Machine Learning-Based MMC Model Predictive Control Method and System

Abstract

The present invention relates to a machine learning-based MMC model predictive control method and system. First, data is collected using an MPC-MMC simulation platform and preprocessed, and then neural network training is performed to obtain a neural network-MPC controller. To improve the neural network training efficiency, random forest is used to optimize the initial weight threshold of the neural network. Finally, a random forest-neural network-MPC controller is obtained to simulate the MPC controller. The results show that RF-NN-MPC is superior to NN-MPC in terms of learning efficiency and learning accuracy; while maintaining good control effects, MPC-MMC is not restricted by the number of sub-modules, and the online calculation amount is always 1 time. The calculation amount is greatly reduced, which is suitable for engineering applications.

Description

Technical Field

[0001] The present invention belongs to the technical field of power engineering. Specifically, it relates to a model predictive control method for a modular multilevel converter with low computational complexity based on machine learning. Background Technique

[0002] Due to the modular design of the modular multilevel converter (MMC), it is very easy to expand its voltage level and power level, and it has been widely used in the grid connection of renewable energy power generation such as wind power and photovoltaic power generation. However, when the MMC is applied to high-voltage and high-power scenarios, the number of sub-modules in the arm reaches hundreds, and the number of sub-module switching signals to be controlled also increases accordingly, bringing a great computational load to the controller and affecting the operation efficiency of the MMC system. Summary of the Invention

[0003] In view of the above problems, the present invention designs a model predictive control (MPC) method for MMC based on machine learning. First, the simulation data of the established MPC-MMC control system is preprocessed, and then the NN-MPC controller is obtained through neural network training, which greatly reduces the computational complexity of the MMC controller and facilitates engineering applications. At the same time, the initial weights and thresholds of the neural network are optimized by the random forest technology, improving the training efficiency and training accuracy of the neural network.

[0004] The technical solution provided by the present invention is as follows:

[0005] A model predictive control method for MMC based on machine learning, characterized by comprising the following steps:

[0006] The MPC-MMC simulation platform collects data and preprocesses the data;

[0007] The data is input into a neural network to train a neural network-MPC controller (Neural Network-MPC, NN-MPC), and a random forest is used to optimize the initial weights and thresholds of the neural network to obtain a random forest-neural network-MPC controller;

[0008] The obtained random forest-neural network-MPC controller is used to simulate the output data of the MPC controller.

[0009] In the above model predictive control method for MMC based on machine learning, relevant data is sampled from the MPC-MMC model, and a set of excellent data sets are obtained through normalization and VMD decomposition operations; Normalization processing:

[0010] .

[0011] In the above-mentioned MMC model predictive control method based on machine learning, variable mode decomposition VMD operation is adopted, and the AC side current variational model is as follows:

[0012] .

[0013] In the above-mentioned MMC model predictive control method based on machine learning, the design of random forest-neural network-MPC controller; the weights and thresholds between the input layer and the hidden layer, and between the hidden layer and the output layer of 11 electrical physical quantities are ω ij , θ ij and ω jk , θ jk The data is recorded as Q1 as the training set of random forest.

[0014] In the above-mentioned MMC model predictive control method based on machine learning, when the sample data set has M characteristic attributes, r characteristic attributes are randomly selected to train N sample subsets; the Gini coefficients of the r characteristic attributes are calculated respectively, and the expression is as follows:

[0015]

[0016] Arrange the feature attributes in order from small to large according to the Gini coefficient, select the best feature attribute from the r feature attributes as the split node of the forest, so that it can grow as much as possible without pruning, forming a decision tree and then a forest;

[0017] Then import the test set data, calculate the output values of all decision trees respectively, and use the mean of all output values as the final output of the random forest; use the root mean square error δ MSE To evaluate the output of random forest;

[0018] .

[0019] In the above-mentioned MMC model predictive control method based on machine learning, each set of data in the training set consists of 11 electrical physical quantities, 11×11×2 ω ij , θ ij value, 11×2×2 value ω jk , θ jk Composition: Extract N sets of data from Q1 with replacement m times to form training subsets N1, N2…Nm Regarding the permutation and combination number of 11 electrical physical quantities as the characteristic attributes of the sample.

[0020] In the above-mentioned MMC model predictive control method based on machine learning, the data collected by the MPC-MMC simulation platform specifically includes:

[0021] Obtaining the number of sub-modules put into the upper and lower bridge arms that meet the phase current tracking at the current moment , ;

[0022] Compensating the same number of levels for the upper and lower bridge arms to make the number of finally put-in sub-modules , Meet the tracking of the AC side current and the suppression balance condition of the inter-phase circulating current.

[0023] In the above-mentioned MMC model predictive control method based on machine learning, the number of finally put-in sub-modules , Meets: , ; is the optimal compensation level number for simultaneous compensation of the upper and lower bridge arms.

[0024] In the above-mentioned MMC model predictive control method based on machine learning, The determination steps are:

[0025] Single-step prediction: Construct a single-step model predictive circulating current control value function, adopt the five-level compensation method to obtain the optimal, sub-optimal, and sub-sub-optimal compensation level numbers;

[0026] Multi-step prediction: Construct a multi-step model predictive circulating current control value function, and obtain the optimal compensation level number according to the compensation level numbers obtained from the single-step prediction.

[0027] A machine learning-based MMC model predictive control system, characterized by including

[0028] The first module: The MPC-MMC simulation platform collects data and preprocesses the data;

[0029] The second module: Input the data into the neural network to train the neural network-MPC controller (Neural Network-MPC, NN-MPC), and use the random forest to optimize the initial weight threshold of the neural network to obtain the random forest-neural network-MPC controller;

[0030] The third module: Use the obtained random forest-neural network-MPC controller to simulate the output data of the MPC controller.

[0031] The present invention has the following advantages: 1. The MMC model predictive control method based on machine learning effectively avoids the core feature of the traditional MPC improved algorithm: the switching signal that requires detailed online evaluation to minimize the objective function. Through machine learning, a large amount of data calculation is effectively avoided, greatly reducing the time required for system calculation and improving the operating efficiency of the MMC system. It neither affects the grid-side current tracking nor suppresses the internal circulating current of the bridge arm, and at the same time effectively avoids the large amount of calculation in the sorting combination of sub-module capacitor voltages. Only through 1 calculation, the optimal compensation level number is obtained, greatly reducing the calculation amount and being applicable to engineering applications. 2. The proposed random forest-neural network effectively solves the problem of low learning efficiency of traditional neural networks. By optimizing a large number of initial values and corresponding thresholds and weights, the initial values of the better GOOD region and BEST region of the neural network controller are obtained, improving the network training accuracy, improving the training efficiency, and reducing the workload of workers. Description of the Drawings

[0032] Figure 1 MMC topology

[0033] Figure 2 Random forest-neural network-MPC control strategy diagram

[0034] Figure 3 Random forest optimized weight threshold diagram

[0035] Figure 4 Random forest-neural network-structure diagram

[0036] Figure 5 Improved multi-step model predictive control schematic diagram Specific implementation method

[0037] The design of the MMC model predictive controller based on machine learning mainly includes the following 3 steps.

[0038] The first step is to construct the MPC-MMC system. The schematic diagram of the MMC main circuit is as Figure 1 shown

[0039] In step (1.1), through single-step model predictive circulating current suppression control, a single-step compensation level set is obtained. First, establish a single-step model predictive function for the circulating current. The schematic diagram of the MMC is as Figure 1 shown, and its differential equation is as follows:

[0040] (1)

[0041] (2)

[0042] where e j is the converter output phase voltage，u uj ，u lj are the output voltages of the upper and lower bridge arms respectively u sj is the grid-side voltage i j is the output phase current. Analyzing the external characteristic equation (1) of the MMC, it can be seen that the output phase current of the converter is related to the difference between the upper and lower bridge arm voltages. From formula (1), we can get t k the number of sub-modules inserted into the upper and lower bridge arms that satisfy phase current tracking at time , . Analyzing the internal characteristic equation (2) of the MMC, it can be seen that the current inside the bridge arm is related to the difference between the DC voltage and the sum of the upper and lower bridge arm voltages. First, the present invention obtains the number of sub-modules inserted into the upper and lower bridge arms that satisfy phase current tracking through formula (1) , , and then by the method of compensating the same number of levels on the upper and lower bridge arms simultaneously, the goal of neither affecting phase current tracking nor suppressing arm circulating current is achieved

[0043] Performing a first-order forward difference on formula (2), the single-step model prediction function (3) of the circulating current is obtained

[0044] (3)

[0045] where , are respectively t k+1 the upper and lower bridge arm voltages that satisfy phase current tracking at time

[0046] Then, construct the single-step model prediction circulating current control value function to obtain its optimal, sub-optimal, and sub-sub-optimal compensation level numbers

[0047] The internal current of each phase is composed of the inter-phase circulating current and one-third of the DC current. In order to suppress the circulating current as much as possible, the reference value of the internal current of each phase is .

[0048] (4)

[0049] The present invention adopts five-level compensation, and the number of compensation levels is denoted as , and the compensation voltage is denoted as . Define the single-step model prediction circulating current control, the upper and lower bridge arm level compensation sets , . To reduce the calculation amount and maintain the phase current tracking accuracy, the number of compensation levels of the upper and lower bridge arms needs to be equal, and the compensation voltage is

[0050] (5)

[0051] Then select the number of compensation levels corresponding to the minimum, the second minimum, and the third minimum of and denote them as d 1, d 2, d 3 respectively. Among them d 1 is the number of compensation levels corresponding to the optimal single-step circulating current prediction, d 2 is the number of compensation levels corresponding to the sub-optimal single-step circulating current prediction, d 3 is the number of compensation levels corresponding to the third-best single-step circulating current prediction.

[0052] In step (1.2), through multi-step model prediction of circulating current suppression control, the final optimal number of compensation levels is obtained. First, define D 2 as the compensation set of the circulating current control level for multi-step model prediction. From the first step, we know that D 2 = d 1 d 2 d 3]. The number of compensation levels predicted by the multi-step model is denoted as , and the compensation voltage , and its value is:

[0053] (6)

[0054] Then, construct the value function of the multi-step model prediction of circulating current control to obtain the final optimal number of compensation levels. The schematic diagram of the improved multi-step model prediction control is shown in Figure 2 .

[0055] Perform a first-order forward difference on formula (2). Based on the first step, with a sampling period of T s， we can obtain t k+2 the predicted value of the internal current at time

[0056] (7)

[0057] , are respectively t k+2 the input voltages of the upper and lower bridge arms that satisfy the phase current tracking at time

[0058] Construct the value function of the multi-step model prediction of circulating current control

[0059] (8)

[0060] Select the number of compensation levels corresponding to the minimum value of and apply it to the converter At this moment, the corresponding compensation level is the optimal compensation level Therefore, the number of upper and lower levels that can satisfy both output AC tracking and circulating current suppression are (9) and (10) respectively.

[0061] (9)

[0062] (10)

[0063] in e a is the converter output phase voltage, U dc is the DC side voltage, u pa ，u na are the output voltages of the upper and lower bridge arms respectively, u sa is the grid-side voltage, i sa is the output phase current, R is the AC side resistance, L is the AC side inductance, L arm is the bridge arm inductance, i diffa is the internal unbalanced current, T s is the sampling time; i sa ( t + T s )、 u zna ( t + T s )、 u zpa ( t + T s )and u sa ( t + T s ) are t + T s AC side current, AC side voltage, upper bridge arm voltage and lower bridge arm voltage at all times. u dc ( t + T s )、 i cira ( t + T s )for t + T s DC side voltage and circulating current at all times; V diffk is the compensation voltage; J k1 and Jk2 respectively represent the objective function of the AC-side current and the objective function of the inter-phase circulating current; i sa * ( t + T s ) and i cira * ( t + T s ) are respectively t + T s the reference values of the AC-side current and the inter-phase circulating current at time

[0064] In summary, the present invention finds the optimal compensation level number with only 8 calculations, realizes multi-step model predictive circulating current suppression, not only ensures the phase current tracking accuracy, but also effectively suppresses the circulating current, and is applicable to engineering applications.

[0065] In the second step, relevant data is collected from the above MPC-MMC control system, and data preprocessing is performed on the data set.

[0066] First, normalization processing is performed:

[0067] (11)

[0068] In the formula x i represents any column data value of the data set.

[0069] Then, during the experiment, it is observed that there are harmonic interferences with different frequencies in the AC-side current i ac , which interfere with the training of the neural network. Common filtering operations will affect the i ac amplitude and phase. And in the MMC system, the tracking of the AC-side current directly affects the tracking of active power and reactive power. Introducing variational mode decomposition (VMD) can solve the problem of high harmonic content and avoid the situation where the AC-side current cannot be tracked due to amplitude and phase changes.

[0070] VMD decomposes the original signal i ac into K modal functions u k ( t ), then i ac can be expressed as:

[0071] (12)

[0072] wherein i ac * represents the fundamental frequency component of the AC-side current i ac ; u kn ( t ) represents the nth mode function

[0073] The expression of the AC-side current variational model is as follows

[0074] (13)

[0075] In the formula is the set of each mode component is the set of the center frequencies corresponding to each mode δ ( t ) is the Dirac function represents shifting the analytical signal spectrum of each mode u k (t) to the baseband

[0076] Initialize the Lagrange multiplier λ , penalty factor α and cyclic operator n , and de-constrain the constrained variational problem. The expression is as shown in Equation (14):

[0077] (14) Use the alternating direction method of multipliers to iteratively update each decomposed mode , center frequency and Lagrange multiplier . Their expressions are as shown in Equations (15) and (16):

[0078] (15)

[0079] (16)

[0080] Repeat the steps of Equations (15) and (16) until the convergence condition (17) is satisfied

[0081] (17)

[0082] Then, through the inverse Fourier transform, the AC-side current value i ac * of the fundamental wave data can be obtained

[0083] The third step: Design of the random forest-neural network-MPC controller

[0084] This design will use the control strategy Figure 2In: AC side voltage u ac , AC side current i ac , DC side voltage u dc , AC side current reference value i ref * , DC side current i dc , upper and lower bridge arm currents i p and i n , upper and lower bridge arm average voltages u p and u n , upper and lower bridge arm total voltages u zp and u zn are used as the inputs of the neural network, and the number of upper and lower bridge arm sub-modules output n pj and n nj are used as the outputs of the neural network.

[0085] Since there are certain defects in the process of neural network training. For example, poor selection of initial weights and thresholds will cause errors in the output results of the neural network and slow network learning speed. The random forest algorithm can solve this problem by optimizing the initial weights. The principle of random forest optimizing neural network is as Figure 3 shown, the process of optimizing the weights and thresholds of the NN network is as Figure 3 shown by the green line in. The initial weights and thresholds have the possibility in all regions. Assuming that the initial weights and thresholds are selected in the Bad and Not Bad regions, the network is continuously trained through the gradient descent method until it reaches the Good region and finally stops training in the Best region. This process requires a large amount of time to calculate; the RF-NN network (shown by the blue line in the figure) classifies all weights through decision trees and limits the initial weight thresholds in the Good and Best regions. When the initial weight thresholds are in the Good region, only a certain amount of time and iteration times are required to obtain the optimal weights and thresholds. This will greatly improve the efficiency of network training.

[0086] The weights and thresholds between the input layer and the hidden layer, and between the hidden layer and the output layer of the 11 electrical physical quantities ω ij , θ ij , ω jk , θ jkThe data is used as the training set of the random forest and denoted as Q1. That is, each set of data in the training set consists of 11 electrical physical quantity values, 11×11×2 ω ij , θ ij values, 11×2×2 values ω jk , θ jk and so on. m times of N groups of data sets are drawn with replacement from Q1 to form a training subset (N1, N2…N m ). The permutation and combination number of 11 electrical physical quantities is regarded as the characteristic attribute of the sample. When there are M characteristic attributes in this sample data set, r characteristic attributes are randomly selected from them to train N sample subsets. The Gini coefficients of r characteristic attributes are calculated respectively, and the expression is as follows:

[0087] (15)

[0088] where P r represents the probability that the sample belongs to the r th group of characteristic attributes. The smaller the Gini coefficient, the greater the probability of belonging to this group of characteristic attributes; otherwise, the probability is smaller.

[0089] These characteristic attributes are arranged in ascending order of the Gini coefficient. The optimal characteristic attribute is selected from r characteristic attributes as the splitting node of the forest, so that it grows as much as possible without pruning, forming a decision tree and then a forest.

[0090] Then the test set data is imported, and the output values of all decision trees are calculated respectively. The mean value of all output values is used as the final output of the random forest. The root mean square error δ MSE is used to evaluate the output value of the random forest.

[0091] (16)

[0092] In the formula, s is the number of samples; out is the output of the random forest; out* is the optimal output of the expected weight and threshold. If the mean square error value does not meet the expectation, r* characteristic attributes are reselected, and the above operations are repeated until the requirements are met. The output of the random forest is used as the initial weight and threshold of the neural network.

[0093] The finally formed random forest - neural network - MPC controller is as Figure 4 shown.

[0094] The following is a specific case, mainly including:

[0095] Step 1: First, determine the structure of the neural network controller and collect corresponding data samples in the MPC-MMC control system. The structure of the neural network consists of an input layer, a hidden layer, and an output layer. The 11 input nodes of the input layer respectively correspond to 11 electrical physical quantities output by the MMC. The number of hidden layer nodes is selected as 11 based on experience. The number of output layer nodes is 2, which are the AC side voltage u ac , the AC side current i ac , the DC side voltage u dc , the AC side current reference value i ref * , the DC side current i dc , the upper and lower arm currents i p and i n , the upper and lower arm average voltages u p and u n , the upper and lower arm total voltages u zp and u zn . The number of output layer nodes is 2, which are the number of sub-modules of the upper and lower arms n pj and n nj .

[0096] Step 2: Preprocess the original sample set. First, normalize the data on the Matlab experimental platform, and then import it into the Python 3.7 simulation platform to build a virtual environment. Decompose the above sample set by VMD. After VMD decomposition and filtering, the THD value of the A-phase AC side current changes from the original 0.63% to 0.26%, and the harmonic content is greatly reduced, obtaining a better sample data set.

[0097] Step 3: 1) The optimal weights and thresholds between the input layer and the hidden layer, and between the hidden layer and the output layer output by the random forest ω ij best 、ω jk best 、θ ij best , θ jk bestAs the initial weights and thresholds between neurons in each layer of the neural network. 2) Propagate the NN1 samples forward in sequence to calculate the weights and thresholds between neurons in each layer. 3) Calculate the output error. If the requirement is not met, backpropagate the error for correction ω ij 、 θ ij and ω jk 、 θ jk ; if the requirement is met, continue with the next group in 2). 4) If the NN1 samples in the dataset are not exhausted, return to 2); otherwise, calculate the network MSE value. If the value of MSE meets the requirement, the neural network training ends; if not, return to 1) to retrain the network.

[0098] Step 4: Apply the trained random forest-neural network-MPC controller to the MMC system.

[0099] In summary, the present invention finds the optimal compensation level number with only 1 calculation, realizes the MMC model predictive control method based on machine learning, not only ensures the phase current tracking accuracy, but also effectively suppresses the circulating current. The capacitor voltage of the sub-module can fluctuate within a stable range and is suitable for engineering applications.

[0100] The present invention also includes an MMC model predictive control system based on machine learning, including

[0101] The first module: The MPC-MMC simulation platform collects data and preprocesses the data;

[0102] The second module: Input the data into the neural network to train the neural network-MPC controller (Neural Network-MPC, NN-MPC), and use the random forest to optimize the initial weight thresholds of the neural network to obtain the random forest-neural network-MPC controller;

[0103] The third module: Use the obtained random forest-neural network-MPC controller to simulate the output data of the MPC controller.

[0104] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Those skilled in the art to which the present invention pertains can make various modifications or supplements to the described specific embodiments or use similar methods for substitution, without departing from the spirit of the present invention or exceeding the scope defined by the appended claims.

Claims

1. A machine learning-based MMC model predictive control method, characterized in that, Including the following steps: The MPC-MMC simulation platform collects data and preprocesses the data, including: sampling relevant data from the MPC-MMC model, and obtaining a set of data sets through normalization and VMD decomposition operations; and performing normalization processing on the data sets: In the formula x i represents any column data value of the data set; The process of the VMD decomposition operation includes: adopting the variational mode decomposition (VMD) operation, and the AC side current variational model is as follows: where u k ( t ) represents the nth modal function, is the set of center frequencies corresponding to each mode; δ ( t ) is the Dirac function, means shifting the analytic signal spectrum of each mode u k (t) to the baseband; The data is input into the neural network training to obtain the neural network-MPC controller, and the random forest is used to optimize the initial weight threshold of the neural network to obtain the random forest-neural network-MPC controller; the design of the random forest-neural network-MPC controller; the weights and thresholds between the input layer and the hidden layer, and between the hidden layer and the output layer of 11 electrical physical quantities are calculated. ω ij , θ ij and ω jk , θ jk The data is recorded as Q1 as the training set of random forest; The training process of the neural network-MPC controller includes: when the sample data set has M feature attributes, randomly select r of them to train N sample subsets; calculate the Gini coefficients of the r feature attributes respectively, and the expression is as follows: where P r represents the probability that the sample belongs to the r group of characteristic attributes; Arrange the feature attributes in ascending order of the Gini coefficient, select the optimal feature attribute from the r feature attributes as the splitting node of the forest, and make it grow as much as possible without pruning to form a decision tree and then form a forest; Then import the test set data again, calculate the output values of all decision trees respectively, and use the mean value of all output values as the final output of the random forest; use the root mean square error δ MSE to evaluate the output value of the random forest; where n is the number of samples; out is the output of the random forest; is the optimal output of the expected weights and thresholds; Use the obtained random forest-neural network-MPC controller to simulate the output data of the MPC controller.

2. The method for predicting and controlling the MMC model based on machine learning according to claim 1, characterized in that, Each set of data in the training set consists of 11 electrical physical quantity values, 11×11×2 ω ij , θ ij values, 11×2×2 values ω jk , θ jk ; m times of N sets of data sets are drawn with replacement from Q1 to form training subsets N1, N2…N m , and the number of permutations and combinations of 11 electrical physical quantities is regarded as the characteristic attribute of the sample.

3. The machine learning-based MMC model predictive control method according to claim 2, characterized in that, The specific data collection by the MPC-MMC simulation platform includes: Obtain the upper and lower bridge arm input sub-module bases that meet the phase current tracking at the current moment , ; Compensate the same number of levels for the upper and lower bridge arms to make the number of finally inserted sub-modules , meet the tracking of the AC-side current and the suppression and balance conditions of the inter-phase circulating current.

4. The method for predicting and controlling the MMC model based on machine learning according to claim 3, characterized in that, The number of sub - modules finally invested , satisfies: , ; is the optimal compensation level number for simultaneous compensation of the upper and lower bridge arms.

5. The method for predicting control of the MMC model based on machine learning according to claim 4, wherein The determination steps are as follows: Single-step prediction: construct a single-step model to predict the circulating current control value function, and adopt the five-level compensation method to obtain the optimal, sub-optimal, and sub-sub-optimal compensation level numbers; Multi-step prediction: construct a multi-step model to predict the circulating current control value function, and obtain the optimal compensation level number according to the compensation level number obtained by the single-step prediction.

6. A machine learning-based MMC model predictive control system, which is used to implement the machine learning-based MMC model predictive control method according to claim 1, and is characterized in that, Including The first module: The MPC-MMC simulation platform collects data and preprocesses the data; The second module: Input the data into the neural network to train the neural network-MPC controller (Neural Network-MPC, NN-MPC), and use the random forest to optimize the initial weight threshold of the neural network to obtain the random forest-neural network-MPC controller; The third module: Use the obtained random forest-neural network-MPC controller to simulate the output data of the MPC controller.

Images