Modular multilevel matrix converter and control method thereof

Abstract

The application discloses a control method of a modular multilevel matrix converter, comprising: obtaining a timing data set of the modular multilevel matrix converter; performing data preprocessing on the timing data set to obtain a standardized timing data tensor, time characteristic tokens and variable characteristic tokens; performing timing and channel feature extraction on the standardized timing data tensor to obtain an enhanced timing feature tensor; performing variable dimension rearrangement on the enhanced timing feature tensor to obtain a rearranged feature tensor organized along a variable dimension; encoding the rearranged feature tensor through an encoder to output an encoded feature tensor; performing inverse dimension rearrangement on the encoded feature tensor to obtain an inverse rearranged feature tensor; mapping the inverse rearranged feature tensor into a control signal through a multilayer perception machine; and controlling the modular multilevel matrix converter based on the control signal. The method can realize real-time predictive control, reduce control delay and maintain high-precision tracking performance.

Description

Technical Field

[0001] This application relates to the field of power electronics technology, and in particular to a modular multilevel matrix converter and its control method. Background Technology

[0002] In related technologies, finite control set model predictive control (FCS-MPC) has the advantages of intuitive structure and strong nonlinearity handling capability. However, when the number of switch combinations increases, its computational complexity increases sharply, making it difficult to meet the real-time requirements of the control cycle. In order to reduce the amount of computation, existing studies often decompose the optimization problem into independent solutions for each bridge arm. However, this ignores the physical coupling information between the bridge arms, resulting in a decrease in control accuracy.

[0003] Neural network-based time series forecasting models are used to accelerate decision-making. For example, recurrent neural networks (RNNs) and their variant LSTM perform well in long-term dependency modeling, but have limited ability to express complex physical coupling relationships. While attention networks and various time series-specific Transformer structures perform well in general forecasting tasks, they still have problems in characterizing cross-variable correlations and physical constraints in power electronic control scenarios with strong physical coupling. Summary of the Invention

[0004] This application provides a modular multilevel matrix converter and its control method. The method enables real-time predictive control of the modular multilevel matrix converter, reducing control delay while maintaining high-precision tracking performance.

[0005] In a first aspect, this application provides a control method for a modular multilevel matrix converter, the method comprising: Obtain the timing dataset of the modular multilevel matrix converter. The timing dataset includes data from multiple sampling points. The data for any sampling point includes the corresponding grid-side voltage, grid-side current, load-side voltage, load-side current, currents of the nine bridge arms, nine historical control signals, and timestamps. Data preprocessing is performed on the time series dataset to obtain a standardized time series data tensor; Temporal and channel features are extracted from the standardized time series data tensor to obtain the enhanced temporal feature tensor. The enhanced temporal feature tensor is organized along the time dimension and contains local temporal dependencies and cross-channel dynamic dependencies. The variable dimensions of the enhanced temporal feature tensor are rearranged to obtain a rearranged feature tensor organized along the variable dimensions. The rearranged feature tensor is encoded by an encoder, and the encoded feature tensor is output. The encoder includes an attention sublayer and a nonlinear feedforward sublayer. The attention sublayer explicitly models the dynamic coupling relationship between variables along the variable dimension, and the nonlinear feedforward sublayer enhances the nonlinear representation capability of the encoded feature tensor, so that the encoded feature tensor reflects the cross-variable physical coupling characteristics between electrical variables. The encoded feature tensor is rearranged in reverse dimension to restore it to its time-dimension-organized form, thus obtaining the inverse rearranged feature tensor. The inverse rearrangement feature tensor is mapped to a control signal using a multilayer perceptron, and the modular multilevel matrix converter is controlled based on the control signal.

[0006] In one embodiment, a convolutional feature extraction module is used to extract temporal and channel features from a standardized temporal data tensor to obtain an enhanced temporal feature tensor. The convolutional feature extraction module includes a channel attention mechanism and a cross-information fusion mechanism. The convolutional feature extraction module extracts features in parallel through the first one-dimensional convolutional path and the second one-dimensional convolutional path to obtain the first channel features and the second channel features that represent local features. The convolutional kernels of the first one-dimensional convolutional path and the second one-dimensional convolutional path are different. The cross-information fusion mechanism is used to cross-fuse the first-channel features and the second-channel features after processing by the channel attention mechanism; The channel attention mechanism is used to learn the importance weights of channel features from the first and second channel features; Based on importance weights, the convolutional feature extraction module performs convolution through a third one-dimensional convolutional path to obtain an enhanced temporal feature tensor. The convolutional kernel of the third one-dimensional convolutional path is larger than the convolutional kernel of the first one-dimensional convolutional path and also larger than the convolutional kernel of the second one-dimensional convolutional path.

[0007] In one embodiment, the channel attention mechanism includes: The first channel feature is subjected to GELU nonlinear processing to obtain a first nonlinear value. The first nonlinear value is then processed by the Sigmoid function and multiplied element-wise by twice the first channel feature to obtain a first self-gated value. Furthermore, the second channel feature is subjected to GELU nonlinear processing to obtain a second nonlinear value. After processing the second nonlinear value with the Sigmoid function, it is multiplied element-wise with twice the second channel feature to obtain a second self-gated value. Cross-information fusion mechanisms include: The first channel feature and the second self-gated value are multiplied element-wise to obtain the first element-wise multiplication value; The second channel feature is multiplied element-wise with the first self-gated value to obtain the second element-wise multiplication value; After adding the multiplication values ​​of the first and second elements, channel weighting is performed through a channel attention mechanism. The channel attention mechanism learns the importance weights of each channel through global average pooling and channel attention.

[0008] In one embodiment, the variable dimension of the enhanced temporal feature tensor is rearranged by reversing the operation, so that the enhanced temporal feature tensor is rearranged from the form of batch multiplied by length multiplied by channel to the form of batch multiplied by channel multiplied by length, thereby obtaining a rearranged feature tensor organized along the variable dimension. The rearranged feature tensor includes multiple variable feature tokens, each variable feature token corresponding to the complete evolution trajectory of a single physical variable over the entire time window.

[0009] In one embodiment, the attention sublayer explicitly models the dynamic coupling relationships between variables along the variable dimension, including: Based on the rearranged feature tensor, the attention operator works along the variable dimension, directly modeling the coupling between voltage signals, current signals and historical control signals, thereby explicitly modeling the dynamic coupling relationship between variables; The attention operator can be any of the following: self-attention operator, linear attention operator, or kernelized attention operator.

[0010] In one embodiment, the encoder uses the iKanformer GR-KAN nonlinear feedforward sublayer to nonlinearly map the output of the attention sublayer, thereby enhancing the nonlinear representation capability of the encoded feature tensor and enabling the encoded feature tensor to reflect the cross-variable physical coupling characteristics between electrical variables.

[0011] In one embodiment, the control method further includes: After the attention sublayer, the GR-KAN nonlinear feedforward sublayer is activated by grouped rational functions and parameters are shared within each group. The polynomial coefficients of the rational function are shared in each channel group.

[0012] In one embodiment, a spline-activated GR-KAN nonlinear feedforward sublayer is employed after the attention sublayer.

[0013] In one embodiment, the encoder uses a Transformer MLP nonlinear feedforward sublayer to nonlinearly map the output of the attention sublayer.

[0014] Secondly, this application also provides a modular multilevel matrix converter, which includes: The control unit is configured to perform a control method for a modular multilevel matrix converter as described in the first aspect.

[0015] The aforementioned control method for a modular multilevel matrix converter (MMC) involves acquiring the time-series dataset of the MMC, followed by preprocessing to obtain a standardized time-series data tensor. Further, time-series and channel features are extracted from the standardized time-series data tensor to obtain an enhanced time-series feature tensor organized along the time dimension and containing local time-series dependencies and cross-channel dynamic dependencies. Further still, the enhanced time-series feature tensor undergoes variable-dimensional rearrangement, transforming it into a variable-dimensional organization. The rearranged feature tensor is processed by an encoder that explicitly models the dynamic coupling between variables along the variable dimension using an attention sublayer and enhances the nonlinear representation using a nonlinear feedforward sublayer, thus enabling the output encoded feature tensor to reflect cross-variable physical coupling characteristics. The encoded feature tensor undergoes inverse-dimensional rearrangement to restore its time-dimensional organization. Finally, a multilayer perceptron maps this feature tensor to a control signal, which is then used to control the converter. This method effectively avoids the bottleneck of excessive computational load in predictive control in related technologies, reduces control latency, and meets real-time requirements. Attached Figure Description

[0016] Figure 1 This is a flowchart of a control method for a modular multilevel matrix converter in one embodiment; Figure 2 This is a schematic diagram of the control system of a modular multilevel matrix converter in one embodiment; Figure 3 This is a circuit diagram of a modular multilevel matrix converter in one embodiment; Figure 4 This is a flowchart of obtaining the enhanced temporal feature tensor in one embodiment; Figure 5 This is an encoder diagram of the GR-KAN nonlinear feedforward sublayer using iKanformer in one embodiment; Figure 6 This is an encoder diagram of an MLP nonlinear feedforward sublayer using Transformer in one embodiment; Figure 7 This is a flowchart of a control method for a modular multilevel matrix converter in another embodiment. Detailed Implementation

[0017] 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.

[0018] It should be noted that, in this document, relational terms such as "first" and "second" are used merely 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 the element.

[0019] In one embodiment, such as Figure 1 As shown, a control method for a modular multilevel matrix converter is provided, which includes the following steps: Step 101: Obtain the timing dataset of the modular multilevel matrix converter. The timing dataset includes data from multiple sampling points. The data for any sampling point includes the corresponding grid-side voltage, grid-side current, load-side voltage, load-side current, currents of the nine bridge arms, nine historical control signals, and timestamps. Specifically, such as Figure 2 As shown, the time-series dataset is obtained in real time by the modular multilevel matrix converter through the sensor module 201. This time-series dataset includes data from multiple sampling points, and the data at each sampling point includes: three-phase grid-side voltage ( ) and grid-side current ( Three-phase grid-side voltage and grid-side current are used to indicate the grid-side status; three-phase load-side voltage ( ) and three-phase load side current ( The three-phase load-side voltage and three-phase load-side current are used to indicate the load-side status; the nine arm currents ( The 9-channel historical control signal can reflect the instantaneous current of the nine bridge arms of the multilevel matrix converter; the 9-channel historical control signal can refer to the switching control signal applied to the nine bridge arms during sampling; the timestamp is used to mark the sampling time corresponding to the sampling point.

[0020] Step 102: Perform data preprocessing on the time series dataset to obtain a standardized time series data tensor. The standardized time series data tensor includes time feature tokens and variable feature tokens. The variable feature tokens include the current signals of the 9 bridge arms, the voltage signals of the 6 bridge arms, and the historical control signals of the 9 bridge arms. Specifically, the time-series dataset collected in chronological order is divided into a series of fixed-length data segments (i.e., time windows). The input window length can be set to 150 sampling points. Each sampling point includes: 1 timestamp, 9 bridge arm currents, 6 voltage signals, and 9 historical control signals, totaling 25 dimensions of data.

[0021] To prevent the large differences in the dimensions and numerical ranges of different physical quantities from affecting model training and convergence, the 25-dimensional data needs to be standardized. For example, Z-score standardization can be used, which involves calculating the mean μ and standard deviation σ for each dimension of the data across the entire training dataset, and applying the transformation (x-μ) / σ to each data point so that the mean of each dimension is 0 and the standard deviation is 1.

[0022] After the above slicing and standardization processes, the time series dataset is constructed into a standardized time series data tensor, which can be represented as: Where B represents the batch dimension; L represents the time dimension; and C represents the number of channels (i.e., the variable dimension).

[0023] In a standardized time-series data tensor, the time feature token is the set of preprocessed timestamps, and the variable feature token is the set of all physical variables (i.e., the current signals of the 9 bridge arms, the voltage signals of the 6 bridge arms, and the historical control signals of the 9 bridge arms) at a certain moment.

[0024] Step 103: Extract time series and channel features from the standardized time series data tensor to obtain the enhanced time series feature tensor. The enhanced time series feature tensor is organized along the time dimension and contains local time series dependencies and cross-channel dynamic dependencies. Specifically, the feature extraction module 203 receives the standardized time-series data tensor. Step 103 involves extracting temporal and channel features to obtain an enhanced temporal feature tensor. Small receptive fields (i.e., convolutional kernels) can be used for small-scale hybrid feature capture to extract local features. Local features include local temporal dependencies and cross-channel dynamic dependencies. Local temporal dependencies refer to the correlation or evolution of the same physical variable (i.e., the same channel) between adjacent or nearby time steps, while cross-channel dynamic dependencies refer to the interactions, couplings, or causal relationships between different physical variables (different channels) during temporal evolution.

[0025] It should be noted that the extracted enhanced temporal feature tensor is organized along the time dimension, that is, with the time step as the main axis of the sequence. The attention modeling objective based on this enhanced temporal feature tensor focuses on the cross-time dependency of the same variable.

[0026] Step 104: Rearrange the variable dimensions of the enhanced temporal feature tensor to obtain the rearranged feature tensor organized along the variable dimensions; Specifically, the rearrangement module 204 permutates the second dimension (time dimension L) and the third dimension (variable dimension C) of the input enhanced temporal feature tensor (shape [B, L, C]) to obtain the rearranged feature tensor (shape [B, C, L]).

[0027] Before rearrangement, the enhanced temporal feature tensor is organized primarily by the time dimension L. Each token corresponds to a specific time point, and each token includes a mixture of features of all variables at that time (e.g., current, voltage, and historical control signals of each bridge arm).

[0028] After rearrangement, a rearranged feature tensor organized along the variable dimension C is obtained. At this point, each token corresponds to a specific feature channel, and the token itself is a sequence of length L, representing the evolution trajectory of that feature channel throughout the entire time window. By rearranging the variable dimension, the subsequent attention modeling objective shifts from "emphasizing the cross-temporal dependence of the same variable" to "emphasizing the dynamic coupling between different variables," thus enabling the solution provided in this application to facilitate the modeling of strong physical coupling.

[0029] Step 105: Encode the rearranged feature tensor through an encoder and output the encoded feature tensor. The encoder includes an attention sublayer and a nonlinear feedforward sublayer. The attention sublayer explicitly models the dynamic coupling relationship between variables along the variable dimension. The nonlinear feedforward sublayer enhances the nonlinear representation capability of the encoded feature tensor, so that the encoded feature tensor reflects the cross-variable physical coupling characteristics between electrical variables. Specifically, the rearranged feature tensor treats the complete temporal evolution trajectory of each physical variable as a token. Using the attention sublayer in encoder 205, an attention mechanism is applied between these tokens, enabling the model to directly and dynamically learn the influence between different variables, thereby effectively capturing the inherent multivariate coupling characteristics of the physical system. Then, a nonlinear feedforward sublayer is used for mapping to improve the nonlinear expression and reduce the number of parameters, so that the encoded feature tensor output by the encoder can reflect the cross-variable physical coupling characteristics between electrical variables.

[0030] The attention sublayer and the nonlinear feedforward sublayer work together, and the output encoded feature tensor is no longer a simple combination of the original signals, but a deeply integrated and highly abstracted cross-variable physical coupling characteristics between electrical variables. This approach is particularly suitable for fitting high-dimensional, nonlinear, and strongly coupled physical control systems such as M3C.

[0031] Step 106: Perform reverse dimensional rearrangement on the encoded feature tensor to restore it to its form organized along the time dimension, and obtain the reverse rearranged feature tensor; Specifically, the inverse rearrangement module 206 outputs the encoded feature tensor (with shape [B, C, L]). By performing a dimension permutation operation, the variable dimension C and the time dimension L are swapped again. After the inverse rearrangement, the shape of the resulting inverse rearranged feature tensor is restored to [B, L, C]. This indicates that the tensor has returned to a form where the variable dimension C is dominant, while the time dimension L is dominant.

[0032] Step 107: Map the inverse rearrangement feature tensor to a control signal using a multilayer perceptron, and control the modular multilevel matrix converter based on the control signal.

[0033] Specifically, signal mapping is performed using a multilayer perceptron 207. The input is an inversely rearranged feature tensor with a shape of [B, L, C]. The multilayer perceptron 207 acts as a mapping head, decoding the inversely rearranged feature tensor and mapping it to a specific, physically executable control signal space. For example, the output of the multilayer perceptron 207 is a multi-step prediction of nine control signals, i.e., predicting the control command sequence that the nine bridge arms should have within a future time period (e.g., an output window length of 100 sampling points).

[0034] The control signal output by the multilayer sensor 207 (usually taking the first step or the first few steps in the multi-step prediction as the immediate command of the current control cycle) is transmitted to the power drive unit of the converter, thereby controlling the switching devices (such as IGBTs) of each bridge arm submodule to turn on and off, thus actually changing the bridge arm voltage and current.

[0035] It should be noted that while outputting control commands, a sorted capacitor voltage balancing strategy needs to be invoked. Since the neural network outputs a macroscopic bridge arm level requirement, the specific selection of which submodule to engage or disengage requires dynamic sorting and selection based on the actual voltage levels of each submodule's capacitors to ensure that all submodule capacitor voltages remain balanced.

[0036] In this embodiment, a time-series dataset of a modular multilevel matrix converter is acquired, and then preprocessed to obtain a normalized time-series data tensor. Further, time-series and channel features are extracted from the normalized time-series data tensor to obtain an enhanced time-series feature tensor organized along the time dimension and containing local time-series dependencies and cross-channel dynamic dependencies. Further still, the enhanced time-series feature tensor undergoes variable-dimensional rearrangement, transforming it into a variable-dimensional organization. The rearranged feature tensor is processed by an encoder that uses an attention sublayer to explicitly model the dynamic coupling between variables along the variable dimension and uses a nonlinear feedforward sublayer to enhance the nonlinear representation, thereby enabling the output encoded feature tensor to reflect cross-variable physical coupling characteristics. The encoded feature tensor undergoes inverse-dimensional rearrangement to restore its time-dimensional organization. Finally, a multilayer perceptron maps this feature tensor to a control signal, which is then used to control the converter. This method effectively avoids the bottleneck of excessive computational load in predictive control in related technologies, reduces control latency, and meets real-time requirements.

[0037] In one embodiment, such as Figure 3 As shown, the grid side includes the three-phase grid-side voltage ( And the equivalent impedance Z of each phase in series s (Used to limit short-circuit current and filter), three-phase load side current ( The load side includes the three-phase load impedance Z. load (Equivalent impedance of the actual load), three-phase load impedance Z load The voltage across the terminals is the voltage on the three-phase load side. ), the current flowing through the output three-phase load side ( There are a total of 9 bridge arms. B 1 to B 9) Each bridge arm is composed of bridge arm inductance. L b (Used to suppress sudden current changes and circulating currents) and consists of several cascaded sub-modules (e.g., SM1, SM2, ..., SM...). N Each submodule contains a capacitor (voltage is...). V cn The output multi-level voltage is generated by switching a controllable switch.

[0038] In one embodiment, the currents of the nine bridge arms are: The voltages of the nine bridge arms are: Let the common mode voltage be... The dynamic relationship between the voltage and current of the bridge arm of the modular multilevel matrix converter is as follows: .

[0039] in, Represents the equivalent voltage vectors at the input and output terminals.

[0040] Adopting dual The transformation decouples the input equivalent voltage, output equivalent voltage, and circulating current component. Let the voltage or current matrix to be transformed be... ,but .

[0041] in, The transformation matrix is ​​the Clarke transform or its extended form. For in double Representation in a coordinate system.

[0042] Voltage or current matrix via double After the transformation, the dynamic equations for the bridge arm voltage, current, and input / output voltage difference can be written as: .

[0043] in, Indicates the difference between input and output voltages. Representation in coordinate system , These are the vectors of the bridge arm current and voltage in this coordinate system, respectively.

[0044] Considering discrete-time implementation, let the sampling period be . Then the grid-side and load-side currents can be approximated as:

[0045] in, For the grid-side inductance and resistance, These are load-side parameters.

[0046] In traditional finite control set MPC, the current tracking errors of nine bridge arms are typically selected as the main optimization objectives. Let the first arm be... The reference value for the current of each bridge arm is... The predicted value is Then the cost function can be defined as: .

[0047] Based on the above cost function formula, traditional FCS-MPC requires model recursion and cost calculation for a large number of candidate switch states in order to obtain the optimal switch state.

[0048] The solution in this application directly uses a neural network to approximate optimally map the input time-series dataset (including the corresponding grid-side voltage, grid-side current, load-side voltage, load-side current, nine bridge arm currents, nine historical control signals, and timestamps). During the online phase, the network directly outputs an approximate optimal switching control signal. Based on this optimal switching control signal, the modular multilevel matrix converter (MMC) outputs the next window of control decisions, drives the power devices, and invokes a sorted capacitor voltage equalization strategy to maintain submodule voltage consistency. Compared to traditional schemes that calculate cost functions using finite sets, the solution provided in this application can achieve high-precision, low-delay control signal prediction, which is beneficial for the real-time control of the MMC.

[0049] In one embodiment, such as Figure 4 As shown, the convolutional feature extraction module extracts temporal and channel features from the standardized temporal data tensor to obtain an enhanced temporal feature tensor (the convolutional feature extraction module includes a channel attention mechanism and a cross-information fusion mechanism), including the following steps: Step 401: The convolutional feature extraction module extracts features in parallel through the first one-dimensional convolutional path and the second one-dimensional convolutional path to obtain the first channel features and the second channel features that represent local features. The convolutional kernels of the first one-dimensional convolutional path and the second one-dimensional convolutional path are different. Step 402: The cross-information fusion mechanism is used to cross-fuse the first channel features and the second channel features after the channel attention mechanism has been processed; Step 403: The channel attention mechanism is used to learn the importance weights of the channel features for the first and second channel features; Step 404: Based on importance weights, the convolutional feature extraction module performs convolution through a third one-dimensional convolutional path to obtain an enhanced temporal feature tensor. The convolutional kernel of the third one-dimensional convolutional path is larger than the convolutional kernel of the first one-dimensional convolutional path and larger than the convolutional kernel of the second one-dimensional convolutional path.

[0050] Specifically, the input standardized time series data tensor is: ; in, For batch dimension, For variable dimensions, The time dimension.

[0051] The convolutional feature extraction module extracts features in parallel through the first one-dimensional convolutional path and the second one-dimensional convolutional path to obtain the first channel features. (Kernel size is 1) and second channel features (Kernel size is 3).

[0052] Furthermore, the channel attention mechanism applies GELU nonlinear activation and Sigmoid-based self-gating to the two feature channels respectively to learn the importance weights of each channel feature, adaptively enhancing key features and suppressing noise. The cross-information fusion mechanism cross-fused the results of the self-gating processing of the first and second channel features, and then performed channel weighting to further improve the feature representation capability.

[0053] Finally, based on the weighted features, the convolutional feature extraction module performs convolution through a third one-dimensional convolutional path. This path uses a convolutional operation with a larger kernel (such as size 5 or 7) to expand the temporal receptive field and outputs an enhanced temporal feature tensor. Enhanced temporal feature tensor It can include local temporal dependencies and cross-channel dynamic dependencies.

[0054] In one embodiment, the channel attention mechanism includes: The first channel feature is subjected to GELU nonlinear processing to obtain a first nonlinear value. The first nonlinear value is then processed by the Sigmoid function and multiplied element-wise by twice the first channel feature to obtain a first self-gated value. Furthermore, the second channel feature is subjected to GELU nonlinear processing to obtain a second nonlinear value. After processing the second nonlinear value with the Sigmoid function, it is multiplied element-wise with twice the second channel feature to obtain a second self-gated value.

[0055] Specifically, the channel attention mechanism then processes the two sets of features separately: for the first channel feature... Perform GELU nonlinear processing to obtain the first nonlinear value. and to Perform self-gating processing (based on the Sigmoid function σ( (element-wise multiplication) to obtain the first self-gated value .

[0056] Features of the second channel Perform GELU nonlinear processing to obtain the second nonlinear value. and for the second nonlinear value Perform self-gating processing to obtain the second self-gated value. .

[0057] in, For the Sigmoid function, This indicates element-wise multiplication. The self-gating mechanism adaptively selects channel features that contribute more to predicting control behavior while suppressing noise components.

[0058] In one embodiment, the cross-information fusion mechanism includes: The first channel feature and the second self-gated value are multiplied element-wise to obtain the first element-wise multiplication value; The second channel feature is multiplied element-wise with the first self-gated value to obtain the second element-wise multiplication value; After adding the multiplication values ​​of the first and second elements, channel weighting is performed through a channel attention mechanism. The channel attention mechanism learns the importance weights of each channel through global average pooling and channel attention.

[0059] Specifically, the characteristics of the first channel With the second self-gated value Perform element-wise multiplication to obtain the multiplication value of the first element. ; Features of the second channel With the first self-gated value Perform element-wise multiplication to obtain the multiplication value of the second element. .

[0060] Furthermore, the multiplication value of the first element Multiplication value with the second element Addition yields fusion features The fusion feature The input channel attention mechanism performs channel weighting. This mechanism compresses the spatial-temporal information of each channel through global average pooling to generate channel descriptors, then learns the importance weights of each channel through a fully connected layer, and recalibrates the fused features to output channel-weighted features. .

[0061] Furthermore, the convolutional feature extraction module performs convolution through a third one-dimensional convolutional path. This path uses a convolutional operation with a larger kernel (such as size 5 or 7) to expand the temporal receptive field and output an enhanced temporal feature tensor. .

[0062] In one embodiment, the variable dimension of the enhanced temporal feature tensor is rearranged by reversing the operation, so that the enhanced temporal feature tensor is rearranged from the form of batch multiplied by length multiplied by channel to the form of batch multiplied by channel multiplied by length, thereby obtaining a rearranged feature tensor organized along the variable dimension. The rearranged feature tensor includes multiple variable feature tokens, each variable feature token corresponding to the complete evolution trajectory of a single physical variable over the entire time window.

[0063] Specifically, the enhanced temporal feature tensor output by the self-gated cross-convolution module... The variable dimensions are rearranged through the inverse operation, which is mathematically defined as using a tensor permutation function to transform the dimensional order from (0, 1, 2) (batch dimension, time dimension, variable dimension) to (0, 2, 1) (batch dimension, variable dimension, time dimension). This results in the rearranged feature tensor. .

[0064] This operation transforms the organization of tensors from being primarily time-dimensional (i.e., each token corresponds to the instantaneous state of all 25 variables at a given time point) to being primarily variable-dimensional. In this case, the tensor contains C=25 variable feature tokens, and each variable feature token (e.g., corresponding to the entire channel of a specific bridge arm current or grid-side voltage) contains the complete evolution trajectory of that single physical variable over the entire L=150 time steps.

[0065] In one embodiment, the attention sublayer explicitly models the dynamic coupling relationships between variables along the variable dimension, including: Based on the rearranged feature tensor, the attention operator works along the variable dimension, directly modeling the coupling between voltage signals, current signals and historical control signals, thereby explicitly modeling the dynamic coupling relationship between variables; The attention operator can be any of the following: self-attention operator, linear attention operator, or kernelized attention operator.

[0066] Specifically, the rearranged feature tensor with shape [B, C, L] after the inverse-transpose operation is... The input attention sublayer is organized in such a way that each variable feature token corresponds to the complete evolution trajectory of a single physical variable (such as a specific bridge arm current or grid-side voltage) over the entire time window. Attention operators in the attention sublayer (such as the standard multi-head self-attention operator MSA) are based on rearranged feature tensors. By generating query (Q), key (K), and value (V) matrices and calculating attention weights on the variable dimension C, they directly model the dynamic coupling relationship between grid-side voltage signals, grid-side current signals, load-side voltage signals, load-side current signals, 9-way bridge arm current signals, and 9-way historical control signals. For example, they can explicitly capture the dynamic impact of a change in the current of a certain bridge arm on the control signals of other bridge arms.

[0067] It should be noted that this attention operator can be replaced by a linear attention operator (to reduce computational complexity) or a kernelized attention operator (to enhance nonlinear expressiveness), but the core mechanism of the operator is to work along the variable dimension (rather than the traditional time dimension).

[0068] In one embodiment, such as Figure 5As shown, the encoder uses the iKanformer GR-KAN nonlinear feedforward sublayer to perform nonlinear mapping on the output of the attention sublayer, thereby enhancing the nonlinear representation capability of the encoded feature tensor and enabling the encoded feature tensor to reflect the cross-variable physical coupling characteristics between electrical variables.

[0069] Specifically, in each layer of the iKanformer encoder, let the feature tensor of the output of the attention sublayer (MSA) after layer normalization (LN) be: The transformation is performed using a GR-KAN nonlinear feedforward sublayer. The GR-KAN nonlinear feedforward sublayer replaces the traditional Transformer feedforward network (MLP) with a GR-KAN structure based on the Kolmogorov-Arnold network (KAN) concept. The mathematical expression of the GR-KAN structure is as follows: .

[0070] For example, the GR-KAN sublayer, activated by a grouped rational function, groups the input tensor by channel dimension and applies a rational function with shared parameters to each group of channels. Activation is performed using P(x) = P(x) / Q(x) (where P(x) and Q(x) are learnable polynomials), which shares the polynomial coefficients of rational functions within each channel group, significantly reducing the number of parameters. At the same time, an independent scaling factor is introduced for each edge (input-output connection) to flexibly adjust the feature contribution weights.

[0071] The implementation of grouped rational function activation can be replaced by spline activation, that is, activation is performed by combining spline basis functions. The spline function is approximated by piecewise polynomials to achieve flexible nonlinear mapping. The parameters of the spline function (such as the node vector and coefficients of B-splines) are shared in each group of channels to maintain parameter efficiency.

[0072] Through the design of the GR-KAN sublayer described above, a stronger nonlinear mapping capability than traditional MLP is achieved with fewer parameters. It is suitable for fitting the high-dimensional nonlinear coupling relationships between electrical variables such as grid-side voltage, grid-side current, load-side voltage, load-side current, bridge arm current, and historical control signals in multilevel matrix converters. The output encoded feature tensor It can reflect the cross-variable physical coupling characteristics between various electrical variables.

[0073] In another embodiment, such as Figure 6 As shown, the encoder uses a Transformer MLP nonlinear feedforward sublayer to perform nonlinear mapping on the output of the attention sublayer.

[0074] Specifically, the formula for the feedforward sublayer of the standard MLP in the traditional Transformer is as follows: .

[0075] in, It is usually GELU or ReLU.

[0076] As an alternative implementation, the encoder can choose to use the aforementioned standard multilayer perceptron (MLP) structure as the encoder's nonlinear feedforward sublayer. After the attention sublayer completes the modeling of the dynamic coupling relationship across physical quantities along the variable dimension, the output features of the attention sublayer will be further processed by the Transformer's MLP nonlinear feedforward sublayer.

[0077] This MLP sublayer is activated by two fully connected linear transformations (W1, W2) and their nonlinear activation functions. σ It provides the model with standard nonlinear transformation and feature space mapping capabilities. The MLP nonlinear feedforward sublayer can effectively cooperate with the variable dimension attention mechanism of the front end to learn and represent the coupling characteristics of the system, thereby enabling explicit modeling of the coupling between variables through the attention mechanism.

[0078] In one embodiment, the encoded feature tensor is reverse-dimensionally rearranged to restore its time-dimension-organized form, resulting in a reverse-rearranged feature tensor, including: Specifically, after iKanformer encoding is completed, the encoded feature tensor is restored to its original shape through an inverse operation to obtain the inverse rearranged feature tensor: ; in, To predict the time dimension, This is for the output control signal dimension. This design allows the attention mechanism to explicitly model the dynamic coupling relationships between physical quantities along the variable dimension.

[0079] In one embodiment, such as Figure 7 As shown, the optimal control method for the modular multilevel matrix converter is as follows: The input dimension module receives timestamps, three-phase grid-side voltage and current, three-phase load-side voltage and current, nine bridge arm currents, and nine historical control signals.

[0080] Furthermore, the feature extraction module processes the standardized temporal data tensor through self-gated cross-convolution to extract local temporal dependencies and cross-channel dynamic relationships, and outputs an enhanced temporal feature tensor.

[0081] Furthermore, a dimensionality rearrangement operation is performed on the enhanced temporal feature tensor, that is, the enhanced temporal feature tensor is rearranged from time dimension priority to variable dimension priority, so that each token corresponds to the complete evolution trajectory of a single physical variable, resulting in a rearranged feature tensor.

[0082] Furthermore, the rearranged feature tensor is input into the Kanformer (i.e., the iKanformer encoder), and the dynamic coupling between voltage, current and control signals is explicitly modeled along the variable dimension through the attention sublayer. The nonlinear representation is enhanced by the GR-KAN nonlinear feedforward sublayer, and the encoded feature tensor is output.

[0083] The encoded feature tensor can be reversed dimensionally to restore it to its time-dimension-organized form, thus obtaining the reverse-rearranged feature tensor.

[0084] After the inverse rearrangement of the feature tensor dimensions is restored to the time dimension organization, the multilayer perceptron of the mapping module maps the features into nine control signals for the next multiple steps and outputs them to directly drive each bridge arm of the converter, thereby achieving high-precision, low-latency real-time predictive control.

[0085] In one embodiment, as shown in the table below, the embodiments of this application have significant advantages over related technologies (such as baseline methods like LSTM, Flowformer, Crossformer, and TimeXer), specifically in terms of improved control accuracy and efficiency. In the real-time control testing of a modular multilevel matrix converter, the method (M1) of this application is compared with existing methods (such as M2-M5).

[0086] Experimental results show that the mean square error (MSE) of this application on 9 control signals is 8.57e-02 and the mean absolute error (MAE) is 1.99e-01, while the mean MSE of the LSTM method (M2) is 1.96e-01 and the mean MAE is 3.08e-01, which relatively reduces the MSE by about 26.4% and the MAE by about 5.69%, confirming that this application optimizes real-time performance while ensuring high-precision control.

[0087]

[0088] Based on the same concept, this application also provides a modular multilevel matrix converter, which includes: The control unit is configured to execute the control method of the modular multilevel matrix converter described above.

[0089] Specifically, the control unit is programmed or embedded to sequentially implement functional modules such as data acquisition and caching, data preprocessing and windowing, self-gated cross-convolution feature extraction, variable dimension rearrangement, iKanformer encoding, mapping output of multi-step control signals, and capacitor voltage equalization control.

[0090] The control unit can receive grid-side voltage, grid-side current, load-side voltage, load-side current, and nine bridge arm current signals from sensors. Based on historical control signals, it generates nine control signals for future multi-step operations in real time. These control signals are converted into specific drive commands and applied to the nine bridge arms of the converter through the power drive unit, thereby achieving precise control over the engagement and disengagement of each bridge arm submodule.

[0091] 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.

[0092] 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 control method for a modular multilevel matrix converter, characterized in that, The control method includes: Obtain the timing dataset of the modular multilevel matrix converter. The timing dataset includes data from multiple sampling points. The data of any sampling point includes the corresponding grid-side voltage, grid-side current, load-side voltage, load-side current, nine bridge arm currents, nine historical control signals, and a timestamp. The time series dataset is preprocessed to obtain a standardized time series data tensor; Temporal and channel features are extracted from the standardized time-series data tensor to obtain an enhanced temporal feature tensor. The enhanced temporal feature tensor is organized along the time dimension and includes local temporal dependencies and cross-channel dynamic dependencies. The variable dimensions of the enhanced temporal feature tensor are rearranged to obtain a rearranged feature tensor organized along the variable dimensions; The rearranged feature tensor is encoded by an encoder to output an encoded feature tensor. The encoder includes an attention sublayer and a nonlinear feedforward sublayer. The attention sublayer explicitly models the dynamic coupling relationship between variables along the variable dimension. The nonlinear feedforward sublayer enhances the nonlinear representation capability of the encoded feature tensor, so that the encoded feature tensor reflects the cross-variable physical coupling characteristics between electrical variables. The encoded feature tensor is reverse-dimensionally rearranged to restore its time-dimension-organized form, thus obtaining the reverse-rearranged feature tensor. The inverse rearrangement feature tensor is mapped to a control signal using a multilayer perceptron, and the modular multilevel matrix converter is controlled based on the control signal.

2. The control method for the modular multilevel matrix converter according to claim 1, characterized in that, The standardized temporal data tensor is subjected to temporal and channel feature extraction using a convolutional feature extraction module to obtain the enhanced temporal feature tensor. This convolutional feature extraction module includes a channel attention mechanism and a cross-information fusion mechanism. The convolutional feature extraction module extracts features in parallel through a first one-dimensional convolutional path and a second one-dimensional convolutional path to obtain first channel features and second channel features representing local features. The convolutional kernels of the first one-dimensional convolutional path and the second one-dimensional convolutional path are different. The cross-information fusion mechanism is used to cross-fuse the first channel features and the second channel features after processing by the channel attention mechanism; The channel attention mechanism is used to learn the importance weights of channel features for the first channel features and the second channel features; Based on the importance weights, the convolutional feature extraction module performs convolution through a third one-dimensional convolutional path to obtain the enhanced temporal feature tensor. The convolutional kernel of the third one-dimensional convolutional path is larger than the convolutional kernel of the first one-dimensional convolutional path and larger than the convolutional kernel of the second one-dimensional convolutional path.

3. The control method for the modular multilevel matrix converter according to claim 2, characterized in that, The channel attention mechanism includes: The first channel feature is subjected to GELU nonlinear processing to obtain a first nonlinear value. The first nonlinear value is then processed by the Sigmoid function and multiplied element-wise by twice the first channel feature to obtain a first self-gated value. Furthermore, the second channel feature is subjected to GELU nonlinear processing to obtain a second nonlinear value. After processing the second nonlinear value with the Sigmoid function, it is multiplied element-wise with twice the second channel feature to obtain a second self-gated value. The cross-information fusion mechanism includes: The first channel feature and the second self-gated value are multiplied element-wise to obtain the first element-wise multiplication value; The second channel feature is multiplied element-wise with the first self-gated value to obtain the second element-wise multiplication value; After adding the multiplication values ​​of the first and second elements, channel weighting is performed through the channel attention mechanism, which learns the importance weights of each channel through global average pooling and channel attention.

4. The control method for the modular multilevel matrix converter according to claim 2, characterized in that, The enhanced temporal feature tensor is rearranged in terms of variable dimensions by reversing the inversion operation, so that the enhanced temporal feature tensor is rearranged from the form of batch multiplied by length multiplied by channel to the form of batch multiplied by channel multiplied by length, thereby obtaining a rearranged feature tensor organized along the variable dimension. The rearranged feature tensor includes multiple variable feature tokens, each of which corresponds to the complete evolution trajectory of a single physical variable over the entire time window.

5. The control method for the modular multilevel matrix converter according to claim 1, characterized in that, The attention sublayer explicitly models the dynamic coupling relationships between variables along the variable dimension, including: Based on the rearranged feature tensor, the attention operator works along the variable dimension to directly model the coupling between the voltage signal, the current signal and the historical control signal, thereby explicitly modeling the dynamic coupling relationship between the variables. The attention operator can be any of the following: self-attention operator, linear attention operator, or kernelized attention operator.

6. The control method for the modular multilevel matrix converter according to claim 5, characterized in that, The encoder employs the iKanformer GR-KAN nonlinear feedforward sublayer to perform nonlinear mapping on the output of the attention sublayer, thereby enhancing the nonlinear representation capability of the encoded feature tensor and enabling the encoded feature tensor to reflect the cross-variable physical coupling characteristics between electrical variables.

7. The control method for the modular multilevel matrix converter according to claim 6, characterized in that, The control method further includes: After the attention sublayer, the GR-KAN nonlinear feedforward sublayer is activated using a grouped rational function and parameters are shared within the group. The polynomial coefficients of the rational function are shared in each group of channels.

8. The control method for the modular multilevel matrix converter according to claim 6, characterized in that, After the attention sublayer, the GR-KAN nonlinear feedforward sublayer is activated using splines.

9. The control method for the modular multilevel matrix converter according to claim 5, characterized in that, The encoder employs a Transformer MLP nonlinear feedforward sublayer to perform nonlinear mapping on the output of the attention sublayer.

10. A modular multilevel matrix converter, characterized in that, The modular multilevel matrix converter includes: A control unit configured to perform a control method for a modular multilevel matrix converter as described in any one of claims 1 to 9.

Images