LSTM (Long Short Term Memory)-based harmonic frequency coupling model optimization method for network construction type converter

Abstract

The invention discloses an LSTM (Long Short Term Memory)-based harmonic frequency coupling model optimization method for a network construction type converter, which considers the influence of a dead zone and establishes a converter model controlled by a virtual synchronous generator based on a harmonic state space method. In the power electronic modeling process, due to approximate simplification, switch neglecting and other nonlinear relations, the model is not accurate enough, meanwhile, in the harmonic state space model modeling process, system state variable parameters need to be considered, and the precision of the model affects harmonic transmission characteristics and coupling analysis, so that the modeling efficiency is greatly improved. Errors caused by power electronic modeling are reduced through an LSTM neural network algorithm, the harmonic amplitude estimation difference is reduced, state variable parameter values in a system model are accurately estimated, the model precision is further improved, and then the harmonic coupling condition under the dead zone influence is effectively and accurately analyzed.

Description

TECHNICAL FIELD

[0001] The application relates to the field of power electronics, in particular to an LSTM-based network-type converter harmonic frequency coupling model optimization method. BACKGROUND

[0002] Large-scale power electronic devices have become a significant feature of power grids. The penetration rate of power electronic equipment represented by photovoltaic inverters and electric vehicle charging piles continues to rise, making the power system exhibit "double high" (high proportion of renewable energy, high proportion of power electronic equipment) characteristics. Virtual synchronous generators have become a key technology for high-proportion new energy grid connection because they can simulate the inertia and damping characteristics of synchronous generators. The nonlinear control characteristics of virtual synchronous generators and the impedance characteristics of power grids interact with each other, resulting in complex harmonic interactions, and the switching frequency harmonics of power electronics form multi-frequency coupling. In addition, the existence of dead zones leads to an increase in some low-frequency harmonics, and under the accumulation of a large number of devices, the multi-frequency harmonic interaction phenomenon becomes more serious. This complex harmonic coupling not only reduces the power quality of the grid-connected converter output, but also seriously affects the stable operation of the power system. Current researches mostly use impedance frequency domain modeling methods, which have insufficient accuracy and cannot explicitly represent the wide frequency coupling phenomenon. The main problems currently faced are how to obtain a method that can accurately describe the wide frequency harmonic coupling phenomenon of the complete system, how to explore the degree of deterioration of the harmonic coupling phenomenon caused by low-frequency harmonic components due to dead zones, and how to effectively improve the model accuracy due to the approximation and simplification in the modeling process of power electronics. SUMMARY

[0003] 1. Technical problems to be solved:

[0004] To solve the above technical problems, the application provides an LSTM-based network-type converter harmonic frequency coupling model optimization method, which uses the harmonic state space method to construct a network-type converter harmonic state space model, derives a harmonic transfer function, uses an LSTM neural network to optimize variable parameters to improve model accuracy, analyzes the influence of dead zones on grid-connected current harmonics, explicitly analyzes the interaction between different frequency harmonics, and the coupling influence of background disturbance harmonic injection on grid-connected current under the influence of dead zones, and quantifies the harmonic coupling phenomenon of harmonic disturbance under the influence of dead zones.

[0005] 2. Technical solutions:

[0006] The LSTM-based network-type converter harmonic frequency coupling model optimization method comprises the following steps:

[0007] Step 1: System preset; the main circuit of the network-type converter adopts a three-phase LC-type converter grid connection, and the control strategy of the control module adopts a virtual synchronous generator control strategy;

[0008] Step two: based on the main circuit and control structure topology of the network-based converter, the harmonic state space method is used to obtain the variable expression relationship;

[0009] Step three: determine the system state variables and input variables, and then derive the harmonic state space equation to determine the harmonic state space frequency domain model; derive the harmonic transfer function through the harmonic state space model to obtain the harmonic transfer function matrix coefficient; the coupling characteristics between different frequency harmonics are used to represent the wideband coupling relationship between the grid harmonic voltage and the grid current under the influence of the dead zone;

[0010] Step four: use the LSTM neural network to learn the mapping relationship between the past time state value of the system variable and the steady-state value, output high-precision steady-state estimation value as the model variable parameter, optimize the model to improve the precision; use the optimized variable parameter to update the harmonic transfer function matrix, and then obtain the coupling characteristics between different frequency harmonics to represent the wideband coupling relationship between the grid harmonic voltage and the grid current under the influence of the dead zone;

[0011] Step five: multiply the frequency domain column vector of the input variable by the matrix to obtain the frequency domain value of part of the output variable, i.e. the grid current. Convert the harmonic frequency response to the time domain, and multiply each time-varying Fourier coefficient of the obtained signal by the corresponding rotation factor to obtain the time domain value of the signal.

[0012] Further, in step one, the three-phase LC converter includes a three-phase bridge rectifier circuit, three sets of inverter side filter inductors L f , parasitic resistors R f and filter capacitors C f connected in sequence;

[0013] The virtual synchronous generator link calculates the active and reactive components of the system through a low-pass filter by the power calculation module to the input common coupling point voltage and current dq axis components; the voltage frequency reference is obtained by adjusting the active and reactive components input by the voltage reference module according to the active and reactive reference values and the active and reactive droop coefficients and damping coefficients; the dq axis voltage V dqref is obtained by the voltage synthesis module and coordinate transformation as the voltage reference of the voltage and current double closed loop;

[0014] The voltage and current double closed loop control link realizes dq axis component decoupling through PI controller combined with decoupling control to obtain output current reference value and output voltage signal;

[0015] The PWM link generates switch tube driving signals s ωa , sωb ωc , and then control the switching action of the inverter bridge to obtain the output voltage of the inverter bridge arm The voltage of the output bridge arm is composed of the DC side voltage and the inverter bridge arm voltage s al and the dead zone harmonic voltage s D (t).

[0016] Further, step two specifically comprises the following steps:

[0017] S21: the variables in the variable expression of the main circuit obtained according to the same include the inverter side three-phase current i Labc , the common coupling point three-phase voltage v pccabc , and the grid three-phase current i gabc ; the corresponding variable expression in the dq reference coordinate system is as follows:

[0018]

[0019] v invd,q = s wd,q e dc + e dc s wd,q (1).

[0020] In the above formula, the subscript dq represents that the corresponding variable is obtained through coordinate transformation in the dq reference coordinate system; i Ldq represents the inverter side current in the dq reference coordinate system; v pccdq represents the common coupling point voltage in the dq reference coordinate system.

[0021] S22: after adding a disturbance to the variables of the main circuit in the dq reference coordinate system, the linearized time domain circuit equation of the following formula is obtained through harmonic linearization:

[0022]

[0023] Δv invd,q = s wd,q Δe dc + e dc Δs wd,q (2).

[0024] In the above formula, Δ represents a small disturbance signal of the corresponding variable.

[0025] S23: the variable relationship of the control module is as follows:

[0026]

[0027] In the above formula: k u , D u are the reactive power droop coefficient and the damping coefficient, respectively; k​p , D p P ref , Q ref , ω ref are active, reactive, frequency reference; P, Q are active, reactive; U od , U oq , I gd , I gq are dq-axis grid-connected voltage and dq-axis grid-connected current steady-state values, which are corresponding to i gdq , v pccdq system grid-connected current, grid-connected voltage signal after small perturbation to reach steady-state value again;

[0028] S24: The relationship of voltage signal under the influence of PWM dead zone is as follows:

[0029]

[0030] In the above formula: M is the modulation ratio; Jn is the Bessel function; n is the frequency of the corresponding harmonic; fc is the carrier frequency; td is the dead time; the corresponding frequency harmonic component is extracted from the above formula and superimposed to obtain the output voltage signal s wdq .

[0031] Further, step three specifically includes:

[0032] S31: Determine the system state variables and input variables, and the equation of the harmonic state space model of the system is the coupling effect under small perturbation, which is based on small signal and its equation form is:

[0033]

[0034] In the above formula, A and B represent state variable coefficient matrix and input variable coefficient matrix, respectively;

[0035] The state variable X and the input variable U in it are respectively:

[0036] X = [ΔI Ld , ΔI Lq , ΔU pccd , ΔU pccq , ΔI gd , ΔI gq , ΔP, ΔQ, ΔE, Δω', Δm1, Δm2, Δm3, Δm4] T

[0037] U = [ΔU ref , ΔU pcc , ΔP ref , ΔQ ref , Δω ref , ΔUgd , ΔU gq ] T (6);

[0038] In the above formula, ΔI Ld , ΔI Lq respectively represent small perturbation of dq-axis lower inverter side inductance current; ΔU pccd , ΔU pccq respectively represent small perturbation of dq-axis lower common coupling point voltage; ΔI gd , ΔI gq respectively represent small perturbation of dq-axis grid-connected current; Δω' represents small perturbation of system frequency; Δm1, Δm2, Δm3, Δm4 are respectively intermediate variables set up for constructing the model; S = diag(s + jwh), h ∈ [-b,..., 0,..., b], b represents the highest harmonic number considered;

[0039] S32: further derivation, the relationship between input variables and state variables can obtain harmonic transfer function matrix:

[0040] X = -(A - N) -1 BU

[0041] H = -(A - N) -1 B (7);

[0042] In the above formula: H is the harmonic transfer function matrix, which represents the relationship between input variables and output variables of different frequency harmonics; N is a diagonal matrix, the dimension depends on the harmonic number considered, for different harmonic number k, the element on the diagonal line is jkw;

[0043] S33: according to the relationship between input variables and output variables, the harmonic transfer function coefficient matrix H(s) of the following formula is derived, the matrix element H n,m between different frequency harmonics in the harmonic transfer function matrix is the coupling coefficient size between different frequency harmonics;

[0044]

[0045] In the above formula, H n,m is the coupling coefficient between input mth harmonic and output nth harmonic.

[0046] Further, in step four, since the harmonic transfer function matrix is derived from matrix A and B, A and B matrix are derived from formula (2) and (3), this process needs to consider i Ldq , v pccdq , i gdqSmall perturbation quantity, so that at different time and under the influence of small perturbation quantity of multiple converters, the harmonic components of each frequency after reaching steady state are taken as input, and through the LSTM neural network, the output high-precision steady-state estimation value is obtained; the steady-state estimation value obtained by the LSTM neural network optimization is substituted into the corresponding A, B matrix or the harmonic transfer function matrix derived from the A, B matrix, and the corresponding matrix elements are updated, so that the coupling coefficient matrix between the input variables and the state variables at steady state can be obtained.

[0047] Further, in step five, the harmonic frequency response is converted to the time domain, and the expression of the signal in the time domain is:

[0048] x(t)=Tr(t)X (9);

[0049] The state variable can be converted from the frequency domain to the time domain by the above formula; wherein Tr(t) is:

[0050]

[0051] Further, the system implemented includes: a network-type converter, a virtual synchronous generator, a voltage and current double-loop control, and a PWM link; the main circuit of the network-type converter is a three-phase LC-type converter grid, and the control strategy is a virtual synchronous generator control strategy; the virtual synchronous generator includes a power calculation module, a low-pass filter, a voltage reference module, a voltage synthesis module, and a coordinate transformation.

[0052] 3. Beneficial effects:

[0053] (1) The network-type converter harmonic frequency coupling model optimization method based on LSTM provided by the application analyzes the harmonic interaction characteristics of the network-type grid-connected converter in the wide frequency domain through the HSS model, combines the dead zone influence, optimizes the matrix elements by using the LSTM neural network, further improves the model precision, studies the harmonic interaction characteristics through the optimized model, and gives the dead zone influence characteristics and the harmonic interaction under the influence of the disturbance.

[0054] (2) The network-type converter harmonic frequency coupling model optimization method based on LSTM provided by the application reduces the error caused by power electronic modeling through the LSTM neural network algorithm, reduces the harmonic amplitude estimation difference, accurately estimates the state variable parameter value in the system model, further improves the model precision, and effectively and accurately analyzes the harmonic coupling under the influence of the dead zone. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 It is a schematic diagram of the main circuit structure of the network-type converter involved in the application;

[0056] Figure 2The control module structure schematic diagram of the grid-connected type converter in the application;

[0057] Figure 3 The LSTM network and optimization flowchart in the application;

[0058] Figure 4 The schematic diagram of the influence of the grid-connected harmonic current dead zone;

[0059] Figure 5 The variable reference value comparison chart before and after optimization in the verification example using the application;

[0060] Figure 6 The coupling coefficient result comparison chart before and after optimization in the verification example considering the influence of the dead zone using the application;

[0061] Figure 7 The grid-connected current coupling analysis comparison chart before and after optimization in the verification example using the application under the grid harmonic disturbance;

[0062] Figure 8 The grid-connected current FFT analysis comparison chart when using the application without harmonic and 2nd harmonic disturbance injection in the verification example;

[0063] Figure 9 The grid-connected current FFT analysis chart when using the application 2nd and 4th harmonic disturbance injection in the verification example. DETAILED DESCRIPTION

[0064] The application will be specifically described below in combination with the drawings.

[0065] The grid-connected type converter harmonic frequency coupling model optimization method based on LSTM includes the following steps:

[0066] Step one: system preset; the main circuit of the grid-connected type converter adopts a three-phase LC type converter grid connection, and the control strategy of the control module adopts a virtual synchronous generator control strategy;

[0067] Step two: based on the main circuit and control structure topology of the grid-connected type converter, the harmonic state space method is used to obtain the variable expression relationship;

[0068] Step three: determine the system state variable and input variable, and then deduce the harmonic state space equation to determine the harmonic state space frequency domain model; the harmonic transfer function is deduced through the harmonic state space model to obtain the harmonic transfer function matrix coefficient; the coupling characteristics between different frequency harmonics are used to represent the broadband coupling relationship between the grid harmonic voltage and the grid-connected current under the influence of the dead zone;

[0069] Step four: using the LSTM neural network, according to the state value of the system variable at the past time as the input, learning the mapping relationship between it and the steady-state value, outputting the high-precision steady-state estimation value as the model variable parameter, optimizing the model to improve the precision; using the optimized variable parameter, updating the harmonic transfer function matrix, and then obtaining the coupling characteristics between different frequency harmonics to represent the wideband coupling relationship between the grid harmonic voltage and the grid current under the influence of the dead zone;

[0070] Step five: by multiplying the frequency domain column vector of the input variable by the matrix, the frequency domain value of the partial output variable, i.e. the grid current, can be obtained, the harmonic frequency response is converted to the time domain, and each time-varying Fourier coefficient of the obtained signal is multiplied by the corresponding rotation factor, so that the time domain value of the signal can be obtained.

[0071] Further, in step one, the three-phase LC type converter includes a three-phase bridge rectifier circuit, three sets of inverter side filter inductors L f , parasitic resistors R f and filter capacitors C f connected in sequence; the virtual synchronous generator control includes a virtual synchronous generator link, a voltage and current double closed loop control link, and a PWM link;

[0072] The virtual synchronous generator link calculates the active and reactive components of the system through a low-pass filter by the power calculation module to the input common coupling point voltage and current dq axis components; the voltage frequency reference is obtained by adjusting the active and reactive components according to the active and reactive reference values and the active and reactive droop coefficients and damping coefficients through the voltage reference module; the dq axis voltage V dqref is obtained as the voltage reference of the voltage and current double closed loop through the voltage synthesis module and the coordinate transformation;

[0073] The voltage and current double closed loop control link realizes dq axis component decoupling through PI controller combined with decoupling control to obtain output current reference value and output voltage signal;

[0074] The PWM link generates switch tube driving signals s ωa , s ωb , s ωc through the PWM module to control the inverter bridge switching action, and obtains the converter bridge arm output voltage The voltage of the output bridge arm is composed of the direct current side voltage and the converter bridge arm voltage s al and the dead zone harmonic voltage s D (t).

[0075] Further, step two specifically includes the following steps:

[0076] S21: the variables in the main circuit variable expression obtained according to the same include the inverter-side three-phase current i Labc , the point-of-common coupling three-phase voltage v pccabc , the grid three-phase current i gabc ; the corresponding variable expression in the dq reference coordinate system is as follows:

[0077]

[0078]

[0079] In the above formula, the subscript dq represents that the corresponding variable is obtained through coordinate transformation in the dq reference coordinate system; i Ldq represents the inverter-side current in the dq reference coordinate system; v pccdq represents the point-of-common coupling voltage in the dq reference coordinate system;

[0080] S22: after adding a disturbance to the variables of the main circuit in the dq reference coordinate system, linearized time-domain circuit equations of the following formula are obtained through harmonic linearization:

[0081]

[0082] In the above formula, Δ represents a small disturbance signal of the corresponding variable;

[0083] S23: the variable relationship of the control module is as follows:

[0084]

[0085]

[0086] In the above formula: k u , D u are the reactive power droop coefficient and the damping coefficient, respectively; k p , D p are the active power droop coefficient and the damping coefficient, respectively; P ref , Q ref , ω ref are the active power, the reactive power, and the frequency reference, respectively; P and Q are the active power and the reactive power, respectively; U od , U oq , I gd , I gq are the dq-axis grid-connected voltage and the dq-axis grid-connected current steady-state values, which are the values of i gdq , v pccdq in the system grid-connected current and the grid-connected voltage signal after a small disturbance and then reaching a steady-state condition;

[0087] S24: the relationship of the voltage signal under the influence of the PWM dead zone is as follows:

[0088]

[0089] In the above formula: M is the modulation ratio; Jn is the Bessel function; n is the frequency of the corresponding harmonic; fc is the carrier frequency; td is the dead time; the corresponding frequency harmonic component superposition extracted from the above formula can obtain the output voltage signal s wdq .

[0090] Further, step three specifically includes:

[0091] S31: determine the system state variable and the input variable, the equation of the harmonic state space model of the system is the coupling effect under small disturbance, which is based on small signal, and the equation form is:

[0092]

[0093] In the above formula, A and B represent the state variable coefficient matrix and the input variable coefficient matrix respectively;

[0094] The state variable X and the input variable U in it are respectively:

[0095] X=[ΔI Ld ,ΔI Lq ,ΔU pccd ,ΔU pccq ,ΔI gd ,ΔI gq ,ΔP,ΔQ,ΔE,Δω',Δm1,Δm2,Δm3,Δm4] T

[0096] U=[ΔU ref ,ΔU pcc ,ΔP ref ,ΔQ ref ,Δω ref ,ΔU gd ,ΔU gq ] T (6);

[0097] In the above formula, ΔI Ld , ΔI Lq respectively represent the small disturbance of the inductance current on the dq axis on the inverter side; ΔU pccd , ΔU pccq respectively represent the small disturbance of the voltage on the dq axis on the common coupling point; ΔI gd , ΔI gq respectively represent the small disturbance of the grid-connected current on the dq axis; Δω' represents the small disturbance of the system frequency; Δm1, Δm2, Δm3, Δm4 are respectively intermediate variables set up for constructing the model; S=diag(s+jwh), h∈[-b, …, 0, …, b], b represents the highest harmonic number considered;

[0098] S32: Further derivation, the relationship between input variables and state variables can obtain the harmonic transfer function matrix:

[0099] X = -(A-N) -1 BU

[0100] H = -(A-N) -1 B (7);

[0101] In the above formula: H is the harmonic transfer function matrix, which represents the relationship between input variables and output variables of different frequencies; N is a diagonal matrix, the dimension depends on the number of harmonics considered, for different harmonic numbers k, the elements on the diagonal are jkw;

[0102] S33: According to the relationship between input variables and output variables, the harmonic transfer function coefficient matrix H(s) of the following formula is derived, and the matrix elements H n,m between different frequency harmonics in the harmonic transfer function matrix are the coupling coefficient sizes between different frequency harmonics;

[0103]

[0104] In the above formula, H n,m is the coupling coefficient between input mth harmonic and output nth harmonic.

[0105] Further, in step four, since the harmonic transfer function matrix is derived from matrix A and B, A and B matrix are derived from formula (2) and (3), and this process needs to consider the small perturbation of i Ldq , v pccdq , i gdq , therefore, the small perturbation of each frequency harmonic component at different time and environment is considered as input, and the high-precision steady-state estimated value is obtained through the LSTM neural network; The steady-state estimated value is obtained by optimizing the LSTM neural network, which is substituted into the corresponding A, B matrix or the harmonic transfer function matrix derived from A, B matrix, and the corresponding matrix elements are updated, so that the coupling coefficient matrix between input variables and state variables at steady state can be obtained.

[0106] Further, in step five, the harmonic frequency response is converted to the time domain, and the expression of the signal in the time domain is:

[0107] x(t) = Tr(t)X (9);

[0108] The state variable can be converted from the frequency domain to the time domain variable by the above formula; wherein Tr(t) is:

[0109]

[0110] Further, the implemented system comprises: a network-forming converter, a virtual synchronous generator, a voltage-current double closed loop control, and a PWM link; the main circuit of the network-forming converter is a three-phase LC converter connected to a power grid, and the control strategy thereof is a virtual synchronous generator control strategy; the virtual synchronous generator comprises a power calculation module, a low-pass filter, a voltage reference module, a voltage synthesis module, and a coordinate transformation.

[0111] Embodiment:

[0112] The embodiment details the specific principles and processes of the scheme.

[0113] The grid-connected converter controlled by the VSG of the scheme is shown in FIG. 1. Figure 1 E dc is a DC power supply; i dc is a DC current; Q1-Q6 are three-phase bridge arms; is a bridge arm output; L f , R f are a filter inductance and a parasitic resistance, and C is a filter capacitance; L g , R g are grid impedances; are an inverter-side current, a grid current, and a point-of-common-coupling voltage, respectively. The main circuit equation is shown in equation (1). A small perturbation is added to the main circuit variable, and a small signal equation is obtained. The harmonic linearization processing is performed on the steady-state operating point, and the linearized time-domain circuit equation of equation (2) is obtained.

[0114] The control module block diagram of the embodiment using the virtual synchronous generator control strategy is shown in FIG. 2. Figure 2 The control module of the VSG-controlled converter comprises a virtual synchronous generator link, a voltage-current double closed loop control link, and a PWM link. The virtual synchronous generator link comprises a power calculation module, a voltage reference module, and a voltage synthesis module. The input signals of the power calculation module are the point-of-common-coupling voltage and the grid current , which represent corresponding phases, and dq-axis components are obtained through abc / dq coordinate transformation. The system active and reactive components are calculated through a low-pass filter. The input of the reference voltage module is the active and reactive components, which are adjusted according to the active and reactive reference values P ref , Q ref , the active and reactive droop coefficients k u , k p , and the damping coefficients D u , D p to obtain the voltage frequency reference. The dq-axis V dqref, as the voltage current loop voltage reference. Voltage current double closed loop control link through PI controller combined with decoupling control to achieve dq axis component decoupling, voltage loop with V dref , V qref as input, compared with PCC voltage V pccd , V pccq , through PI controller adjustment, get output current reference value. In the current loop, using input current reference I d , I q , compared with actual inductance current I Ld , I Lq , and through cross coupling item combined with PCC voltage to realize current loop control, and then through coordinate transformation and voltage coefficient to get output voltage signal.

[0115] PWM link will generate switch tube drive signal s ωa , s ωb , s ωc through PWM module of voltage signal output by control loop, and then control inverter bridge switch action, get converter bridge arm output voltage Generally, the relationship between the dc side voltage obtained by the PWM link and the converter bridge arm voltage is s al is a general function. After considering the dead zone factor of equation (4), the low frequency harmonic component is increased accordingly, and the result is affected by the dead zone and the superposition of PWM signal:

[0116] V inv* = (s al + s D ) E dc .

[0117] As shown in the accompanying Figure 2 , the variable relationship of the control module is shown in equation (3). The variable relationship of the main circuit and the control module is obtained, and the equation of the matrix form of the harmonic state space model HSS model is shown in equation (5), wherein the state variables and input variables are shown in equation (6).

[0118] There are system state variable parameters i Ldq , V pccdq , i gdq in the harmonic transfer function matrix. The matrix elements containing state variables need to obtain their steady state values, but due to the approximation and simplification in the process of power electronic modeling, the accuracy of part of the model is not enough, and the steady state values of the state variables at different times under the influence of small disturbance exist fluctuations, which leads to errors, further affecting the model accuracy. Therefore, the LSTM neural network makes up for the gradient vanishing / explosion problem of traditional network, and through the synergistic effect of triple gating, it can adaptively retain key information and improve the modeling ability of complex sequence. Through iteration or direct mapping calculation, the steady state value of the input variable is more accurate, and the model accuracy is improved.

[0119] In the LSTM model, the update of the memory cell depends on the input gate, the forget gate, and its own state. The model first gets a new candidate value through the input gate and the current input, then determines the amount of old information to be retained by combining the previous memory cell state through the forget gate, and finally gets the updated memory cell state by adding the new candidate value to the retained old information. In addition to updating the memory cell content, the LSTM model also uses the output gate, the memory cell state, and the current input to calculate the new hidden state. The process first multiplies the memory cell state by the output gate to generate a candidate hidden state. Then the candidate hidden state is processed by an activation function (usually tanh) to limit its value to the range of -1 to 1. Finally, this result is multiplied by the result of the output gate to get the final hidden state. The specific optimization process is as shown in Figure 3 The LSTM has a forget gate f t , an input gate i t , a previous cell state h t-1 , and an output c t , h t , y t .

[0120] The state variable of the system at a certain time is obtained, and the LSTM neural network is used to optimize the system state variable to obtain the steady-state estimated value. Replace the original steady-state value (approximation) obtained by randomly selecting a certain time, optimize the model precision, and use the model with higher precision to study the harmonic interaction relationship. The elements H in the matrix can describe the relationship between different frequency harmonics, and the harmonic transfer function matrix H is:

[0121]

[0122] where H n,m are the coupling coefficients between the input mth harmonic and the output nth harmonic.

[0123] The steady-state estimated value obtained by neural network optimization is substituted into the harmonic transfer function matrix, and the corresponding matrix elements are updated, so that the coupling coefficient matrix between the input variable and the state variable at steady state can be obtained. By multiplying the frequency domain column vector of the input variable by the matrix, the frequency domain value of part of the output variable, i.e., the grid-connected current, can be obtained. The harmonic frequency response is converted to the time domain, and each time-varying Fourier coefficient of the obtained signal is multiplied by the corresponding rotation factor, so that the time domain value of the signal can be obtained.

[0124] x(t) = Tr(t)X (9)

[0125] The state variable can be converted from the frequency domain to the time domain by equation (80). Where Tr(t) is:

[0126]

[0127] Take h=2 as an example: The frequency harmonic size of the state variable is converted into a time domain variable through Tr(t).

[0128] The matrix elements corresponding to the position, i.e. the part of the input variable and the state variable, are extracted from the known harmonic transfer function matrix, and the coupling relationship of the required harmonic component is obtained. When the rth harmonic current is additionally output on the output side, the time domain expression is transformed into the Fourier series form as follows:

[0129]

[0130] In the formula: I gξ,r is the grid-connected current harmonic amplitude; θ ξ,r is the phase angle of the harmonic current; I gξ,(-r) , I gξ,(+r) are the Fourier coefficients of the rth grid-connected harmonic current respectively.

[0131] The grid-connected current harmonic amplitude i gξ,p (t) is expressed as follows:

[0132]

[0133] In the formula: V pccξ,p is the grid voltage amplitude; θ ξ,p is the grid voltage phase angle; H ξ(±r,±p) represents the coupling relationship between the grid-connected harmonic current with the Fourier coefficient of ±r and the grid harmonic voltage with the Fourier coefficient of ±p; is the coupling coefficient phase angle, the rth state variable is multiplied by the input variable corresponding to the hth to the hth harmonic through the harmonic transfer function matrix coefficient, assuming that the matrix coefficient is n, then the final state variable is the result sum of the subscript n+h=r.

[0134] In this scheme, the coupling characteristics between the background harmonic voltage and the grid-connected harmonic current can be studied through the harmonic transfer function matrix, the background harmonic is injected into the grid, the corresponding changes of the grid-connected harmonic current are analyzed, and the coupling response of the background harmonic injection under the influence of the dead zone is analyzed. Further, the system parameters can be adjusted, the influence of the parameters on the coupling is analyzed through the change of the harmonic transfer function matrix elements, and the system parameters can be optimized and the influence size of different frequency harmonics can be considered for targeted suppression.

[0135] Verification example

[0136] In order to verify the superiority and feasibility of the scheme, the embodiment of the application is simulated in Matlab / Simulink, and the system parameters can be set as shown in Table 1. The system is simulated under different conditions respectively.

[0137] Table 1 Grid-connected converter parameters

[0138]

[0139] According to the following dead zone harmonic voltage signal formula, the dead zone will generate 1, 3, 5 and other odd multiple frequency harmonic components under the influence of the fundamental frequency. If there is a disturbance of a certain frequency in the system, the disturbance will couple to generate harmonic components of other frequencies under the influence of the dead zone, that is, 2, 6, 10 and other frequency harmonics will be generated under the influence of the double frequency coupling, resulting in more complex coupling. The size of the harmonic generated under the influence of the dead zone is proportional to the dead zone time, and the higher the frequency, the smaller the harmonic amplitude generated.

[0140]

[0141] Figure 4 The figure is the influence diagram of the grid-connected harmonic current dead zone. Different dead zone times (0, 2, 3, 6 μs) are set for the low-frequency harmonic current size of the grid-connected point. Since the model is based on the dq axis coordinate system, the increase of odd harmonics is reflected in 2, 4, 6 and other frequencies. It can be seen that due to the influence of the dead zone, the grid-connected current amplitude increases, and with the increase of the dead zone time, the harmonic components of the corresponding frequency will also increase, verifying the influence of the dead zone on the harmonics and the trend of the influence of the dead zone time. Therefore, the PWM link is further combined in the scheme to strengthen the coupling between harmonics.

[0142] Figure 5 The figure is the comparison between the approximate value of the system state variable at each time before and after the optimization of the model using the scheme and the steady-state estimated value obtained by the neural network optimization. The state variable in the figure is V pccd As an example, it can be seen from Figure 5 that the interval of the approximate value of the system state variable disturbance before optimization is [309.12, 312.41], and the fluctuation amplitude is 3.29; the estimated value of the state variable after neural network optimization is about 310.93, and the amplitude of the upper and lower deviation is about ±0.2. By comparing the size of the steady-state value of the common point voltage before and after optimization, it is found that the error is significantly reduced, the HSS model and the harmonic transfer function matrix updated are more accurate, the accuracy will be higher, and the subsequent harmonic coupling response analysis will be more accurate.

[0143] Figure 6For the analysis of coupling coefficients before and after the optimization of the model considering the influence of dead zone, the matrix is truncated based on 10 times frequency, and the coupling coefficient matrix is obtained to clearly show the harmonic coupling under the influence of dead zone. As can be seen from the comparison of figures (a) and (b), after the optimization of the state variable parameters in the harmonic transfer function matrix by the neural network, the 2 times frequency disturbance component increases the 10 times harmonic coupling coefficient, and the 3 times frequency and 5 times frequency voltage increases and decreases the coupling degree of the same frequency current, with an error of 0.0392 and 0.0185, and the accuracy increases. Taking 1, 2, 3, and 4 times frequency in the optimized harmonic coupling coefficient in figure (b) as an example, they are 0.4765, 0.287, 0.1637, and 0.1437 respectively. Considering the dead zone factor in the model, its main influence is on low frequency odd harmonics, corresponding to the grid-connected current side in the figure, under the injection of base frequency disturbance, the grid-connected current will generate corresponding 3, 5, 7, and 9 frequency harmonics due to coupling, which are 0.0802, 0.0642, 0.0595, and 0.0429 respectively; under the injection of 2 times frequency disturbance, the influence degree caused by the coupling of the grid-connected current side is 0.0599 and 0.0447; under the influence of 3 times base frequency disturbance, the coupling degree of the 9 times harmonic generated by the grid-connected current is 0.0328; it can also be known from the dead zone relationship that the injection of base frequency disturbance will lead to the generation of corresponding odd harmonics in voltage harmonics, therefore, under different times frequency disturbance, the voltage will also generate odd harmonics under the influence of times frequency, due to the shortcomings of the control strategy, it will lead to the harmonic coupling of different frequencies on the grid-connected current side, increase the THD of the grid-connected current, cause the distortion of the grid-connected current, and affect the stability of the system. The comparison of the harmonic transfer function matrix proves that the HSS model optimized by the neural network has improved to a certain extent in accuracy, which is beneficial to future research and analysis.

[0144] Figure 7 For the comparison of the grid harmonic disturbance before and after the optimization of the grid-connected current coupling, the left and right graphs in the figure are the analysis of 2 and 3 times frequency harmonic voltage injection, at this time, Figure 7 the base frequency amplitude of the left graph is 0.4176, and due to the influence of dead zone coupling, the 3, 5, and 7 odd harmonics also increase; Figure 7It can be seen from the right figure of that the harmonic coupling will produce odd harmonic components corresponding to 2 times frequency, that is, the 6th, 10th, etc. harmonic content will increase. Taking the 3rd harmonic at 2 times frequency as an example, the harmonic amplitudes before and after optimization are 0.907 and 0.723, respectively, and the simulation value is 0.685, and the errors are 0.222 and 0.038, respectively; taking the 2nd harmonic at 3 times frequency as an example, the harmonic amplitudes before and after optimization are 0.286 and 0.273, respectively, and the simulation value is 0.269, and the errors are 0.017 and 0.004, respectively; it can be seen that the error is greatly reduced, and the error range after optimization is reduced by about 80%. The model calculation value and the simulation value prove that there is a coupling relationship between the grid background harmonic disturbance and the grid-connected current, and the coupling degree is related to the dead zone effect. At the same time, the comparison of the harmonic current amplitudes before and after optimization with the simulation value shows that the harmonic amplitudes of the grid-connected current calculated by the optimized model are closer to the simulation value, and the proposed neural network optimization model method can more accurately provide the steady-state value of the parameters, improve the accuracy of the model, and facilitate the analysis of the harmonic coupling characteristics.

[0145] Figure 8 、 9 are the FFT analysis of the grid-connected current when there is no harmonic and 2nd harmonic disturbance injection in the present scheme, and the analysis when 2 and 4 times frequency disturbances are injected at the same time, Figure 8 (a) and (b) show the analysis results without disturbance, and the harmonic content in the grid-connected current is small, only the harmonic content at 100 Hz and 250 Hz reaches 0.0157% and 0.0133%, and the harmonic content at other frequencies is less than 0.08%, and the THD of the grid-connected current is 0.42%. Figure 8 (c) and (d) show that the FFT analysis results show that the superimposed grid background harmonic makes the harmonic content at 2, 6, 10, 14, 18, etc. frequencies increase, among which the harmonic content at 2, 10, 14 frequencies increases to 4.243%, 0.203%, and 0.412%, and the harmonic content at 6 and 18 frequencies reaches 0.431% and 0.402, and the background harmonic disturbance under the influence of the dead zone will produce corresponding harmonic components. At the same time Figure 9The 2 and 4 times frequency disturbance injection results are shown, at this time, the harmonic proportion corresponding to the disturbance injection frequency rises to 4.273%, 2.034%. In addition, the harmonic coupling effect causes the harmonic proportion of 300, 500, 600, 700 and other frequencies to increase by 0.809%, 0.612%, 0.391%, 0.408% respectively, and the current THD rises to 5.17%. Usually, the harmonic distortion rate of grid-connected current should be below 3%, and the harmonic distortion rate of grid-connected current has exceeded the requirement. The analysis of the experimental operation results proves that the dead zone influence leads to further intensified harmonic interaction. Therefore, the method can calculate relatively accurate harmonic coupling response results under the condition of changes in simulation simulation environment, parameters and the like, and the subsequent harmonic suppression strategy method is formulated according to the harmonic coupling situation.

[0146] Although the present application has been disclosed with the preferred embodiments as above, they are not intended to limit the present application, and any person skilled in the art can make various changes or modifications without departing from the spirit and scope of the present application, and therefore the protection scope of the present application should be defined by the protection scope of the claims of the present application.

Claims

1. An optimization method for harmonic frequency coupling model of a network converter based on LSTM, characterized in that: Includes the following steps: Step 1: System preset; The main circuit of the grid-connected converter adopts a three-phase LC converter connected to the grid, and the control strategy of its control module adopts a virtual synchronous generator control strategy; Step 2: Based on the main circuit and control structure topology of the grid-type converter, the variable expression relationship is obtained using the harmonic state-space method; Step 3: Determine the system state variables and input variables, then derive the harmonic state-space equations and determine the harmonic state-space frequency domain model; derive the harmonic transfer function through the harmonic state-space model and obtain the harmonic transfer function matrix coefficients; The coupling characteristics between harmonics of different frequencies were investigated to characterize the broadband coupling relationship between grid harmonic voltage and grid-connected current under the influence of dead zone. Step 4: Using an LSTM neural network, the system learns the mapping relationship between the past state values ​​of the system variables and the steady-state values, and outputs high-precision steady-state estimates as model variable parameters to optimize the model and improve accuracy. By using the optimized variable parameters, the harmonic transfer function matrix is ​​updated, and the coupling characteristics between different frequency harmonics are obtained to characterize the broadband coupling relationship between grid harmonic voltage and grid-connected current under the influence of dead zone. Step 5: By multiplying the frequency domain column vectors of the input variables by the matrix, the frequency domain values ​​of some output variables, namely the grid-connected current, can be obtained. The harmonic frequency response is then converted to the time domain, and each time-varying Fourier coefficient of the obtained signal is multiplied by the corresponding rotation factor to obtain the time domain value of the signal.

2. The LSTM-based harmonic frequency coupling model optimization method for grid-type converters according to claim 1, characterized in that: In step one, the three-phase LC converter includes a three-phase bridge rectifier circuit and three sets of inverter-side filter inductors L connected in sequence. f Parasitic resistance R f and filter capacitor C f The virtual synchronous generator control includes a virtual synchronous generator stage, a voltage and current dual closed-loop control stage, and a PWM stage. The virtual synchronous generator stage calculates the active and reactive power components of the system by using a power calculation module to process the input common coupling point voltage and current dq-axis components through a low-pass filter; it then uses a voltage reference module to input the active and reactive power components and adjusts them according to the active and reactive power reference values, droop coefficients, and damping coefficients to obtain the voltage frequency reference; finally, it uses a voltage synthesis module and coordinate transformation to obtain the dq-axis voltage V. dqref , serving as a voltage reference for the voltage and current dual closed-loop; The voltage and current dual closed-loop control loop uses a PI controller combined with decoupling control to achieve decoupling of the dq axis components, thereby obtaining the output current reference value and the output voltage signal. The PWM stage generates a switching transistor drive signal s from the voltage signal output by the voltage and current dual closed-loop control stage through the PWM module. ωa s ωb s ωc This, in turn, controls the inverter bridge switch to operate, thereby obtaining the converter bridge arm output voltage. The voltage of the output bridge arm is determined by the DC side voltage and the converter bridge arm voltage s. al and dead zone harmonic voltages D (t) constitutes.

3. The LSTM-based harmonic frequency coupling model optimization method for grid-type converters according to claim 2, characterized in that: Step two specifically includes the following steps: S21: The variables in the main circuit variable expression obtained from it include the inverter side three-phase current i. Labc Three-phase voltage at common coupling point v pccabc Three-phase current i in the power grid gabc The variable; the corresponding variable expression in the dq reference coordinate system is as follows: v invd,q =s wd,q yes dc +e dc s wd,q (1); In the above formula, the subscript dq indicates that the corresponding variable in the dq reference coordinate system is obtained through coordinate transformation; i Ldq This represents the inverter-side current in the dq reference coordinate system; v pccdq This represents the voltage at the common coupling point in the dq reference coordinate system; S22: After adding perturbations to the variables of the main circuit in the dq reference coordinate system, the linearized time-domain circuit equation is obtained through harmonic linearization as follows: Δv invd,q =s wd,q Δe dc +e dc Δs wd,q (2); In the above formula, Δ represents the small perturbation signal of the corresponding variable; S23: The variable relationships of the control module are as follows: In the above formula: k u D u These are the reactive power loop droop coefficient and damping coefficient, respectively; k p D p P represents the active power loop droop factor and damping factor. ref Q ref ω ref These represent active power, reactive power, and frequency reference, respectively; P and Q represent active power and reactive power, respectively; U od U oq I gd I gq These are the steady-state values ​​of the grid-connected voltage and current along the dq axis, respectively, which correspond to i in equation (2). gdq v pccdq The system grid-connected current and grid-connected voltage signals return to their steady-state values ​​after a small disturbance. S24: The relationship between the voltage signal and the dead time of the PWM stage is as follows: In the above formula: M is the modulation ratio; Jn is the Bessel function; n is the frequency of the corresponding harmonic; fc is the carrier frequency; td is the dead time; the output voltage signal s can be obtained by extracting the corresponding frequency harmonic components from the above formula and superimposing them. wdq .

4. The LSTM-based harmonic frequency coupling model optimization method for grid-type converters according to claim 3, characterized in that: Step three specifically includes: S31: Determine the system state variables and input variables. The equations of the system's harmonic state-space model are based on the coupling effects under small disturbances and small signals. The equation form is as follows: In the above formula, A and B represent the state variable coefficient matrix and the input variable coefficient matrix, respectively; The state variable X and the input variable U are respectively: X=[ΔI Ld ,D I Lq ,D U pccd ,D U pccq ,D I gd ,D I gq ,ΔP,ΔQ,ΔE,Δω',Δm1,Δm2,Δm3,Δm4] T U=[ΔU ref ,D U pcc ,ΔP ref ,ΔQ ref ,Here ref ,D U gd ,D U gq ] T (6); In the above formula, ΔI Ld ΔI Lq These represent the small disturbances in the inverter-side inductor current along the dq axis; ΔU pccd , ΔU pccq These represent the small perturbations in the voltage at the common coupling point along the dq axes; ΔI gd ΔI gq Δm1, Δm2, Δm3, and Δm4 represent the small disturbance of the grid-connected current under the dq axis; Δω' represents the small disturbance of the system frequency; Δm1, Δm2, Δm3, and Δm4 are the intermediate variables set up for building the model; S = diag(s + jwh), h ∈ [-b, ..., 0, ..., b], b represents the highest harmonic order considered; S32: Further derivation yields the harmonic transfer function matrix by obtaining the relationship between the input variables and the state variables. X=-(AN) -1 THIS H=-(A-N) -1 B (7); In the above formula: H is the harmonic transfer function matrix, which represents the relationship between different frequency harmonics between the input and output variables; N is a diagonal matrix whose dimension depends on the harmonic order being considered. For different harmonic orders k, the elements on the diagonal are jkw. S33: Based on the relationship between input and output variables, the harmonic transfer function coefficient matrix H(s) is derived, where H is the matrix element H between different frequency harmonics in the harmonic transfer function matrix. n,m This refers to the magnitude of the coupling coefficient between harmonics of different frequencies; In the above formula, H n,m The coupling coefficient between the input m-th harmonic and the output n-th harmonic is given.

5. The LSTM-based harmonic frequency coupling model optimization method for grid-type converters according to claim 4, characterized in that: In step four, since the harmonic transfer function matrix is ​​derived from matrices A and B, and matrices A and B are derived from equations (2) and (3), this process needs to consider i. Ldq v pccdq i gdq Since the small disturbances are considered, the harmonic components of each frequency after the small disturbances of multiple converters reach steady state under different times and environments are used as inputs. Through the LSTM neural network, a high-precision steady-state estimate is obtained. The steady-state estimate is obtained by optimizing the LSTM neural network, and then substituted into the corresponding A and B matrices or the harmonic transfer function matrix derived from the A and B matrices to update the corresponding matrix elements. The coupling coefficient matrix between the input variables and the state variables in steady state can then be obtained.

6. The LSTM-based harmonic frequency coupling model optimization method for grid-type converters according to claim 5, characterized in that: In step five, the harmonic frequency response is converted to the time domain, and the expression for the signal in the time domain is obtained as follows: x(t)=Tr(t)X (9); The state variable in the frequency domain can be converted into a time domain variable using the above formula; where Tr(t) is:

7. The LSTM-based harmonic frequency coupling model optimization method for grid-type converters according to claim 1, characterized in that: The system includes: a grid-connected converter, a virtual synchronous generator, dual closed-loop voltage and current control, and a PWM circuit; the main circuit of the grid-connected converter is a three-phase LC converter connected to the grid, and its control strategy is a virtual synchronous generator control strategy; the virtual synchronous generator includes a power calculation module, a low-pass filter, a voltage reference module, a voltage synthesis module, and a coordinate transformation.

Images