Flying capacitor voltage prediction method based on neural network

Abstract

The application discloses a flying capacitor voltage prediction method based on a neural network, and comprises the following steps: S10, collecting the switch signal, output voltage and real value of the flying capacitor voltage of a flying capacitor multilevel converter (FCMC) in the process of running from no load to full load through a model predictive controller (MPC); S20, constructing a double-layer feedforward neural network, recording defined input vectors and output vectors, wherein the input vectors comprise the switch signal, output voltage at the current moment and the estimated value of the flying capacitor voltage at the previous moment, and the output vectors are the estimated value of the flying capacitor voltage at the current moment; S30, repeating S10 and S20 with different flying capacitor voltages as initial conditions; S40, performing offline training on the neural network by using a back propagation (BP) algorithm, optimizing the network weight and bias, and making the network output approximate to the real voltage of the flying capacitor in operation; and S50, deploying the trained neural network to an FPGA controller, and combining a model predictive control (MPC) algorithm to realize online real-time estimation of the flying capacitor voltage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power electronics, and particularly relates to a flying capacitor voltage prediction method based on a neural network. BACKGROUND

[0002] The flying capacitor topology has good scalability. This technology enables low-cost and small-power semiconductor switches to be effectively applied to larger power occasions and better performance indicators are obtained. However, compared with the traditional two-level topology, the flying capacitor multi-level topology needs to additionally adjust the flying capacitor voltage. No matter linear control or nonlinear control strategy is adopted, to realize active balancing of the flying capacitor voltage, a necessary prerequisite is to accurately obtain the capacitor voltage value in real time. This requirement leads to a sharp increase in the number of sensors in the system.

[0003] Specifically, for an n-level flying capacitor converter, x(n-2) flying capacitor voltage sensors are needed for each phase, where x represents the number of phases. With the development of power electronic converters towards high frequency and high power density, higher requirements are put forward for the quality and performance of the sensors. The increase in the number of sensors not only leads to an increase in cost and size, but also may cause hidden dangers to the reliability of the system.

[0004] In order to reduce the dependence on sensors, the existing solution is to approximately obtain the flying capacitor voltage value by numerical approximation method. However, these methods usually have limited convergence speed and insufficient accuracy, and cannot well adapt to the high frequency demand of multi-level converters. For example, G. Farivar et al. realized the balance of the flying capacitor voltage by using a single voltage sensor, but this method is limited to a working frequency of 1 kHz. R. P. Aguilera et al. realized the estimation of the four-level flying capacitor voltage by using Kalman filter, and the working frequency was increased to 10 kHz. Professor Miguel F. Escalante's team used Petri Nets (PNs) to predict the actual value of each flying capacitor voltage, and achieved success in a five-level FCMC, with a working frequency of 20 kHz. However, the convergence speed and accuracy of these methods still need to be improved.

[0005] In view of the above problems, it is particularly important to study a technology capable of quickly and accurately observing the flying capacitor voltage. SUMMARY

[0006] The main purpose of the present application is to provide a flying capacitor voltage prediction method based on a neural network, to solve the problems of low observation accuracy, slow convergence speed and large number of sensors in the prior art.

[0007] To achieve the above object, the present application provides the following technical solutions, comprising the following steps:

[0008] S10, collecting, by a model predictive controller MPC, a switching signal, an output voltage and a real value of a flying capacitor voltage of a flying capacitor multi-level converter FCMC in a process of running from no load to full load;

[0009] S20, constructing a double-layer feedforward neural network, and recording a defined input vector and an output vector, wherein the input vector comprises the switching signal, the output voltage and an estimated value of the flying capacitor voltage at a previous moment, and the output vector is an estimated value of the flying capacitor voltage at the current moment;

[0010] S30, repeating S10 and S20 with different capacitor voltages as initial conditions;

[0011] S40, performing offline training on the neural network by using a back propagation algorithm BP, and optimizing network weights and biases, so that the network output approximates to a real voltage of the flying capacitor running at this time;

[0012] S50, deploying the trained neural network to an FPGA controller, and combining a model predictive control MPC algorithm to online and real-time estimate the flying capacitor voltage, and replacing a traditional voltage sensor.

[0013] In a possible implementation, the S10 specifically comprises:

[0014] In a simulation environment, a FCMC model is established, and a load current dynamic equation of the FCMC is represented as:

[0015]

[0016] wherein the current i is an inherent state variable, v N is a value of the output voltage, R is a resistance value of the load, L is an inductance value of the load, and a first-order Euler approximation is used to obtain:

[0017]

[0018] wherein T is a sampling period, and k is a moment, the FCMC is controlled by using the MPC, and bus voltage, flying capacitor voltage, flying capacitor capacitance and load parameters of the FCMC are recorded.

[0019] In a possible implementation, the S20 specifically comprises:

[0020] The current output voltage, the capacitor voltage at the previous moment and the switching signal are used as independent variables, the estimated value of the current capacitor voltage is used as a dependent variable, and a mapping relationship therebetween is shown in formula (3):

[0021]

[0022] where v k is the current capacitor voltage estimate, s k-1 is the switching signal, v k-1 is the previous capacitor voltage estimate, is the previous output voltage.

[0023] To this end, a neural network model is established to learn the operation data of the multi-level converter, so that the neural network approximates the mapping relationship of formula (3), as shown in formula (4). A two-layer BP network structure is adopted, and the output layer directly outputs the capacitor voltage estimate, and the number of nodes is equal to the number of flying capacitors,

[0024]

[0025] where I k is the input vector of the neural network, H k is the output vector of the hidden layer of the neural network, W [1] and W [2] are the weight matrices of the hidden layer and the output layer respectively, b [1] and b [2] are the bias vectors of the hidden layer and the output layer respectively, g [1] and g [2] are the activation functions of the hidden layer and the output layer respectively. Considering that the capacitor voltage is a positive real number, and the value range of the Relu function is consistent, the Relu is selected as the activation function g [2] of the output layer; the activation function g [1] of the hidden layer of the neural network adopts tanh, and the value range is (-1, 1) about the origin center symmetry.

[0026] Record the sample and the corresponding label at each time point, where, is the switching signal, is the output voltage, is the previous capacitor voltage estimate.

[0027] In a possible implementation, the S30 specifically includes:

[0028] The initial capacitor voltage value selection includes typical values of the capacitor voltage balance and unbalance states of the flying capacitor multi-level inverter in the normal operation range. By selecting multiple groups of representative initial values, rich operation data is obtained to improve the accuracy and adaptability of the neural network for flying capacitor voltage estimation.

[0029] In a possible implementation, the S40 specifically includes:

[0030] The collected samples are used to train the neural network, and the obtained sample data is divided into a training set, a validation set and a test set; the training set is used for preliminary parameter adjustment and learning of the neural network; the validation set is used to monitor the performance of the network during the training process to prevent overfitting; and the test set is used to finally evaluate the generalization ability of the trained network.

[0031] During the training process, the accuracy of the network is evaluated periodically using the validation set. In the continuous training and evaluation, the hyperparameters of the neural network are constantly adjusted until the accuracy of the network on the validation set reaches a pre-set threshold.

[0032] The present application has at least the following beneficial effects:

[0033] (1) Through simulation and experimental verification, the method of the present application has high prediction accuracy. For example, in some embodiments, the prediction accuracy of the method can reach more than 99%, which can meet the accuracy requirements of flying capacitor voltage prediction in actual applications. And the working frequency can reach 100kHz, which means that the prediction and estimation of flying capacitor voltage can be completed in a very short time, which helps to enhance the stability and reliability of the system, and avoids system instability or failure caused by capacitor voltage fluctuation.

[0034] (2) The method of the present application has universality and scalability, and can be applied to flying capacitor multi-level topology with different number of levels, so it can be applied to various sizes and types of power electronic converters, providing strong support for the development of power electronic technology. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The step flow chart of the flying capacitor voltage prediction method based on neural network of an embodiment of the present application;

[0036] Figure 2 The general topology structure of an n-level flying capacitor converter applied in an embodiment of the present application;

[0037] Figure 3 The principle diagram of the capacitor voltage estimator (neural network) based on neural network in an embodiment of the present application;

[0038] Figure 4 The connection diagram of a flying capacitor five-level inverter for analyzing the estimation accuracy of the neural network in a specific application example;

[0039] Figure 5 The connection diagram of a flying capacitor five-level inverter for verifying the implementation effect of the neural network in the control loop in another specific application example. DETAILED DESCRIPTION

[0040] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described, obviously, the described embodiments are a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0041] Referring to Figure 1 , a step flow chart of a flying capacitor voltage prediction method based on a neural network is shown, comprising the following steps:

[0042] S10, collecting the switching signal, output voltage and real value of the flying capacitor voltage of the flying capacitor multilevel converter FCMC in the process of running from no load to full load by the model predictive controller MPC;

[0043] S20, constructing a double-layer feedforward neural network, recording the defined input vector and output vector, wherein the input vector comprises the switching signal, output voltage at the current time and the estimated value of the flying capacitor voltage at the previous time, and the output vector is the estimated value of the flying capacitor voltage at the current time;

[0044] S30, repeating S10 and S20 with different flying capacitor voltages as initial conditions;

[0045] S40, using the back propagation algorithm BP to perform offline training on the neural network, optimizing the network weight and bias, so that the network output approximates the real voltage of the flying capacitor running at this time;

[0046] S50, deploying the trained neural network to the FPGA controller, combining the model predictive control MPC algorithm to estimate the flying capacitor voltage online and in real time, replacing the traditional voltage sensor.

[0047] In a specific application example, as shown in Figure 2 , a structure schematic diagram of a n-level flying capacitor multilevel converter in the specific application example is shown, which is composed of n-1 switching units in series, each unit containing a pair of complementary semiconductor switching devices S and corresponding flying capacitor C, the flying capacitor C i is connected between the switching nodes i and i+1. Under the balance condition, the voltage on each flying capacitor can be expressed as:

[0048]

[0049] In the simulation environment, the FCMC model is established, and the load current dynamic equation of the FCMC is expressed as

[0050]

[0051] where i is the inherent state variable. v N is the value of output voltage, R is the resistance value of load, L is the inductance value of load. Using the first-order Euler approximation, we get

[0052]

[0053] where T is the sampling period, k is the time. Using MPC to control flying capacitor converter FCMC, record the bus voltage, flying capacitor voltage, flying capacitor capacitance and load parameters of FCMC.

[0054] Further, in S20, referring to the schematic diagram of Figure 3 , the current output voltage, the capacitor voltage at the previous time, the switch signal as the independent variable, the current capacitor voltage estimation value as the dependent variable, the mapping relationship between them is shown in equation (3):

[0055]

[0056] where v k is the current capacitor voltage estimation value, s k-1 is the switch signal, v k-1 is the capacitor voltage estimation value at the previous time, is the output voltage at the previous time;

[0057] To this end, a neural network model is established to learn the operation data of the multi-level converter, so that the neural network approximates the mapping relationship of equation (3). As shown in equation (4), the neural network adopts a two-layer BP network structure, and the output layer directly outputs the capacitor voltage estimation value, and the number of nodes is equal to the number of flying capacitors.

[0058]

[0059] where I k is the input vector of neural network, H k is the output vector of neural network hidden layer, W [1] and W [2] are the weight matrix of hidden layer and output layer respectively, b [1] and b [2] are the bias vector of hidden layer and output layer respectively, g [1] and g [2] are the activation function of hidden layer and output layer respectively. Considering that the capacitor voltage is a positive real number, and the value range of Relu function is consistent, therefore, Relu is selected as the activation function g [2] of output layer; the activation function g [1] of neural network hidden layer adopts tanh, and its value range is (-1, 1) and is central symmetric about the origin;

[0060] Record the sample at each time and the corresponding label wherein, is a switching signal, is an output voltage, is a previous time capacitor voltage estimate.

[0061] In a specific application example, in S30, the initial value of the capacitor voltage is selected to include but not limited to the typical values of the flying capacitor multilevel inverter in the normal operating range that may occur in the balanced and unbalanced state of the capacitor voltage, such as [50V, 100V, 150V]. By selecting multiple sets of representative initial values, rich operating data is obtained to improve the accuracy and adaptability of the neural network for flying capacitor voltage estimation;

[0062] Then, in S40, the neural network is trained using the collected samples, and the obtained sample data is divided into a training set, a validation set, and a test set. The training set is used for preliminary parameter adjustment and learning of the neural network; the validation set is used to monitor the performance of the network during training to prevent overfitting; and the test set is used to evaluate the generalization ability of the trained network.

[0063] During training, the accuracy of the network is evaluated periodically using the validation set. The evaluation indicators of accuracy can be selected according to the specific application scenario, such as accuracy (Accuracy), recall (Recall), F1 value, mean square error (MSE), root mean square error (RMSE), etc. In continuous training and evaluation, the hyperparameters of the neural network (such as learning rate, batch size, number of iterations, etc.) are constantly adjusted until the accuracy of the network on the validation set reaches a pre-set threshold that can meet the actual application requirements. The setting of this threshold should consider factors such as application scenario, performance requirements, cost constraints, etc.

[0064] In a specific application example, in order to further verify the practicality of the proposed neural network flying capacitor voltage prediction technology, a five-level FCMC prototype machine is designed and built for simulation verification, as shown in Figure 4 and Figure 5 , Figure 4 is a connection diagram of a flying capacitor five-level inverter used to analyze the estimation accuracy of the neural network in a specific application example; Figure 5 is a connection diagram of a flying capacitor five-level inverter used to verify the implementation effect of the neural network in the control loop in another specific application example.

[0065] 1. System parameters: bus voltage 200V, flying capacitor C1-C3 capacitance 200μF, reference voltage 50V / 100V / 150V, load 5Ω+1mH.

[0066] 2. Data acquisition: MPC control converter operated under a sinusoidal load current with amplitude 6-18 A, 10,000 sets of data recorded, covering the range of capacitor voltage 0-170 V.

[0067] 3. Network training: input layer 8 nodes (4 switching signals + 3 historical voltages + 1 node voltage), hidden layer 15 nodes, output layer 3 nodes; training time 120 s, test set MSE = 0.12 V 2 .

[0068] 4. Performance verification:

[0069] Steady-state accuracy: at 100 kHz, V c2 estimated error 0.35 V (actual value of sensor 100 V);

[0070] Dynamic response: V c2 reference value jumps from 100 V to 120 V, tracking time 1.5 ms, total harmonic distortion (THD) of current decreases from 2.1% to 1.8%.

[0071] It should be understood that the example embodiments described herein are illustrative and non-limiting. Although one or more embodiments of the present application are described in conjunction with the attached figures, it is understood that various alterations, modifications, and variations can be made thereof, and it is intended to cover all such alterations, modifications, and variations, which fall within the scope of the present application as defined by the appended claims.

Claims

1. A neural network-based flying capacitor voltage prediction method, characterized by, The method comprises the following steps: S10, collecting, by a model predictive controller MPC, real values of switching signals, output voltages and capacitor voltages of a flying capacitor multilevel converter FCMC during no-load to full-load operation; S20, constructing a double-layer feedforward neural network, and recording defined input vectors and output vectors, wherein the input vectors comprise the switching signals, the output voltages and the estimated value of the flying capacitor voltage at the previous moment, and the output vectors are the estimated value of the flying capacitor voltage at the current moment; S30, repeating S10 and S20 with different capacitor voltage initial conditions; S40, performing offline training on the neural network by using a back propagation algorithm BP, optimizing network weights and biases, and making the network output approximate to the real voltage of the flying capacitor in operation; S50, deploying the trained neural network to an FPGA controller, and combining a model predictive control MPC algorithm to online and real-time estimate the flying capacitor voltage, replacing a traditional voltage sensor; The S20 specifically comprises: The current output voltage, the capacitor voltage at the previous moment and the switching signals are taken as independent variables, and the estimated value of the current capacitor voltage is taken as a dependent variable, and a mapping relationship therebetween is shown in formula (3): (3) wherein is a current capacitor voltage estimate, is a switching signal, is a previous capacitor voltage estimate, is a previous output voltage; To this end, a neural network model is established to learn the operation data of the multilevel converter, and the neural network is caused to approximate the mapping relationship of formula (3), as shown in formula (4), the neural network adopts a two-layer BP network structure, the output layer directly outputs the estimated value of the capacitor voltage, and the number of nodes thereof is equal to the number of the flying capacitors, (4) wherein is an input vector of the neural network, is an output vector of the neural network hidden layer, and are weight matrices of the hidden layer and the output layer, respectively, b [1] and b [2] are bias vectors of the hidden layer and the output layer, respectively, g [1] and g [2] are activation functions of the hidden layer and the output layer, respectively, considering that the capacitance voltage is a positive real number and the value range of the Relu function is consistent, the Relu is selected as the activation function g [2] of the output layer; [1] the activation function g of the neural network hidden layer adopts the tanh, whose value range is (-1, 1) and is central symmetric about the origin. record the sample I at each time instant k [ , , , , , , , ] T and the corresponding label v k [ , , ] T wherein, is a switching signal, is an output voltage, is a previous time instant capacitor voltage estimate.

2. The neural network-based flying-capacitor-voltage prediction method of claim 1, wherein, The S10 specifically comprises: An FCMC model is established in a simulation environment, and a load current dynamic equation of the FCMC is represented as: (1) where the current is an intrinsic state variable, is the value of the output voltage, R is the value of the load resistance, L is the value of the load inductance, using a first order Euler approximation: (2) wherein, is the sampling period, k is the time, using MPC control flying capacitor converter FCMC operation, record the bus voltage of FCMC, flying capacitor voltage, flying capacitor capacity and load parameters.

3. The neural network-based flying-capacitor-voltage prediction method of claim 1, wherein, The S30 specifically comprises: The initial value of the capacitor voltage is selected to include typical values of the capacitor voltage in balanced and unbalanced states in a normal operation range of the flying capacitor multilevel inverter, a plurality of groups of representative initial values are selected, rich operation data are obtained, and the accuracy and adaptability of the neural network in estimating the flying capacitor voltage are improved.

4. The neural network-based flying-capacitor-voltage-forecasting method of claim 1, wherein, The S40 specifically comprises: The neural network is trained by using the collected samples, the obtained sample data are divided into a training set, a verification set and a test set, the training set is used for preliminary parameter adjustment and learning of the neural network, the verification set is used for monitoring the performance of the network during the training process to prevent overfitting, and the test set is used for finally evaluating the generalization ability of the trained network; During the training process, the accuracy of the network is evaluated by using the verification set regularly, the hyperparameters of the neural network are continuously adjusted until the accuracy of the network on the verification set reaches a pre-set threshold value in the continuous training and evaluation.

Images