoothing processes are used to perform limited corrections on the predicted current for the next control cycle, ensuring that the corrected predicted current is formed before the reference voltage is calculated. This forms a collaborative predictive control mechanism of "timing alignment—sensitivity weighted evaluation—limited correction—suppression before voltage calculation." It should be noted that this invention does not solely employ single-step delay compensation, fixed error feedback, or offline sensitivity calibration. Instead, it allows the offline parameter sensitivity prior to be passed step-by-step through online error indices, correction factors, and predicted current corrections, ultimately acting on the predicted current reconstruction process before the reference voltage is calculated. Through this process, the offline sensitivity prior is no longer merely a simple weighting coefficient but is transformed into a model mismatch correction basis that runs through the entire process of prediction error evaluation, compensation intensity determination, and predicted current reconstruction.

[0085] The beneficial effects of this invention are as follows:

[0086] (1) This invention does not require the introduction of additional observers, recursive estimation algorithms or additional signal injection links. It directly uses the deviation between the predicted current and the actual current to characterize the mismatch state of the motor model, eliminating the need for online identification of parameters such as stator resistance, d-axis and q-axis inductance, and permanent magnet flux linkage. This reduces the computational burden of the controller and the difficulty of algorithm parameter tuning, and meets the real-time control requirements of high-bandwidth current loop.

[0087] (2) Based on the parameter sensitivity weights obtained from offline calibration, the drift of different parameters such as stator resistance, inductance, permanent magnet flux linkage, and the influence of different axial errors of d and q on the current prediction accuracy are quantitatively distinguished. This achieves differentiated matching between correction intensity and prediction error sensitivity, making the compensation strategy more adaptable to the influence of different parameter drifts on prediction error and improving the prediction error correction effect.

[0088] (3) Set saturation boundary constraints for adaptive correction factors, and dynamically adjust the compensation intensity according to the magnitude of the prediction error: when the error is large, increase the correction intensity to quickly suppress the accumulation of deviation; when the error is small, automatically weaken the compensation effect, effectively avoid overcompensation problems, and suppress measurement noise amplification, voltage command sudden change and PWM saturation phenomenon to ensure the smooth operation of the control loop.

[0089] (4) In response to various types of parameter drift conditions caused by motor temperature rise, magnetic saturation, and slight demagnetization of permanent magnets, this compensation strategy can reduce the steady-state current error of the system, suppress current ripple and output torque fluctuation, and improve the steady-state operation quality of the motor.

[0090] (5) In this invention, single-step delay compensation and sensitivity-weighted error correction are set in the same predictive control link before the reference voltage is generated. Single-step delay compensation is used to ensure that the timing of the predicted current state is consistent with the actual voltage action time of the inverter. The sensitivity-weighted online error index is used to convert the influence of different parameter drifts on the prediction error into the correction intensity of the current control cycle. The restricted correction factor is used to perform restricted reconstruction of the predicted current after delay compensation. Thus, the offline parameter sensitivity prior is not used as an isolated weight, but runs through the entire process of prediction error evaluation, compensation intensity generation, prediction current correction and state reconstruction before reference voltage solution, thereby reducing the prediction control error caused by single sampling cycle input delay and parameter drift at the same time.

[0091] (6) This invention does not rely on additional temperature sensors, online multi-parameter identifiers or external injection signals, and can be directly integrated into existing dead-beat predictive current control and model predictive current control frameworks. Through partitioned sensitivity weights, error dead zones, saturation boundaries and first-order smoothing, the correction intensity can be adaptively adjusted according to the operating conditions and prediction error, while reducing the risks of overcompensation, voltage surges, PWM saturation and current oscillation. Attached Figure Description

[0092] Figure 1 This is a block diagram of the parameter drift compensation device for permanent magnet synchronous motor based on sensitivity weighting as described in this invention.

[0093] Figure 2 This is a flowchart of the offline parameter sensitivity weight calibration process of the present invention;

[0094] Figure 3 This is an overall control block diagram of the sensitivity-weighted permanent magnet synchronous motor parameter drift compensation device described in this invention in the closed-loop control system of a permanent magnet synchronous motor.

[0095] Figure 4The diagram shows the dq-axis current response and offline sensitivity analysis results when relative disturbances are applied to the stator resistance, d-axis inductance, q-axis inductance and permanent magnet flux linkage under the same steady-state operating conditions. (a) is the result of the stator resistance disturbance, (b) is the result of the d-axis inductance disturbance, (c) is the result of the q-axis inductance disturbance, and (d) is the result of the permanent magnet flux linkage disturbance.

[0096] Figure 5 The results are simulations of the dynamic tracking of the method of the present invention under a triangular wave velocity command.

[0097] Figure 6 The results are simulations of the dynamic tracking of the method of this invention under a sinusoidal velocity command.

[0098] Figure 7 The results of the present invention method and the traditional deadbeat predictive current control method in a progressive intermittent speed change experiment are compared, where (a) is the speed, phase current and position waveform of the traditional deadbeat predictive current control method, and (b) is the corresponding waveform of the present invention method.

[0099] Figure 8 The results of the present invention method and the traditional deadbeat predictive current control method in the forward and reverse switching experiment are compared, where (a) is the speed, phase current and position waveform of the traditional deadbeat predictive current control, and (b) is the corresponding waveform of the present invention method.

[0100] Figure 9 The results of the present invention method and the traditional deadbeat predictive current control method in a triangular wave dynamic speed change experiment are compared. (a) shows the speed, phase current and q-axis current waveforms of the traditional deadbeat predictive current control method, and (b) shows the corresponding waveforms of the present invention method. Detailed Implementation

[0101] I. Device Structure

[0102] like Figure 1 As shown, the device in this embodiment includes a data acquisition unit, a delay prediction buffer unit, a single-step delay compensation unit, a prediction error acquisition unit, a sensitivity weight calibration unit, an online error index calculation unit, a correction factor generation unit, a correction current prediction unit, a reference voltage generation unit, and a limiting stabilization unit.

[0103] The data acquisition unit is used to acquire the phase current, rotor position and electric angular velocity, inverter output status and DC bus voltage of the permanent magnet synchronous motor, and obtain the dq axis current through coordinate transformation.

[0104] The delay prediction buffer unit is used to store the model prediction current calculated in the previous control cycle and corresponding to the current sampling time, as well as the voltage command output in the previous control cycle.

[0105] The single-step delay compensation unit is used to predict the dq-axis predicted current corresponding to the actual voltage action time of the next control cycle based on the dq-axis current, electric angular velocity, motor discrete prediction model, and voltage command calculated and cached in the previous control cycle and actually applied in the current control cycle. It also forms a delay compensation predicted current for solving the reference voltage in the next control cycle, so that the predicted current state corresponds to the actual voltage action time of the inverter.

[0106] The prediction error acquisition unit is used to subtract the model prediction current obtained in the previous control cycle and corresponding to the current sampling time from the actual current obtained in the current sampling to obtain the d-axis prediction current error and the q-axis prediction current error.

[0107] The sensitivity weight calibration unit is used to offline perturb the stator resistance, d-axis inductance, q-axis inductance and permanent magnet flux linkage respectively, and statistically predict the root mean square error of the current and normalize it to obtain the parameter sensitivity weight.

[0108] The online error index calculation unit is used to construct a weighted quadratic online error index based on the predicted current error of the dq axis, the error weight matrix of the dq axis obtained by the parameter sensitivity weight, and the normalization factor, so that the parameter drift sensitivity information obtained offline can participate in the error evaluation of the current control cycle.

[0109] The correction factor generation unit is used to perform error dead zone judgment, saturation boundary constraint and smoothing processing based on the online error index to generate a restricted and continuously changing correction factor.

[0110] The current correction prediction unit is used to perform limited reconstruction of the next control cycle prediction current obtained by single-step delay compensation based on the correction factor and the dq axis prediction current error, so as to obtain the corrected next control cycle prediction current, and the prediction error correction amount is introduced before the reference voltage is solved.

[0111] The reference voltage generation unit is used to substitute the corrected predicted current of the next control cycle into the predicted current control voltage equation to generate a reference voltage, so that the prediction error caused by parameter drift is suppressed before participating in the voltage solution.

[0112] The limiting and stabilizing unit is used to constrain the correction factor and reference voltage to avoid voltage surges, PWM saturation or current oscillations caused by overcompensation.

[0113] II. Method and Flow

[0114] like Figure 2As shown, the offline sensitivity weight calibration process includes: under typical steady-state conditions, selecting a parameter that has not yet been calibrated, applying a relative perturbation to the selected parameter, and keeping the remaining parameters at their nominal values; after the transient process decays, opening the steady-state data window and collecting the predicted current error. Calculate the root mean square value of the error vector and the sensitivity index of the current parameter; determine whether all parameters have been calibrated. If not, select the next parameter that has not yet been calibrated and repeat the above process; if all parameters have been calibrated, normalize the sensitivity index of each parameter to obtain the parameter sensitivity weights, and convert them into a dq-axis error weight matrix for online error index calculation. Stored to the controller.

[0115] like Figure 3 As shown, this invention embeds an improved dead-beat prediction current control module based on sensitivity weighting into the closed-loop control system of a permanent magnet synchronous motor. Given an electrical angular velocity... With feedback electrical angular velocity The speed error is obtained by subtraction, and the speed error is processed by the speed loop proportional-integral (PI) controller to generate the stator current amplitude command. The d-axis and q-axis current reference values ​​are allocated by the Maximum Torque per Ampere (MTPA) module. , The encoder outputs the rotor electrical angle. This is used for the transformation of three-phase current to the dq coordinate system and the inverse transformation of the dq-axis reference voltage to the αβ coordinate system. Simultaneously, the rotor electrical angle... The electric angular velocity is obtained after processing by the differential module d / dt. An improved dead-beat prediction current control module based on sensitivity weighting uses d-axis and q-axis current reference values. , Actual dq-axis current vector and electric angular velocity Predictive control calculations are performed, and d-axis and q-axis reference voltages are generated based on the predicted current error correction results. , The α and β axis reference voltages are obtained by inverse dq / αβ transformation. , The input is then processed by a Space Vector Pulse Width Modulation (SVPWM) module to generate the inverter drive signal to drive the permanent magnet synchronous motor. The single-step delay compensation, online error index calculation, correction factor generation, and correction current prediction processes within this module are as follows: Figure 1 As shown. Among them, Indicates the current number The actual current vector along the dq axis for each sampling period. This represents the predicted dq-axis current vector obtained from the model and corresponding to the current sampling time. This represents the predicted current error vector; This represents the uncorrected predicted current vector for the next cycle. This represents the corrected predicted current vector for the next cycle, which is used to replace the uncorrected predicted current and substitute it into the voltage equation to obtain the final reference voltage.

[0116] Step 1: Collect the operating status parameters of the permanent magnet synchronous motor, including phase current, rotor position and electric angular velocity, inverter output status and DC bus voltage, and establish a discrete predictive current model of the dq axis of the permanent magnet synchronous motor.

[0117] In the synchronously rotating dq coordinate system, the differential equation for the stator current of the permanent magnet synchronous motor can be expressed as:

[0118] (1);

[0119] Discretization is performed using the first-order forward Euler method, with a sampling period of... , obtained the Single-step prediction model for each sampling period

[0120] (2);

[0121] Traditional deadbeat predictive current control , The reference voltage can be obtained by inverse solving:

[0122] (3);

[0123] in, , They are respectively with the first Each control cycle corresponds to a d-axis and q-axis reference voltage command. However, in actual digital control systems, there is a time delay, and parameter drift can cause a deviation between the predicted current and the actual current.

[0124] Step 2: Based on the dq axis current, electric angular velocity, motor discrete prediction model, and voltage command calculated and cached in the previous control cycle and actually applied in the current control cycle, predict the dq axis predicted current corresponding to the actual voltage action time in the next control cycle, and form a delay compensation predicted current for solving the reference voltage in the next control cycle, thus realizing single-step delay compensation.

[0125] To compensate for the one-beat delay, this invention employs a single-step advance prediction strategy. In the k-th sampling cycle, based on the current sampling information and voltage command, a discrete prediction model is used to predict the dq-axis predicted current corresponding to the actual voltage application moment of the next control cycle.

[0126] In the presence of a single-sampling-cycle input delay, the voltage command actually applied in the current control cycle is the voltage command calculated and cached in the previous control cycle. This invention uses this voltage command and, combined with the currently sampled dq-axis current and electric angular velocity, obtains the predicted dq-axis current corresponding to the actual voltage application time in the next control cycle through a discrete prediction model.

[0127] The specific implementation of delay compensation is as follows: based on the current sampling information, the discrete model is used to predict the first... Periodic current:

[0128] (4);

[0129] Among them, in equation (4) , The reference voltage command is shown in equation (3); considering the input delay of a single sampling period, the reference voltage command is calculated and cached by the previous control period and is actually applied in the k-th control period; , According to the first The sampled value of the first control cycle and the voltage command prediction obtained above. The uncorrected d-axis and q-axis currents for each control cycle are then used to construct the reference voltage solution for the next control cycle based on the delayed compensation predicted current.

[0130] (5);

[0131] Equation (5) is used to calculate the reference voltage for the next control cycle based on the predicted current after delay compensation. Equations (4) and (5) above achieve control delay compensation. However, parameter drift will still cause a deviation between the predicted current and the actual current, requiring further error correction.

[0132] Step 3: Read the model-predicted current obtained from the previous control cycle and corresponding to the current sampling time, and subtract it from the actual current obtained from the current sampling to obtain the dq-axis predicted current error; define the predicted current error vector as:

[0133] (6);

[0134] in, , The model predicts the current based on the voltage command of the previous cycle and corresponding to the current sampling time. , This refers to the d-axis and q-axis predicted current errors. When using the corrected predicted current closed-loop recursion, the corrected predicted current written into the delay prediction buffer unit in the previous control cycle is used as the model predicted current corresponding to the current sampling time in the current control cycle to participate in the prediction error calculation.

[0135] Step 4: Offline calibration of the sensitivity weights of each parameter.

[0136] During the offline calibration phase, perturbations are applied to the stator resistance, d-axis inductance, q-axis inductance, and permanent magnet flux linkage, respectively. The predicted current error under different parameter perturbations is statistically analyzed and normalized to obtain the parameter sensitivity weights. Furthermore, the sensitivity weights can be calibrated according to the motor operating range. Specifically, the motor operating range can be divided into multiple sub-ranges based on speed, load current, or field weakening state, and a sensitivity weight table corresponding to each sub-range can be obtained.

[0137] For example, the operating range can be divided into low-speed, medium-speed, high-speed, and weak magnetic field zones; it can also be divided according to light load, medium load, and heavy load. Using partitioned weights allows the correction intensity to better adapt to the impact of parameter drift on prediction errors under different operating conditions. At the boundaries of adjacent operating ranges, hysteresis switching, linear interpolation, or first-order smoothing can be used to transition the sensitivity weights between adjacent ranges to reduce abrupt changes in online error indicators and correction factors caused by weight switching. Since different parameters have varying degrees of influence on prediction errors, this invention performs parameter sensitivity analysis offline.

[0138] Specifically, regarding the stator resistance d-axis inductance q-axis inductance and permanent magnet magnetic flux Apply relative perturbations respectively Only one parameter is changed each time, while the others remain at their nominal values. Within the steady-state window after transient decay, the root mean square value of the predicted current error vector is measured and denoted as... The sensitivity index is defined as follows:

[0139] (7);

[0140] When all parameters use the same relative disturbance amplitude Then, the sensitivity weights of each parameter are obtained through normalization:

[0141] (8);

[0142] in, This represents the summation of sensitivity indices over all parameters. , These are all parameter indices. The root mean square value of the steady-state error differs under different parameter perturbations, and the weighting coefficients are calculated accordingly. Used for subsequent online error index calculation. The parameter sensitivity weights This is used to characterize the impact of different motor parameter drifts on prediction errors, and the d-axis error weight can be calculated based on the sensitivity of the d-axis and q-axis prediction error components. and q-axis error weights This forms the dq-axis error weight matrix used for online error index calculation. .

[0143] Furthermore, to convert the sensitivity weights of each parameter into dq-axis error weights for online error index calculation, the first and second parameters can be statistically analyzed separately. The root mean square values ​​of the d-axis prediction error and the root mean square value of the q-axis prediction error under each parameter perturbation are denoted as follows: and .when When this parameter perturbation contributes to the prediction error components of the d-axis and q-axis, the contribution ratio is defined as follows:

[0144] (9);

[0145] for The parameter perturbations do not participate in the dq axis error weighting calculation.

[0146] Then based on the parameter sensitivity weight Calculate the d-axis error weights and q-axis error weights:

[0147] (10);

[0148] in, , They represent the first The proportion of each parameter perturbation to the prediction error components of the d-axis and q-axis; , The error weights for the d-axis and q-axis are calculated based on the sensitivity weights of the offline parameters, thus forming the dq-axis error weight matrix for online error index calculation. .

[0149] Step 5: During the online operation phase, a weighted quadratic online error index is constructed based on the dq-axis predicted current error, the dq-axis error weight matrix calculated from the parameter sensitivity weights, and the normalization factor. In one embodiment, a weighted quadratic form is used:

[0150] (11);

[0151] in, For the first Online error index for each sampling period. For the predicted current error vector The transpose of . The normalization factor can be the rated value of the quadratic form of the prediction error, the square value of the rated current, or obtained through offline calibration. , These are the d-axis and q-axis error weights calculated based on the sensitivity weights of the offline parameters.

[0152] Step 6: Determine the error dead zone based on the online error index, and generate a correction factor constrained by a saturation boundary. Limit the correction factor to between zero and a preset maximum correction factor, where the preset maximum correction factor is less than 1. Then, smooth the correction factor.

[0153] (12);

[0154] in, The maximum correction factor (e.g., 0.8) is preset and tuned based on the motor's rated parameters and the controller's stability margin. When the prediction error is large, the correction strength is increased; when the prediction error is small, it is automatically reduced to avoid noise amplification. Furthermore, to prevent frequent adjustments of the correction factor due to small steady-state errors or current sampling noise, an error dead zone threshold is set. The setting can be based on current sampling noise, motor rated current, and controller allowable steady-state error. When the online error index... Less than When, let the correction factor ;when Greater than or equal to Then, a correction factor is generated according to the saturation function. That is:

[0155] when hour, ;

[0156] when hour, .

[0157] Furthermore, to suppress abrupt changes in the correction factor, a first-order smoothing process can be performed on the correction factor to obtain the smoothed correction factor that actually participates in the correction of the predicted current: .

[0158] in, The value of the smoothing coefficient is preferably in the range of 0.6 to 0.9. The larger the value, the more gradual the change in the correction factor; The smaller the value, the faster the correction response. In practical control, it can be... replace Used for correcting the predicted current in the next control cycle. Since the correction factor is generated by a sensitivity-weighted online error index, the sensitivity of different parameter drifts to the prediction error can indirectly affect the predicted current correction process in the next control cycle through the correction factor.

[0159] Step 7: Correct the predicted current for the next control cycle. The predicted current for the next control cycle is corrected using a correction factor. This invention employs the following correction method:

[0160] (13);

[0161] in, , This is the uncorrected predicted current. , This is the corrected predicted current. This correction method utilizes the currently measurable prediction error to suppress the accumulation of current deviation in the next cycle. For ease of explanation, the subsequent use of predicted current correction and boundedness analysis... Both represent the actual correction factor after first-order smoothing. The corrected predicted current for the next control cycle is simultaneously written into the delayed prediction buffer unit and used as the model predicted current corresponding to the current sampling time in the next sampling cycle for prediction error calculation.

[0162] Step 8: Generate a reference voltage and ensure its boundedness.

[0163] Substituting the corrected predicted current into the dead-beat predicted current control voltage equation, we obtain the reference voltage for the next cycle:

[0164] (14);

[0165] When the reference current changes slowly or zero-order hold is used within a sampling period, the reference current in the next control period can be made consistent with the reference current in the current sampling period. When the reference current update is predictable, the reference current in the next control period can also be used in the reference voltage solution. To limit the correction strength and keep the reference voltage within the inverter's output range, this invention limits the correction factor and the generated reference voltage. The uncorrected one-step prediction error is defined as:

[0166] (15);

[0167] in, In the first The prediction obtained in the sampling period is the first Uncorrected periodic predicted current vector. For the first The current vector obtained from the actual sampling of the period. It should be noted that in equation (15) To reach the The actual sampled current obtained after the sampling time is only used to define the uncorrected one-step prediction error and to perform boundedness analysis, and is not used as the first step prediction. The input quantity for online solution of the sampling period reference voltage. The d-axis and q-axis components of the uncorrected predicted current vector are respectively in equation (13) and .

[0168] And assume the existence of a constant. Make From the corrected definition of the predicted current, the error recursion can be obtained:

[0169] (16);

[0170] in, For the first The periodic corrected prediction error vector, For the first The current vector is predicted after periodic correction. It should be noted that in equation (16)... This refers to the measurable prediction error formed by the predicted current after correction in the previous control cycle at the current sampling time. This is the prediction error formed at the next sampling time by the predicted current after correction for the current control cycle. This represents the uncorrected one-step prediction error.

[0171] Take the norm and use ,have:

[0172] (17);

[0173] The recursive solution yields:

[0174] (18);

[0175] therefore:

[0176] (19);

[0177] From equation (19), it can be seen that when the uncorrected one-step prediction error is a bounded quantity, and the actual correction factor after first-order smoothing satisfies At this point, the error of the prediction error correction subsystem eventually becomes uniformly bounded. Therefore, under the above assumptions, the introduced error correction term will not cause the prediction error to diverge and will help reduce the risk of current oscillations caused by parameter drift. and They represent the first Period and the The norm of the periodic prediction current error vector. for of Power of 1. This represents the predicted current error vector. For the current number The actual current vector along the dq axis for each sampling period. This is the predicted dq-axis current vector obtained from the model and corresponding to the current sampling time. The corrected predicted current vector for the next cycle is used to replace the uncorrected predicted current and is substituted into the voltage equation to obtain the final reference voltage. This is the uncorrected dq-axis current vector for the next cycle, predicted by the model.

[0178] Step 9: Update cyclically.

[0179] In each sampling period, the operating status parameter acquisition in step 1 is performed first, followed by steps 2, 3, 5 through 8 in sequence to form a closed-loop online calibration. The discrete predictive current model establishment in step 1 and the parameter sensitivity weight calibration in step 4 are completed in the initialization or offline phase, and the parameter sensitivity weight is used as the offline calibration result in the online operation phase.

[0180] This invention enables offline identification , , and In such cases, the correction strength is adaptively adjusted based on the measurable predicted current error. For parameter drift caused by temperature rise, magnetic saturation, or changes in operating point, this invention can suppress the continuous accumulation of prediction errors, reduce steady-state current deviation and current ripple, and maintain low computational complexity.

[0181] This invention can be applied to dead-beat predictive current control and other model predictive current control frameworks. In the aforementioned control framework, single-step delay compensation is used to resolve one-beat timing misalignment in the digital control link, and sensitivity-weighted error correction is used to suppress prediction errors caused by parameter drift. The two functions act on two different aspects, timing matching and model accuracy, respectively, and synergistically improve the robustness of predictive current control during the reference voltage generation process.

[0182] Experimental verification:

[0183] In this embodiment, the method of the present invention is simulated and experimentally verified based on a current loop control model of a permanent magnet synchronous motor. The parameters of the permanent magnet synchronous motor used are: stator resistance... d-axis inductance q-axis inductance Permanent magnet magnetic flux Number of pole pairs of motor The rated voltage is 80V. A two-level PWM inverter is used in the simulation, with a DC bus voltage of... The PWM switching frequency is 8kHz, and MTPA control is used. Among these, Figures 7 to 9 The speed feedback, position signal, phase current, and q-axis current are used to characterize the speed tracking, current stability, and position response of the control system under different dynamic operating conditions.

[0184] First, offline sensitivity analysis was performed on different motor parameters under the same operating conditions. In the simulation, the permanent magnet synchronous motor accelerated from 0 r / min to 1000 r / min within 0–200 ms, and then maintained steady-state operation, with the load torque set to 10 N·m. While keeping other parameters constant at their nominal values, ±50% relative perturbations were applied to the stator resistance, d-axis inductance, q-axis inductance, and permanent magnet flux linkage, respectively, i.e., the variation range of each parameter was set to 0.5–1.5 times the corresponding nominal value.

[0185] like Figure 4 As shown, under the same disturbance amplitude and operating conditions, changes in stator resistance, d-axis inductance, q-axis inductance, and permanent magnet flux linkage have varying degrees of impact on the d- and q-axis current responses, indicating that the sensitivity of different motor parameter drifts to current prediction and tracking performance differs. Figure 4 (a) shows the result of the stator resistance disturbance. Figure 4 (b) shows the result of the d-axis inductance disturbance. Figure 4 (c) shows the result of the q-axis inductance disturbance. Figure 4 (d) shows the result of the permanent magnet flux disturbance.

[0186] After the transient process of each parameter disturbance test decays, a steady-state data window is selected. The dq-axis predicted current error is constructed using the model predicted current and the actual sampled current. The root mean square value of the d-axis prediction error and the root mean square value of the q-axis prediction error are calculated respectively. The root mean square value of the error and the parameter disturbance amplitude are further normalized according to equations (7) to (10) to obtain the sensitivity weights corresponding to each parameter and the dq-axis error weight matrix used for online error index calculation.

[0187] Secondly, the dynamic tracking performance of the method of this invention was verified under a triangular wave velocity command. The simulation conditions were: DC bus voltage... The PWM switching frequency is 8kHz, MTPA control is used, and the load torque is 10N·m. The motor accelerates from 0 r / min to 1200 r / min within 0–2.0s, then decelerates to 0 r / min, and then accelerates back to 1200 r / min, repeating this cycle 5 times before stopping. Figure 5 As shown, the method of the present invention has a faster speed response during acceleration and deceleration, and the d-axis and q-axis current fluctuation range is smaller than that of the traditional dead-beat predictive current control method.

[0188] Next, the speed tracking performance of the method of the present invention was verified under a sinusoidal speed command. The simulation conditions were the same as those for the triangular wave speed command. Figure 6 It can be seen that the method of the present invention can follow the changes of sinusoidal speed command and is superior to the traditional dead-beat predictive current control method in terms of speed tracking and current stability.

[0189] Furthermore, an experimental platform based on the RX24T control module was built for experimental verification. The experimental platform includes a power supply circuit, a main control circuit, an inverter drive circuit, a current sampling circuit, and a communication circuit, and consists of a power supply, a permanent magnet synchronous motor, a mechanical transmission device, a host computer, and an encoder to form an AC servo control system for the permanent magnet synchronous motor.

[0190] (1) In the progressive intermittent speed change experiment, such as Figure 7 As shown in (a) and (b), the motor operates under a 10 N·m positive load, with speed commands sequentially set to 500 r / min, 1000 r / min, 1500 r / min, 1000 r / min, and 500 r / min. Each speed stage lasts for 5–6 seconds, after which the motor stops. Speed, phase current, and position signals are recorded experimentally. Experimental results show that the method of this invention can maintain stable operation during multi-stage speed changes and reduce current fluctuations during the speed change process.

[0191] (2) In the forward and reverse switching experiment, such as Figure 8 As shown in (a) and (b), the motor operates under a load of 10 N·m, initially running at 1000 r / min in the forward direction, switching to -1000 r / min after 7–8 seconds, and then stopping. Experimental results show that the method of the present invention can maintain stable control even when the speed direction changes abruptly, avoiding sudden instability in speed and current.

[0192] (3) In dynamic periodic speed change experiments, such as Figure 9 As shown in (a) and (b), the motor operates under a 10 N·m positive load. The speed command is from 0 r / min to 1200 r / min within 2 seconds, then decelerates to 0 r / min, and then accelerates back to 1200 r / min, repeating this cycle. Experimental results show that the motor maintains rapid response and stable operation during sudden speed changes and cyclic speed variations. The phase current maintains sinusoidal characteristics without significant distortion, and no sudden jumps occur in current or speed. These results indicate that the method of this invention helps improve the stability of predicted current and dynamic tracking performance under parameter drift conditions.

[0193] The simulation and experimental results above demonstrate that the method of the present invention can suppress the accumulation of prediction errors and improve the operational stability of predictive current control under conditions of parameter drift and dynamic speed variation.

[0194] The foregoing has shown and described the basic principles and main features of the present invention. Those skilled in the art should understand that the above embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention; various changes and modifications can be made by those skilled in the art without departing from the spirit and essence of the present invention, and all such changes and modifications should fall within the scope of protection of the present invention.

Claims

1. A parameter drift compensation device for a permanent magnet synchronous motor based on sensitivity weighting, characterized in that, It includes a data acquisition unit, a delay prediction buffer unit, a single-step delay compensation unit, a prediction error acquisition unit, a sensitivity weight calibration unit, an online error index calculation unit, a correction factor generation unit, a correction current prediction unit, a reference voltage generation unit, and a limiting and stabilizing unit; The data acquisition unit is used to acquire the phase current, rotor position and electric angular velocity, inverter output status and DC bus voltage of the permanent magnet synchronous motor. The delay prediction cache unit is used to store the model prediction current calculated in the previous control cycle and corresponding to the current sampling time, as well as the voltage command output in the previous control cycle. The single-step delay compensation unit is used to predict the dq axis predicted current corresponding to the actual voltage action time of the next control cycle based on the dq axis current, electric angular velocity, motor discrete prediction model, and voltage command calculated and cached in the previous control cycle and actually applied in the current control cycle, and to form a delay compensation predicted current for the reference voltage solution of the next control cycle, so that the predicted current state corresponds to the actual voltage action time of the inverter. The prediction error acquisition unit is used to subtract the model prediction current obtained in the previous control cycle and corresponding to the current sampling time from the actual current obtained in the current sampling to obtain the dq axis prediction current error. The sensitivity weight calibration unit is used to apply perturbations to the stator resistance, d-axis inductance, q-axis inductance and permanent magnet flux during the offline calibration stage, and to statistically analyze the predicted current error under different parameter perturbations and normalize it to obtain the parameter sensitivity weights. The online error index calculation unit is used to construct a weighted quadratic online error index based on the dq axis predicted current error, the dq axis error weight matrix obtained by the parameter sensitivity weight, and the normalization factor, so that the parameter drift sensitivity information obtained offline can participate in the error evaluation of the current control cycle. The correction factor generation unit is used to perform error dead zone judgment, saturation boundary constraint and smoothing processing according to the online error index, generate the correction factor for the current control cycle, and limit the correction factor to between zero and a preset maximum correction factor, wherein the preset maximum correction factor is less than 1. The correction current prediction unit is used to perform limited reconstruction of the next control cycle prediction current obtained by single-step delay compensation based on the correction factor and the dq axis prediction current error, so as to obtain the corrected next control cycle prediction current, and the prediction error correction amount is introduced before the reference voltage is solved. The reference voltage generation unit is used to substitute the corrected next control cycle predicted current into the predicted current control voltage equation after single-step delay compensation to generate the reference voltage for the next control cycle. The limiting and stabilizing unit is used to limit the correction factor and the reference voltage of the next control cycle, so that the correction factor and the reference voltage are kept within a preset range and the prediction error correction amount is kept bounded. The single-step delay compensation unit is used to compensate for the single-sampling cycle input delay caused by sampling, calculation and PWM update, so that the predicted current of the next control cycle is aligned with the actual voltage action time of the inverter. The sensitivity weighting calibration unit is used to form a priori parameters for drift sensitivity. The online error index calculation unit is used to convert the priori parameters for drift sensitivity into an error evaluation quantity for the current control cycle. The correction factor generation unit is used to convert the error evaluation quantity into a constrained correction strength. The correction current prediction unit is used to reconstruct the predicted current after delay compensation based on the constrained correction strength and input the reconstructed predicted current into the reference voltage generation unit. Thus, the parameter drift sensitivity information is continuously transmitted along the link of "sensitivity calibration - error weighting - correction strength generation - predicted current reconstruction - reference voltage solution".

2. The parameter drift compensation device for permanent magnet synchronous motor based on sensitivity weighting according to claim 1, characterized in that, The sensitivity weight calibration unit, under the condition that only one parameter is disturbed each time while the other parameters remain unchanged, collects the root mean square value of the norm of the prediction current error vector within the steady-state window, and normalizes the root mean square value corresponding to each parameter according to the disturbance amplitude to obtain the parameter sensitivity index; the sensitivity weight calibration unit divides each parameter sensitivity index by the sum of all parameter sensitivity indices to obtain a sensitivity weight that is non-negative and sums to 1, and forms a weight matrix for weighting the dq axis prediction error based on the sensitivity weight, the weight matrix consisting of d-axis error weight and q-axis error weight.

3. The parameter drift compensation device for permanent magnet synchronous motor based on sensitivity weighting according to claim 1, characterized in that, The online error index calculation unit constructs a weighted quadratic online error index based on the dq-axis predicted current error vector, the dq-axis error weight matrix obtained by converting parameter sensitivity weights, and a normalization factor. The correction factor generation unit performs error dead zone judgment based on the online error index. When the online error index is less than a preset error dead zone threshold, the correction factor is set to zero. When the online error index is greater than or equal to the preset error dead zone threshold, a correction factor that increases with the online error index is generated based on a saturation function, and the correction factor is limited to between zero and a preset maximum correction factor, wherein the preset maximum correction factor is less than 1. The correction current prediction unit uses the correction factor to restrict the introduction of dq-axis predicted current error, and subtracts the corresponding error correction amount from the predicted current of the next control cycle obtained after single-step delay compensation to obtain the corrected predicted current of the next control cycle.

4. A method for compensating for parameter drift in a permanent magnet synchronous motor based on sensitivity weighting, characterized in that, Includes the following steps: Step 1: Collect the operating status parameters of the permanent magnet synchronous motor, including phase current, rotor position and electric angular velocity, inverter output status and DC bus voltage, and establish a discrete predictive current model of the dq axis of the permanent magnet synchronous motor. Step 2: Based on the dq axis current, electric angular velocity, motor discrete prediction model, and voltage command calculated and cached in the previous control cycle and actually applied in the current control cycle, predict the dq axis predicted current corresponding to the actual voltage action time in the next control cycle, and form a delay compensation predicted current for solving the reference voltage in the next control cycle, thus realizing single-step delay compensation. Step 3: Read the model predicted current obtained in the previous control cycle and corresponding to the current sampling time, and subtract it from the actual current obtained in the current sampling to obtain the dq axis predicted current error; Step 4: In the offline calibration stage, perturbations are applied to the stator resistance, d-axis inductance, q-axis inductance and permanent magnet flux linkage respectively. The predicted current error under different parameter perturbations is statistically analyzed and normalized to obtain the parameter sensitivity weights. Step 5: During the online operation phase, a weighted quadratic online error index is constructed based on the dq-axis predicted current error, the dq-axis error weight matrix calculated from the parameter sensitivity weights, and the normalization factor. Step 6: Determine the error dead zone based on the online error index, and generate a correction factor constrained by the saturation boundary, so that the correction factor is limited to between zero and a preset maximum correction factor, and the preset maximum correction factor is less than 1, and smooth the correction factor. Step 7: Correct the predicted current of the next control cycle obtained after single-step delay compensation based on the correction factor after smoothing and the dq axis predicted current error; Step 8: Substitute the corrected predicted current for the next control cycle into the predicted current control voltage equation to generate the reference voltage for the next control cycle, and output the reference voltage after limiting it to drive the inverter. Step 9: Enter the next sampling cycle and repeat the acquisition of running status parameters in Step 1, as well as Steps 2, 3, 5 to 8; the establishment of the discrete predictive current model in Step 1 and the calibration of parameter sensitivity weights in Step 4 are completed in the initialization or offline stage, and the parameter sensitivity weights are called as offline calibration results in the online running stage; Specifically, the single-step delay compensation is used to align the predicted current state with the actual voltage application time; the parameter sensitivity weight is used to construct an online error index, which is used to generate a correction factor constrained by error dead zone, saturation boundary, and smoothing processing; the correction factor is used to correct the predicted current of the next control cycle obtained by single-step delay compensation, so that parameter drift sensitivity information is passed step by step through online error evaluation, correction strength determination, and predicted current correction; and the corrected predicted current of the next control cycle is formed before the reference voltage is solved, and is used to replace the uncorrected predicted current in the calculation of the predicted current control voltage equation, so that the prediction error correction is introduced in a limited manner before the reference voltage is solved, and the prediction error caused by parameter drift is suppressed before participating in the voltage solution.

5. The method for compensating for parameter drift of a permanent magnet synchronous motor based on sensitivity weighting according to claim 4, characterized in that, The sensitivity weights in step 4 are obtained through offline parameter perturbation experiments. These experiments include changing the stator resistance, d-axis inductance, q-axis inductance, and permanent magnet flux linkage, and calculating the root mean square (RMS) value of the norm of the predicted current error vector under different parameter perturbations. The RMS values ​​corresponding to different parameters are normalized according to the perturbation amplitude to obtain parameter sensitivity indices. Each parameter sensitivity index is divided by the sum of all parameter sensitivity indices to obtain the parameter sensitivity weight.

6. The method for compensating for parameter drift of a permanent magnet synchronous motor based on sensitivity weighting according to claim 4, characterized in that, In step 6, it is determined whether the online error index is less than a preset error dead zone threshold. When the online error index is less than the preset error dead zone threshold, the correction factor is set to zero. When the online error index is greater than or equal to the preset error dead zone threshold, a correction factor that increases with the online error index is generated according to the saturation function, and the correction factor is limited to between zero and a preset maximum correction factor, wherein the preset maximum correction factor is less than 1. The correction factor is further subjected to first-order smoothing to obtain a smoothed correction factor that actually participates in the prediction current correction of the next control cycle.

7. The method for compensating for parameter drift of a permanent magnet synchronous motor based on sensitivity weighting according to claim 4, characterized in that, In step 8, after substituting the corrected next control cycle predicted current into the predicted current control voltage equation to generate the next control cycle reference voltage, the next control cycle reference voltage is limited to keep the voltage command within the inverter's output voltage range and reduce the risk of voltage surges, PWM saturation, or current oscillations caused by overcompensation.

Images