A three-phase four-bridge-arm inverter fixed-frequency model predictive control method and device

Abstract

The application discloses a three-phase four-bridge-arm inverter constant-frequency model prediction control method and device; a teacher model is constructed according to the operating state quantity of the three-phase four-bridge-arm inverter, the teacher model pre-selects a plurality of direction-optimal effective vectors, and the optimal duty ratio combination is obtained in the space of the effective vectors through local grid search; the student model is regressed and trained through a training data set, and the upper and lower limit safety constraints of the inverter hardware are met while generating continuous duty ratios through a soft physical limiter; the weight parameters of the student model are deployed on a simulation controller, the continuous duty ratio instruction is obtained through the simulation controller, and the inverter is driven according to the continuous duty ratio instruction; in the case that the calculation amount of the student model is unchanged, constant-frequency modulation output is realized, and the total current harmonic distortion is reduced.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment condition monitoring, and in particular to a fixed-frequency model predictive control method and device for a three-phase four-bridge inverter. Background Technology

[0002] Distributed power sources (such as photovoltaic and wind power) and electrified loads (such as electric vehicle charging stations) are being connected to the distribution network on a large scale. Due to the randomness of distributed power output and the high degree of asymmetry in the spatial and temporal distribution of terminal loads, modern distribution networks are facing increasingly severe three-phase imbalance problems. Three-phase imbalance not only leads to excessive neutral current and increased transformer losses, but also seriously threatens the operational safety of off-grid power supply systems and precision sensitive loads in microgrids.

[0003] To address the demands for imbalance management and flexible power flow control in complex power grid environments, the three-phase four-leg (3P4L) inverter has emerged as a key power conversion topology. Compared to traditional three-phase three-leg inverters, the three-phase four-leg inverter, by adding an independent fourth leg (neutral leg), provides a path for zero-sequence current, thereby achieving completely independent decoupled control of the power in each phase. This offers significant advantages when handling three-phase unbalanced loads and nonlinear loads.

[0004] In the control strategy of three-phase four-arm inverters, Finite-Control-Set Model Predictive Control (FCS-MPC) has attracted widespread attention due to its extremely fast dynamic response, lack of pulse width modulation (PWM) module, and ease of handling multi-constraint nonlinear problems. However, traditional FCS-MPC faces two core problems in practical engineering applications:

[0005] I. Heavy Online Computational Burden: The three-phase four-arm bridge has 16 discrete switching states. To improve the steady-state performance of the system, a long-horizon prediction strategy is often required. However, the computational load increases exponentially with the prediction step size. For example, with a step size of 3, thousands of value function calculations are required per sampling period, which is almost impossible for conventional low-cost DSPs or MCUs.

[0006] Second, the switching frequency is not fixed: FCS-MPC is essentially based on direct control of discrete switching states, which causes the inverter switching frequency to fluctuate drastically over a wide frequency band. This not only makes it difficult to filter out high-frequency switching ripple, but also generates serious electromagnetic interference (EMI), limiting its application in high-quality power supply fields.

[0007] To address the computational burden issue, existing research proposes using artificial neural networks (ANNs) or one-dimensional convolutional neural networks (1D-CNNs) to learn the nonlinear control laws of MPC through offline training. During online execution, only simple forward inference is needed to generate control commands. However, if such "CNN replacing MPC" schemes still use the logic of FCS-MPC, their output remains a discrete switching state, failing to address the inherent defect of indeterminate frequency.

[0008] In addition, although modulation model predictive control achieves a fixed switching frequency through vector synthesis technology, its core vector selection and duty cycle calculation process (especially the grid optimization method used to obtain extremely low THD) involves an extremely large number of iterative cycles, and its computing power bottleneck is unbearable for conventional DSPs or MCUs. Summary of the Invention

[0009] The following is an overview of the topics described in detail in this article.

[0010] The purpose of this application is to at least partially solve one of the technical problems existing in the related technologies. The embodiments of this application provide a fixed-frequency model predictive control method and device for a three-phase four-arm inverter.

[0011] An embodiment of the first aspect of this application provides a fixed-frequency model predictive control method for a three-phase four-arm inverter, characterized in that it includes:

[0012] A teacher model is constructed based on the operating state variables of a three-phase four-arm inverter. The teacher model pre-screens multiple optimal effective vectors in multiple directions, and the optimal duty cycle combination is obtained through local grid search within the space of the effective vectors.

[0013] The training dataset is constructed by using the operating state variables of a three-phase four-arm inverter as the dataset input and the continuous duty cycle after balancing the neutral arm potential as the output label.

[0014] A student model is constructed based on the teacher model. The student model is trained by regression using the training dataset. The activation function of the output layer of the student model is set to a soft physical limiter, which satisfies the upper and lower limit safety constraints of the inverter hardware while generating continuous duty cycles.

[0015] The weight parameters of the student model are deployed on the simulation controller. The real-time state feature vector of the three-phase four-arm inverter is input to the simulation controller to calculate the continuous duty cycle command. The inverter is driven according to the continuous duty cycle command.

[0016] According to certain embodiments of the first aspect of this application, a teacher model is constructed based on the operating state variables of a three-phase four-arm inverter, including:

[0017] Based on the duty cycle of the previous sampling time, a one-beat delay compensation is performed to predict the current value at the next sampling time.

[0018] Based on the evaluation function, the optimal effective vectors in multiple directions are selected from multiple basic effective vectors;

[0019] Under the constraint that the sum of the duty cycles of the effective vectors participating in the synthesis is equal to 1, local grid optimization is performed in the duty cycle space formed by the effective vectors and the zero vector with a fixed step size to find the optimal duty cycle combination that minimizes the mean square error of the current.

[0020] By dynamically allocating zero vector time, the duty cycle of the centerline arm is brought closer to the preset duty cycle, generating a target equivalent four-arm continuous duty cycle label.

[0021] According to certain embodiments of the first aspect of this application, predicting the current value at the next sampling time by performing one-beat delay compensation based on the duty cycle of the previous sampling time includes:

[0022] Based on the actual duty cycle of each phase arm in the previous sampling period, the actual duty cycle of the neutral arm in the previous sampling period, and the DC bus voltage, calculate the average output voltage of each phase arm relative to the neutral arm in the previous sampling period.

[0023] The equivalent voltage difference between each phase for current prediction is calculated based on the difference between the average output voltage of each phase arm relative to the neutral arm and the voltage at the load end of each phase during the previous sampling period.

[0024] Calculate the system common-mode compensation voltage caused by inductive coupling of the neutral branch based on the three-phase equivalent voltage difference, the neutral line filter inductance and the three-phase AC side filter inductance.

[0025] Using the Euler discretization model, the current value at the next sampling time is predicted based on the current of each phase filter inductor at the previous sampling time, the control sampling period, the three-phase filter inductors, and the difference between the equivalent voltage difference of each phase and the common-mode compensation voltage of the system.

[0026] According to certain embodiments of the first aspect of this application, the value function for local mesh optimization is expressed as follows:

[0027] ;

[0028] in, ;

[0029] ;

[0030] ;

[0031] In the formula, This represents the current prediction error value function used to evaluate the quality of duty cycle combinations. This indicates the prediction obtained under the combined effect of candidate duty cycles. Phase inductor current, It represents any one of the three phases: phase a, phase b, and phase c. This indicates the prediction obtained under the combined effect of candidate duty cycles. Phase inductor current, This indicates the predicted next sampling time after one-beat delay compensation. Phase filter inductor current, This represents the duty cycle of the m-th basic effective vector participating in the optimization. This represents the change in current caused by the corresponding fundamental effective vector within one sampling period. This indicates the control sampling period. This represents a three-phase filter inductor. Represents the m-th fundamental effective vector. Equivalent voltage difference, This represents the common-mode compensation voltage corresponding to the m-th fundamental effective vector. In the m-th fundamental effective vector The switching status of the phase bridge arm, This indicates the switching state of the centerline arm in the m-th basic effective vector. Indicates the DC bus voltage. express Phase load terminal voltage.

[0032] According to certain embodiments of the first aspect of this application, the student model is a one-dimensional convolutional neural network;

[0033] The hierarchical structure of the student model includes: a first fully connected layer, a one-dimensional convolutional layer, a batch normalization layer and an activation layer, a max pooling layer, a feature fusion layer, and an output layer;

[0034] The first fully connected layer maps the input features to a high-dimensional feature space, and after ReLU activation, it is reshaped into a 1-channel sequence with a length of 18.

[0035] The one-dimensional convolutional layer compresses the sequence length and extracts local features.

[0036] The local features are normalized and then activated by ReLU through the batch normalization layer and activation layer.

[0037] The output features of the batch normalization layer and the activation layer are subjected to max pooling through a max pooling layer, which halves the sequence length.

[0038] The pooled 16-dimensional features are flattened through the feature fusion layer and then nonlinearly fused through the second fully connected layer.

[0039] The output layer utilizes a fully connected layer and a sigmoid activation function to output a continuous duty cycle.

[0040] According to certain embodiments of the first aspect of this application, the soft physical limiter maps the network-predicted infinite-domain output value to a restricted physical domain that conforms to PWM modulation requirements.

[0041] According to certain embodiments of the first aspect of this application, during regression training, a regression fitting strategy is adopted to minimize the mean square error loss between the model prediction duty cycle and the teacher model label, so that the student model can use its generalization performance to perform second-order smoothing on the discrete stepping error generated by grid optimization, thereby reducing the total harmonic distortion of the current during online operation.

[0042] According to certain embodiments of the first aspect of this application, the state feature vector includes: the three-phase reference current output by the voltage outer loop, the three-phase measured inductor current, the three-phase output load voltage, and the duty cycle of the four bridge arms at the previous sampling time; the state feature vector is standardized based on the mean and variance obtained from the teacher model dataset before being input into the model.

[0043] According to a second aspect of this application, an electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the fixed-frequency model predictive control method for a three-phase four-arm inverter as described in an embodiment of the first aspect of this application.

[0044] According to a third aspect of this application, a computer storage medium stores computer-executable instructions for performing the fixed-frequency model predictive control method for a three-phase four-arm inverter as described in an embodiment of the first aspect of this application.

[0045] The above scheme has at least the following beneficial effects: it simplifies the tens of thousands of logical judgments and cost function comparisons in traditional long-range optimization to a fixed-length matrix operation in a single forward propagation of a neural network. The online computational load is reduced by more than two orders of magnitude compared to the traditional grid optimization method, and the computation time is constant, not changing with the improvement of optimization accuracy or the increase of prediction step size, which greatly reduces the requirements for the computing power of the control chip;

[0046] By using a multi-vector output with a continuous duty cycle instead of a single-vector discrete state, this invention, combined with PWM modulation technology, achieves fixed switching frequency operation of the inverter, significantly improving the output waveform of the off-grid system. Simultaneously, by utilizing the nonlinear interpolation and regression smoothing effect of neural networks on discrete grid labels, the online generated control commands are more continuous than the original teacher model, and the measured current THD index even outperforms the grid optimization method used to generate the data.

[0047] By decoupling the computing architecture from the physical parameters of the controlled object, and by changing different training datasets and fine-tuning the loss function type, it is possible to flexibly switch between different performance tendencies (such as fixed frequency, variable frequency, long step size, and short step size) under the same hardware topology, thereby improving engineering versatility and robustness. Attached Figure Description

[0048] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0049] Figure 1 This is a flowchart illustrating the steps of the fixed-frequency model predictive control method for a three-phase four-arm inverter.

[0050] Figure 2 This is a schematic diagram of the overall system topology of a fixed-frequency model predictive control system for an off-grid three-phase four-arm inverter.

[0051] Figure 3 It is a multi-vector model predictive control block diagram;

[0052] Figure 4 This is a flowchart of the training process for a 1-D CNN model;

[0053] Figure 5 This is a waveform diagram of the three-phase inductor current output by a fixed-frequency 1-D CNN;

[0054] Figure 6 This is a performance comparison chart under different control methods. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, or the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0057] The embodiments of this application provide a fixed-frequency model predictive control method and apparatus for a three-phase four-arm inverter.

[0058] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0059] Reference Figure 1 A fixed-frequency model predictive control method for a three-phase four-arm inverter includes the following steps:

[0060] Step S100: Construct a teacher model based on the operating state variables of the three-phase four-arm inverter. The teacher model pre-selects the optimal effective vectors in multiple directions. The optimal duty cycle combination is obtained by local grid search within the space of the effective vectors.

[0061] Step S200: Use the operating state variables of the three-phase four-arm inverter as the dataset input and the continuous duty cycle after balancing the neutral arm potential as the output label to construct the training dataset.

[0062] Step S300: Construct a student model based on the teacher model, perform regression training on the student model using the training dataset, and set the output layer activation function of the student model to a soft physical limiter to meet the upper and lower limit safety constraints of the inverter hardware while generating continuous duty cycles.

[0063] In step S400, the weight parameters of the student model are deployed to the simulation controller, the real-time state feature vector of the three-phase four-arm inverter is input to the simulation controller to calculate the continuous duty cycle command, and the inverter is driven according to the continuous duty cycle command.

[0064] Reference Figure 2 For a three-phase four-arm inverter, the converter is powered by a DC-side power supply V. DC Three-phase four-arm inverter bridge, output filter circuit and load R load The circuit consists of three bridge arms: the first three are phase bridge arms (phases a, b, and c), and the fourth is the neutral point bridge arm (phase n). The output filter circuit uses an LC structure and includes a three-phase filter inductor L. f , Neutral filter inductor L n and three-phase filter capacitor C f This structure, by adding a fourth bridge arm and a neutral point inductor, provides a zero-sequence current path, effectively addressing three-phase unbalanced load conditions.

[0065] Before constructing a high-performance teacher model based on multi-vector grid optimization, we first conducted a feasibility study on the strong coupling and high nonlinearity characteristics of the three-phase four-arm converter control system, and established a control performance comparison benchmark.

[0066] Specifically, based on the derived three-phase four-arm bridge mathematical model, a prediction step size of N is adopted. pThe Finite Control Set Model Predictive Control (FCS-MPC) algorithm with a step size of 3 is used as the benchmark. The purpose of choosing a long step size prediction is to induce more complex characteristics in the system by increasing the discrete search space and computational complexity of the control law. The algorithm is used for simulation, and the system's state input vector and corresponding optimal switching actions are collected in real time, thereby constructing a basic dataset containing the "state-action" mapping relationship.

[0067] This phase of the work has multiple objectives:

[0068] 1. Feasibility verification: By observing the classification accuracy of the neural network for discrete switch action selection, the ability of the convolution operator to fit complex predictive control logic is verified.

[0069] 2. Performance Comparison Benchmark between Fixed-Frequency and Variable-Frequency Controls: Since the FCS-MPC outputs only a single switching state in each sampling period, its corresponding neural network controller essentially falls under the category of "variable-frequency control." This model aims to provide a crucial comparative reference for subsequent "fixed-frequency regression neural networks" based on MV-MPCC. It also more intuitively demonstrates the technical advantages of fixed-frequency CNN controllers in improving power quality and optimizing filter design.

[0070] 3. Using two different teacher models, we verify the engineering generality and robustness of the proposed 1-D CNN.

[0071] The verification and comparison work in this stage has demonstrated the potential of neural networks to handle three-phase four-arm control problems and clarified the necessity of evolving from indeterminate frequency discrete control to fixed frequency continuous duty cycle control.

[0072] The teacher model is constructed based on the operating state variables of the three-phase four-arm inverter, including the following steps:

[0073] Based on the duty cycle of the previous sampling time, a one-beat delay compensation is performed to predict the current value at the next sampling time.

[0074] Based on the evaluation function, the optimal effective vectors in multiple directions are selected from multiple basic effective vectors;

[0075] Under the constraint that the sum of the duty cycles of the effective vectors participating in the synthesis is equal to 1, local grid optimization is performed in the duty cycle space composed of the effective vectors and the zero vector with a fixed step size to find the optimal duty cycle combination that minimizes the mean square error of the current.

[0076] By dynamically allocating zero vector time, the duty cycle of the centerline arm is brought closer to the preset duty cycle, generating a target equivalent four-arm continuous duty cycle label.

[0077] Based on the duty cycle of the previous sampling time, a one-beat delay compensation is performed to predict the current value at the next sampling time, including the following steps:

[0078] Step S110: Based on the actual duty cycle of each phase bridge arm in the previous sampling period, the actual duty cycle of the neutral bridge arm in the previous sampling period, and the DC bus voltage, calculate the average output voltage of each phase bridge arm relative to the neutral bridge arm in the previous sampling period.

[0079] Step S120: Calculate the equivalent voltage difference of each phase for current prediction based on the difference between the average output voltage of each phase arm relative to the neutral arm and the voltage at the load end of each phase during the previous sampling period.

[0080] Step S130: Calculate the system common-mode compensation voltage caused by inductive coupling of the neutral branch based on the three-phase equivalent voltage difference, the neutral filter inductance, and the three-phase AC side filter inductance.

[0081] Step S140: Using the Euler discretization model, predict the current value at the next sampling time based on the current of each phase filter inductor at the previous sampling time, the control sampling period, the three-phase filter inductors, and the difference between the equivalent voltage difference of each phase and the system common-mode compensation voltage.

[0082] Specifically, to achieve model predictive control, a predictive model of the system in the abc coordinate system is first established. According to Kirchhoff's laws and the average switching cycle method, the rate of change of the system current is determined by the difference between the inverter output voltage vector and the load voltage, as follows:

[0083] ;

[0084] ;

[0085] In the formula, It represents any one of the three phases. Indicates the period within the kth sampling period The average output voltage of the phase bridge arm relative to the neutral bridge arm. express Phase load terminal voltage, This represents the x-phase equivalent voltage difference used for current prediction. Indicates the previous sampling period The actual duty cycle of the phase bridge arm. This indicates the actual duty cycle of the midline bridge arm in the previous sampling period. This indicates the DC bus voltage.

[0086] Each duty cycle satisfies: 0 ≤ ≤1, 0≤ ≤1.

[0087] The system common-mode compensation voltage caused by inductive coupling in the neutral branch can be expressed by the following formula: ;

[0088] In the formula, This represents the system common-mode compensation voltage caused by inductive coupling in the neutral branch; This indicates the three-phase AC side filter inductor. Indicates the neutral line filter inductance; , , These represent the equivalent voltage differences between phases a, b, and c, respectively.

[0089] The predicted current value at the next sampling time is obtained by calculating using the Euler discretization model:

[0090] ;

[0091] In the formula, Indicates the first Each sampling time Phase filter inductor current; This indicates the predicted next sampling time after one-beat delay compensation. Phase filter inductor current; Indicates the control sampling period; This represents the filter inductor on the three-phase AC side; This indicates the sequence number of the discrete sampling time.

[0092] Based on feasibility verification and benchmark establishment, in order to further improve the steady-state performance of the system and provide more accurate teacher signals for the 1D-CNN student model, a multi-vector model predictive control (MV-MPC) based on heuristic selection and grid optimization is used as the final teacher model.

[0093] To ensure that the converter output voltage can accurately track the reference voltage and reduce harmonic content, the geometric dimension of the required number of synthesized vectors was derived:

[0094] Two-dimensional spatial analysis: In a traditional three-phase three-arm converter, the reference voltage vector vref lies in a two-dimensional plane. To synthesize any vector within the two-dimensional plane region, at least three basic voltage vectors (two effective vectors plus one zero vector) are required for combination.

[0095] Three-dimensional spatial recursion: Since the three-phase four-arm converter adds a neutral point arm, the system dimension is expanded from two-dimensional to three-dimensional, and its switching vector forms a spatial dodecahedron composed of 16 vertices in space.

[0096] The synthesis rule is determined: Based on spatial geometry theory, to synthesize any reference voltage vector in three-dimensional space without error, it is necessary to use four basic voltage vectors (i.e., three non-coplanar effective vectors and one zero vector) in a linear combination.

[0097] Through the above simple derivation, a four-vector synthesis strategy using three effective vectors in conjunction with a zero vector was determined, so that the trajectory range of the synthesized voltage vector is exactly the same as the range of the three-dimensional reference voltage vector.

[0098] Reference Figure 3 The implementation process of the MV-MPC teacher model is as follows:

[0099] One-beat delay compensation: In digital control systems, computation time causes a one-beat delay in the control action. This delay is eliminated by predicting the state at time k+1. The duty cycle d from the previous execution time is used. (k-1) Calculate the predicted value This serves as the starting point for the search for optimization.

[0100] Effective vector screening and value function: Three-phase four-bridge arm has 2 4 =16 basic spatial vectors. To reduce the complexity of offline optimization, a simple value function is first used to select the three basic effective vectors closest to the reference vector. The cost function is:

[0101] ;

[0102] Where m∈Ωv represents the set of 14 effective vectors after removing two zero vectors from the 16 basic vectors of the three-phase four-arm inverter; the three vectors with the smallest g are selected as candidate effective vectors for subsequent local mesh optimization. i x,m Let be the change in current caused by the m-th base vector within one sampling period.

[0103] The three optimal effective vectors with the lowest cost are selected by using the instantaneous current error cost function and then used in the subsequent grid optimization.

[0104] Local mesh optimization: Within the space defined by the three selected effective vectors and the zero vector, perform a detailed traversal of duty cycles d1, d2, d3, and d0. Constraints: Final value function The objective is to minimize the mean square error of the current at time k+2.

[0105] The value function for local grid optimization is expressed as:

[0106] ;

[0107] in, ;

[0108] ;

[0109] ;

[0110] In the formula, This represents the current prediction error value function used to evaluate the quality of duty cycle combinations. This indicates the prediction obtained under the combined effect of candidate duty cycles. Phase inductor current, It represents any one of the three phases: phase a, phase b, and phase c. This indicates the prediction obtained under the combined effect of candidate duty cycles. Phase inductor current, This indicates the predicted next sampling time after one-beat delay compensation. Phase filter inductor current, This represents the duty cycle of the m-th basic effective vector participating in the optimization. This represents the change in current caused by the corresponding fundamental effective vector within one sampling period. This indicates the control sampling period. This represents a three-phase filter inductor. Represents the m-th fundamental effective vector. Equivalent voltage difference, This represents the common-mode compensation voltage corresponding to the m-th fundamental effective vector. In the m-th fundamental effective vector The switching status of the phase bridge arm, This indicates the switching state of the centerline arm in the m-th basic effective vector. Indicates the DC bus voltage. express Phase load terminal voltage.

[0111] Iterate through all duty cycle combinations and take... The combination of duty cycles corresponding to the minimum value is taken as the optimal vector allocation scheme.

[0112] An equivalent transformation is performed on the vector duty cycle obtained from mesh optimization, the duration of the zero vector is dynamically allocated, the duty cycle of the midline arm is forced to approach 0.5, the neutral point potential is balanced, and the final 4D continuous duty cycle label D = [d a , d b , d c ,d n ].

[0113] The entire real-time system operating status is collected, constituting the network input: the three-phase reference current output from the voltage outer loop, the three-phase measured inductor current, the three-phase output load voltage, and the duty cycle of the four bridge arms at the previous sampling time. The state feature vectors are standardized based on the mean and variance obtained from the teacher model dataset before being input into the model. The 13-dimensional input is standardized based on the statistical mean and variance of the entire dataset to eliminate differences in the dimensions of voltage, current, and duty cycle, accelerating network training convergence. This results in a 13-dimensional standardized regression training dataset.

[0114] The student model is a one-dimensional convolutional neural network; a 1D-CNN structure, which can greatly reduce the computational complexity of online inference while ensuring control accuracy.

[0115] The hierarchical structure of the student model includes: Input Layer, First Fully Connected Layer (FC1), One-Dimensional Convolutional Layer (Conv1d), Batch Normalization and Activation Layer, Max Pooling Layer, Feature Fusion Layer (FC2), and Output Layer.

[0116] Through the input layer: 13-dimensional feature vectors are received, including the reference current i. ref Measuring current i meas Load voltage u load Critical system status, etc.

[0117] The first fully connected layer maps the input features to a high-dimensional feature space, and after ReLU activation, reshapes them into a 1-channel sequence of length 18; specifically, it maps the 13-dimensional input to an 18-dimensional high-dimensional space. Its function is to perform preliminary feature fusion on the original input electrical signal. Furthermore, the ReLU activation function introduces a nonlinear activation mechanism, enabling the network to fit complex nonlinear control laws while effectively mitigating the gradient vanishing problem in deep networks.

[0118] A one-dimensional convolutional layer is used to compress the sequence length and extract local features. Specifically, after reshaping the data into a sequence form, a one-dimensional convolution operator with 2 channels and a kernel size of 3 is used. The function of this layer is to extract local correlation features between the reference current, measured current, load voltage, and historical duty cycle along the input feature dimension, and to extract transient features.

[0119] By using batch normalization layers and activation layers, local features are normalized and then activated by ReLU; this aims to accelerate network convergence and enhance the model's robustness to load perturbations.

[0120] Max pooling is used to perform max pooling on the output features of the batch normalization layer and the activation layer, halving the sequence length. Specifically, a pooling operation with a stride and kernel size of 2 is used to reduce the feature length by half. Its purpose is to extract the most significant features while reducing computational cost, thus enhancing the model's robustness against sampling noise.

[0121] The feature fusion layer flattens the pooled 16-dimensional features and performs non-linear fusion through the second fully connected layer, enabling deep logical reasoning and feature integration.

[0122] Through the output layer, a continuous duty cycle is output using a fully connected layer and a sigmoid activation function. The continuous duty cycle consists of four signals, each corresponding to an equivalent control command for one of the four bridge arms.

[0123] The Sigmoid activation function no longer represents classification probability, but instead serves as a physical limiter in the inverter hardware system. It is used to force the infinite domain output value predicted by the network to be mapped to a restricted physical domain that meets the requirements of PWM modulation, ensuring that the output value is within the legal range (0, 1) of the physical duty cycle.

[0124] The Sigmoid activation function is expressed as: ;

[0125] In the formula, This represents the 16-dimensional hidden feature vector output by the second fully connected layer; ∈R 4×16 This represents the output layer weight matrix; ∈R 4 This represents the output layer bias vector; =[da,db,dc,dn] T This represents the continuous duty cycle vector of the four bridge arms of the neural network output; the Sigmoid function is used to map the output to the (0,1) interval to meet the physical range requirements of the PWM modulation duty cycle.

[0126] Through this network structure, it is possible to achieve N p Accurate fitting of the predictive control law was achieved under the complex operating condition of 3, reducing the computational cost from approximately 130,000 FLOPS to approximately 1,600 FLOPS while maintaining the same performance as the original FCS-MPC. This verifies the ability of the 1D-CNN model to handle the three-phase four-arm strongly coupled control problem. The computational complexity remains unchanged with increases in the teacher model's prediction step size or improvements in grid optimization accuracy.

[0127] Reference Figure 4 To transform the output of the student model from a discrete switching state with an indefinite frequency to a continuous duty cycle with a fixed frequency, the mean squared error function is used as the loss function while retaining the original CNN hierarchical structure.

[0128] The mean square error function is expressed as: ;

[0129] In the formula, Indicates the number of samples in a single training batch; Indicates the training sample number; This indicates the duty cycle output channel number of the four bridge arms, corresponding to the four bridge arms a, b, c, and n respectively; The first number predicted by the neural network is... The first sample Duty cycle of each bridge arm; This indicates the corresponding duty cycle label generated by the MV-MPC teacher model.

[0130] This regression strategy enables the network to fit an extremely smooth control law surface, thereby eliminating the step quantization error caused by mesh optimization during online operation.

[0131] During regression training, a regression fitting strategy is adopted. By minimizing the mean squared error loss between the model's predicted duty cycle and the teacher model's label, the student model utilizes its generalization performance to perform second-order smoothing on the discrete stepping error generated by grid optimization, thereby reducing the total harmonic distortion of the current during online operation, which is lower than that of the teacher model.

[0132] Reference Figure 5 , Figure 5 This is a waveform diagram of the three-phase inductor current output by a fixed-frequency 1-D CNN.

[0133] After offline training convergence, the weight parameters of the student model are extracted and reconstructed in C language and deployed in the simulation controller. During the online operation phase, the controller collects and inputs the current system's state feature vector in real time. Relying on the network's single forward inference, it directly calculates four continuous duty cycle commands, which are then modulated by the PWM module to drive the inverter. This replaces the traditional online rolling optimization process of MPC with a fixed switching frequency.

[0134] The performance of several control methods (FCS-MPC, variable frequency 1-DCNN, MV-MPC, and fixed frequency 1-DCNN) is compared, and the simulation parameters used are shown in Table 1.

[0135] Table 1 Simulation Parameter List

[0136]

[0137] Reference Figure 6As shown in Table 1, FCS-MPC and MV-MPC, due to their ergonomic optimization methods, require computations in the tens of thousands of FLOPS. The variable-frequency 1-D CNN, while maintaining the same THD, simplifies the computation to approximately 1600 FLOPS. Meanwhile, the fixed-frequency 1-D CNN, while maintaining the same 1600 FLOPS computation, surpasses the teacher model in performance, reducing the THD from 2.85% to 1.75%, thus resolving the computational bottleneck and variable-frequency control problem inherent in traditional FCS-MPC.

[0138] In this embodiment, the tens of thousands of logical judgments and cost function comparisons in traditional long-range optimization are simplified to a fixed-length matrix operation in a single forward propagation of the neural network. The online computational load is reduced by more than two orders of magnitude compared to the traditional grid optimization method, and the computation time remains constant, not changing with the improvement of optimization accuracy or the increase of prediction step size, which greatly reduces the computing power requirements of the control chip.

[0139] By using a multi-vector output with a continuous duty cycle instead of a single-vector discrete state, this invention, combined with PWM modulation technology, achieves fixed switching frequency operation of the inverter, significantly improving the output waveform of the off-grid system. Simultaneously, by utilizing the nonlinear interpolation and regression smoothing effect of neural networks on discrete grid labels, the online generated control commands are more continuous than the original teacher model, and the measured current THD index even outperforms the grid optimization method used to generate the data.

[0140] By decoupling the computing architecture from the physical parameters of the controlled object, and by changing different training datasets and fine-tuning the loss function type, it is possible to flexibly switch between different performance tendencies (such as fixed frequency, variable frequency, long step size, and short step size) under the same hardware topology, thereby improving engineering versatility and robustness.

[0141] Embodiments of this application provide an electronic device. The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the fixed-frequency model predictive control method for a three-phase four-arm inverter as described above.

[0142] This electronic device can be any smart terminal, including computers.

[0143] In general, for the hardware structure of electronic devices, the processor can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits to execute relevant programs and implement the technical solutions provided in the embodiments of this application.

[0144] The memory can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory and called by the processor to execute the methods of the embodiments of this application.

[0145] Input / output interfaces are used to implement information input and output.

[0146] The communication interface is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0147] The bus transmits information between various components of a device, such as the processor, memory, input / output interfaces, and communication interfaces. The processor, memory, input / output interfaces, and communication interfaces communicate with each other within the device via the bus.

[0148] Embodiments of this application provide a computer storage medium. The computer storage medium stores computer-executable instructions for executing the three-phase four-arm inverter fixed-frequency model predictive control method described above.

[0149] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium. In the foregoing description of this specification, references to terms such as "one embodiment," "another embodiment," or "some embodiments," etc., indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of this application. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0150] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0151] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0152] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0153] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0154] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed between each other may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms. Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0155] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A fixed-frequency model predictive control method for a three-phase four-arm inverter, characterized in that, include: A teacher model is constructed based on the operating state variables of a three-phase four-arm inverter. The teacher model pre-screens multiple optimal effective vectors in multiple directions, and the optimal duty cycle combination is obtained through local grid search within the space of the effective vectors. The training dataset is constructed by using the operating state variables of a three-phase four-arm inverter as the dataset input and the continuous duty cycle after balancing the neutral arm potential as the output label. A student model is constructed based on the teacher model. The student model is trained by regression using the training dataset. The activation function of the output layer of the student model is set to a soft physical limiter, which satisfies the upper and lower limit safety constraints of the inverter hardware while generating continuous duty cycles. The weight parameters of the student model are deployed on the simulation controller. The real-time state feature vector of the three-phase four-arm inverter is input to the simulation controller to calculate the continuous duty cycle command. The inverter is driven according to the continuous duty cycle command.

2. The fixed-frequency model predictive control method for a three-phase four-arm inverter according to claim 1, characterized in that, A teacher model is constructed based on the operating state variables of a three-phase four-arm inverter, including: Based on the duty cycle of the previous sampling time, a one-beat delay compensation is performed to predict the current value at the next sampling time. Based on the evaluation function, the optimal effective vectors in multiple directions are selected from multiple basic effective vectors; Under the constraint that the sum of the duty cycles of the effective vectors participating in the synthesis is equal to 1, local grid optimization is performed in the duty cycle space formed by the effective vectors and the zero vector with a fixed step size to find the optimal duty cycle combination that minimizes the mean square error of the current. By dynamically allocating zero vector time, the duty cycle of the centerline arm is brought closer to the preset duty cycle, generating a target equivalent four-arm continuous duty cycle label.

3. The fixed-frequency model predictive control method for a three-phase four-arm inverter according to claim 2, characterized in that, Based on the duty cycle of the previous sampling time, a one-beat delay compensation is performed to predict the current value at the next sampling time, including: Based on the actual duty cycle of each phase arm in the previous sampling period, the actual duty cycle of the neutral arm in the previous sampling period, and the DC bus voltage, calculate the average output voltage of each phase arm relative to the neutral arm in the previous sampling period. The equivalent voltage difference between each phase for current prediction is calculated based on the difference between the average output voltage of each phase arm relative to the neutral arm and the voltage at the load end of each phase during the previous sampling period. Calculate the system common-mode compensation voltage caused by inductive coupling of the neutral branch based on the three-phase equivalent voltage difference, the neutral line filter inductance and the three-phase AC side filter inductance. Using the Euler discretization model, the current value at the next sampling time is predicted based on the current of each phase filter inductor at the previous sampling time, the control sampling period, the three-phase filter inductors, and the difference between the equivalent voltage difference of each phase and the common-mode compensation voltage of the system.

4. The fixed-frequency model predictive control method for a three-phase four-arm inverter according to claim 2, characterized in that, The value function for local grid optimization is expressed as: ； in, ; ； ； In the formula, This represents the current prediction error value function used to evaluate the quality of duty cycle combinations. This indicates the prediction obtained under the combined effect of candidate duty cycles. Phase inductor current, It represents any one of the three phases: phase a, phase b, and phase c. This indicates the prediction obtained under the combined effect of candidate duty cycles. Phase inductor current, This indicates the predicted next sampling time after one-beat delay compensation. Phase filter inductor current, This represents the duty cycle of the m-th basic effective vector participating in the optimization. This represents the change in current caused by the corresponding fundamental effective vector within one sampling period. This indicates the control sampling period. This represents a three-phase filter inductor. Represents the m-th fundamental effective vector. Equivalent voltage difference, This represents the common-mode compensation voltage corresponding to the m-th fundamental effective vector. In the m-th fundamental effective vector The switching status of the phase bridge arm, This indicates the switching state of the centerline arm in the m-th basic effective vector. Indicates the DC bus voltage. express Phase load terminal voltage.

5. The fixed-frequency model predictive control method for a three-phase four-arm inverter according to claim 1, characterized in that, The student model is a one-dimensional convolutional neural network; The hierarchical structure of the student model includes: a first fully connected layer, a one-dimensional convolutional layer, a batch normalization layer and an activation layer, a max pooling layer, a feature fusion layer, and an output layer; The first fully connected layer maps the input features to a high-dimensional feature space, and after ReLU activation, it is reshaped into a 1-channel sequence with a length of 18. The one-dimensional convolutional layer compresses the sequence length and extracts local features. The local features are normalized and then activated by ReLU through the batch normalization layer and activation layer. The output features of the batch normalization layer and the activation layer are subjected to max pooling through a max pooling layer, which halves the sequence length. The pooled 16-dimensional features are flattened through the feature fusion layer and then nonlinearly fused through the second fully connected layer. The output layer utilizes a fully connected layer and a sigmoid activation function to output a continuous duty cycle.

6. The fixed-frequency model predictive control method for a three-phase four-arm inverter according to claim 5, characterized in that, The soft physical limiter maps the infinite domain output value predicted by the network to a restricted physical domain that meets the requirements of PWM modulation.

7. The fixed-frequency model predictive control method for a three-phase four-arm inverter according to claim 1, characterized in that, During regression training, a regression fitting strategy is adopted. By minimizing the mean squared error loss between the model prediction duty cycle and the teacher model label, the student model can use its generalization performance to perform second-order smoothing on the discrete stepping error generated by grid optimization, thereby reducing the total harmonic distortion of the current during online operation.

8. The fixed-frequency model predictive control method for a three-phase four-arm inverter according to claim 1, characterized in that, The state feature vector includes: the three-phase reference current output by the voltage outer loop, the three-phase measured inductor current, the three-phase output load voltage, and the duty cycle of the four bridge arms at the previous sampling time; the state feature vector is standardized based on the mean and variance obtained from the teacher model dataset before being input into the model.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the fixed-frequency model predictive control method for a three-phase four-arm inverter as described in any one of claims 1 to 8.

10. A computer storage medium, characterized in that, The device stores computer-executable instructions for performing the fixed-frequency model predictive control method for a three-phase four-arm inverter as described in any one of claims 1 to 8.

Images