Online Identification Method and System for Parameters of Power Electronic Converters Based on Neural Networks

Abstract

The present invention discloses an online parameter identification method and system for a power electronic converter based on a neural network, including: performing modal analysis and establishing a state equation through the state space averaging method; discretizing the state equation to obtain a discretized model of the converter; constructing a neural network based on the mutual relationship between the discretized model and parameters, and constructing a square loss function during the dynamic response process of the converter; taking the partial derivative of the loss function to obtain the gradient expressions corresponding to the inductance and capacitance parameters, and adjusting the weight values of the discretized model in each switching cycle according to the gradient descent algorithm; calculating the inductance value and capacitance value based on the discrete model and the weight values. Without adding additional circuits and sensors, the present invention can perform online identification of the parameters of dynamic components such as inductors and capacitors in a power electronic converter, which is beneficial to the life prediction of energy storage components, the state monitoring of the converter, and the optimization of control loop parameters.

Description

Technical Field

[0001] The present invention relates to the field of power electronics, and particularly to an online parameter identification method and system for a power electronic converter based on a neural network. Background Art

[0002] The application fields of power electronic converters are expanding day by day. While the performance requirements are getting higher and higher, the operating conditions of the converters are becoming more and more demanding. At present, power electronic converters are developing towards higher power density, but the heat generation problem brought by high power density is becoming more and more serious. In different application scenarios, the reliability of power electronic converters also faces different challenges, such as the random fluctuation of the equivalent load in photovoltaic applications; the random fluctuation of wind energy and the frequent switching of wind turbines in wind power generation; in the field of electric vehicles, due to the frequent acceleration and braking of the vehicle, the converter is in a non-stationary operating state for a long time, etc.

[0003] Harsh operating conditions increase the requirements for the reliability of power electronic converters, and at the same time greatly increase the probability of main circuit failures. Most of the main circuit failures are caused by device failures, which can generally be divided into structural failures and parametric failures. Structural failures refer to failures caused by complete device failure, which will bring very serious consequences and even directly lead to converter damage. Parametric failures refer to failures caused by device parameter degradation, which will reduce system performance, such as ripple size, dynamic response, and control loop performance, etc.

[0004] If parametric failures are not taken seriously, it is very likely that device parameters will degrade to complete failure, resulting in parametric failures developing into structural failures and bringing catastrophic consequences. Capacitors and inductors are key devices in the converter. Their energy charging and discharging processes enable the operating state of the converter to switch between different modes, enabling the converter to obtain corresponding outputs. In addition, changes in inductance values and capacitance values will also affect various performances of the converter. If the parameter information of inductors and capacitors in the converter can be accurately obtained, it will provide a prerequisite for device life prediction, converter state monitoring, and closed-loop control performance optimization, etc. Therefore, it is necessary to identify the inductance and capacitance parameters in the converter online.

[0005] According to the influence degree of the parameter identification method on the operation of the converter, it can be divided into an offline identification method, an intrusive online identification method, and a non-intrusive online identification method. The offline identification method does not require additional circuits, has a low cost, and can obtain a large capacitor voltage fluctuation with high identification accuracy. However, it can only perform parameter identification when the converter stops working and cannot reflect the parameter changes in real time, so it is not suitable for converters operating for a long time. The intrusive online parameter identification method refers to using a device that affects the operation of the converter or temporarily changing its operation mode during the normal operation of the converter to achieve parameter identification. It has a simple structure and high accuracy. However, it is necessary to measure the voltage and current of the capacitor simultaneously, increasing the number of sensors and the cost, and it will affect the normal operation of the converter during parameter identification. The non-intrusive online parameter identification method can calculate the capacitance value and ESR by using known electrical quantities and combining the high-frequency and low-frequency models of the capacitor, without affecting the normal operation of the converter and having strong real-time performance. However, in order to obtain the ripple information of the switching frequency, a high sampling frequency is required, and the identification accuracy may also be affected by other parasitic parameters in the circuit. For some non-intrusive methods based on intelligent algorithms, they are generally supervised learning methods, which require a large number of training samples, with a huge workload and poor generalization ability.

[0006] In addition to the above single-device parameter identification methods, it is also possible to simultaneously identify the parameters of multiple devices in the converter. According to the influence degree of the identification method on the operation of the converter, the multi-device parameter identification method can also be divided into two categories: intrusive and non-intrusive. The multi-device parameter identification method can obtain richer converter state information, which can not only provide information for state monitoring, but also guide the design and adjustment of the controller through the parameter information in the converter, improving the controller performance. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide an online parameter identification method and system for a power electronic converter based on a neural network, which can conveniently and quickly identify the parameters of multiple devices in the converter online, without adding additional circuits and sensors, and can achieve online parameter identification only relying on the electrical quantities required in the original closed-loop control, with a low cost, providing a research basis for the life prediction of energy storage elements, the state monitoring of converters, and the optimization of control loop parameters.

[0008] The technical solution adopted by the present invention to solve the above technical problem is:

[0009] On the one hand, an online parameter identification method for a power electronic converter based on a neural network includes:

[0010] S101, performing modal analysis and establishing a state equation by the state-space averaging method;

[0011] S102. Discretize the state equation to obtain the discretized model of the converter;

[0012] S103. Based on the mutual relationship between the discretized model of the converter and the parameters, construct a neural network. During the dynamic response process of the converter, use the actual sampling value as the reference value and the output value of the discrete model as the estimated value to construct a squared loss function;

[0013] S104. Take the partial derivative of the loss function to obtain the gradient expressions corresponding to the inductor and capacitor parameters. According to the gradient descent algorithm, adjust the weight values of the discretized model in each switching period;

[0014] S105. Calculate the inductor value and capacitor value based on the discrete model of the converter and the weight values.

[0015] Preferably, S101 specifically includes:

[0016] Taking the inductor current i L and the output voltage U o as state variables, the converter establishes a state equation through the state space averaging method as follows:

[0017]

[0018] where, i L (t) represents the inductor current at time t; U o (t) represents the input voltage at time t; L represents the inductor value in the converter circuit; C represents the filter capacitor value in the converter circuit; R represents the load resistance value in the converter circuit; D represents the duty cycle, and 0 ≤ D ≤ 1; t is the time variable.

[0019] Preferably, S102 specifically includes:

[0020] Arrange the state equation of the converter into the Taylor series expansion form of a matrix. Considering the first-order approximation, perform discretization to obtain the following model;

[0021]

[0022] where, i L ((k + 1)T represents the average value of the inductor current in the (k + 1)-th switching period; U o ((k + 1)T) represents the average value of the output voltage in the (k + 1)-th switching period; i L (kT) represents the average value of the inductor current in the k-th switching period; U o(kT) represents the average value of the output voltage in the k-th switching period; T represents the sampling period; L represents the inductance value in the converter circuit; C represents the filter capacitance value in the converter circuit; R represents the load resistance value in the converter circuit; D represents the duty cycle, and 0 ≤ D ≤ 1; U s (kT) represents the input voltage in the k-th switching period.

[0023] Preferably, S103 to S105 specifically include:

[0024] Taking the average value of the inductor current i L (kT), the average value of the output voltage U o (kT) and the duty cycle D as the inputs of the neural network; the dynamic process includes voltage jump, load jump, input jump and start-up phase process;

[0025] Taking the output of the discretized model of the converter as the predicted value of the electrical quantity corresponding to the (k + 1)-th switching period, and the expression of the predicted value is:

[0026]

[0027] where, i L-sim ((k + 1)T) represents the predicted value of the average value of the inductor current in the (k + 1)-th switching period in the dynamic process; U o-sim ((k + 1)T) represents the predicted value of the average value of the output voltage in the (k + 1)-th switching period in the dynamic process; U s (kT) represents the input voltage in the k-th switching period; ω i is the weight value, i ∈ [1, 4], where

[0028]

[0029] b j is the bias value, j ∈ [1, 2], where

[0030]

[0031] Constructing a squared loss function m with the error between the predicted value and the actual value as follows:

[0032] m = (i L-sim ((k + 1)T) - i L ((k + 1)T)) 2 + (U o-sim ((k + 1)T) - U o ((k + 1)T)) 2

[0033] Take the partial derivatives of the loss function \(m\) with respect to the weight values and bias values respectively. According to the chain rule, the gradients of \(m\) in the directions of inductance \(L\) and capacitance \(C\) are as follows:

[0034]

[0035] After substituting the specific expressions of the weight values and bias values, the gradients of \(m\) in the directions of inductance \(L\) and capacitance \(C\) in the discretized model are as follows:

[0036]

[0037] Use the gradient descent method to iteratively solve for \(L\) and \(C\) in each switching period. Let \(\eta\) j be the learning rate, and satisfy \(0 \lt \eta\) j \(\lt 1\). Then the iterative formulas for \(L\) and \(C\) are as follows:

[0038]

[0039] Stop the iteration when the distance \(d\) of each descent is less than the set error, as follows:

[0040]

[0041] On the other hand, a power electronic converter parameter online identification system based on a neural network includes:

[0042] A state equation establishment module, which is used to perform modal analysis and establish a state equation by the state space averaging method;

[0043] A discretized model acquisition module, which is used to discretize the state equation to obtain the discretized model of the converter;

[0044] A squared loss function construction module, which is used to construct a neural network based on the mutual relationship between the discretized model of the converter and the parameters. During the dynamic response process of the converter, taking the actual sampling value as the reference value and the output value of the discrete model as the estimated value, construct a squared loss function;

[0045] A weight value adjustment module, which is used to take the partial derivative of the loss function to obtain the gradient expressions corresponding to the inductance and capacitance parameters, and adjust the weight values of the discretized model in each switching period according to the gradient descent algorithm;

[0046] A parameter calculation module, which is used to calculate the inductance value and capacitance value based on the discrete model of the converter and the weight values.

[0047] The beneficial effects of the present invention are:

[0048] (1) The online identification method and system for power electronic converter parameters based on a neural network according to the present invention realizes high-precision parameter estimation of inductance and capacitance;

[0049] (2) The on-line parameter identification method and system of the power electronic converter based on neural network of the present invention do not require additional circuits and sensors, and can realize on-line parameter identification only by relying on the electrical quantities required in the original closed-loop control, with low cost, providing a research basis for the life prediction of energy storage elements, the state monitoring of converters, and the optimization of control loop parameters. Description of the Drawings

[0050] Figure 1 is the circuit topology diagram of the Buck converter in the embodiment of the present invention;

[0051] Figure 2 is the basic flowchart of the on-line parameter identification method of the power electronic converter based on neural network in the embodiment of the present invention;

[0052] Figure 3 is the detailed flowchart of parameter identification in the embodiment of the present invention;

[0053] Figure 4 is the simulation result diagram in the embodiment of the present invention;

[0054] Figure 5 is the structural block diagram of the on-line parameter identification system of the power electronic converter based on neural network in the embodiment of the present invention;

[0055] Among them, S represents the switch of the Buck converter circuit; D represents the diode of the Buck converter circuit; L represents the energy storage inductor of the Buck converter circuit; C represents the filter capacitor of the Buck converter circuit; R represents the load of the Buck converter circuit; U s represents the input voltage; i L represents the inductor current; U o represents the load voltage. Detailed Embodiments

[0056] The following further elaborates the present invention in conjunction with the drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, various equivalent forms of modification of the present invention fall within the scope defined by the appended claims of this application.

[0057] See Figure 1 As shown, this embodiment will take the Buck converter as an example to illustrate the parameter identification process of the inductor and filter capacitor therein.

[0058] Specifically, the Buck converter circuit is composed of a voltage source U s , a switch S, a diode D, an inductor L, a filter capacitor C, and a load resistor R. Among them, one end of the switch S is connected to the voltage source Us The positive electrode; the other end is connected to the negative electrode of the diode D and one end of the inductor L; the other end of the inductor is connected to one end of the filter capacitor C, and the other end of the filter capacitor is simultaneously connected to the positive electrode of the diode D and the negative electrode of the voltage source U s ; the load resistor R is connected in parallel with the filter capacitor C. Among them, the switching tube S can be a switching device such as a MOSFET or an IGBT.

[0059] See Figure 2 and Figure 3 As shown in, an online parameter identification method for a power electronic converter based on a neural network in this embodiment includes:

[0060] S101, perform modal analysis and establish a state equation by the state-space averaging method;

[0061] S102, discretize the state equation to obtain a discretized model of the Buck converter;

[0062] S103, based on the relationship between the discretized model of the Buck converter and the parameters, construct a neural network. During the dynamic response process of the Buck converter, use the actual sampling value as the reference value and the output value of the discrete model as the estimated value to construct a squared loss function;

[0063] S104, take the partial derivative of the loss function to obtain the gradient expressions corresponding to the inductor and capacitor parameters, and adjust the weight values of the discretized model in each switching period according to the gradient descent algorithm;

[0064] S105, calculate the inductor value and capacitor value based on the discrete model of the Buck converter and the weight values.

[0065] Specifically, taking the inductor current i L and the output voltage U o as state variables, the converter establishes a state equation by the state-space averaging method as follows:

[0066]

[0067] In the formula, i L (t) represents the inductor current at time t; U o (t) represents the input voltage at time t; L, C, and R are the inductor value, filter capacitor value, and load resistor value in the Buck converter circuit respectively; D is the duty cycle and 0 ≤ D ≤ 1; t is the time variable.

[0068] Furthermore, organize the state equation of the Buck converter into a matrix form and the discretized model obtained is;

[0069]

[0070] Among them, i L ((k + 1)T represents the average value of the inductor current in the (k + 1)-th switching period; U o ((k + 1)T) represents the average value of the output voltage in the (k + 1)-th switching period; i L (kT) represents the average value of the inductor current in the k-th switching period; U o (kT) represents the average value of the output voltage in the k-th switching period; T represents the sampling period; U s (kT) represents the input voltage in the k-th switching period.

[0071] The construction of the neural network and the calculation of the inductor value L and the filter capacitor value C include the following steps:

[0072] Take the average value i of the inductor current in the k-th switching period during the dynamic process of the Buck converter L , the average value U of the output voltage o and the duty cycle D as the inputs of the neural network; the dynamic process includes processes such as voltage jump, load jump, input jump, start-up phase, etc.

[0073] Take the output of the discrete model of the Buck converter circuit as the predicted value of the electrical quantity corresponding to the (k + 1)-th switching period. The expression of the predicted value is:

[0074]

[0075] Among them, i L-sim ((k + 1)T) represents the predicted value of the average value of the inductor current in the (k + 1)-th switching period during the dynamic process; U o-sim ((k + 1)T) represents the predicted value of the average value of the output voltage in the (k + 1)-th switching period during the dynamic process; U s (kT) represents the input voltage in the k-th switching period; ω i is the weight value, i ∈ [1, 4], where:

[0076]

[0077] b j is the bias value, j ∈ [1, 2], where:

[0078]

[0079] Construct a squared loss function m with the error between the predicted value and the actual value as follows:

[0080] m = (i L-sim ((k + 1)T) - i L ((k + 1)T)) 2 + (Uo-sim ((k + 1)T) - U o ((k + 1)T)) 2

[0081] Take the partial derivatives of the loss function m with respect to the weight value and the bias value respectively. According to the chain rule, the gradients of m in the directions of the inductor L and the capacitor C are obtained as follows:

[0082]

[0083] After substituting the specific expressions of the weight value and the bias value, the gradients of m in the directions of the inductor L and the capacitor C in the Buck circuit model can be obtained as

[0084]

[0085] Use the gradient descent method to iteratively solve for L and C in each switching period. Let η j be the learning rate, and satisfy 0 < η j < 1. Then the iterative formulas for L and C are as follows:

[0086]

[0087] Stop the iteration when the distance d of each descent is less than the set error ε, that is

[0088]

[0089] In this embodiment, the simulation parameters of the Buck converter circuit are shown in Table 1. The input voltage U s = 60V, the inductor L = 2mH, the filter capacitor C = 1000 μF, the load resistor R = 50, the switching frequency f s is 10 kHz, the sampling period T = 0.00001 s, the output voltage U o = 30V. At the moment of 0.1 s in the simulation time, the output voltage jumps to 40V. And during the period when the output voltage jump starts and reaches the steady state again, the online identification of the inductor and capacitor parameters is carried out according to the above steps.

[0090] Table 1

[0091]

[0092]

[0093] See Figure 4 shown, for the simulation results of parameter identification. Before 0.1 s, the output voltage U o= 30V, the output voltage jumped at 0.1s, and the parameter identification program also ran simultaneously. The inductance value L finally converged to approximately 1.977 mH, and the filter capacitor C finally converged to approximately 1000.7 μF. The errors of the two were approximately 1.15% and 0.07% respectively. It can be seen that this method has high identification accuracy. Moreover, this method does not require additional circuits and sensors, does not need to collect a large amount of data in advance, and can perform online parameter identification while the circuit is running.

[0094] See Figure 5 As shown, according to another aspect of the present invention, this embodiment also discloses an online parameter identification system for a power electronic converter based on a neural network, including:

[0095] A state equation establishment module 501, configured to perform modal analysis and establish a state equation by the state space averaging method;

[0096] A discretized model acquisition module 502, configured to perform discretization processing on the state equation to obtain a discretized model of the converter;

[0097] A squared loss function construction module 503, configured to construct a neural network based on the mutual relationship between the discretized model of the converter and the parameters, and during the dynamic response process of the converter, use the actual sampled value as the reference value and the output value of the discrete model as the estimated value to construct a squared loss function;

[0098] A weight value adjustment module 504, configured to take the partial derivative of the loss function to obtain the gradient expressions corresponding to the inductance and capacitance parameters, and according to the gradient descent algorithm, adjust the weight values of the discretized model in each switching cycle;

[0099] A parameter calculation module 505, configured to calculate the inductance value and capacitance value based on the discretized model of the converter and the weight values.

[0100] The specific implementation of an online parameter identification system for a power electronic converter based on a neural network is the same as that of an online parameter identification method for a power electronic converter based on a neural network, and this embodiment will not be repeated here.

[0101] The above is only a specific implementation manner of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantive modification made to the present invention using this concept shall fall within the scope of infringement of the protection of the present invention.

Claims

1. An online parameter identification method for power electronic converters based on neural networks, characterized in that, Including: S101, perform modal analysis and establish a state equation by the state - space averaging method; S102, discretize the state equation to obtain a discretized model of the converter; S103, based on the mutual relationship between the discretized model of the converter and the parameters, construct a neural network. During the dynamic response process of the converter, taking the actual sampling value as the reference value and the output value of the discretized model as the estimated value, construct a squared - loss function; S104, take the partial derivative of the loss function to obtain the gradient expressions corresponding to the inductance and capacitance parameters. According to the gradient - descent algorithm, adjust the weight values of the discretized model in each switching period; S105, calculate the inductance value and capacitance value based on the discretized model of the converter and the weight values; The S103 - S105 specifically include: Taking the average inductor current \(i\) L (kT), the average output voltage \(U\) o (kT), and the duty cycle \(D\) in the \(k\)-th switching period during the dynamic process of the converter as the inputs of the neural network; the dynamic process includes voltage jump, load jump, input jump, and startup phase process; Take the output of the discretized model of the converter as the predicted value of the electrical quantity corresponding to the (k + 1)-th switching period; Construct a squared - loss function m with the error between the predicted value and the actual value; Take the partial derivatives of the loss function m with respect to the weight value and the bias value respectively. According to the chain rule, the gradients of m in the directions of inductance L and capacitance C are: where ω i is the weight value, i ∈ [1, 4], b j is the bias value, j ∈ [1, 2]; After substituting the specific expressions of the weight value and the bias value, the gradients of m in the directions of inductance L and capacitance C in the discretized model are: Among them, i L-sim ((k + 1)T) represents the predicted value of the average inductor current in the (k + 1)-th switching period during the dynamic process; U o-sim ((k + 1)T) represents the predicted value of the average output voltage in the (k + 1)-th switching period during the dynamic process; U s (kT) represents the input voltage in the k-th switching period; i L ((k + 1))T represents the average inductor current in the (k + 1)-th switching period; U o ((k + 1)T) represents the average output voltage in the (k + 1)-th switching period; T represents the sampling period; R represents the load resistance value in the converter circuit; Iteratively solve for L and C in each switching period using the gradient descent method, and let η j be the learning rate, and satisfy 0 < η j < 1. Then the iterative formulas for L and C are as follows: Stop iteration when the distance d of each descent is less than the set error, as follows:

2. The on-line parameter identification method of the power electronic converter based on neural network according to claim 1, characterized in that The S101 specifically includes: Taking the inductor current i L and the output voltage U o as state variables, the converter establishes the state equation by the state-space averaging method as follows: where i L (t) represents the inductor current at time t; U o (t) represents the input voltage at time t; t is the time variable.

3. The on-line parameter identification method of the power electronic converter based on neural network according to claim 1, characterized in that The S102 specifically includes: Arrange the state equation of the converter into the Taylor - series expansion form of a matrix. Considering the first - order approximation and performing discretization, the obtained model is as follows; 4. The online parameter identification method of a power - electronic converter based on a neural network according to claim 1, characterized in that The expression of the predicted value is: Where The squared - loss function m is as follows: m = (i L-sim ((k + 1)T) - i L ((k + 1)T)) 2 + (U o-sim ((k + 1)T) - U o ((k + 1)T)) 2 。 5. An on-line parameter identification system for a power electronic converter based on a neural network, characterized in that, Based on the online parameter identification method of a power - electronic converter based on a neural network according to any one of claims 1 - 4, including: A state - equation establishment module, used to perform modal analysis and establish a state equation by the state - space averaging method; A discretized - model acquisition module, used to discretize the state equation to obtain a discretized model of the converter; A squared - loss - function construction module, used to construct a neural network based on the mutual relationship between the discretized model of the converter and the parameters. During the dynamic response process of the converter, taking the actual sampling value as the reference value and the output value of the discretized model as the estimated value, construct a squared - loss function; A weight - value adjustment module, used to take the partial derivative of the loss function to obtain the gradient expressions corresponding to the inductance and capacitance parameters. According to the gradient - descent algorithm, adjust the weight values of the discretized model in each switching period; A parameter - calculation module, used to calculate the inductance value and capacitance value based on the discretized model of the converter and the weight values.

Images