Neural network prediction control method for modular multilevel matrix converter

Abstract

The invention discloses a neural network prediction control method for a modular multilevel matrix converter, and belongs to the field of power electronics. The M3C is a power electronic topology for realizing direct alternating current-alternating current electric energy conversion, and has the advantages of modularization, flexible control, high electric energy quality and the like. A traditional FCS-MPC needs to traverse all possible switch state combinations in a control period, huge calculation burden is caused, meanwhile, the control precision of the FCS-MPC depends on the accuracy of a mathematical model, and robustness is poor. The invention provides a neural network predictive control method for M3C, and the method comprises the steps: collecting the input and output data of an FCS-MPC controller under different working conditions, carrying out the offline training of a neural network through the collected data, replacing the FCS-MPC with the trained network, and carrying out the real-time control, and also provides a delay compensation scheme of the method. According to the method, the online calculation burden of the digital controller is greatly reduced, and disturbance of M3C system parameters is introduced in the training process of the neural network, so that the neural network prediction controller has higher adaptability to parameter changes, and the robustness of the control system is further improved.

Description

Technical Field

[0001] This invention relates to the field of power electronics technology, and more specifically to a neural network predictive control method for a modular multilevel matrix converter. Background Technology

[0002] The Modular Multilevel Matrix Converter (M3C) is a power electronic topology capable of direct AC-AC power conversion. It has attracted widespread attention due to its modularity, flexible control, and high power quality. This topology has broad application prospects in medium- and high-voltage fields such as frequency division transmission, offshore wind power conversion, power system interconnection, and AC motor drives.

[0003] The unique structure of the M3C system results in highly coupled internal electrical quantities, posing technical challenges for control. The introduction of dual αβ0 coordinate transformation decouples the M3C's input, output, and circulating currents, allowing the use of multiple proportional-integral (PI) controllers. However, the use of multiple controllers complicates the control loop structure, and adjusting the control parameters is a highly complex task. In contrast, Model Predictive Control (MPC), especially Finite Control Set MPC (FCS-MPC), offers flexibility, simplicity, and fast response, enabling simultaneous adjustment of multiple targets and making it well-suited for multi-input multi-output systems like the M3C.

[0004] Even though FCS-MPC has many advantages in M3C applications, it still suffers from drawbacks such as high computational cost, difficulty in adjusting weight factors, and poor robustness. To address these issues, various optimization algorithms for FCS-MPC have been proposed. For example, indirect FCS-MPC removes the submodule capacitor voltage balance control from the cost function and implements it using an external sorting algorithm, reducing computational cost while avoiding weight factor adjustment. Long-time-domain sequential FCS-MPC extends the prediction time domain to multiple future control cycles, changing only the switching state of one submodule in each bridge arm in each control cycle, further reducing computational cost. However, these optimization algorithms do not improve the core problem of FCS-MPC: online evaluation of all possible switching state combinations to find the optimal switching sequence that minimizes the cost function. This online optimization process is the fundamental reason for the high computational cost of FCS-MPC; the computational burden increases dramatically with the number of submodules. Meanwhile, FCS-MPC is a predictive algorithm based on a mathematical model. Its control accuracy is highly dependent on the accuracy of the mathematical model. When the system and the mathematical model parameters do not match, it will affect the control effect of FCS-MPC, and in severe cases, it may even threaten the stable operation of the M3C system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a neural network predictive control method for modular multilevel matrix converters. Its purpose is to reduce the computational burden on the controller compared to the existing FCS-MPC method, eliminating the need for weight factors and online optimization processes, and ensuring that the computational load does not increase with the number of sub-modules. By introducing random fluctuations in system parameters during training, the neural network exhibits stronger parameter adaptability, thereby improving the system's robustness. Furthermore, this invention addresses the inherent computational delay problem of digital controllers such as DSPs by providing a delay compensation scheme based on a neural network predictive controller, thus improving the system's steady-state performance.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] Step 1: Use the dual closed-loop structure of average capacitor voltage control and active / reactive power control outer loop and FCS-MPC inner loop to control M3C, and change the active power of the system from 0 to large to make M3C operate under different power levels, and collect the input and output data of FCS-MPC controller.

[0008] Step 2: Use the collected FCS-MPC data to train the neural network offline, where the network's input and output variables correspond one-to-one with the input and output variables of FCS-MPC;

[0009] Step 3: Replace the FCS-MPC in the control system with a trained neural network, and combine it with a delay compensation scheme to achieve real-time control of the M3C using a neural network predictive controller.

[0010] Specifically, in step 1, the M3C system's input side achieves average capacitor voltage control through a closed-loop algorithm and power feedforward compensation, while the output side achieves active and reactive power control. The outer loop output serves as the current reference input for the inner loop FCS-MPC controller. The inner loop generates the number of sub-modules to be engaged and balances the capacitor voltages of the sub-modules through a sorting algorithm. Under this control structure, the active power of the M3C is gradually adjusted to ensure that the M3C operates within its full range of operating points from no-load to rated power. At different power levels, a certain range of perturbations are applied to the AC voltage, input / output inductance, bridge arm inductance, and sub-module capacitors of the M3C system based on their rated values. Input and output data of the FCS-MPC controller are collected, including input voltage, output voltage, bridge arm current and its reference value, input current and its reference value, output current and its reference value, and the number of engaged sub-modules.

[0011] Specifically, in step 2, the collected data is divided into a training set and a test set according to a certain ratio. The neural network model is trained using the training set, employing the backpropagation algorithm. Before training, the data is normalized to improve model accuracy. During training, a loss function is defined to measure the error between the network's predicted and actual values. The error gradient of each neuron is calculated starting from the output layer and backpropagated to the previous layer, thereby updating the model's weights, biases, and other parameters. The test set data is used to verify the model's training effect.

[0012] Specifically, in step 3, the trained neural network replaces the FCS-MPC and forms a complete M3C control system with the outer loop controller. Furthermore, to compensate for the computational delay caused by the digital controller, the number of input sub-modules output from the previous control cycle is used as the input for the current cycle, thus constructing a neural network predictive controller with delay compensation functionality.

[0013] In summary, the present invention achieves the following beneficial effects through the above three steps:

[0014] 1. A neural network predictive control method for M3C is provided. This method does not require the online rolling optimization process in the FCS-MPC method. It can directly calculate the optimal number of input sub-modules based on the input variables. The computational load is small and does not increase with the number of sub-modules, making it very suitable for practical engineering.

[0015] 2. By subjecting the system parameters to a certain degree of fluctuation during the training of the neural network model, the trained network becomes more adaptable to parameter changes, thereby improving the robustness of the M3C system.

[0016] 3. To address the inherent computational delay problem of digital controllers, a delay compensation scheme based on a neural network predictive controller is proposed, which improves the control performance of M3C, according to the two-step forward prediction idea. Attached Figure Description

[0017] Figure 1 The topology of a modular multilevel matrix converter; Figure 2 The diagram shows the indirect FCS-MPC control structure of the M3C. Figure 3 Flowchart of the sub-module sorting and equalization algorithm; Figure 4 The diagram shows the predictive control structure of a neural network. Figure 5 This is a schematic diagram of the neural network predictive control implementation in M3C. Figure 6 This is the overall control block diagram of M3C. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the following description is merely exemplary and not intended to limit the scope of use and protection of the present invention. It should be noted that since M3C has various FCS-MPC methods and various neural network types, the selection of indirect FCS-MPC and artificial neural networks in this invention is merely illustrative; the use of various types of FCS-MPC and neural networks is within the scope of protection of this invention.

[0019] Figure 1 Here is a topology diagram of M3C, where u a u b u c i represents the input grid voltage. a i b i c For the input-side grid current, u u u v u w i represents the output grid voltage. u i v i w For the output side grid current, L s For the input-side inductance, L r For the output-side inductance, u com This represents the common-mode voltage. Two three-phase AC systems are connected via nine bridge arms, each arm consisting of N full-bridge sub-modules (SMs) and one bridge arm inductor L. b Composition, i xy(x=a, b, c; y=u, v, w) represents the current flowing through the bridge arm xy. Each full-bridge SM consists of 4 switching devices and 1 capacitor in parallel, which can be used as an energy storage unit.

[0020] Figure 2 The diagram below shows the indirect FCS-MPC control structure of M3C. The principle of this method is briefly described below.

[0021] First, taking the bridge arm as a unit, a dynamic model of M3C is established, which can be described as:

[0022]

[0023] Ignoring common-mode voltage, a discrete prediction model for the M3C is established using the forward Euler discretization method:

[0024]

[0025] Among them, T s For the sampling period, u xy (k) = n xy (k)*u* C, where u* C is the rated voltage of the capacitor.

[0026] To minimize the current tracking error on the input / output side, based on the deadbeat principle, i x (k+1) and i y (k+1) can be replaced by i* x(k+1) and i* y(k+1) respectively.

[0027] The reference value of the bridge arm current at the next moment is defined as:

[0028]

[0029] Where i* xyz(k+1) is the circulating current reference value of bridge arm xy, which is used to balance the submodule capacitor voltage between bridge arms under special operating conditions, and is set to 0 here.

[0030] The cost function is constructed as follows:

[0031]

[0032] For all possible n xy (n xy Rolling optimization is performed within the range [-N, N] to select the n that minimizes the cost function J. xy As the output of FCS-MPC, denoted as N xy .

[0033] Figure 3The flowchart illustrates the submodule voltage equalization algorithm. After determining the optimal number of submodules to be deployed in the bridge arm, voltage equalization within the bridge arm can be achieved by combining the capacitor voltage sorting algorithm. The specific implementation process is as follows: First, the capacitor voltages of the N submodules are sorted by voltage boosting. When N... xy When i > 0, if i xy If the value is >0, it indicates that the bridge arm is in a charging state, then the lower capacitor voltage will be applied to the first |N... xy | Each submodule is put into positive operation; conversely, the one with the higher capacitor voltage is put into negative operation. | N xy | Each submodule is actively being implemented; when N xy When i < 0, if i xy If the value is greater than 0, it indicates that the bridge arm is in a discharging state. Therefore, the capacitor with the higher voltage (|N) will be connected to the next higher voltage level. xy | Sub-modules are negatively fed in; conversely, the first N modules with lower capacitor voltages are fed in. xy Each sub-module is subject to negative investment.

[0034] Figure 4 This is a diagram of a neural network predictive control structure. Taking an Artificial Neural Network (ANN) as an example, an ANN consists of three parts: an input layer, hidden layers, and an output layer. The input and output layers are single-layer structures, while the hidden layers can be single-layer or multi-layered, each consisting of several neurons. The neurons in the hidden and output layers each have a specific activation function, and there are weight coefficients and biases between layers. Since this invention uses a neural network to simulate and replace FCS-MPC, the input of FCS-MPC is used as the network's input, and the output of FCS-MPC is used as the network's output. Therefore, the ANN input layer consists of 8 neurons, representing: bridge arm current i xy Bridge arm current reference value i* xy, input current i x Input current reference value i*x, output current i y Output current reference value i*y, input voltage u x Output voltage u y The output layer consists of one neuron, representing the number N submodules that need to be added to the pontine arm. xy Since the direct output of an ANN is not necessarily an integer, and the sorting algorithm requires that the number of input submodules must be an integer, the round function is used to round the output of the ANN model to the nearest integer.

[0035] Figure 5This diagram illustrates the implementation of neural network predictive control in the M3C system. First, under FCS-MPC control, the active power is gradually increased to allow the M3C to operate under different conditions. Input and output data of the FCS-MPC at different power levels are collected, and this data is used to train the neural network offline. Simultaneously, during data collection, the system parameters are subjected to a certain degree of random perturbation to enhance the network's parameter adaptability. Specifically, the input grid-side voltage fluctuates randomly by 10% above and below its rated value, and the input / output inductors, bridge arm inductors, and submodule capacitors each fluctuate randomly by 20% above and below their rated values. This allows the neural network trained using the collected data to better withstand grid fluctuations and adapt to parameter mismatches when system parameters are disturbed. Finally, the trained network replaces the FCS-MPC inner loop in the control system, enabling online operation of the neural network predictive controller. Furthermore, to enable delay compensation, one neuron is added to the input layer, representing the optimal number of submodules to be engaged in the previous sampling period. Therefore, the neural network with delay compensation has nine input layer neurons and one output layer neuron.

[0036] Figure 6 This is the overall control block diagram of M3C. The average capacitor voltage control and output active / reactive power control of M3C form the outer loop. The output of the outer loop PI controller serves as the reference value for the input and output currents. The neural network predictive controller serves as the inner loop, receiving the reference value from the outer loop and calculating the number of submodules that need to be put into operation for each bridge arm online. Then, the switching sequence of each submodule of M3C is obtained through the capacitor voltage sorting algorithm.

[0037] The formula for calculating the average capacitor voltage is:

[0038]

[0039] Among them, u Cxyi This represents the capacitor voltage of the i-th submodule of bridge arm xy.

[0040] Instantaneous active and reactive power are:

[0041]

[0042] Transforming the above equation to the dq coordinate system, we obtain the expressions for active and reactive power:

[0043]

[0044] The above description illustrates a specific embodiment and advantages of the present invention, but the scope of protection of the present invention is not limited thereto. For those skilled in the art, variations and modifications can be made to the above embodiments without fundamentally departing from the technical spirit and principles described herein, and these variations and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A neural network predictive control method of a modular multilevel matrix converter, characterized by, The modular multilevel matrix converter (MFC) is controlled using the FCS-MPC method. Input and output data of the FCS-MPC controller are collected under different power levels. The collected data is used to train the neural network offline. The trained network is then fed into the control system, replacing the inner loop of the FCS-MPC, thus enabling the online operation of the neural network predictive controller. By using the output of the previous control cycle as the input of the current cycle, a neural network predictive controller with delay compensation function is constructed. The specific steps include: Step 1: Use the dual closed-loop structure of average capacitor voltage control and active / reactive power control outer loop and FCS-MPC inner loop to control M3C, and change the active power of the system from 0 to large to make M3C operate under different power levels, and collect the input and output data of FCS-MPC controller. Step 2: Use the collected FCS-MPC data to train the neural network offline, where the network's input and output variables correspond one-to-one with the input and output variables of FCS-MPC; Step 3: Replace the FCS-MPC in the control system with a trained neural network, and combine it with a delay compensation scheme to achieve real-time control of the M3C using a neural network predictive controller.

2. The method of claim 1, wherein, In step 1, the M3C system input side achieves average capacitor voltage control through a closed-loop algorithm and power feedforward compensation, while the output side achieves active and reactive power control. The outer loop output serves as the current reference input for the inner loop FCS-MPC controller. The inner loop generates the number of sub-modules to be engaged and balances the capacitor voltages of the sub-modules through a sorting algorithm. Under this control structure, the active power of the M3C is gradually adjusted so that the M3C operates within the full range of operating points from no-load to rated power. At different power levels, a certain range of disturbances are applied to the AC voltage, input / output inductance, bridge arm inductance, sub-module capacitors, and other parameters of the M3C system based on their rated values. Input and output data of the FCS-MPC controller are collected, including input voltage, output voltage, bridge arm current and its reference value, input current and its reference value, output current and its reference value, and the number of sub-modules engaged.

3. The method according to claim 1, characterized in that, In step 2, the collected data is divided into training set and test set according to a certain ratio. The training set is used to train the neural network model. The training method adopts the backpropagation algorithm. Before training, the data is normalized to improve the accuracy of the model. During the training process, a loss function is defined to measure the error between the predicted value and the actual value of the neural network. The error gradient of each neuron is calculated from the output layer and backpropagated back to the previous layer to update the model's weights, biases and other parameters. Use test set data to validate the training effect of the model.

4. The method according to claim 1, characterized in that, In step 3, the trained neural network replaces the FCS-MPC and forms a complete M3C control structure with the outer loop controller. In addition, to compensate for the computational delay caused by the digital controller, the number of input sub-modules output in the previous control cycle is used as the input of the current cycle to form a neural network predictive controller with delay compensation function.

Images