A MMC multi-step model predictive control method and system based on neural network approximation

Abstract

The application discloses a MMC multi-step model predictive control method and system based on neural network approximation, and relates to the field of grid-connected power electronic converter control. The method establishes a discrete state space model of MMC, deduces a multi-step prediction matrix, constructs a cost function containing output current tracking error and bridge arm circulating current error, and solves optimal samples of three-phase upper and lower bridge arm reference voltages through offline conversion into a quadratic programming; neural network is trained with three-phase grid-side current, current reference value, grid-side voltage, bridge arm circulating current and circulating current reference value as input features and optimal bridge arm voltage as a label. In the online stage, only forward reasoning is performed, bridge arm voltage reference value is directly output, and sub-module switching signals are generated in combination with the nearest level approximation modulation, capacitor voltage sorting and safety fallback mechanism. The application reduces online calculation amount while maintaining multi-step MPC control performance, improves MMC real-time control and engineering deployment capability, and is suitable for high-voltage and large-capacity flexible equipment.

Description

Technical Field

[0001] This invention relates to the field of power electronic converter control technology, and in particular to a method and system for multi-step continuous set model predictive control, artificial neural network approximation control, bridge arm voltage reference generation and sub-module modulation drive collaborative implementation for modular multilevel converters. It can be applied to real-time control scenarios of high voltage DC transmission, new energy grid connection, flexible AC transmission and high voltage large capacity grid-connected converters. Background Technology

[0002] With the advancement of new power system construction, the proportion of new energy sources such as wind power and photovoltaics connected to the grid via flexible DC systems is continuously increasing. Modular multilevel converters (MMCs), with their advantages of modular structure, high voltage and large capacity, good output waveform quality and strong scalability, have been widely used in high-voltage DC transmission and new energy grid connection.

[0003] MMC (Multi-Module Control) consists of a large number of cascaded submodules, characterized by high state dimension, strong variable coupling, complex bridge arm energy dynamics, and the need for balanced control of submodule capacitor voltages. In actual operation, the control system not only needs to achieve rapid tracking of AC side output current, but also needs to suppress bridge arm circulating current, maintain energy balance between upper and lower bridge arms, and ensure stable submodule capacitor voltages. Therefore, it places high demands on the dynamic performance, steady-state accuracy, and real-time computing capabilities of the control algorithm.

[0004] Traditional MMC control methods often employ PI controllers, PR controllers, or cascaded control structures. While simple to implement, these methods have shortcomings in multi-objective coordination, parameter tuning, and adaptation to complex disturbances. Finite control set model predictive control can directly consider switching states, but it requires traversing a large number of switching combinations. The computational complexity increases rapidly with the number of submodules, making it unsuitable for high-voltage, high-capacity MMCs.

[0005] Continuous set model predictive control transforms discrete switching optimization into continuous control quantity optimization. The bridge arm reference voltage or duty cycle can be obtained through quadratic programming, exhibiting good output current and bridge arm circulating current control performance. However, for multi-step predictive control, the longer the prediction time domain, the larger the optimization variables and matrix size, significantly increasing the computational load of solving the quadratic programming problem online. In embedded controllers with short control cycles and limited hardware resources, this can easily lead to computation timeouts, control delays, and degraded harmonic performance.

[0006] Existing technologies typically reduce computational burden by shortening the prediction step size, simplifying the model, or using lookup tables. However, shortening the prediction time domain weakens the ability of multi-step prediction to describe future states, simplifying the model reduces control accuracy, and lookup tables consume significant storage space and are difficult to cover continuous operating conditions. Therefore, how to retain the dynamic and steady-state performance of multi-step continuous set model predictive control while avoiding the computational bottleneck caused by online quadratic programming is a pressing technical problem to be solved in the engineering application of MMC real-time control. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies, such as high computational complexity and difficulty in balancing control performance and real-time performance in multi-step continuous set model predictive control. Furthermore, existing technical solutions that use neural networks or machine learning methods to approximate model predictive control laws often focus on reducing online optimization solution time. They typically do not establish a dedicated input-output mapping relationship for the MMC arm voltage generation mechanism, nor do they fully consider safety constraints such as the feasible range of arm voltage, the number of submodules deployed, arm circulating current deviation, and the dispersion of submodule capacitor voltage during MMC operation. When the neural network input state exceeds the coverage range of the training samples, or when the output control quantity is in the model extrapolation region, problems such as arm voltage reference value exceeding limits, abnormal number of submodules deployed, decreased circulating current suppression effect, or insufficient capacitor voltage balancing capability may occur, thereby affecting the real-time control safety and engineering reliability of high-voltage, high-capacity flexible DC converter equipment. Therefore, simply replacing online MPC solving with neural networks cannot fully meet the engineering application requirements of MMC. It is still necessary to construct a dedicated neural network mapping structure for MMC arm voltage control, and introduce input / output boundary detection, arm voltage limiting, submodule deployment verification, and backup control backoff mechanisms to ensure the physical executability and operational safety of the control results while reducing the online computational burden. This paper presents a multi-step model predictive control method and system for MMC based on neural network approximation. The optimal control law of multi-step CCS-MPC is approximated through offline training of the neural network. During online operation, the neural network directly maps the output control quantity, reducing the online computational complexity to a constant level while ensuring dynamic and steady-state control performance, achieving microsecond-level real-time computation.

[0008] To achieve the above objectives, the present invention provides a multi-step model predictive control method for MMC based on neural network approximation, comprising the following steps: S1. Based on the target modular multilevel converter topology, a state-space model including AC side output current, bridge arm circulating current, grid side voltage and DC side voltage is established, and the state-space model is discretized to obtain the MMC discrete mathematical model. S2. Based on the MMC discrete mathematical model, derive the multi-step prediction state matrix, construct a cost function that includes the three-phase output current tracking error and the three-phase bridge arm circulating current tracking error, transform the cost function into the standard form of quadratic programming, solve offline to obtain the optimal control quantity sequence of the reference voltage of the three-phase upper and lower bridge arms, and take the corresponding control quantity as the training label. S3. Using the three-phase grid-side current, the three-phase grid-side current reference value, the grid-side voltage in the two-phase stationary coordinate system, the three-phase bridge arm circulating current and the three-phase bridge arm circulating current reference value as input features, and the optimal control quantity of the three-phase upper and lower bridge arm reference voltages as output labels, a multi-condition sample dataset is constructed, and an artificial neural network is trained offline to obtain a neural network controller for approximating the predictive control law of a multi-step continuous set model. S4. Connect the trained neural network controller to the MMC online control system, collect state variables and reference values ​​in real time and input them into the neural network controller, directly map and output the reference values ​​of the three-phase upper and lower bridge arm voltages for the next control cycle, and generate sub-module switching signals by combining the nearest level approximation modulation and sub-module capacitor voltage sorting strategy to drive the MMC to run.

[0009] Preferably, step S2 includes: S21. Select state variables, control variables, and output vectors to construct discrete state-space equations; S22. Iterate and recursively derive the discrete state space equation to obtain the predicted set of output quantities in the prediction time domain, and transform the single-step prediction model into a multi-step prediction matrix form. S23. Construct a cost function that includes the output current error term and the bridge arm circulating current error term. Combine the multi-step prediction matrix to organize the cost function into a standard quadratic programming form and solve to obtain the optimal control quantity sequence of the bridge arm voltage.

[0010] Preferably, the state variables include the output current of each phase on the AC side, the circulating current of each phase arm, the grid-side voltage in the two-phase stationary coordinate system, and the DC side voltage; the control variables include the upper arm voltage and the lower arm voltage of each phase; and the output vector includes the output current of each phase and the circulating current of each phase arm.

[0011] Preferably, step S2 further includes a reference value calculation step, which includes: The phase angle of the grid connection point voltage is obtained by phase-locked loop, and the three-phase stationary coordinate system is transformed into the dq rotating coordinate system by coordinate transformation. The dq axis current reference value is calculated based on the active power command and reactive power command, and then the three-phase output current reference value is obtained by inverse coordinate transformation. A two-frequency circulating current injection method combined with a bridge arm energy balancing strategy is used to generate a three-phase bridge arm circulating current reference value.

[0012] Preferably, the bridge arm energy balancing strategy includes: Calculate the input power or energy storage state of the upper and lower bridge arms of a single phase respectively, sum them to obtain the common-mode component of the bridge arm energy, and calculate the difference to obtain the differential-mode component of the bridge arm energy. The common-mode component and differential-mode component of the arm energy are compared with their corresponding reference values ​​and then input into the PI controller. The controller outputs the circulating current adjustment component, which is then superimposed on the basic circulating current reference value to obtain the final arm circulating current reference value.

[0013] Preferably, a third-order Lagrange interpolation polynomial is used to extrapolate the three-phase output current reference value and the three-phase bridge arm circulating current reference value in multiple steps to generate a reference trajectory sequence in the prediction time domain, and the reference trajectory sequence is used as the optimization target of multi-step model predictive control.

[0014] Preferably, the artificial neural network adopts a feedforward network structure consisting of an input layer, a hidden layer, and an output layer; The input layer includes 14 input features, namely, three-phase grid-side current, three-phase grid-side current reference value, grid-side voltage in two-phase stationary coordinate system, three-phase bridge arm circulating current, and three-phase bridge arm circulating current reference value; The output layer includes six output quantities, namely the three-phase upper bridge arm reference voltage and the three-phase lower bridge arm reference voltage.

[0015] Preferably, the artificial neural network includes 14 input neurons, 12 hidden layer neurons, and 6 output neurons; before training, the input features and output labels are subjected to maximum-minimum normalization; during the training process, mean squared error is used as the loss function, and gradient descent is used to update the connection weights and bias thresholds until the loss function meets the preset accuracy requirements or reaches the maximum number of iterations.

[0016] Preferably, the nearest-level approximation modulation in step S4 includes: Divide the reference values ​​of the three-phase upper and lower bridge arm voltages by the reference voltage of the submodule capacitors to obtain the number of submodules that need to be put into the corresponding bridge arm. The voltage of all submodule capacitors in the current bridge arm is sorted, and the corresponding number of submodules are selected to be put into or cut off according to the direction of the bridge arm current, so as to achieve the balance between the bridge arm voltage output and the submodule capacitor voltage.

[0017] Preferably, boundary detection is performed on the neural network input, neural network output, and modulation results during online operation; When the input state quantity exceeds the coverage of the training samples, the output bridge arm voltage reference value exceeds the allowable range, the number of sub-modules to be put into the bridge arm exceeds the limit, the bridge arm circulating current error exceeds the set threshold, or the sub-module capacitor voltage deviation exceeds the set threshold, the bridge arm voltage reference value is limited and corrected, or switched to standby deadbeat model predictive control, multi-step model predictive control, or PI control mode.

[0018] Preferably, another aspect of the present invention provides a multi-step model prediction and control system for MMC based on neural network approximation, including a model building module, a multi-step prediction and optimization module, a reference value calculation module, an offline training module, an online control module, and a safety protection module; The model building module is used to establish a state-space model based on the target modular multilevel converter topology, including AC side output current, bridge arm circulating current, grid side voltage and DC side voltage, and to discretize it to obtain the MMC discrete mathematical model. The multi-step prediction optimization module is used to derive the multi-step prediction state matrix based on the MMC discrete mathematical model, construct a cost function that includes the three-phase output current tracking error and the three-phase bridge arm circulating current tracking error, transform the cost function into the standard form of quadratic programming, and solve it offline to obtain the optimal control quantity sequence of the reference voltage of the three-phase upper and lower bridge arms. The reference value calculation module is used to generate three-phase output current reference values ​​and three-phase bridge arm circulating current reference values, and to perform multi-step extrapolation on the three-phase output current reference values ​​and three-phase bridge arm circulating current reference values. The offline training module is used to construct a multi-condition sample dataset by taking the three-phase grid-side current, the three-phase grid-side current reference value, the grid-side voltage in the two-phase stationary coordinate system, the three-phase bridge arm circulating current and the three-phase bridge arm circulating current reference value as input features, and the optimal control quantity of the three-phase upper and lower bridge arm reference voltage as output labels, and to train the artificial neural network offline. The online control module is used to collect state variables and reference values ​​in real time, and directly map and output the reference values ​​of the three-phase upper and lower bridge arm voltages for the next control cycle through the trained neural network controller. It also generates submodule switching signals by combining the nearest level approximation modulation and submodule capacitor voltage sorting strategy. The security protection module is used to perform boundary detection on the neural network input, neural network output and modulation results, and to perform amplitude limiting correction or switch to backup control mode in abnormal situations.

[0019] Preferably, the offline training module includes a data preprocessing unit, a network configuration unit, and an iterative training unit; The data preprocessing unit is used to normalize the input features and output labels in the multi-condition sample data. The network configuration unit is used to set the input layer, hidden layer, output layer, activation function, and loss function of the artificial neural network; The iterative training unit is used to iteratively update the network parameters using the gradient descent method to complete the training of the neural network controller.

[0020] Preferably, the online control module includes a signal acquisition unit, a neural network inference unit, and a modulation drive unit; The signal acquisition unit is used to acquire the three-phase grid-side current, grid-side voltage in the two-phase stationary coordinate system, three-phase bridge arm circulating current, DC-side voltage and submodule capacitor voltage in real time, and calculate the corresponding reference values. The neural network inference unit is used to input the state variables and reference values ​​into the trained neural network controller and output the reference values ​​of the three-phase upper and lower bridge arm voltages. The modulation drive unit is used to generate submodule switching signals based on the reference values ​​of the three-phase upper and lower bridge arm voltages, the sorting results of the submodule capacitor voltages, and the direction of the bridge arm current, and to drive the power devices to operate.

[0021] Preferably, the safety protection module is used to perform input / output boundary detection, bridge arm voltage limiting, submodule number exceeding the limit judgment, bridge arm circulating current error judgment, submodule capacitor voltage deviation judgment, and backup control mode switching; When the neural network input state quantity exceeds the coverage range of the training samples, the output bridge arm voltage reference value exceeds the allowable range, the number of sub-modules deployed exceeds the limit, the bridge arm circulating current error exceeds the set threshold, or the sub-module capacitor voltage deviation exceeds the set threshold, the safety protection module performs amplitude limiting correction on the bridge arm voltage reference value, or switches the control mode to standby deadbeat model predictive control, multi-step model predictive control, or PI control mode.

[0022] The present invention has the following beneficial effects: 1. This invention fully retains the multi-objective coordination capability and excellent dynamic and steady-state performance of multi-step continuous set model predictive control, while reducing the online computational complexity from that of traditional CCS-MPC through offline approximation using neural networks. Reduce to a constant The online computation time reaches the microsecond level, which completely solves the problem of exponential growth in computational load caused by the increase in prediction step size and greatly reduces the hardware computing power requirements.

[0023] 2. This invention simultaneously achieves accurate output current tracking and bridge arm circulating current suppression, and combines a bridge arm energy balancing strategy to ensure stable submodule capacitor voltage, taking into account both grid-connected power quality and converter operation reliability.

[0024] 3. The method framework of this invention can be adapted to various MMC topologies such as half-bridge, full-bridge, and hybrid submodules, and can also be extended to other types of grid-connected converters, making it widely applicable in engineering.

[0025] 4. The neural network training process of this invention is completed entirely offline. Online control only needs to perform forward inference, without the need for a complex optimization solver, which makes it easy to deploy in embedded controllers.

[0026] 5. This invention demonstrates a clear improvement in real-time and steady-state performance. With a 50μs control sampling period, the average online computation time using the described ANN-MPC method is reduced from 20.01μs in the traditional MS-MPC method to 3.51μs, a reduction of approximately 82.46%. This significantly increases the time margin for sampling, communication, modulation, and protection logic. Simultaneously, the AC output current THD is approximately 1.15%, close to the 1.06% of the MS-MPC method. This indicates that while significantly reducing the online computation burden, this invention maintains good output current quality, steady-state accuracy, and dynamic response capabilities, making it suitable for real-time engineering deployment in embedded controllers and high-voltage, high-capacity MMC flexible DC equipment. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] Figure 1 A roadmap for MMC multi-step model predictive control technology based on neural network approximation; Figure 2 This is a schematic diagram of a half-bridge grid-connected MMC topology; Figure 3 A schematic diagram of the long prediction time domain CCS-MPC control structure; Figure 4 Calculate the control block diagram for the bridge arm circulation reference value; Figure 5 This is a schematic diagram of an artificial neural network structure; Figure 6 Flowchart of offline training process for neural networks; Figure 7 A schematic diagram of the overall control structure of an ANN based on CCS-MPC approximation; Figure 8 The simulation waveform of the three-phase output current is shown below. Figure 9 Simulation waveforms of the current in the upper and lower bridge arms of phase A; Figure 10 The simulation waveform of the circulating current in phase A bridge arm; Figure 11 The simulation waveforms of the capacitor voltages of the upper and lower bridge arm submodules in phase A are shown.

[0029] Figure 12 A schematic diagram of a dedicated ANN structure for bridge arm voltage mapping; Figure 13 Flowchart for online boundary detection, amplitude limiting correction, and backup control rollback; Figure 14 This is a performance comparison chart of DB-MPC, MS-MPC, and ANN-MPC.

[0030] Table 1 Simulation parameters of MMC system Table 2 Neural Network Training Parameters Table 3 Performance Comparison of DB-MPC, MS-MPC and ANN-MPC in, Figure 1 This invention is used to illustrate the overall technical route of the present invention, from MMC modeling, multi-step prediction optimization, QP optimal bridge arm voltage sample generation, ANN offline training to online inference control. Figure 12 This is used to illustrate the dedicated neural network mapping structure with 14 input features and 6 output bridge arm voltages used in this invention. Figure 13 This is used to illustrate the input boundary detection, output limiting, input submodule quantity verification, and backup control fallback mechanism during the online control phase. Figure 14 This is used to illustrate the performance comparison results of the present invention compared with DB-MPC and MS-MPC in terms of average calculation time and current THD. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings and multiple embodiments. The described embodiments are only some implementations of the present invention, and not all implementations. Based on the technical solutions of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0032] Example 1: A multi-step model predictive control method for modular multilevel converters based on neural network approximation: like Figure 1 As shown, the overall technical approach of this invention includes MMC discrete state-space modeling, multi-step prediction matrix derivation, offline quadratic programming to solve for optimal bridge arm voltage samples, offline training of artificial neural networks, online forward inference to output bridge arm voltage reference values, nearest-level approximation modulation and submodule capacitor voltage sorting, and safety backoff control under abnormal operating conditions. This technical approach transfers the computationally intensive online optimization solution in traditional multi-step CCS-MPC to the offline stage. The online stage only performs neural network forward inference and modulation execution, thus balancing control performance and real-time performance. This embodiment uses a three-phase grid-connected half-bridge modular multilevel converter as the controlled object to fully implement the control method described in this invention. The specific implementation process is as follows: Step S1: Establish a discrete state-space model This step involves constructing a mathematical model that accurately describes the electrical dynamic characteristics of the converter, providing a reliable model foundation for subsequent multi-step prediction calculations and optimization solutions.

[0033] Based on the topology of the target modular multilevel converter, the AC equivalent circuit and DC equivalent circuit are decomposed by combining Kirchhoff's voltage law and current law. The output current of each phase on the AC side, the circulating current of each phase arm, the grid-side voltage in the two-phase stationary coordinate system, and the DC side voltage are selected as state variables. The lower arm voltage and upper arm voltage of each phase are selected as control variables. The output current of each phase and the circulating current of each phase arm are selected as output vectors. The continuous domain state space equation is constructed and the corresponding system matrix, input matrix, and output matrix are generated. The Euler forward difference formula is used to discretize the continuous state space equation to obtain the discrete mathematical model and discrete system matrix for the corresponding sampling period.

[0034] The obtained discrete mathematical model can accurately reflect the state evolution law of the converter, ensure the calculation accuracy of subsequent multi-step prediction, and provide underlying support for the steady-state performance of the control strategy.

[0035] Step S2: Multi-step prediction optimization and optimal control sample generation This step obtains the optimal control quantity for multi-step continuous set model predictive control through multi-step prediction derivation and quadratic programming solution. At the same time, it generates sample data covering various operating conditions, providing labels and feature data for neural network training.

[0036] Based on the discrete mathematical model, the state and output quantities at all times in the prediction time domain are iteratively predicted. The prediction state matrix and prediction control matrix are constructed, and the single-step state equation is transformed into a matrix expression of the prediction output. A cost function containing the output current tracking error term and the arm circulating current tracking error term is constructed, and the weights of the two types of control objectives are allocated through the penalty matrix.

[0037] Synchronous calculation of control reference values: The phase angle of the grid-connected voltage is acquired by phase-locked loop, and the three-phase stationary coordinate system is transformed into a dq rotating coordinate system through coordinate transformation. The dq axis current reference value is calculated according to the active and reactive power commands, and the three-phase output current reference value is obtained through inverse coordinate transformation. The basic circulating current reference value is calculated by using the second-harmonic circulating current injection method. At the same time, the input power of the upper and lower bridge arms of a single phase is calculated, and the common-mode component and differential-mode component of the bridge arm energy are extracted. The circulating current adjustment component is output by two PI controllers and superimposed with the basic circulating current reference value to obtain the final circulating current reference value. The current and circulating current reference values ​​are extrapolated in multiple steps using a third-order Lagrange interpolation polynomial to generate a reference trajectory sequence in the prediction time domain.

[0038] Substituting the prediction matrix and reference trajectory into the cost function, the result is rearranged into a standard quadratic programming form and solved to obtain the optimal arm voltage control sequence at each time step. This process covers various steady-state and dynamic disturbance conditions, generating multiple sets of sample data with one-to-one correspondence between "state quantity - reference value - optimal control quantity".

[0039] This step fully reproduces the optimal control law of multi-step continuous set model predictive control. The generated samples cover the entire operating range of the converter, providing sufficient and high-quality training data for the neural network to achieve high-precision approximation.

[0040] Step S3: Offline training of artificial neural networks This step enables the neural network to learn the input-output mapping relationship of multi-step model predictive control through offline training, transforming the complex online quadratic programming solution into a simple neural network forward computation, thus reducing the dimensionality and burden of online control.

[0041] A three-layer feedforward neural network structure consisting of an input layer, a hidden layer, and an output layer is constructed. The input layer receives three-phase grid-side current, three-phase grid-side current reference value, two-phase stationary coordinate system grid-side voltage, three-phase bridge arm circulating current, and three-phase bridge arm circulating current reference value as input features. The output layer outputs the upper and lower bridge arm reference voltages corresponding to the three phases. The hidden layer achieves nonlinear mapping through activation functions. Each neuron in the layer weights and sums the input signals, adds a bias threshold, and outputs the signal to the next layer after transformation by the activation function.

[0042] Max-min normalization is performed on all sample features and target quantities to map each physical quantity to a uniform numerical range. The samples are divided into training set, validation set and test set. Mean square error is selected as the loss function, and gradient descent algorithm is used to iteratively update the connection weights and bias thresholds of the network. Training stops when the validation set loss meets the preset accuracy requirement or reaches the maximum number of iterations, and the trained neural network controller model is saved.

[0043] The trained neural network can approximate the optimal control law of multi-step continuous set model predictive control with high precision, transforming the optimization problem that originally required iterative solution into forward inference with a fixed amount of computation, thereby reducing the complexity of online computation from the root.

[0044] Step S4: Implementation of online closed-loop control This step connects the trained neural network controller to the actual control system of the converter, achieving real-time closed-loop control with low computational burden.

[0045] The trained neural network controller is deployed to the converter embedded controller to build an online closed-loop control system. The system's operating status variables are collected in real time, and the three-phase current reference values ​​and circulating current reference values ​​are calculated synchronously. After normalization, they are input into the neural network controller. The three-phase upper and lower bridge arm voltage reference values ​​are directly output through neural network forward inference.

[0046] The switching signal is generated using the nearest level approximation modulation strategy: the reference value of the bridge arm voltage is divided by the reference voltage of the submodule capacitor to obtain the number of submodules to be put into operation for the corresponding bridge arm; the capacitor voltages of all submodules in the current bridge arm are sorted, and the corresponding number of submodules are selected for operation in combination with the direction of the bridge arm current, so as to simultaneously achieve the balance between the level output and the capacitor voltage of the submodule, and finally generate the switching drive signal of the submodule power device.

[0047] By eliminating the need for complex quadratic programming solutions during the online control process, the computational load remains constant. While fully preserving the dynamic and steady-state performance of multi-step model predictive control, the real-time performance of control is significantly improved, and the hardware computing power requirements of the controller are reduced.

[0048] Example 2: See Figure 12 This invention provides a method and system for multi-step model predictive control (MMC) based on neural network approximation, comprising the following steps: This embodiment uses a three-phase grid-connected half-bridge MMC as the controlled object, fully implements the control method of this invention, and provides all derivation formulas. The MMC topology is as follows: Figure 1 As shown, each phase consists of upper and lower bridge arms, and each bridge arm is composed of... The system consists of a half-bridge module connected in series with a bridge arm inductor. The AC side is connected to the power grid through a filter inductor, and the DC side is connected to the DC bus.

[0049] Step 1: Establish a discrete state-space model according to Figure 2 The MMC topology shown, combined with voltage and current reference directions and Kirchhoff's current law, can be analyzed for single-phase circuits to obtain the following: (1); (2); In the formula: , indicating the corresponding phase sequence; For MMC communication side j Phase output current; and They are respectively j Current in the upper and lower bridge arms; for j Phase bridge arm circulation.

[0050] According to equations (1) and (2), the MMC single-phase circuit can be decomposed into an AC equivalent circuit and a DC equivalent circuit, and the corresponding voltage equations are as follows: (3); (4); In the formula: For equivalent AC inductance, satisfy ,in For grid-side filter inductors, For bridge arm inductance; , They are respectively j Output voltage of upper and lower bridge arms; Let j be the voltage of the grid phase; This is the DC side voltage.

[0051] Based on Euler's forward difference formula, the above MMC dynamic characteristic equation can be discretized to obtain: (5); (6); In the formula: The sampling period; Indicates the current Constantly sampled AC output current value. express The alternating current output at any given moment; similarly This represents the current circulating sample value. express The circulating sample value at a given time.

[0052] Selecting state variables Control variables Output vector The corresponding continuous state-space equation can then be expressed as: (7); in A Phase constant matrix , , They are respectively: ; In the formula Represents the grid angular velocity. B and C phase constant matrices. , The phase constant matrices of phase A are all different from those of phase A, and are as follows: ; ; Output matrices of phases B and C , and The structure is consistent.

[0053] Discretizing the continuous state-space equations using the Euler forward difference method yields the discrete state equations: ; in , , It is an identity matrix.

[0054] Step 2: Multi-step prediction derivation and quadratic programming solution To estimate the prediction time domain The future system state within the range is solved iteratively using discrete state-space equations. The calculation process is as follows: ; To facilitate the solution of quadratic programming matrix operations, the prediction time domain is set as follows: The set of all state variables and output vectors is represented as: ; The single-step discrete space state equation can then be rewritten in prediction matrix form: (9); In the formula: ; Cost function Considering both the output current error and the bridge arm circulating current error, the expression is as follows: (10); In the formula: For reference value vector, This is the actual output value vector; Predict the number of steps for MPC; This is the penalty matrix, used to represent the weight between the output current and the circulating current error.

[0055] To solve quadratic programming problems using the optimization toolbox, the cost function is... Combined with the prediction model, the standard form of the QP solution is further derived as follows: (11); In the formula: , , The penalty matrix for block diagonalization. Reference value obtained by extrapolation , A set of vectors of size . , This is a constant term independent of the control input. Solving this quadratic programming problem yields the optimal control input sequence, and the control input at the first time step is taken as the output at the current time step.

[0056] Reference value calculation and multi-step extrapolation: The voltage phase angle at PCC is obtained through a phase-locked loop. The three-phase stationary coordinate system abc is transformed into a rotating coordinate system dq using coordinate transformation. The q-axis component of the AC side voltage when the three phases are symmetrical is then defined. The value is 0. Since the line resistance is small, neglecting losses in the MMC bridge arms and the line, the active power on the DC side of the MMC should be equal to the active power on the AC side. The reference values ​​for the phase currents in the dq coordinate system are as follows: (12) In the formula: , These are the command values ​​for active and reactive power of the MMC system, respectively. This represents the d-axis component of the voltage at PCC. Due to three-phase symmetry... The above formula can be simplified to , .

[0057] By performing an inverse coordinate transformation on the dq-axis current reference values, the reference values ​​for the three-phase currents can be obtained. coordinate transformation matrix for: ; In addition, MMC circulating current suppression is achieved by using a frequency-doubled circulating current injection method. In this case, the basic circulating current reference value is: (13); In the formula: This is the reference value for the output phase voltage; This is a reference value for the output phase current.

[0058] To maintain the stable and efficient operation of the MMC system, it is necessary to ensure the conservation of MMC bridge arm voltage and DC side voltage, and the voltage balance between the upper and lower bridge arms. Bridge arm inductance. The energy stored in the middle is negligible compared to the energy stored in the submodule capacitors. Therefore, the input power of the single-phase upper and lower bridge arms can be expressed as follows: (14); (15); The common-mode component can be obtained by adding the energy of each phase of the upper and lower bridge arms. Subtracting them yields the differential modulus component. Since the output phase voltage and phase current of MMC can be approximated as sinusoidal signals, that is... , The change in bridge arm energy can then be further expressed as: (16); (17); Will , The circulating current reference value component can be obtained by comparing it with the reference value and using a PI controller. , , and the baseline reference value The actual circulation reference value can be obtained by adding them together. .

[0059] To provide a reference trajectory for multi-step MPC at multiple future sampling times, a third-order Lagrange interpolation polynomial is used for extrapolation calculation. The formula is as follows: (18); in For the prediction step number, For the first time in history Reference values ​​for steps.

[0060] Step 3: Artificial Neural Network Construction and Offline Training Construct a neural network structure consisting of an input layer, hidden layers, and an output layer, such as... Figure 4 As shown. The input layer includes the grid-side current. Reference grid-side current Grid-side voltage Circulating current Reference circulating current ( The output layer includes the voltages of the upper and lower bridge arms of the three phases. The output layer has a size of 6, representing the reference voltages for the upper and lower arms of the three-phase bridge. .

[0061] The signals can be nonlinearly processed through activation functions between neurons in each layer, and weights and thresholds can be assigned. The input-output relationship of each neuron in each layer can be represented as follows: (19); In the formula: Represents the previous level The output of each neuron; Represents the first layer The output of each neuron; For activation functions; For the next layer of neurons The weight, The threshold value is used.

[0062] Because the dimensions of different characteristic quantities such as voltage and current are inconsistent, the sampled data is normalized: (20); in Represents each set of sample data. , These represent the minimum and maximum values ​​of each characteristic quantity.

[0063] In this embodiment, the error function is the mean square error (MSE), and its expression is as follows: (twenty one); in: The number of samples in the training set; For the true value of the target quantity, These are the predicted values ​​from the neural network. The training process uses gradient descent to iteratively update the weights and thresholds.

[0064] Step 4: Online control and modulation implementation The trained neural network is used as an approximate MPC controller and substituted into the MMC control system. The overall control structure is as follows: Figure 6 As shown, the reference value of the bridge arm voltage at the next moment is obtained based on the mapping between the system sampled value and the reference value, and the switching signal is generated by combining the nearest level approximation modulation strategy.

[0065] Divide the bridge arm voltage reference value by the submodule capacitor reference voltage. This allows you to calculate the number of submodules needed for the upper and lower bridge arms: (twenty two); Taking the upper arm of phase a as an example, arrange the current voltage of the arm submodule capacitors from largest to smallest. If the current direction is negative at this time, the submodule capacitor is discharging, and the previous phase is engaged. Each submodule; conversely, if the current direction is positive, the submodule capacitor charges, and then after being put into operation... Each sub-module is used to achieve capacitor voltage equalization control.

[0066] Example 3: This embodiment provides a multi-step model predictive control system for MMC based on neural network approximation, which implements the control method described in Embodiment 1. It is applied to a three-phase grid-connected half-bridge modular multilevel converter. The system includes a model building module, a multi-step prediction optimization module, a reference value calculation module, an offline training module, and an online control module. Data interaction and communication between the modules are achieved to jointly complete the offline controller training and online closed-loop control of MMC.

[0067] 1. Model building module For the topology of the target modular multilevel converter, a system state-space model is established and discretized to obtain a discrete mathematical model.

[0068] Specifically, based on the topology and electrical parameters of the half-bridge MMC, this module derives the voltage and current equations for the single-phase AC equivalent circuit and DC equivalent circuit using Kirchhoff's laws. It selects the output current of each phase on the AC side, the circulating current of each phase arm, the grid-side voltage in the two-phase stationary coordinate system, and the DC-side voltage as state variables; selects the lower and upper arm voltages of each phase as control variables; and selects the output current of each phase and the circulating current of each phase arm as output vectors. This constructs a continuous-domain state-space equation and generates the corresponding system matrix, input matrix, and output matrix. Based on Euler's forward difference formula, the continuous-space equation is discretized to obtain discrete state equations and a discrete system matrix, which are then output to the multi-step prediction and optimization module.

[0069] 2. Multi-step prediction optimization module This is used to derive the multi-step predictive state matrix based on a discrete mathematical model, extrapolate and calculate the predicted values ​​of all state variables of the system within a preset prediction time domain; construct the cost function of model predictive control with the converter output current tracking error and the bridge arm circulating current tracking error as the control objectives, transform the cost function into a quadratic programming problem and solve it to obtain the optimal control quantity sample.

[0070] Specifically, the module receives a discrete mathematical model, iteratively recursively derives the discrete state-space equations to obtain a set of predicted state variables and a set of predicted output variables for all times in the prediction time domain, and transforms the single-step state equations into a prediction matrix form; it constructs a cost function containing output current error terms and bridge arm circulating current error terms, configures a penalty matrix to adjust the weights of the two types of control objectives; combines the prediction matrix to transform the cost function into a standard quadratic programming form, calls a quadratic programming solver to solve for the optimal control variable sequence, and outputs the optimal control variable as a sample label to the offline training module.

[0071] 3. Reference value calculation module The reference value calculation module is used to generate current reference values ​​and circulating current reference values ​​for system operation and output them to the multi-step prediction optimization module and the online control module; the reference value calculation module includes a current reference calculation unit and a circulating current reference calculation unit.

[0072] Current reference calculation unit: The phase angle of the grid connection point voltage is obtained through a phase-locked loop. The three-phase stationary coordinate system is transformed into a dq rotating coordinate system through coordinate transformation. The dq axis current reference value is calculated based on the active and reactive power commands. Then, the three-phase output current reference value is obtained through inverse coordinate transformation.

[0073] The circulating current reference calculation unit calculates the basic circulating current reference value using a second-harmonic circulating current injection method; it calculates the input power of the upper and lower single-phase bridge arms separately, sums them to obtain the common-mode component of the bridge arm energy, and calculates the difference to obtain the differential-mode component of the bridge arm energy; it compares the common-mode component and the differential-mode component with the corresponding reference value and inputs them into the PI controller, outputs the circulating current reference component, and superimposes it with the basic circulating current reference value to obtain the final bridge arm circulating current reference value.

[0074] Meanwhile, the reference value calculation module uses a third-order Lagrange interpolation polynomial to extrapolate the output current reference value and the bridge arm circulating current reference value in multiple steps, generating a reference trajectory sequence in the prediction time domain, which is then output to the multi-step prediction optimization module as the optimization target.

[0075] 4. Offline Training Module This system collects state sample values ​​and corresponding reference values ​​during operation as feature quantities, and uses the optimal control quantity obtained from the corresponding solution as the target value to construct a sample dataset. Based on the sample dataset, an artificial neural network is trained offline to obtain a neural network controller that approximates the predictive control law of a multi-step continuous set model. The offline training module includes a data preprocessing unit, a network configuration unit, and an iterative training unit.

[0076] Data preprocessing unit: Collects system state sample values ​​and corresponding reference values ​​under multiple operating conditions such as steady-state operation, power step, and voltage disturbance as input features, matches the corresponding optimal control quantity as output label, and constructs sample dataset; performs maximum-minimum normalization processing on all feature quantities and target quantities respectively, and maps each physical quantity to a unified numerical range of [0,1].

[0077] Network configuration unit: Configure the artificial neural network as a three-layer feedforward network structure consisting of an input layer, a hidden layer, and an output layer; set the input layer features to include three-phase grid-side current, three-phase grid-side current reference value, two-phase stationary coordinate system grid-side voltage, three-phase bridge arm circulating current, and three-phase bridge arm circulating current reference value; set the output layer output to include the upper and lower bridge arm reference voltages corresponding to the three phases; configure the number of neurons in the hidden layer, the activation function type, and select mean square error as the loss function.

[0078] Iterative training unit: The gradient descent method is used to iteratively update the connection weights and bias thresholds of the network. The loss function value is calculated in each iteration until the loss function meets the preset accuracy requirements or reaches the maximum number of iterations. The trained neural network controller model is then saved and output to the online control module.

[0079] 5. Online control module This online control module is used to connect a trained neural network controller to the online control system of a modular multilevel converter. It collects system state variables and reference values ​​in real time and inputs them into the neural network controller, mapping and outputting the bridge arm voltage reference value for the next moment. Based on the bridge arm voltage reference value and a modulation strategy, it generates submodule switching signals to drive the converter. The online control module includes a signal acquisition unit, a neural network inference unit, and a modulation drive unit.

[0080] Signal acquisition unit: Real-time acquisition of state quantities such as three-phase output current, three-phase bridge arm circulating current, grid-side voltage, and DC-side voltage during converter operation. Combined with the three-phase current reference value and circulating current reference value generated by the reference value calculation module, the data is normalized and then output to the neural network inference unit.

[0081] Neural network inference unit: Loads the trained neural network controller model, performs forward propagation calculations on the input state variables and reference values, and directly maps and outputs the reference values ​​of the three-phase upper and lower bridge arm voltages at the next moment.

[0082] Modulation drive unit: Adopting the nearest level approximation modulation strategy, the bridge arm voltage reference value is divided by the sub-module capacitor reference voltage to obtain the number of sub-modules to be put into operation for the corresponding bridge arm; the capacitor voltages of all sub-modules in the current bridge arm are sorted, and the corresponding number of sub-modules are selected for operation in combination with the bridge arm current direction, generating the switching drive signal of the sub-module power device, and outputting it to the MMC main circuit to realize closed-loop control.

[0083] To verify the control performance and feasibility of the method of this invention, simulation verification was performed in a three-phase grid-connected half-bridge MMC system. The main simulation parameters are shown in Table 1; the neural network training parameters are shown in Table 2. The training samples consist of MS-MPC offline optimization data under different power command values ​​and control cycle conditions.

[0084] Table 1 Main Simulation Parameters

[0085] Table 2 Neural Network Training Parameters

[0086] To examine dynamic response capability, the active power command was stepped from 0.12 MW to 0.24 MW at 0.6 s, while the reactive power command remained unchanged. Simulation results show that after the power step, the three-phase output current can quickly track the new reference value, and the steady-state AC output current THD is approximately 1.04%; the circulating current fluctuation amplitude of the bridge arm before the step is approximately 5 A, and the circulating current amplitude only increases slightly after the step; the voltage fluctuation amplitude of the upper and lower bridge arm submodule capacitors is less than 5 V, and equilibrium can be quickly restored after the power step. Further, an RTLab hardware-in-the-loop experimental platform was built. The MMC main circuit model was deployed in the RTLab real-time simulation equipment, and the neural network approximation control algorithm was deployed in the DSPF28335 control chip. The two interact through a real-time signal interface to form a closed-loop control. Comparisons of the three control methods are shown in Table 3 and... Figure 14 As shown in Table 3, the average computation time of the ANN-MPC method described in this invention is 3.51 μs, which is only about 17.5% of the 20.01 μs of the MS-MPC method, representing a reduction of approximately 82.46% in computation time. Simultaneously, its current THD is 1.15%, only 0.09 percentage points higher than the 1.06% of MS-MPC, and significantly better than the 1.32% of DB-MPC. Therefore, this invention, while essentially maintaining the steady-state accuracy and dynamic response capability of MS-MPC, replaces online complex quadratic programming solutions with neural network forward inference with a fixed computational load, thus meeting the real-time operation requirements within a 50 μs control cycle.

[0087] Table 3 Comparison of Three Control Methods

[0088] This system adopts an architecture that separates offline training and online inference. The computationally intensive multi-step prediction and quadratic programming solution are transferred to the offline stage. The online stage only needs to perform neural network forward computation. While fully preserving the multi-step CCS-MPC control performance, it significantly reduces the computing power requirements of the online controller. It can be directly deployed on conventional embedded control chips and is suitable for the real-time control needs of high-voltage, high-capacity MMC.

[0089] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is not limited to these embodiments. All similar variations related to these embodiments are within the protection scope of the present invention. All modifications that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should fall within the protection scope of the present invention.

Claims

1. A multi-step model predictive control method for MMC based on neural network approximation, characterized in that, The steps include the following: S1. Based on the target modular multilevel converter topology, a state-space model including AC side output current, bridge arm circulating current, grid side voltage and DC side voltage is established, and the state-space model is discretized to obtain the MMC discrete mathematical model. S2. Based on the MMC discrete mathematical model, derive the multi-step prediction state matrix, construct a cost function that includes the three-phase output current tracking error and the three-phase bridge arm circulating current tracking error, transform the cost function into the standard form of quadratic programming, solve offline to obtain the optimal control quantity sequence of the reference voltage of the three-phase upper and lower bridge arms, and take the corresponding control quantity as the training label. S3. Using the three-phase grid-side current, the three-phase grid-side current reference value, the grid-side voltage in the two-phase stationary coordinate system, the three-phase bridge arm circulating current and the three-phase bridge arm circulating current reference value as input features, and the optimal control quantity of the three-phase upper and lower bridge arm reference voltages as output labels, a multi-condition sample dataset is constructed, and an artificial neural network is trained offline to obtain a neural network controller for approximating the predictive control law of a multi-step continuous set model. S4. Connect the trained neural network controller to the MMC online control system, collect state variables and reference values ​​in real time and input them into the neural network controller, directly map and output the reference values ​​of the three-phase upper and lower bridge arm voltages for the next control cycle, and generate sub-module switching signals by combining the nearest level approximation modulation and sub-module capacitor voltage sorting strategy to drive the MMC to run.

2. The MMC multi-step model predictive control method based on neural network approximation according to claim 1, characterized in that, Step S2 includes: S21. Select state variables, control variables, and output vectors to construct discrete state-space equations; S22. Iterate and recursively derive the discrete state space equation to obtain the predicted set of output quantities in the prediction time domain, and transform the single-step prediction model into a multi-step prediction matrix form. S23. Construct a cost function that includes the output current error term and the bridge arm circulating current error term. Combine the multi-step prediction matrix to organize the cost function into a standard quadratic programming form and solve to obtain the optimal control quantity sequence of the bridge arm voltage.

3. The MMC multi-step model predictive control method based on neural network approximation according to claim 2, characterized in that, The state variables include the output current of each phase on the AC side, the circulating current of each phase arm, the grid-side voltage in the two-phase stationary coordinate system, and the DC side voltage; the control variables include the upper arm voltage and the lower arm voltage of each phase; the output vector includes the output current of each phase and the circulating current of each phase arm.

4. The MMC multi-step model predictive control method based on neural network approximation according to claim 1, characterized in that, Step S2 further includes a reference value calculation step, which includes: The phase angle of the grid connection point voltage is obtained by phase-locked loop, and the three-phase stationary coordinate system is transformed into the dq rotating coordinate system by coordinate transformation. The dq axis current reference value is calculated based on the active power command and reactive power command, and then the three-phase output current reference value is obtained by inverse coordinate transformation. A two-frequency circulating current injection method combined with a bridge arm energy balancing strategy is used to generate a three-phase bridge arm circulating current reference value.

5. The MMC multi-step model predictive control method based on neural network approximation according to claim 4, characterized in that, The bridge arm energy balancing strategy includes: Calculate the input power or energy storage state of the upper and lower bridge arms of a single phase respectively, sum them to obtain the common-mode component of the bridge arm energy, and calculate the difference to obtain the differential-mode component of the bridge arm energy. The common-mode component and differential-mode component of the arm energy are compared with their corresponding reference values ​​and then input into the PI controller. The controller outputs the circulating current adjustment component, which is then superimposed on the basic circulating current reference value to obtain the final arm circulating current reference value.

6. The MMC multi-step model predictive control method based on neural network approximation according to claim 4, characterized in that, A third-order Lagrange interpolation polynomial is used to extrapolate the three-phase output current reference value and the three-phase bridge arm circulating current reference value in multiple steps to generate a reference trajectory sequence in the prediction time domain, and the reference trajectory sequence is used as the optimization target of multi-step model predictive control.

7. The MMC multi-step model predictive control method based on neural network approximation according to claim 1, characterized in that, The artificial neural network adopts a feedforward network structure consisting of an input layer, a hidden layer, and an output layer; The input layer includes 14 input features, namely, three-phase grid-side current, three-phase grid-side current reference value, grid-side voltage in two-phase stationary coordinate system, three-phase bridge arm circulating current, and three-phase bridge arm circulating current reference value; The output layer includes six output quantities, namely the three-phase upper bridge arm reference voltage and the three-phase lower bridge arm reference voltage.

8. The MMC multi-step model predictive control method based on neural network approximation according to claim 7, characterized in that, The artificial neural network includes 14 input neurons, 12 hidden layer neurons, and 6 output neurons; before training, the input features and output labels are subjected to max-min normalization. The training process uses mean squared error as the loss function and gradient descent to update the connection weights and bias thresholds until the loss function meets the preset accuracy requirements or reaches the maximum number of iterations.

9. The MMC multi-step model predictive control method based on neural network approximation according to claim 1, characterized in that, The nearest-level approximation modulation in step S4 includes: Divide the reference values ​​of the three-phase upper and lower bridge arm voltages by the reference voltage of the submodule capacitors to obtain the number of submodules that need to be put into the corresponding bridge arm. The voltage of all submodule capacitors in the current bridge arm is sorted, and the corresponding number of submodules are selected to be put into or cut off according to the direction of the bridge arm current, so as to achieve the balance between the bridge arm voltage output and the submodule capacitor voltage.

10. The MMC multi-step model predictive control method based on neural network approximation according to claim 1, characterized in that, Boundary detection is performed on the neural network input, output, and modulation results during online runtime; When the input state quantity exceeds the coverage of the training samples, the output bridge arm voltage reference value exceeds the allowable range, the number of sub-modules to be put into the bridge arm exceeds the limit, the bridge arm circulating current error exceeds the set threshold, or the sub-module capacitor voltage deviation exceeds the set threshold, the bridge arm voltage reference value is limited and corrected, or switched to standby deadbeat model predictive control, multi-step model predictive control, or PI control mode.

11. A multi-step model predictive control system (MMC) based on neural network approximation, characterized in that, It includes a model building module, a multi-step prediction and optimization module, a reference value calculation module, an offline training module, an online control module, and a security protection module; The model building module is used to establish a state-space model based on the target modular multilevel converter topology, including AC side output current, bridge arm circulating current, grid side voltage and DC side voltage, and to discretize it to obtain the MMC discrete mathematical model. The multi-step prediction optimization module is used to derive the multi-step prediction state matrix based on the MMC discrete mathematical model, construct a cost function that includes the three-phase output current tracking error and the three-phase bridge arm circulating current tracking error, transform the cost function into the standard form of quadratic programming, and solve it offline to obtain the optimal control quantity sequence of the reference voltage of the three-phase upper and lower bridge arms. The reference value calculation module is used to generate three-phase output current reference values ​​and three-phase bridge arm circulating current reference values, and to perform multi-step extrapolation on the three-phase output current reference values ​​and three-phase bridge arm circulating current reference values. The offline training module is used to construct a multi-condition sample dataset by taking the three-phase grid-side current, the three-phase grid-side current reference value, the grid-side voltage in the two-phase stationary coordinate system, the three-phase bridge arm circulating current and the three-phase bridge arm circulating current reference value as input features, and the optimal control quantity of the three-phase upper and lower bridge arm reference voltage as output labels, and to train the artificial neural network offline. The online control module is used to collect state variables and reference values ​​in real time, and directly map and output the reference values ​​of the three-phase upper and lower bridge arm voltages for the next control cycle through the trained neural network controller. It also generates submodule switching signals by combining the nearest level approximation modulation and submodule capacitor voltage sorting strategy. The security protection module is used to perform boundary detection on the neural network input, neural network output and modulation results, and to perform amplitude limiting correction or switch to backup control mode in abnormal situations.

12. The MMC multi-step model predictive control system based on neural network approximation according to claim 11, characterized in that, The offline training module includes a data preprocessing unit, a network configuration unit, and an iterative training unit. The data preprocessing unit is used to normalize the input features and output labels in the multi-condition sample data. The network configuration unit is used to set the input layer, hidden layer, output layer, activation function, and loss function of the artificial neural network; The iterative training unit is used to iteratively update the network parameters using the gradient descent method to complete the training of the neural network controller.

13. The MMC multi-step model predictive control system based on neural network approximation according to claim 11, characterized in that, The online control module includes a signal acquisition unit, a neural network inference unit, and a modulation drive unit; The signal acquisition unit is used to acquire the three-phase grid-side current, grid-side voltage in the two-phase stationary coordinate system, three-phase bridge arm circulating current, DC-side voltage and submodule capacitor voltage in real time, and calculate the corresponding reference values. The neural network inference unit is used to input the state variables and reference values ​​into the trained neural network controller and output the reference values ​​of the three-phase upper and lower bridge arm voltages. The modulation drive unit is used to generate submodule switching signals based on the reference values ​​of the three-phase upper and lower bridge arm voltages, the sorting results of the submodule capacitor voltages, and the direction of the bridge arm current, and to drive the power devices to operate.

14. The MMC multi-step model predictive control system based on neural network approximation according to claim 11, characterized in that, The safety protection module is used to perform input / output boundary detection, bridge arm voltage limiting, submodule number exceeding the limit judgment, bridge arm circulating current error judgment, submodule capacitor voltage deviation judgment, and backup control mode switching. When the neural network input state quantity exceeds the coverage range of the training samples, the output bridge arm voltage reference value exceeds the allowable range, the number of sub-modules deployed exceeds the limit, the bridge arm circulating current error exceeds the set threshold, or the sub-module capacitor voltage deviation exceeds the set threshold, the safety protection module performs amplitude limiting correction on the bridge arm voltage reference value, or switches the control mode to standby deadbeat model predictive control, multi-step model predictive control, or PI control mode.

Images