A Multi-Objective Optimization Control Method for Grid-Connected Converters Based on Deep Neural Networks

By predicting the switching state of grid-connected converters using a deep neural network model, the real-time performance and scalability issues of traditional MPC in power electronic converter control are solved, achieving faster response speed and higher current tracking accuracy.

CN123122287APending Publication Date: 2026-11-17TAIHANG LABORATORY +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611611374.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional MPCs suffer from real-time bottlenecks and insufficient scalability in power electronic converter control, making it difficult to meet the requirements of rapid response and multi-objective control.

Method used

A multi-objective optimization control method for grid-connected inverters based on deep neural networks is adopted. By constructing a simulation model of a three-phase four-arm grid-connected inverter, simulation operation status data is obtained, a training dataset is built, and a deep neural network model is trained. This enables direct prediction of switching states during online control, reducing the burden of enumeration and real-time calculation.

Benefits of technology

It improves the dynamic response speed and adaptability to structural changes of the three-phase four-bridge grid-connected inverter, reduces control delay, and enhances current tracking accuracy and stability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122287A_ABST
    Figure CN123122287A_ABST
Patent Text Reader

Abstract

This application provides a multi-objective optimization control method for grid-connected inverters based on deep neural networks, belonging to the field of power electronics technology. Specifically, it includes acquiring simulation operation state data of a three-phase four-arm grid-connected inverter simulation model; processing the simulation operation state data to construct a training dataset; constructing a deep neural network model for multi-objective optimization control of the grid-connected inverter; training the deep neural network model using the training dataset to obtain the trained deep neural network model; processing the real-time acquired actual operation state data of the three-phase four-arm grid-connected inverter; and inputting the processed data into the trained deep neural network model to output the final complete on / off state combination. Through this application's processing scheme, the dynamic response speed of the three-phase four-arm grid-connected inverter and its adaptability to changes in the inverter's structure and constraints are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of power electronics technology, and in particular to a multi-objective optimization control method for grid-connected converters based on deep neural networks. Background Technology

[0002] In the field of power electronic converter control, MPC can directly adapt to the nonlinear characteristics and multivariable constraints of the system, and has been widely used in core control scenarios such as current tracking and voltage regulation. Its typical technical approach involves predefining all feasible combinations of IGBT switching devices. Within each control cycle, based on system parameters such as inductance and capacitance, combined with real-time current and voltage states, the future behavior of the system under each switching state is predicted. The control effect of each state is evaluated through a cost function, and finally, the optimal switching state is selected to drive the actuator. MPC stands for Model Predictive Control, and IGBT stands for Insulated Gate Bipolar Transistor.

[0003] However, this type of finite set MPC based on enumeration has two major limitations that severely restrict its practical application in power electronic control. The first is the real-time bottleneck, where there is a contradiction between its online computational burden and the real-time control requirements. Traditional MPC requires multiple forward predictions and cost function evaluations for all switch state combinations in the finite control set within each control cycle, a typical sequential enumeration solution. Although the calculation for a single state evaluation is relatively simple, the number of combinations increases dramatically with the number of prediction steps, resulting in significant cumulative computational delays. This makes it difficult to output control commands in a timely manner, affecting the accuracy of switch timing and causing stability problems such as increased current tracking deviation and deteriorated output ripple.

[0004] Secondly, there is insufficient scalability. Expanding the system scale leads to a significant increase in computational complexity. Since the number of switching states increases exponentially with the number of IGBTs, and the amount of online computation is positively correlated with this, traditional MPC becomes extremely cumbersome when facing system upgrades or multi-objective control constraints. More importantly, MPC typically relies on fixed weighting coefficients to coordinate different control objectives, making it difficult to adapt to the dynamic adjustment requirements of objective priorities when operating conditions change. Summary of the Invention

[0005] In view of this, this application provides a multi-objective optimization control method for grid-connected inverters based on deep neural networks, which solves the problems in the prior art and improves the dynamic response speed of three-phase four-arm grid-connected inverters and their adaptability to changes in the structure and constraints of three-phase four-arm grid-connected inverters.

[0006] This application provides a multi-objective optimization control method for grid-connected converters based on deep neural networks, which adopts the following technical solution: A multi-objective optimization control method for grid-connected converters based on deep neural networks includes the following steps: A simulation model of a three-phase four-arm grid-connected inverter was built. Simulation experiments were conducted on the simulation model of the three-phase four-arm grid-connected inverter under the model predictive control algorithm. Simulation operation status data of the three-phase four-arm grid-connected inverter simulation model in each control cycle under multiple operating conditions were obtained, as well as the simulation optimal on / off state combination of the four insulated gate bipolar transistors of the upper arm in the simulation model of the three-phase four-arm grid-connected inverter. Constructing a training dataset: The simulation operation state data of the three-phase four-arm grid-connected inverter simulation model in each control cycle is used as a set of simulation raw data. Each set of simulation raw data is used as initial data. Data processing is performed on each set of initial data, and the processed data is paired one by one with the simulation optimal on / off state combination. The data processing methods include extracting original state features and operating condition coupling derived features from the initial data, normalizing the original state features and operating condition coupling derived features to obtain feature samples of the corresponding initial data; attaching corresponding event labels to the event samples in the initial data according to predefined event classification rules, with event classification including steady-state events and operating condition feature mutation events; constructing a steady-state sliding time series window for steady-state events, and constructing a time series window including the occurrence time of the operating condition feature mutation event for operating condition feature mutation events. Construct a deep neural network model for multi-objective optimization control of grid-connected converters; The deep neural network model is trained using the training dataset to obtain the trained deep neural network model; Online control: Real-time acquisition of actual operating status data of the three-phase four-arm grid-connected inverter; using the actual operating status data of the current control cycle of the three-phase four-arm grid-connected inverter as initial data for data processing; and inputting the data processed by the current control cycle and the data processed by all control cycles within the timing window into the trained deep neural network model. The trained deep neural network model outputs the final complete on / off state combination of the eight insulated gate bipolar transistors in the three-phase four-arm grid-connected inverter.

[0007] Optionally, the simulation operation status data of the three-phase four-arm grid-connected inverter simulation model in each control cycle under multiple operating conditions, and the operating conditions during the simulation of the optimal on / off state combination of the four insulated gate bipolar transistors of the upper arm in the three-phase four-arm grid-connected inverter simulation model include: Operating conditions with reference current varying from 10A to 50A at different amplitudes; Operating conditions with different step changes and different slope changes in reference current; Disturbance conditions where load parameters suddenly increase or decrease; Operating conditions including grid voltage fluctuations and grid voltage within a preset zero-crossing voltage range; The simulation model of the three-phase four-arm grid-connected inverter transitions from a steady state to a transient operating condition during the dynamic adjustment process.

[0008] Optionally, the simulation operation status data of the three-phase four-arm grid-connected inverter simulation model in each control cycle includes: Reference value for phase load current tracking , Reference value for phase load current tracking , Reference value for phase load current tracking , Phase grid side inductor current Phase grid side inductor current , Phase grid side inductor current , Phase-connected grid current , Phase-connected grid current , Phase-connected grid current , Phase grid side voltage , Phase grid side voltage , Phase grid side voltage , Phase load side back electromotive force , Phase load side back electromotive force , Phase load side back electromotive force , Phase filter capacitor voltage , Phase filter capacitor voltage , Phase filter capacitor voltage ,in, This indicates the current control cycle.

[0009] Optionally, when processing the original simulation data, the original state characteristics are the simulation operation state data of the three-phase four-arm grid-connected inverter simulation model in the current control cycle. When processing the actual operation state data, the original state characteristics are the real operation state data of the three-phase four-arm grid-connected inverter in the current control cycle, which is collected in real time. The characteristics derived from operating condition coupling include state increment characteristics, error coupling characteristics, switching cost characteristics, and resonance sensitivity characteristics; The state increment characteristics include at least one of the grid-connected current increment, grid-side inductor current increment, and filter capacitor voltage increment. Error coupling characteristics include at least one of the following: current tracking error, rate of change of current tracking error, coupling term between current tracking error and filter capacitor voltage, and coupling term between current tracking error and grid-side inductor current. When processing the simulation raw data, the switching cost feature is constructed based on the simulation optimal on / off state combination of the first control cycle before the current control cycle and several historical control cycles. When processing the actual operating state data, the switching cost feature is constructed based on the on / off state combination actually executed in the first control cycle before the current control cycle and the on / off state combination of the four insulated gate bipolar transistors of the upper bridge arm actually executed in several historical control cycles. The switching cost feature includes the number of switch position flips and the bridge arm switching intensity in the first control cycle before the current control cycle, as well as the cumulative switch flip trend in several historical control cycles. The resonance sensitivity feature is a resonance sensitivity index constructed using the filter capacitor voltage, grid-side inductor current, and grid-connected current. The resonance sensitivity feature is used to characterize the degree to which the LCL filter network of a three-phase four-bridge-arm grid-connected inverter is in the oscillation risk range. During the training and online control of the deep neural network model, when inputting data into the deep neural network model, the event labels of each control cycle within the time window are encoded to form event label codes. The original state features, operating condition coupled derived features, and event label codes are arranged in the time dimension according to the time window to form an event-driven composite feature tensor, which is then input into the deep neural network model.

[0010] Optionally, a steady-state event is an event in which, within a consecutive preset number of control cycles, the change in reference current, current tracking error, state increment characteristics, and resonance sensitivity characteristics are all within the corresponding preset ranges, and the number of switch flips per unit time is within the preset normal switching range. Abrupt operating condition events include reference current abrupt change events, current tracking error exceeding threshold events, state increment characteristic exceeding threshold events, switch flipping events, phase switching events, and resonance sensitive events; When the change in reference current between adjacent control cycles exceeds a preset threshold, it is considered a reference current abrupt change event. When the absolute value of the current tracking error between the grid-connected current and the reference current exceeds a preset threshold, it is a current tracking error exceeding threshold event. When the rate of change of the filter capacitor voltage, grid-side inductor current, or grid-connected current exceeds the corresponding preset threshold in adjacent sampling periods, it is a state increment feature threshold over-threshold event. When constructing the training dataset, based on the difference between the simulated optimal on / off state combination of the first control cycle before the current control cycle and the simulated optimal on / off state combination of the second control cycle before the current control cycle, it is determined whether at least one switch position has flipped in the first control cycle before the current control cycle compared to the second control cycle before the current control cycle. If at least one switch position has flipped in the first control cycle before the current control cycle, and the number of switch flips per unit time exceeds the preset normal switching range, then the event sample of the first control cycle before the current control cycle is determined as a switch flipping event, and a switch flipping event is added to the event sample of the current control cycle. During online control, after the final complete on / off state combination is output in the current control cycle, if the final complete on / off state combination output in the current control cycle has at least one switch position flipped compared to the first control cycle before the current control cycle, and the number of switch flips per unit time exceeds the preset normal switching range, a switch flipping event is added to the event sample of the current control cycle. When processing the data of the current control cycle, if the event sample of the first control cycle before the current control cycle is a switch flipping event, then a switch flipping event is added to the event sample of the current control cycle. A phase switching event occurs when the grid voltage phase is within a preset phase neighborhood or when the instantaneous grid voltage value is within a preset zero-crossing voltage range. When the resonance sensitivity index exceeds the preset threshold, it is determined to be a resonance sensitivity event.

[0011] Optionally, the timing window for steady-state events is a fixed-length timing window, including the current control period and the periods preceding the current control period. One control cycle, The value of is an integer between 1 and 10; The timing window for a sudden change in operating condition characteristics is a timing window ending with the current control cycle. The timing window for a sudden change in operating condition characteristics includes the current control cycle and the periods preceding the current control cycle. One control cycle; in, Set the event type to be the same or different. It is a positive integer; For reference current abrupt events The range is 3 to 10; For switch toggling events The range is 10 to 30; For current tracking error exceeding the threshold event and phase switching event The range is 5 to 20; For state increment characteristic exceeding threshold events and resonance sensitive events. The range is 15 to 50.

[0012] Optional deep neural network models include: The input layer is used to receive event-driven composite feature tensors composed of data from each control cycle within a defined time window after data processing. The constraint-aware local feature extraction layer is used to perform local temporal feature extraction and constraint-aware weighted processing on the event-driven composite feature tensor received by the input layer to obtain a constraint-aware local feature sequence. The temporal dependency modeling layer is used to perform temporal dependency modeling on the constraint-aware local feature sequence output by the constraint-aware local feature extraction layer to obtain the temporal latent feature sequence. Switch the sensitive attention layer to perform time-biased attention weighting on the temporal latent feature sequence output by the temporal dependency modeling layer to obtain global control features; The output layer is used to map the global control characteristics of the switching sensitive attention layer output to obtain the initial turn-on probability of the four insulated gate bipolar transistors in the upper arm of the three-phase four-arm grid-connected inverter. The constraint consistency processing module is used to discretize and correct the initial conduction probability of the output layer according to the preset correction rules, output the final on / off state combination of the four insulated gate bipolar transistors of the upper arm in the three-phase four-bridge-arm grid-connected inverter, and generate the on / off state combination of the four insulated gate bipolar transistors of the lower arm in the three-phase four-bridge-arm grid-connected inverter according to complementary logic, so as to obtain the final complete on / off state combination of the eight insulated gate bipolar transistors in the three-phase four-bridge-arm grid-connected inverter.

[0013] Optionally, the constraint-aware local feature extraction layer includes: A one-dimensional convolutional sublayer is used to perform convolution operations on the data received from the input layer along the time dimension to extract local dynamic features; The batch normalization sublayer is used to batch normalize the local dynamic features output by the one-dimensional convolutional sublayer to obtain scale-normalized local dynamic features. The nonlinear activation sublayer is used to perform nonlinear mapping on the scale-normalized local dynamic features output by the batch normalization sublayer to obtain the activated local dynamic features. The gated weighted sublayer is used to generate gate weights based on the constraint description vector, and apply the gate weights to the activated local dynamic features output by the nonlinear activation sublayer to obtain the constraint-aware local feature sequence. The constraint description vector includes current tracking error, rate of change of current tracking error, switch flipping risk, bridge arm complementary constraint risk, and resonance sensitivity index.

[0014] Optionally, the switching sensitive attention layer includes a query vector generation sub-layer, a key vector generation sub-layer, a value vector generation sub-layer, a time bias generation sub-layer, and an attention weighting sub-layer; The query vector generation sublayer, key vector generation sublayer, and value vector generation sublayer generate query vectors based on the temporal latent feature sequences output by the temporal dependency modeling layer, respectively. Key vector Sum value vector ; The time bias generation sublayer is used to generate a time bias matrix based on at least one of the following: the switching risk of the time step within the timing window, the rate of change of the current tracking error, the resonance sensitivity index, and the distance between the time step and the center of the abrupt change event of the operating condition characteristics in the historical control cycle. ; Attention-weighted sublayer based on query vector Key vector Value vector and time bias matrix Calculate the attention weight matrix And based on the attention weight matrix value vector Weighting is performed to obtain the global control characteristics; Attention weight matrix Calculated using the following formula: ; Global control features It can be obtained through the following formula: ; in, Represents the normalization function. This represents the matrix transpose operation. Indicates the current control cycle. The feature dimensions represent the query vector and key vector. This indicates pooling, flattening, or feature fusion processing.

[0015] Optionally, the initial conduction probability output by the output layer. ; in, Indicates the current number In the three-phase four-arm grid-connected inverter of the first control cycle, the first... The initial turn-on probability of each upper bridge arm insulated gate bipolar transistor. , This represents the matrix transpose operation; it determines the candidate on / off state combinations of the four insulated-gate bipolar transistors in the upper arm of a three-phase four-arm grid-connected inverter based on the initial on-probability. ; The constraint consistency processing module determines the final on / off state combination for the current control cycle according to the following correction rules: ; in, Indicates the current number The final on / off state combination of each control cycle, for Any candidate on / off state combination, which consists of four binary switch states. This indicates the final on / off state combination actually executed in the first control cycle preceding the current control cycle. This indicates the final on / off state combination actually executed in the second control cycle preceding the current control cycle. Indicates the resonance sensitivity index, , , These are the weighting coefficients. Describes the norm 1. Represents the L2 norm; The on / off state combinations of the four insulated-gate bipolar transistors in the lower arm of a three-phase four-arm grid-connected inverter. Generated according to the following complementary logic: ; The final complete on / off state combination of the eight insulated gate bipolar transistors in the three-phase four-bridge-arm grid-connected inverter is obtained.

[0016] In summary, this application includes the following beneficial technical effects: This application transforms the optimization problem solved online in traditional MPC into an offline training and online forward inference mode, significantly reducing the real-time computational burden on the controller and meeting the rapid response requirements in power electronic conversion scenarios. The deep neural network model has the ability to perform time-series modeling and nonlinear mapping, reducing the real-time pressure caused by enumerating multiple switching states one by one and calculating the cost function online in traditional model predictive control algorithms. Compared with ordinary neural networks, the deep neural network model of this application has outstanding advantages in nonlinear mapping and time-series feature extraction due to its deeper structure. It can effectively capture long-term dependencies, further improve prediction accuracy, and make the controller more generalizable and robust under unknown operating conditions and disturbances, ultimately achieving a more stable and accurate current tracking effect. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a multi-objective optimization control method for grid-connected converters based on deep neural networks; Figure 2This is a block diagram illustrating the principle of a multi-objective optimization control method for grid-connected converters based on deep neural networks. Detailed Implementation

[0019] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0020] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0022] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The illustrations only show the components related to this application and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0023] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0024] This application provides a multi-objective optimization control method for grid-connected converters based on deep neural networks.

[0025] like Figure 1 and Figure 2As shown, a multi-objective optimization control method for grid-connected converters based on deep neural networks includes: Step 1: Obtain simulation operation status data of the three-phase four-arm grid-connected inverter simulation model: Build a three-phase four-arm grid-connected inverter simulation model, conduct simulation experiments on the three-phase four-arm grid-connected inverter simulation model under the model predictive control algorithm, obtain the simulation operation status data of the three-phase four-arm grid-connected inverter simulation model in each control cycle under multiple operating conditions, and the simulation optimal on / off state combination of the four insulated gate bipolar transistors of the upper arm in the three-phase four-arm grid-connected inverter simulation model. The three-phase four-arm grid-connected inverter simulation model includes a switching module composed of eight insulated gate bipolar transistors, an LCL filter circuit, and a signal monitoring module. Step 2, Constructing the Training Dataset: The simulation operation state data of the three-phase four-arm grid-connected inverter simulation model in each control cycle is used as a set of simulation raw data. Each set of simulation raw data is used as initial data. Data processing is performed on each set of initial data, and the processed data is paired one by one with the simulation optimal on / off state combination. The data processing method includes extracting original state features and operating condition coupling derived features from the initial data, normalizing the original state features and operating condition coupling derived features to obtain feature samples of the corresponding initial data; attaching corresponding event labels to the event samples in the initial data according to predefined event classification rules. Event classification includes steady-state events and operating condition feature mutation events; constructing a steady-state sliding time series window for steady-state events, and constructing a time series window including the occurrence time of the operating condition feature mutation event for operating condition feature mutation events. Step 3: Construct a deep neural network model for multi-objective optimization control of grid-connected converters; Step 4: Use the training dataset to train the deep neural network model to obtain the trained deep neural network model. During the training process, the training dataset is divided into training set, validation set and test set according to the ratio of 70%, 15% and 15% respectively. Step 5, Online Control: Real-time acquisition of actual operating status data of the three-phase four-arm grid-connected inverter. Using the actual operating status data of the current control cycle as initial data, data processing is performed on the actual operating status data. The processed data of the current control cycle, along with the processed data of all control cycles within the timing window, are input into the trained deep neural network model. The trained deep neural network model outputs the final complete on / off state combination of the eight insulated-gate bipolar transistors (IGBTs) in the three-phase four-arm grid-connected inverter. By learning the mapping relationship between the system operating state and the optimal switching state output by the model predictive control algorithm, the deep neural network model can directly predict the switching state of the eight IGBTs corresponding to the current control cycle without performing the enumeration optimization process in the model predictive control algorithm during the online control stage.

[0026] Specifically: In step 1, the simulation operation status data of the three-phase four-arm grid-connected inverter simulation model in each control cycle under multiple operating conditions are obtained. The operating conditions in the simulation optimal on / off state combination process of the four insulated gate bipolar transistors of the upper arm in the three-phase four-arm grid-connected inverter simulation model include: operating conditions with reference current varying from 10A to 50A with different amplitudes; operating conditions with reference current varying with different step changes and different slopes; disturbance conditions with sudden increases or decreases in load parameters; operating conditions with grid voltage fluctuations and grid voltage within the preset zero-crossing voltage range; and transient operating conditions where the three-phase four-arm grid-connected inverter simulation model transitions from steady state to dynamic adjustment.

[0027] The simulation operation status data of the three-phase four-arm grid-connected inverter simulation model in each control cycle in step 2 includes: Reference value for phase load current tracking , Reference value for phase load current tracking , Reference value for phase load current tracking , Phase grid side inductor current Phase grid side inductor current , Phase grid side inductor current , Phase-connected grid current , Phase-connected grid current , Phase-connected grid current , Phase grid side voltage , Phase grid side voltage , Phase grid side voltage , Phase load side back electromotive force , Phase load side back electromotive force , Phase load side back electromotive force , Phase filter capacitor voltage , Phase filter capacitor voltage , Phase filter capacitor voltage ,in, This indicates the current control cycle.

[0028] In step 2, when processing the original simulation data: The original state characteristics are the simulation operation state data of the three-phase four-arm grid-connected inverter simulation model within the current control cycle; the operating condition coupling derived characteristics include state increment characteristics, error coupling characteristics, switching cost characteristics, and resonance sensitivity characteristics.

[0029] The state increment characteristics include at least one of the following: grid-connected current increment, grid-side inductor current increment, and filter capacitor voltage increment. The error coupling characteristics include at least one of the following: current tracking error, rate of change of current tracking error, coupling term between current tracking error and filter capacitor voltage, and coupling term between current tracking error and grid-side inductor current. The switching cost characteristics are constructed based on the simulated optimal on / off state combinations of the first control cycle before the current control cycle and several historical control cycles. These switching cost characteristics include the number of switch position flips and the bridge arm switching intensity in the first control cycle before the current control cycle, as well as the cumulative switch flipping trend in several historical control cycles. The resonance sensitivity characteristics are resonance sensitivity indices constructed using filter capacitor voltage, grid-side inductor current, and grid-connected current. These characteristics characterize the degree to which the LCL filter network of a three-phase four-bridge-arm grid-connected inverter is in the oscillation risk range. LCL stands for Inductor-Capacitor-Inductor, which is an inductor-capacitor-inductor type filter circuit. A steady-state event is defined as an event in which, within a preset number of consecutive control cycles, the reference current change, current tracking error, state increment characteristics, and resonance sensitivity characteristics are all within their respective preset ranges, and the number of switch flips per unit time is within a preset normal switching range. The timing window for a steady-state event is a fixed-length window, including the current control cycle and the periods preceding it. One control cycle, The value of is an integer between 1 and 10; Operating condition characteristic abrupt events include reference current abrupt events, current tracking error exceeding threshold events, state increment characteristic exceeding threshold events, switch reversal events, phase switching events, and resonance sensitive events. A reference current abrupt event occurs when the change in reference current exceeds a preset threshold in an adjacent control cycle. A current tracking error exceeding threshold event occurs when the absolute value of the current tracking error between the grid current and the reference current exceeds a preset threshold. A state increment characteristic exceeding threshold event occurs when the rate of change of the filter capacitor voltage, grid-side inductor current, or grid current exceeds the corresponding preset threshold in an adjacent sampling cycle. The optimal on / off state combination is determined based on the simulation of the first control cycle preceding the current control cycle and the current... The difference between the simulated optimal on / off state combinations of the second control cycle before the current control cycle is used to determine whether at least one switch position flipped in the first control cycle before the current control cycle. If at least one switch position flipped in the first control cycle before the current control cycle, and the number of switch flips per unit time exceeds the preset normal switching range, then the event sample of the first control cycle before the current control cycle is determined as a switch flipping event, and a switch flipping event is added to the event sample of the current control cycle. When the grid voltage phase is in the preset phase neighborhood or the instantaneous value of the grid voltage is in the preset zero-crossing voltage range, it is a phase switching event. When the resonance sensitivity index exceeds the preset threshold, it is determined as a resonance sensitivity event.

[0030] The timing window for a sudden change in operating condition characteristics is a timing window ending with the current control cycle. The timing window for a sudden change in operating condition characteristics includes the current control cycle and the periods preceding the current control cycle. One control cycle; Set the event type to be the same or different. It is a positive integer; for a sudden change in the reference current, The value is 3 to 10; for switch toggle events, The value is 10 to 30; for current tracking error exceeding the threshold event and phase switching event, The value is 5 to 20; for state increment characteristic over-threshold events and resonance-sensitive events, The range is 15 to 50.

[0031] In step 5, the data processing of the actual operating status data differs from that in step 2 in the following ways: The original state characteristics are the real operating state data of the three-phase four-arm grid-connected inverter in the current control cycle, which are collected in real time. The switching cost characteristics are constructed based on the on / off state combinations actually executed in the first control cycle before the current control cycle and the on / off state combinations of the four insulated-gate bipolar transistors of the upper arm actually executed in several historical control cycles. During the constraint consistency processing of the deep neural network model, the switching cost is further calculated based on the difference between the candidate on / off state combinations of the four insulated-gate bipolar transistors of the upper arm in the three-phase four-arm grid-connected inverter and the historical on / off state combinations.

[0032] After the current control cycle outputs the final complete on / off state combination, if the final complete on / off state combination output by the current control cycle has at least one switch position flipped compared to the first control cycle, and the number of switch flips per unit time exceeds the preset normal switching range, then a switch flipping event is added to the event sample of the current control cycle. When processing data for the current control cycle, since the deep neural network model has not yet output the final complete on / off state combination of the eight insulated gate bipolar transistors in the three-phase four-arm grid-connected inverter of the current control cycle, if the event sample of the first control cycle is a switch flipping event, then a switch flipping event is added to the event sample of the current control cycle.

[0033] The timing window for sudden changes in operating condition characteristics includes the current control cycle and the periods preceding the current control cycle. One control cycle is used to avoid using future sampling data.

[0034] When the number of control cycles within different time windows differs, or when the number of historical control cycles in the initial control phase is insufficient, the time windows are standardized according to a preset input length. If the number of control cycles within a time window is less than the preset input length, at least one of the following methods is applied to the missing historical control cycle positions: zero-padding, boundary value padding, or repeated padding, with a mask identifier set for the padding positions. If the number of control cycles within a time window is greater than the preset input length, a preset number of control cycles are truncated in order of increasing distance from the current control cycle. After standardization, an event-driven composite feature tensor with a preset unified time dimension is formed.

[0035] In steps 4 and 5, during the training and online control of the deep neural network model, when inputting data into the model, a time-series window is constructed based on the event labels of the data in the current control cycle. The event labels of each control cycle within the time-series window are then encoded to form event label codes. The original state features, operating condition coupling-derived features, and event label codes from each control cycle within the time-series window are arranged along the time dimension according to the time-series window, forming an event-driven composite feature tensor, which is then input into the deep neural network model. The input to the deep neural network model is no longer limited to general time-series state quantities but can explicitly include control semantic information related to control errors, switching costs, resonance risks, and event categories.

[0036] This application enhances the learning ability of deep neural network models to understand transient processes and control key points by identifying key events, thereby focusing subsequent data construction and deep neural network model training on critical periods that affect switching decisions. By constructing error coupling features, switching cost features, and resonance sensitivity features, it enhances the ability of deep neural network models to express control semantics and dynamic risks. Through the event-driven historical time-series window construction method, it can retain local dynamic information related to control decisions more accurately than ordinary multi-time-time splicing.

[0037] The deep neural network model in this application includes: The input layer is used to receive event-driven composite feature tensors composed of data from each control cycle within a defined time window after data processing. The constraint-aware local feature extraction layer is used to perform local temporal feature extraction and constraint-aware weighted processing on the event-driven composite feature tensor received by the input layer to obtain a constraint-aware local feature sequence. The temporal dependency modeling layer is used to perform temporal dependency modeling on the constraint-aware local feature sequence output by the constraint-aware local feature extraction layer to obtain the temporal latent feature sequence. Switch the sensitive attention layer to perform time-biased attention weighting on the temporal latent feature sequence output by the temporal dependency modeling layer to obtain global control features; The output layer is used to map the global control characteristics of the switching sensitive attention layer output to obtain the initial turn-on probability of the four insulated gate bipolar transistors in the upper arm of the three-phase four-arm grid-connected inverter. The constraint consistency processing module is used to discretize and correct the initial conduction probability of the output layer according to the preset correction rules, output the final on / off state combination of the four insulated gate bipolar transistors of the upper arm in the three-phase four-bridge-arm grid-connected inverter, and generate the on / off state combination of the four insulated gate bipolar transistors of the lower arm in the three-phase four-bridge-arm grid-connected inverter according to complementary logic, so as to obtain the final complete on / off state combination of the eight insulated gate bipolar transistors in the three-phase four-bridge-arm grid-connected inverter.

[0038] For specific deep neural network models: The constraint-aware local feature extraction layer includes: A one-dimensional convolutional sublayer is used to perform convolution operations on the data received from the input layer along the time dimension to extract local dynamic features; The batch normalization sublayer is used to batch normalize the local dynamic features output by the one-dimensional convolutional sublayer to obtain scale-normalized local dynamic features. The nonlinear activation sublayer is used to perform nonlinear mapping on the scale-normalized local dynamic features output by the batch normalization sublayer to obtain the activated local dynamic features. The gated weighted sublayer is used to generate gate weights based on the constraint description vector, and apply the gate weights to the activated local dynamic features output by the nonlinear activation sublayer to obtain the constraint-aware local feature sequence. The constraint description vector includes current tracking error, rate of change of current tracking error, switch flipping risk, bridge arm complementary constraint risk, and resonance sensitivity index.

[0039] The switching-sensitive attention layer includes a query vector generation sublayer, a key vector generation sublayer, a value vector generation sublayer, a temporal bias generation sublayer, and an attention weighting sublayer. The query vector generation sublayer, key vector generation sublayer, and value vector generation sublayer generate query vectors based on the temporal latent feature sequence output by the temporal dependency modeling layer, respectively. Key vector Sum value vector ; The time bias generation sublayer is used to generate a time bias matrix based on at least one of the following: the switching risk of the time step within the timing window, the rate of change of the current tracking error, the resonance sensitivity index, and the distance between the time step and the center of the abrupt change event of the operating condition characteristics in the historical control cycle. Wherein, the center of the sudden change event of the operating condition is the control cycle that is determined to be a sudden change event of the operating condition within the time window, and the distance is the number of control cycle intervals between each time step within the time window and the control cycle. Attention-weighted sublayer based on query vector Key vector Value vector and time bias matrix Calculate the attention weight matrix And based on the attention weight matrix value vector Weighting is performed to obtain the global control characteristics; Attention weight matrix Calculated using the following formula: ; Global control features It can be obtained through the following formula: ; in, Represents the normalization function. This represents the matrix transpose operation. Indicates the current control cycle. The feature dimensions represent the query vector and key vector. This indicates pooling, flattening, or feature fusion processing.

[0040] Initial conduction probability of the output layer ; in, Indicates the current number In the three-phase four-arm grid-connected inverter of the first control cycle, the first... The initial turn-on probability of each upper bridge arm insulated gate bipolar transistor. , This represents the matrix transpose operation; it determines the candidate on / off state combinations of the four insulated-gate bipolar transistors in the upper arm of a three-phase four-arm grid-connected inverter based on the initial on-probability. Specifically, threshold judgments are performed on the four insulated-gate bipolar transistors (IGBTs) of the upper arm in a three-phase four-arm grid-connected inverter. When the initial turn-on probability is greater than a first preset probability threshold, the corresponding IGBT candidate state is determined to be on. When the initial turn-on probability is less than a second preset probability threshold, the corresponding IGBT candidate state is determined to be off. When the initial turn-on probability is between the second and first preset probability thresholds, both the on and off states are used as candidate states for the corresponding IGBTs. The candidate states of the four upper arm IGBTs are arranged and combined, and invalid combinations that do not meet the preset hardware and safety constraints are eliminated to obtain a set of candidate on / off state combinations. .

[0041] The constraint consistency processing module determines the final on / off state combination for the current control cycle according to the following correction rules: ; in, Indicates the current number The final on / off state combination of each control cycle, for Any candidate on / off state combination, which consists of four binary switch states. This indicates the final on / off state combination actually executed in the first control cycle preceding the current control cycle. This indicates the final on / off state combination actually executed in the second control cycle preceding the current control cycle. Indicates the resonance sensitivity index, , , These are the weighting coefficients. Describes the norm 1. Represents the L2 norm; The on / off state combinations of the four insulated-gate bipolar transistors in the lower arm of a three-phase four-arm grid-connected inverter. Generated according to the following complementary logic: ; The final complete on / off state combination of the eight insulated gate bipolar transistors in the three-phase four-bridge-arm grid-connected inverter is obtained.

[0042] The deep neural network model of this application improves the ability of the deep neural network model to extract key local patterns and key time steps by using a constraint-aware local feature extraction layer and a switching sensitive attention layer; and through a constraint consistency processing module, the output results of the deep neural network model are made to better conform to the bridge arm complementary logic and actual engineering control requirements, and reduce invalid high-frequency flips.

[0043] Grid-connected inverters exhibit the following characteristics that distinguish them from ordinary timing prediction tasks during operation: the controlled object has obvious discrete switching characteristics, and the optimal control result is not a continuous quantity, but rather a combination of the on / off states of an insulated-gate bipolar transistor; the system experiences significant critical event moments under conditions such as sudden changes in reference current, load disturbances, and grid-side voltage variations, and the state evolution before and after these moments has a significant impact on switching decisions; grid-connected inverters typically contain LCL filter networks, and the coupling between state variables is strong, which may exhibit dynamic characteristics such as rapid changes, resonance sensitivity, and sudden error changes in local time periods; the control output must not only meet the current tracking accuracy requirements, but also take into account the switching cost, bridge arm complementary constraints, and dynamic stability. Therefore, the network learning objective is not ordinary classification, but a decision mapping that satisfies physical control constraints.

[0044] This application constructs an event-driven state reconstruction data processing method, which introduces key control events of grid-connected inverters into the dataset organization and feature construction process; it also constructs a constraint-aware hybrid deep neural network structure, which introduces an adaptive modulation mechanism related to switching risk, error evolution and bridge arm constraints in the local feature extraction layer and attention layer.

[0045] Compared to a conventional feedforward neural network that does not specifically introduce temporal modeling, event-driven composite feature tensors, and control constraint processing mechanisms, and mainly relies on direct mapping based on the current control cycle state, this application explicitly introduces historical state evolution, key events, switching risks, and execution constraints related to control decisions into the entire decision-making chain. The difference between the two is not only the network size, but also the different ways of organizing input information and the different decision-making mechanisms.

[0046] At the input end, this application encodes the original state features, operating condition coupling-derived features, and event labels of the current control cycle and historical control cycles, arranging them along the time dimension according to a time-series window to form an event-driven composite feature tensor. The operating condition coupling-derived features include state increment features, error coupling features, switching cost features, and resonance sensitivity features. Event labels are used to distinguish between steady-state events and sudden changes in operating condition features such as reference current mutations, current tracking error exceeding thresholds, state increment feature exceeding thresholds, switch flipping, phase switching, and resonance sensitivity. Therefore, the deep neural network model can not only obtain the instantaneous states such as current and voltage, but also information on state changes, error evolution, recent switching conditions, and current dynamic events. The constraint-aware local feature extraction layer further extracts local dynamic features along the time dimension and generates gating weights based on current tracking error and its rate of change, switch reversal risk, bridge arm complementary constraint risk, and resonance sensitivity index. The time-dependent modeling layer preserves the state evolution relationship across control cycles. The switching-sensitive attention layer generates a time bias using at least one of the following: switch reversal risk, current tracking error rate of change, resonance sensitivity index, and distance between the time step and the center of abrupt change events in historical operating conditions. This differential weighting applies to different historical time steps. This structural design enables the deep neural network model to select local and historical information more relevant to the current switching decision based on the current control risk and key events, rather than relying solely on the state at a single moment for mapping.

[0047] At the output end, this application does not directly use the initial on-probability of the four insulated-gate bipolar transistors (IGBTs) of the upper bridge arm as the final switching command. When the initial on-probability is between the second preset probability threshold and the first preset probability threshold, both on and off candidate states are retained. Subsequently, the candidate states of the four IGBTs of the upper bridge arm are arranged and combined, and invalid combinations that do not meet the preset hardware constraints and safety constraints are eliminated. Then, the constraint consistency processing module determines the final on-off state combination of the current control cycle by combining the final on-off state of the historical actual execution, the switching-related cost, and the resonance sensitivity index, and generates the on-off state of the lower bridge arm according to complementary logic. Therefore, there is still candidate state screening and constraint consistency processing between the output probability of the deep neural network model and the final execution command, and the final decision is not entirely determined by the result of a single probability classification.

[0048] Therefore, the advantages of the aforementioned ordinary feedforward neural network mainly stem from its structure, which matches the control characteristics of grid-connected converters: event-driven composite feature input, local dynamics and constraint-aware feature extraction, cross-cycle temporal dependency modeling, switching-sensitive attention weighting, and candidate state and constraint consistency processing. This structure, mechanistically speaking, has the foundation to more fully express nonlinear dynamics, historical dependencies, key events, and discrete switching constraint relationships, making it more targeted than networks that simply classify switches based on their current state.

[0049] In terms of generalization, the training data in this application is generated by MPC under multiple operating conditions, including reference current variations from 10A to 50A with different amplitudes, different step changes and slopes, sudden increases or decreases in load parameters, grid voltage fluctuations and zero-crossing intervals, and transient processes transitioning from steady state to dynamic adjustment. The corresponding simulated optimal on / off state combinations serve as the supervision target. Therefore, the deep neural network model training is not limited to a single steady-state operating point. Simultaneously, features such as state increment, error rate of change, historical switch-flipping trends, and resonance sensitivity indicators describe the system's relationship to changes and risk information relative to the current target and historical states. Event labels and event-related time-series windows organize processes with similar control significance under different operating conditions, such as steady-state, abrupt changes, error exceeding thresholds, switch flipping, and resonance sensitivity, according to unified rules. This reduces the network's tendency to only memorize instantaneous values ​​at a single absolute operating point, providing a structural foundation for learning reusable state evolution-switching decision relationships across different operating conditions.

[0050] The ability of a deep neural network model to function under unknown operating conditions does not mean that the network can infinitely extrapolate to any unseen state. Rather, for specific samples or combinations of operating conditions that did not appear one by one during training, as long as their state changes, error evolution, event types, and constraint relationships are still within the scope covered by existing multi-operating-condition training data and can be expressed by the existing features of the original application, the deep neural network model has the basis for rapid inference using the learned control mapping. For example, even if a specific combination of reference current, load change, and grid state does not appear exactly as it was in the training set, the resulting state increment, error change, switching risk, and resonance risk are still within the range of patterns covered by existing training samples, thus enabling this application to be applied to cross-operating-condition generalization scenarios.

[0051] In terms of robustness, this application reduces the dependence of control decisions on individual instantaneous states through historical timing windows and timing dependency modeling. During reference current surges, load disturbances, or grid voltage changes, even if the instantaneous states of a certain control cycle are similar, they may have previously been in different evolutionary stages such as rapid error increase, error decline, or just a switch flip. This application can distinguish these dynamic backgrounds by combining state increments, error change rates, event labels, and historical switching information, and adjusts feature contributions based on current risk through gating weighting and switching-sensitive attention. On the other hand, when the initial turn-on probability is between two preset probability thresholds, this application retains both turn-on and turn-off candidate states, and then determines the final state through hardware constraints, safety constraints, and constraint consistency processing. Therefore, local output changes within the intermediate probability interval are not necessarily directly equivalent to an immediate switch flip; the final state is further constrained by the candidate set and historical control constraints, reducing ineffective high-frequency flips. These two aspects together constitute the structural robustness basis of this application in the face of operating condition disturbances and local network output fluctuations.

[0052] Compared to ordinary feedforward neural networks, this application, within the scope of operating conditions and preset constraints covered by the training data, can more fully utilize historical information across operating conditions, key transients, and control risks, and further constrain the network's probability output. Therefore, it possesses a more explicit mechanism for cross-operating condition feature reuse and disturbance adaptation. Compared to traditional finite control set MPC, the main advantage of this application lies in its online computation path. Traditional MPC requires predicting the future behavior of feasible switch states in each control cycle and evaluating them one by one using a cost function. This application, however, uses the simulation-optimal on / off state combination obtained from MPC as offline training supervision information. In the online phase, it mainly performs feature construction, network forward inference, and constraint consistency processing of candidate states, eliminating the need to perform MPC-style forward prediction and control cost evaluation for all feasible switch states, thus reducing the online computational burden. It should be noted that MPC relies on the explicit model to be re-solved in the current state; its adaptation to unknown states is not training generalization in the sense of a neural network. Therefore, it is not stated that the generalization ability of this application under arbitrary unknown operating conditions is necessarily superior to MPC.

[0053] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-objective optimization control method for grid-connected converters based on deep neural networks, characterized in that, Includes the following steps: A simulation model of a three-phase four-arm grid-connected inverter was built. Simulation experiments were conducted on the simulation model of the three-phase four-arm grid-connected inverter under the model predictive control algorithm. Simulation operation status data of the three-phase four-arm grid-connected inverter simulation model in each control cycle under multiple operating conditions were obtained, as well as the simulation optimal on / off state combination of the four insulated gate bipolar transistors of the upper arm in the simulation model of the three-phase four-arm grid-connected inverter. Constructing a training dataset: The simulation operation state data of the three-phase four-arm grid-connected inverter simulation model in each control cycle is used as a set of simulation raw data. Each set of simulation raw data is used as initial data. Data processing is performed on each set of initial data, and the processed data is paired one by one with the simulation optimal on / off state combination. The data processing methods include extracting original state features and operating condition coupling derived features from the initial data, normalizing the original state features and operating condition coupling derived features to obtain feature samples of the corresponding initial data; attaching corresponding event labels to the event samples in the initial data according to predefined event classification rules, with event classification including steady-state events and operating condition feature mutation events; constructing a steady-state sliding time series window for steady-state events, and constructing a time series window including the occurrence time of the operating condition feature mutation event for operating condition feature mutation events. Construct a deep neural network model for multi-objective optimization control of grid-connected converters; The deep neural network model is trained using the training dataset to obtain the trained deep neural network model; Online control: Real-time acquisition of the actual operating status data of the three-phase four-arm grid-connected inverter, using the actual operating status data of the current control cycle of the three-phase four-arm grid-connected inverter as the initial data, and performing data processing on the actual operating status data; The data processed by the current control cycle and the data processed by all control cycles within the timing window are input into the trained deep neural network model. The trained deep neural network model outputs the final complete on / off state combination of the eight insulated gate bipolar transistors in the three-phase four-bridge-arm grid-connected inverter.

2. The multi-objective optimization control method for grid-connected converters based on deep neural networks according to claim 1, characterized in that, The simulation operation status data of the three-phase four-arm grid-connected inverter simulation model in each control cycle under multiple operating conditions are obtained. The operating conditions during the simulation of the optimal on / off state combination of the four insulated-gate bipolar transistors in the upper arm of the three-phase four-arm grid-connected inverter simulation model include: Operating conditions with reference current varying from 10A to 50A at different amplitudes; Operating conditions with different step changes and different slope changes in reference current; Disturbance conditions where load parameters suddenly increase or decrease; Operating conditions including grid voltage fluctuations and grid voltage within a preset zero-crossing voltage range; The simulation model of the three-phase four-arm grid-connected inverter transitions from a steady state to a transient operating condition during the dynamic adjustment process.

3. The multi-objective optimization control method for grid-connected converters based on deep neural networks according to claim 1, characterized in that, The simulation operation status data of the three-phase four-arm grid-connected inverter simulation model in each control cycle includes: Reference value for phase load current tracking , Reference value for phase load current tracking , Reference value for phase load current tracking , Phase grid side inductor current Phase grid side inductor current , Phase grid side inductor current , Phase-connected grid current , Phase-connected grid current , Phase-connected grid current , Phase grid side voltage , Phase grid side voltage , Phase grid side voltage , Phase load side back electromotive force , Phase load side back electromotive force , Phase load side back electromotive force , Phase filter capacitor voltage , Phase filter capacitor voltage , Phase filter capacitor voltage ,in, This indicates the current control cycle.

4. The multi-objective optimization control method for grid-connected converters based on deep neural networks according to claim 1, characterized in that, When processing the original simulation data, the original state characteristics are the simulation operation state data of the three-phase four-arm grid-connected inverter simulation model in the current control cycle. When processing the actual operation state data, the original state characteristics are the real operation state data of the three-phase four-arm grid-connected inverter collected in real time in the current control cycle. The characteristics derived from operating condition coupling include state increment characteristics, error coupling characteristics, switching cost characteristics, and resonance sensitivity characteristics; The state increment characteristics include at least one of the grid-connected current increment, grid-side inductor current increment, and filter capacitor voltage increment. Error coupling characteristics include at least one of the following: current tracking error, rate of change of current tracking error, coupling term between current tracking error and filter capacitor voltage, and coupling term between current tracking error and grid-side inductor current. When processing the simulation raw data, the switching cost feature is constructed based on the simulation optimal on / off state combination of the first control cycle before the current control cycle and several historical control cycles. When processing the actual operating state data, the switching cost feature is constructed based on the on / off state combination actually executed in the first control cycle before the current control cycle and the on / off state combination of the four insulated gate bipolar transistors of the upper bridge arm actually executed in several historical control cycles. The switching cost feature includes the number of switch position flips and the bridge arm switching intensity in the first control cycle before the current control cycle, as well as the cumulative switch flip trend in several historical control cycles. The resonance sensitivity feature is a resonance sensitivity index constructed using the filter capacitor voltage, grid-side inductor current, and grid-connected current. The resonance sensitivity feature is used to characterize the degree to which the LCL filter network of a three-phase four-bridge-arm grid-connected inverter is in the oscillation risk range. During the training and online control of the deep neural network model, when inputting data into the deep neural network model, the event labels of each control cycle within the time window are encoded to form event label codes. The original state features, operating condition coupled derived features, and event label codes are arranged in the time dimension according to the time window to form an event-driven composite feature tensor, which is then input into the deep neural network model.

5. The multi-objective optimization control method for grid-connected converters based on deep neural networks according to claim 4, characterized in that, A steady-state event is an event in which, within a consecutive preset number of control cycles, the change in reference current, current tracking error, state increment characteristics, and resonance sensitivity characteristics are all within their respective preset ranges, and the number of switch flips per unit time is within a preset normal switching range. Abrupt operating condition events include reference current abrupt change events, current tracking error exceeding threshold events, state increment characteristic exceeding threshold events, switch flipping events, phase switching events, and resonance sensitive events; When the change in reference current between adjacent control cycles exceeds a preset threshold, it is considered a reference current abrupt change event. When the absolute value of the current tracking error between the grid-connected current and the reference current exceeds a preset threshold, it is a current tracking error exceeding threshold event. When the rate of change of the filter capacitor voltage, grid-side inductor current, or grid-connected current exceeds the corresponding preset threshold in adjacent sampling periods, it is a state increment feature threshold over-threshold event. When constructing the training dataset, based on the difference between the simulated optimal on / off state combination of the first control cycle before the current control cycle and the simulated optimal on / off state combination of the second control cycle before the current control cycle, it is determined whether at least one switch position has flipped in the first control cycle before the current control cycle compared to the second control cycle before the current control cycle. If at least one switch position has flipped in the first control cycle before the current control cycle, and the number of switch flips per unit time exceeds the preset normal switching range, then the event sample of the first control cycle before the current control cycle is determined as a switch flipping event, and a switch flipping event is added to the event sample of the current control cycle. During online control, after the final complete on / off state combination is output in the current control cycle, if the final complete on / off state combination output in the current control cycle has at least one switch position flipped compared to the first control cycle before the current control cycle, and the number of switch flips per unit time exceeds the preset normal switching range, a switch flipping event is added to the event sample of the current control cycle. When processing the data of the current control cycle, if the event sample of the first control cycle before the current control cycle is a switch flipping event, then a switch flipping event is added to the event sample of the current control cycle. A phase switching event occurs when the grid voltage phase is within a preset phase neighborhood or when the instantaneous grid voltage value is within a preset zero-crossing voltage range. When the resonance sensitivity index exceeds the preset threshold, it is determined to be a resonance sensitivity event.

6. The multi-objective optimization control method for grid-connected converters based on deep neural networks according to claim 5, characterized in that, The timing window for steady-state events is a fixed-length timing window, including the current control cycle and the time preceding the current control cycle. One control cycle, The value of is an integer between 1 and 10; The timing window for a sudden change in operating condition characteristics is a timing window ending with the current control cycle. The timing window for a sudden change in operating condition characteristics includes the current control cycle and the periods preceding the current control cycle. One control cycle; in, Set the event type to be the same or different. It is a positive integer; For reference current abrupt events The range is 3 to 10; For switch toggling events The range is 10 to 30; For current tracking error exceeding the threshold event and phase switching event The range is 5 to 20; For state increment characteristic exceeding threshold events and resonance sensitive events. The range is 15 to 50.

7. The multi-objective optimization control method for grid-connected converters based on deep neural networks according to claim 1, characterized in that, Deep neural network models include: The input layer is used to receive event-driven composite feature tensors composed of data from each control cycle within a defined time window after data processing. The constraint-aware local feature extraction layer is used to perform local temporal feature extraction and constraint-aware weighted processing on the event-driven composite feature tensor received by the input layer to obtain a constraint-aware local feature sequence. The temporal dependency modeling layer is used to perform temporal dependency modeling on the constraint-aware local feature sequence output by the constraint-aware local feature extraction layer to obtain the temporal latent feature sequence. Switch the sensitive attention layer to perform time-biased attention weighting on the temporal latent feature sequence output by the temporal dependency modeling layer to obtain global control features; The output layer is used to map the global control characteristics of the switching sensitive attention layer output to obtain the initial turn-on probability of the four insulated gate bipolar transistors in the upper arm of the three-phase four-arm grid-connected inverter. The constraint consistency processing module is used to discretize and correct the initial conduction probability of the output layer according to the preset correction rules, output the final on / off state combination of the four insulated gate bipolar transistors of the upper arm in the three-phase four-bridge-arm grid-connected inverter, and generate the on / off state combination of the four insulated gate bipolar transistors of the lower arm in the three-phase four-bridge-arm grid-connected inverter according to complementary logic, so as to obtain the final complete on / off state combination of the eight insulated gate bipolar transistors in the three-phase four-bridge-arm grid-connected inverter.

8. The multi-objective optimization control method for grid-connected converters based on deep neural networks according to claim 7, characterized in that, The constraint-aware local feature extraction layer includes: A one-dimensional convolutional sublayer is used to perform convolution operations on the data received from the input layer along the time dimension to extract local dynamic features; The batch normalization sublayer is used to batch normalize the local dynamic features output by the one-dimensional convolutional sublayer to obtain scale-normalized local dynamic features. The nonlinear activation sublayer is used to perform nonlinear mapping on the scale-normalized local dynamic features output by the batch normalization sublayer to obtain the activated local dynamic features. The gated weighted sublayer is used to generate gate weights based on the constraint description vector, and apply the gate weights to the activated local dynamic features output by the nonlinear activation sublayer to obtain the constraint-aware local feature sequence. The constraint description vector includes current tracking error, rate of change of current tracking error, switch flipping risk, bridge arm complementary constraint risk, and resonance sensitivity index.

9. The multi-objective optimization control method for grid-connected converters based on deep neural networks according to claim 7, characterized in that, The switching sensitive attention layer includes a query vector generation sublayer, a key vector generation sublayer, a value vector generation sublayer, a time bias generation sublayer, and an attention weighting sublayer; The query vector generation sublayer, key vector generation sublayer, and value vector generation sublayer generate query vectors based on the temporal latent feature sequences output by the temporal dependency modeling layer, respectively. Key vector Sum value vector ; The time bias generation sublayer is used to generate a time bias matrix based on at least one of the following: the switching risk of the time step within the timing window, the rate of change of the current tracking error, the resonance sensitivity index, and the distance between the time step and the center of the abrupt change event of the operating condition characteristics in the historical control cycle. ; Attention-weighted sublayer based on query vector Key vector Value vector and time bias matrix Calculate the attention weight matrix And based on the attention weight matrix value vector Weighting is performed to obtain the global control characteristics; Attention weight matrix Calculated using the following formula: ; Global control features It can be obtained through the following formula: ; in, Represents the normalization function. This represents the matrix transpose operation. Indicates the current control cycle. The feature dimensions represent the query vector and key vector. This indicates pooling, flattening, or feature fusion processing.

10. The multi-objective optimization control method for grid-connected converters based on deep neural networks according to claim 7, characterized in that, Initial conduction probability of the output layer ; in, Indicates the current number In the three-phase four-arm grid-connected inverter of the first control cycle, the first... The initial turn-on probability of each upper bridge arm insulated gate bipolar transistor. , This represents the matrix transpose operation; it determines the candidate on / off state combinations of the four insulated-gate bipolar transistors in the upper arm of a three-phase four-arm grid-connected inverter based on the initial on-probability. ; The constraint consistency processing module determines the final on / off state combination for the current control cycle according to the following correction rules: ; in, Indicates the current number The final on / off state combination of each control cycle, for Any candidate on / off state combination, which consists of four binary switch states. This indicates the final on / off state combination actually executed in the first control cycle preceding the current control cycle. This indicates the final on / off state combination actually executed in the second control cycle preceding the current control cycle. Indicates the resonance sensitivity index, , , These are the weighting coefficients. Describes the norm 1. Represents the L2 norm; The on / off state combinations of the four insulated-gate bipolar transistors in the lower arm of a three-phase four-arm grid-connected inverter. Generated according to the following complementary logic: ; The final complete on / off state combination of the eight insulated gate bipolar transistors in the three-phase four-bridge-arm grid-connected inverter is obtained.