A Model-Free Predictive Current Control Method Based on an Improved ARX Model
By improving the ARX model, explicitly removing cross-coupling terms and introducing electric angular velocity, and combining SCAD-QRRLS and second-order feedforward error compensation, the problem of insufficient dynamic adaptability and robustness of traditional ARX-MFPCC in permanent magnet synchronous motor control is solved, achieving faster dynamic response and higher steady-state accuracy.
Patent Information
- Application Number
- CN202611020709.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-09
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional ARX-MFPCC has problems in permanent magnet synchronous motor control, such as incomplete model structure, difficulty in parameter identification, and inability to eliminate residual prediction errors, resulting in insufficient dynamic adaptability and robustness.
By improving the ARX model, explicitly removing the cross-coupling terms between the d and q axes, introducing electric angular velocity as an extended input, using the SCAD-QRRLS algorithm for parameter identification, and designing a second-order feedforward error compensation circuit, an improved model-free predictive current control method is constructed.
It improves the dynamic response speed under varying speed and load conditions, reduces current overshoot and steady-state current ripple, enhances the robustness and steady-state accuracy of the control system, and is suitable for real-time implementation on low-cost hardware.
Smart Images

Figure CN123122355A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of permanent magnet synchronous motor control technology, and in particular to a model-free predictive current control method based on an improved ARX model. Background Technology
[0002] Permanent magnet synchronous motors (PMSMs) are widely used in high-end equipment manufacturing fields such as industrial robots, electric vehicles, and CNC machine tools due to their high power density, high efficiency, and excellent speed regulation performance. As the core of the PMSM drive system, current loop control directly determines the overall system performance through its dynamic response speed and steady-state accuracy.
[0003] Traditional field-oriented control relies on proportional-integral (PI) controllers, which are cumbersome to tune and whose dynamic performance is easily affected by changes in motor parameters and load disturbances. Model predictive current control uses a physical model of the motor to predict future current, offering advantages such as fast dynamic response and flexible control objectives. However, its control performance is highly dependent on the accuracy of physical parameters such as motor resistance, inductance, and flux linkage. In actual motor operation, factors such as temperature rise, magnetic saturation, and load changes can cause parameters to deviate from nominal values, resulting in predictive model mismatch and consequently, steady-state current error and dynamic tracking error.
[0004] To overcome the dependence on physical parameters, model-free predictive control (MFPC) has emerged, which uses only the system's input and output data to construct a data-driven model. Among them, the MFPC scheme based on the ARX model is considered to have good engineering application potential due to its clear model structure, online parameter updates through recursive least squares, and ability to capture system dynamics through higher-order terms.
[0005] However, traditional ARX-MFPCC technology still has the following inherent drawbacks:
[0006] First, the model structure is incomplete: traditional ARX models only describe the input-output relationship between voltage and current, without explicitly considering the speed-related back EMF term in the PMSM and the cross-coupling effect between the d and q axes. This results in a longer reconvergence time under variable speed or load conditions, and significant current overshoot and tracking errors during dynamic processes.
[0007] Secondly, high-order models are difficult to identify: to improve model accuracy, high-order ARX models are often required. However, high-order models introduce redundant lag terms and spurious non-zero parameters, making the model sensitive to noise, increasing the variance of parameter estimation, and even causing numerical ill-conditioning, thus affecting the stability of control.
[0008] Third, residual prediction errors cannot be eliminated: even after parameter identification converges, factors such as sampling noise, inverter nonlinearity, and unmodeled disturbances will still introduce residual prediction errors. Especially during steady-state operation, the forgetting factor set to suppress parameter fluctuations will slow down the parameter update speed, causing the model to lag in response to changes in operating conditions and have insufficient dynamic compensation capability.
[0009] In summary, the existing ARX-MFPCC technology has significant shortcomings in terms of dynamic adaptability, parameter identification robustness, and steady-state error suppression. There is an urgent need for an improved model-free predictive current control method that can comprehensively solve the above problems. Summary of the Invention
[0010] In view of this, the purpose of this invention is to provide a model-free predictive current control method based on an improved ARX model. This method systematically innovates the traditional ARX-MFPCC from three dimensions: model structure extension, parameter identification algorithm improvement, and error feedforward compensation, so as to achieve high-performance and robust control of the PMSM current loop.
[0011] The present invention solves the technical problem by adopting the following technical solution:
[0012] A model-free predictive current control method based on an improved ARX model includes the following steps:
[0013] Step S1: Collect the dq-axis stator current of the permanent magnet synchronous motor at the current time k. , Stator voltage , and electric angular velocity ;
[0014] Step S2: Construct an improved ARX prediction model. The improved ARX prediction model takes the current and historical dq-axis voltages, historical dq-axis currents, and current and historical electric angular velocities as inputs, and predicts the dq-axis current at the next moment as output. Specifically, the known cross-coupling terms between the dq axes are explicitly removed from the improved ARX prediction model structure. and and electric angular velocity As an independent extended input variable;
[0015] Step S3: The orthogonal triangular recursive least squares algorithm based on smooth truncation of absolute bias regularization is used to update the parameter vector of the improved ARX prediction model online using the collected input and output data;
[0016] Step S4: Design a second-order feedforward error compensation circuit, calculate the current prediction error at the current moment, and filter it through two cascaded first-order low-pass filters to obtain the second-order filtering error, which is then converted into the feedforward compensation amount after gain and limiting.
[0017] Step S5: Use the feedforward compensation amount to correct and improve the predicted current of the ARX prediction model, or correct the current command value, to obtain the compensated current prediction value or the compensated current command.
[0018] Step S6: Based on the compensated current prediction value or the compensated current command, and combined with the improved ARX prediction model, the dq axis voltage command to be applied at the next moment is solved in reverse and used to control the permanent magnet synchronous motor.
[0019] Furthermore, the improved ARX prediction model constructed in step S2 has parameter vectors for its d-axis and q-axis. Extended definition: ;
[0020] in, for The autoregressive coefficients in the shaft current model that are related to past output current. In the formula, Input for electric angular velocity right The influence coefficient of the shaft-identified output current, where for The influence coefficient of the shaft back electromotive force term on current dynamics. Used as The velocity-related effects caused by axis parameter uncertainties or residual coupling are diluted; and It is the order of the polynomial;
[0021] Extended regression vectors along the d-axis and q-axis and for: ;
[0022] in, for Stator voltage in a synchronous rotating coordinate system; for Stator current in a synchronous rotating coordinate system For discrete time intervals.
[0023] Furthermore, based on the improved ARX model, targeting... shaft and The axes each have known coupling terms. and To obtain the predicted current value at the next moment. and : ;
[0024] At higher sampling frequencies, the rotational speed is approximately considered constant within one sampling period, and the output voltage is calculated using the following formula: ;
[0025] in, The corresponding d-axis and q-axis are respectively at Current command value at any time This is the current control cycle.
[0026] Furthermore, the cost function of the orthogonal triangular recursive least squares algorithm with smooth truncation and absolute bias regularization in step S3 is... Represented as: ;in, Forgetting factor, Let φ(t) be the system output at time t, and φ(t) be the regression vector at time t. The j-th parameter in the parameter vector, The length of the parameter vector. For SCAD regularization functions, Let the vector be the parameters to be determined.
[0027] The SCAD regular function is shown in the following formula: ;
[0028] Its derivative is:
[0029] ;
[0030] in, These are the curve shape parameters; This is the regularization parameter.
[0031] Furthermore, in step S3, an orthogonal triangular decomposition is achieved using a recursive update method based on Givens rotation, so as to recursively save and update the upper triangular regularization matrix. And regularized orthogonal trigonometric decomposition of the right-hand vector This avoids the direct inversion and recursion of the covariance matrix; parameter estimation is achieved by solving the upper triangular equations. get.
[0032] Furthermore, during the recursive update using Givens rotations, a data row QR update is performed on the sampled data at time k: ;
[0033] in, This is the middle upper triangular regularized matrix. The right-hand vector of the intermediate regularized QR decomposition. For the k-th sampled data, It is a zero row vector. For scalar terms that are independent of the parameter solution;
[0034] Then, row-by-row regularization is used for updating. Let the current regularization direction be the i-th row. One parameter direction, For the first Given unit vectors, construct SCAD regularized rows: ;
[0035] And perform a second QR update:
[0036] in, For the first Weights in each parameter direction;
[0037] By order exist The parameters change cyclically, and the regularization update of all parameter directions is gradually completed within multiple sampling periods. Finally, the parameters are determined by the formula. It was obtained by going back to the source.
[0038] Furthermore, the discrete transfer function of the second-order feedforward error compensation stage in step S4 is: ;in, These are the filter coefficients, and 0 < <1;
[0039] The filtered prediction error is used to construct the feedforward compensation: ;in, To compensate for the gain; To compensate for the bandwidth limitation; This is the amplitude limiting function.
[0040] Furthermore, in step S5, the compensated predicted current for: ;
[0041] Simultaneously, the reference value is compensated accordingly in the voltage inverse solution to offset the influence of model residual error on the control voltage; the compensated current command is defined. for: .
[0042] The model-free predictive current control method based on an improved ARX model disclosed in this invention has the following beneficial effects:
[0043] (1) This invention solves the problems of slow dynamic response and large current overshoot caused by the need for parameter re-convergence under variable speed and load conditions in traditional ARX-MFPCC by introducing a velocity term and removing coupling terms through model extension. Simulation and experimental results show that, compared with traditional ARX-MFPCC, the method of this invention reduces the ITAE index of q-axis current by more than 30% and the current overshoot by more than 80% when the speed changes abruptly, achieving a dynamic response speed close to that of DPCC that relies on an accurate physical model, without overshoot.
[0044] (2) The present invention realizes the sparse identification of high-order models through the SCAD-QRRLS algorithm, eliminates redundant parameters, reduces the sensitivity of the model to noise, and combines the second-order feedforward error compensation to smooth the residual error, so that the steady-state current ripple is significantly reduced, the total harmonic distortion rate is lower, and the running stability and efficiency of the motor are improved.
[0045] (3) This invention is entirely data-driven and does not rely on physical parameters such as motor resistance, inductance, and flux linkage. The improved model structure and SCAD regularization algorithm ensure the consistency and effectiveness of model parameters under different operating conditions. Even if the motor parameters change significantly due to temperature rise, magnetic saturation, etc., the control system can maintain excellent current tracking performance, which is unmatched by the traditional MPCC that relies on physical models.
[0046] (4) This invention uses QR-RLS instead of traditional RLS, combined with Givens rotational recursive updates, which greatly improves the numerical stability of parameter identification under finite word length conditions and avoids the risk of ill-conditioned divergence of the covariance matrix. This makes the algorithm easy to implement in real time on low-cost fixed-point DSPs or FPGAs, and has high engineering application value.
[0047] (5) The three improvement measures proposed in this invention, namely “model extension, SCAD-QRRLS identification and second-order feedforward compensation”, can be implemented as a complete solution to obtain the best performance, or can be modularly selected and integrated according to specific application requirements and hardware conditions, which is highly flexible. Attached Figure Description
[0048] Figure 1 This is a flowchart of the method of the present invention.
[0049] Figure 2 This is the SCAD regular function curve.
[0050] Figure 3 This is a flowchart of the SCAD-QRRLS algorithm implementation.
[0051] Figure 4 The flowchart shows the model-free predictive current control strategy based on the improved ARX model.
[0052] Figure 5 The waveform diagram shows the dynamic response simulation of the ARX-MFPCC at varying speeds.
[0053] Figure 6 The waveform diagram is a simulation result of the dynamic response of the EARX-MFPCC at varying speeds.
[0054] Figure 7 A comparison chart of performance evaluation indicators for ARX-MFPCC and EARX-MFPCC control strategies under variable speed conditions.
[0055] Figure 8 The waveforms of current and speed changes corresponding to the DPCC algorithm under the condition of sudden speed change are shown.
[0056] Figure 9 The waveforms of current and speed changes corresponding to the ARX-MFPCC algorithm under the condition of sudden speed change are shown.
[0057] Figure 10 The waveforms of current and speed changes corresponding to the EARX-MFPCC algorithm under the condition of sudden speed change are shown.
[0058] Figure 11 This is a comparison chart of the dynamic performance indicators of different control strategies under variable speed conditions. Detailed Implementation
[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0060] Traditional ARX model-based model-free predictive current control methods describe the dynamics of motor current by establishing an ARX model driven by system input and output data, thereby achieving current prediction and control and eliminating dependence on motor physical parameters.
[0061] SPMSM in The equations of the synchronously rotating coordinate system are:
[0062] In the formula, for Stator voltage in a synchronous rotating coordinate system; for Stator current in a synchronous rotating coordinate system; for Stator flux linkage in a synchronous rotating coordinate system;
[0063] The stator flux linkage equation is:
[0064] The electromagnetic torque equation is:
[0065] In the formula, For SPMSM electromagnetic torque, It is an extreme logarithm.
[0066] Discretizing the above equation using the zero-order holder (ZOH) method yields the discrete-time current dynamic equation:
[0067] In the formula, The current control period is selected in this invention as being equal to the current sampling period. For discrete time intervals.
[0068] For a multi-input multi-output (MIMO) system, the input vector All elements in the vector are strongly coupled; they all affect the output vector simultaneously. The discretized dynamic model of this system can be expressed as:
[0069] in It is the denominator of the transfer function. It is the numerator of the transfer function corresponding to each input and output element, and its form is:
[0070] here and It is the order of the polynomial. and These are the coefficients of the polynomial, where the subscripts are... Indicates the number of the input element. This indicates the number of the output element. Typically, the order of the denominator is set to be greater than or equal to the order of the numerator (i.e.,...). This is to ensure that the system is causal. and The choice between order and model accuracy requires a trade-off: increasing the order can yield a more accurate model, but the computational burden will increase dramatically due to the increased matrix dimension of the RLS algorithm.
[0071] For the current control system of a permanent magnet synchronous motor, its input is shaft voltage, output is Shaft current. The discrete dynamic model of the permanent magnet synchronous motor is as follows:
[0072] In the formula , for Shaft voltage; , for shaft current; , , , The numerator polynomial of the discrete transfer function; , It is a denominator polynomial.
[0073] The above polynomial is obtained by using the polynomial in the above formula. and Replace with and The above model can simultaneously describe... The coupling effect of shaft voltage on current. The coefficients in the discrete dynamic model of a permanent magnet synchronous motor can be obtained through RLS iteration. RLS minimizes the weighted sum of squared errors from the initial time to the current time by recursively updating the parameter estimates, and has the advantages of fast convergence speed and low computational resource consumption. Its recursive formula is as follows:
[0074] In the formula, This is the model parameter estimation vector; This represents the prediction error; For regression vectors; It is the covariance matrix; Update the gain for the parameters; It is a forgetting factor.
[0075] The RLS algorithm can update the ARX model parameters in real time using the latest input and output data in each sampling period, thereby enabling online learning of the motor's dynamic characteristics.
[0076] for The parameter vector of the shaft current system can be expressed as:
[0077] In the formula, They are respectively shaft and The parameter vector of the axis. for The autoregressive coefficients in the shaft current model that are related to past output current. ; Input voltage For output current The influence coefficient, where The corresponding regression vectors along the d-axis and q-axis. for:
[0078] Based on the ARX model described above, the predicted current value for the next time step can be obtained. and for:
[0079] Based on the identified ARX model and current reference value, the output voltage at the next moment is obtained using the following formula. and for:
[0080] In the formula, and They are respectively Current command value at any given time.
[0081] The flow of the model-free predictive current control algorithm based on the ARX model is as follows: Figure 1 As shown, this algorithm is completely independent of any physical parameters and can identify disturbances and changes in system parameters through online iteration, thereby greatly improving the control performance of the system under parameter mismatch. Furthermore, the higher-order terms of the ARX model can more accurately capture the higher-order dynamic and temporal characteristics of the system, giving ARX-MFPCC high control accuracy in model-free predictive current control.
[0082] Although the MFPCC based on the ARX model achieves high current prediction accuracy under steady-state conditions, thus exhibiting good steady-state control performance, the lack of explicit modeling of the speed term makes the algorithm quite sensitive to speed loop disturbances. When reference values, loads, or operating states change, deviations in the model's identification will affect prediction performance, making it difficult for the original model to maintain consistent prediction capabilities over short periods, potentially impacting control effectiveness. Therefore, in practical applications, it is necessary to further improve and refine the algorithm to enhance its adaptability and robustness in dynamic processes, thereby better avoiding potential transient instability problems and improving overall control performance.
[0083] Although traditional ARX-MFPCC models typically achieve high current prediction accuracy under steady-state operating conditions, the model identified by the RLS algorithm may not necessarily be consistent with the actual motor model. On one hand, the discrete dynamic model of a permanent magnet synchronous motor (PMSM) only models the input-output relationship between voltage and current, while the actual electrical model of a PMSM also includes a speed-related back electromotive force term and... Cross-coupling terms between axes. These unmodeled dynamic factors cause the system to exhibit different equivalent dynamic characteristics under different operating conditions, resulting in fundamental differences in the identified model parameters at different operating points. When the system's operating state changes, such as dynamic changes in speed or load, the parameter vector often needs to be readjusted, leading to a secondary convergence process. During the stage before the parameters have reconverged, the model prediction error will increase significantly, thus affecting the dynamic performance of the controller.
[0084] On the other hand, to fully capture the dynamic characteristics of the motor under different operating conditions, traditional methods often choose a high ARX model order. However, high-order ARX models also have problems. First, overparameterization of the model may lead to instability in the obtained model, even if the identification process itself is stable. Second, since the motor itself is a low-order model, the ideal parameter vector should be sparse, but the identification of high-order models often generates multiple canceling zeros and poles. Due to noise, sampling delay, and zero-pole cancellation, these parameters that should theoretically be zero or close to zero will be estimated as non-zero values. At this time, the model is more sensitive to noise, the coefficient fluctuations are aggravated, the convergence speed decreases, and it may even affect the stability of the voltage inverse solution. Finally, during the steady-state operation of the motor, when the input signal does not have continuous excitation characteristics or collinearity exists, the covariance matrix of the input signal may become ill-conditioned or even singular, which will lead to the risk of divergence and increased variance of the estimated coefficients.
[0085] Therefore, this invention addresses the shortcomings of traditional ARX-MFPCC by improving it from three aspects: model structure, parameter identification, and error compensation. First, known coupling terms are explicitly removed, and electric angular velocity is introduced into the ARX model as an extended input. Then, the SCAD-QRRLS algorithm is used to identify model parameters online to suppress redundant parameters and improve numerical stability under finite word length conditions. Finally, a second-order feedforward error compensation stage is designed to smooth the residual prediction error of ARX, thereby further improving the steady-state prediction accuracy.
[0086] refer to Figure 1 This invention provides a model-free predictive current control method based on an improved ARX model, comprising the following steps:
[0087] Step S1: Collect the dq-axis stator current of the permanent magnet synchronous motor at the current time k. , Stator voltage , and electric angular velocity ;
[0088] Step S2: Construct an improved ARX prediction model. The improved ARX prediction model takes the current and historical dq-axis voltages, historical dq-axis currents, and current and historical electric angular velocities as inputs, and predicts the dq-axis current at the next moment as output. Specifically, the known cross-coupling terms between the dq axes are explicitly removed from the improved ARX prediction model structure. and and electric angular velocity As an independent extended input variable;
[0089] Step S3: The orthogonal triangular recursive least squares algorithm based on smooth truncation of absolute bias regularization is used to update the parameter vector of the improved ARX prediction model online using the collected input and output data;
[0090] Step S4: Design a second-order feedforward error compensation circuit, calculate the current prediction error at the current moment, and filter it through two cascaded first-order low-pass filters to obtain the second-order filtering error, which is then converted into the feedforward compensation amount after gain and limiting.
[0091] Step S5: Use the feedforward compensation amount to correct and improve the predicted current of the ARX prediction model, or correct the current command value, to obtain the compensated current prediction value or the compensated current command.
[0092] Step S6: Based on the compensated current prediction value or the compensated current command, and combined with the improved ARX prediction model, the dq axis voltage command to be applied at the next moment is solved in reverse and used to control the permanent magnet synchronous motor.
[0093] This invention introduces electric angular velocity As an extended input, the ARX model can directly learn and characterize the impact of speed changes on current dynamics. This allows for accurate current prediction without significant parameter adjustments during speed changes or load abrupt changes, significantly shortening parameter reconvergence time in dynamic processes and suppressing current overshoot and oscillations. By removing known cross-coupling terms, the ARX model no longer needs to fit these strongly nonlinear coupling relationships through complex parameter combinations. This not only simplifies the model structure and reduces the required model order, but more importantly, it concentrates identification resources on compensating for truly uncertain residuals, thereby improving the efficiency and accuracy of parameter identification. After these modifications, the new model exhibits more consistent dynamic characteristic representation under different speed and load conditions, avoiding the fundamental differences in traditional models at different operating points, laying the foundation for subsequent global stability control.
[0094] Further optimize the technical solution; the improved ARX prediction model constructed in step S2 has parameter vectors for its d-axis and q-axis. Extended definition: ;
[0095] in, for The autoregressive coefficients in the shaft current model that are related to past output current. In the formula, Input for electric angular velocity right The influence coefficient of the shaft-identified output current, where for The influence coefficient of the shaft back electromotive force term on current dynamics. Used as The velocity-related effects caused by axis parameter uncertainties or residual coupling are diluted; and It is the order of the polynomial;
[0096] Extended regression vectors along the d-axis and q-axis and for: ;
[0097] in, for Stator voltage in a synchronous rotating coordinate system; for Stator current in a synchronous rotating coordinate system For discrete time intervals.
[0098] Further optimize the technical solution, based on the improved ARX model, targeting... shaft and The axes each have known coupling terms. and To obtain the predicted current value at the next moment. and : ;
[0099] At higher sampling frequencies, the rotational speed is approximately considered constant within one sampling period, and the output voltage is calculated using the following formula: ;
[0100] in, The corresponding d-axis and q-axis are respectively at Current command value at any time This is the current control cycle.
[0101] Further optimize the technical solution by using the orthogonal triangular recursive least squares algorithm with smooth truncation and absolute bias regularization in step S3, whose cost function... Represented as: ;in, Forgetting factor, Let φ(t) be the system output at time t, and φ(t) be the regression vector at time t. The j-th parameter in the parameter vector, The length of the parameter vector. For SCAD regularization functions, Let the vector be the parameters to be determined.
[0102] SCAD regular function curve as shown Figure 2 As shown in the following formula: ;
[0103] Its derivative is:
[0104] ;
[0105] in, These are the curve shape parameters; This is the regularization parameter.
[0106] Curve shape parameters satisfy The value is typically set to 3.7. As shown by the SCAD regularization function curve, when the parameter magnitude is small, the penalty function exhibits approximately linear growth near the zero point with a large slope, indicating that the regularization term exerts a strong contraction effect on small parameters. Therefore, it can compress redundant parameters, noise parameters, and spurious non-zero terms introduced by higher-order modeling to zero. As the parameter magnitude further increases, the slope of the SCAD curve gradually decreases, indicating that the penalty intensity gradually weakens. When the parameter magnitude exceeds a certain threshold, the penalty term tends to flatten, indicating that almost no further contraction is applied to large parameters. Thus, SCAD regularization can achieve parameter sparsity while avoiding continuous bias to the truly effective parameters. For higher-order ARX models, this characteristic helps eliminate redundant lag terms and approximate zero-pole cancellation, reduces the model's sensitivity to noise, improves parameter convergence speed and identification stability, and further improves the control performance of current prediction and voltage inverse solution.
[0107] To further optimize the technical solution, in step S3, an orthogonal triangular decomposition is achieved using a recursive update method based on Givens rotation, so as to recursively save and update the upper triangular regularized matrix. And regularized orthogonal trigonometric decomposition of the right-hand vector This avoids the direct inversion and recursion of the covariance matrix; parameter estimation is achieved by solving the upper triangular equations. get.
[0108] Traditional Regression-Solution Modeling (RLS) relies on covariance matrix recursion, which can lead to numerical ill-conditioning, matrix asymmetry, and parameter jitter under higher-order models, strongly correlated inputs, or finite word lengths. QR decomposition, based on orthogonal transformations to achieve least-squares recursion, eliminates the need for explicit inversion, improves numerical stability under finite word lengths, and is suitable for hardware and real-time control implementations. Specifically, QR decomposition is the process of decomposing a matrix into orthogonal matrices and upper triangular matrices. For regression matrices... ,have: ,in, The regression matrix is obtained by stacking the regression vectors from multiple sampling times. It is an orthogonal matrix. It is an upper triangular matrix.
[0109] Since orthogonal transformations do not amplify vector norms, QR decomposition exhibits better numerical stability compared to least-squares solutions based on normal equations. QR decomposition can be implemented using methods such as Gram-Schmidt orthogonalization, Householder reflection, and Givenns rotation. Among these, the Gram-Schmidt method has relatively weak numerical stability, Householder reflection is more suitable for batch decomposition of complete matrices, and Givenns rotation can eliminate specified elements one by one, making it suitable for incorporating new data rows into an existing upper triangular matrix. Therefore, this invention employs Givenns rotation to implement single-row recursive updates of QR-RLS, avoiding the need to re-perform the complete QR decomposition in each sampling period and keeping the complexity within acceptable limits. This implementation also facilitates the subsequent embedding of SCAD regularization terms as additional regularized rows into the QR-RLS recursive process.
[0110] The classical RLS normal equation obtained from the least squares optimality condition is: ;
[0111] in: In the formula, For RLS information matrix, is the right-hand vector of the normal equation.
[0112] Solving directly from the classical RLS normal equations : ;
[0113] Directly solving the normal equations requires matrix inversion or solving general linear equations, resulting in poor numerical stability. Classical RLS can avoid direct inversion using the Woodbury matrix identity; however, when the cost function contains complex regularization terms such as SCAD regularization terms, the inversion process cannot be directly simplified by the Woodbury matrix identity. Therefore, it is necessary to adopt the QR-RLS algorithm, which utilizes orthogonal transformations to achieve recursive parameter identification, avoiding explicit inversion and increasing numerical stability.
[0114] QR-RLS does not directly recursively deduce Instead, it stores the associated upper triangular matrix. And the right-hand vector of QR decomposition ,satisfy: ;
[0115] In the formula, and Let be the upper triangular matrix and the right-hand side vector of the QR decomposition, respectively. Substituting the above equation into the classical RLS normal equation, the parameter estimate can be given by the equivalent upper triangular equation: ;
[0116] When the Each sampled data Upon arrival, considering the influence of the forgetting factor, construct the augmented matrix: In the formula, To complete the upper triangular matrix before updating the current sampled data, The right-hand vector of the QR decomposition before the update.
[0117] The updated result is obtained through QR decomposition and retriangulation. and : In the formula, It is a zero row vector; This is a scalar term that is independent of the parameter solution.
[0118] Since the orthogonal transformations in QR decomposition do not change the solution to the least squares problem, the above equation is equivalent to a recursive update of the classical RLS normal equation. Finally, the parameter vector is obtained by substituting the upper triangular equation back into the upper triangular equation. Because Since it is an upper triangular matrix, the computational complexity of back-substitution is O(n log n). It is much smaller than the complexity of matrix inversion or QR decomposition. .
[0119] The SCAD regularization recursive normal equation obtained from the cost function shown in the SCAD regularization function formula is: ;in, In the formula, The regularized information matrix after introducing the SCAD regularization term; Let be the right-hand vector of the regularized normal equation.
[0120] The SCAD matrix is defined as follows: ;
[0121] Among them, the first Weights in each parameter direction for:
[0122] To avoid direct solution In matrix inversion, QR-RLS does not directly use recursion. and Instead, it recursively stores the equivalent upper triangular regularized matrix. And regularized QR decomposition right-hand side vector To satisfy: ;
[0123] In the formula, It is an upper triangular regularized matrix. This is the right-hand vector of the regularized QR decomposition. (Subscript) This indicates that the variable corresponds to the recursive process after SCAD regularization. Therefore, the SCAD regularized recursive normal equation can be equivalently transformed into an upper triangular equation: ;
[0124] Depend on It can be seen that, The update includes two main new features: and .in, It is a rank-1 update generated by a single regression vector, corresponding to one QR update of the data row, with a computational complexity of O(n log n). .and Typically includes If all non-zero regularization directions are added at once, the QR update complexity will reach [value missing]. Therefore, SCAD-QRRLS employs a two-step QR update implementation. First, a data row QR update is performed on the newly sampled data:
[0125] ;
[0126] in, This is the middle upper triangular regularized matrix. The right-hand vector of the intermediate regularized QR decomposition. For the k-th sampled data, It is a zero row vector. For scalar terms that are independent of the parameter solution;
[0127] Then, row-by-row regularization is used for updating. Let the current regularization direction be the i-th row. One parameter direction, For the first Given unit vectors, construct SCAD regularized rows: ;
[0128] And perform a second QR update:
[0129] in, For the first Weights in each parameter direction;
[0130] By order exist The parameters change cyclically, and the regularization update of all parameter directions is gradually completed within multiple sampling periods. Finally, the parameters are determined by the formula. Back-substitution yields the result. Therefore, SCAD-QRRLS reduces the complexity of a fully regularized QR update from... Reduce to While maintaining the numerical stability of QR-RLS and the sparse selection capability of SCAD. Figure 3 The flowchart of the SCAD-QRRLS algorithm implementation is shown.
[0131] Compared to traditional L1 or L2 regularization, the core advantage of SCAD regularization lies in its sparsity and unbiasedness. For redundant lag coefficients that should be zero in high-order ARX models, or spurious non-zero parameters caused by noise, SCAD applies a strong contraction effect to compress them to zero, thus automatically achieving model order reduction and feature selection. For truly important non-zero parameters, the penalty of SCAD tends to flatten as the parameter magnitude increases, thereby avoiding the biased estimation problem caused by L2 regularization and ensuring the estimation accuracy of the main loop parameters. Replacing traditional RLS with QR-RLS is key to solving numerical ill-conditioning under finite word length conditions. Traditional RLS recursively calculates the covariance matrix, which easily becomes ill-conditioned or even singular when the input signal is insufficiently excited or has strong correlation, leading to parameter estimation divergence. QR decomposition avoids matrix inversion through orthogonal transformation, and the Givens rotation recursive update method itself has good numerical robustness, effectively suppressing the accumulation of rounding errors. This is crucial for online identification of high-order ARX models implementing fixed-point arithmetic on embedded controllers.
[0132] Further optimization of the technical solution is needed. Although decoupled extended ARX models and SCAD-QRRLS can improve parameter identification accuracy, model prediction errors are still difficult to completely avoid in actual motor control systems. On the one hand, sampling noise, dead-zone effects, inverter non-ideal factors, and incompletely modeled high-frequency disturbances introduce residual prediction errors. On the other hand, during steady-state operation, insufficient input excitation may cause slow drift in online identified parameters. These effects are reflected in the prediction error of current during the control process.
[0133] In addition, to reduce the total harmonic distortion (THD) of steady-state current, a forgetting factor is used. Typically, the value is set to 1. At this setting, historical data carries a larger weight, slowing down parameter updates. While this helps suppress steady-state parameter fluctuations, it also makes it difficult for the model parameters to adapt promptly to changes in speed, load, or reference current. This results in persistent prediction errors during dynamic processes, ultimately weakening current tracking performance. Therefore, relying solely on RLS online identification cannot simultaneously achieve both steady-state smoothness and dynamic compensation capabilities. It is necessary to introduce a prediction error feedforward compensation stage to correct residual model errors.
[0134] To balance steady-state smoothness and dynamic compensation capability, this invention designs a second-order feedforward error compensation stage.
[0135] First, define the current prediction error as: ;
[0136] By cascading two identical first-order low-pass filters, the second-order filtering error is obtained: In the formula, These are the filter coefficients.
[0137] The discrete transfer function of a second-order filter is: ;in, These are the filter coefficients, and 0 < <1;
[0138] The filtered prediction error is used to construct the feedforward compensation: ;in, To compensate for the gain; To compensate for the bandwidth limitation; This is the amplitude limiting function.
[0139] To further optimize the technical solution, in step S5, the compensated predicted current... for: ;
[0140] Simultaneously, the reference value is compensated accordingly in the voltage inverse solution to offset the influence of model residual error on the control voltage; the compensated current command is defined. for: .
[0141] The proposed improved strategy, through coupling term processing and electric angular velocity (EAC) extension input, enables the model to maintain good prediction consistency under conditions of speed variation and load disturbance. Simultaneously, SCAD-QRRLS online identification can suppress redundant parameters and spurious non-zero terms in high-order ARX models, reducing parameter fluctuations and noise sensitivity, thereby improving the stability of the prediction model. Furthermore, second-order feedforward error compensation can smoothly correct residual prediction errors, reducing the impact of model mismatch on control voltage. In summary, the proposed strategy can simultaneously improve current dynamic response, steady-state tracking accuracy, and current waveform quality without relying on precise motor parameters. The flow of the model-free predictive current control strategy based on the improved ARX model is as follows: Figure 4 As shown.
[0142] Traditional feedforward compensation directly uses instantaneous errors, which easily introduces high-frequency noise, leading to system oscillations. This scheme uses a second-order low-pass filter to smooth the prediction error. Compared to first-order filtering, second-order filtering has a steeper amplitude-frequency characteristic, which can more effectively filter out high-frequency disturbances while preserving the low-frequency trend of the error. This ensures that the feedforward compensation is smooth and stable, thus effectively eliminating steady-state current error and reducing total harmonic distortion without introducing additional noise interference. During steady-state operation, to obtain better steady-state accuracy, the forgetting factor of RLS is usually set to 1, which slows down parameter updates. When the operating conditions change abruptly, although Schemes 1 and 2 can respond quickly, there are still small residual errors. The second-order feedforward compensation loop can directly sense changes in the prediction error and quickly generate a compensation amount to act on the control output, thereby compensating for the inertia of parameter identification response and significantly reducing current tracking error and overshoot in the dynamic process. As an independent closed loop, this compensation loop can effectively cope with disturbances that are fully compensated by any scheme. This further enhances the robustness of the entire control system.
[0143] This invention ensures that the calculated voltage command can more accurately offset the dynamic coupling and nonlinear factors in the actual motor by explicitly removing known coupling terms and introducing feedforward compensation in the inverse equation, thereby driving the actual current to closely follow the command value and achieving a full-link performance improvement from model, parameters to output.
[0144] Experimental example:
[0145] To verify the effectiveness of the proposed method, a Matlab / Simulink simulation was performed using a two-level inverter-based SPMSM. The simulation conditions were as follows: sampling frequency set to 10kHz; motor speed at 0s: 400 r / min, load 0.5 N·m; load at 1s: 1.9 N·m; speed at 0.5s: 300 r / min; speed at 1.5s: 800 r / min; load at 2s: 3 N·m. The physical parameters of the motor are shown in Table 1.
[0146] symbol parameter numerical values <![CDATA[P N ]]> Rated power 1kW <![CDATA[U N ]]> bus voltage 150V <![CDATA[n p ]]> Extreme logarithm 4 <![CDATA[R s ]]> Stator resistance 0.55Ω <![CDATA[L d ]]> Stator inductor 1.65mH <![CDATA[L q ]]> Stator inductor 1.65mH <![CDATA[Ψ f ]]> Permanent magnet flux 0.13Wb
[0147] Table 1
[0148] The integral of time and absolute error (ITAE), which reflects the overall performance throughout the dynamic process, is defined as follows: ;
[0149] in It represents the error between the specified variable and the target value. Because SPMSM uses... Control, jumps, and errors are usually not large, therefore... The changes in various physical quantities of the shaft are used as the evaluation criteria for dynamic performance.
[0150] At 1.5s, the speed command jumps from 300r / min to 800r / min. Figure 5 and Figure 6 The simulation waveforms of the variable speed dynamic response of ARX-MFPCC and EARX-MFPCC are shown respectively. Figure 7 The results of comparing the performance evaluation indicators of the two control strategies under variable speed conditions are presented.
[0151] Depend on Figure 7 It can be seen that ARX-MFPCC's The shaft current ITAe is 0.03, and the current overshoot reaches 1.20A, indicating that it still has certain transient prediction errors and parameter re-convergence processes during sudden changes in operating conditions. In contrast, the EARX-MFPCC... The shaft current ITAE was reduced to 0.021, and the current overshoot was only 0.20A. This indicates that the proposed EARX-MFPCC can effectively reduce dynamic tracking error and suppress current overshoot, and has better adaptability to changing operating conditions and dynamic control performance.
[0152] To verify the actual operational effectiveness of the control strategy proposed in this invention, an experimental platform for a permanent magnet synchronous motor based on the RTU-BOX205 real-time digital controller was built. The experimental platform mainly consists of a host computer, an RTU-BOX205 controller, a three-phase voltage source inverter, a permanent magnet synchronous motor, an encoder, a current sampling circuit, a DC power supply, and a load device. The host computer is used for control algorithm modeling, program downloading, parameter debugging, and experimental data recording. The main processor of the RTU-BOX205 uses a 32-bit TMS320C28346 processor from TI, which performs current sampling, rotor position and speed acquisition, control algorithm calculation, and PWM signal output. The inverter drives the permanent magnet synchronous motor according to the PWM signal output by the controller.
[0153] To verify the dynamic tracking capability of different control strategies under speed change conditions, the load torque was set to 1.9 N·m, and the given speed was stepped from 300 r / min to 800 r / min. Figures 8 to 10 The waveforms of current and speed changes corresponding to different algorithms under the condition of sudden speed change are shown. Figure 11 Comparison of different algorithms under variable speed conditions Axis current ITAE and current overshoot.
[0154] Depend on Figures 8 to 11 It is known that under variable speed conditions, DPCC can achieve rapid tracking of command current with virtually no overshoot during dynamic operation. However, since DPCC relies on fixed motor parameters, parameter mismatch can lead to discrepancies between the prediction model and the actual situation when the motor's operating conditions change, resulting in unavoidable prediction errors. The shaft current ITAe reaches 0.0365. In contrast, the ARX-MFPCC identifies model parameters online via RLS, enabling it to progressively correct model biases as data is updated, thus its... The shaft current ITAe decreased to 0.0348; however, in the initial stage of sudden speed changes, the current overshoot reached as high as 1.21A during dynamic processes. Furthermore, because the model parameters still needed to reconverge, the ARX-MFPCC still exhibited significant transient prediction bias under varying operating conditions. The EARX-MFPCC further improved the dynamic current response under varying operating conditions, reducing the current overshoot to 0.85A. The shaft current ITAe was further reduced to 0.0238. These results demonstrate that the proposed EARX-MFPCC can effectively reduce current overshoot and dynamic tracking error during sudden speed changes, verifying the effectiveness of the proposed algorithm under variable speed conditions.
[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A model-free predictive current control method based on an improved ARX model, characterized in that, Includes the following steps: Step S1: Collect the dq-axis stator current of the permanent magnet synchronous motor at the current time k. , Stator voltage , and electric angular velocity ; Step S2: Construct an improved ARX prediction model. The improved ARX prediction model takes the current and historical dq-axis voltages, historical dq-axis currents, and current and historical electric angular velocities as inputs, and predicts the dq-axis current at the next moment as output. Specifically, the known cross-coupling terms between the dq axes are explicitly removed from the improved ARX prediction model structure. and and electric angular velocity As an independent extended input variable; Step S3: The orthogonal triangular recursive least squares algorithm based on smooth truncation of absolute bias regularization is used to update the parameter vector of the improved ARX prediction model online using the collected input and output data; Step S4: Design a second-order feedforward error compensation circuit, calculate the current prediction error at the current moment, and filter it through two cascaded first-order low-pass filters to obtain the second-order filtering error, which is then converted into the feedforward compensation amount after gain and limiting. Step S5: Use the feedforward compensation amount to correct and improve the predicted current of the ARX prediction model, or correct the current command value, to obtain the compensated current prediction value or the compensated current command. Step S6: Based on the compensated current prediction value or the compensated current command, and combined with the improved ARX prediction model, the dq axis voltage command to be applied at the next moment is solved in reverse and used to control the permanent magnet synchronous motor.
2. The model-free predictive current control method based on an improved ARX model according to claim 1, characterized in that, The improved ARX prediction model constructed in step S2 has parameter vectors for its d-axis and q-axis. Extended definition: ; in, for The autoregressive coefficients in the shaft current model that are related to past output current. In the formula, Input for electric angular velocity right The influence coefficient of the shaft-identified output current, where for The influence coefficient of the shaft back electromotive force term on current dynamics. Used as The velocity-related effects caused by axis parameter uncertainties or residual coupling are diluted; and It is the order of the polynomial; Extended regression vectors along the d-axis and q-axis and for: ; in, for Stator voltage in a synchronous rotating coordinate system; for Stator current in a synchronous rotating coordinate system For discrete time intervals.
3. The model-free predictive current control method based on an improved ARX model according to claim 2, characterized in that, Based on the improved ARX model, targeting... shaft and The axes each have known coupling terms. and To obtain the predicted current value at the next moment. and : ; At higher sampling frequencies, the rotational speed is approximately considered constant within one sampling period, and the output voltage is calculated using the following formula: ; in, The corresponding d-axis and q-axis are respectively at Current command value at any time This is the current control cycle.
4. The model-free predictive current control method based on an improved ARX model according to claim 3, characterized in that, The cost function of the orthogonal triangular recursive least squares algorithm with smooth truncation and absolute bias regularization in step S3 is... Represented as: ;in, Forgetting factor, Let φ(t) be the system output at time t, and φ(t) be the regression vector at time t. The j-th parameter in the parameter vector, The length of the parameter vector. For SCAD regularization functions, Let the vector be the parameters to be determined. The SCAD regular function is shown in the following formula: ; Its derivative is: ; in, These are the curve shape parameters; This is the regularization parameter.
5. The model-free predictive current control method based on an improved ARX model according to claim 4, characterized in that, In step S3, an orthogonal triangular decomposition is achieved using a recursive update method based on Givens rotation, so as to recursively save and update the upper triangular regularization matrix. And regularized orthogonal trigonometric decomposition of the right-hand vector This avoids the direct inversion and recursion of the covariance matrix; parameter estimation is achieved by solving the upper triangular equations. get.
6. The model-free predictive current control method based on an improved ARX model according to claim 5, characterized in that, When performing recursive updates via Givens rotations, a data row QR update is performed on the sampled data at time k: ; in, This is the middle upper triangular regularized matrix. The right-hand vector of the intermediate regularized QR decomposition. For the k-th sampled data, It is a zero row vector. For scalar terms that are independent of the parameter solution; Then, row-by-row regularization is used for updating. Let the current regularization direction be the i-th row. One parameter direction, For the first Given unit vectors, construct SCAD regularized rows: ; And perform a second QR update: ; in, For the first Weights in each parameter direction; By order exist The parameters change cyclically, and the regularization update of all parameter directions is gradually completed within multiple sampling periods. Finally, the parameters are determined by the formula. It was obtained by going back to the source.
7. The model-free predictive current control method based on an improved ARX model according to claim 6, characterized in that, The discrete transfer function of the second-order feedforward error compensation stage in step S4 is: ;in, These are the filter coefficients, and 0 < <1; The filtered prediction error is used to construct the feedforward compensation: ;in, To compensate for the gain; To compensate for the bandwidth limitation; This is the amplitude limiting function.
8. The model-free predictive current control method based on an improved ARX model according to claim 7, characterized in that, In step S5, the compensated predicted current for: ; Simultaneously, the reference value is compensated accordingly in the voltage inverse solution to offset the influence of model residual error on the control voltage; the compensated current command is defined. for: .
