A MMC rectifier predictive control strategy based on artificial neural network

Abstract

The application discloses a kind of based on artificial neural network's MMC rectifier prediction control strategy, belong to power electronics technical field.The online operation of FCS-MPC will occupy a lot of microprocessor resources, cause huge computing burden.Meanwhile, since FCS-MPC is the control method based on the mathematical model of controlled object, robustness is not high.The present application is based on existing FCS-MPC control strategy, under various different conditions, collect the input and output data of FCS-MPC, train ANN offline to the input and output data collected.Then replace FCS-MPC with offline training complete ANN.The present application greatly simplifies the amount of operation compared with existing FCS-MPC, reduces the computing burden of controller.And, since MMC parameters are fluctuated in a certain range during training, the ANN has stronger inclusiveness for parameters, so the entire controller has stronger robustness.Meanwhile, the application provides a kind of ANN delay compensation strategy.

Description

Technical Field

[0001] This invention relates to the field of power electronics technology, and more specifically to an artificial neural network control method for MMC rectifiers. Background Technology

[0002] Model predictive control (MPC), especially finite control set MPC (FCS-MPC), has been widely used in power electronics since its inception. Due to its flexibility, simplicity, and ability to control multiple targets simultaneously, it is well-suited for multi-input multi-output systems such as modular multilevel converters (MMCs). Compared to traditional linear control strategies, FCS-MPC is not only simpler and more convenient to design, but also offers faster dynamic response and better output characteristics. Therefore, MMC control algorithms based on FCS-MPC have been extensively researched and applied.

[0003] However, the drawbacks of FCS-MPC are also obvious, such as the difficulty in adjusting the weighting factor and the excessive computational load. To address these issues, several improved algorithms for FCS-MPC have been proposed. These algorithms reduce the online computational load of FCS-MPC and avoid the problem of difficult weighting factor adjustment. For example, indirect FCS-MPC uses a sorting algorithm to balance the voltage of the submodule capacitors, greatly reducing the computational load; hierarchical FCS-MPC places the MMC output current and interphase circulating current in two separate cost functions, avoiding the need to adjust the weighting factor. However, these improvements do not change the core of FCS-MPC, which is to evaluate all switching combinations online to find the optimal combination that minimizes the cost function. Online computation consumes a large amount of microprocessor resources, creating a huge computational burden. Furthermore, since FCS-MPC is a predictive control method based on the mathematical model of the controlled object, its robustness is low. Therefore, when system parameters are mismatched, the control effect of FCS-MPC will be greatly reduced, affecting the overall control effect of the system.

[0004] The invention with patent number CN116203843A proposes a machine learning-based MMC model predictive control method. It utilizes a neural network to simulate MPC control of MMC, significantly reducing the computational load on the controller. Furthermore, it employs random forest technology to optimize the initial weights and thresholds of the neural network, improving its training accuracy. However, because it does not consider different actual operating conditions and parameter mismatches during training, the trained neural network lacks strong applicability and robustness, failing to achieve good control performance under various operating conditions and parameter mismatches. Simultaneously, due to the inherent characteristics of digital processors, computational delays can degrade the control performance of the neural network, potentially affecting the reliability of the entire system in severe cases. Summary of the Invention

[0005] To address the shortcomings and problems of existing technologies, this invention provides a predictive control strategy for MMC rectifiers based on Artificial Neural Networks (ANNs). Its purpose is to significantly simplify the computational load and reduce the computational burden on the controller compared to the existing FCS-MPC. Furthermore, because the system parameters are subjected to regular fluctuations during training, the ANN exhibits greater tolerance to parameter variations, thus improving the system's robustness. Simultaneously, this invention provides a delay compensation strategy for the ANN, resolving the control performance degradation problem caused by the computational delay of the digital arithmetic unit.

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

[0007] Step 1: Under various operating conditions from full load to no load, in the case of dual closed loop, the outer loop adopts a DC voltage loop, and the inner loop adopts an FCS-MPC control MMC rectifier to collect the input and output data of FCS-MPC.

[0008] Step 2: Use the input of FCS-MPC as the input of ANN and the output of FCS-MPC as the output of ANN to train ANN offline.

[0009] Step 3: Replace FCS-MPC with the offline-trained ANN and run the ANN online under the complete double-loop condition.

[0010] Specifically, in step 1, the transition from full load to no load involves starting the MMC DC-side resistance from its minimum value, indicating full power operation. Then, the resistance value is continuously increased by a fixed amount, thus continuously reducing power, until the DC side is open, indicating no load. Under various operating conditions, the grid-side voltage, DC-side voltage, bridge arm inductance, grid-side inductance, and submodule capacitor of the MMC rectifier fluctuate within a certain range above and below their rated parameters.

[0011] Specifically, in step 2, the collected FCS-MPC data is randomly divided into three categories according to a certain proportion: 70% is used as training data, 15% as testing data, and 15% as validation data. The training data is used to train the ANN, the testing data is used to observe the ANN's fit and the error rate between the output and the target output data, and the validation data is used to prevent the ANN from overfitting. Offline training uses the backpropagation (BP) algorithm to train the ANN.

[0012] Specifically, in step 3, the complete dual closed loop consists of an outer loop (DC voltage loop controlled by a proportional-integral (PI) regulator) and an inner loop (ANN).

[0013] Specifically, the delay compensation strategy of ANN is to use the number of sub-modules that are turned on in the upper and lower bridge arms output in the previous control cycle (at time k) as the input in the next control cycle (at time k+1) to form an ANN structure with delay compensation function.

[0014] In summary, the present invention achieves the following beneficial effects through the above three steps:

[0015] 1. An artificial neural network control algorithm is proposed, which has very low computational complexity, avoiding the online rolling optimization process of FCS-MPC, making programming simpler and more convenient, and suitable for computation on digital processing units such as DSPs. Furthermore, the computational load does not increase with the number of MMC submodules, solving the problem that the computational load of FCS-MPC increases dramatically with the number of submodules, making it unsuitable for practical engineering problems.

[0016] 2. Because the parameters are subjected to waveforms within a certain range during training, the ANN has a stronger tolerance for parameter variations. Therefore, even when the accuracy of the MMC parameters in the control is not high, or when parameter mismatch occurs, the ANN still exhibits good control characteristics, solving the problem of MMC control performance degradation in FCS-MPC under parameter mismatch conditions.

[0017] 3. To address the issue that the control instructions of digital arithmetic units can only be issued in the next control cycle due to the inherent characteristics of digital arithmetic units, resulting in a delay of one control cycle and thus degrading the control performance of ANNs, a delay compensation strategy for ANNs is proposed based on the basic principle of two-step prediction. Attached Figure Description

[0018] Figure 1 This refers to the topology of an MMC rectifier. Figure 2 This is a mapping diagram between FCS-MPC and ANN; Figure 3This is a hierarchical FCS-MPC control block diagram; Figure 4 A schematic diagram showing how ANN replaces FCS-MPC; Figure 5 This is a diagram of an ANN structure with delay compensation functionality; Figure 6 This is a block diagram of the overall dual closed-loop control of the MMC rectifier. Detailed Implementation

[0019] To make the basic principles, technical solutions, and advantages of this invention clearer, a detailed description of an MMC rectifier predictive control strategy based on artificial neural networks, as involved in this invention, is provided below. It should be understood that the following description is merely exemplary and not intended to limit the scope of protection or application of this invention. It should be noted that, since various FCS-MPC structures exist, the hierarchical FCS-MPC chosen in this invention is merely an example, and the use of various types of FCS-MPC is within the scope of protection of this invention.

[0020] Figure 1 This is a topology diagram of an MMC rectifier, V dc The DC side voltage, v ju and v jl Represents the voltage of the upper and lower bridge arms of phase j, i ju and i jl i represents the current in the upper and lower arms of phase j. j U represents the output current of phase j. j Represents the AC side voltage of phase j, R o and L o These are the load resistance and load inductance, respectively.

[0021] Figure 2 This illustrates the mapping relationship between FCS-MPC and ANN. An ANN consists of 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 or multi-layered. Each layer contains several neurons, and there are weight coefficients W and a threshold b between each layer. The output of each layer has an activation function σ(x). Since ANN is used to simulate and replace FCS-MPC, the input of FCS-MPC is used as the input of ANN, and the output of FCS-MPC is used as the output of ANN. Therefore, the input layer of the ANN consists of 7 neurons, representing: the output reference current i... j * (k+1), Upper bridge arm current i ju Lower bridge arm current i jl Upper arm voltage V ju Upper arm voltage V jl Grid-side voltage u j Unbalanced current reference value i * diffjThe output layer consists of two neurons, representing the number of submodules that need to be cut into the upper and lower pons arms, respectively: n uj n lj Since the number of submodules must be an integer, the round function is used here to round the ANN output layer to the nearest integer.

[0022] Figure 3 The control block diagram of the hierarchical FCS-MPC is mainly divided into three stages. The first stage is grid-side current control. By tracing all switching combinations of all upper and lower bridge arms, the number of conducting sub-modules of the upper and lower bridge arms that minimizes the AC side current control cost function is selected, which is the number of conducting sub-modules N of the upper bridge arm. uj1 (k) and the number of conductions N of the lower bridge arm submodule lj1 (k) The second stage is interphase circulating current suppression. Building upon the first stage, it involves simultaneously adding or removing δ submodules from both the upper and lower bridge arms to suppress circulating current. The third stage, based on the second stage, uses a sorting algorithm to balance the capacitor voltages of the submodules. The principles of each stage are briefly described below:

[0023] ① Grid-side current control

[0024] If the reference current on the grid side at time k+1 is i j * (k+1), given by the DC voltage loop, then the cost function for the first stage is:

[0025]

[0026] Among them, i j (k+1) is the predicted value of the grid-side current at time k+1.

[0027] ② Interphase circulation suppression

[0028] Interphase circulating current consists of DC and AC components. To reduce unnecessary losses, the AC component should be suppressed. Therefore, the reference value for interphase circulating current is:

[0029]

[0030] Among them, I dc For DC side current, P AC The active power transmitted by the MMC rectifier, V dc This is the DC-side voltage. Therefore, the cost function for the second stage is:

[0031]

[0032] Among them, i diffj (k+1) is the predicted value of the interphase circulation at time k+1, i *diffj (k+1) is the reference value for the interphase circulation at time k+1.

[0033] ③ Submodule capacitor voltage balancing

[0034] The submodule capacitor voltages are balanced using a sorting algorithm. The optimal number N of conducting submodules for the upper and lower bridge arms can be obtained from the second stage. ju_opt and N jl_opt This allows for the sequential switching in or out of submodules based on the direction of the bridge arm current. If the bridge arm current i... jx If the value is greater than 0, then the capacitor with the lowest voltage, n, will be selected. xj When a submodule is switched in, the capacitor is in a charging state; if the bridge arm current i jx If <0, then the capacitor with the highest voltage n will be... xj When a submodule is switched on, the capacitor is in a discharging state. Therefore, the switch state S... rji (k) can then be obtained and applied to the MMC rectifier.

[0035] Figure 4 To illustrate how an ANN can replace the FCS-MPC, the following steps are taken: First, input-output data from the FCS-MPC is collected under various operating conditions, from full load to no load. To improve the robustness of the ANN, the FCS-MPC needs to be run under different operating conditions and system parameters, and then the input-output data is used to train the ANN offline. According to the "Power Quality Supply Voltage Permissible Deviation" (GB12325-90), the voltage fluctuation of 10kV and below high-voltage power supply and low-voltage power is ±7%. Therefore, the grid voltage fluctuation is ±7% to ensure the ANN has the ability to resist grid voltage disturbances. Each set of parameters is randomly selected within the fluctuation range of each parameter in Table 1. Then, this data is used to train the ANN offline, making the ANN's input-output characteristics basically consistent with those of the FCS-MPC. Finally, the well-trained ANN is run online to replace the FCS-MPC.

[0036] Table 1. Fluctuation range of simulation parameters

[0037]

[0038] Figure 5This is a diagram of an ANN structure with delay compensation. The delay compensation strategy of FCS-MPC involves two-step prediction: calculating the predicted value of the control variable at time k+2 based on the switching combinations of the previous control cycle, and then minimizing the cost function to obtain the optimal switching combination. Essentially, it uses the switching combination of the previous control cycle (time k+1) as the input for the next control cycle (time k+2) for prediction. Therefore, the delay compensation of the ANN also adopts the same approach, using the switching combination of the previous control cycle (the number of submodules in the upper and lower bridge arms that are conducting) as the input of the ANN. Thus, the input layer of the ANN with delay compensation has 9 neurons, requiring not only the control variable mentioned above but also the number of conducting submodules n in the upper and lower bridge arms of the previous control cycle. uj (k-1), n lj (k-1). An ANN with delay compensation function, such as... Figure 5 As shown.

[0039] Figure 6 The diagram shows the overall control block diagram of a dual-closed-loop MMC rectifier. The inner loop (ANN) calculates the number of submodules that need to be turned on in the upper and lower bridge arms and passes this information to the sorting algorithm, thus obtaining the switching sequence of the MMC rectifier. The outer loop is the DC voltage loop, which uses a PI regulator to adjust the DC side voltage V... dc Stabilizes at the given voltage value V * dc The output of the voltage outer loop PI regulator is the d-axis current reference i. * d Based on instantaneous reactive power:

[0040]

[0041] Transforming the above equation into the dq coordinate system, the expressions for active power and reactive power are:

[0042]

[0043] If the grid voltage is directed onto the d-axis, then u q =0, the above formula can be simplified to

[0044]

[0045] Therefore, the q-axis current is given by i * d For Q * / (-1.5v d ), Q * The reactive power is given, which is set to 0 here, v d Let be the d-axis component of the grid voltage. Based on the above formula, the active power P and reactive power Q of the system can be calculated in real time. i can be calculated using the unbalanced current equation. *diffj (k+1).

[0046] The above description represents 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 in the present invention, and these variations and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A predictive control method for an MMC rectifier based on an artificial neural network, characterized in that, Under various operating conditions, input and output data of FCS-MPC are collected. The collected input and output data are used to train an ANN offline. Then, the offline-trained ANN replaces FCS-MPC, and the ANN is run online under complete double-loop conditions. This method includes the following steps: Step 1: Under various operating conditions from full load to no load, in the case of dual closed loop, the outer loop adopts DC voltage loop, and the inner loop adopts FCS-MPC to control MMC, and collects the input and output data of FCS-MPC. Step 2: Use the input of FCS-MPC as the input of ANN and the output of FCS-MPC as the output of ANN. Train ANN offline based on the input and output data of FCS-MPC. The ANN structure consists of an input layer, hidden layers, and an output layer. The input and output layers are single-layer structures, while the hidden layers are single-layer or multi-layer structures. The input layer neurons of the ANN are consistent with the input variables of FCS-MPC, and the output layer neurons of the ANN are consistent with the output variables of FCS-MPC. The round function is used to round the ANN output layer to the nearest integer. The collected FCS-MPC data is randomly divided into three categories according to the proportion: training data, detection data, and validation data. The training data is used to train the ANN, the detection data is used to observe the goodness of fit of the ANN and the error rate between the output and the target output data, and the validation data is used to prevent the ANN from overfitting. The backpropagation algorithm is used to train the ANN offline. Step 3: Replace FCS-MPC with the offline-trained ANN and run the ANN online under the complete double closed-loop condition; the ANN, as the inner loop, calculates the number of sub-modules that need to be cut into the upper and lower bridge arms and passes them to the sorting algorithm to obtain the corresponding MMC rectifier switching pulses; the outer loop is the DC voltage loop, which uses a proportional-integral regulator to stabilize the DC side voltage at the given value of the DC side voltage. Step 4: Use the number of active submodules in the upper and lower bridge arms output by the ANN in the previous control cycle as the input for the next control cycle to form an ANN structure with delay compensation function. In addition to the input variables of FCS-MPC, the input layer of the ANN with delay compensation function also needs the number of active submodules in the upper and lower bridge arms of MMC calculated by the ANN in the previous control cycle. The output layer is consistent with the output of FCS-MPC.

Images