Bidirectional DC / DC Converter Parameter Identification Method Based on Physical Information Machine Learning

Abstract

The present invention provides a method for parameter identification of a bidirectional DC / DC converter based on physical information machine learning, and relates to the technical field of parameter identification. The method first completes the establishment of a system dynamics model of a bidirectional DC / DC converter, and then uses an implicit Runge-Kutta method to construct the relationship between the data intermediate state and the initial state and the final state, and applies it to the system dynamics model, and then inputs the historical training data into a deep neural network to obtain the intermediate state of the data, and substitutes it into the physical model as input to achieve seamless coupling of the physical model to the training of the deep neural network. Then the trained PINN is used to achieve online parameter identification. This hybrid design can seamlessly combine the data of the BDDC physical model, effectively modulate the training phase, and force it to be enforced to the underlying physical principles and constraints of the converter, greatly reducing the required training data.

Description

Technical Field

[0001] The present invention belongs to the technical field of parameter identification, and in particular relates to a bidirectional DC / DC converter parameter identification method based on physical information machine learning. Background Art

[0002] Since it is difficult for a single energy source to meet the requirements of high energy density and high power density at the same time in the energy storage system of electric vehicles, it is necessary to use high energy density lithium batteries and high power density supercapacitors for reasonable matching. Under the cooperation of energy management strategy, reasonable power distribution is carried out for batteries and supercapacitors to give full play to the energy characteristics of batteries and the power characteristics of supercapacitors. While saving energy, it can also meet the requirements of electric vehicles for endurance and power performance. Due to the different charging and discharging characteristics of the two energy sources, in order to reduce the pressure of high-power charging and discharging of batteries, it is necessary to add a bidirectional DC / DC converter to the composite energy storage system to control and adjust the charging and discharging current and input and output voltage between batteries and supercapacitors to optimize the working coordination between the two. The use of two bidirectional DC / DC converters can achieve more effective and controllable power distribution for batteries and supercapacitors, and effectively stabilize the DC bus voltage.

[0003] When a bidirectional DC / DC converter works in different application environments for a long time, its key components such as electrolytic capacitors, inductors, insulated gate bipolar transistors, metal oxide semiconductor field effect transistors, etc., under the combined effects of external environmental stress (such as operating temperature, humidity, vibration and radiation, etc.) and internal electrical stress, the parameter values ​​of the components will gradually deviate from the normal tolerance range and cause soft faults. In the early stage of soft faults, they are often ignored because the fault characteristics are weak and difficult to detect; as the deviation of components gradually increases, when the parameter deviation exceeds the failure threshold of the components, there will be a huge deviation between the actual circuit and the theoretical model, which will affect the control strategy effect of the system and the accuracy of the fault diagnosis method, resulting in misdiagnosis. Therefore, it is of great significance to perform parameter identification on electronic devices in bidirectional DC / DC converters. Summary of the invention

[0004] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a bidirectional DC / DC converter parameter identification method based on physical information machine learning. The method can train the neural network with fewer training samples, effectively reducing the computational burden and enhancing the prediction transparency.

[0005] Specifically, the present invention provides a method for identifying parameters of a bidirectional DC / DC converter based on physical information machine learning, the method comprising the following steps:

[0006] S1. Constructing a system dynamics model of a bidirectional DC / DC converter, wherein the bidirectional DC / DC converter is a converter for controlling and regulating the charge and discharge current and input and output voltage between a battery and a supercapacitor;

[0007] S2. Using the implicit Runge-Kutta method, the relationship between the intermediate state of the data and the initial state and the final state is constructed, and the relationship is applied to the system dynamics model to solve the unknown quantity therein;

[0008] S3, constructing a physical information neural network, which consists of two parts: a data-driven part and a physical model composed of a deep neural network; wherein the physical model uses the solution method of step S2 to obtain the predicted value of the physical information neural network according to the output of the data-driven part;

[0009] Using the parameters to be identified in the bidirectional DC / DC converter as parameters of the data driving part in the physical information neural network, and training all parameters in the physical information neural network by back propagation method;

[0010] S4. Using the trained physical information neural network to realize online parameter identification of the bidirectional DC / DC converter.

[0011] As a further explanation of the present invention, the system dynamics model building process of step S1 specifically includes the following steps:

[0012] S11. General system dynamics model:

[0013] ut+N[u;λ]=0,x∈Ω,t∈[0,T] (1)

[0014] f=ut+N[u;λ] (2)

[0015] Where u = u (x, t) is the solution of the differential equation, x is the spatial coordinate, t is the time coordinate;

[0016] S12. Construct a system dynamics model of the bidirectional DC / DC converter based on the system dynamics model of S11:

[0017]

[0018] Where λ = {L 1 , L 2 , R 1 , R 2 , C o} is the parameter of the component to be identified;

[0019] In the formula, R 1 , L 1 、i l1They are the resistance, inductance and current of the battery pack respectively;

[0020] R 2 , L 2 、i l2 are the resistance, inductance and current of the supercapacitor group respectively; V bat 、V sc are the voltage of the battery and the voltage of the supercapacitor respectively; U o is the DC bus output voltage, i o is the DC bus output current value, C o is the DC bus filter capacitor; μ 01 , μ 23 These are the switching signals for the two bidirectional DC / DC converters.

[0021] As a further explanation of the present invention, the specific steps of S2 are as follows:

[0022] S21. Regarding the implicit Runge-Kutta method, u is the time-varying data in the bidirectional DC / DC converter, so its spatial coordinate x is ignored. It is assumed that at T s Internally, the observable state u(t n ) and u(t n+1 ) between which there are q intermediate states u(t n+ci ), where t n+ci =t n +ci, ci∈[0,1], i=1,…,q, can be compared with the observable state u(t n ) and u(t n+1 ) coupling, and its functional relationship can be expressed as:

[0023]

[0024] where i,j=1,2,…,q,u i (t n )=u(t n ),u i (t n+1 )=u(t n+1 ), parameter {a ij , b j , c j} is provided by the Butcher table for a given order q, and its expression is as follows:

[0025]

[0026] S22, the end state u(t n+1 ) with the intermediate state u(t n+ci ) in the form of:

[0027]

[0028] S23. Use the implicit Runge-Kutta method to solve the system dynamics equations in S12.

[0029] As a further explanation of the present invention, the specific steps of S3 are as follows:

[0030] S31. Build a physical information neural network, which consists of two parts: data drive and physical model composed of deep neural network.

[0031] The input of the data driving part is the battery pack current i l1 (t n ), supercapacitor bank current i l2 (t n ), output voltage U o (t n ) and the switching signals μ of the two bidirectional DC / DC converters 01 and μ 23 The output of the data driving part is the intermediate state of the input signal [i l1 (t n+c1 ),…,i l1 (t n+cq ),i l2 (t n+c1 ),…,i l2 (t n+cq ), U o (t n+c1 ), …, U o (t n+cq )];

[0032] The physical model part is used to take the intermediate state output by the data driving part as input, and substitute the data into the forward equation formula (9) and the backward equation formula (6) according to the dynamic model of BDDC to obtain the predicted value i of the physical information neural network l1 '(t n ),i l1 '(t n+1 ),i l2 '(t n ),i l2 '(t n+1 ), U 'o (t n ), U 'o (t n+1 ), where the physical information neural network parameter set θ = {w, b, λ}, {w, b} is the weight and bias of the data-driven part;

[0033] S32, constructing the error function of the physical information neural network:

[0034]

[0035] S33, collecting the training sample set {i l1 、i l2 , U o , μ 01 , μ 23}, the parameters to be identified in the bidirectional DC / DC converter are used as the parameters of the data-driven part in the physical information neural network, all parameters in the physical information neural network are trained by the back-propagation method, and the supervised learning error finally output is propagated through the physical model part and the data-driven part respectively to update the learning gradient.

[0036] As a further explanation of the present invention, the specific steps of S4 are as follows:

[0037] The real-time sampling value i l1 、i l2 , U o Substitute it into the trained physical information neural network to realize online parameter identification of the bidirectional DC / DC converter.

[0038] Compared with the prior art, the present invention has the following beneficial technical effects:

[0039] The bidirectional DC / DC converter parameter identification method based on physical information machine learning provided by the present invention can realize the training of the neural network with fewer training samples, effectively reducing the calculation burden and enhancing the prediction transparency.

[0040] The technical solution of the present technical solution is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are used to provide further understanding of the present technical solution and constitute a part of the specification. Together with the embodiments of the present technical solution, they are used to explain the present technical solution and do not constitute a limitation to the present technical solution.

[0042] Figure 1 A topological structure diagram of a bidirectional DC / DC converter provided by the present invention;

[0043] Figure 2 is the observable state u(t n ),u(t n+1 ) and the intermediate state u(t n+ci ) between the explicit relationship diagram;

[0044] Figure 3This is a physical information neural network structure diagram of the bidirectional DC / DC converter provided by the present invention. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present technical solution are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present technical solution, and are not used to limit the present technical solution.

[0046] When a bidirectional DC / DC converter works in different application environments for a long time, its key components such as electrolytic capacitors, inductors, insulated gate bipolar transistors, metal oxide semiconductor field effect transistors, etc., under the combined effects of external environmental stress (such as operating temperature, humidity, vibration and radiation, etc.) and internal electrical stress, the parameter values ​​of the components will gradually deviate from the normal tolerance range and cause soft faults. In the early stage of soft faults, they are often ignored because the fault characteristics are weak and difficult to detect; as the deviation of components gradually increases, when the parameter deviation exceeds the failure threshold of the components, there will be a huge deviation between the actual circuit and the theoretical model, which will affect the control strategy effect of the system and the accuracy of the fault diagnosis method, resulting in misdiagnosis. Therefore, it is of great significance to perform parameter identification on electronic devices in bidirectional DC / DC converters.

[0047] Based on this, the present invention provides a bidirectional DC / DC converter parameter identification method based on physical information machine learning. l1 、i l2 , U o Substitute it into the trained physical information neural network to realize the online parameter identification of the bidirectional DC / DC converter. The neural network can be trained with fewer training samples, which effectively reduces the computational burden and enhances the prediction transparency.

[0048] Specifically, Figure 1 A topological structure diagram of a bidirectional DC / DC converter provided by the present invention; Figure 1 As shown, the hybrid energy system consists of a battery and a supercapacitor, two independent bidirectional DC / DC converters as interface circuits and a five-phase permanent magnet synchronous motor as a load.

[0049] Based on the above bidirectional DC / DC converter structure, the present invention proposes a bidirectional DC / DC converter parameter identification method based on physical information machine learning, which includes the following steps:

[0050] S1. Construct a system dynamics model of a bidirectional DC / DC converter. The bidirectional DC / DC converter is a converter used to control and regulate the charge and discharge current and input and output voltage between a battery and a supercapacitor.

[0051] Specifically, the system dynamics model building process in step S1 above specifically includes the following steps:

[0052] S11. General system dynamics model:

[0053] ut+N[u;λ]=0,x∈Ω,t∈[0,T] (1)

[0054] f=ut+N[u;λ] (2)

[0055] Where u=u(x,t) is the solution of the differential equation, x is the spatial coordinate, and t is the time coordinate.

[0056] S12. Based on the system dynamics model of S11, a system dynamics model of a bidirectional DC / DC converter is constructed:

[0057]

[0058] Where λ = {L 1 , L 2 , R 1 , R 2 , C o} is the component parameter to be identified.

[0059] In the formula, R 1 , L 1 、i l1 are the resistance, inductance and current of the battery pack respectively; R 2 , L 2 、i l2 are the resistance, inductance and current of the supercapacitor group respectively; V bat 、V sc are the voltage of the battery and the voltage of the supercapacitor respectively; U o is the DC bus output voltage, i o is the DC bus output current value, C o is the DC bus filter capacitor; μ 01 , μ 23 These are the switching signals for the two bidirectional DC / DC converters.

[0060] S2, such as Figure 2 As shown, the implicit Runge-Kutta method is used to construct the relationship between the intermediate state of the data and the initial state and the final state, and it is used in the system dynamics model to solve the unknown quantities.

[0061] The specific steps of the above step S2 are as follows:

[0062] S21. Regarding the implicit Runge-Kutta method, u is the time-varying data in the bidirectional DC / DC converter, so its spatial coordinate x is ignored. It is assumed that at Ts Internally, the observable state u(t n ) and u(t n+1 ) between which there are q intermediate states u(t n+ci ), where t n+ci =t n +ci, ci∈[0,1], i=1,…,q. In the framework of implicit Runge-Kutta method, it can be compared with the observable state u(t n ) and u(t n+1 ) coupling. Its functional relationship can be expressed as:

[0063]

[0064] where i,j=1,2,…,q,u i (t n )=u(t n ),u i (t n+1 )=u(t n+1 ), parameter {a ij , b j , c j} is provided by the Butcher table for a given order q, and its expression is as follows:

[0065]

[0066] S22, the end state u(t n+1 ) with the intermediate state u(t n+ci ) in the form of:

[0067]

[0068] S23. Use the implicit Runge-Kutta method to solve the system dynamics equations in S12.

[0069] S3, such as Figure 3 As shown, a physical information neural network is constructed, which consists of two parts: a data-driven part and a physical model composed of a deep neural network; wherein the physical model obtains a predicted value of the physical information neural network according to the output of the data-driven part using the solution method of step S2; the parameters to be identified in the bidirectional DC / DC converter are used as parameters of the data-driven part in the physical information neural network, and all parameters in the physical information neural network are trained by the back propagation method.

[0070] The specific steps of the above step S3 are as follows:

[0071] S31. Build a physical information neural network, which consists of two parts: data drive and physical model composed of a deep neural network. The input of the data drive part is the battery pack current i l1 (tn ), supercapacitor bank current i l2 (t n ), output voltage U o (t n ) and the switching signals μ of the two bidirectional DC / DC converters 01 and μ 23 The network performance can be effectively changed by adjusting the number of hidden layers and neurons in the deep neural network, so reasonable configuration can maximize the efficiency of resource utilization. The output of the data-driven part is the intermediate state of the input signal [i l1 (t n+c1 ),…,i l1 (t n+cq ),i l2 (t n+c1 ),…,i l2 (t n+cq ), U o (t n+c1 ), …, U o (t n+cq )].

[0072] The physical model part is used to take the intermediate state output by the data-driven part as input. According to the dynamic model of BDDC, the data is substituted into the forward equation formula (9) and the backward equation formula (6) to obtain the predicted value i of the physical information neural network. l1 '(t n ),i l1 '(t n+1 ),i l2 '(t n ),i l2 '(t n+1 ), U 'o (t n ), U 'o (t n+1 ). The physical information neural network parameter set θ = {w, b, λ}, {w, b} is the weight and bias of the data-driven part.

[0073] S32. Construct the error function of the physical information neural network:

[0074]

[0075] S33, collect the training sample set required by the physical information neural network {i l1 、i l2 , U o , μ 01 , μ 23}, the parameters to be identified in the bidirectional DC / DC converter are used as the parameters of the data-driven part in the physical information neural network, and all parameters in the physical information neural network are trained by the back-propagation method. The supervised learning error finally output is propagated through the physical model part and the data-driven part respectively to update the learning gradient.

[0076] The physical information neural network designed by the present invention seamlessly combines the physical information neural network with the physical model of the bidirectional DC / DC converter, effectively modulating the training phase so that it is enforced to the underlying physical principles and constraints of the converter.

[0077] Since the supervised learning error of the final output can be propagated through the physical model part and the data-driven part respectively to update the learning gradient, the chain rule of the typical back-propagation training scheme can be easily extended to the physical part, so that all parameters of the physical information neural network can be trained simultaneously through the standard back-propagation method.

[0078] S4. Use the trained physical information neural network to realize online parameter identification of the bidirectional DC / DC converter.

[0079] The real-time sampling value i l1 、i l2 , U o Substitute it into the trained physical information neural network to realize online parameter identification of bidirectional DC / DC converter.

[0080] Obviously, those skilled in the art can make various changes and modifications to the technical solution without departing from the spirit and scope of the technical solution. Thus, if these modifications and variations of the technical solution fall within the scope of the claims of the technical solution and their equivalents, the technical solution is also intended to include these modifications and variations.

Claims

1. A bidirectional DC / DC converter parameter identification method based on physical information machine learning, characterized in that: The method comprises the following steps: S1. Constructing a system dynamics model of a bidirectional DC / DC converter, wherein the bidirectional DC / DC converter is a converter for controlling and regulating the charge and discharge current and input and output voltage between a battery and a supercapacitor; S2. Using the implicit Runge-Kutta method, the relationship between the intermediate state of the data and the initial state and the final state is constructed, and the relationship is applied to the system dynamics model to solve the unknown quantity therein; S3, constructing a physical information neural network, which consists of two parts: a data-driven part and a physical model composed of a deep neural network; wherein the physical model uses the solution method of step S2 to obtain the predicted value of the physical information neural network according to the output of the data-driven part; Using the parameters to be identified in the bidirectional DC / DC converter as parameters of the data driving part in the physical information neural network, and training all parameters in the physical information neural network by back propagation method; S4. Using the trained physical information neural network to realize online parameter identification of the bidirectional DC / DC converter.

2. The bidirectional DC / DC converter parameter identification method based on physical information machine learning according to claim 1, characterized in that: The system dynamics model building process of step S1 specifically includes the following steps: S11. General system dynamics model: u t +N[u;λ]=0,x∈Ω,t∈[0,T] (1) f=u t +N[u;λ] (2) Where u = u (x, t) is the solution of the differential equation, x is the spatial coordinate, t is the time coordinate; S12. Construct a system dynamics model of the bidirectional DC / DC converter based on the system dynamics model of S11: Where λ = {L1, L2, R1, R2, C o } is the parameter of the component to be identified; In the formula, R1, L1, i l1 are the resistance, inductance and current of the battery pack respectively; R2, L2, i l2 are the resistance, inductance and current of the supercapacitor group respectively; V bat 、V sc are the voltage of the battery and the voltage of the supercapacitor respectively; U o is the DC bus output voltage, i o is the DC bus output current value, C o is the DC bus filter capacitor; μ 01 , μ 23 These are the switching signals for the two bidirectional DC / DC converters.

3. The bidirectional DC / DC converter parameter identification method based on physical information machine learning according to claim 2, characterized in that: The specific steps of S2 are as follows: S21. Regarding the implicit Runge-Kutta method, u is the time-varying data in the bidirectional DC / DC converter, so its spatial coordinate x is ignored. It is assumed that at T s Internally, the observable state u(t n ) and u(t n+1 ) between which there are q intermediate states u(t n+ci ), where t n+ci =t n +ci, ci∈[0,1], i=1,…,q, can be compared with the observable state u(t n ) and u(t n+1 ) coupling, and its functional relationship can be expressed as: where i,j=1,2,…,q,u i (t n )=u(t n ),u i (t n+1 )=u(t n+1 ), parameter {a ij , b j , c j } is provided by the Butcher table for a given order q, and its expression is as follows: S22, the end state u(t n+1 ) with the intermediate state u(t n+ci ) in the form of: S23. Use the implicit Runge-Kutta method to solve the system dynamics equations in S12.

4. The bidirectional DC / DC converter parameter identification method based on physical information machine learning according to claim 3, characterized in that: The specific steps of S3 are as follows: S31. Build a physical information neural network, which consists of two parts: data drive and physical model composed of deep neural network. The input of the data driving part is the battery pack current i l1 (t n ), supercapacitor bank current i l2 (t n ), output voltage U o (t n ) and the switching signals μ of the two bidirectional DC / DC converters 01 and μ 23 The output of the data driving part is the intermediate state of the input signal [i l1 (t n+c1 ),…,i l1 (t n+cq ),i l2 (t n+c1 ),…,i l2 (t n+cq ), U o (t n+c1 ), …, U o (t n+cq )]; The physical model part is used to take the intermediate state output by the data driving part as input, and substitute the data into the forward equation formula (9) and the backward equation formula (6) according to the dynamic model of BDDC to obtain the predicted value i of the physical information neural network l1 '(t n ),i l1 '(t n+1 ),i l2 '(t n ),i l2 '(t n+1 ), U' o (t n ), U' o (t n+1 ), where the physical information neural network parameter set θ = {w, b, λ}, {w, b} is the weight and bias of the data-driven part; S32, constructing the error function of the physical information neural network: S33, collecting the training sample set {i l1 、i l2 , U o , μ 01 , μ 23 }, the parameters to be identified in the bidirectional DC / DC converter are used as the parameters of the data-driven part in the physical information neural network, all parameters in the physical information neural network are trained by the back-propagation method, and the supervised learning error finally output is propagated through the physical model part and the data-driven part respectively to update the learning gradient.

5. The bidirectional DC / DC converter parameter identification method based on physical information machine learning according to claim 4, characterized in that: The specific steps of S4 are as follows: The real-time sampling value i l1 、i l2 , U o Substitute it into the trained physical information neural network to realize online parameter identification of the bidirectional DC / DC converter.

Images