Active front-end rectifier and predictive control method thereof

Abstract

The invention discloses an active front-end rectifier and a prediction control method thereof, and the method comprises the steps: calculating the active power and reactive power of a power grid according to a voltage model of a three-level active front-end rectifier and an instantaneous power theory; according to the voltage model, combining the active power and the reactive power to calculate a power gradient; converting the power gradient into a second-order state-space equation, predicting the control input of the active front-end rectifier by using a super-spiral observer, and obtaining the input voltage of the active front-end rectifier according to the predicted control input and the grid-side voltage; and based on the input voltage, obtaining an optimal voltage vector by adopting a finite set model predictive control framework, and controlling the active front-end rectifier based on the switching state corresponding to the optimal voltage vector. According to the method, power gradient modeling, superhelix observer disturbance compensation and finite set prediction control are combined, so that the rectifier overcomes the influence of parameter mismatch, and high-sine-degree current, stable direct-current voltage and power tracking control are realized.

Description

Technical Field

[0001] This application relates to the field of predictive control of rectifiers, and more particularly to an active front-end rectifier and its predictive control method. Background Technology

[0002] With the development of power electronics technology, active front-end rectifiers are increasingly widely used in industrial drives, new energy power generation, and other fields. Finite set model predictive control based on instantaneous power theory has become one of the mainstream solutions due to its superior dynamic performance. However, this solution faces the following key technical challenges in practical applications: control performance is highly dependent on the accurate mathematical model of the system; model mismatch occurs when parameters such as filter inductors drift; linear observation structures such as extended state observers in related technologies cannot accurately describe the nonlinear dynamic characteristics of the rectifier; and traditional solutions do not fully consider the voltage saturation constraints of power switching devices. These technical challenges lead to increased current harmonics and decreased power point tracking accuracy under parameter variations and extreme operating conditions, hindering the widespread application of active front-end rectifiers in complex industrial environments. Therefore, it is urgent to develop a high-performance control method that can effectively overcome parameter sensitivity and saturation nonlinearity problems. Summary of the Invention

[0003] This application provides an active front-end rectifier and its predictive control method, which can realize adaptive predictive control of the active front-end rectifier without relying on precise model parameters.

[0004] In a first aspect, embodiments of this application provide a predictive control method for an active front-end rectifier, the method comprising: Based on the voltage model of the three-level active front-end rectifier, the active power and reactive power of the power grid are calculated according to the instantaneous power theory. Based on the voltage model, and combining active and reactive power, the power gradient is calculated. The power gradient is converted into a second-order state-space equation. The control input of the active front-end rectifier is predicted using a super-spiral observer. The input voltage of the active front-end rectifier is obtained based on the predicted control input and the grid-side voltage. Based on the input voltage, a finite set model predictive control framework is used to obtain the optimal voltage vector, and the active front-end rectifier is controlled based on the switching state corresponding to the optimal voltage vector.

[0005] In one embodiment, the predictive control method further includes adaptively updating the input gain of the superspiral observer using a fast gradient method, including: When there is an input gain mismatch in the superspiral observer, the second-order state-space equation can be expressed as: ; In the formula, , , , , , P g Indicates active power. Q g Indicates reactive power. Indicates the system's filter resistance. Indicates the filter inductance. , Indicates grid-side voltage Axial components, Indicates grid-side voltage Axial components, Indicates the angular frequency on the power grid side. Indicates input gain The estimated value; Calculate state variables X The normal unit vector of 1, the transpose of the normal unit vector and the expansion state X 2. Perform the inner product and construct the error equation for the fast gradient algorithm based on the inner product result; Based on the error equation, the input gain of the superspiral observer is estimated, and the input gain of the superspiral observer is updated based on the estimated input gain.

[0006] In one embodiment, constructing the error equation for the fast gradient algorithm based on the inner product result includes: Remove some known quantities from the inner product result using the following formula: ; In the formula, Represents state variables X The normal unit vector of 1, Indicates the angular frequency on the power grid side. The error equation is configured to calculate the square of the difference between zero and the inner product after removing some known quantities.

[0007] In one embodiment, the input gain of the superspiral observer is estimated based on the error equation, including the following steps: Based on the error equation, the input gain estimate is obtained. The gradient, based on the input gain estimate. The gradient is used to adaptively update the input gain estimate using a discrete-form fast gradient method. .

[0008] In one embodiment, an actuator is used to trim the control input of the active front-end rectifier output by the superspiral observer, and the input voltage of the active front-end rectifier is obtained based on the trimmed control input and the grid-side voltage. Specifically, a neural network is used to fit the execution deviation of the actuator to obtain an estimated value of the execution deviation. The estimated value of the execution deviation is then subtracted from the control input of the active front-end rectifier output by the superspiral observer to obtain the trimmed control input.

[0009] In one embodiment, the neural network is a two-layer neural network, which is represented by the following formula: ; In the formula, V T This represents the fixed and known weights of the first layer of the neural network. This represents the weights of the second layer of the neural network. This represents a known activation function. X NN =[ R*X 1] T , , P g * A reference value representing active power. Q g * This represents a reference value for reactive power.

[0010] In one embodiment, the predictive control method further includes: Set the update rate for the second layer weights of the neural network. Based on the update rate, update the second layer weights of the neural network using the first-order Euler discretization method. Calculate the execution bias estimate based on the updated neural network.

[0011] In one embodiment, the update rate of the second layer weights of the neural network is configured as follows: ; In the formula, , is a constant matrix. For positive integers, E This represents the tracking error of the system. b This indicates the input gain.

[0012] In one embodiment, predicting the control input of the active front-end rectifier using a superhelical observer includes: A superspiral observer is used to establish the nonlinear sliding membrane function for the state variables in the second-order state-space equation. X 1 and expansion state X 2. Make an estimate to obtain the first observation state and the second observation state; Discretize the first and second observation states, and tune the superspiral observer so that the second observation state approximates the extended state. X 2, System status The control input of the superspiral observer is obtained by subtracting the estimate of the extended state from the derivative of the reference value, then subtracting the system tracking error modulated by the error coefficient, adding the robust control term of the system, and finally dividing by the input gain.

[0013] Secondly, this application also provides an active front-end rectifier configured to apply a predictive control method for an active front-end rectifier as described in the first aspect.

[0014] The aforementioned active front-end rectifier and its predictive control method calculate the active and reactive power of the power grid based on the rectifier voltage model and instantaneous power theory; calculate the power gradient according to the voltage model and power values; after converting the power gradient into a second-order state-space equation, use a superspiral observer to predict the control input and obtain the rectifier input voltage by combining it with the grid-side voltage; obtain the optimal voltage vector based on the input voltage using finite set model predictive control, and control the rectifier through the corresponding switching states. This method, through the organic combination of power gradient modeling, superspiral observer disturbance compensation, and finite set predictive control, enables the rectifier to overcome the influence of parameter mismatch and achieve high sinusoidal current, fast and stable DC voltage, and accurate power point tracking control. Attached Figure Description

[0015] Figure 1 This is a flowchart of a predictive control method for an active front-end rectifier in one embodiment; Figure 2 This is a schematic diagram of a three-level active front-end rectifier circuit topology in one embodiment; Figure 3 This is a flowchart illustrating the prediction of control inputs to an active front-end rectifier in one embodiment; Figure 4 This is a waveform diagram of the three-phase current on the AC side in one embodiment; Figure 5 The waveform of the DC-side output voltage is shown in one embodiment. Figure 6 The waveform diagram of grid-side voltage power in one embodiment is shown. Figure 7 This is a current harmonic spectrum in one embodiment. Detailed Implementation

[0016] The present application will be described in detail below with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present application. Any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the protection scope of the present application.

[0017] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0018] In one embodiment, such as Figure 1 As shown, a predictive control method for an active front-end rectifier is provided, which includes the following steps: Step 101: Based on the voltage model of the three-level active front-end rectifier, calculate the active power and reactive power of the power grid according to the instantaneous power theory; Specifically, the three-level active front-end rectifier circuit topology is as follows: Figure 2 As shown, Arm-1, Arm-2, and Arm-3 represent the three arms of the three-phase system; each arm contains four power switches, where S... a1 S a2 S a3 S a4 The four switching transistors representing phase A bridge arm, S b1 S b2 S b3 S b4 Represents phase B bridge arm, S c1 S c2 S c3 S c4 This represents the C-phase bridge arm; the DC side includes two equalizing capacitors C1 and C2 and a load resistor R. L i dc For DC-side output current; i cp This could refer to the current flowing into the voltage equalization capacitor C1; i cl This could refer to the current flowing into the voltage equalization capacitor C2; R g and Lg These are the filter resistor and the filter inductor, respectively. i a The AC current flowing into phase A bridge arm, i b For the AC current flowing into phase B bridge arm, i c This refers to the alternating current flowing into the C-phase bridge arm. u ra , u rb , u rc These represent the input phase voltages of the rectifier on the three-phase AC side; v p It can represent the positive DC bus voltage (i.e., the upper capacitor). C 1. Voltage across both ends); v l Capacitance can be represented C 2. Voltage at both ends. This topology, through precise control of the on / off states of 12 switching transistors, achieves stable control of the DC bus voltage while effectively managing circulating current and ensuring power quality on the AC side.

[0019] The voltage model of a three-level active front-end rectifier in the stationary αβ coordinate system is as follows: ; in, e gα、 e gβ These represent the grid-side voltage at... Components of the axis, u rα、 u rβ These represent the rectifier input voltage at... Components of the axis, i gα、 i gβ Indicates the current in the line Components of the axis, R g and L g These are the filter resistor and filter inductor, respectively. This voltage model can represent the grid-side voltage. e gα、 e gβ Line current i gα、 i gβ With rectifier input voltage u rα、 u rβThe dynamic relationship between them. In order to directly control the power, a power calculation model based on instantaneous power theory is introduced. Its core idea is to use the instantaneous values ​​of voltage and current in the αβ coordinate system to perform algebraic operations, thereby avoiding complex coordinate rotation and filtering links, and realizing fast reactive power control.

[0020] Active power on the grid side P g and reactive power Q g The calculation formula is as follows: ; It should be noted that active power P g It can represent the actual rate at which electrical energy is transferred from the grid to the DC side; reactive power can reflect the energy that is constantly exchanged but not consumed between the grid and the rectifier. This power calculation method based on instantaneous values ​​has a fast response speed, requires no filtering, provides accurate real-time feedback signals for subsequent predictive control algorithms, and is naturally suitable for digital controller implementation.

[0021] Step 102: Calculate the power gradient based on the voltage model, combined with active and reactive power; Specifically, regarding active power P g Differentiating reactive power Q g The formula for calculating the power gradient can be obtained as follows: ; Among them, active power power gradient Rectifier input voltage u rα and input voltage Components of the axis u rβ Filter resistor R g Grid-side voltage at Square of the components of the axis and grid angular frequency ω With reactive power The coupling effect has a combined influence; reactive power power gradient Then subject to the rectifier input voltage u rα and input voltage Components of the axis u rβ Filter resistor R g and grid angular frequency ω With active power The coupling effect determines this.

[0022] By analyzing the rectifier's voltage model and current power state, the changes in active and reactive power in the next second under a given control command are calculated. This calculation process comprehensively considers the rectifier's voltage regulation effect, the loss effect of circuit resistance, the fundamental driving force of the grid voltage, and the mutual influence between active and reactive power.

[0023] Step 103: Convert the power gradient into a second-order state-space equation, use a super-spiral observer to predict the control input of the active front-end rectifier, and obtain the input voltage of the active front-end rectifier based on the predicted control input and the grid-side voltage. Specifically, the power gradient is expressed as a first-order state-space equation, with the following formula: ; in, a =- R g / L g , b =1.5 / L g , , , .

[0024] Furthermore, to address the issues of parameter uncertainty, unmodeled dynamics, and external disturbances in practical applications, an extended state is introduced to unify and integrate the uncertainties that are difficult to model precisely into an extended state X2. This state reconstruction method shifts the focus of the control problem from pursuing precise modeling to achieving accurate observation. By defining disturbances as new state variables, it creates the necessary conditions for subsequent feedforward compensation control based on disturbance estimation. Therefore, it is necessary to extend the first-order state-space equation to a second-order state-space equation, the formula of which is: ; Let V represent the rate of change of the lumped disturbance and the extended state X2, then the second-order state-space equation is further as follows: ; ; .

[0025] Second-order state-space equations are typically observed by linear observers such as extended state observers and universal proportional-input-output observers. However, power converters are mathematically classified as nonlinear systems, and using linear observers can lead to inaccurate modeling and a lack of stability. Superspiral observers address this problem by converting discontinuous sliding mode variables that cause jitter into higher-order derivatives. Therefore, this application employs a superspiral observer to establish a nonlinear sliding mode function, overcoming the inaccurate estimation of nonlinear systems by traditional linear observers, and achieving accurate estimation of the state and disturbances through the characteristics of nonlinear sliding mode.

[0026] Among them, the superspiral observer is a high-order sliding mode observer that, through its unique nonlinear structure, can estimate the system state quickly, smoothly, and without overshoot. The superspiral observer can estimate the state variables in the second-order state-space equations. X 1 and expansion state X 2. Estimate the first and second observation states. Further, discretize the first and second observation states to facilitate digital implementation, forming an algorithm that can iteratively calculate within each sampling period, ensuring good estimation performance under various operating conditions. The system state is then... The control input of the superspiral observer is obtained by subtracting the estimate of the extended state from the derivative of the reference value, then subtracting the system tracking error modulated by the error coefficient, adding the robust control term of the system, and finally dividing by the input gain.

[0027] Based on this, and according to the feedback control principle, in order to make the tracking error converge to 0, the control input is designed in the following form: ; in, Representing state X The reference value is 1. This can represent tracking error. For robust control terms, k t Usually, 1 / T s , K r It is a normal number.

[0028] It should be noted that, It can be an ideal instruction calculated in the control algorithm, existing in the digital world of the controller.

[0029] because .

[0030] This definition can establish mathematical control quantities. U Compared to the input voltage of an active front-end rectifier in the physical world and The mapping relationship between them. Further, the actual voltage is solved through inverse transformation, since the control algorithm provides the ideal... ,and U And also related to the input voltage of the active front-end rectifier and There is a clear mathematical relationship; by inverting the matrix, the input voltage of the active front-end rectifier can be calculated from the ideal control quantities. .

[0031] Step 104: Based on the input voltage, the optimal voltage vector is obtained using a finite set model predictive control framework, and the active front-end rectifier is controlled based on the switching state corresponding to the optimal voltage vector.

[0032] Specifically, a finite set model predictive control framework is used to achieve direct decision-making and execution of the optimal switching vector. The core of this control strategy is to make full use of the inherent characteristic of the finite switching states of the three-level rectifier, and to form a definite finite search set of 27 possible switching combinations. In each control cycle, all candidate switching vectors are traversed, and the key controlled variables that each vector will generate at the next sampling time, such as active power and reactive power, are predicted based on the discretized mathematical model.

[0033] Furthermore, the predicted key controlled variables are fed into a pre-designed value function for evaluation. This value function assigns a score representing the control effect to each candidate vector by quantifying the tracking error (e.g., the deviation between the predicted power and the reference value) and comprehensively considering constraints such as switching frequency. By comparing all scores, the switching state that minimizes the value function is selected as the optimal voltage vector for the current cycle. Further, this optimal switching combination is applied to the power switch of the rectifier and maintained in this state throughout the entire sampling cycle until re-optimization in the next cycle.

[0034] In this embodiment, the method calculates the active and reactive power of the power grid based on the rectifier voltage model and instantaneous power theory; calculates the power gradient according to the voltage model and power values; after converting the power gradient into a second-order state-space equation, it uses a superspiral observer to predict the control input and obtains the rectifier input voltage by combining it with the grid-side voltage; based on the input voltage, it uses a finite set model predictive control to obtain the optimal voltage vector, and controls the rectifier through the corresponding switching state. This method, through the organic combination of power gradient modeling, superspiral observer disturbance compensation, and finite set predictive control, enables the rectifier to overcome the influence of parameter mismatch and achieve high sinusoidal current, fast and stable DC voltage, and accurate power point tracking control.

[0035] In one embodiment, such as Figure 3As shown, predicting the control input of an active front-end rectifier using a superhelical observer includes the following steps: Step 301: Establish the nonlinear sliding membrane function for the state variables in the second-order state-space equation using a superspiral observer. X 1 and expansion state X 2. Make an estimate to obtain the first observation state and the second observation state; Specifically, by establishing a nonlinear sliding mode function, the observer can monitor the state variables in the second-order state-space equations. X 1 and expansion state X 2. Perform estimation to obtain the first observed state. Z 1 and second observation states Z 2. Its continuous-time dynamics are described by the following formula: ; This method introduces a fractional power term of the state error and a sign function to form a nonlinear sliding surface, enabling the observer dynamics to converge rapidly.

[0036] Step 302: Discretize the first and second observation states, and tune the superspiral observer so that the second observation state approximates the extended state. X 2; Specifically, the continuous-time observer dynamics are discretized to obtain the discrete form shown in the following formula: ; This discrete algorithm enables the observer to perform each sampling period. T s Internally, based on current measurements X ( k ) and control input U ( k Iteratively update its state. X 1 and expansion state X 2 Observational state values Z 1 and Z 2.

[0037] By tuning the observer gain k 1 and k 2 (Usually based on the system's dynamic range and sampling period, the optimal range is determined through stability analysis or simulation), ensuring that the discretized superspiral observer not only remains stable but also rapidly forces the state estimation error to change within a finite time. X ( k )− Z 1( k Its derivative converges to the zero neighborhood. Once the superhelical observer is well tuned, its internal dynamics will drive the second observation state. Z2. Approximates the extended state with high precision (within the accuracy range of the discrete system). X 2.

[0038] Step 303: Set system status The control input of the superspiral observer is obtained by subtracting the estimate of the extended state from the derivative of the reference value, then subtracting the system tracking error modulated by the error coefficient, adding the robust control term of the system, and finally dividing by the input gain.

[0039] Specifically, the system state X 1 Reference value R ∗ ( k The derivative of +1) minus the expansion state provided in real time by the superspiral observer X Estimator of 2 Z 2( k +1), its purpose is to actively offset all identified uncertainties and disturbances in the system through a feedforward approach.

[0040] Further subtract the instantaneous tracking error of the system after the error coefficient modulation. This feedback ensures that the system output accurately tracks the reference trajectory; coupled with the system's robust control terms... This term, as a nonlinear correction based on the sliding mode principle, is used to suppress unmodeled dynamics and residual disturbances; the algebraic sum of all the above terms is divided by the estimated input gain of the system. (Updated online using the fast gradient method), the control input is normalized to obtain the final output control input of the superspiral observer. This control input integrates feedforward disturbance compensation, feedback error adjustment, and nonlinear robust correction, forming the core instruction for achieving high-performance control of the active front-end rectifier.

[0041] In one embodiment, the predictive control method further includes adaptively updating the input gain of the superspiral observer using a fast gradient method, including: When there is an input gain mismatch in the superspiral observer, the second-order state-space equation can be expressed as: ; In the formula, , , , , , P g Indicates active power. Q g Indicates reactive power. Indicates the system's filter resistance. Indicates the filter inductance. , Indicates grid-side voltage Axial components, Indicates grid-side voltage Axial components, Indicates the angular frequency on the power grid side. Indicates input gain The estimated value; Calculate state variables X The normal unit vector of 1, the transpose of the normal unit vector and the expansion state X 2. Perform the inner product and construct the error equation for the fast gradient algorithm based on the inner product result; Based on the error equation, the input gain of the superspiral observer is estimated, and the input gain of the superspiral observer is updated based on the estimated input gain.

[0042] Specifically, when the preset input gain estimate inside the controller Compared with actual value b When mismatch exists, the second-order state-space equations characterizing the system power dynamics and lumped disturbances are reformulated as follows: This formula can represent the parameter mismatch term ( b - ) U This is a significant source of observation and control errors. To achieve... The algorithm calculates the state variables for self-correction. X 1 (i.e., power vector) The normal unit vector of ) The normal unit vector is perpendicular to the current power operating point.

[0043] Furthermore, the normal unit vector S Transpose and expansion states X 2. Perform inner product operations S T X 2, can be obtained ,in, , can represent the control input vector U In the state normal unit vector S Projection components in the direction. The error equation for the fast gradient algorithm is constructed based on the inner product result: F s = .

[0044] Through the error equation The core purpose of processing the original inner product result is to... Mixed signals (including known nonlinear couplings) and some known input gains Calculated partial effects Subtracting known or estimated quantities from the input gain mismatch yields a pure result. b - The error components contributed by ) .

[0045] Based on this result, the error equation is configured to calculate the inner product of zero and the result after removing some known quantities. The square of the difference, i.e. F ( b )=(0− F s ) 2 =[( - b ()( )] 2 The error function F ( b When it reaches its minimum value (zero), it must satisfy the following condition. =0, under continuous incentives (i.e. Under the condition that =0), it means that = b The minimum point of this function corresponds to the correct... b value.

[0046] Based on the error equation Input gain of the superspiral observer Online estimation and adaptive updating are performed. Based on the error equation, the estimated value of the input gain is obtained by differentiation. gradient∇ F ( The gradient passes through the error function. F ( For the estimated value Differentiating, we get: .in, , is an effective excitation signal, in which To control the input vector U In the state normal unit vector S The projection component along the direction. This gradient ∇ F ( This can be expressed as adjusting to reduce estimation error. The direction and magnitude.

[0047] Due to the actual parameters b The unknown needs to be utilized in actual calculations, specifically the output of the observer. F s As a feedback signal, it is used to replace the gradient in practice. Furthermore, the fast gradient method is used to input the gain estimate. It can be estimated as follows: ; in, θ For adaptive step size, λ and ρ This is an adjustable coefficient.

[0048] Furthermore, by calculating the gradient ∇ of this error function... F ( And using the discrete form of the fast gradient method is: ; in, As an adaptive step size factor, it itself is based on the gradient norm. Dynamic adjustment (by coefficient) Driven increase, by (Controlled decay), continuously fine-tuning within the normalized gradient direction. This allows it to converge quickly to the true input gain. b This ensures that the superspiral observer is even when the system filter inductor is in operation. L g Even when the situation is unknown or changes, it can still maintain accurate disturbance estimation capabilities.

[0049] In one embodiment, the predictive control method further includes: An actuator is used to trim the control input of the active front-end rectifier output from the superspiral observer. Based on the trimmed control input and the grid-side voltage, the input voltage of the active front-end rectifier is obtained. Specifically, a neural network is used to fit the execution deviation of the actuator to obtain an estimated value of the execution deviation. The estimated value of the execution deviation is then subtracted from the control input of the active front-end rectifier output by the superspiral observer to obtain the trimmed control input.

[0050] Specifically, in a practical control system, the ideal control input output of the superspiral observer is... When actually executed by an actuator (such as a power switching device), a saturation effect may occur due to physical limitations, that is, when the desired input exceeds the upper limit of the actuator's output. U max or lower limit U min At that time, the actuator will trim it, resulting in the actual output. U ( k ) and expected output Γ( k ) produces an execution deviation Ψ( k ), execution deviation Ψ( k )as follows: ; in, m This is the scaling factor between the expected input and the actual input when the system is not saturated; it is usually 1.

[0051] In one embodiment, to actively compensate for this nonlinear saturation effect, this application introduces a neural network to fit the execution deviation online. The neural network is a two-layer neural network, and the neural network is represented by the following formula: ; In the formula, V T This represents the fixed and known weights of the first layer of the neural network. This represents the weights of the second layer of the neural network. This represents a known activation function. , A reference value representing active power. Reference value indicating reactive power; Set the update rate of the second layer weights of the neural network. Based on the update rate, update the second layer weights of the neural network using the first-order Euler discretization method. Calculate the execution bias estimate based on the updated neural network. The update rate of the weights in the second layer of the neural network is configured as follows: ; In the formula, , is a constant matrix. For positive integers, E This represents the tracking error of the system. b This indicates the input gain.

[0052] Specifically, the neural networks used in practice are ideal neural networks, as shown by the formula... It means that among them V T This represents a fixed and known weight matrix for the first layer of the neural network, which remains unchanged after network initialization. This represents the adjustable weight matrix of the second layer of the neural network, which is a key parameter for the network to learn online. σ (⋅) represents the selected activation function. In this application, the Sigmoid function is specifically used, which has smooth saturation characteristics and can effectively simulate the saturation behavior of the actuator. Network input vector From system power reference value and actual power state vector Together they constitute.

[0053] To ensure that the neural network can dynamically track time-varying execution deviation characteristics, the weights of the second layer are... A specific adaptive update rate was designed. This update rate has a clear physical meaning, namely, the first term on the right side of the equation. This constitutes a gradient descent learning term based on Lyapunov stability, where is a constant matrix, and represents the learning rate of the neural network, which controls the speed of weight updates. This term drives the neural network weights along a path that reduces the system's tracking error. E Adjust the direction; the second item It can be a specially designed damping term, through normal numbers k nn Adjusting the damping strength prevents excessive numerical fluctuations in the weight matrix during the adaptive process, ensuring the numerical stability and convergence of the learning process.

[0054] In practical digital control systems, the first-order Euler discretization method can be used to convert the continuous-time update rate into a discrete form that can be directly programmed: ; So that in each sampling period Ts The system can update the weight parameters based on the latest state. Based on the updated weights... ( k +1), the estimated execution bias at the current moment can be obtained in real time through the forward computation of the neural network. This estimate will be directly used in the feedforward compensation stage of the control input to effectively suppress the saturation nonlinearity effect of the actuator.

[0055] In one embodiment, the active front-end rectifier system parameters are as follows: AC grid output frequency. Set to 50Hz, DC bus voltage reference value Set to 500V, control system sampling time Set to 100μs, DC side capacitor value Set to 2700μF, grid-side filter inductor Set to 10mH, its series equivalent resistance Set to 1Ω, load resistance value Set to 40Ω.

[0056] Figure 4 The horizontal axis represents time, and the vertical axis represents the instantaneous value of the three-phase current, which shows the quality of the AC side current waveform. The displayed AC side three-phase current waveform exhibits a high degree of sinusoidality and symmetry. Figure 5 The horizontal axis represents time, and the vertical axis represents DC voltage value, reflecting the dynamic response of DC side voltage. The dynamic process of DC side output voltage shows that it can quickly converge to the 500V reference value after a short start-up phase, with small overshoot and short adjustment time. Figure 6The horizontal axis represents time, and the vertical axis represents power value, presenting the active / reactive power tracking process. The active and reactive power response curves on the grid side show that both can smoothly and accurately track the given reference value with minimal steady-state fluctuations, achieving precise power decoupling control. Figure 7 The horizontal axis represents the harmonic order, and the vertical axis represents the harmonic content rate, showing the characteristics of current harmonic distribution. It further shows that the amplitude of each harmonic component is suppressed to a low level, effectively improving the power grid current quality.

[0057] These simulation results collectively demonstrate that the method described in this application exhibits superior overall performance in terms of dynamic response, steady-state accuracy, and power quality.

[0058] Based on the same concept, this application also provides an active front-end rectifier configured to apply the above-described predictive control method.

[0059] Specifically, the active front-end rectifier protected in this application calculates instantaneous active and reactive power in a stationary αβ coordinate system based on the acquired grid voltage and current signals. It utilizes a superspiral observer, a nonlinear estimator, to capture power dynamics and observe all model uncertainties and external disturbances as lumped expansion states in real time. It uses a fast gradient method to adaptively update the observer's key parameters (input gain) online, completely eliminating prior dependence on system physical parameters (such as filter inductors). An integrated neural network compensator fits and compensates for nonlinear execution deviations caused by the voltage saturation limitation of the power switches. The controller converts the processed control commands into optimal switching vectors within the framework of finite set model predictive control and directly drives the rectifier's power switching devices (such as IGBTs).

[0060] Through collaborative design combining hardware and software, the active front-end rectifier can consistently demonstrate excellent steady-state accuracy, fast dynamic response, and strong anti-interference capability in practical industrial applications, especially under complex operating conditions with parameter drift and model uncertainty, thus becoming a high-performance and highly reliable power conversion device.

[0061] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a solid-state disk (SSD), etc.

[0062] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The above are merely preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A predictive control method for an active front-end rectifier, characterized in that, The predictive control method includes: Based on the voltage model of the three-level active front-end rectifier, the active power and reactive power of the power grid are calculated according to the instantaneous power theory. Based on the voltage model, and in conjunction with the active power and the reactive power, the power gradient is calculated; The power gradient is converted into a second-order state-space equation. The control input of the active front-end rectifier is predicted using a super-spiral observer. Based on the predicted control input and the grid-side voltage, the input voltage of the active front-end rectifier is obtained. Based on the input voltage, an optimal voltage vector is obtained using a finite set model predictive control framework, and the active front-end rectifier is controlled based on the switching state corresponding to the optimal voltage vector.

2. The predictive control method for an active front-end rectifier according to claim 1, characterized in that, The predictive control method further includes using a fast gradient method to adaptively update the input gain of the superspiral observer, including: When the input gain of the superspiral observer is mismatched, the second-order state-space equation can be expressed as: ； In the formula, , , , , , P g This indicates the active power. Q g This indicates the reactive power. Indicates the system's filter resistance. Indicates the filter inductance. , Indicates grid-side voltage Axial components, Indicates grid-side voltage Axial components, Indicates the angular frequency on the power grid side. Indicates input gain The estimated value; Calculate state variables X The normal unit vector of 1, the transpose of the normal unit vector and the expansion state X 2. Perform the inner product and construct the error equation for the fast gradient algorithm based on the inner product result; Based on the error equation, the input gain of the superspiral observer is estimated, and the input gain of the superspiral observer is updated based on the estimated input gain.

3. The predictive control method for an active front-end rectifier according to claim 2, characterized in that, The error equation for constructing the fast gradient algorithm based on the inner product result includes: Remove some known quantities from the inner product result using the following formula: ； In the formula, Represents state variables X The normal unit vector of 1, Indicates the angular frequency on the power grid side. The error equation is configured to calculate the square of the difference between zero and the inner product result after removing some known quantities.

4. The predictive control method for an active front-end rectifier according to claim 3, characterized in that, Based on the error equation, the input gain of the superspiral observer is estimated, including the following steps: Based on the error equation, an estimate of the input gain is obtained. The gradient, based on the input gain estimate. The gradient is used to adaptively update the input gain estimate using a discrete-form fast gradient method. .

5. The predictive control method for an active front-end rectifier according to claim 1, characterized in that, The predictive control method further includes: An actuator is used to trim the control input of the active front-end rectifier output from the superspiral observer. Based on the trimmed control input and the grid-side voltage, the input voltage of the active front-end rectifier is obtained. Specifically, a neural network is used to fit the execution deviation of the actuator to obtain an estimated execution deviation value. The estimated execution deviation value is then subtracted from the control input of the active front-end rectifier output by the superspiral observer to obtain the trimmed control input.

6. The predictive control method for an active front-end rectifier according to claim 5, characterized in that, The neural network is a two-layer neural network, and the neural network is represented by the following formula: ； In the formula, V T This represents the fixed and known weights of the first layer of the neural network. This represents the weights of the second layer of the neural network. This represents a known activation function. X NN =[ R*X 1] T , , P g * This represents a reference value for the active power. Q g * This represents a reference value for the reactive power.

7. The predictive control method for the active front-end rectifier according to claim 4, characterized in that, The predictive control method further includes: The update rate of the second layer weights of the neural network is set, and the second layer weights of the neural network are updated using the first-order Euler discretization method based on the update rate. The execution bias estimate is calculated based on the updated neural network.

8. The predictive control method for an active front-end rectifier according to claim 5, characterized in that, The update rate of the second layer weights of the neural network is configured as follows: ； In the formula, , is a constant matrix. For positive integers, E This represents the tracking error of the system. b This indicates the input gain.

9. The predictive control method for an active front-end rectifier according to claim 1, characterized in that, The prediction of the control input of the active front-end rectifier using a superhelical observer includes: The superspiral observer is used to establish a nonlinear sliding membrane function for the state variables in the second-order state-space equation. X 1 and expansion state X 2. Make an estimate to obtain the first observation state and the second observation state; Discretize the first and second observation states, and tune the superspiral observer so that the second observation state approximates the expanded state. X 2, System status The control input of the superspiral observer is obtained by subtracting the estimate of the extended state from the derivative of the reference value, then subtracting the system tracking error modulated by the error coefficient, adding the robust control term of the system, and finally dividing by the input gain.

10. An active front-end rectifier, characterized in that, The active front-end rectifier is configured to apply the predictive control method as described in any one of claims 1 to 9.

Images