Nonlinear frequency domain modeling method for sampling and modulation links of H-bridge converter

Abstract

The invention discloses a nonlinear frequency domain modeling method for sampling and modulation links of an H-bridge converter, and relates to the technical field of PWM (Pulse Width Modulation) of the H-bridge converter. According to the method, harmonic analysis is expanded to the whole frequency range by adopting a sampling link combined with AE and a modulation link combined with SE; a control model is divided into an FIP and an FCP, and an artificial neural network is designed to predict the output of the FCP so as to process complex nonlinearity.

Description

Technical Field

[0001] The invention relates to the technical field of H-bridge converter PWM technology, and in particular to a nonlinear frequency domain modeling method for sampling and modulation links of an H-bridge converter. Background Art

[0002] At present, the harmonic analysis research of grid-connected converters mainly focuses on the frequency range below the Nyquist sampling frequency (NSF) or the switching frequency (SF), making full use of linearization methods. However, in some practical high-power industrial applications, such as electric railways or wind power generation, the harmonic problem is very challenging in a wide frequency range up to several thousand Hertz. This is due to the aliasing effect and sideband effect, which can be ignored at low frequencies, but at high frequencies, due to frequency coupling and nonlinear characteristics, it will produce a big difference.

[0003] Therefore, it is necessary to find a nonlinear characteristic characterization method to expand the harmonic analysis frequency to the entire frequency range, so as to comprehensively and clearly analyze the harmonic performance of the grid-connected converter from aspects such as harmonic instability, harmonic resonance, harmonic overvoltage, and filter design. The defects and shortcomings of the existing technology are mainly:

[0004] 1)f hr Setting it to 2.5kHz is f sw =f sam = one-quarter of 10kHz, indicating that it may not be sufficient for low f sw and high f hr high power applications.

[0005] 2) You can get (0,f sw / 4) has a wider damping region, but for practical applications where the frequency is much higher than f sw The width of the high-order harmonic resonance is also insufficient.

[0006] 3) The frequency range of positive equivalent resistance is only f sw / 2.

[0007] It can be seen that the above three methods cannot handle high-power GCC. sw Only a few hundred Hz, and f hr It may reach several thousand Hertz.

[0008] 4) is based on f sam =f sw or 2f sam =f sw The assumption is made and the PWM process is simplified to a proportional component. Although the applicable frequency of GCC is expanded, in practical applications, due to the multi-channel or modular cascade structure of GCC, f sam and f swThe multiplication relationship between them can be 0.5 times, one times, two times, four times or even any multiple.

[0009] It can be seen that the above four methods are limited in their applicable frequency range. Summary of the invention

[0010] In view of the above-mentioned deficiencies in the prior art, the present invention provides a sampling link combined with AE and a modulation link combined with SE, making them universal, and by using a regression method based on an artificial neural network to map the input-output relationship of the pulse width modulation process, the harmonic analysis frequency is expanded to the entire frequency range, and the harmonic performance of GCC can be comprehensively and clearly analyzed from aspects such as harmonic instability, harmonic resonance, harmonic overvoltage, and filter design. In order to achieve the above-mentioned invention purpose, the technical solution adopted by the present invention is:

[0011] The present invention provides a nonlinear frequency domain modeling method for sampling and modulation links of an H-bridge converter, which comprises the following steps:

[0012] S1, collecting voltage and current signals of the grid-connected converter;

[0013] S2, using the AE sampling link to pre-process the collected signal;

[0014] S3, using SE modulation link to modulate the pre-processed signal;

[0015] S4, dividing the control model into linear and nonlinear parts;

[0016] S5, establish nonlinear characteristics;

[0017] S6, using artificial neural network to fit the frequency coupling part FCP spectrum transfer function;

[0018] S7, forming a control model in the entire frequency range including the frequency independent part FIP ​​and the frequency coupled part FCP;

[0019] S8, complete wide frequency domain analysis;

[0020] Preferably, in the above-mentioned nonlinear frequency domain modeling method of the H-bridge converter sampling and modulation link, the step S2 takes into account the need to address the FCE under the influence of the aliasing effect (AE), and models the sampling process as a multi-input-multi-output matrix.

[0021] Preferably, in the above-mentioned nonlinear frequency domain modeling method of the sampling and modulation link of the H-bridge converter, the step S3 considers the influence of the sideband effect (SE), and uses a natural number group (m, k) to array the sideband harmonics, and determines the sideband harmonic frequency range of the converter output voltage vc at different modulation wave frequencies;

[0022] Preferably, in the above-mentioned nonlinear frequency domain modeling method for sampling and modulation of the H-bridge converter, the step S4 further comprises:

[0023] S4-1, process the low frequency part and the high frequency part separately;

[0024] S4-2, the aliasing effect and sideband effect of the low-frequency part can be ignored and regarded as a linear part, using the traditional ZOH model;

[0025] S4-3, the high frequency part is frequency coupled and is regarded as a nonlinear part.

[0026] Preferably, in the above-mentioned nonlinear frequency domain modeling method for sampling and modulation of the H-bridge converter, the step S6 further comprises:

[0027] S6-1, generating input and output data of FCP, wherein SIMULINK, a visual simulation tool in MATLAB, can be selected to generate the input and output data of FCP;

[0028] S6-2, then, at 50~5000hz(f sw Characterize the input and output data in a frequency field (2.5 times of );

[0029] S6-3, training data, verification data and test data, select the elastic back propagation algorithm, which can select different learning rates according to different weights to accelerate the convergence of the algorithm;

[0030] S6-4. Use ANN to directly replace the PWM link to predict the harmonic spectrum output in a wide frequency range.

[0031] Preferably, in the above-mentioned nonlinear frequency domain modeling method for the sampling and modulation link of the H-bridge converter, the specific process of step S7 is:

[0032] In FIP, the signal transmission process is linear, and the n×n matrix in the transfer function is simplified to a simple diagonal matrix; the control model is generated by combining the transfer function of S6.

[0033] The beneficial effects of the present invention are as follows: the present invention proposes a sampling link combined with AE and a modulation link combined with SE, making it universal, and designs an artificial neural network to predict the output of FCP, dealing with complex nonlinear problems. The present invention extends the harmonic analysis frequency to the entire frequency range, and can comprehensively and clearly analyze the harmonic performance of GCC from aspects such as harmonic instability, harmonic resonance, harmonic overvoltage, and filter design. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the modeling method;

[0035] Figure 2 It is a schematic diagram of the artificial neural network processing process;

[0036] Figure 3 This is a general control block diagram for the full frequency range;

[0037] Figure 4 The signal flow diagram in the vehicle-grid coupling system when considering the aliasing effect and sideband effect;

[0038] Figure 5 To derivate the sampling transfer function considering multiple frequencies;

[0039] Figure 6 The figure is a schematic diagram of frequency spectrum comparison between actual value and predicted value using an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0041] Figure 1 It is a nonlinear frequency domain modeling method for the sampling and modulation links of an H-bridge converter. Figure 2 It is a schematic diagram of the artificial neural network processing process; Figure 2 In the neural network, the frequency-independent part is the linear part, and the frequency-coupled part is the nonlinear part. The artificial neural network is divided into three layers: input layer, intermediate layer, and output layer to process the nonlinear part.

[0042] Figure 3 This is a general control block diagram for the full frequency range; Figure 3 Where Hcv(s) is the capacitor voltage controller, Hci(s) is the current loop controller, Hsam(s) is the sampling link transfer function, Hffv(s) is the grid voltage feedforward transfer function, Gd(s) is the total system delay, Hpwm(s) is the NSPWM link transfer function, and Hfbi(s) is the current sampling sensor correction transfer function.

[0043] The signals of the entire control system can be divided into three parts: fundamental component, low-order harmonic component, and high-order harmonic component. Among them, the fundamental component is the most important part of all signals. For example, the control of reactive current and active current will produce some high-frequency components through the aliasing effect and sideband effect. Taking the vehicle-grid coupling system as an example, the interaction between the double frequency fluctuation of the DC side voltage and the phase-locked loop (PLL) will produce low-order harmonics such as the 3rd, 5th, and 7th order. The low-order harmonics themselves usually do not cause overvoltage in the traction network.

[0044] This part of low-order harmonic components will also generate some high-frequency components after the action of AE and SE. Since the current command tracking transient error determined by the locomotive running condition, traction network noise, and sensor noise mainly contain high-frequency components, low-frequency components will be generated after the action of AE and SE. All harmonic components will be Figure 4 The signal flows shown continue to interact, with low-frequency components generating high-frequency components and high-frequency components generating low-frequency components, and harmonic interference will eventually cover a very wide frequency range. Among them:

[0045] 1) The source of fundamental wave component, which is the most important part of the signal we want to process, such as reactive and active current control. After AE and SE, high-frequency components will be generated.

[0046] 2) The voltage 2x frequency fluctuation and PLL generate 2, 3, 4, 5... and other low-order harmonics. The DC capacitor voltage fluctuation and AC current harmonics are repeatedly generated. This is due to the power balance restrictions on the DC and AC sides of the capacitor. The low-order harmonics themselves usually do not generate overvoltage.

[0047] 3) Current command step tracking transient error. This is determined by the locomotive running conditions. At the moment of transient change, it is impossible to keep up completely, and the error shows a rich component.

[0048] 4) Background noise, including traction network noise, sensor noise, and interference from other vehicles transmitted through the traction network. These harmonic sources will further enrich the harmonic spectrum according to the aliasing of sampling and the sideband effect of modulation. Low frequencies will generate high frequencies, and high frequencies will generate low frequencies. When passing through the entire loop feedback system including the internal impedance of the traction network, it is possible to generate overvoltage at a certain frequency.

[0049] Figure 4 The signal flow diagram in the vehicle-grid coupling system when considering the aliasing effect and sideband effect. The aliasing effect and sideband effect couple the single-frequency basic signals in the low-frequency model together to form a set of basic signals.

[0050] The present embodiment is described in detail below with reference to the accompanying drawings.

[0051] like Figure 1 As shown, a nonlinear frequency domain modeling method for sampling and modulation of an H-bridge converter is provided, comprising the following steps:

[0052] S1, collecting voltage and current signals of the grid-connected converter;

[0053] S2, using the aliasing effect AE sampling link to pre-process the collected signal;

[0054] S3, using the sideband effect SE modulation link to modulate the pre-processed signal;

[0055] S4, dividing the control model into linear and nonlinear parts;

[0056] S5, establishing nonlinear features for the nonlinear part;

[0057] S6. Use artificial neural network to fit the FCP spectrum transfer function;

[0058] S7, forming a control model over the entire frequency range including FIP and FCP;

[0059] S8, complete wide frequency domain analysis;

[0060] In the step S2, when the AE sampling link is used to pre-process the collected signal, it is considered that under the influence of the aliasing effect (AE), the frequency coupling effect FCE needs to be addressed, and the sampling process is modeled as a multi-input-multi-output matrix, as shown in formula (1). In formula (1), the main diagonal a ii The elements on the diagonal are the self-coupling coefficients, which represent the input-output relationship of signals with the same frequency; the elements on the off-diagonal lines are the mutual coupling coefficients, which represent the input-output relationship of signals at different frequencies.

[0061]

[0062] In order to determine A = (a ij ) n×n According to the superposition theorem, the input signal is set to have only one non-zero coefficient each time, that is:

[0063]

[0064] In the formula, a kl refers to row k and column l of A, r k (output signal) refers to the kth row before A, x i (Input signal) refers to the i-th row of the matrix following A, x l Refers to the lth row (other rows) of the matrix after A. In the formula, ω sam is the sampling frequency, 2πfsam is the angular frequency, T sis the sampling period. x(t) is the original signal in the time domain, r(jω) is the output signal of ZOH, and the frequency of the input signal is w x .

[0065] where x l (ω l )=δ(ω-ω l ), now we take solving the first component and the second component conduction coefficient as an example to illustrate. Figure 5 As shown in (a), first let the frequency of the input signal be ω x The amplitude of the component is 1, and the other components are 0, that is, δ(ω-ω x ), according to the signal x after the impact string p The expression of (t) and the transfer function of the zero-order holder can directly obtain the amplitude and phase angle curves of each component in the output signal r(t). Figure 5 (a) shows |r(ω)| and ∠r(ω). According to Figure 5 The component represented by the bold curve in (a) can be used to obtain the autocoupler conduction coefficient a 11 , and the mutual coupling conduction coefficient a 12 、a 13 ...The first column element of the conduction matrix in equation (1) can be obtained. Similarly, when solving the second column element of the matrix, let the frequency be ω sam -ω x The amplitude of the harmonic component of is 1, and the other components are 0, such as Figure 5 (b) Repeat the above steps until all elements of the entire matrix are obtained, as shown in formula (2).

[0066]

[0067] In the formula

[0068]

[0069] is the transfer function of the zero-order holder, Ts is the sampling period, ω sam is the sampling frequency, ω is the angular frequency, T s is the sampling period.

[0070] Among them, in step S3, the SE modulation link is used to modulate the preprocessed signal, that is, considering the influence of the sideband effect (SE), the natural number group (m, k) is used to perform array analysis on the sideband harmonics to determine the sideband harmonic frequency range of the converter output voltage vc at different modulation wave frequencies. Considering the influence of the sideband effect (SE), the natural number group (m, k) is used to perform array analysis on the sideband harmonics. The main purpose is to determine the sideband harmonic frequency range of the converter output voltage. Steps for determining the sideband harmonic frequency range: (1) Determine the carrier frequency fc and the modulation frequency fm; (2) Calculate the sideband frequency: The sideband frequency is given by the formula Δf=|f c ±k·f m |Determine, where k is a natural number representing the serial number of the sideband; the positive and negative signs correspond to the upper sideband and the lower sideband respectively; (3) Select the range of m and k 4. Analyze the frequency range of the sideband harmonics: Through the above steps, the frequency range of the sideband harmonics of the converter output voltage under a certain modulation wave frequency can be determined. For different modulation wave frequencies, repeat this process to obtain a series of sideband harmonic frequency ranges.

[0071] Step S4 further comprises:

[0072] S4-1, process the low frequency part and the high frequency part separately;

[0073] S4-2, the aliasing effect and sideband effect of the low-frequency part can be ignored and regarded as a linear part, using the traditional ZOH model;

[0074] S4-3, high frequency part frequency coupling, regarded as nonlinear part

[0075] Wherein, the step S6 of using an artificial neural network to fit the FCP spectrum transfer function further comprises:

[0076] S6-1. Generate input and output data of FCP, wherein SIMULINK, a visual simulation tool in MATLAB, may be selected to generate the FCP input and output data.

[0077] S6-2, then, at 50~5000hz(f sw The input and output data are characterized in a frequency field (2.5 times of ). The ratio of training data, validation data, and test data is set to 5:3:2, and the elastic back propagation algorithm is selected. This algorithm can select different learning rates according to different weights to accelerate the convergence of the algorithm.

[0078] The number of hidden layers is set to 1, and the number of neurons is set to 212. The final neural network has a fitting error of 0.02 and a generalization error of 0.0693.

[0079] S6-3, training data, verification data and test data, select the resilient backpropagation algorithm, which can select different learning rates according to different weights to accelerate the convergence of the algorithm. Resilient Backpropagation (Rprop) is an optimization algorithm for training artificial neural networks. Compared with the traditional gradient descent algorithm, the Rprop algorithm accelerates the network training process by dynamically adjusting the learning rate while maintaining stability. The core idea of ​​the Rprop algorithm is to update the weights according to the sign of the gradient, rather than the absolute value of the gradient, which can avoid the problem of gradient disappearance or gradient explosion.

[0080] In the Rprop algorithm, each weight has an independent update value that is adjusted based on the sign change of the gradient. If the sign of the gradient is the same for several consecutive iterations, the update value is increased to speed up the learning process; if the sign of the gradient changes, the update value is decreased to prevent the learning process from overshooting. The Rprop algorithm usually sets two fixed multipliers, one for increasing the update value (elastic acceleration) and the other for decreasing the update value (elastic deceleration).

[0081] The advantage of the Rprop algorithm is that it can automatically adjust the learning rate without the user having to manually set the learning rate, which makes the algorithm more adaptable and robust on different problems. In addition, the Rprop algorithm usually converges faster than the standard back-propagation algorithm.

[0082] In practical applications, the Rprop algorithm has been used to train a variety of neural network models, including tasks such as handwritten digit recognition, and has been implemented on platforms such as MATLAB.

[0083] S6-4. Use artificial neural network (ANN) to directly replace the PWM link to predict the harmonic spectrum output in a wide frequency range.

[0084] The specific process of forming the control model in the entire frequency range including FIP and FCP in step S7 is as follows:

[0085] In FIP, the signal transmission process is linear, and the n×n matrix in the transfer function is simplified to a simple diagonal matrix; the control model is generated by combining the transfer function of S6.

[0086] In one embodiment of the present invention, f sam =f sw Set to 2kHz. A total of 100,000 sets of simulation data were obtained. Then, at 50~5000hz(f swThe input and output data are characterized in a frequency field of 2.5 times of the training data, verification data and test data are set to 5:3:2, and the elastic back propagation algorithm is selected. The algorithm can select different learning rates according to different weights to accelerate the convergence of the algorithm. The number of hidden layers is set to 1, and the number of neurons is set to 212. The final neural network fitting error is 0.02, and the generalization error is 0.0693. In order to more clearly illustrate the fitting effect of the artificial neural network, the present invention explains the frequency spectrum comparison between the actual value and the predicted value of a sample, such as Figure 6 shown.

[0087] It can be seen that the use of artificial neural network to predict FCP spectrum has higher accuracy.

[0088] In the present invention, artificial neural network (ANN), grid-connected converter (GCC), harmonics, resonant overvoltage, sampling frequency, sideband effect (SE), switching frequency (SF), grid-connected converters (GCCs), Nyquist sampling frequency (NSF),

[0089] Switching frequency (SF), digital signal processor (DSP), artificial neural network (ANN), pulse width modulation (PWM),

[0090] Common grid point (PCC), feed forward (FF), proportional integral (PI), frequency coupling effect (FCE), time sampling PWM (RSPWM), zero order holder (ZOH), frequency independent part (FIP), frequency coupling part (FCP), natural sampling PWM (NSPWM), low pass filter (LPF), aliasing effect (AE), sideband effect (SE), phase locked loop (PLL), ideal power supply (ve), internal impedance (Ze), converter current (ig), PCC voltage (vg PCC), converter filter impedance (Zc), converter output voltage (vc), reference current (iref), sampling frequency (fsam), converter control frequency (fc), carrier frequency / switching frequency (fsw), calculation delay (tc), signal update delay (tu), Nyquist sampling frequency (fNy), GCC harmonic resonance frequency (fhr), frequency domain original signal (x(jω)), sampling angular frequency (ωsam), Nyquist sampling angular frequency (ωNy), time domain original signal (x(t)), pulse sequence (p(t)), ideal LPF (h(t)), pulse sequence The formed signal (xp(t)), the output signal of ZOH (xr(jω)), the signal used to execute the control strategy (v(jω)), the arm output voltage of the dual-level converter (uan(t)), the harmonic coefficients (Cmn, Bmn), the angular frequency of the modulation wave (ωm), the angular frequency of the carrier (ωca), the converter DC capacitor voltage (Udc), the modulation wave frequency (fm), the harmonic frequency of VC (fh), the baseband modulation ratio (dr), the input-output relationship of RSPWM (Q(jω)), the PCC voltage limit (Vlim), and the proportional coefficient (Kp) are obtained by the double Fourier series analytical method.

[0091] In summary, the present invention proposes a sampling link combined with AE and a modulation link combined with SE to make it universal, and designs an artificial neural network to predict the output of FCP and handle complex nonlinear problems. The present invention extends the harmonic analysis frequency to the entire frequency range, and can comprehensively and clearly analyze the harmonic performance of GCC from aspects such as harmonic instability, harmonic resonance, harmonic overvoltage, and filter design.

Claims

1. A nonlinear frequency domain modeling method for sampling and modulation of an H-bridge converter, characterized in that: The following steps are involved: S1, collecting voltage and current signals of the grid-connected converter; S2, using the aliasing effect AE sampling link to pre-process the collected signal; S3, using the sideband effect SE modulation link to modulate the pre-processed signal; S4, dividing the control model into linear and nonlinear parts; S5, establish nonlinear characteristics; S6. Use artificial neural network to fit the FCP spectrum transfer function; S7, forming a control model over the entire frequency range including FIP and FCP; S8. Use the control model to complete wide frequency domain analysis.

2. The nonlinear frequency domain modeling method for sampling and modulation of an H-bridge converter according to claim 1 is characterized in that: In step S2, it is considered that under the influence of the aliasing effect, the FCE needs to be addressed, and the sampling process is modeled as a multi-input-multi-output matrix.

3. The nonlinear frequency domain modeling method for sampling and modulation of an H-bridge converter according to claim 2 is characterized in that: In step S3, the sideband harmonics are arrayed using a natural number group (m, k) under the influence of the sideband effect SE, and the sideband harmonic frequency range after judgment of the converter output voltage vc at different modulation wave frequencies is determined.

4. The nonlinear frequency domain modeling method for sampling and modulation of an H-bridge converter according to claim 1, characterized in that: In step S4, the low-frequency part and the high-frequency part are processed separately, wherein the aliasing effect and the sideband effect of the low-frequency part can be ignored and is regarded as a linear part and processed using a traditional ZOH model; the high-frequency part is frequency coupled and is regarded as a nonlinear part.

5. The nonlinear frequency domain modeling method for sampling and modulation of an H-bridge converter according to claim 3, characterized in that: The step S6 further comprises: S6-1. Use SIMULINK to generate input and output data for FCP; S6-2. Then, the input and output data are characterized in a frequency field of 50 to 5000 Hz. S6-3, training data, verification data and test data, select the elastic back propagation algorithm, which can select different learning rates according to different weights to accelerate the convergence of the algorithm. S6-4. Use artificial neural network to directly replace PWM link to predict harmonic spectrum output in a wide frequency range.

6. The nonlinear frequency domain modeling method for sampling and modulation of an H-bridge converter according to claim 4, characterized in that: The specific process of step S7 is as follows: In FIP, the signal transmission process is linear, and the n×n matrix in the transfer function is simplified to a simple diagonal matrix; the control model is generated by combining the transfer function of S6.

Images